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 adjust_managed_page_count(page, nr_pages);
1299 } else {
1300 for (loop = 0; loop < nr_pages; loop++, p++) {
1301 __ClearPageReserved(p);
1302 set_page_count(p, 0);
1303 }
1304
1305 /* memblock adjusts totalram_pages() manually. */
1306 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1307 }
1308
1309 if (page_contains_unaccepted(page, order)) {
1310 if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1311 return;
1312
1313 accept_memory(page_to_phys(page), PAGE_SIZE << order);
1314 }
1315
1316 /*
1317 * Bypass PCP and place fresh pages right to the tail, primarily
1318 * relevant for memory onlining.
1319 */
1320 __free_pages_ok(page, order, FPI_TO_TAIL);
1321 }
1322
1323 /*
1324 * Check that the whole (or subset of) a pageblock given by the interval of
1325 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1326 * with the migration of free compaction scanner.
1327 *
1328 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1329 *
1330 * It's possible on some configurations to have a setup like node0 node1 node0
1331 * i.e. it's possible that all pages within a zones range of pages do not
1332 * belong to a single zone. We assume that a border between node0 and node1
1333 * can occur within a single pageblock, but not a node0 node1 node0
1334 * interleaving within a single pageblock. It is therefore sufficient to check
1335 * the first and last page of a pageblock and avoid checking each individual
1336 * page in a pageblock.
1337 *
1338 * Note: the function may return non-NULL struct page even for a page block
1339 * which contains a memory hole (i.e. there is no physical memory for a subset
1340 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1341 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1342 * even though the start pfn is online and valid. This should be safe most of
1343 * the time because struct pages are still initialized via init_unavailable_range()
1344 * and pfn walkers shouldn't touch any physical memory range for which they do
1345 * not recognize any specific metadata in struct pages.
1346 */
__pageblock_pfn_to_page(unsigned long start_pfn,unsigned long end_pfn,struct zone * zone)1347 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1348 unsigned long end_pfn, struct zone *zone)
1349 {
1350 struct page *start_page;
1351 struct page *end_page;
1352
1353 /* end_pfn is one past the range we are checking */
1354 end_pfn--;
1355
1356 if (!pfn_valid(end_pfn))
1357 return NULL;
1358
1359 start_page = pfn_to_online_page(start_pfn);
1360 if (!start_page)
1361 return NULL;
1362
1363 if (page_zone(start_page) != zone)
1364 return NULL;
1365
1366 end_page = pfn_to_page(end_pfn);
1367
1368 /* This gives a shorter code than deriving page_zone(end_page) */
1369 if (page_zone_id(start_page) != page_zone_id(end_page))
1370 return NULL;
1371
1372 return start_page;
1373 }
1374
1375 /*
1376 * The order of subdivision here is critical for the IO subsystem.
1377 * Please do not alter this order without good reasons and regression
1378 * testing. Specifically, as large blocks of memory are subdivided,
1379 * the order in which smaller blocks are delivered depends on the order
1380 * they're subdivided in this function. This is the primary factor
1381 * influencing the order in which pages are delivered to the IO
1382 * subsystem according to empirical testing, and this is also justified
1383 * by considering the behavior of a buddy system containing a single
1384 * large block of memory acted on by a series of small allocations.
1385 * This behavior is a critical factor in sglist merging's success.
1386 *
1387 * -- nyc
1388 */
expand(struct zone * zone,struct page * page,int low,int high,int migratetype)1389 static inline unsigned int expand(struct zone *zone, struct page *page, int low,
1390 int high, int migratetype)
1391 {
1392 unsigned int size = 1 << high;
1393 unsigned int nr_added = 0;
1394
1395 while (high > low) {
1396 high--;
1397 size >>= 1;
1398 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1399
1400 /*
1401 * Mark as guard pages (or page), that will allow to
1402 * merge back to allocator when buddy will be freed.
1403 * Corresponding page table entries will not be touched,
1404 * pages will stay not present in virtual address space
1405 */
1406 if (set_page_guard(zone, &page[size], high))
1407 continue;
1408
1409 __add_to_free_list(&page[size], zone, high, migratetype, false);
1410 set_buddy_order(&page[size], high);
1411 nr_added += size;
1412 }
1413
1414 return nr_added;
1415 }
1416
page_del_and_expand(struct zone * zone,struct page * page,int low,int high,int migratetype)1417 static __always_inline void page_del_and_expand(struct zone *zone,
1418 struct page *page, int low,
1419 int high, int migratetype)
1420 {
1421 int nr_pages = 1 << high;
1422
1423 __del_page_from_free_list(page, zone, high, migratetype);
1424 nr_pages -= expand(zone, page, low, high, migratetype);
1425 account_freepages(zone, -nr_pages, migratetype);
1426 }
1427
check_new_page_bad(struct page * page)1428 static void check_new_page_bad(struct page *page)
1429 {
1430 if (unlikely(page->flags & __PG_HWPOISON)) {
1431 /* Don't complain about hwpoisoned pages */
1432 if (PageBuddy(page))
1433 __ClearPageBuddy(page);
1434 return;
1435 }
1436
1437 bad_page(page,
1438 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1439 }
1440
1441 /*
1442 * This page is about to be returned from the page allocator
1443 */
check_new_page(struct page * page)1444 static bool check_new_page(struct page *page)
1445 {
1446 if (likely(page_expected_state(page,
1447 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1448 return false;
1449
1450 check_new_page_bad(page);
1451 return true;
1452 }
1453
check_new_pages(struct page * page,unsigned int order)1454 static inline bool check_new_pages(struct page *page, unsigned int order)
1455 {
1456 if (is_check_pages_enabled()) {
1457 for (int i = 0; i < (1 << order); i++) {
1458 struct page *p = page + i;
1459
1460 if (check_new_page(p))
1461 return true;
1462 }
1463 }
1464
1465 return false;
1466 }
1467
should_skip_kasan_unpoison(gfp_t flags)1468 static inline bool should_skip_kasan_unpoison(gfp_t flags)
1469 {
1470 /* Don't skip if a software KASAN mode is enabled. */
1471 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1472 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1473 return false;
1474
1475 /* Skip, if hardware tag-based KASAN is not enabled. */
1476 if (!kasan_hw_tags_enabled())
1477 return true;
1478
1479 /*
1480 * With hardware tag-based KASAN enabled, skip if this has been
1481 * requested via __GFP_SKIP_KASAN.
1482 */
1483 return flags & __GFP_SKIP_KASAN;
1484 }
1485
should_skip_init(gfp_t flags)1486 static inline bool should_skip_init(gfp_t flags)
1487 {
1488 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1489 if (!kasan_hw_tags_enabled())
1490 return false;
1491
1492 /* For hardware tag-based KASAN, skip if requested. */
1493 return (flags & __GFP_SKIP_ZERO);
1494 }
1495
post_alloc_hook(struct page * page,unsigned int order,gfp_t gfp_flags)1496 inline void post_alloc_hook(struct page *page, unsigned int order,
1497 gfp_t gfp_flags)
1498 {
1499 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1500 !should_skip_init(gfp_flags);
1501 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1502 int i;
1503
1504 set_page_private(page, 0);
1505
1506 arch_alloc_page(page, order);
1507 debug_pagealloc_map_pages(page, 1 << order);
1508
1509 /*
1510 * Page unpoisoning must happen before memory initialization.
1511 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1512 * allocations and the page unpoisoning code will complain.
1513 */
1514 kernel_unpoison_pages(page, 1 << order);
1515
1516 /*
1517 * As memory initialization might be integrated into KASAN,
1518 * KASAN unpoisoning and memory initializion code must be
1519 * kept together to avoid discrepancies in behavior.
1520 */
1521
1522 /*
1523 * If memory tags should be zeroed
1524 * (which happens only when memory should be initialized as well).
1525 */
1526 if (zero_tags) {
1527 /* Initialize both memory and memory tags. */
1528 for (i = 0; i != 1 << order; ++i)
1529 tag_clear_highpage(page + i);
1530
1531 /* Take note that memory was initialized by the loop above. */
1532 init = false;
1533 }
1534 if (!should_skip_kasan_unpoison(gfp_flags) &&
1535 kasan_unpoison_pages(page, order, init)) {
1536 /* Take note that memory was initialized by KASAN. */
1537 if (kasan_has_integrated_init())
1538 init = false;
1539 } else {
1540 /*
1541 * If memory tags have not been set by KASAN, reset the page
1542 * tags to ensure page_address() dereferencing does not fault.
1543 */
1544 for (i = 0; i != 1 << order; ++i)
1545 page_kasan_tag_reset(page + i);
1546 }
1547 /* If memory is still not initialized, initialize it now. */
1548 if (init)
1549 kernel_init_pages(page, 1 << order);
1550
1551 set_page_owner(page, order, gfp_flags);
1552 page_table_check_alloc(page, order);
1553 pgalloc_tag_add(page, current, 1 << order);
1554 }
1555
prep_new_page(struct page * page,unsigned int order,gfp_t gfp_flags,unsigned int alloc_flags)1556 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1557 unsigned int alloc_flags)
1558 {
1559 post_alloc_hook(page, order, gfp_flags);
1560
1561 if (order && (gfp_flags & __GFP_COMP))
1562 prep_compound_page(page, order);
1563
1564 /*
1565 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1566 * allocate the page. The expectation is that the caller is taking
1567 * steps that will free more memory. The caller should avoid the page
1568 * being used for !PFMEMALLOC purposes.
1569 */
1570 if (alloc_flags & ALLOC_NO_WATERMARKS)
1571 set_page_pfmemalloc(page);
1572 else
1573 clear_page_pfmemalloc(page);
1574 }
1575
1576 /*
1577 * Go through the free lists for the given migratetype and remove
1578 * the smallest available page from the freelists
1579 */
1580 static __always_inline
__rmqueue_smallest(struct zone * zone,unsigned int order,int migratetype)1581 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1582 int migratetype)
1583 {
1584 unsigned int current_order;
1585 struct free_area *area;
1586 struct page *page;
1587
1588 /* Find a page of the appropriate size in the preferred list */
1589 for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
1590 area = &(zone->free_area[current_order]);
1591 page = get_page_from_free_area(area, migratetype);
1592 if (!page)
1593 continue;
1594
1595 page_del_and_expand(zone, page, order, current_order,
1596 migratetype);
1597 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1598 pcp_allowed_order(order) &&
1599 migratetype < MIGRATE_PCPTYPES);
1600 return page;
1601 }
1602
1603 return NULL;
1604 }
1605
1606
1607 /*
1608 * This array describes the order lists are fallen back to when
1609 * the free lists for the desirable migrate type are depleted
1610 *
1611 * The other migratetypes do not have fallbacks.
1612 */
1613 static int fallbacks[MIGRATE_PCPTYPES][MIGRATE_PCPTYPES - 1] = {
1614 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1615 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1616 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1617 };
1618
1619 #ifdef CONFIG_CMA
__rmqueue_cma_fallback(struct zone * zone,unsigned int order)1620 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1621 unsigned int order)
1622 {
1623 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1624 }
1625 #else
__rmqueue_cma_fallback(struct zone * zone,unsigned int order)1626 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1627 unsigned int order) { return NULL; }
1628 #endif
1629
1630 /*
1631 * Change the type of a block and move all its free pages to that
1632 * type's freelist.
1633 */
__move_freepages_block(struct zone * zone,unsigned long start_pfn,int old_mt,int new_mt)1634 static int __move_freepages_block(struct zone *zone, unsigned long start_pfn,
1635 int old_mt, int new_mt)
1636 {
1637 struct page *page;
1638 unsigned long pfn, end_pfn;
1639 unsigned int order;
1640 int pages_moved = 0;
1641
1642 VM_WARN_ON(start_pfn & (pageblock_nr_pages - 1));
1643 end_pfn = pageblock_end_pfn(start_pfn);
1644
1645 for (pfn = start_pfn; pfn < end_pfn;) {
1646 page = pfn_to_page(pfn);
1647 if (!PageBuddy(page)) {
1648 pfn++;
1649 continue;
1650 }
1651
1652 /* Make sure we are not inadvertently changing nodes */
1653 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1654 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1655
1656 order = buddy_order(page);
1657
1658 move_to_free_list(page, zone, order, old_mt, new_mt);
1659
1660 pfn += 1 << order;
1661 pages_moved += 1 << order;
1662 }
1663
1664 set_pageblock_migratetype(pfn_to_page(start_pfn), new_mt);
1665
1666 return pages_moved;
1667 }
1668
prep_move_freepages_block(struct zone * zone,struct page * page,unsigned long * start_pfn,int * num_free,int * num_movable)1669 static bool prep_move_freepages_block(struct zone *zone, struct page *page,
1670 unsigned long *start_pfn,
1671 int *num_free, int *num_movable)
1672 {
1673 unsigned long pfn, start, end;
1674
1675 pfn = page_to_pfn(page);
1676 start = pageblock_start_pfn(pfn);
1677 end = pageblock_end_pfn(pfn);
1678
1679 /*
1680 * The caller only has the lock for @zone, don't touch ranges
1681 * that straddle into other zones. While we could move part of
1682 * the range that's inside the zone, this call is usually
1683 * accompanied by other operations such as migratetype updates
1684 * which also should be locked.
1685 */
1686 if (!zone_spans_pfn(zone, start))
1687 return false;
1688 if (!zone_spans_pfn(zone, end - 1))
1689 return false;
1690
1691 *start_pfn = start;
1692
1693 if (num_free) {
1694 *num_free = 0;
1695 *num_movable = 0;
1696 for (pfn = start; pfn < end;) {
1697 page = pfn_to_page(pfn);
1698 if (PageBuddy(page)) {
1699 int nr = 1 << buddy_order(page);
1700
1701 *num_free += nr;
1702 pfn += nr;
1703 continue;
1704 }
1705 /*
1706 * We assume that pages that could be isolated for
1707 * migration are movable. But we don't actually try
1708 * isolating, as that would be expensive.
1709 */
1710 if (PageLRU(page) || __PageMovable(page))
1711 (*num_movable)++;
1712 pfn++;
1713 }
1714 }
1715
1716 return true;
1717 }
1718
move_freepages_block(struct zone * zone,struct page * page,int old_mt,int new_mt)1719 static int move_freepages_block(struct zone *zone, struct page *page,
1720 int old_mt, int new_mt)
1721 {
1722 unsigned long start_pfn;
1723
1724 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
1725 return -1;
1726
1727 return __move_freepages_block(zone, start_pfn, old_mt, new_mt);
1728 }
1729
1730 #ifdef CONFIG_MEMORY_ISOLATION
1731 /* Look for a buddy that straddles start_pfn */
find_large_buddy(unsigned long start_pfn)1732 static unsigned long find_large_buddy(unsigned long start_pfn)
1733 {
1734 int order = 0;
1735 struct page *page;
1736 unsigned long pfn = start_pfn;
1737
1738 while (!PageBuddy(page = pfn_to_page(pfn))) {
1739 /* Nothing found */
1740 if (++order > MAX_PAGE_ORDER)
1741 return start_pfn;
1742 pfn &= ~0UL << order;
1743 }
1744
1745 /*
1746 * Found a preceding buddy, but does it straddle?
1747 */
1748 if (pfn + (1 << buddy_order(page)) > start_pfn)
1749 return pfn;
1750
1751 /* Nothing found */
1752 return start_pfn;
1753 }
1754
1755 /**
1756 * move_freepages_block_isolate - move free pages in block for page isolation
1757 * @zone: the zone
1758 * @page: the pageblock page
1759 * @migratetype: migratetype to set on the pageblock
1760 *
1761 * This is similar to move_freepages_block(), but handles the special
1762 * case encountered in page isolation, where the block of interest
1763 * might be part of a larger buddy spanning multiple pageblocks.
1764 *
1765 * Unlike the regular page allocator path, which moves pages while
1766 * stealing buddies off the freelist, page isolation is interested in
1767 * arbitrary pfn ranges that may have overlapping buddies on both ends.
1768 *
1769 * This function handles that. Straddling buddies are split into
1770 * individual pageblocks. Only the block of interest is moved.
1771 *
1772 * Returns %true if pages could be moved, %false otherwise.
1773 */
move_freepages_block_isolate(struct zone * zone,struct page * page,int migratetype)1774 bool move_freepages_block_isolate(struct zone *zone, struct page *page,
1775 int migratetype)
1776 {
1777 unsigned long start_pfn, pfn;
1778
1779 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
1780 return false;
1781
1782 /* No splits needed if buddies can't span multiple blocks */
1783 if (pageblock_order == MAX_PAGE_ORDER)
1784 goto move;
1785
1786 /* We're a tail block in a larger buddy */
1787 pfn = find_large_buddy(start_pfn);
1788 if (pfn != start_pfn) {
1789 struct page *buddy = pfn_to_page(pfn);
1790 int order = buddy_order(buddy);
1791
1792 del_page_from_free_list(buddy, zone, order,
1793 get_pfnblock_migratetype(buddy, pfn));
1794 set_pageblock_migratetype(page, migratetype);
1795 split_large_buddy(zone, buddy, pfn, order, FPI_NONE);
1796 return true;
1797 }
1798
1799 /* We're the starting block of a larger buddy */
1800 if (PageBuddy(page) && buddy_order(page) > pageblock_order) {
1801 int order = buddy_order(page);
1802
1803 del_page_from_free_list(page, zone, order,
1804 get_pfnblock_migratetype(page, pfn));
1805 set_pageblock_migratetype(page, migratetype);
1806 split_large_buddy(zone, page, pfn, order, FPI_NONE);
1807 return true;
1808 }
1809 move:
1810 __move_freepages_block(zone, start_pfn,
1811 get_pfnblock_migratetype(page, start_pfn),
1812 migratetype);
1813 return true;
1814 }
1815 #endif /* CONFIG_MEMORY_ISOLATION */
1816
change_pageblock_range(struct page * pageblock_page,int start_order,int migratetype)1817 static void change_pageblock_range(struct page *pageblock_page,
1818 int start_order, int migratetype)
1819 {
1820 int nr_pageblocks = 1 << (start_order - pageblock_order);
1821
1822 while (nr_pageblocks--) {
1823 set_pageblock_migratetype(pageblock_page, migratetype);
1824 pageblock_page += pageblock_nr_pages;
1825 }
1826 }
1827
1828 /*
1829 * When we are falling back to another migratetype during allocation, try to
1830 * steal extra free pages from the same pageblocks to satisfy further
1831 * allocations, instead of polluting multiple pageblocks.
1832 *
1833 * If we are stealing a relatively large buddy page, it is likely there will
1834 * be more free pages in the pageblock, so try to steal them all. For
1835 * reclaimable and unmovable allocations, we steal regardless of page size,
1836 * as fragmentation caused by those allocations polluting movable pageblocks
1837 * is worse than movable allocations stealing from unmovable and reclaimable
1838 * pageblocks.
1839 */
can_steal_fallback(unsigned int order,int start_mt)1840 static bool can_steal_fallback(unsigned int order, int start_mt)
1841 {
1842 /*
1843 * Leaving this order check is intended, although there is
1844 * relaxed order check in next check. The reason is that
1845 * we can actually steal whole pageblock if this condition met,
1846 * but, below check doesn't guarantee it and that is just heuristic
1847 * so could be changed anytime.
1848 */
1849 if (order >= pageblock_order)
1850 return true;
1851
1852 /*
1853 * Movable pages won't cause permanent fragmentation, so when you alloc
1854 * small pages, you just need to temporarily steal unmovable or
1855 * reclaimable pages that are closest to the request size. After a
1856 * while, memory compaction may occur to form large contiguous pages,
1857 * and the next movable allocation may not need to steal. Unmovable and
1858 * reclaimable allocations need to actually steal pages.
1859 */
1860 if (order >= pageblock_order / 2 ||
1861 start_mt == MIGRATE_RECLAIMABLE ||
1862 start_mt == MIGRATE_UNMOVABLE ||
1863 page_group_by_mobility_disabled)
1864 return true;
1865
1866 return false;
1867 }
1868
boost_watermark(struct zone * zone)1869 static inline bool boost_watermark(struct zone *zone)
1870 {
1871 unsigned long max_boost;
1872
1873 if (!watermark_boost_factor)
1874 return false;
1875 /*
1876 * Don't bother in zones that are unlikely to produce results.
1877 * On small machines, including kdump capture kernels running
1878 * in a small area, boosting the watermark can cause an out of
1879 * memory situation immediately.
1880 */
1881 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1882 return false;
1883
1884 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1885 watermark_boost_factor, 10000);
1886
1887 /*
1888 * high watermark may be uninitialised if fragmentation occurs
1889 * very early in boot so do not boost. We do not fall
1890 * through and boost by pageblock_nr_pages as failing
1891 * allocations that early means that reclaim is not going
1892 * to help and it may even be impossible to reclaim the
1893 * boosted watermark resulting in a hang.
1894 */
1895 if (!max_boost)
1896 return false;
1897
1898 max_boost = max(pageblock_nr_pages, max_boost);
1899
1900 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1901 max_boost);
1902
1903 return true;
1904 }
1905
1906 /*
1907 * This function implements actual steal behaviour. If order is large enough, we
1908 * can claim the whole pageblock for the requested migratetype. If not, we check
1909 * the pageblock for constituent pages; if at least half of the pages are free
1910 * or compatible, we can still claim the whole block, so pages freed in the
1911 * future will be put on the correct free list. Otherwise, we isolate exactly
1912 * the order we need from the fallback block and leave its migratetype alone.
1913 */
1914 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)1915 steal_suitable_fallback(struct zone *zone, struct page *page,
1916 int current_order, int order, int start_type,
1917 unsigned int alloc_flags, bool whole_block)
1918 {
1919 int free_pages, movable_pages, alike_pages;
1920 unsigned long start_pfn;
1921 int block_type;
1922
1923 block_type = get_pageblock_migratetype(page);
1924
1925 /*
1926 * This can happen due to races and we want to prevent broken
1927 * highatomic accounting.
1928 */
1929 if (is_migrate_highatomic(block_type))
1930 goto single_page;
1931
1932 /* Take ownership for orders >= pageblock_order */
1933 if (current_order >= pageblock_order) {
1934 unsigned int nr_added;
1935
1936 del_page_from_free_list(page, zone, current_order, block_type);
1937 change_pageblock_range(page, current_order, start_type);
1938 nr_added = expand(zone, page, order, current_order, start_type);
1939 account_freepages(zone, nr_added, start_type);
1940 return page;
1941 }
1942
1943 /*
1944 * Boost watermarks to increase reclaim pressure to reduce the
1945 * likelihood of future fallbacks. Wake kswapd now as the node
1946 * may be balanced overall and kswapd will not wake naturally.
1947 */
1948 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1949 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1950
1951 /* We are not allowed to try stealing from the whole block */
1952 if (!whole_block)
1953 goto single_page;
1954
1955 /* moving whole block can fail due to zone boundary conditions */
1956 if (!prep_move_freepages_block(zone, page, &start_pfn, &free_pages,
1957 &movable_pages))
1958 goto single_page;
1959
1960 /*
1961 * Determine how many pages are compatible with our allocation.
1962 * For movable allocation, it's the number of movable pages which
1963 * we just obtained. For other types it's a bit more tricky.
1964 */
1965 if (start_type == MIGRATE_MOVABLE) {
1966 alike_pages = movable_pages;
1967 } else {
1968 /*
1969 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1970 * to MOVABLE pageblock, consider all non-movable pages as
1971 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1972 * vice versa, be conservative since we can't distinguish the
1973 * exact migratetype of non-movable pages.
1974 */
1975 if (block_type == MIGRATE_MOVABLE)
1976 alike_pages = pageblock_nr_pages
1977 - (free_pages + movable_pages);
1978 else
1979 alike_pages = 0;
1980 }
1981 /*
1982 * If a sufficient number of pages in the block are either free or of
1983 * compatible migratability as our allocation, claim the whole block.
1984 */
1985 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1986 page_group_by_mobility_disabled) {
1987 __move_freepages_block(zone, start_pfn, block_type, start_type);
1988 return __rmqueue_smallest(zone, order, start_type);
1989 }
1990
1991 single_page:
1992 page_del_and_expand(zone, page, order, current_order, block_type);
1993 return page;
1994 }
1995
1996 /*
1997 * Check whether there is a suitable fallback freepage with requested order.
1998 * If only_stealable is true, this function returns fallback_mt only if
1999 * we can steal other freepages all together. This would help to reduce
2000 * fragmentation due to mixed migratetype pages in one pageblock.
2001 */
find_suitable_fallback(struct free_area * area,unsigned int order,int migratetype,bool only_stealable,bool * can_steal)2002 int find_suitable_fallback(struct free_area *area, unsigned int order,
2003 int migratetype, bool only_stealable, bool *can_steal)
2004 {
2005 int i;
2006 int fallback_mt;
2007
2008 if (area->nr_free == 0)
2009 return -1;
2010
2011 *can_steal = false;
2012 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
2013 fallback_mt = fallbacks[migratetype][i];
2014 if (free_area_empty(area, fallback_mt))
2015 continue;
2016
2017 if (can_steal_fallback(order, migratetype))
2018 *can_steal = true;
2019
2020 if (!only_stealable)
2021 return fallback_mt;
2022
2023 if (*can_steal)
2024 return fallback_mt;
2025 }
2026
2027 return -1;
2028 }
2029
2030 /*
2031 * Reserve the pageblock(s) surrounding an allocation request for
2032 * exclusive use of high-order atomic allocations if there are no
2033 * empty page blocks that contain a page with a suitable order
2034 */
reserve_highatomic_pageblock(struct page * page,int order,struct zone * zone)2035 static void reserve_highatomic_pageblock(struct page *page, int order,
2036 struct zone *zone)
2037 {
2038 int mt;
2039 unsigned long max_managed, flags;
2040
2041 /*
2042 * The number reserved as: minimum is 1 pageblock, maximum is
2043 * roughly 1% of a zone. But if 1% of a zone falls below a
2044 * pageblock size, then don't reserve any pageblocks.
2045 * Check is race-prone but harmless.
2046 */
2047 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
2048 return;
2049 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
2050 if (zone->nr_reserved_highatomic >= max_managed)
2051 return;
2052
2053 spin_lock_irqsave(&zone->lock, flags);
2054
2055 /* Recheck the nr_reserved_highatomic limit under the lock */
2056 if (zone->nr_reserved_highatomic >= max_managed)
2057 goto out_unlock;
2058
2059 /* Yoink! */
2060 mt = get_pageblock_migratetype(page);
2061 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2062 if (!migratetype_is_mergeable(mt))
2063 goto out_unlock;
2064
2065 if (order < pageblock_order) {
2066 if (move_freepages_block(zone, page, mt, MIGRATE_HIGHATOMIC) == -1)
2067 goto out_unlock;
2068 zone->nr_reserved_highatomic += pageblock_nr_pages;
2069 } else {
2070 change_pageblock_range(page, order, MIGRATE_HIGHATOMIC);
2071 zone->nr_reserved_highatomic += 1 << order;
2072 }
2073
2074 out_unlock:
2075 spin_unlock_irqrestore(&zone->lock, flags);
2076 }
2077
2078 /*
2079 * Used when an allocation is about to fail under memory pressure. This
2080 * potentially hurts the reliability of high-order allocations when under
2081 * intense memory pressure but failed atomic allocations should be easier
2082 * to recover from than an OOM.
2083 *
2084 * If @force is true, try to unreserve pageblocks even though highatomic
2085 * pageblock is exhausted.
2086 */
unreserve_highatomic_pageblock(const struct alloc_context * ac,bool force)2087 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2088 bool force)
2089 {
2090 struct zonelist *zonelist = ac->zonelist;
2091 unsigned long flags;
2092 struct zoneref *z;
2093 struct zone *zone;
2094 struct page *page;
2095 int order;
2096 int ret;
2097
2098 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2099 ac->nodemask) {
2100 /*
2101 * Preserve at least one pageblock unless memory pressure
2102 * is really high.
2103 */
2104 if (!force && zone->nr_reserved_highatomic <=
2105 pageblock_nr_pages)
2106 continue;
2107
2108 spin_lock_irqsave(&zone->lock, flags);
2109 for (order = 0; order < NR_PAGE_ORDERS; order++) {
2110 struct free_area *area = &(zone->free_area[order]);
2111 int mt;
2112
2113 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2114 if (!page)
2115 continue;
2116
2117 mt = get_pageblock_migratetype(page);
2118 /*
2119 * In page freeing path, migratetype change is racy so
2120 * we can counter several free pages in a pageblock
2121 * in this loop although we changed the pageblock type
2122 * from highatomic to ac->migratetype. So we should
2123 * adjust the count once.
2124 */
2125 if (is_migrate_highatomic(mt)) {
2126 unsigned long size;
2127 /*
2128 * It should never happen but changes to
2129 * locking could inadvertently allow a per-cpu
2130 * drain to add pages to MIGRATE_HIGHATOMIC
2131 * while unreserving so be safe and watch for
2132 * underflows.
2133 */
2134 size = max(pageblock_nr_pages, 1UL << order);
2135 size = min(size, zone->nr_reserved_highatomic);
2136 zone->nr_reserved_highatomic -= size;
2137 }
2138
2139 /*
2140 * Convert to ac->migratetype and avoid the normal
2141 * pageblock stealing heuristics. Minimally, the caller
2142 * is doing the work and needs the pages. More
2143 * importantly, if the block was always converted to
2144 * MIGRATE_UNMOVABLE or another type then the number
2145 * of pageblocks that cannot be completely freed
2146 * may increase.
2147 */
2148 if (order < pageblock_order)
2149 ret = move_freepages_block(zone, page, mt,
2150 ac->migratetype);
2151 else {
2152 move_to_free_list(page, zone, order, mt,
2153 ac->migratetype);
2154 change_pageblock_range(page, order,
2155 ac->migratetype);
2156 ret = 1;
2157 }
2158 /*
2159 * Reserving the block(s) already succeeded,
2160 * so this should not fail on zone boundaries.
2161 */
2162 WARN_ON_ONCE(ret == -1);
2163 if (ret > 0) {
2164 spin_unlock_irqrestore(&zone->lock, flags);
2165 return ret;
2166 }
2167 }
2168 spin_unlock_irqrestore(&zone->lock, flags);
2169 }
2170
2171 return false;
2172 }
2173
2174 /*
2175 * Try finding a free buddy page on the fallback list and put it on the free
2176 * list of requested migratetype, possibly along with other pages from the same
2177 * block, depending on fragmentation avoidance heuristics. Returns true if
2178 * fallback was found so that __rmqueue_smallest() can grab it.
2179 *
2180 * The use of signed ints for order and current_order is a deliberate
2181 * deviation from the rest of this file, to make the for loop
2182 * condition simpler.
2183 */
2184 static __always_inline struct page *
__rmqueue_fallback(struct zone * zone,int order,int start_migratetype,unsigned int alloc_flags)2185 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2186 unsigned int alloc_flags)
2187 {
2188 struct free_area *area;
2189 int current_order;
2190 int min_order = order;
2191 struct page *page;
2192 int fallback_mt;
2193 bool can_steal;
2194
2195 /*
2196 * Do not steal pages from freelists belonging to other pageblocks
2197 * i.e. orders < pageblock_order. If there are no local zones free,
2198 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2199 */
2200 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2201 min_order = pageblock_order;
2202
2203 /*
2204 * Find the largest available free page in the other list. This roughly
2205 * approximates finding the pageblock with the most free pages, which
2206 * would be too costly to do exactly.
2207 */
2208 for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2209 --current_order) {
2210 area = &(zone->free_area[current_order]);
2211 fallback_mt = find_suitable_fallback(area, current_order,
2212 start_migratetype, false, &can_steal);
2213 if (fallback_mt == -1)
2214 continue;
2215
2216 /*
2217 * We cannot steal all free pages from the pageblock and the
2218 * requested migratetype is movable. In that case it's better to
2219 * steal and split the smallest available page instead of the
2220 * largest available page, because even if the next movable
2221 * allocation falls back into a different pageblock than this
2222 * one, it won't cause permanent fragmentation.
2223 */
2224 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2225 && current_order > order)
2226 goto find_smallest;
2227
2228 goto do_steal;
2229 }
2230
2231 return NULL;
2232
2233 find_smallest:
2234 for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
2235 area = &(zone->free_area[current_order]);
2236 fallback_mt = find_suitable_fallback(area, current_order,
2237 start_migratetype, false, &can_steal);
2238 if (fallback_mt != -1)
2239 break;
2240 }
2241
2242 /*
2243 * This should not happen - we already found a suitable fallback
2244 * when looking for the largest page.
2245 */
2246 VM_BUG_ON(current_order > MAX_PAGE_ORDER);
2247
2248 do_steal:
2249 page = get_page_from_free_area(area, fallback_mt);
2250
2251 /* take off list, maybe claim block, expand remainder */
2252 page = steal_suitable_fallback(zone, page, current_order, order,
2253 start_migratetype, alloc_flags, can_steal);
2254
2255 trace_mm_page_alloc_extfrag(page, order, current_order,
2256 start_migratetype, fallback_mt);
2257
2258 return page;
2259 }
2260
2261 /*
2262 * Do the hard work of removing an element from the buddy allocator.
2263 * Call me with the zone->lock already held.
2264 */
2265 static __always_inline struct page *
__rmqueue(struct zone * zone,unsigned int order,int migratetype,unsigned int alloc_flags)2266 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2267 unsigned int alloc_flags)
2268 {
2269 struct page *page;
2270
2271 if (IS_ENABLED(CONFIG_CMA)) {
2272 /*
2273 * Balance movable allocations between regular and CMA areas by
2274 * allocating from CMA when over half of the zone's free memory
2275 * is in the CMA area.
2276 */
2277 if (alloc_flags & ALLOC_CMA &&
2278 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2279 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2280 page = __rmqueue_cma_fallback(zone, order);
2281 if (page)
2282 return page;
2283 }
2284 }
2285
2286 page = __rmqueue_smallest(zone, order, migratetype);
2287 if (unlikely(!page)) {
2288 if (alloc_flags & ALLOC_CMA)
2289 page = __rmqueue_cma_fallback(zone, order);
2290
2291 if (!page)
2292 page = __rmqueue_fallback(zone, order, migratetype,
2293 alloc_flags);
2294 }
2295 return page;
2296 }
2297
2298 /*
2299 * Obtain a specified number of elements from the buddy allocator, all under
2300 * a single hold of the lock, for efficiency. Add them to the supplied list.
2301 * Returns the number of new pages which were placed at *list.
2302 */
rmqueue_bulk(struct zone * zone,unsigned int order,unsigned long count,struct list_head * list,int migratetype,unsigned int alloc_flags)2303 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2304 unsigned long count, struct list_head *list,
2305 int migratetype, unsigned int alloc_flags)
2306 {
2307 unsigned long flags;
2308 int i;
2309
2310 spin_lock_irqsave(&zone->lock, flags);
2311 for (i = 0; i < count; ++i) {
2312 struct page *page = __rmqueue(zone, order, migratetype,
2313 alloc_flags);
2314 if (unlikely(page == NULL))
2315 break;
2316
2317 /*
2318 * Split buddy pages returned by expand() are received here in
2319 * physical page order. The page is added to the tail of
2320 * caller's list. From the callers perspective, the linked list
2321 * is ordered by page number under some conditions. This is
2322 * useful for IO devices that can forward direction from the
2323 * head, thus also in the physical page order. This is useful
2324 * for IO devices that can merge IO requests if the physical
2325 * pages are ordered properly.
2326 */
2327 list_add_tail(&page->pcp_list, list);
2328 }
2329 spin_unlock_irqrestore(&zone->lock, flags);
2330
2331 return i;
2332 }
2333
2334 /*
2335 * Called from the vmstat counter updater to decay the PCP high.
2336 * Return whether there are addition works to do.
2337 */
decay_pcp_high(struct zone * zone,struct per_cpu_pages * pcp)2338 int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2339 {
2340 int high_min, to_drain, batch;
2341 int todo = 0;
2342
2343 high_min = READ_ONCE(pcp->high_min);
2344 batch = READ_ONCE(pcp->batch);
2345 /*
2346 * Decrease pcp->high periodically to try to free possible
2347 * idle PCP pages. And, avoid to free too many pages to
2348 * control latency. This caps pcp->high decrement too.
2349 */
2350 if (pcp->high > high_min) {
2351 pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2352 pcp->high - (pcp->high >> 3), high_min);
2353 if (pcp->high > high_min)
2354 todo++;
2355 }
2356
2357 to_drain = pcp->count - pcp->high;
2358 if (to_drain > 0) {
2359 spin_lock(&pcp->lock);
2360 free_pcppages_bulk(zone, to_drain, pcp, 0);
2361 spin_unlock(&pcp->lock);
2362 todo++;
2363 }
2364
2365 return todo;
2366 }
2367
2368 #ifdef CONFIG_NUMA
2369 /*
2370 * Called from the vmstat counter updater to drain pagesets of this
2371 * currently executing processor on remote nodes after they have
2372 * expired.
2373 */
drain_zone_pages(struct zone * zone,struct per_cpu_pages * pcp)2374 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2375 {
2376 int to_drain, batch;
2377
2378 batch = READ_ONCE(pcp->batch);
2379 to_drain = min(pcp->count, batch);
2380 if (to_drain > 0) {
2381 spin_lock(&pcp->lock);
2382 free_pcppages_bulk(zone, to_drain, pcp, 0);
2383 spin_unlock(&pcp->lock);
2384 }
2385 }
2386 #endif
2387
2388 /*
2389 * Drain pcplists of the indicated processor and zone.
2390 */
drain_pages_zone(unsigned int cpu,struct zone * zone)2391 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2392 {
2393 struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2394 int count;
2395
2396 do {
2397 spin_lock(&pcp->lock);
2398 count = pcp->count;
2399 if (count) {
2400 int to_drain = min(count,
2401 pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
2402
2403 free_pcppages_bulk(zone, to_drain, pcp, 0);
2404 count -= to_drain;
2405 }
2406 spin_unlock(&pcp->lock);
2407 } while (count);
2408 }
2409
2410 /*
2411 * Drain pcplists of all zones on the indicated processor.
2412 */
drain_pages(unsigned int cpu)2413 static void drain_pages(unsigned int cpu)
2414 {
2415 struct zone *zone;
2416
2417 for_each_populated_zone(zone) {
2418 drain_pages_zone(cpu, zone);
2419 }
2420 }
2421
2422 /*
2423 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2424 */
drain_local_pages(struct zone * zone)2425 void drain_local_pages(struct zone *zone)
2426 {
2427 int cpu = smp_processor_id();
2428
2429 if (zone)
2430 drain_pages_zone(cpu, zone);
2431 else
2432 drain_pages(cpu);
2433 }
2434
2435 /*
2436 * The implementation of drain_all_pages(), exposing an extra parameter to
2437 * drain on all cpus.
2438 *
2439 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2440 * not empty. The check for non-emptiness can however race with a free to
2441 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2442 * that need the guarantee that every CPU has drained can disable the
2443 * optimizing racy check.
2444 */
__drain_all_pages(struct zone * zone,bool force_all_cpus)2445 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2446 {
2447 int cpu;
2448
2449 /*
2450 * Allocate in the BSS so we won't require allocation in
2451 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2452 */
2453 static cpumask_t cpus_with_pcps;
2454
2455 /*
2456 * Do not drain if one is already in progress unless it's specific to
2457 * a zone. Such callers are primarily CMA and memory hotplug and need
2458 * the drain to be complete when the call returns.
2459 */
2460 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2461 if (!zone)
2462 return;
2463 mutex_lock(&pcpu_drain_mutex);
2464 }
2465
2466 /*
2467 * We don't care about racing with CPU hotplug event
2468 * as offline notification will cause the notified
2469 * cpu to drain that CPU pcps and on_each_cpu_mask
2470 * disables preemption as part of its processing
2471 */
2472 for_each_online_cpu(cpu) {
2473 struct per_cpu_pages *pcp;
2474 struct zone *z;
2475 bool has_pcps = false;
2476
2477 if (force_all_cpus) {
2478 /*
2479 * The pcp.count check is racy, some callers need a
2480 * guarantee that no cpu is missed.
2481 */
2482 has_pcps = true;
2483 } else if (zone) {
2484 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2485 if (pcp->count)
2486 has_pcps = true;
2487 } else {
2488 for_each_populated_zone(z) {
2489 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2490 if (pcp->count) {
2491 has_pcps = true;
2492 break;
2493 }
2494 }
2495 }
2496
2497 if (has_pcps)
2498 cpumask_set_cpu(cpu, &cpus_with_pcps);
2499 else
2500 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2501 }
2502
2503 for_each_cpu(cpu, &cpus_with_pcps) {
2504 if (zone)
2505 drain_pages_zone(cpu, zone);
2506 else
2507 drain_pages(cpu);
2508 }
2509
2510 mutex_unlock(&pcpu_drain_mutex);
2511 }
2512
2513 /*
2514 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2515 *
2516 * When zone parameter is non-NULL, spill just the single zone's pages.
2517 */
drain_all_pages(struct zone * zone)2518 void drain_all_pages(struct zone *zone)
2519 {
2520 __drain_all_pages(zone, false);
2521 }
2522
nr_pcp_free(struct per_cpu_pages * pcp,int batch,int high,bool free_high)2523 static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2524 {
2525 int min_nr_free, max_nr_free;
2526
2527 /* Free as much as possible if batch freeing high-order pages. */
2528 if (unlikely(free_high))
2529 return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2530
2531 /* Check for PCP disabled or boot pageset */
2532 if (unlikely(high < batch))
2533 return 1;
2534
2535 /* Leave at least pcp->batch pages on the list */
2536 min_nr_free = batch;
2537 max_nr_free = high - batch;
2538
2539 /*
2540 * Increase the batch number to the number of the consecutive
2541 * freed pages to reduce zone lock contention.
2542 */
2543 batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2544
2545 return batch;
2546 }
2547
nr_pcp_high(struct per_cpu_pages * pcp,struct zone * zone,int batch,bool free_high)2548 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2549 int batch, bool free_high)
2550 {
2551 int high, high_min, high_max;
2552
2553 high_min = READ_ONCE(pcp->high_min);
2554 high_max = READ_ONCE(pcp->high_max);
2555 high = pcp->high = clamp(pcp->high, high_min, high_max);
2556
2557 if (unlikely(!high))
2558 return 0;
2559
2560 if (unlikely(free_high)) {
2561 pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2562 high_min);
2563 return 0;
2564 }
2565
2566 /*
2567 * If reclaim is active, limit the number of pages that can be
2568 * stored on pcp lists
2569 */
2570 if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2571 int free_count = max_t(int, pcp->free_count, batch);
2572
2573 pcp->high = max(high - free_count, high_min);
2574 return min(batch << 2, pcp->high);
2575 }
2576
2577 if (high_min == high_max)
2578 return high;
2579
2580 if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2581 int free_count = max_t(int, pcp->free_count, batch);
2582
2583 pcp->high = max(high - free_count, high_min);
2584 high = max(pcp->count, high_min);
2585 } else if (pcp->count >= high) {
2586 int need_high = pcp->free_count + batch;
2587
2588 /* pcp->high should be large enough to hold batch freed pages */
2589 if (pcp->high < need_high)
2590 pcp->high = clamp(need_high, high_min, high_max);
2591 }
2592
2593 return high;
2594 }
2595
free_frozen_page_commit(struct zone * zone,struct per_cpu_pages * pcp,struct page * page,int migratetype,unsigned int order)2596 static void free_frozen_page_commit(struct zone *zone,
2597 struct per_cpu_pages *pcp, struct page *page, int migratetype,
2598 unsigned int order)
2599 {
2600 int high, batch;
2601 int pindex;
2602 bool free_high = false;
2603
2604 /*
2605 * On freeing, reduce the number of pages that are batch allocated.
2606 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2607 * allocations.
2608 */
2609 pcp->alloc_factor >>= 1;
2610 __count_vm_events(PGFREE, 1 << order);
2611 pindex = order_to_pindex(migratetype, order);
2612 list_add(&page->pcp_list, &pcp->lists[pindex]);
2613 pcp->count += 1 << order;
2614
2615 batch = READ_ONCE(pcp->batch);
2616 /*
2617 * As high-order pages other than THP's stored on PCP can contribute
2618 * to fragmentation, limit the number stored when PCP is heavily
2619 * freeing without allocation. The remainder after bulk freeing
2620 * stops will be drained from vmstat refresh context.
2621 */
2622 if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2623 free_high = (pcp->free_count >= batch &&
2624 (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2625 (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2626 pcp->count >= READ_ONCE(batch)));
2627 pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2628 } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2629 pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2630 }
2631 if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2632 pcp->free_count += (1 << order);
2633 high = nr_pcp_high(pcp, zone, batch, free_high);
2634 if (pcp->count >= high) {
2635 free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high),
2636 pcp, pindex);
2637 if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2638 zone_watermark_ok(zone, 0, high_wmark_pages(zone),
2639 ZONE_MOVABLE, 0))
2640 clear_bit(ZONE_BELOW_HIGH, &zone->flags);
2641 }
2642 }
2643
2644 /*
2645 * Free a pcp page
2646 */
free_frozen_pages(struct page * page,unsigned int order)2647 void free_frozen_pages(struct page *page, unsigned int order)
2648 {
2649 unsigned long __maybe_unused UP_flags;
2650 struct per_cpu_pages *pcp;
2651 struct zone *zone;
2652 unsigned long pfn = page_to_pfn(page);
2653 int migratetype;
2654
2655 if (!pcp_allowed_order(order)) {
2656 __free_pages_ok(page, order, FPI_NONE);
2657 return;
2658 }
2659
2660 if (!free_pages_prepare(page, order))
2661 return;
2662
2663 /*
2664 * We only track unmovable, reclaimable and movable on pcp lists.
2665 * Place ISOLATE pages on the isolated list because they are being
2666 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2667 * get those areas back if necessary. Otherwise, we may have to free
2668 * excessively into the page allocator
2669 */
2670 zone = page_zone(page);
2671 migratetype = get_pfnblock_migratetype(page, pfn);
2672 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2673 if (unlikely(is_migrate_isolate(migratetype))) {
2674 free_one_page(zone, page, pfn, order, FPI_NONE);
2675 return;
2676 }
2677 migratetype = MIGRATE_MOVABLE;
2678 }
2679
2680 pcp_trylock_prepare(UP_flags);
2681 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2682 if (pcp) {
2683 free_frozen_page_commit(zone, pcp, page, migratetype, order);
2684 pcp_spin_unlock(pcp);
2685 } else {
2686 free_one_page(zone, page, pfn, order, FPI_NONE);
2687 }
2688 pcp_trylock_finish(UP_flags);
2689 }
2690
2691 /*
2692 * Free a batch of folios
2693 */
free_unref_folios(struct folio_batch * folios)2694 void free_unref_folios(struct folio_batch *folios)
2695 {
2696 unsigned long __maybe_unused UP_flags;
2697 struct per_cpu_pages *pcp = NULL;
2698 struct zone *locked_zone = NULL;
2699 int i, j;
2700
2701 /* Prepare folios for freeing */
2702 for (i = 0, j = 0; i < folios->nr; i++) {
2703 struct folio *folio = folios->folios[i];
2704 unsigned long pfn = folio_pfn(folio);
2705 unsigned int order = folio_order(folio);
2706
2707 if (!free_pages_prepare(&folio->page, order))
2708 continue;
2709 /*
2710 * Free orders not handled on the PCP directly to the
2711 * allocator.
2712 */
2713 if (!pcp_allowed_order(order)) {
2714 free_one_page(folio_zone(folio), &folio->page,
2715 pfn, order, FPI_NONE);
2716 continue;
2717 }
2718 folio->private = (void *)(unsigned long)order;
2719 if (j != i)
2720 folios->folios[j] = folio;
2721 j++;
2722 }
2723 folios->nr = j;
2724
2725 for (i = 0; i < folios->nr; i++) {
2726 struct folio *folio = folios->folios[i];
2727 struct zone *zone = folio_zone(folio);
2728 unsigned long pfn = folio_pfn(folio);
2729 unsigned int order = (unsigned long)folio->private;
2730 int migratetype;
2731
2732 folio->private = NULL;
2733 migratetype = get_pfnblock_migratetype(&folio->page, pfn);
2734
2735 /* Different zone requires a different pcp lock */
2736 if (zone != locked_zone ||
2737 is_migrate_isolate(migratetype)) {
2738 if (pcp) {
2739 pcp_spin_unlock(pcp);
2740 pcp_trylock_finish(UP_flags);
2741 locked_zone = NULL;
2742 pcp = NULL;
2743 }
2744
2745 /*
2746 * Free isolated pages directly to the
2747 * allocator, see comment in free_frozen_pages.
2748 */
2749 if (is_migrate_isolate(migratetype)) {
2750 free_one_page(zone, &folio->page, pfn,
2751 order, FPI_NONE);
2752 continue;
2753 }
2754
2755 /*
2756 * trylock is necessary as folios may be getting freed
2757 * from IRQ or SoftIRQ context after an IO completion.
2758 */
2759 pcp_trylock_prepare(UP_flags);
2760 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2761 if (unlikely(!pcp)) {
2762 pcp_trylock_finish(UP_flags);
2763 free_one_page(zone, &folio->page, pfn,
2764 order, FPI_NONE);
2765 continue;
2766 }
2767 locked_zone = zone;
2768 }
2769
2770 /*
2771 * Non-isolated types over MIGRATE_PCPTYPES get added
2772 * to the MIGRATE_MOVABLE pcp list.
2773 */
2774 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2775 migratetype = MIGRATE_MOVABLE;
2776
2777 trace_mm_page_free_batched(&folio->page);
2778 free_frozen_page_commit(zone, pcp, &folio->page, migratetype,
2779 order);
2780 }
2781
2782 if (pcp) {
2783 pcp_spin_unlock(pcp);
2784 pcp_trylock_finish(UP_flags);
2785 }
2786 folio_batch_reinit(folios);
2787 }
2788
2789 /*
2790 * split_page takes a non-compound higher-order page, and splits it into
2791 * n (1<<order) sub-pages: page[0..n]
2792 * Each sub-page must be freed individually.
2793 *
2794 * Note: this is probably too low level an operation for use in drivers.
2795 * Please consult with lkml before using this in your driver.
2796 */
split_page(struct page * page,unsigned int order)2797 void split_page(struct page *page, unsigned int order)
2798 {
2799 int i;
2800
2801 VM_BUG_ON_PAGE(PageCompound(page), page);
2802 VM_BUG_ON_PAGE(!page_count(page), page);
2803
2804 for (i = 1; i < (1 << order); i++)
2805 set_page_refcounted(page + i);
2806 split_page_owner(page, order, 0);
2807 pgalloc_tag_split(page_folio(page), order, 0);
2808 split_page_memcg(page, order, 0);
2809 }
2810 EXPORT_SYMBOL_GPL(split_page);
2811
__isolate_free_page(struct page * page,unsigned int order)2812 int __isolate_free_page(struct page *page, unsigned int order)
2813 {
2814 struct zone *zone = page_zone(page);
2815 int mt = get_pageblock_migratetype(page);
2816
2817 if (!is_migrate_isolate(mt)) {
2818 unsigned long watermark;
2819 /*
2820 * Obey watermarks as if the page was being allocated. We can
2821 * emulate a high-order watermark check with a raised order-0
2822 * watermark, because we already know our high-order page
2823 * exists.
2824 */
2825 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2826 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2827 return 0;
2828 }
2829
2830 del_page_from_free_list(page, zone, order, mt);
2831
2832 /*
2833 * Set the pageblock if the isolated page is at least half of a
2834 * pageblock
2835 */
2836 if (order >= pageblock_order - 1) {
2837 struct page *endpage = page + (1 << order) - 1;
2838 for (; page < endpage; page += pageblock_nr_pages) {
2839 int mt = get_pageblock_migratetype(page);
2840 /*
2841 * Only change normal pageblocks (i.e., they can merge
2842 * with others)
2843 */
2844 if (migratetype_is_mergeable(mt))
2845 move_freepages_block(zone, page, mt,
2846 MIGRATE_MOVABLE);
2847 }
2848 }
2849
2850 return 1UL << order;
2851 }
2852
2853 /**
2854 * __putback_isolated_page - Return a now-isolated page back where we got it
2855 * @page: Page that was isolated
2856 * @order: Order of the isolated page
2857 * @mt: The page's pageblock's migratetype
2858 *
2859 * This function is meant to return a page pulled from the free lists via
2860 * __isolate_free_page back to the free lists they were pulled from.
2861 */
__putback_isolated_page(struct page * page,unsigned int order,int mt)2862 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2863 {
2864 struct zone *zone = page_zone(page);
2865
2866 /* zone lock should be held when this function is called */
2867 lockdep_assert_held(&zone->lock);
2868
2869 /* Return isolated page to tail of freelist. */
2870 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2871 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2872 }
2873
2874 /*
2875 * Update NUMA hit/miss statistics
2876 */
zone_statistics(struct zone * preferred_zone,struct zone * z,long nr_account)2877 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2878 long nr_account)
2879 {
2880 #ifdef CONFIG_NUMA
2881 enum numa_stat_item local_stat = NUMA_LOCAL;
2882
2883 /* skip numa counters update if numa stats is disabled */
2884 if (!static_branch_likely(&vm_numa_stat_key))
2885 return;
2886
2887 if (zone_to_nid(z) != numa_node_id())
2888 local_stat = NUMA_OTHER;
2889
2890 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2891 __count_numa_events(z, NUMA_HIT, nr_account);
2892 else {
2893 __count_numa_events(z, NUMA_MISS, nr_account);
2894 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2895 }
2896 __count_numa_events(z, local_stat, nr_account);
2897 #endif
2898 }
2899
2900 static __always_inline
rmqueue_buddy(struct zone * preferred_zone,struct zone * zone,unsigned int order,unsigned int alloc_flags,int migratetype)2901 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2902 unsigned int order, unsigned int alloc_flags,
2903 int migratetype)
2904 {
2905 struct page *page;
2906 unsigned long flags;
2907
2908 do {
2909 page = NULL;
2910 spin_lock_irqsave(&zone->lock, flags);
2911 if (alloc_flags & ALLOC_HIGHATOMIC)
2912 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2913 if (!page) {
2914 page = __rmqueue(zone, order, migratetype, alloc_flags);
2915
2916 /*
2917 * If the allocation fails, allow OOM handling and
2918 * order-0 (atomic) allocs access to HIGHATOMIC
2919 * reserves as failing now is worse than failing a
2920 * high-order atomic allocation in the future.
2921 */
2922 if (!page && (alloc_flags & (ALLOC_OOM|ALLOC_NON_BLOCK)))
2923 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2924
2925 if (!page) {
2926 spin_unlock_irqrestore(&zone->lock, flags);
2927 return NULL;
2928 }
2929 }
2930 spin_unlock_irqrestore(&zone->lock, flags);
2931 } while (check_new_pages(page, order));
2932
2933 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2934 zone_statistics(preferred_zone, zone, 1);
2935
2936 return page;
2937 }
2938
nr_pcp_alloc(struct per_cpu_pages * pcp,struct zone * zone,int order)2939 static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
2940 {
2941 int high, base_batch, batch, max_nr_alloc;
2942 int high_max, high_min;
2943
2944 base_batch = READ_ONCE(pcp->batch);
2945 high_min = READ_ONCE(pcp->high_min);
2946 high_max = READ_ONCE(pcp->high_max);
2947 high = pcp->high = clamp(pcp->high, high_min, high_max);
2948
2949 /* Check for PCP disabled or boot pageset */
2950 if (unlikely(high < base_batch))
2951 return 1;
2952
2953 if (order)
2954 batch = base_batch;
2955 else
2956 batch = (base_batch << pcp->alloc_factor);
2957
2958 /*
2959 * If we had larger pcp->high, we could avoid to allocate from
2960 * zone.
2961 */
2962 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
2963 high = pcp->high = min(high + batch, high_max);
2964
2965 if (!order) {
2966 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
2967 /*
2968 * Double the number of pages allocated each time there is
2969 * subsequent allocation of order-0 pages without any freeing.
2970 */
2971 if (batch <= max_nr_alloc &&
2972 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
2973 pcp->alloc_factor++;
2974 batch = min(batch, max_nr_alloc);
2975 }
2976
2977 /*
2978 * Scale batch relative to order if batch implies free pages
2979 * can be stored on the PCP. Batch can be 1 for small zones or
2980 * for boot pagesets which should never store free pages as
2981 * the pages may belong to arbitrary zones.
2982 */
2983 if (batch > 1)
2984 batch = max(batch >> order, 2);
2985
2986 return batch;
2987 }
2988
2989 /* Remove page from the per-cpu list, caller must protect the list */
2990 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)2991 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2992 int migratetype,
2993 unsigned int alloc_flags,
2994 struct per_cpu_pages *pcp,
2995 struct list_head *list)
2996 {
2997 struct page *page;
2998
2999 do {
3000 if (list_empty(list)) {
3001 int batch = nr_pcp_alloc(pcp, zone, order);
3002 int alloced;
3003
3004 alloced = rmqueue_bulk(zone, order,
3005 batch, list,
3006 migratetype, alloc_flags);
3007
3008 pcp->count += alloced << order;
3009 if (unlikely(list_empty(list)))
3010 return NULL;
3011 }
3012
3013 page = list_first_entry(list, struct page, pcp_list);
3014 list_del(&page->pcp_list);
3015 pcp->count -= 1 << order;
3016 } while (check_new_pages(page, order));
3017
3018 return page;
3019 }
3020
3021 /* 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)3022 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3023 struct zone *zone, unsigned int order,
3024 int migratetype, unsigned int alloc_flags)
3025 {
3026 struct per_cpu_pages *pcp;
3027 struct list_head *list;
3028 struct page *page;
3029 unsigned long __maybe_unused UP_flags;
3030
3031 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
3032 pcp_trylock_prepare(UP_flags);
3033 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3034 if (!pcp) {
3035 pcp_trylock_finish(UP_flags);
3036 return NULL;
3037 }
3038
3039 /*
3040 * On allocation, reduce the number of pages that are batch freed.
3041 * See nr_pcp_free() where free_factor is increased for subsequent
3042 * frees.
3043 */
3044 pcp->free_count >>= 1;
3045 list = &pcp->lists[order_to_pindex(migratetype, order)];
3046 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3047 pcp_spin_unlock(pcp);
3048 pcp_trylock_finish(UP_flags);
3049 if (page) {
3050 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3051 zone_statistics(preferred_zone, zone, 1);
3052 }
3053 return page;
3054 }
3055
3056 /*
3057 * Allocate a page from the given zone.
3058 * Use pcplists for THP or "cheap" high-order allocations.
3059 */
3060
3061 /*
3062 * Do not instrument rmqueue() with KMSAN. This function may call
3063 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
3064 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3065 * may call rmqueue() again, which will result in a deadlock.
3066 */
3067 __no_sanitize_memory
3068 static inline
rmqueue(struct zone * preferred_zone,struct zone * zone,unsigned int order,gfp_t gfp_flags,unsigned int alloc_flags,int migratetype)3069 struct page *rmqueue(struct zone *preferred_zone,
3070 struct zone *zone, unsigned int order,
3071 gfp_t gfp_flags, unsigned int alloc_flags,
3072 int migratetype)
3073 {
3074 struct page *page;
3075
3076 if (likely(pcp_allowed_order(order))) {
3077 page = rmqueue_pcplist(preferred_zone, zone, order,
3078 migratetype, alloc_flags);
3079 if (likely(page))
3080 goto out;
3081 }
3082
3083 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3084 migratetype);
3085
3086 out:
3087 /* Separate test+clear to avoid unnecessary atomics */
3088 if ((alloc_flags & ALLOC_KSWAPD) &&
3089 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3090 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3091 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3092 }
3093
3094 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3095 return page;
3096 }
3097
__zone_watermark_unusable_free(struct zone * z,unsigned int order,unsigned int alloc_flags)3098 static inline long __zone_watermark_unusable_free(struct zone *z,
3099 unsigned int order, unsigned int alloc_flags)
3100 {
3101 long unusable_free = (1 << order) - 1;
3102
3103 /*
3104 * If the caller does not have rights to reserves below the min
3105 * watermark then subtract the free pages reserved for highatomic.
3106 */
3107 if (likely(!(alloc_flags & ALLOC_RESERVES)))
3108 unusable_free += READ_ONCE(z->nr_free_highatomic);
3109
3110 #ifdef CONFIG_CMA
3111 /* If allocation can't use CMA areas don't use free CMA pages */
3112 if (!(alloc_flags & ALLOC_CMA))
3113 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3114 #endif
3115
3116 return unusable_free;
3117 }
3118
3119 /*
3120 * Return true if free base pages are above 'mark'. For high-order checks it
3121 * will return true of the order-0 watermark is reached and there is at least
3122 * one free page of a suitable size. Checking now avoids taking the zone lock
3123 * to check in the allocation paths if no pages are free.
3124 */
__zone_watermark_ok(struct zone * z,unsigned int order,unsigned long mark,int highest_zoneidx,unsigned int alloc_flags,long free_pages)3125 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3126 int highest_zoneidx, unsigned int alloc_flags,
3127 long free_pages)
3128 {
3129 long min = mark;
3130 int o;
3131
3132 /* free_pages may go negative - that's OK */
3133 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3134
3135 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
3136 /*
3137 * __GFP_HIGH allows access to 50% of the min reserve as well
3138 * as OOM.
3139 */
3140 if (alloc_flags & ALLOC_MIN_RESERVE) {
3141 min -= min / 2;
3142
3143 /*
3144 * Non-blocking allocations (e.g. GFP_ATOMIC) can
3145 * access more reserves than just __GFP_HIGH. Other
3146 * non-blocking allocations requests such as GFP_NOWAIT
3147 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
3148 * access to the min reserve.
3149 */
3150 if (alloc_flags & ALLOC_NON_BLOCK)
3151 min -= min / 4;
3152 }
3153
3154 /*
3155 * OOM victims can try even harder than the normal reserve
3156 * users on the grounds that it's definitely going to be in
3157 * the exit path shortly and free memory. Any allocation it
3158 * makes during the free path will be small and short-lived.
3159 */
3160 if (alloc_flags & ALLOC_OOM)
3161 min -= min / 2;
3162 }
3163
3164 /*
3165 * Check watermarks for an order-0 allocation request. If these
3166 * are not met, then a high-order request also cannot go ahead
3167 * even if a suitable page happened to be free.
3168 */
3169 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3170 return false;
3171
3172 /* If this is an order-0 request then the watermark is fine */
3173 if (!order)
3174 return true;
3175
3176 /* For a high-order request, check at least one suitable page is free */
3177 for (o = order; o < NR_PAGE_ORDERS; o++) {
3178 struct free_area *area = &z->free_area[o];
3179 int mt;
3180
3181 if (!area->nr_free)
3182 continue;
3183
3184 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3185 if (!free_area_empty(area, mt))
3186 return true;
3187 }
3188
3189 #ifdef CONFIG_CMA
3190 if ((alloc_flags & ALLOC_CMA) &&
3191 !free_area_empty(area, MIGRATE_CMA)) {
3192 return true;
3193 }
3194 #endif
3195 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3196 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3197 return true;
3198 }
3199 }
3200 return false;
3201 }
3202
zone_watermark_ok(struct zone * z,unsigned int order,unsigned long mark,int highest_zoneidx,unsigned int alloc_flags)3203 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3204 int highest_zoneidx, unsigned int alloc_flags)
3205 {
3206 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3207 zone_page_state(z, NR_FREE_PAGES));
3208 }
3209
zone_watermark_fast(struct zone * z,unsigned int order,unsigned long mark,int highest_zoneidx,unsigned int alloc_flags,gfp_t gfp_mask)3210 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3211 unsigned long mark, int highest_zoneidx,
3212 unsigned int alloc_flags, gfp_t gfp_mask)
3213 {
3214 long free_pages;
3215
3216 free_pages = zone_page_state(z, NR_FREE_PAGES);
3217
3218 /*
3219 * Fast check for order-0 only. If this fails then the reserves
3220 * need to be calculated.
3221 */
3222 if (!order) {
3223 long usable_free;
3224 long reserved;
3225
3226 usable_free = free_pages;
3227 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3228
3229 /* reserved may over estimate high-atomic reserves. */
3230 usable_free -= min(usable_free, reserved);
3231 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3232 return true;
3233 }
3234
3235 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3236 free_pages))
3237 return true;
3238
3239 /*
3240 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3241 * when checking the min watermark. The min watermark is the
3242 * point where boosting is ignored so that kswapd is woken up
3243 * when below the low watermark.
3244 */
3245 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3246 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3247 mark = z->_watermark[WMARK_MIN];
3248 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3249 alloc_flags, free_pages);
3250 }
3251
3252 return false;
3253 }
3254
zone_watermark_ok_safe(struct zone * z,unsigned int order,unsigned long mark,int highest_zoneidx)3255 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3256 unsigned long mark, int highest_zoneidx)
3257 {
3258 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3259
3260 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3261 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3262
3263 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3264 free_pages);
3265 }
3266
3267 #ifdef CONFIG_NUMA
3268 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3269
zone_allows_reclaim(struct zone * local_zone,struct zone * zone)3270 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3271 {
3272 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3273 node_reclaim_distance;
3274 }
3275 #else /* CONFIG_NUMA */
zone_allows_reclaim(struct zone * local_zone,struct zone * zone)3276 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3277 {
3278 return true;
3279 }
3280 #endif /* CONFIG_NUMA */
3281
3282 /*
3283 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3284 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3285 * premature use of a lower zone may cause lowmem pressure problems that
3286 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3287 * probably too small. It only makes sense to spread allocations to avoid
3288 * fragmentation between the Normal and DMA32 zones.
3289 */
3290 static inline unsigned int
alloc_flags_nofragment(struct zone * zone,gfp_t gfp_mask)3291 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3292 {
3293 unsigned int alloc_flags;
3294
3295 /*
3296 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3297 * to save a branch.
3298 */
3299 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3300
3301 #ifdef CONFIG_ZONE_DMA32
3302 if (!zone)
3303 return alloc_flags;
3304
3305 if (zone_idx(zone) != ZONE_NORMAL)
3306 return alloc_flags;
3307
3308 /*
3309 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3310 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3311 * on UMA that if Normal is populated then so is DMA32.
3312 */
3313 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3314 if (nr_online_nodes > 1 && !populated_zone(--zone))
3315 return alloc_flags;
3316
3317 alloc_flags |= ALLOC_NOFRAGMENT;
3318 #endif /* CONFIG_ZONE_DMA32 */
3319 return alloc_flags;
3320 }
3321
3322 /* 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)3323 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3324 unsigned int alloc_flags)
3325 {
3326 #ifdef CONFIG_CMA
3327 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3328 alloc_flags |= ALLOC_CMA;
3329 #endif
3330 return alloc_flags;
3331 }
3332
3333 /*
3334 * get_page_from_freelist goes through the zonelist trying to allocate
3335 * a page.
3336 */
3337 static struct page *
get_page_from_freelist(gfp_t gfp_mask,unsigned int order,int alloc_flags,const struct alloc_context * ac)3338 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3339 const struct alloc_context *ac)
3340 {
3341 struct zoneref *z;
3342 struct zone *zone;
3343 struct pglist_data *last_pgdat = NULL;
3344 bool last_pgdat_dirty_ok = false;
3345 bool no_fallback;
3346
3347 retry:
3348 /*
3349 * Scan zonelist, looking for a zone with enough free.
3350 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3351 */
3352 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3353 z = ac->preferred_zoneref;
3354 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3355 ac->nodemask) {
3356 struct page *page;
3357 unsigned long mark;
3358
3359 if (cpusets_enabled() &&
3360 (alloc_flags & ALLOC_CPUSET) &&
3361 !__cpuset_zone_allowed(zone, gfp_mask))
3362 continue;
3363 /*
3364 * When allocating a page cache page for writing, we
3365 * want to get it from a node that is within its dirty
3366 * limit, such that no single node holds more than its
3367 * proportional share of globally allowed dirty pages.
3368 * The dirty limits take into account the node's
3369 * lowmem reserves and high watermark so that kswapd
3370 * should be able to balance it without having to
3371 * write pages from its LRU list.
3372 *
3373 * XXX: For now, allow allocations to potentially
3374 * exceed the per-node dirty limit in the slowpath
3375 * (spread_dirty_pages unset) before going into reclaim,
3376 * which is important when on a NUMA setup the allowed
3377 * nodes are together not big enough to reach the
3378 * global limit. The proper fix for these situations
3379 * will require awareness of nodes in the
3380 * dirty-throttling and the flusher threads.
3381 */
3382 if (ac->spread_dirty_pages) {
3383 if (last_pgdat != zone->zone_pgdat) {
3384 last_pgdat = zone->zone_pgdat;
3385 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3386 }
3387
3388 if (!last_pgdat_dirty_ok)
3389 continue;
3390 }
3391
3392 if (no_fallback && nr_online_nodes > 1 &&
3393 zone != zonelist_zone(ac->preferred_zoneref)) {
3394 int local_nid;
3395
3396 /*
3397 * If moving to a remote node, retry but allow
3398 * fragmenting fallbacks. Locality is more important
3399 * than fragmentation avoidance.
3400 */
3401 local_nid = zonelist_node_idx(ac->preferred_zoneref);
3402 if (zone_to_nid(zone) != local_nid) {
3403 alloc_flags &= ~ALLOC_NOFRAGMENT;
3404 goto retry;
3405 }
3406 }
3407
3408 cond_accept_memory(zone, order);
3409
3410 /*
3411 * Detect whether the number of free pages is below high
3412 * watermark. If so, we will decrease pcp->high and free
3413 * PCP pages in free path to reduce the possibility of
3414 * premature page reclaiming. Detection is done here to
3415 * avoid to do that in hotter free path.
3416 */
3417 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3418 goto check_alloc_wmark;
3419
3420 mark = high_wmark_pages(zone);
3421 if (zone_watermark_fast(zone, order, mark,
3422 ac->highest_zoneidx, alloc_flags,
3423 gfp_mask))
3424 goto try_this_zone;
3425 else
3426 set_bit(ZONE_BELOW_HIGH, &zone->flags);
3427
3428 check_alloc_wmark:
3429 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3430 if (!zone_watermark_fast(zone, order, mark,
3431 ac->highest_zoneidx, alloc_flags,
3432 gfp_mask)) {
3433 int ret;
3434
3435 if (cond_accept_memory(zone, order))
3436 goto try_this_zone;
3437
3438 /*
3439 * Watermark failed for this zone, but see if we can
3440 * grow this zone if it contains deferred pages.
3441 */
3442 if (deferred_pages_enabled()) {
3443 if (_deferred_grow_zone(zone, order))
3444 goto try_this_zone;
3445 }
3446 /* Checked here to keep the fast path fast */
3447 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3448 if (alloc_flags & ALLOC_NO_WATERMARKS)
3449 goto try_this_zone;
3450
3451 if (!node_reclaim_enabled() ||
3452 !zone_allows_reclaim(zonelist_zone(ac->preferred_zoneref), zone))
3453 continue;
3454
3455 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3456 switch (ret) {
3457 case NODE_RECLAIM_NOSCAN:
3458 /* did not scan */
3459 continue;
3460 case NODE_RECLAIM_FULL:
3461 /* scanned but unreclaimable */
3462 continue;
3463 default:
3464 /* did we reclaim enough */
3465 if (zone_watermark_ok(zone, order, mark,
3466 ac->highest_zoneidx, alloc_flags))
3467 goto try_this_zone;
3468
3469 continue;
3470 }
3471 }
3472
3473 try_this_zone:
3474 page = rmqueue(zonelist_zone(ac->preferred_zoneref), zone, order,
3475 gfp_mask, alloc_flags, ac->migratetype);
3476 if (page) {
3477 prep_new_page(page, order, gfp_mask, alloc_flags);
3478
3479 /*
3480 * If this is a high-order atomic allocation then check
3481 * if the pageblock should be reserved for the future
3482 */
3483 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3484 reserve_highatomic_pageblock(page, order, zone);
3485
3486 return page;
3487 } else {
3488 if (cond_accept_memory(zone, order))
3489 goto try_this_zone;
3490
3491 /* Try again if zone has deferred pages */
3492 if (deferred_pages_enabled()) {
3493 if (_deferred_grow_zone(zone, order))
3494 goto try_this_zone;
3495 }
3496 }
3497 }
3498
3499 /*
3500 * It's possible on a UMA machine to get through all zones that are
3501 * fragmented. If avoiding fragmentation, reset and try again.
3502 */
3503 if (no_fallback) {
3504 alloc_flags &= ~ALLOC_NOFRAGMENT;
3505 goto retry;
3506 }
3507
3508 return NULL;
3509 }
3510
warn_alloc_show_mem(gfp_t gfp_mask,nodemask_t * nodemask)3511 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3512 {
3513 unsigned int filter = SHOW_MEM_FILTER_NODES;
3514
3515 /*
3516 * This documents exceptions given to allocations in certain
3517 * contexts that are allowed to allocate outside current's set
3518 * of allowed nodes.
3519 */
3520 if (!(gfp_mask & __GFP_NOMEMALLOC))
3521 if (tsk_is_oom_victim(current) ||
3522 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3523 filter &= ~SHOW_MEM_FILTER_NODES;
3524 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3525 filter &= ~SHOW_MEM_FILTER_NODES;
3526
3527 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3528 }
3529
warn_alloc(gfp_t gfp_mask,nodemask_t * nodemask,const char * fmt,...)3530 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3531 {
3532 struct va_format vaf;
3533 va_list args;
3534 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3535
3536 if ((gfp_mask & __GFP_NOWARN) ||
3537 !__ratelimit(&nopage_rs) ||
3538 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3539 return;
3540
3541 va_start(args, fmt);
3542 vaf.fmt = fmt;
3543 vaf.va = &args;
3544 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3545 current->comm, &vaf, gfp_mask, &gfp_mask,
3546 nodemask_pr_args(nodemask));
3547 va_end(args);
3548
3549 cpuset_print_current_mems_allowed();
3550 pr_cont("\n");
3551 dump_stack();
3552 warn_alloc_show_mem(gfp_mask, nodemask);
3553 }
3554
3555 static inline struct page *
__alloc_pages_cpuset_fallback(gfp_t gfp_mask,unsigned int order,unsigned int alloc_flags,const struct alloc_context * ac)3556 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3557 unsigned int alloc_flags,
3558 const struct alloc_context *ac)
3559 {
3560 struct page *page;
3561
3562 page = get_page_from_freelist(gfp_mask, order,
3563 alloc_flags|ALLOC_CPUSET, ac);
3564 /*
3565 * fallback to ignore cpuset restriction if our nodes
3566 * are depleted
3567 */
3568 if (!page)
3569 page = get_page_from_freelist(gfp_mask, order,
3570 alloc_flags, ac);
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_result = COMPACT_SKIPPED;
4247 compact_priority = DEF_COMPACT_PRIORITY;
4248 cpuset_mems_cookie = read_mems_allowed_begin();
4249 zonelist_iter_cookie = zonelist_iter_begin();
4250
4251 /*
4252 * The fast path uses conservative alloc_flags to succeed only until
4253 * kswapd needs to be woken up, and to avoid the cost of setting up
4254 * alloc_flags precisely. So we do that now.
4255 */
4256 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4257
4258 /*
4259 * We need to recalculate the starting point for the zonelist iterator
4260 * because we might have used different nodemask in the fast path, or
4261 * there was a cpuset modification and we are retrying - otherwise we
4262 * could end up iterating over non-eligible zones endlessly.
4263 */
4264 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4265 ac->highest_zoneidx, ac->nodemask);
4266 if (!zonelist_zone(ac->preferred_zoneref))
4267 goto nopage;
4268
4269 /*
4270 * Check for insane configurations where the cpuset doesn't contain
4271 * any suitable zone to satisfy the request - e.g. non-movable
4272 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4273 */
4274 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4275 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4276 ac->highest_zoneidx,
4277 &cpuset_current_mems_allowed);
4278 if (!zonelist_zone(z))
4279 goto nopage;
4280 }
4281
4282 if (alloc_flags & ALLOC_KSWAPD)
4283 wake_all_kswapds(order, gfp_mask, ac);
4284
4285 /*
4286 * The adjusted alloc_flags might result in immediate success, so try
4287 * that first
4288 */
4289 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4290 if (page)
4291 goto got_pg;
4292
4293 /*
4294 * For costly allocations, try direct compaction first, as it's likely
4295 * that we have enough base pages and don't need to reclaim. For non-
4296 * movable high-order allocations, do that as well, as compaction will
4297 * try prevent permanent fragmentation by migrating from blocks of the
4298 * same migratetype.
4299 * Don't try this for allocations that are allowed to ignore
4300 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4301 */
4302 if (can_direct_reclaim && can_compact &&
4303 (costly_order ||
4304 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4305 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4306 page = __alloc_pages_direct_compact(gfp_mask, order,
4307 alloc_flags, ac,
4308 INIT_COMPACT_PRIORITY,
4309 &compact_result);
4310 if (page)
4311 goto got_pg;
4312
4313 /*
4314 * Checks for costly allocations with __GFP_NORETRY, which
4315 * includes some THP page fault allocations
4316 */
4317 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4318 /*
4319 * If allocating entire pageblock(s) and compaction
4320 * failed because all zones are below low watermarks
4321 * or is prohibited because it recently failed at this
4322 * order, fail immediately unless the allocator has
4323 * requested compaction and reclaim retry.
4324 *
4325 * Reclaim is
4326 * - potentially very expensive because zones are far
4327 * below their low watermarks or this is part of very
4328 * bursty high order allocations,
4329 * - not guaranteed to help because isolate_freepages()
4330 * may not iterate over freed pages as part of its
4331 * linear scan, and
4332 * - unlikely to make entire pageblocks free on its
4333 * own.
4334 */
4335 if (compact_result == COMPACT_SKIPPED ||
4336 compact_result == COMPACT_DEFERRED)
4337 goto nopage;
4338
4339 /*
4340 * Looks like reclaim/compaction is worth trying, but
4341 * sync compaction could be very expensive, so keep
4342 * using async compaction.
4343 */
4344 compact_priority = INIT_COMPACT_PRIORITY;
4345 }
4346 }
4347
4348 retry:
4349 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4350 if (alloc_flags & ALLOC_KSWAPD)
4351 wake_all_kswapds(order, gfp_mask, ac);
4352
4353 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4354 if (reserve_flags)
4355 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4356 (alloc_flags & ALLOC_KSWAPD);
4357
4358 /*
4359 * Reset the nodemask and zonelist iterators if memory policies can be
4360 * ignored. These allocations are high priority and system rather than
4361 * user oriented.
4362 */
4363 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4364 ac->nodemask = NULL;
4365 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4366 ac->highest_zoneidx, ac->nodemask);
4367 }
4368
4369 /* Attempt with potentially adjusted zonelist and alloc_flags */
4370 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4371 if (page)
4372 goto got_pg;
4373
4374 /* Caller is not willing to reclaim, we can't balance anything */
4375 if (!can_direct_reclaim)
4376 goto nopage;
4377
4378 /* Avoid recursion of direct reclaim */
4379 if (current->flags & PF_MEMALLOC)
4380 goto nopage;
4381
4382 /* Try direct reclaim and then allocating */
4383 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4384 &did_some_progress);
4385 if (page)
4386 goto got_pg;
4387
4388 /* Try direct compaction and then allocating */
4389 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4390 compact_priority, &compact_result);
4391 if (page)
4392 goto got_pg;
4393
4394 /* Do not loop if specifically requested */
4395 if (gfp_mask & __GFP_NORETRY)
4396 goto nopage;
4397
4398 /*
4399 * Do not retry costly high order allocations unless they are
4400 * __GFP_RETRY_MAYFAIL and we can compact
4401 */
4402 if (costly_order && (!can_compact ||
4403 !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4404 goto nopage;
4405
4406 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4407 did_some_progress > 0, &no_progress_loops))
4408 goto retry;
4409
4410 /*
4411 * It doesn't make any sense to retry for the compaction if the order-0
4412 * reclaim is not able to make any progress because the current
4413 * implementation of the compaction depends on the sufficient amount
4414 * of free memory (see __compaction_suitable)
4415 */
4416 if (did_some_progress > 0 && can_compact &&
4417 should_compact_retry(ac, order, alloc_flags,
4418 compact_result, &compact_priority,
4419 &compaction_retries))
4420 goto retry;
4421
4422
4423 /*
4424 * Deal with possible cpuset update races or zonelist updates to avoid
4425 * a unnecessary OOM kill.
4426 */
4427 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4428 check_retry_zonelist(zonelist_iter_cookie))
4429 goto restart;
4430
4431 /* Reclaim has failed us, start killing things */
4432 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4433 if (page)
4434 goto got_pg;
4435
4436 /* Avoid allocations with no watermarks from looping endlessly */
4437 if (tsk_is_oom_victim(current) &&
4438 (alloc_flags & ALLOC_OOM ||
4439 (gfp_mask & __GFP_NOMEMALLOC)))
4440 goto nopage;
4441
4442 /* Retry as long as the OOM killer is making progress */
4443 if (did_some_progress) {
4444 no_progress_loops = 0;
4445 goto retry;
4446 }
4447
4448 nopage:
4449 /*
4450 * Deal with possible cpuset update races or zonelist updates to avoid
4451 * a unnecessary OOM kill.
4452 */
4453 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4454 check_retry_zonelist(zonelist_iter_cookie))
4455 goto restart;
4456
4457 /*
4458 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4459 * we always retry
4460 */
4461 if (unlikely(nofail)) {
4462 /*
4463 * Lacking direct_reclaim we can't do anything to reclaim memory,
4464 * we disregard these unreasonable nofail requests and still
4465 * return NULL
4466 */
4467 if (!can_direct_reclaim)
4468 goto fail;
4469
4470 /*
4471 * Help non-failing allocations by giving some access to memory
4472 * reserves normally used for high priority non-blocking
4473 * allocations but do not use ALLOC_NO_WATERMARKS because this
4474 * could deplete whole memory reserves which would just make
4475 * the situation worse.
4476 */
4477 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4478 if (page)
4479 goto got_pg;
4480
4481 cond_resched();
4482 goto retry;
4483 }
4484 fail:
4485 warn_alloc(gfp_mask, ac->nodemask,
4486 "page allocation failure: order:%u", order);
4487 got_pg:
4488 return page;
4489 }
4490
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)4491 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4492 int preferred_nid, nodemask_t *nodemask,
4493 struct alloc_context *ac, gfp_t *alloc_gfp,
4494 unsigned int *alloc_flags)
4495 {
4496 ac->highest_zoneidx = gfp_zone(gfp_mask);
4497 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4498 ac->nodemask = nodemask;
4499 ac->migratetype = gfp_migratetype(gfp_mask);
4500
4501 if (cpusets_enabled()) {
4502 *alloc_gfp |= __GFP_HARDWALL;
4503 /*
4504 * When we are in the interrupt context, it is irrelevant
4505 * to the current task context. It means that any node ok.
4506 */
4507 if (in_task() && !ac->nodemask)
4508 ac->nodemask = &cpuset_current_mems_allowed;
4509 else
4510 *alloc_flags |= ALLOC_CPUSET;
4511 }
4512
4513 might_alloc(gfp_mask);
4514
4515 if (should_fail_alloc_page(gfp_mask, order))
4516 return false;
4517
4518 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4519
4520 /* Dirty zone balancing only done in the fast path */
4521 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4522
4523 /*
4524 * The preferred zone is used for statistics but crucially it is
4525 * also used as the starting point for the zonelist iterator. It
4526 * may get reset for allocations that ignore memory policies.
4527 */
4528 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4529 ac->highest_zoneidx, ac->nodemask);
4530
4531 return true;
4532 }
4533
4534 /*
4535 * __alloc_pages_bulk - Allocate a number of order-0 pages to an array
4536 * @gfp: GFP flags for the allocation
4537 * @preferred_nid: The preferred NUMA node ID to allocate from
4538 * @nodemask: Set of nodes to allocate from, may be NULL
4539 * @nr_pages: The number of pages desired in the array
4540 * @page_array: 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 the page_array.
4544 *
4545 * Note that only NULL elements are populated with pages and nr_pages
4546 * is the maximum number of pages that will be stored in the array.
4547 *
4548 * Returns the number of pages in the array.
4549 */
alloc_pages_bulk_noprof(gfp_t gfp,int preferred_nid,nodemask_t * nodemask,int nr_pages,struct page ** page_array)4550 unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
4551 nodemask_t *nodemask, int nr_pages,
4552 struct page **page_array)
4553 {
4554 struct page *page;
4555 unsigned long __maybe_unused UP_flags;
4556 struct zone *zone;
4557 struct zoneref *z;
4558 struct per_cpu_pages *pcp;
4559 struct list_head *pcp_list;
4560 struct alloc_context ac;
4561 gfp_t alloc_gfp;
4562 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4563 int nr_populated = 0, nr_account = 0;
4564
4565 /*
4566 * Skip populated array elements to determine if any pages need
4567 * to be allocated before disabling IRQs.
4568 */
4569 while (nr_populated < nr_pages && page_array[nr_populated])
4570 nr_populated++;
4571
4572 /* No pages requested? */
4573 if (unlikely(nr_pages <= 0))
4574 goto out;
4575
4576 /* Already populated array? */
4577 if (unlikely(nr_pages - nr_populated == 0))
4578 goto out;
4579
4580 /* Bulk allocator does not support memcg accounting. */
4581 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4582 goto failed;
4583
4584 /* Use the single page allocator for one page. */
4585 if (nr_pages - nr_populated == 1)
4586 goto failed;
4587
4588 #ifdef CONFIG_PAGE_OWNER
4589 /*
4590 * PAGE_OWNER may recurse into the allocator to allocate space to
4591 * save the stack with pagesets.lock held. Releasing/reacquiring
4592 * removes much of the performance benefit of bulk allocation so
4593 * force the caller to allocate one page at a time as it'll have
4594 * similar performance to added complexity to the bulk allocator.
4595 */
4596 if (static_branch_unlikely(&page_owner_inited))
4597 goto failed;
4598 #endif
4599
4600 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4601 gfp &= gfp_allowed_mask;
4602 alloc_gfp = gfp;
4603 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4604 goto out;
4605 gfp = alloc_gfp;
4606
4607 /* Find an allowed local zone that meets the low watermark. */
4608 z = ac.preferred_zoneref;
4609 for_next_zone_zonelist_nodemask(zone, z, ac.highest_zoneidx, ac.nodemask) {
4610 unsigned long mark;
4611
4612 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4613 !__cpuset_zone_allowed(zone, gfp)) {
4614 continue;
4615 }
4616
4617 if (nr_online_nodes > 1 && zone != zonelist_zone(ac.preferred_zoneref) &&
4618 zone_to_nid(zone) != zonelist_node_idx(ac.preferred_zoneref)) {
4619 goto failed;
4620 }
4621
4622 cond_accept_memory(zone, 0);
4623 retry_this_zone:
4624 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4625 if (zone_watermark_fast(zone, 0, mark,
4626 zonelist_zone_idx(ac.preferred_zoneref),
4627 alloc_flags, gfp)) {
4628 break;
4629 }
4630
4631 if (cond_accept_memory(zone, 0))
4632 goto retry_this_zone;
4633
4634 /* Try again if zone has deferred pages */
4635 if (deferred_pages_enabled()) {
4636 if (_deferred_grow_zone(zone, 0))
4637 goto retry_this_zone;
4638 }
4639 }
4640
4641 /*
4642 * If there are no allowed local zones that meets the watermarks then
4643 * try to allocate a single page and reclaim if necessary.
4644 */
4645 if (unlikely(!zone))
4646 goto failed;
4647
4648 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4649 pcp_trylock_prepare(UP_flags);
4650 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4651 if (!pcp)
4652 goto failed_irq;
4653
4654 /* Attempt the batch allocation */
4655 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4656 while (nr_populated < nr_pages) {
4657
4658 /* Skip existing pages */
4659 if (page_array[nr_populated]) {
4660 nr_populated++;
4661 continue;
4662 }
4663
4664 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4665 pcp, pcp_list);
4666 if (unlikely(!page)) {
4667 /* Try and allocate at least one page */
4668 if (!nr_account) {
4669 pcp_spin_unlock(pcp);
4670 goto failed_irq;
4671 }
4672 break;
4673 }
4674 nr_account++;
4675
4676 prep_new_page(page, 0, gfp, 0);
4677 set_page_refcounted(page);
4678 page_array[nr_populated++] = page;
4679 }
4680
4681 pcp_spin_unlock(pcp);
4682 pcp_trylock_finish(UP_flags);
4683
4684 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4685 zone_statistics(zonelist_zone(ac.preferred_zoneref), zone, nr_account);
4686
4687 out:
4688 return nr_populated;
4689
4690 failed_irq:
4691 pcp_trylock_finish(UP_flags);
4692
4693 failed:
4694 page = __alloc_pages_noprof(gfp, 0, preferred_nid, nodemask);
4695 if (page)
4696 page_array[nr_populated++] = page;
4697 goto out;
4698 }
4699 EXPORT_SYMBOL_GPL(alloc_pages_bulk_noprof);
4700
4701 /*
4702 * This is the 'heart' of the zoned buddy allocator.
4703 */
__alloc_frozen_pages_noprof(gfp_t gfp,unsigned int order,int preferred_nid,nodemask_t * nodemask)4704 struct page *__alloc_frozen_pages_noprof(gfp_t gfp, unsigned int order,
4705 int preferred_nid, nodemask_t *nodemask)
4706 {
4707 struct page *page;
4708 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4709 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4710 struct alloc_context ac = { };
4711
4712 /*
4713 * There are several places where we assume that the order value is sane
4714 * so bail out early if the request is out of bound.
4715 */
4716 if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
4717 return NULL;
4718
4719 gfp &= gfp_allowed_mask;
4720 /*
4721 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4722 * resp. GFP_NOIO which has to be inherited for all allocation requests
4723 * from a particular context which has been marked by
4724 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4725 * movable zones are not used during allocation.
4726 */
4727 gfp = current_gfp_context(gfp);
4728 alloc_gfp = gfp;
4729 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4730 &alloc_gfp, &alloc_flags))
4731 return NULL;
4732
4733 /*
4734 * Forbid the first pass from falling back to types that fragment
4735 * memory until all local zones are considered.
4736 */
4737 alloc_flags |= alloc_flags_nofragment(zonelist_zone(ac.preferred_zoneref), gfp);
4738
4739 /* First allocation attempt */
4740 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4741 if (likely(page))
4742 goto out;
4743
4744 alloc_gfp = gfp;
4745 ac.spread_dirty_pages = false;
4746
4747 /*
4748 * Restore the original nodemask if it was potentially replaced with
4749 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4750 */
4751 ac.nodemask = nodemask;
4752
4753 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4754
4755 out:
4756 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4757 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4758 free_frozen_pages(page, order);
4759 page = NULL;
4760 }
4761
4762 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4763 kmsan_alloc_page(page, order, alloc_gfp);
4764
4765 return page;
4766 }
4767 EXPORT_SYMBOL(__alloc_frozen_pages_noprof);
4768
__alloc_pages_noprof(gfp_t gfp,unsigned int order,int preferred_nid,nodemask_t * nodemask)4769 struct page *__alloc_pages_noprof(gfp_t gfp, unsigned int order,
4770 int preferred_nid, nodemask_t *nodemask)
4771 {
4772 struct page *page;
4773
4774 page = __alloc_frozen_pages_noprof(gfp, order, preferred_nid, nodemask);
4775 if (page)
4776 set_page_refcounted(page);
4777 return page;
4778 }
4779 EXPORT_SYMBOL(__alloc_pages_noprof);
4780
__folio_alloc_noprof(gfp_t gfp,unsigned int order,int preferred_nid,nodemask_t * nodemask)4781 struct folio *__folio_alloc_noprof(gfp_t gfp, unsigned int order, int preferred_nid,
4782 nodemask_t *nodemask)
4783 {
4784 struct page *page = __alloc_pages_noprof(gfp | __GFP_COMP, order,
4785 preferred_nid, nodemask);
4786 return page_rmappable_folio(page);
4787 }
4788 EXPORT_SYMBOL(__folio_alloc_noprof);
4789
4790 /*
4791 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4792 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4793 * you need to access high mem.
4794 */
get_free_pages_noprof(gfp_t gfp_mask,unsigned int order)4795 unsigned long get_free_pages_noprof(gfp_t gfp_mask, unsigned int order)
4796 {
4797 struct page *page;
4798
4799 page = alloc_pages_noprof(gfp_mask & ~__GFP_HIGHMEM, order);
4800 if (!page)
4801 return 0;
4802 return (unsigned long) page_address(page);
4803 }
4804 EXPORT_SYMBOL(get_free_pages_noprof);
4805
get_zeroed_page_noprof(gfp_t gfp_mask)4806 unsigned long get_zeroed_page_noprof(gfp_t gfp_mask)
4807 {
4808 return get_free_pages_noprof(gfp_mask | __GFP_ZERO, 0);
4809 }
4810 EXPORT_SYMBOL(get_zeroed_page_noprof);
4811
4812 /**
4813 * __free_pages - Free pages allocated with alloc_pages().
4814 * @page: The page pointer returned from alloc_pages().
4815 * @order: The order of the allocation.
4816 *
4817 * This function can free multi-page allocations that are not compound
4818 * pages. It does not check that the @order passed in matches that of
4819 * the allocation, so it is easy to leak memory. Freeing more memory
4820 * than was allocated will probably emit a warning.
4821 *
4822 * If the last reference to this page is speculative, it will be released
4823 * by put_page() which only frees the first page of a non-compound
4824 * allocation. To prevent the remaining pages from being leaked, we free
4825 * the subsequent pages here. If you want to use the page's reference
4826 * count to decide when to free the allocation, you should allocate a
4827 * compound page, and use put_page() instead of __free_pages().
4828 *
4829 * Context: May be called in interrupt context or while holding a normal
4830 * spinlock, but not in NMI context or while holding a raw spinlock.
4831 */
__free_pages(struct page * page,unsigned int order)4832 void __free_pages(struct page *page, unsigned int order)
4833 {
4834 /* get PageHead before we drop reference */
4835 int head = PageHead(page);
4836 struct alloc_tag *tag = pgalloc_tag_get(page);
4837
4838 if (put_page_testzero(page))
4839 free_frozen_pages(page, order);
4840 else if (!head) {
4841 pgalloc_tag_sub_pages(tag, (1 << order) - 1);
4842 while (order-- > 0)
4843 free_frozen_pages(page + (1 << order), order);
4844 }
4845 }
4846 EXPORT_SYMBOL(__free_pages);
4847
free_pages(unsigned long addr,unsigned int order)4848 void free_pages(unsigned long addr, unsigned int order)
4849 {
4850 if (addr != 0) {
4851 VM_BUG_ON(!virt_addr_valid((void *)addr));
4852 __free_pages(virt_to_page((void *)addr), order);
4853 }
4854 }
4855
4856 EXPORT_SYMBOL(free_pages);
4857
make_alloc_exact(unsigned long addr,unsigned int order,size_t size)4858 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4859 size_t size)
4860 {
4861 if (addr) {
4862 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4863 struct page *page = virt_to_page((void *)addr);
4864 struct page *last = page + nr;
4865
4866 split_page_owner(page, order, 0);
4867 pgalloc_tag_split(page_folio(page), order, 0);
4868 split_page_memcg(page, order, 0);
4869 while (page < --last)
4870 set_page_refcounted(last);
4871
4872 last = page + (1UL << order);
4873 for (page += nr; page < last; page++)
4874 __free_pages_ok(page, 0, FPI_TO_TAIL);
4875 }
4876 return (void *)addr;
4877 }
4878
4879 /**
4880 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4881 * @size: the number of bytes to allocate
4882 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4883 *
4884 * This function is similar to alloc_pages(), except that it allocates the
4885 * minimum number of pages to satisfy the request. alloc_pages() can only
4886 * allocate memory in power-of-two pages.
4887 *
4888 * This function is also limited by MAX_PAGE_ORDER.
4889 *
4890 * Memory allocated by this function must be released by free_pages_exact().
4891 *
4892 * Return: pointer to the allocated area or %NULL in case of error.
4893 */
alloc_pages_exact_noprof(size_t size,gfp_t gfp_mask)4894 void *alloc_pages_exact_noprof(size_t size, gfp_t gfp_mask)
4895 {
4896 unsigned int order = get_order(size);
4897 unsigned long addr;
4898
4899 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4900 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4901
4902 addr = get_free_pages_noprof(gfp_mask, order);
4903 return make_alloc_exact(addr, order, size);
4904 }
4905 EXPORT_SYMBOL(alloc_pages_exact_noprof);
4906
4907 /**
4908 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4909 * pages on a node.
4910 * @nid: the preferred node ID where memory should be allocated
4911 * @size: the number of bytes to allocate
4912 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4913 *
4914 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4915 * back.
4916 *
4917 * Return: pointer to the allocated area or %NULL in case of error.
4918 */
alloc_pages_exact_nid_noprof(int nid,size_t size,gfp_t gfp_mask)4919 void * __meminit alloc_pages_exact_nid_noprof(int nid, size_t size, gfp_t gfp_mask)
4920 {
4921 unsigned int order = get_order(size);
4922 struct page *p;
4923
4924 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4925 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4926
4927 p = alloc_pages_node_noprof(nid, gfp_mask, order);
4928 if (!p)
4929 return NULL;
4930 return make_alloc_exact((unsigned long)page_address(p), order, size);
4931 }
4932
4933 /**
4934 * free_pages_exact - release memory allocated via alloc_pages_exact()
4935 * @virt: the value returned by alloc_pages_exact.
4936 * @size: size of allocation, same value as passed to alloc_pages_exact().
4937 *
4938 * Release the memory allocated by a previous call to alloc_pages_exact.
4939 */
free_pages_exact(void * virt,size_t size)4940 void free_pages_exact(void *virt, size_t size)
4941 {
4942 unsigned long addr = (unsigned long)virt;
4943 unsigned long end = addr + PAGE_ALIGN(size);
4944
4945 while (addr < end) {
4946 free_page(addr);
4947 addr += PAGE_SIZE;
4948 }
4949 }
4950 EXPORT_SYMBOL(free_pages_exact);
4951
4952 /**
4953 * nr_free_zone_pages - count number of pages beyond high watermark
4954 * @offset: The zone index of the highest zone
4955 *
4956 * nr_free_zone_pages() counts the number of pages which are beyond the
4957 * high watermark within all zones at or below a given zone index. For each
4958 * zone, the number of pages is calculated as:
4959 *
4960 * nr_free_zone_pages = managed_pages - high_pages
4961 *
4962 * Return: number of pages beyond high watermark.
4963 */
nr_free_zone_pages(int offset)4964 static unsigned long nr_free_zone_pages(int offset)
4965 {
4966 struct zoneref *z;
4967 struct zone *zone;
4968
4969 /* Just pick one node, since fallback list is circular */
4970 unsigned long sum = 0;
4971
4972 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4973
4974 for_each_zone_zonelist(zone, z, zonelist, offset) {
4975 unsigned long size = zone_managed_pages(zone);
4976 unsigned long high = high_wmark_pages(zone);
4977 if (size > high)
4978 sum += size - high;
4979 }
4980
4981 return sum;
4982 }
4983
4984 /**
4985 * nr_free_buffer_pages - count number of pages beyond high watermark
4986 *
4987 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4988 * watermark within ZONE_DMA and ZONE_NORMAL.
4989 *
4990 * Return: number of pages beyond high watermark within ZONE_DMA and
4991 * ZONE_NORMAL.
4992 */
nr_free_buffer_pages(void)4993 unsigned long nr_free_buffer_pages(void)
4994 {
4995 return nr_free_zone_pages(gfp_zone(GFP_USER));
4996 }
4997 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4998
zoneref_set_zone(struct zone * zone,struct zoneref * zoneref)4999 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5000 {
5001 zoneref->zone = zone;
5002 zoneref->zone_idx = zone_idx(zone);
5003 }
5004
5005 /*
5006 * Builds allocation fallback zone lists.
5007 *
5008 * Add all populated zones of a node to the zonelist.
5009 */
build_zonerefs_node(pg_data_t * pgdat,struct zoneref * zonerefs)5010 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5011 {
5012 struct zone *zone;
5013 enum zone_type zone_type = MAX_NR_ZONES;
5014 int nr_zones = 0;
5015
5016 do {
5017 zone_type--;
5018 zone = pgdat->node_zones + zone_type;
5019 if (populated_zone(zone)) {
5020 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5021 check_highest_zone(zone_type);
5022 }
5023 } while (zone_type);
5024
5025 return nr_zones;
5026 }
5027
5028 #ifdef CONFIG_NUMA
5029
__parse_numa_zonelist_order(char * s)5030 static int __parse_numa_zonelist_order(char *s)
5031 {
5032 /*
5033 * We used to support different zonelists modes but they turned
5034 * out to be just not useful. Let's keep the warning in place
5035 * if somebody still use the cmd line parameter so that we do
5036 * not fail it silently
5037 */
5038 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5039 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5040 return -EINVAL;
5041 }
5042 return 0;
5043 }
5044
5045 static char numa_zonelist_order[] = "Node";
5046 #define NUMA_ZONELIST_ORDER_LEN 16
5047 /*
5048 * sysctl handler for numa_zonelist_order
5049 */
numa_zonelist_order_handler(const struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)5050 static int numa_zonelist_order_handler(const struct ctl_table *table, int write,
5051 void *buffer, size_t *length, loff_t *ppos)
5052 {
5053 if (write)
5054 return __parse_numa_zonelist_order(buffer);
5055 return proc_dostring(table, write, buffer, length, ppos);
5056 }
5057
5058 static int node_load[MAX_NUMNODES];
5059
5060 /**
5061 * find_next_best_node - find the next node that should appear in a given node's fallback list
5062 * @node: node whose fallback list we're appending
5063 * @used_node_mask: nodemask_t of already used nodes
5064 *
5065 * We use a number of factors to determine which is the next node that should
5066 * appear on a given node's fallback list. The node should not have appeared
5067 * already in @node's fallback list, and it should be the next closest node
5068 * according to the distance array (which contains arbitrary distance values
5069 * from each node to each node in the system), and should also prefer nodes
5070 * with no CPUs, since presumably they'll have very little allocation pressure
5071 * on them otherwise.
5072 *
5073 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5074 */
find_next_best_node(int node,nodemask_t * used_node_mask)5075 int find_next_best_node(int node, nodemask_t *used_node_mask)
5076 {
5077 int n, val;
5078 int min_val = INT_MAX;
5079 int best_node = NUMA_NO_NODE;
5080
5081 /*
5082 * Use the local node if we haven't already, but for memoryless local
5083 * node, we should skip it and fall back to other nodes.
5084 */
5085 if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5086 node_set(node, *used_node_mask);
5087 return node;
5088 }
5089
5090 for_each_node_state(n, N_MEMORY) {
5091
5092 /* Don't want a node to appear more than once */
5093 if (node_isset(n, *used_node_mask))
5094 continue;
5095
5096 /* Use the distance array to find the distance */
5097 val = node_distance(node, n);
5098
5099 /* Penalize nodes under us ("prefer the next node") */
5100 val += (n < node);
5101
5102 /* Give preference to headless and unused nodes */
5103 if (!cpumask_empty(cpumask_of_node(n)))
5104 val += PENALTY_FOR_NODE_WITH_CPUS;
5105
5106 /* Slight preference for less loaded node */
5107 val *= MAX_NUMNODES;
5108 val += node_load[n];
5109
5110 if (val < min_val) {
5111 min_val = val;
5112 best_node = n;
5113 }
5114 }
5115
5116 if (best_node >= 0)
5117 node_set(best_node, *used_node_mask);
5118
5119 return best_node;
5120 }
5121
5122
5123 /*
5124 * Build zonelists ordered by node and zones within node.
5125 * This results in maximum locality--normal zone overflows into local
5126 * DMA zone, if any--but risks exhausting DMA zone.
5127 */
build_zonelists_in_node_order(pg_data_t * pgdat,int * node_order,unsigned nr_nodes)5128 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5129 unsigned nr_nodes)
5130 {
5131 struct zoneref *zonerefs;
5132 int i;
5133
5134 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5135
5136 for (i = 0; i < nr_nodes; i++) {
5137 int nr_zones;
5138
5139 pg_data_t *node = NODE_DATA(node_order[i]);
5140
5141 nr_zones = build_zonerefs_node(node, zonerefs);
5142 zonerefs += nr_zones;
5143 }
5144 zonerefs->zone = NULL;
5145 zonerefs->zone_idx = 0;
5146 }
5147
5148 /*
5149 * Build __GFP_THISNODE zonelists
5150 */
build_thisnode_zonelists(pg_data_t * pgdat)5151 static void build_thisnode_zonelists(pg_data_t *pgdat)
5152 {
5153 struct zoneref *zonerefs;
5154 int nr_zones;
5155
5156 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5157 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5158 zonerefs += nr_zones;
5159 zonerefs->zone = NULL;
5160 zonerefs->zone_idx = 0;
5161 }
5162
build_zonelists(pg_data_t * pgdat)5163 static void build_zonelists(pg_data_t *pgdat)
5164 {
5165 static int node_order[MAX_NUMNODES];
5166 int node, nr_nodes = 0;
5167 nodemask_t used_mask = NODE_MASK_NONE;
5168 int local_node, prev_node;
5169
5170 /* NUMA-aware ordering of nodes */
5171 local_node = pgdat->node_id;
5172 prev_node = local_node;
5173
5174 memset(node_order, 0, sizeof(node_order));
5175 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5176 /*
5177 * We don't want to pressure a particular node.
5178 * So adding penalty to the first node in same
5179 * distance group to make it round-robin.
5180 */
5181 if (node_distance(local_node, node) !=
5182 node_distance(local_node, prev_node))
5183 node_load[node] += 1;
5184
5185 node_order[nr_nodes++] = node;
5186 prev_node = node;
5187 }
5188
5189 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5190 build_thisnode_zonelists(pgdat);
5191 pr_info("Fallback order for Node %d: ", local_node);
5192 for (node = 0; node < nr_nodes; node++)
5193 pr_cont("%d ", node_order[node]);
5194 pr_cont("\n");
5195 }
5196
5197 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5198 /*
5199 * Return node id of node used for "local" allocations.
5200 * I.e., first node id of first zone in arg node's generic zonelist.
5201 * Used for initializing percpu 'numa_mem', which is used primarily
5202 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5203 */
local_memory_node(int node)5204 int local_memory_node(int node)
5205 {
5206 struct zoneref *z;
5207
5208 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5209 gfp_zone(GFP_KERNEL),
5210 NULL);
5211 return zonelist_node_idx(z);
5212 }
5213 #endif
5214
5215 static void setup_min_unmapped_ratio(void);
5216 static void setup_min_slab_ratio(void);
5217 #else /* CONFIG_NUMA */
5218
build_zonelists(pg_data_t * pgdat)5219 static void build_zonelists(pg_data_t *pgdat)
5220 {
5221 struct zoneref *zonerefs;
5222 int nr_zones;
5223
5224 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5225 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5226 zonerefs += nr_zones;
5227
5228 zonerefs->zone = NULL;
5229 zonerefs->zone_idx = 0;
5230 }
5231
5232 #endif /* CONFIG_NUMA */
5233
5234 /*
5235 * Boot pageset table. One per cpu which is going to be used for all
5236 * zones and all nodes. The parameters will be set in such a way
5237 * that an item put on a list will immediately be handed over to
5238 * the buddy list. This is safe since pageset manipulation is done
5239 * with interrupts disabled.
5240 *
5241 * The boot_pagesets must be kept even after bootup is complete for
5242 * unused processors and/or zones. They do play a role for bootstrapping
5243 * hotplugged processors.
5244 *
5245 * zoneinfo_show() and maybe other functions do
5246 * not check if the processor is online before following the pageset pointer.
5247 * Other parts of the kernel may not check if the zone is available.
5248 */
5249 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5250 /* These effectively disable the pcplists in the boot pageset completely */
5251 #define BOOT_PAGESET_HIGH 0
5252 #define BOOT_PAGESET_BATCH 1
5253 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5254 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5255
__build_all_zonelists(void * data)5256 static void __build_all_zonelists(void *data)
5257 {
5258 int nid;
5259 int __maybe_unused cpu;
5260 pg_data_t *self = data;
5261 unsigned long flags;
5262
5263 /*
5264 * The zonelist_update_seq must be acquired with irqsave because the
5265 * reader can be invoked from IRQ with GFP_ATOMIC.
5266 */
5267 write_seqlock_irqsave(&zonelist_update_seq, flags);
5268 /*
5269 * Also disable synchronous printk() to prevent any printk() from
5270 * trying to hold port->lock, for
5271 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5272 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5273 */
5274 printk_deferred_enter();
5275
5276 #ifdef CONFIG_NUMA
5277 memset(node_load, 0, sizeof(node_load));
5278 #endif
5279
5280 /*
5281 * This node is hotadded and no memory is yet present. So just
5282 * building zonelists is fine - no need to touch other nodes.
5283 */
5284 if (self && !node_online(self->node_id)) {
5285 build_zonelists(self);
5286 } else {
5287 /*
5288 * All possible nodes have pgdat preallocated
5289 * in free_area_init
5290 */
5291 for_each_node(nid) {
5292 pg_data_t *pgdat = NODE_DATA(nid);
5293
5294 build_zonelists(pgdat);
5295 }
5296
5297 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5298 /*
5299 * We now know the "local memory node" for each node--
5300 * i.e., the node of the first zone in the generic zonelist.
5301 * Set up numa_mem percpu variable for on-line cpus. During
5302 * boot, only the boot cpu should be on-line; we'll init the
5303 * secondary cpus' numa_mem as they come on-line. During
5304 * node/memory hotplug, we'll fixup all on-line cpus.
5305 */
5306 for_each_online_cpu(cpu)
5307 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5308 #endif
5309 }
5310
5311 printk_deferred_exit();
5312 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5313 }
5314
5315 static noinline void __init
build_all_zonelists_init(void)5316 build_all_zonelists_init(void)
5317 {
5318 int cpu;
5319
5320 __build_all_zonelists(NULL);
5321
5322 /*
5323 * Initialize the boot_pagesets that are going to be used
5324 * for bootstrapping processors. The real pagesets for
5325 * each zone will be allocated later when the per cpu
5326 * allocator is available.
5327 *
5328 * boot_pagesets are used also for bootstrapping offline
5329 * cpus if the system is already booted because the pagesets
5330 * are needed to initialize allocators on a specific cpu too.
5331 * F.e. the percpu allocator needs the page allocator which
5332 * needs the percpu allocator in order to allocate its pagesets
5333 * (a chicken-egg dilemma).
5334 */
5335 for_each_possible_cpu(cpu)
5336 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5337
5338 mminit_verify_zonelist();
5339 cpuset_init_current_mems_allowed();
5340 }
5341
5342 /*
5343 * unless system_state == SYSTEM_BOOTING.
5344 *
5345 * __ref due to call of __init annotated helper build_all_zonelists_init
5346 * [protected by SYSTEM_BOOTING].
5347 */
build_all_zonelists(pg_data_t * pgdat)5348 void __ref build_all_zonelists(pg_data_t *pgdat)
5349 {
5350 unsigned long vm_total_pages;
5351
5352 if (system_state == SYSTEM_BOOTING) {
5353 build_all_zonelists_init();
5354 } else {
5355 __build_all_zonelists(pgdat);
5356 /* cpuset refresh routine should be here */
5357 }
5358 /* Get the number of free pages beyond high watermark in all zones. */
5359 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5360 /*
5361 * Disable grouping by mobility if the number of pages in the
5362 * system is too low to allow the mechanism to work. It would be
5363 * more accurate, but expensive to check per-zone. This check is
5364 * made on memory-hotadd so a system can start with mobility
5365 * disabled and enable it later
5366 */
5367 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5368 page_group_by_mobility_disabled = 1;
5369 else
5370 page_group_by_mobility_disabled = 0;
5371
5372 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5373 nr_online_nodes,
5374 str_off_on(page_group_by_mobility_disabled),
5375 vm_total_pages);
5376 #ifdef CONFIG_NUMA
5377 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5378 #endif
5379 }
5380
zone_batchsize(struct zone * zone)5381 static int zone_batchsize(struct zone *zone)
5382 {
5383 #ifdef CONFIG_MMU
5384 int batch;
5385
5386 /*
5387 * The number of pages to batch allocate is either ~0.1%
5388 * of the zone or 1MB, whichever is smaller. The batch
5389 * size is striking a balance between allocation latency
5390 * and zone lock contention.
5391 */
5392 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5393 batch /= 4; /* We effectively *= 4 below */
5394 if (batch < 1)
5395 batch = 1;
5396
5397 /*
5398 * Clamp the batch to a 2^n - 1 value. Having a power
5399 * of 2 value was found to be more likely to have
5400 * suboptimal cache aliasing properties in some cases.
5401 *
5402 * For example if 2 tasks are alternately allocating
5403 * batches of pages, one task can end up with a lot
5404 * of pages of one half of the possible page colors
5405 * and the other with pages of the other colors.
5406 */
5407 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5408
5409 return batch;
5410
5411 #else
5412 /* The deferral and batching of frees should be suppressed under NOMMU
5413 * conditions.
5414 *
5415 * The problem is that NOMMU needs to be able to allocate large chunks
5416 * of contiguous memory as there's no hardware page translation to
5417 * assemble apparent contiguous memory from discontiguous pages.
5418 *
5419 * Queueing large contiguous runs of pages for batching, however,
5420 * causes the pages to actually be freed in smaller chunks. As there
5421 * can be a significant delay between the individual batches being
5422 * recycled, this leads to the once large chunks of space being
5423 * fragmented and becoming unavailable for high-order allocations.
5424 */
5425 return 0;
5426 #endif
5427 }
5428
5429 static int percpu_pagelist_high_fraction;
zone_highsize(struct zone * zone,int batch,int cpu_online,int high_fraction)5430 static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5431 int high_fraction)
5432 {
5433 #ifdef CONFIG_MMU
5434 int high;
5435 int nr_split_cpus;
5436 unsigned long total_pages;
5437
5438 if (!high_fraction) {
5439 /*
5440 * By default, the high value of the pcp is based on the zone
5441 * low watermark so that if they are full then background
5442 * reclaim will not be started prematurely.
5443 */
5444 total_pages = low_wmark_pages(zone);
5445 } else {
5446 /*
5447 * If percpu_pagelist_high_fraction is configured, the high
5448 * value is based on a fraction of the managed pages in the
5449 * zone.
5450 */
5451 total_pages = zone_managed_pages(zone) / high_fraction;
5452 }
5453
5454 /*
5455 * Split the high value across all online CPUs local to the zone. Note
5456 * that early in boot that CPUs may not be online yet and that during
5457 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5458 * onlined. For memory nodes that have no CPUs, split the high value
5459 * across all online CPUs to mitigate the risk that reclaim is triggered
5460 * prematurely due to pages stored on pcp lists.
5461 */
5462 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5463 if (!nr_split_cpus)
5464 nr_split_cpus = num_online_cpus();
5465 high = total_pages / nr_split_cpus;
5466
5467 /*
5468 * Ensure high is at least batch*4. The multiple is based on the
5469 * historical relationship between high and batch.
5470 */
5471 high = max(high, batch << 2);
5472
5473 return high;
5474 #else
5475 return 0;
5476 #endif
5477 }
5478
5479 /*
5480 * pcp->high and pcp->batch values are related and generally batch is lower
5481 * than high. They are also related to pcp->count such that count is lower
5482 * than high, and as soon as it reaches high, the pcplist is flushed.
5483 *
5484 * However, guaranteeing these relations at all times would require e.g. write
5485 * barriers here but also careful usage of read barriers at the read side, and
5486 * thus be prone to error and bad for performance. Thus the update only prevents
5487 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5488 * should ensure they can cope with those fields changing asynchronously, and
5489 * fully trust only the pcp->count field on the local CPU with interrupts
5490 * disabled.
5491 *
5492 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5493 * outside of boot time (or some other assurance that no concurrent updaters
5494 * exist).
5495 */
pageset_update(struct per_cpu_pages * pcp,unsigned long high_min,unsigned long high_max,unsigned long batch)5496 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5497 unsigned long high_max, unsigned long batch)
5498 {
5499 WRITE_ONCE(pcp->batch, batch);
5500 WRITE_ONCE(pcp->high_min, high_min);
5501 WRITE_ONCE(pcp->high_max, high_max);
5502 }
5503
per_cpu_pages_init(struct per_cpu_pages * pcp,struct per_cpu_zonestat * pzstats)5504 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5505 {
5506 int pindex;
5507
5508 memset(pcp, 0, sizeof(*pcp));
5509 memset(pzstats, 0, sizeof(*pzstats));
5510
5511 spin_lock_init(&pcp->lock);
5512 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5513 INIT_LIST_HEAD(&pcp->lists[pindex]);
5514
5515 /*
5516 * Set batch and high values safe for a boot pageset. A true percpu
5517 * pageset's initialization will update them subsequently. Here we don't
5518 * need to be as careful as pageset_update() as nobody can access the
5519 * pageset yet.
5520 */
5521 pcp->high_min = BOOT_PAGESET_HIGH;
5522 pcp->high_max = BOOT_PAGESET_HIGH;
5523 pcp->batch = BOOT_PAGESET_BATCH;
5524 pcp->free_count = 0;
5525 }
5526
__zone_set_pageset_high_and_batch(struct zone * zone,unsigned long high_min,unsigned long high_max,unsigned long batch)5527 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
5528 unsigned long high_max, unsigned long batch)
5529 {
5530 struct per_cpu_pages *pcp;
5531 int cpu;
5532
5533 for_each_possible_cpu(cpu) {
5534 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5535 pageset_update(pcp, high_min, high_max, batch);
5536 }
5537 }
5538
5539 /*
5540 * Calculate and set new high and batch values for all per-cpu pagesets of a
5541 * zone based on the zone's size.
5542 */
zone_set_pageset_high_and_batch(struct zone * zone,int cpu_online)5543 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5544 {
5545 int new_high_min, new_high_max, new_batch;
5546
5547 new_batch = max(1, zone_batchsize(zone));
5548 if (percpu_pagelist_high_fraction) {
5549 new_high_min = zone_highsize(zone, new_batch, cpu_online,
5550 percpu_pagelist_high_fraction);
5551 /*
5552 * PCP high is tuned manually, disable auto-tuning via
5553 * setting high_min and high_max to the manual value.
5554 */
5555 new_high_max = new_high_min;
5556 } else {
5557 new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
5558 new_high_max = zone_highsize(zone, new_batch, cpu_online,
5559 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
5560 }
5561
5562 if (zone->pageset_high_min == new_high_min &&
5563 zone->pageset_high_max == new_high_max &&
5564 zone->pageset_batch == new_batch)
5565 return;
5566
5567 zone->pageset_high_min = new_high_min;
5568 zone->pageset_high_max = new_high_max;
5569 zone->pageset_batch = new_batch;
5570
5571 __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
5572 new_batch);
5573 }
5574
setup_zone_pageset(struct zone * zone)5575 void __meminit setup_zone_pageset(struct zone *zone)
5576 {
5577 int cpu;
5578
5579 /* Size may be 0 on !SMP && !NUMA */
5580 if (sizeof(struct per_cpu_zonestat) > 0)
5581 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5582
5583 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5584 for_each_possible_cpu(cpu) {
5585 struct per_cpu_pages *pcp;
5586 struct per_cpu_zonestat *pzstats;
5587
5588 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5589 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5590 per_cpu_pages_init(pcp, pzstats);
5591 }
5592
5593 zone_set_pageset_high_and_batch(zone, 0);
5594 }
5595
5596 /*
5597 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5598 * page high values need to be recalculated.
5599 */
zone_pcp_update(struct zone * zone,int cpu_online)5600 static void zone_pcp_update(struct zone *zone, int cpu_online)
5601 {
5602 mutex_lock(&pcp_batch_high_lock);
5603 zone_set_pageset_high_and_batch(zone, cpu_online);
5604 mutex_unlock(&pcp_batch_high_lock);
5605 }
5606
zone_pcp_update_cacheinfo(struct zone * zone,unsigned int cpu)5607 static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu)
5608 {
5609 struct per_cpu_pages *pcp;
5610 struct cpu_cacheinfo *cci;
5611
5612 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5613 cci = get_cpu_cacheinfo(cpu);
5614 /*
5615 * If data cache slice of CPU is large enough, "pcp->batch"
5616 * pages can be preserved in PCP before draining PCP for
5617 * consecutive high-order pages freeing without allocation.
5618 * This can reduce zone lock contention without hurting
5619 * cache-hot pages sharing.
5620 */
5621 spin_lock(&pcp->lock);
5622 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
5623 pcp->flags |= PCPF_FREE_HIGH_BATCH;
5624 else
5625 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
5626 spin_unlock(&pcp->lock);
5627 }
5628
setup_pcp_cacheinfo(unsigned int cpu)5629 void setup_pcp_cacheinfo(unsigned int cpu)
5630 {
5631 struct zone *zone;
5632
5633 for_each_populated_zone(zone)
5634 zone_pcp_update_cacheinfo(zone, cpu);
5635 }
5636
5637 /*
5638 * Allocate per cpu pagesets and initialize them.
5639 * Before this call only boot pagesets were available.
5640 */
setup_per_cpu_pageset(void)5641 void __init setup_per_cpu_pageset(void)
5642 {
5643 struct pglist_data *pgdat;
5644 struct zone *zone;
5645 int __maybe_unused cpu;
5646
5647 for_each_populated_zone(zone)
5648 setup_zone_pageset(zone);
5649
5650 #ifdef CONFIG_NUMA
5651 /*
5652 * Unpopulated zones continue using the boot pagesets.
5653 * The numa stats for these pagesets need to be reset.
5654 * Otherwise, they will end up skewing the stats of
5655 * the nodes these zones are associated with.
5656 */
5657 for_each_possible_cpu(cpu) {
5658 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5659 memset(pzstats->vm_numa_event, 0,
5660 sizeof(pzstats->vm_numa_event));
5661 }
5662 #endif
5663
5664 for_each_online_pgdat(pgdat)
5665 pgdat->per_cpu_nodestats =
5666 alloc_percpu(struct per_cpu_nodestat);
5667 }
5668
zone_pcp_init(struct zone * zone)5669 __meminit void zone_pcp_init(struct zone *zone)
5670 {
5671 /*
5672 * per cpu subsystem is not up at this point. The following code
5673 * relies on the ability of the linker to provide the
5674 * offset of a (static) per cpu variable into the per cpu area.
5675 */
5676 zone->per_cpu_pageset = &boot_pageset;
5677 zone->per_cpu_zonestats = &boot_zonestats;
5678 zone->pageset_high_min = BOOT_PAGESET_HIGH;
5679 zone->pageset_high_max = BOOT_PAGESET_HIGH;
5680 zone->pageset_batch = BOOT_PAGESET_BATCH;
5681
5682 if (populated_zone(zone))
5683 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5684 zone->present_pages, zone_batchsize(zone));
5685 }
5686
5687 static void setup_per_zone_lowmem_reserve(void);
5688
adjust_managed_page_count(struct page * page,long count)5689 void adjust_managed_page_count(struct page *page, long count)
5690 {
5691 atomic_long_add(count, &page_zone(page)->managed_pages);
5692 totalram_pages_add(count);
5693 setup_per_zone_lowmem_reserve();
5694 }
5695 EXPORT_SYMBOL(adjust_managed_page_count);
5696
free_reserved_area(void * start,void * end,int poison,const char * s)5697 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5698 {
5699 void *pos;
5700 unsigned long pages = 0;
5701
5702 start = (void *)PAGE_ALIGN((unsigned long)start);
5703 end = (void *)((unsigned long)end & PAGE_MASK);
5704 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5705 struct page *page = virt_to_page(pos);
5706 void *direct_map_addr;
5707
5708 /*
5709 * 'direct_map_addr' might be different from 'pos'
5710 * because some architectures' virt_to_page()
5711 * work with aliases. Getting the direct map
5712 * address ensures that we get a _writeable_
5713 * alias for the memset().
5714 */
5715 direct_map_addr = page_address(page);
5716 /*
5717 * Perform a kasan-unchecked memset() since this memory
5718 * has not been initialized.
5719 */
5720 direct_map_addr = kasan_reset_tag(direct_map_addr);
5721 if ((unsigned int)poison <= 0xFF)
5722 memset(direct_map_addr, poison, PAGE_SIZE);
5723
5724 free_reserved_page(page);
5725 }
5726
5727 if (pages && s)
5728 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5729
5730 return pages;
5731 }
5732
free_reserved_page(struct page * page)5733 void free_reserved_page(struct page *page)
5734 {
5735 clear_page_tag_ref(page);
5736 ClearPageReserved(page);
5737 init_page_count(page);
5738 __free_page(page);
5739 adjust_managed_page_count(page, 1);
5740 }
5741 EXPORT_SYMBOL(free_reserved_page);
5742
page_alloc_cpu_dead(unsigned int cpu)5743 static int page_alloc_cpu_dead(unsigned int cpu)
5744 {
5745 struct zone *zone;
5746
5747 lru_add_drain_cpu(cpu);
5748 mlock_drain_remote(cpu);
5749 drain_pages(cpu);
5750
5751 /*
5752 * Spill the event counters of the dead processor
5753 * into the current processors event counters.
5754 * This artificially elevates the count of the current
5755 * processor.
5756 */
5757 vm_events_fold_cpu(cpu);
5758
5759 /*
5760 * Zero the differential counters of the dead processor
5761 * so that the vm statistics are consistent.
5762 *
5763 * This is only okay since the processor is dead and cannot
5764 * race with what we are doing.
5765 */
5766 cpu_vm_stats_fold(cpu);
5767
5768 for_each_populated_zone(zone)
5769 zone_pcp_update(zone, 0);
5770
5771 return 0;
5772 }
5773
page_alloc_cpu_online(unsigned int cpu)5774 static int page_alloc_cpu_online(unsigned int cpu)
5775 {
5776 struct zone *zone;
5777
5778 for_each_populated_zone(zone)
5779 zone_pcp_update(zone, 1);
5780 return 0;
5781 }
5782
page_alloc_init_cpuhp(void)5783 void __init page_alloc_init_cpuhp(void)
5784 {
5785 int ret;
5786
5787 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5788 "mm/page_alloc:pcp",
5789 page_alloc_cpu_online,
5790 page_alloc_cpu_dead);
5791 WARN_ON(ret < 0);
5792 }
5793
5794 /*
5795 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5796 * or min_free_kbytes changes.
5797 */
calculate_totalreserve_pages(void)5798 static void calculate_totalreserve_pages(void)
5799 {
5800 struct pglist_data *pgdat;
5801 unsigned long reserve_pages = 0;
5802 enum zone_type i, j;
5803
5804 for_each_online_pgdat(pgdat) {
5805
5806 pgdat->totalreserve_pages = 0;
5807
5808 for (i = 0; i < MAX_NR_ZONES; i++) {
5809 struct zone *zone = pgdat->node_zones + i;
5810 long max = 0;
5811 unsigned long managed_pages = zone_managed_pages(zone);
5812
5813 /* Find valid and maximum lowmem_reserve in the zone */
5814 for (j = i; j < MAX_NR_ZONES; j++) {
5815 if (zone->lowmem_reserve[j] > max)
5816 max = zone->lowmem_reserve[j];
5817 }
5818
5819 /* we treat the high watermark as reserved pages. */
5820 max += high_wmark_pages(zone);
5821
5822 if (max > managed_pages)
5823 max = managed_pages;
5824
5825 pgdat->totalreserve_pages += max;
5826
5827 reserve_pages += max;
5828 }
5829 }
5830 totalreserve_pages = reserve_pages;
5831 }
5832
5833 /*
5834 * setup_per_zone_lowmem_reserve - called whenever
5835 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5836 * has a correct pages reserved value, so an adequate number of
5837 * pages are left in the zone after a successful __alloc_pages().
5838 */
setup_per_zone_lowmem_reserve(void)5839 static void setup_per_zone_lowmem_reserve(void)
5840 {
5841 struct pglist_data *pgdat;
5842 enum zone_type i, j;
5843
5844 for_each_online_pgdat(pgdat) {
5845 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5846 struct zone *zone = &pgdat->node_zones[i];
5847 int ratio = sysctl_lowmem_reserve_ratio[i];
5848 bool clear = !ratio || !zone_managed_pages(zone);
5849 unsigned long managed_pages = 0;
5850
5851 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5852 struct zone *upper_zone = &pgdat->node_zones[j];
5853
5854 managed_pages += zone_managed_pages(upper_zone);
5855
5856 if (clear)
5857 zone->lowmem_reserve[j] = 0;
5858 else
5859 zone->lowmem_reserve[j] = managed_pages / ratio;
5860 }
5861 }
5862 }
5863
5864 /* update totalreserve_pages */
5865 calculate_totalreserve_pages();
5866 }
5867
__setup_per_zone_wmarks(void)5868 static void __setup_per_zone_wmarks(void)
5869 {
5870 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5871 unsigned long lowmem_pages = 0;
5872 struct zone *zone;
5873 unsigned long flags;
5874
5875 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5876 for_each_zone(zone) {
5877 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5878 lowmem_pages += zone_managed_pages(zone);
5879 }
5880
5881 for_each_zone(zone) {
5882 u64 tmp;
5883
5884 spin_lock_irqsave(&zone->lock, flags);
5885 tmp = (u64)pages_min * zone_managed_pages(zone);
5886 tmp = div64_ul(tmp, lowmem_pages);
5887 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5888 /*
5889 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5890 * need highmem and movable zones pages, so cap pages_min
5891 * to a small value here.
5892 *
5893 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5894 * deltas control async page reclaim, and so should
5895 * not be capped for highmem and movable zones.
5896 */
5897 unsigned long min_pages;
5898
5899 min_pages = zone_managed_pages(zone) / 1024;
5900 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5901 zone->_watermark[WMARK_MIN] = min_pages;
5902 } else {
5903 /*
5904 * If it's a lowmem zone, reserve a number of pages
5905 * proportionate to the zone's size.
5906 */
5907 zone->_watermark[WMARK_MIN] = tmp;
5908 }
5909
5910 /*
5911 * Set the kswapd watermarks distance according to the
5912 * scale factor in proportion to available memory, but
5913 * ensure a minimum size on small systems.
5914 */
5915 tmp = max_t(u64, tmp >> 2,
5916 mult_frac(zone_managed_pages(zone),
5917 watermark_scale_factor, 10000));
5918
5919 zone->watermark_boost = 0;
5920 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
5921 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5922 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5923
5924 spin_unlock_irqrestore(&zone->lock, flags);
5925 }
5926
5927 /* update totalreserve_pages */
5928 calculate_totalreserve_pages();
5929 }
5930
5931 /**
5932 * setup_per_zone_wmarks - called when min_free_kbytes changes
5933 * or when memory is hot-{added|removed}
5934 *
5935 * Ensures that the watermark[min,low,high] values for each zone are set
5936 * correctly with respect to min_free_kbytes.
5937 */
setup_per_zone_wmarks(void)5938 void setup_per_zone_wmarks(void)
5939 {
5940 struct zone *zone;
5941 static DEFINE_SPINLOCK(lock);
5942
5943 spin_lock(&lock);
5944 __setup_per_zone_wmarks();
5945 spin_unlock(&lock);
5946
5947 /*
5948 * The watermark size have changed so update the pcpu batch
5949 * and high limits or the limits may be inappropriate.
5950 */
5951 for_each_zone(zone)
5952 zone_pcp_update(zone, 0);
5953 }
5954
5955 /*
5956 * Initialise min_free_kbytes.
5957 *
5958 * For small machines we want it small (128k min). For large machines
5959 * we want it large (256MB max). But it is not linear, because network
5960 * bandwidth does not increase linearly with machine size. We use
5961 *
5962 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5963 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5964 *
5965 * which yields
5966 *
5967 * 16MB: 512k
5968 * 32MB: 724k
5969 * 64MB: 1024k
5970 * 128MB: 1448k
5971 * 256MB: 2048k
5972 * 512MB: 2896k
5973 * 1024MB: 4096k
5974 * 2048MB: 5792k
5975 * 4096MB: 8192k
5976 * 8192MB: 11584k
5977 * 16384MB: 16384k
5978 */
calculate_min_free_kbytes(void)5979 void calculate_min_free_kbytes(void)
5980 {
5981 unsigned long lowmem_kbytes;
5982 int new_min_free_kbytes;
5983
5984 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5985 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5986
5987 if (new_min_free_kbytes > user_min_free_kbytes)
5988 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
5989 else
5990 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5991 new_min_free_kbytes, user_min_free_kbytes);
5992
5993 }
5994
init_per_zone_wmark_min(void)5995 int __meminit init_per_zone_wmark_min(void)
5996 {
5997 calculate_min_free_kbytes();
5998 setup_per_zone_wmarks();
5999 refresh_zone_stat_thresholds();
6000 setup_per_zone_lowmem_reserve();
6001
6002 #ifdef CONFIG_NUMA
6003 setup_min_unmapped_ratio();
6004 setup_min_slab_ratio();
6005 #endif
6006
6007 khugepaged_min_free_kbytes_update();
6008
6009 return 0;
6010 }
postcore_initcall(init_per_zone_wmark_min)6011 postcore_initcall(init_per_zone_wmark_min)
6012
6013 /*
6014 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6015 * that we can call two helper functions whenever min_free_kbytes
6016 * changes.
6017 */
6018 static int min_free_kbytes_sysctl_handler(const struct ctl_table *table, int write,
6019 void *buffer, size_t *length, loff_t *ppos)
6020 {
6021 int rc;
6022
6023 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6024 if (rc)
6025 return rc;
6026
6027 if (write) {
6028 user_min_free_kbytes = min_free_kbytes;
6029 setup_per_zone_wmarks();
6030 }
6031 return 0;
6032 }
6033
watermark_scale_factor_sysctl_handler(const struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)6034 static int watermark_scale_factor_sysctl_handler(const struct ctl_table *table, int write,
6035 void *buffer, size_t *length, loff_t *ppos)
6036 {
6037 int rc;
6038
6039 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6040 if (rc)
6041 return rc;
6042
6043 if (write)
6044 setup_per_zone_wmarks();
6045
6046 return 0;
6047 }
6048
6049 #ifdef CONFIG_NUMA
setup_min_unmapped_ratio(void)6050 static void setup_min_unmapped_ratio(void)
6051 {
6052 pg_data_t *pgdat;
6053 struct zone *zone;
6054
6055 for_each_online_pgdat(pgdat)
6056 pgdat->min_unmapped_pages = 0;
6057
6058 for_each_zone(zone)
6059 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6060 sysctl_min_unmapped_ratio) / 100;
6061 }
6062
6063
sysctl_min_unmapped_ratio_sysctl_handler(const struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)6064 static int sysctl_min_unmapped_ratio_sysctl_handler(const struct ctl_table *table, int write,
6065 void *buffer, size_t *length, loff_t *ppos)
6066 {
6067 int rc;
6068
6069 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6070 if (rc)
6071 return rc;
6072
6073 setup_min_unmapped_ratio();
6074
6075 return 0;
6076 }
6077
setup_min_slab_ratio(void)6078 static void setup_min_slab_ratio(void)
6079 {
6080 pg_data_t *pgdat;
6081 struct zone *zone;
6082
6083 for_each_online_pgdat(pgdat)
6084 pgdat->min_slab_pages = 0;
6085
6086 for_each_zone(zone)
6087 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6088 sysctl_min_slab_ratio) / 100;
6089 }
6090
sysctl_min_slab_ratio_sysctl_handler(const struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)6091 static int sysctl_min_slab_ratio_sysctl_handler(const struct ctl_table *table, int write,
6092 void *buffer, size_t *length, loff_t *ppos)
6093 {
6094 int rc;
6095
6096 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6097 if (rc)
6098 return rc;
6099
6100 setup_min_slab_ratio();
6101
6102 return 0;
6103 }
6104 #endif
6105
6106 /*
6107 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6108 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6109 * whenever sysctl_lowmem_reserve_ratio changes.
6110 *
6111 * The reserve ratio obviously has absolutely no relation with the
6112 * minimum watermarks. The lowmem reserve ratio can only make sense
6113 * if in function of the boot time zone sizes.
6114 */
lowmem_reserve_ratio_sysctl_handler(const struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)6115 static int lowmem_reserve_ratio_sysctl_handler(const struct ctl_table *table,
6116 int write, void *buffer, size_t *length, loff_t *ppos)
6117 {
6118 int i;
6119
6120 proc_dointvec_minmax(table, write, buffer, length, ppos);
6121
6122 for (i = 0; i < MAX_NR_ZONES; i++) {
6123 if (sysctl_lowmem_reserve_ratio[i] < 1)
6124 sysctl_lowmem_reserve_ratio[i] = 0;
6125 }
6126
6127 setup_per_zone_lowmem_reserve();
6128 return 0;
6129 }
6130
6131 /*
6132 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6133 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6134 * pagelist can have before it gets flushed back to buddy allocator.
6135 */
percpu_pagelist_high_fraction_sysctl_handler(const struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)6136 static int percpu_pagelist_high_fraction_sysctl_handler(const struct ctl_table *table,
6137 int write, void *buffer, size_t *length, loff_t *ppos)
6138 {
6139 struct zone *zone;
6140 int old_percpu_pagelist_high_fraction;
6141 int ret;
6142
6143 mutex_lock(&pcp_batch_high_lock);
6144 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6145
6146 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6147 if (!write || ret < 0)
6148 goto out;
6149
6150 /* Sanity checking to avoid pcp imbalance */
6151 if (percpu_pagelist_high_fraction &&
6152 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6153 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6154 ret = -EINVAL;
6155 goto out;
6156 }
6157
6158 /* No change? */
6159 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6160 goto out;
6161
6162 for_each_populated_zone(zone)
6163 zone_set_pageset_high_and_batch(zone, 0);
6164 out:
6165 mutex_unlock(&pcp_batch_high_lock);
6166 return ret;
6167 }
6168
6169 static const struct ctl_table page_alloc_sysctl_table[] = {
6170 {
6171 .procname = "min_free_kbytes",
6172 .data = &min_free_kbytes,
6173 .maxlen = sizeof(min_free_kbytes),
6174 .mode = 0644,
6175 .proc_handler = min_free_kbytes_sysctl_handler,
6176 .extra1 = SYSCTL_ZERO,
6177 },
6178 {
6179 .procname = "watermark_boost_factor",
6180 .data = &watermark_boost_factor,
6181 .maxlen = sizeof(watermark_boost_factor),
6182 .mode = 0644,
6183 .proc_handler = proc_dointvec_minmax,
6184 .extra1 = SYSCTL_ZERO,
6185 },
6186 {
6187 .procname = "watermark_scale_factor",
6188 .data = &watermark_scale_factor,
6189 .maxlen = sizeof(watermark_scale_factor),
6190 .mode = 0644,
6191 .proc_handler = watermark_scale_factor_sysctl_handler,
6192 .extra1 = SYSCTL_ONE,
6193 .extra2 = SYSCTL_THREE_THOUSAND,
6194 },
6195 {
6196 .procname = "percpu_pagelist_high_fraction",
6197 .data = &percpu_pagelist_high_fraction,
6198 .maxlen = sizeof(percpu_pagelist_high_fraction),
6199 .mode = 0644,
6200 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6201 .extra1 = SYSCTL_ZERO,
6202 },
6203 {
6204 .procname = "lowmem_reserve_ratio",
6205 .data = &sysctl_lowmem_reserve_ratio,
6206 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6207 .mode = 0644,
6208 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6209 },
6210 #ifdef CONFIG_NUMA
6211 {
6212 .procname = "numa_zonelist_order",
6213 .data = &numa_zonelist_order,
6214 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6215 .mode = 0644,
6216 .proc_handler = numa_zonelist_order_handler,
6217 },
6218 {
6219 .procname = "min_unmapped_ratio",
6220 .data = &sysctl_min_unmapped_ratio,
6221 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6222 .mode = 0644,
6223 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6224 .extra1 = SYSCTL_ZERO,
6225 .extra2 = SYSCTL_ONE_HUNDRED,
6226 },
6227 {
6228 .procname = "min_slab_ratio",
6229 .data = &sysctl_min_slab_ratio,
6230 .maxlen = sizeof(sysctl_min_slab_ratio),
6231 .mode = 0644,
6232 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6233 .extra1 = SYSCTL_ZERO,
6234 .extra2 = SYSCTL_ONE_HUNDRED,
6235 },
6236 #endif
6237 };
6238
page_alloc_sysctl_init(void)6239 void __init page_alloc_sysctl_init(void)
6240 {
6241 register_sysctl_init("vm", page_alloc_sysctl_table);
6242 }
6243
6244 #ifdef CONFIG_CONTIG_ALLOC
6245 /* Usage: See admin-guide/dynamic-debug-howto.rst */
alloc_contig_dump_pages(struct list_head * page_list)6246 static void alloc_contig_dump_pages(struct list_head *page_list)
6247 {
6248 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6249
6250 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6251 struct page *page;
6252
6253 dump_stack();
6254 list_for_each_entry(page, page_list, lru)
6255 dump_page(page, "migration failure");
6256 }
6257 }
6258
6259 /*
6260 * [start, end) must belong to a single zone.
6261 * @migratetype: using migratetype to filter the type of migration in
6262 * trace_mm_alloc_contig_migrate_range_info.
6263 */
__alloc_contig_migrate_range(struct compact_control * cc,unsigned long start,unsigned long end,int migratetype)6264 static int __alloc_contig_migrate_range(struct compact_control *cc,
6265 unsigned long start, unsigned long end, int migratetype)
6266 {
6267 /* This function is based on compact_zone() from compaction.c. */
6268 unsigned int nr_reclaimed;
6269 unsigned long pfn = start;
6270 unsigned int tries = 0;
6271 int ret = 0;
6272 struct migration_target_control mtc = {
6273 .nid = zone_to_nid(cc->zone),
6274 .gfp_mask = cc->gfp_mask,
6275 .reason = MR_CONTIG_RANGE,
6276 };
6277 struct page *page;
6278 unsigned long total_mapped = 0;
6279 unsigned long total_migrated = 0;
6280 unsigned long total_reclaimed = 0;
6281
6282 lru_cache_disable();
6283
6284 while (pfn < end || !list_empty(&cc->migratepages)) {
6285 if (fatal_signal_pending(current)) {
6286 ret = -EINTR;
6287 break;
6288 }
6289
6290 if (list_empty(&cc->migratepages)) {
6291 cc->nr_migratepages = 0;
6292 ret = isolate_migratepages_range(cc, pfn, end);
6293 if (ret && ret != -EAGAIN)
6294 break;
6295 pfn = cc->migrate_pfn;
6296 tries = 0;
6297 } else if (++tries == 5) {
6298 ret = -EBUSY;
6299 break;
6300 }
6301
6302 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6303 &cc->migratepages);
6304 cc->nr_migratepages -= nr_reclaimed;
6305
6306 if (trace_mm_alloc_contig_migrate_range_info_enabled()) {
6307 total_reclaimed += nr_reclaimed;
6308 list_for_each_entry(page, &cc->migratepages, lru) {
6309 struct folio *folio = page_folio(page);
6310
6311 total_mapped += folio_mapped(folio) *
6312 folio_nr_pages(folio);
6313 }
6314 }
6315
6316 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6317 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6318
6319 if (trace_mm_alloc_contig_migrate_range_info_enabled() && !ret)
6320 total_migrated += cc->nr_migratepages;
6321
6322 /*
6323 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6324 * to retry again over this error, so do the same here.
6325 */
6326 if (ret == -ENOMEM)
6327 break;
6328 }
6329
6330 lru_cache_enable();
6331 if (ret < 0) {
6332 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6333 alloc_contig_dump_pages(&cc->migratepages);
6334 putback_movable_pages(&cc->migratepages);
6335 }
6336
6337 trace_mm_alloc_contig_migrate_range_info(start, end, migratetype,
6338 total_migrated,
6339 total_reclaimed,
6340 total_mapped);
6341 return (ret < 0) ? ret : 0;
6342 }
6343
split_free_pages(struct list_head * list,gfp_t gfp_mask)6344 static void split_free_pages(struct list_head *list, gfp_t gfp_mask)
6345 {
6346 int order;
6347
6348 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6349 struct page *page, *next;
6350 int nr_pages = 1 << order;
6351
6352 list_for_each_entry_safe(page, next, &list[order], lru) {
6353 int i;
6354
6355 post_alloc_hook(page, order, gfp_mask);
6356 set_page_refcounted(page);
6357 if (!order)
6358 continue;
6359
6360 split_page(page, order);
6361
6362 /* Add all subpages to the order-0 head, in sequence. */
6363 list_del(&page->lru);
6364 for (i = 0; i < nr_pages; i++)
6365 list_add_tail(&page[i].lru, &list[0]);
6366 }
6367 }
6368 }
6369
__alloc_contig_verify_gfp_mask(gfp_t gfp_mask,gfp_t * gfp_cc_mask)6370 static int __alloc_contig_verify_gfp_mask(gfp_t gfp_mask, gfp_t *gfp_cc_mask)
6371 {
6372 const gfp_t reclaim_mask = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
6373 const gfp_t action_mask = __GFP_COMP | __GFP_RETRY_MAYFAIL | __GFP_NOWARN |
6374 __GFP_ZERO | __GFP_ZEROTAGS | __GFP_SKIP_ZERO;
6375 const gfp_t cc_action_mask = __GFP_RETRY_MAYFAIL | __GFP_NOWARN;
6376
6377 /*
6378 * We are given the range to allocate; node, mobility and placement
6379 * hints are irrelevant at this point. We'll simply ignore them.
6380 */
6381 gfp_mask &= ~(GFP_ZONEMASK | __GFP_RECLAIMABLE | __GFP_WRITE |
6382 __GFP_HARDWALL | __GFP_THISNODE | __GFP_MOVABLE);
6383
6384 /*
6385 * We only support most reclaim flags (but not NOFAIL/NORETRY), and
6386 * selected action flags.
6387 */
6388 if (gfp_mask & ~(reclaim_mask | action_mask))
6389 return -EINVAL;
6390
6391 /*
6392 * Flags to control page compaction/migration/reclaim, to free up our
6393 * page range. Migratable pages are movable, __GFP_MOVABLE is implied
6394 * for them.
6395 *
6396 * Traditionally we always had __GFP_RETRY_MAYFAIL set, keep doing that
6397 * to not degrade callers.
6398 */
6399 *gfp_cc_mask = (gfp_mask & (reclaim_mask | cc_action_mask)) |
6400 __GFP_MOVABLE | __GFP_RETRY_MAYFAIL;
6401 return 0;
6402 }
6403
6404 /**
6405 * alloc_contig_range() -- tries to allocate given range of pages
6406 * @start: start PFN to allocate
6407 * @end: one-past-the-last PFN to allocate
6408 * @migratetype: migratetype of the underlying pageblocks (either
6409 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6410 * in range must have the same migratetype and it must
6411 * be either of the two.
6412 * @gfp_mask: GFP mask. Node/zone/placement hints are ignored; only some
6413 * action and reclaim modifiers are supported. Reclaim modifiers
6414 * control allocation behavior during compaction/migration/reclaim.
6415 *
6416 * The PFN range does not have to be pageblock aligned. The PFN range must
6417 * belong to a single zone.
6418 *
6419 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6420 * pageblocks in the range. Once isolated, the pageblocks should not
6421 * be modified by others.
6422 *
6423 * Return: zero on success or negative error code. On success all
6424 * pages which PFN is in [start, end) are allocated for the caller and
6425 * need to be freed with free_contig_range().
6426 */
alloc_contig_range_noprof(unsigned long start,unsigned long end,unsigned migratetype,gfp_t gfp_mask)6427 int alloc_contig_range_noprof(unsigned long start, unsigned long end,
6428 unsigned migratetype, gfp_t gfp_mask)
6429 {
6430 unsigned long outer_start, outer_end;
6431 int ret = 0;
6432
6433 struct compact_control cc = {
6434 .nr_migratepages = 0,
6435 .order = -1,
6436 .zone = page_zone(pfn_to_page(start)),
6437 .mode = MIGRATE_SYNC,
6438 .ignore_skip_hint = true,
6439 .no_set_skip_hint = true,
6440 .alloc_contig = true,
6441 };
6442 INIT_LIST_HEAD(&cc.migratepages);
6443
6444 gfp_mask = current_gfp_context(gfp_mask);
6445 if (__alloc_contig_verify_gfp_mask(gfp_mask, (gfp_t *)&cc.gfp_mask))
6446 return -EINVAL;
6447
6448 /*
6449 * What we do here is we mark all pageblocks in range as
6450 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6451 * have different sizes, and due to the way page allocator
6452 * work, start_isolate_page_range() has special handlings for this.
6453 *
6454 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6455 * migrate the pages from an unaligned range (ie. pages that
6456 * we are interested in). This will put all the pages in
6457 * range back to page allocator as MIGRATE_ISOLATE.
6458 *
6459 * When this is done, we take the pages in range from page
6460 * allocator removing them from the buddy system. This way
6461 * page allocator will never consider using them.
6462 *
6463 * This lets us mark the pageblocks back as
6464 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6465 * aligned range but not in the unaligned, original range are
6466 * put back to page allocator so that buddy can use them.
6467 */
6468
6469 ret = start_isolate_page_range(start, end, migratetype, 0);
6470 if (ret)
6471 goto done;
6472
6473 drain_all_pages(cc.zone);
6474
6475 /*
6476 * In case of -EBUSY, we'd like to know which page causes problem.
6477 * So, just fall through. test_pages_isolated() has a tracepoint
6478 * which will report the busy page.
6479 *
6480 * It is possible that busy pages could become available before
6481 * the call to test_pages_isolated, and the range will actually be
6482 * allocated. So, if we fall through be sure to clear ret so that
6483 * -EBUSY is not accidentally used or returned to caller.
6484 */
6485 ret = __alloc_contig_migrate_range(&cc, start, end, migratetype);
6486 if (ret && ret != -EBUSY)
6487 goto done;
6488
6489 /*
6490 * When in-use hugetlb pages are migrated, they may simply be released
6491 * back into the free hugepage pool instead of being returned to the
6492 * buddy system. After the migration of in-use huge pages is completed,
6493 * we will invoke replace_free_hugepage_folios() to ensure that these
6494 * hugepages are properly released to the buddy system.
6495 */
6496 ret = replace_free_hugepage_folios(start, end);
6497 if (ret)
6498 goto done;
6499
6500 /*
6501 * Pages from [start, end) are within a pageblock_nr_pages
6502 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6503 * more, all pages in [start, end) are free in page allocator.
6504 * What we are going to do is to allocate all pages from
6505 * [start, end) (that is remove them from page allocator).
6506 *
6507 * The only problem is that pages at the beginning and at the
6508 * end of interesting range may be not aligned with pages that
6509 * page allocator holds, ie. they can be part of higher order
6510 * pages. Because of this, we reserve the bigger range and
6511 * once this is done free the pages we are not interested in.
6512 *
6513 * We don't have to hold zone->lock here because the pages are
6514 * isolated thus they won't get removed from buddy.
6515 */
6516 outer_start = find_large_buddy(start);
6517
6518 /* Make sure the range is really isolated. */
6519 if (test_pages_isolated(outer_start, end, 0)) {
6520 ret = -EBUSY;
6521 goto done;
6522 }
6523
6524 /* Grab isolated pages from freelists. */
6525 outer_end = isolate_freepages_range(&cc, outer_start, end);
6526 if (!outer_end) {
6527 ret = -EBUSY;
6528 goto done;
6529 }
6530
6531 if (!(gfp_mask & __GFP_COMP)) {
6532 split_free_pages(cc.freepages, gfp_mask);
6533
6534 /* Free head and tail (if any) */
6535 if (start != outer_start)
6536 free_contig_range(outer_start, start - outer_start);
6537 if (end != outer_end)
6538 free_contig_range(end, outer_end - end);
6539 } else if (start == outer_start && end == outer_end && is_power_of_2(end - start)) {
6540 struct page *head = pfn_to_page(start);
6541 int order = ilog2(end - start);
6542
6543 check_new_pages(head, order);
6544 prep_new_page(head, order, gfp_mask, 0);
6545 set_page_refcounted(head);
6546 } else {
6547 ret = -EINVAL;
6548 WARN(true, "PFN range: requested [%lu, %lu), allocated [%lu, %lu)\n",
6549 start, end, outer_start, outer_end);
6550 }
6551 done:
6552 undo_isolate_page_range(start, end, migratetype);
6553 return ret;
6554 }
6555 EXPORT_SYMBOL(alloc_contig_range_noprof);
6556
__alloc_contig_pages(unsigned long start_pfn,unsigned long nr_pages,gfp_t gfp_mask)6557 static int __alloc_contig_pages(unsigned long start_pfn,
6558 unsigned long nr_pages, gfp_t gfp_mask)
6559 {
6560 unsigned long end_pfn = start_pfn + nr_pages;
6561
6562 return alloc_contig_range_noprof(start_pfn, end_pfn, MIGRATE_MOVABLE,
6563 gfp_mask);
6564 }
6565
pfn_range_valid_contig(struct zone * z,unsigned long start_pfn,unsigned long nr_pages)6566 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6567 unsigned long nr_pages)
6568 {
6569 unsigned long i, end_pfn = start_pfn + nr_pages;
6570 struct page *page;
6571
6572 for (i = start_pfn; i < end_pfn; i++) {
6573 page = pfn_to_online_page(i);
6574 if (!page)
6575 return false;
6576
6577 if (page_zone(page) != z)
6578 return false;
6579
6580 if (PageReserved(page))
6581 return false;
6582
6583 if (PageHuge(page))
6584 return false;
6585 }
6586 return true;
6587 }
6588
zone_spans_last_pfn(const struct zone * zone,unsigned long start_pfn,unsigned long nr_pages)6589 static bool zone_spans_last_pfn(const struct zone *zone,
6590 unsigned long start_pfn, unsigned long nr_pages)
6591 {
6592 unsigned long last_pfn = start_pfn + nr_pages - 1;
6593
6594 return zone_spans_pfn(zone, last_pfn);
6595 }
6596
6597 /**
6598 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6599 * @nr_pages: Number of contiguous pages to allocate
6600 * @gfp_mask: GFP mask. Node/zone/placement hints limit the search; only some
6601 * action and reclaim modifiers are supported. Reclaim modifiers
6602 * control allocation behavior during compaction/migration/reclaim.
6603 * @nid: Target node
6604 * @nodemask: Mask for other possible nodes
6605 *
6606 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6607 * on an applicable zonelist to find a contiguous pfn range which can then be
6608 * tried for allocation with alloc_contig_range(). This routine is intended
6609 * for allocation requests which can not be fulfilled with the buddy allocator.
6610 *
6611 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6612 * power of two, then allocated range is also guaranteed to be aligned to same
6613 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6614 *
6615 * Allocated pages can be freed with free_contig_range() or by manually calling
6616 * __free_page() on each allocated page.
6617 *
6618 * Return: pointer to contiguous pages on success, or NULL if not successful.
6619 */
alloc_contig_pages_noprof(unsigned long nr_pages,gfp_t gfp_mask,int nid,nodemask_t * nodemask)6620 struct page *alloc_contig_pages_noprof(unsigned long nr_pages, gfp_t gfp_mask,
6621 int nid, nodemask_t *nodemask)
6622 {
6623 unsigned long ret, pfn, flags;
6624 struct zonelist *zonelist;
6625 struct zone *zone;
6626 struct zoneref *z;
6627
6628 zonelist = node_zonelist(nid, gfp_mask);
6629 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6630 gfp_zone(gfp_mask), nodemask) {
6631 spin_lock_irqsave(&zone->lock, flags);
6632
6633 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6634 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6635 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6636 /*
6637 * We release the zone lock here because
6638 * alloc_contig_range() will also lock the zone
6639 * at some point. If there's an allocation
6640 * spinning on this lock, it may win the race
6641 * and cause alloc_contig_range() to fail...
6642 */
6643 spin_unlock_irqrestore(&zone->lock, flags);
6644 ret = __alloc_contig_pages(pfn, nr_pages,
6645 gfp_mask);
6646 if (!ret)
6647 return pfn_to_page(pfn);
6648 spin_lock_irqsave(&zone->lock, flags);
6649 }
6650 pfn += nr_pages;
6651 }
6652 spin_unlock_irqrestore(&zone->lock, flags);
6653 }
6654 return NULL;
6655 }
6656 #endif /* CONFIG_CONTIG_ALLOC */
6657
free_contig_range(unsigned long pfn,unsigned long nr_pages)6658 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6659 {
6660 unsigned long count = 0;
6661 struct folio *folio = pfn_folio(pfn);
6662
6663 if (folio_test_large(folio)) {
6664 int expected = folio_nr_pages(folio);
6665
6666 if (nr_pages == expected)
6667 folio_put(folio);
6668 else
6669 WARN(true, "PFN %lu: nr_pages %lu != expected %d\n",
6670 pfn, nr_pages, expected);
6671 return;
6672 }
6673
6674 for (; nr_pages--; pfn++) {
6675 struct page *page = pfn_to_page(pfn);
6676
6677 count += page_count(page) != 1;
6678 __free_page(page);
6679 }
6680 WARN(count != 0, "%lu pages are still in use!\n", count);
6681 }
6682 EXPORT_SYMBOL(free_contig_range);
6683
6684 /*
6685 * Effectively disable pcplists for the zone by setting the high limit to 0
6686 * and draining all cpus. A concurrent page freeing on another CPU that's about
6687 * to put the page on pcplist will either finish before the drain and the page
6688 * will be drained, or observe the new high limit and skip the pcplist.
6689 *
6690 * Must be paired with a call to zone_pcp_enable().
6691 */
zone_pcp_disable(struct zone * zone)6692 void zone_pcp_disable(struct zone *zone)
6693 {
6694 mutex_lock(&pcp_batch_high_lock);
6695 __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
6696 __drain_all_pages(zone, true);
6697 }
6698
zone_pcp_enable(struct zone * zone)6699 void zone_pcp_enable(struct zone *zone)
6700 {
6701 __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
6702 zone->pageset_high_max, zone->pageset_batch);
6703 mutex_unlock(&pcp_batch_high_lock);
6704 }
6705
zone_pcp_reset(struct zone * zone)6706 void zone_pcp_reset(struct zone *zone)
6707 {
6708 int cpu;
6709 struct per_cpu_zonestat *pzstats;
6710
6711 if (zone->per_cpu_pageset != &boot_pageset) {
6712 for_each_online_cpu(cpu) {
6713 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6714 drain_zonestat(zone, pzstats);
6715 }
6716 free_percpu(zone->per_cpu_pageset);
6717 zone->per_cpu_pageset = &boot_pageset;
6718 if (zone->per_cpu_zonestats != &boot_zonestats) {
6719 free_percpu(zone->per_cpu_zonestats);
6720 zone->per_cpu_zonestats = &boot_zonestats;
6721 }
6722 }
6723 }
6724
6725 #ifdef CONFIG_MEMORY_HOTREMOVE
6726 /*
6727 * All pages in the range must be in a single zone, must not contain holes,
6728 * must span full sections, and must be isolated before calling this function.
6729 *
6730 * Returns the number of managed (non-PageOffline()) pages in the range: the
6731 * number of pages for which memory offlining code must adjust managed page
6732 * counters using adjust_managed_page_count().
6733 */
__offline_isolated_pages(unsigned long start_pfn,unsigned long end_pfn)6734 unsigned long __offline_isolated_pages(unsigned long start_pfn,
6735 unsigned long end_pfn)
6736 {
6737 unsigned long already_offline = 0, flags;
6738 unsigned long pfn = start_pfn;
6739 struct page *page;
6740 struct zone *zone;
6741 unsigned int order;
6742
6743 offline_mem_sections(pfn, end_pfn);
6744 zone = page_zone(pfn_to_page(pfn));
6745 spin_lock_irqsave(&zone->lock, flags);
6746 while (pfn < end_pfn) {
6747 page = pfn_to_page(pfn);
6748 /*
6749 * The HWPoisoned page may be not in buddy system, and
6750 * page_count() is not 0.
6751 */
6752 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6753 pfn++;
6754 continue;
6755 }
6756 /*
6757 * At this point all remaining PageOffline() pages have a
6758 * reference count of 0 and can simply be skipped.
6759 */
6760 if (PageOffline(page)) {
6761 BUG_ON(page_count(page));
6762 BUG_ON(PageBuddy(page));
6763 already_offline++;
6764 pfn++;
6765 continue;
6766 }
6767
6768 BUG_ON(page_count(page));
6769 BUG_ON(!PageBuddy(page));
6770 VM_WARN_ON(get_pageblock_migratetype(page) != MIGRATE_ISOLATE);
6771 order = buddy_order(page);
6772 del_page_from_free_list(page, zone, order, MIGRATE_ISOLATE);
6773 pfn += (1 << order);
6774 }
6775 spin_unlock_irqrestore(&zone->lock, flags);
6776
6777 return end_pfn - start_pfn - already_offline;
6778 }
6779 #endif
6780
6781 /*
6782 * This function returns a stable result only if called under zone lock.
6783 */
is_free_buddy_page(const struct page * page)6784 bool is_free_buddy_page(const struct page *page)
6785 {
6786 unsigned long pfn = page_to_pfn(page);
6787 unsigned int order;
6788
6789 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6790 const struct page *head = page - (pfn & ((1 << order) - 1));
6791
6792 if (PageBuddy(head) &&
6793 buddy_order_unsafe(head) >= order)
6794 break;
6795 }
6796
6797 return order <= MAX_PAGE_ORDER;
6798 }
6799 EXPORT_SYMBOL(is_free_buddy_page);
6800
6801 #ifdef CONFIG_MEMORY_FAILURE
add_to_free_list(struct page * page,struct zone * zone,unsigned int order,int migratetype,bool tail)6802 static inline void add_to_free_list(struct page *page, struct zone *zone,
6803 unsigned int order, int migratetype,
6804 bool tail)
6805 {
6806 __add_to_free_list(page, zone, order, migratetype, tail);
6807 account_freepages(zone, 1 << order, migratetype);
6808 }
6809
6810 /*
6811 * Break down a higher-order page in sub-pages, and keep our target out of
6812 * buddy allocator.
6813 */
break_down_buddy_pages(struct zone * zone,struct page * page,struct page * target,int low,int high,int migratetype)6814 static void break_down_buddy_pages(struct zone *zone, struct page *page,
6815 struct page *target, int low, int high,
6816 int migratetype)
6817 {
6818 unsigned long size = 1 << high;
6819 struct page *current_buddy;
6820
6821 while (high > low) {
6822 high--;
6823 size >>= 1;
6824
6825 if (target >= &page[size]) {
6826 current_buddy = page;
6827 page = page + size;
6828 } else {
6829 current_buddy = page + size;
6830 }
6831
6832 if (set_page_guard(zone, current_buddy, high))
6833 continue;
6834
6835 add_to_free_list(current_buddy, zone, high, migratetype, false);
6836 set_buddy_order(current_buddy, high);
6837 }
6838 }
6839
6840 /*
6841 * Take a page that will be marked as poisoned off the buddy allocator.
6842 */
take_page_off_buddy(struct page * page)6843 bool take_page_off_buddy(struct page *page)
6844 {
6845 struct zone *zone = page_zone(page);
6846 unsigned long pfn = page_to_pfn(page);
6847 unsigned long flags;
6848 unsigned int order;
6849 bool ret = false;
6850
6851 spin_lock_irqsave(&zone->lock, flags);
6852 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6853 struct page *page_head = page - (pfn & ((1 << order) - 1));
6854 int page_order = buddy_order(page_head);
6855
6856 if (PageBuddy(page_head) && page_order >= order) {
6857 unsigned long pfn_head = page_to_pfn(page_head);
6858 int migratetype = get_pfnblock_migratetype(page_head,
6859 pfn_head);
6860
6861 del_page_from_free_list(page_head, zone, page_order,
6862 migratetype);
6863 break_down_buddy_pages(zone, page_head, page, 0,
6864 page_order, migratetype);
6865 SetPageHWPoisonTakenOff(page);
6866 ret = true;
6867 break;
6868 }
6869 if (page_count(page_head) > 0)
6870 break;
6871 }
6872 spin_unlock_irqrestore(&zone->lock, flags);
6873 return ret;
6874 }
6875
6876 /*
6877 * Cancel takeoff done by take_page_off_buddy().
6878 */
put_page_back_buddy(struct page * page)6879 bool put_page_back_buddy(struct page *page)
6880 {
6881 struct zone *zone = page_zone(page);
6882 unsigned long flags;
6883 bool ret = false;
6884
6885 spin_lock_irqsave(&zone->lock, flags);
6886 if (put_page_testzero(page)) {
6887 unsigned long pfn = page_to_pfn(page);
6888 int migratetype = get_pfnblock_migratetype(page, pfn);
6889
6890 ClearPageHWPoisonTakenOff(page);
6891 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6892 if (TestClearPageHWPoison(page)) {
6893 ret = true;
6894 }
6895 }
6896 spin_unlock_irqrestore(&zone->lock, flags);
6897
6898 return ret;
6899 }
6900 #endif
6901
6902 #ifdef CONFIG_ZONE_DMA
has_managed_dma(void)6903 bool has_managed_dma(void)
6904 {
6905 struct pglist_data *pgdat;
6906
6907 for_each_online_pgdat(pgdat) {
6908 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6909
6910 if (managed_zone(zone))
6911 return true;
6912 }
6913 return false;
6914 }
6915 #endif /* CONFIG_ZONE_DMA */
6916
6917 #ifdef CONFIG_UNACCEPTED_MEMORY
6918
6919 /* Counts number of zones with unaccepted pages. */
6920 static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6921
6922 static bool lazy_accept = true;
6923
accept_memory_parse(char * p)6924 static int __init accept_memory_parse(char *p)
6925 {
6926 if (!strcmp(p, "lazy")) {
6927 lazy_accept = true;
6928 return 0;
6929 } else if (!strcmp(p, "eager")) {
6930 lazy_accept = false;
6931 return 0;
6932 } else {
6933 return -EINVAL;
6934 }
6935 }
6936 early_param("accept_memory", accept_memory_parse);
6937
page_contains_unaccepted(struct page * page,unsigned int order)6938 static bool page_contains_unaccepted(struct page *page, unsigned int order)
6939 {
6940 phys_addr_t start = page_to_phys(page);
6941
6942 return range_contains_unaccepted_memory(start, PAGE_SIZE << order);
6943 }
6944
__accept_page(struct zone * zone,unsigned long * flags,struct page * page)6945 static void __accept_page(struct zone *zone, unsigned long *flags,
6946 struct page *page)
6947 {
6948 bool last;
6949
6950 list_del(&page->lru);
6951 last = list_empty(&zone->unaccepted_pages);
6952
6953 account_freepages(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6954 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6955 __ClearPageUnaccepted(page);
6956 spin_unlock_irqrestore(&zone->lock, *flags);
6957
6958 accept_memory(page_to_phys(page), PAGE_SIZE << MAX_PAGE_ORDER);
6959
6960 __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
6961
6962 if (last)
6963 static_branch_dec(&zones_with_unaccepted_pages);
6964 }
6965
accept_page(struct page * page)6966 void accept_page(struct page *page)
6967 {
6968 struct zone *zone = page_zone(page);
6969 unsigned long flags;
6970
6971 spin_lock_irqsave(&zone->lock, flags);
6972 if (!PageUnaccepted(page)) {
6973 spin_unlock_irqrestore(&zone->lock, flags);
6974 return;
6975 }
6976
6977 /* Unlocks zone->lock */
6978 __accept_page(zone, &flags, page);
6979 }
6980
try_to_accept_memory_one(struct zone * zone)6981 static bool try_to_accept_memory_one(struct zone *zone)
6982 {
6983 unsigned long flags;
6984 struct page *page;
6985
6986 spin_lock_irqsave(&zone->lock, flags);
6987 page = list_first_entry_or_null(&zone->unaccepted_pages,
6988 struct page, lru);
6989 if (!page) {
6990 spin_unlock_irqrestore(&zone->lock, flags);
6991 return false;
6992 }
6993
6994 /* Unlocks zone->lock */
6995 __accept_page(zone, &flags, page);
6996
6997 return true;
6998 }
6999
has_unaccepted_memory(void)7000 static inline bool has_unaccepted_memory(void)
7001 {
7002 return static_branch_unlikely(&zones_with_unaccepted_pages);
7003 }
7004
cond_accept_memory(struct zone * zone,unsigned int order)7005 static bool cond_accept_memory(struct zone *zone, unsigned int order)
7006 {
7007 long to_accept;
7008 bool ret = false;
7009
7010 if (!has_unaccepted_memory())
7011 return false;
7012
7013 if (list_empty(&zone->unaccepted_pages))
7014 return false;
7015
7016 /* How much to accept to get to promo watermark? */
7017 to_accept = promo_wmark_pages(zone) -
7018 (zone_page_state(zone, NR_FREE_PAGES) -
7019 __zone_watermark_unusable_free(zone, order, 0) -
7020 zone_page_state(zone, NR_UNACCEPTED));
7021
7022 while (to_accept > 0) {
7023 if (!try_to_accept_memory_one(zone))
7024 break;
7025 ret = true;
7026 to_accept -= MAX_ORDER_NR_PAGES;
7027 }
7028
7029 return ret;
7030 }
7031
__free_unaccepted(struct page * page)7032 static bool __free_unaccepted(struct page *page)
7033 {
7034 struct zone *zone = page_zone(page);
7035 unsigned long flags;
7036 bool first = false;
7037
7038 if (!lazy_accept)
7039 return false;
7040
7041 spin_lock_irqsave(&zone->lock, flags);
7042 first = list_empty(&zone->unaccepted_pages);
7043 list_add_tail(&page->lru, &zone->unaccepted_pages);
7044 account_freepages(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
7045 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
7046 __SetPageUnaccepted(page);
7047 spin_unlock_irqrestore(&zone->lock, flags);
7048
7049 if (first)
7050 static_branch_inc(&zones_with_unaccepted_pages);
7051
7052 return true;
7053 }
7054
7055 #else
7056
page_contains_unaccepted(struct page * page,unsigned int order)7057 static bool page_contains_unaccepted(struct page *page, unsigned int order)
7058 {
7059 return false;
7060 }
7061
cond_accept_memory(struct zone * zone,unsigned int order)7062 static bool cond_accept_memory(struct zone *zone, unsigned int order)
7063 {
7064 return false;
7065 }
7066
__free_unaccepted(struct page * page)7067 static bool __free_unaccepted(struct page *page)
7068 {
7069 BUILD_BUG();
7070 return false;
7071 }
7072
7073 #endif /* CONFIG_UNACCEPTED_MEMORY */
7074