]>
Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
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 | ||
1da177e4 LT |
18 | #include <linux/stddef.h> |
19 | #include <linux/mm.h> | |
ca79b0c2 | 20 | #include <linux/highmem.h> |
1da177e4 | 21 | #include <linux/interrupt.h> |
10ed273f | 22 | #include <linux/jiffies.h> |
1da177e4 | 23 | #include <linux/compiler.h> |
9f158333 | 24 | #include <linux/kernel.h> |
b8c73fc2 | 25 | #include <linux/kasan.h> |
b073d7f8 | 26 | #include <linux/kmsan.h> |
1da177e4 LT |
27 | #include <linux/module.h> |
28 | #include <linux/suspend.h> | |
a238ab5b | 29 | #include <linux/ratelimit.h> |
5a3135c2 | 30 | #include <linux/oom.h> |
1da177e4 LT |
31 | #include <linux/topology.h> |
32 | #include <linux/sysctl.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/cpuset.h> | |
90491d87 | 35 | #include <linux/pagevec.h> |
bdc8cb98 | 36 | #include <linux/memory_hotplug.h> |
1da177e4 | 37 | #include <linux/nodemask.h> |
a6cccdc3 | 38 | #include <linux/vmstat.h> |
933e312e | 39 | #include <linux/fault-inject.h> |
56de7263 | 40 | #include <linux/compaction.h> |
0d3d062a | 41 | #include <trace/events/kmem.h> |
d379f01d | 42 | #include <trace/events/oom.h> |
268bb0ce | 43 | #include <linux/prefetch.h> |
6e543d57 | 44 | #include <linux/mm_inline.h> |
f920e413 | 45 | #include <linux/mmu_notifier.h> |
041d3a8c | 46 | #include <linux/migrate.h> |
5b3cc15a | 47 | #include <linux/sched/mm.h> |
48c96a36 | 48 | #include <linux/page_owner.h> |
df4e817b | 49 | #include <linux/page_table_check.h> |
4949148a | 50 | #include <linux/memcontrol.h> |
42c269c8 | 51 | #include <linux/ftrace.h> |
d92a8cfc | 52 | #include <linux/lockdep.h> |
eb414681 | 53 | #include <linux/psi.h> |
4aab2be0 | 54 | #include <linux/khugepaged.h> |
5bf18281 | 55 | #include <linux/delayacct.h> |
362d37a1 | 56 | #include <linux/cacheinfo.h> |
ac924c60 | 57 | #include <asm/div64.h> |
1da177e4 | 58 | #include "internal.h" |
e900a918 | 59 | #include "shuffle.h" |
36e66c55 | 60 | #include "page_reporting.h" |
1da177e4 | 61 | |
f04a5d5d DH |
62 | /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */ |
63 | typedef int __bitwise fpi_t; | |
64 | ||
65 | /* No special request */ | |
66 | #define FPI_NONE ((__force fpi_t)0) | |
67 | ||
68 | /* | |
69 | * Skip free page reporting notification for the (possibly merged) page. | |
70 | * This does not hinder free page reporting from grabbing the page, | |
71 | * reporting it and marking it "reported" - it only skips notifying | |
72 | * the free page reporting infrastructure about a newly freed page. For | |
73 | * example, used when temporarily pulling a page from a freelist and | |
74 | * putting it back unmodified. | |
75 | */ | |
76 | #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0)) | |
77 | ||
47b6a24a DH |
78 | /* |
79 | * Place the (possibly merged) page to the tail of the freelist. Will ignore | |
80 | * page shuffling (relevant code - e.g., memory onlining - is expected to | |
81 | * shuffle the whole zone). | |
82 | * | |
83 | * Note: No code should rely on this flag for correctness - it's purely | |
84 | * to allow for optimizations when handing back either fresh pages | |
85 | * (memory onlining) or untouched pages (page isolation, free page | |
86 | * reporting). | |
87 | */ | |
88 | #define FPI_TO_TAIL ((__force fpi_t)BIT(1)) | |
89 | ||
c8e251fa CS |
90 | /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */ |
91 | static DEFINE_MUTEX(pcp_batch_high_lock); | |
74f44822 | 92 | #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8) |
c8e251fa | 93 | |
4b23a68f MG |
94 | #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT) |
95 | /* | |
96 | * On SMP, spin_trylock is sufficient protection. | |
97 | * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP. | |
98 | */ | |
99 | #define pcp_trylock_prepare(flags) do { } while (0) | |
100 | #define pcp_trylock_finish(flag) do { } while (0) | |
101 | #else | |
102 | ||
103 | /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */ | |
104 | #define pcp_trylock_prepare(flags) local_irq_save(flags) | |
105 | #define pcp_trylock_finish(flags) local_irq_restore(flags) | |
106 | #endif | |
107 | ||
01b44456 MG |
108 | /* |
109 | * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid | |
110 | * a migration causing the wrong PCP to be locked and remote memory being | |
111 | * potentially allocated, pin the task to the CPU for the lookup+lock. | |
112 | * preempt_disable is used on !RT because it is faster than migrate_disable. | |
113 | * migrate_disable is used on RT because otherwise RT spinlock usage is | |
114 | * interfered with and a high priority task cannot preempt the allocator. | |
115 | */ | |
116 | #ifndef CONFIG_PREEMPT_RT | |
117 | #define pcpu_task_pin() preempt_disable() | |
118 | #define pcpu_task_unpin() preempt_enable() | |
119 | #else | |
120 | #define pcpu_task_pin() migrate_disable() | |
121 | #define pcpu_task_unpin() migrate_enable() | |
122 | #endif | |
c8e251fa | 123 | |
01b44456 MG |
124 | /* |
125 | * Generic helper to lookup and a per-cpu variable with an embedded spinlock. | |
126 | * Return value should be used with equivalent unlock helper. | |
127 | */ | |
128 | #define pcpu_spin_lock(type, member, ptr) \ | |
129 | ({ \ | |
130 | type *_ret; \ | |
131 | pcpu_task_pin(); \ | |
132 | _ret = this_cpu_ptr(ptr); \ | |
133 | spin_lock(&_ret->member); \ | |
134 | _ret; \ | |
135 | }) | |
136 | ||
57490774 | 137 | #define pcpu_spin_trylock(type, member, ptr) \ |
01b44456 MG |
138 | ({ \ |
139 | type *_ret; \ | |
140 | pcpu_task_pin(); \ | |
141 | _ret = this_cpu_ptr(ptr); \ | |
57490774 | 142 | if (!spin_trylock(&_ret->member)) { \ |
01b44456 MG |
143 | pcpu_task_unpin(); \ |
144 | _ret = NULL; \ | |
145 | } \ | |
146 | _ret; \ | |
147 | }) | |
148 | ||
149 | #define pcpu_spin_unlock(member, ptr) \ | |
150 | ({ \ | |
151 | spin_unlock(&ptr->member); \ | |
152 | pcpu_task_unpin(); \ | |
153 | }) | |
154 | ||
01b44456 MG |
155 | /* struct per_cpu_pages specific helpers. */ |
156 | #define pcp_spin_lock(ptr) \ | |
157 | pcpu_spin_lock(struct per_cpu_pages, lock, ptr) | |
158 | ||
57490774 MG |
159 | #define pcp_spin_trylock(ptr) \ |
160 | pcpu_spin_trylock(struct per_cpu_pages, lock, ptr) | |
01b44456 MG |
161 | |
162 | #define pcp_spin_unlock(ptr) \ | |
163 | pcpu_spin_unlock(lock, ptr) | |
164 | ||
72812019 LS |
165 | #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID |
166 | DEFINE_PER_CPU(int, numa_node); | |
167 | EXPORT_PER_CPU_SYMBOL(numa_node); | |
168 | #endif | |
169 | ||
4518085e KW |
170 | DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key); |
171 | ||
7aac7898 LS |
172 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES |
173 | /* | |
174 | * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. | |
175 | * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. | |
176 | * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() | |
177 | * defined in <linux/topology.h>. | |
178 | */ | |
179 | DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ | |
180 | EXPORT_PER_CPU_SYMBOL(_numa_mem_); | |
181 | #endif | |
182 | ||
8b885f53 | 183 | static DEFINE_MUTEX(pcpu_drain_mutex); |
bd233f53 | 184 | |
38addce8 | 185 | #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY |
58bea414 | 186 | volatile unsigned long latent_entropy __latent_entropy; |
38addce8 ER |
187 | EXPORT_SYMBOL(latent_entropy); |
188 | #endif | |
189 | ||
1da177e4 | 190 | /* |
13808910 | 191 | * Array of node states. |
1da177e4 | 192 | */ |
13808910 CL |
193 | nodemask_t node_states[NR_NODE_STATES] __read_mostly = { |
194 | [N_POSSIBLE] = NODE_MASK_ALL, | |
195 | [N_ONLINE] = { { [0] = 1UL } }, | |
196 | #ifndef CONFIG_NUMA | |
197 | [N_NORMAL_MEMORY] = { { [0] = 1UL } }, | |
198 | #ifdef CONFIG_HIGHMEM | |
199 | [N_HIGH_MEMORY] = { { [0] = 1UL } }, | |
20b2f52b | 200 | #endif |
20b2f52b | 201 | [N_MEMORY] = { { [0] = 1UL } }, |
13808910 CL |
202 | [N_CPU] = { { [0] = 1UL } }, |
203 | #endif /* NUMA */ | |
204 | }; | |
205 | EXPORT_SYMBOL(node_states); | |
206 | ||
dcce284a | 207 | gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; |
6471384a | 208 | |
bb14c2c7 VB |
209 | /* |
210 | * A cached value of the page's pageblock's migratetype, used when the page is | |
211 | * put on a pcplist. Used to avoid the pageblock migratetype lookup when | |
212 | * freeing from pcplists in most cases, at the cost of possibly becoming stale. | |
213 | * Also the migratetype set in the page does not necessarily match the pcplist | |
214 | * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any | |
215 | * other index - this ensures that it will be put on the correct CMA freelist. | |
216 | */ | |
217 | static inline int get_pcppage_migratetype(struct page *page) | |
218 | { | |
219 | return page->index; | |
220 | } | |
221 | ||
222 | static inline void set_pcppage_migratetype(struct page *page, int migratetype) | |
223 | { | |
224 | page->index = migratetype; | |
225 | } | |
226 | ||
d9c23400 | 227 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE |
d00181b9 | 228 | unsigned int pageblock_order __read_mostly; |
d9c23400 MG |
229 | #endif |
230 | ||
7fef431b DH |
231 | static void __free_pages_ok(struct page *page, unsigned int order, |
232 | fpi_t fpi_flags); | |
a226f6c8 | 233 | |
1da177e4 LT |
234 | /* |
235 | * results with 256, 32 in the lowmem_reserve sysctl: | |
236 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) | |
237 | * 1G machine -> (16M dma, 784M normal, 224M high) | |
238 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA | |
239 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL | |
84109e15 | 240 | * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA |
a2f1b424 AK |
241 | * |
242 | * TBD: should special case ZONE_DMA32 machines here - in those we normally | |
243 | * don't need any ZONE_NORMAL reservation | |
1da177e4 | 244 | */ |
62069aac | 245 | static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = { |
4b51d669 | 246 | #ifdef CONFIG_ZONE_DMA |
d3cda233 | 247 | [ZONE_DMA] = 256, |
4b51d669 | 248 | #endif |
fb0e7942 | 249 | #ifdef CONFIG_ZONE_DMA32 |
d3cda233 | 250 | [ZONE_DMA32] = 256, |
fb0e7942 | 251 | #endif |
d3cda233 | 252 | [ZONE_NORMAL] = 32, |
e53ef38d | 253 | #ifdef CONFIG_HIGHMEM |
d3cda233 | 254 | [ZONE_HIGHMEM] = 0, |
e53ef38d | 255 | #endif |
d3cda233 | 256 | [ZONE_MOVABLE] = 0, |
2f1b6248 | 257 | }; |
1da177e4 | 258 | |
9420f89d | 259 | char * const zone_names[MAX_NR_ZONES] = { |
4b51d669 | 260 | #ifdef CONFIG_ZONE_DMA |
2f1b6248 | 261 | "DMA", |
4b51d669 | 262 | #endif |
fb0e7942 | 263 | #ifdef CONFIG_ZONE_DMA32 |
2f1b6248 | 264 | "DMA32", |
fb0e7942 | 265 | #endif |
2f1b6248 | 266 | "Normal", |
e53ef38d | 267 | #ifdef CONFIG_HIGHMEM |
2a1e274a | 268 | "HighMem", |
e53ef38d | 269 | #endif |
2a1e274a | 270 | "Movable", |
033fbae9 DW |
271 | #ifdef CONFIG_ZONE_DEVICE |
272 | "Device", | |
273 | #endif | |
2f1b6248 CL |
274 | }; |
275 | ||
c999fbd3 | 276 | const char * const migratetype_names[MIGRATE_TYPES] = { |
60f30350 VB |
277 | "Unmovable", |
278 | "Movable", | |
279 | "Reclaimable", | |
280 | "HighAtomic", | |
281 | #ifdef CONFIG_CMA | |
282 | "CMA", | |
283 | #endif | |
284 | #ifdef CONFIG_MEMORY_ISOLATION | |
285 | "Isolate", | |
286 | #endif | |
287 | }; | |
288 | ||
1da177e4 | 289 | int min_free_kbytes = 1024; |
42aa83cb | 290 | int user_min_free_kbytes = -1; |
e95d372c KW |
291 | static int watermark_boost_factor __read_mostly = 15000; |
292 | static int watermark_scale_factor = 10; | |
0ee332c1 TH |
293 | |
294 | /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ | |
295 | int movable_zone; | |
296 | EXPORT_SYMBOL(movable_zone); | |
c713216d | 297 | |
418508c1 | 298 | #if MAX_NUMNODES > 1 |
b9726c26 | 299 | unsigned int nr_node_ids __read_mostly = MAX_NUMNODES; |
ce0725f7 | 300 | unsigned int nr_online_nodes __read_mostly = 1; |
418508c1 | 301 | EXPORT_SYMBOL(nr_node_ids); |
62bc62a8 | 302 | EXPORT_SYMBOL(nr_online_nodes); |
418508c1 MS |
303 | #endif |
304 | ||
dcdfdd40 KS |
305 | static bool page_contains_unaccepted(struct page *page, unsigned int order); |
306 | static void accept_page(struct page *page, unsigned int order); | |
307 | static bool try_to_accept_memory(struct zone *zone, unsigned int order); | |
308 | static inline bool has_unaccepted_memory(void); | |
309 | static bool __free_unaccepted(struct page *page); | |
310 | ||
9ef9acb0 MG |
311 | int page_group_by_mobility_disabled __read_mostly; |
312 | ||
3a80a7fa | 313 | #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT |
3c0c12cc WL |
314 | /* |
315 | * During boot we initialize deferred pages on-demand, as needed, but once | |
316 | * page_alloc_init_late() has finished, the deferred pages are all initialized, | |
317 | * and we can permanently disable that path. | |
318 | */ | |
9420f89d | 319 | DEFINE_STATIC_KEY_TRUE(deferred_pages); |
3c0c12cc | 320 | |
94ae8b83 | 321 | static inline bool deferred_pages_enabled(void) |
3c0c12cc | 322 | { |
94ae8b83 | 323 | return static_branch_unlikely(&deferred_pages); |
3c0c12cc WL |
324 | } |
325 | ||
3a80a7fa | 326 | /* |
9420f89d MRI |
327 | * deferred_grow_zone() is __init, but it is called from |
328 | * get_page_from_freelist() during early boot until deferred_pages permanently | |
329 | * disables this call. This is why we have refdata wrapper to avoid warning, | |
330 | * and to ensure that the function body gets unloaded. | |
3a80a7fa | 331 | */ |
9420f89d MRI |
332 | static bool __ref |
333 | _deferred_grow_zone(struct zone *zone, unsigned int order) | |
3a80a7fa | 334 | { |
9420f89d | 335 | return deferred_grow_zone(zone, order); |
3a80a7fa MG |
336 | } |
337 | #else | |
94ae8b83 | 338 | static inline bool deferred_pages_enabled(void) |
2c335680 | 339 | { |
94ae8b83 | 340 | return false; |
2c335680 | 341 | } |
9420f89d | 342 | #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */ |
3a80a7fa | 343 | |
0b423ca2 | 344 | /* Return a pointer to the bitmap storing bits affecting a block of pages */ |
ca891f41 | 345 | static inline unsigned long *get_pageblock_bitmap(const struct page *page, |
0b423ca2 MG |
346 | unsigned long pfn) |
347 | { | |
348 | #ifdef CONFIG_SPARSEMEM | |
f1eca35a | 349 | return section_to_usemap(__pfn_to_section(pfn)); |
0b423ca2 MG |
350 | #else |
351 | return page_zone(page)->pageblock_flags; | |
352 | #endif /* CONFIG_SPARSEMEM */ | |
353 | } | |
354 | ||
ca891f41 | 355 | static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn) |
0b423ca2 MG |
356 | { |
357 | #ifdef CONFIG_SPARSEMEM | |
358 | pfn &= (PAGES_PER_SECTION-1); | |
0b423ca2 | 359 | #else |
4f9bc69a | 360 | pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn); |
0b423ca2 | 361 | #endif /* CONFIG_SPARSEMEM */ |
399b795b | 362 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; |
0b423ca2 MG |
363 | } |
364 | ||
a04d12c2 KS |
365 | /** |
366 | * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages | |
367 | * @page: The page within the block of interest | |
368 | * @pfn: The target page frame number | |
369 | * @mask: mask of bits that the caller is interested in | |
370 | * | |
371 | * Return: pageblock_bits flags | |
372 | */ | |
373 | unsigned long get_pfnblock_flags_mask(const struct page *page, | |
374 | unsigned long pfn, unsigned long mask) | |
0b423ca2 MG |
375 | { |
376 | unsigned long *bitmap; | |
377 | unsigned long bitidx, word_bitidx; | |
378 | unsigned long word; | |
379 | ||
380 | bitmap = get_pageblock_bitmap(page, pfn); | |
381 | bitidx = pfn_to_bitidx(page, pfn); | |
382 | word_bitidx = bitidx / BITS_PER_LONG; | |
383 | bitidx &= (BITS_PER_LONG-1); | |
1c563432 MK |
384 | /* |
385 | * This races, without locks, with set_pfnblock_flags_mask(). Ensure | |
386 | * a consistent read of the memory array, so that results, even though | |
387 | * racy, are not corrupted. | |
388 | */ | |
389 | word = READ_ONCE(bitmap[word_bitidx]); | |
d93d5ab9 | 390 | return (word >> bitidx) & mask; |
0b423ca2 MG |
391 | } |
392 | ||
ca891f41 MWO |
393 | static __always_inline int get_pfnblock_migratetype(const struct page *page, |
394 | unsigned long pfn) | |
0b423ca2 | 395 | { |
a04d12c2 | 396 | return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK); |
0b423ca2 MG |
397 | } |
398 | ||
399 | /** | |
400 | * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages | |
401 | * @page: The page within the block of interest | |
402 | * @flags: The flags to set | |
403 | * @pfn: The target page frame number | |
0b423ca2 MG |
404 | * @mask: mask of bits that the caller is interested in |
405 | */ | |
406 | void set_pfnblock_flags_mask(struct page *page, unsigned long flags, | |
407 | unsigned long pfn, | |
0b423ca2 MG |
408 | unsigned long mask) |
409 | { | |
410 | unsigned long *bitmap; | |
411 | unsigned long bitidx, word_bitidx; | |
04ec0061 | 412 | unsigned long word; |
0b423ca2 MG |
413 | |
414 | BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4); | |
125b860b | 415 | BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits)); |
0b423ca2 MG |
416 | |
417 | bitmap = get_pageblock_bitmap(page, pfn); | |
418 | bitidx = pfn_to_bitidx(page, pfn); | |
419 | word_bitidx = bitidx / BITS_PER_LONG; | |
420 | bitidx &= (BITS_PER_LONG-1); | |
421 | ||
422 | VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page); | |
423 | ||
d93d5ab9 WY |
424 | mask <<= bitidx; |
425 | flags <<= bitidx; | |
0b423ca2 MG |
426 | |
427 | word = READ_ONCE(bitmap[word_bitidx]); | |
04ec0061 UB |
428 | do { |
429 | } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags)); | |
0b423ca2 | 430 | } |
3a80a7fa | 431 | |
ee6f509c | 432 | void set_pageblock_migratetype(struct page *page, int migratetype) |
b2a0ac88 | 433 | { |
5d0f3f72 KM |
434 | if (unlikely(page_group_by_mobility_disabled && |
435 | migratetype < MIGRATE_PCPTYPES)) | |
49255c61 MG |
436 | migratetype = MIGRATE_UNMOVABLE; |
437 | ||
d93d5ab9 | 438 | set_pfnblock_flags_mask(page, (unsigned long)migratetype, |
535b81e2 | 439 | page_to_pfn(page), MIGRATETYPE_MASK); |
b2a0ac88 MG |
440 | } |
441 | ||
13e7444b | 442 | #ifdef CONFIG_DEBUG_VM |
c6a57e19 | 443 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page) |
1da177e4 | 444 | { |
82d9b8c8 | 445 | int ret; |
bdc8cb98 DH |
446 | unsigned seq; |
447 | unsigned long pfn = page_to_pfn(page); | |
b5e6a5a2 | 448 | unsigned long sp, start_pfn; |
c6a57e19 | 449 | |
bdc8cb98 DH |
450 | do { |
451 | seq = zone_span_seqbegin(zone); | |
b5e6a5a2 CS |
452 | start_pfn = zone->zone_start_pfn; |
453 | sp = zone->spanned_pages; | |
82d9b8c8 | 454 | ret = !zone_spans_pfn(zone, pfn); |
bdc8cb98 DH |
455 | } while (zone_span_seqretry(zone, seq)); |
456 | ||
b5e6a5a2 | 457 | if (ret) |
613813e8 DH |
458 | pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n", |
459 | pfn, zone_to_nid(zone), zone->name, | |
460 | start_pfn, start_pfn + sp); | |
b5e6a5a2 | 461 | |
bdc8cb98 | 462 | return ret; |
c6a57e19 DH |
463 | } |
464 | ||
c6a57e19 DH |
465 | /* |
466 | * Temporary debugging check for pages not lying within a given zone. | |
467 | */ | |
5bb14214 | 468 | static bool __maybe_unused bad_range(struct zone *zone, struct page *page) |
c6a57e19 DH |
469 | { |
470 | if (page_outside_zone_boundaries(zone, page)) | |
5bb14214 | 471 | return true; |
5b855aa3 | 472 | if (zone != page_zone(page)) |
5bb14214 | 473 | return true; |
c6a57e19 | 474 | |
5bb14214 | 475 | return false; |
1da177e4 | 476 | } |
13e7444b | 477 | #else |
5bb14214 | 478 | static inline bool __maybe_unused bad_range(struct zone *zone, struct page *page) |
13e7444b | 479 | { |
5bb14214 | 480 | return false; |
13e7444b NP |
481 | } |
482 | #endif | |
483 | ||
82a3241a | 484 | static void bad_page(struct page *page, const char *reason) |
1da177e4 | 485 | { |
d936cf9b HD |
486 | static unsigned long resume; |
487 | static unsigned long nr_shown; | |
488 | static unsigned long nr_unshown; | |
489 | ||
490 | /* | |
491 | * Allow a burst of 60 reports, then keep quiet for that minute; | |
492 | * or allow a steady drip of one report per second. | |
493 | */ | |
494 | if (nr_shown == 60) { | |
495 | if (time_before(jiffies, resume)) { | |
496 | nr_unshown++; | |
497 | goto out; | |
498 | } | |
499 | if (nr_unshown) { | |
ff8e8116 | 500 | pr_alert( |
1e9e6365 | 501 | "BUG: Bad page state: %lu messages suppressed\n", |
d936cf9b HD |
502 | nr_unshown); |
503 | nr_unshown = 0; | |
504 | } | |
505 | nr_shown = 0; | |
506 | } | |
507 | if (nr_shown++ == 0) | |
508 | resume = jiffies + 60 * HZ; | |
509 | ||
ff8e8116 | 510 | pr_alert("BUG: Bad page state in process %s pfn:%05lx\n", |
3dc14741 | 511 | current->comm, page_to_pfn(page)); |
d2f07ec0 | 512 | dump_page(page, reason); |
3dc14741 | 513 | |
4f31888c | 514 | print_modules(); |
1da177e4 | 515 | dump_stack(); |
d936cf9b | 516 | out: |
8cc3b392 | 517 | /* Leave bad fields for debug, except PageBuddy could make trouble */ |
22b751c3 | 518 | page_mapcount_reset(page); /* remove PageBuddy */ |
373d4d09 | 519 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
1da177e4 LT |
520 | } |
521 | ||
44042b44 MG |
522 | static inline unsigned int order_to_pindex(int migratetype, int order) |
523 | { | |
44042b44 MG |
524 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
525 | if (order > PAGE_ALLOC_COSTLY_ORDER) { | |
526 | VM_BUG_ON(order != pageblock_order); | |
5d0a661d | 527 | return NR_LOWORDER_PCP_LISTS; |
44042b44 MG |
528 | } |
529 | #else | |
530 | VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER); | |
531 | #endif | |
532 | ||
c1dc69e6 | 533 | return (MIGRATE_PCPTYPES * order) + migratetype; |
44042b44 MG |
534 | } |
535 | ||
536 | static inline int pindex_to_order(unsigned int pindex) | |
537 | { | |
538 | int order = pindex / MIGRATE_PCPTYPES; | |
539 | ||
540 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
5d0a661d | 541 | if (pindex == NR_LOWORDER_PCP_LISTS) |
44042b44 | 542 | order = pageblock_order; |
44042b44 MG |
543 | #else |
544 | VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER); | |
545 | #endif | |
546 | ||
547 | return order; | |
548 | } | |
549 | ||
550 | static inline bool pcp_allowed_order(unsigned int order) | |
551 | { | |
552 | if (order <= PAGE_ALLOC_COSTLY_ORDER) | |
553 | return true; | |
554 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
555 | if (order == pageblock_order) | |
556 | return true; | |
557 | #endif | |
558 | return false; | |
559 | } | |
560 | ||
21d02f8f MG |
561 | static inline void free_the_page(struct page *page, unsigned int order) |
562 | { | |
44042b44 MG |
563 | if (pcp_allowed_order(order)) /* Via pcp? */ |
564 | free_unref_page(page, order); | |
21d02f8f MG |
565 | else |
566 | __free_pages_ok(page, order, FPI_NONE); | |
567 | } | |
568 | ||
1da177e4 LT |
569 | /* |
570 | * Higher-order pages are called "compound pages". They are structured thusly: | |
571 | * | |
1d798ca3 | 572 | * The first PAGE_SIZE page is called the "head page" and have PG_head set. |
1da177e4 | 573 | * |
1d798ca3 KS |
574 | * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded |
575 | * in bit 0 of page->compound_head. The rest of bits is pointer to head page. | |
1da177e4 | 576 | * |
1d798ca3 | 577 | * The first tail page's ->compound_order holds the order of allocation. |
41d78ba5 | 578 | * This usage means that zero-order pages may not be compound. |
1da177e4 | 579 | */ |
d98c7a09 | 580 | |
d00181b9 | 581 | void prep_compound_page(struct page *page, unsigned int order) |
18229df5 AW |
582 | { |
583 | int i; | |
584 | int nr_pages = 1 << order; | |
585 | ||
18229df5 | 586 | __SetPageHead(page); |
5b24eeef JM |
587 | for (i = 1; i < nr_pages; i++) |
588 | prep_compound_tail(page, i); | |
1378a5ee | 589 | |
5b24eeef | 590 | prep_compound_head(page, order); |
18229df5 AW |
591 | } |
592 | ||
5375336c MWO |
593 | void destroy_large_folio(struct folio *folio) |
594 | { | |
dd6fa0b6 | 595 | if (folio_test_hugetlb(folio)) { |
454a00c4 | 596 | free_huge_folio(folio); |
dd6fa0b6 MWO |
597 | return; |
598 | } | |
8dc4a8f1 | 599 | |
de53c05f | 600 | if (folio_test_large_rmappable(folio)) |
8dc4a8f1 | 601 | folio_undo_large_rmappable(folio); |
5375336c | 602 | |
0f2f43fa MWO |
603 | mem_cgroup_uncharge(folio); |
604 | free_the_page(&folio->page, folio_order(folio)); | |
5375336c MWO |
605 | } |
606 | ||
ab130f91 | 607 | static inline void set_buddy_order(struct page *page, unsigned int order) |
6aa3001b | 608 | { |
4c21e2f2 | 609 | set_page_private(page, order); |
676165a8 | 610 | __SetPageBuddy(page); |
1da177e4 LT |
611 | } |
612 | ||
5e1f0f09 MG |
613 | #ifdef CONFIG_COMPACTION |
614 | static inline struct capture_control *task_capc(struct zone *zone) | |
615 | { | |
616 | struct capture_control *capc = current->capture_control; | |
617 | ||
deba0487 | 618 | return unlikely(capc) && |
5e1f0f09 MG |
619 | !(current->flags & PF_KTHREAD) && |
620 | !capc->page && | |
deba0487 | 621 | capc->cc->zone == zone ? capc : NULL; |
5e1f0f09 MG |
622 | } |
623 | ||
624 | static inline bool | |
625 | compaction_capture(struct capture_control *capc, struct page *page, | |
626 | int order, int migratetype) | |
627 | { | |
628 | if (!capc || order != capc->cc->order) | |
629 | return false; | |
630 | ||
631 | /* Do not accidentally pollute CMA or isolated regions*/ | |
632 | if (is_migrate_cma(migratetype) || | |
633 | is_migrate_isolate(migratetype)) | |
634 | return false; | |
635 | ||
636 | /* | |
f0953a1b | 637 | * Do not let lower order allocations pollute a movable pageblock. |
5e1f0f09 MG |
638 | * This might let an unmovable request use a reclaimable pageblock |
639 | * and vice-versa but no more than normal fallback logic which can | |
640 | * have trouble finding a high-order free page. | |
641 | */ | |
642 | if (order < pageblock_order && migratetype == MIGRATE_MOVABLE) | |
643 | return false; | |
644 | ||
645 | capc->page = page; | |
646 | return true; | |
647 | } | |
648 | ||
649 | #else | |
650 | static inline struct capture_control *task_capc(struct zone *zone) | |
651 | { | |
652 | return NULL; | |
653 | } | |
654 | ||
655 | static inline bool | |
656 | compaction_capture(struct capture_control *capc, struct page *page, | |
657 | int order, int migratetype) | |
658 | { | |
659 | return false; | |
660 | } | |
661 | #endif /* CONFIG_COMPACTION */ | |
662 | ||
6ab01363 AD |
663 | /* Used for pages not on another list */ |
664 | static inline void add_to_free_list(struct page *page, struct zone *zone, | |
665 | unsigned int order, int migratetype) | |
666 | { | |
667 | struct free_area *area = &zone->free_area[order]; | |
668 | ||
bf75f200 | 669 | list_add(&page->buddy_list, &area->free_list[migratetype]); |
6ab01363 AD |
670 | area->nr_free++; |
671 | } | |
672 | ||
673 | /* Used for pages not on another list */ | |
674 | static inline void add_to_free_list_tail(struct page *page, struct zone *zone, | |
675 | unsigned int order, int migratetype) | |
676 | { | |
677 | struct free_area *area = &zone->free_area[order]; | |
678 | ||
bf75f200 | 679 | list_add_tail(&page->buddy_list, &area->free_list[migratetype]); |
6ab01363 AD |
680 | area->nr_free++; |
681 | } | |
682 | ||
293ffa5e DH |
683 | /* |
684 | * Used for pages which are on another list. Move the pages to the tail | |
685 | * of the list - so the moved pages won't immediately be considered for | |
686 | * allocation again (e.g., optimization for memory onlining). | |
687 | */ | |
6ab01363 AD |
688 | static inline void move_to_free_list(struct page *page, struct zone *zone, |
689 | unsigned int order, int migratetype) | |
690 | { | |
691 | struct free_area *area = &zone->free_area[order]; | |
692 | ||
bf75f200 | 693 | list_move_tail(&page->buddy_list, &area->free_list[migratetype]); |
6ab01363 AD |
694 | } |
695 | ||
696 | static inline void del_page_from_free_list(struct page *page, struct zone *zone, | |
697 | unsigned int order) | |
698 | { | |
36e66c55 AD |
699 | /* clear reported state and update reported page count */ |
700 | if (page_reported(page)) | |
701 | __ClearPageReported(page); | |
702 | ||
bf75f200 | 703 | list_del(&page->buddy_list); |
6ab01363 AD |
704 | __ClearPageBuddy(page); |
705 | set_page_private(page, 0); | |
706 | zone->free_area[order].nr_free--; | |
707 | } | |
708 | ||
5d671eb4 MRI |
709 | static inline struct page *get_page_from_free_area(struct free_area *area, |
710 | int migratetype) | |
711 | { | |
712 | return list_first_entry_or_null(&area->free_list[migratetype], | |
1bf61092 | 713 | struct page, buddy_list); |
5d671eb4 MRI |
714 | } |
715 | ||
a2129f24 AD |
716 | /* |
717 | * If this is not the largest possible page, check if the buddy | |
718 | * of the next-highest order is free. If it is, it's possible | |
719 | * that pages are being freed that will coalesce soon. In case, | |
720 | * that is happening, add the free page to the tail of the list | |
721 | * so it's less likely to be used soon and more likely to be merged | |
722 | * as a higher order page | |
723 | */ | |
724 | static inline bool | |
725 | buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn, | |
726 | struct page *page, unsigned int order) | |
727 | { | |
8170ac47 ZY |
728 | unsigned long higher_page_pfn; |
729 | struct page *higher_page; | |
a2129f24 | 730 | |
5e0a760b | 731 | if (order >= MAX_PAGE_ORDER - 1) |
a2129f24 AD |
732 | return false; |
733 | ||
8170ac47 ZY |
734 | higher_page_pfn = buddy_pfn & pfn; |
735 | higher_page = page + (higher_page_pfn - pfn); | |
a2129f24 | 736 | |
8170ac47 ZY |
737 | return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1, |
738 | NULL) != NULL; | |
a2129f24 AD |
739 | } |
740 | ||
1da177e4 LT |
741 | /* |
742 | * Freeing function for a buddy system allocator. | |
743 | * | |
744 | * The concept of a buddy system is to maintain direct-mapped table | |
745 | * (containing bit values) for memory blocks of various "orders". | |
746 | * The bottom level table contains the map for the smallest allocatable | |
747 | * units of memory (here, pages), and each level above it describes | |
748 | * pairs of units from the levels below, hence, "buddies". | |
749 | * At a high level, all that happens here is marking the table entry | |
750 | * at the bottom level available, and propagating the changes upward | |
751 | * as necessary, plus some accounting needed to play nicely with other | |
752 | * parts of the VM system. | |
753 | * At each level, we keep a list of pages, which are heads of continuous | |
6e292b9b MW |
754 | * free pages of length of (1 << order) and marked with PageBuddy. |
755 | * Page's order is recorded in page_private(page) field. | |
1da177e4 | 756 | * So when we are allocating or freeing one, we can derive the state of the |
5f63b720 MN |
757 | * other. That is, if we allocate a small block, and both were |
758 | * free, the remainder of the region must be split into blocks. | |
1da177e4 | 759 | * If a block is freed, and its buddy is also free, then this |
5f63b720 | 760 | * triggers coalescing into a block of larger size. |
1da177e4 | 761 | * |
6d49e352 | 762 | * -- nyc |
1da177e4 LT |
763 | */ |
764 | ||
48db57f8 | 765 | static inline void __free_one_page(struct page *page, |
dc4b0caf | 766 | unsigned long pfn, |
ed0ae21d | 767 | struct zone *zone, unsigned int order, |
f04a5d5d | 768 | int migratetype, fpi_t fpi_flags) |
1da177e4 | 769 | { |
a2129f24 | 770 | struct capture_control *capc = task_capc(zone); |
dae37a5d | 771 | unsigned long buddy_pfn = 0; |
a2129f24 | 772 | unsigned long combined_pfn; |
a2129f24 AD |
773 | struct page *buddy; |
774 | bool to_tail; | |
d9dddbf5 | 775 | |
d29bb978 | 776 | VM_BUG_ON(!zone_is_initialized(zone)); |
6e9f0d58 | 777 | VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page); |
1da177e4 | 778 | |
ed0ae21d | 779 | VM_BUG_ON(migratetype == -1); |
d9dddbf5 | 780 | if (likely(!is_migrate_isolate(migratetype))) |
8f82b55d | 781 | __mod_zone_freepage_state(zone, 1 << order, migratetype); |
ed0ae21d | 782 | |
76741e77 | 783 | VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page); |
309381fe | 784 | VM_BUG_ON_PAGE(bad_range(zone, page), page); |
1da177e4 | 785 | |
5e0a760b | 786 | while (order < MAX_PAGE_ORDER) { |
5e1f0f09 MG |
787 | if (compaction_capture(capc, page, order, migratetype)) { |
788 | __mod_zone_freepage_state(zone, -(1 << order), | |
789 | migratetype); | |
790 | return; | |
791 | } | |
13ad59df | 792 | |
8170ac47 ZY |
793 | buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn); |
794 | if (!buddy) | |
d9dddbf5 | 795 | goto done_merging; |
bb0e28eb ZY |
796 | |
797 | if (unlikely(order >= pageblock_order)) { | |
798 | /* | |
799 | * We want to prevent merge between freepages on pageblock | |
800 | * without fallbacks and normal pageblock. Without this, | |
801 | * pageblock isolation could cause incorrect freepage or CMA | |
802 | * accounting or HIGHATOMIC accounting. | |
803 | */ | |
b5ffd297 | 804 | int buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn); |
bb0e28eb ZY |
805 | |
806 | if (migratetype != buddy_mt | |
807 | && (!migratetype_is_mergeable(migratetype) || | |
808 | !migratetype_is_mergeable(buddy_mt))) | |
809 | goto done_merging; | |
810 | } | |
811 | ||
c0a32fc5 SG |
812 | /* |
813 | * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, | |
814 | * merge with it and move up one order. | |
815 | */ | |
b03641af | 816 | if (page_is_guard(buddy)) |
2847cf95 | 817 | clear_page_guard(zone, buddy, order, migratetype); |
b03641af | 818 | else |
6ab01363 | 819 | del_page_from_free_list(buddy, zone, order); |
76741e77 VB |
820 | combined_pfn = buddy_pfn & pfn; |
821 | page = page + (combined_pfn - pfn); | |
822 | pfn = combined_pfn; | |
1da177e4 LT |
823 | order++; |
824 | } | |
d9dddbf5 VB |
825 | |
826 | done_merging: | |
ab130f91 | 827 | set_buddy_order(page, order); |
6dda9d55 | 828 | |
47b6a24a DH |
829 | if (fpi_flags & FPI_TO_TAIL) |
830 | to_tail = true; | |
831 | else if (is_shuffle_order(order)) | |
a2129f24 | 832 | to_tail = shuffle_pick_tail(); |
97500a4a | 833 | else |
a2129f24 | 834 | to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order); |
97500a4a | 835 | |
a2129f24 | 836 | if (to_tail) |
6ab01363 | 837 | add_to_free_list_tail(page, zone, order, migratetype); |
a2129f24 | 838 | else |
6ab01363 | 839 | add_to_free_list(page, zone, order, migratetype); |
36e66c55 AD |
840 | |
841 | /* Notify page reporting subsystem of freed page */ | |
f04a5d5d | 842 | if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY)) |
36e66c55 | 843 | page_reporting_notify_free(order); |
1da177e4 LT |
844 | } |
845 | ||
b2c9e2fb ZY |
846 | /** |
847 | * split_free_page() -- split a free page at split_pfn_offset | |
848 | * @free_page: the original free page | |
849 | * @order: the order of the page | |
850 | * @split_pfn_offset: split offset within the page | |
851 | * | |
86d28b07 ZY |
852 | * Return -ENOENT if the free page is changed, otherwise 0 |
853 | * | |
b2c9e2fb ZY |
854 | * It is used when the free page crosses two pageblocks with different migratetypes |
855 | * at split_pfn_offset within the page. The split free page will be put into | |
856 | * separate migratetype lists afterwards. Otherwise, the function achieves | |
857 | * nothing. | |
858 | */ | |
86d28b07 ZY |
859 | int split_free_page(struct page *free_page, |
860 | unsigned int order, unsigned long split_pfn_offset) | |
b2c9e2fb ZY |
861 | { |
862 | struct zone *zone = page_zone(free_page); | |
863 | unsigned long free_page_pfn = page_to_pfn(free_page); | |
864 | unsigned long pfn; | |
865 | unsigned long flags; | |
866 | int free_page_order; | |
86d28b07 ZY |
867 | int mt; |
868 | int ret = 0; | |
b2c9e2fb | 869 | |
88ee1343 | 870 | if (split_pfn_offset == 0) |
86d28b07 | 871 | return ret; |
88ee1343 | 872 | |
b2c9e2fb | 873 | spin_lock_irqsave(&zone->lock, flags); |
86d28b07 ZY |
874 | |
875 | if (!PageBuddy(free_page) || buddy_order(free_page) != order) { | |
876 | ret = -ENOENT; | |
877 | goto out; | |
878 | } | |
879 | ||
b5ffd297 | 880 | mt = get_pfnblock_migratetype(free_page, free_page_pfn); |
86d28b07 ZY |
881 | if (likely(!is_migrate_isolate(mt))) |
882 | __mod_zone_freepage_state(zone, -(1UL << order), mt); | |
883 | ||
b2c9e2fb ZY |
884 | del_page_from_free_list(free_page, zone, order); |
885 | for (pfn = free_page_pfn; | |
886 | pfn < free_page_pfn + (1UL << order);) { | |
887 | int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn); | |
888 | ||
86d28b07 | 889 | free_page_order = min_t(unsigned int, |
88ee1343 ZY |
890 | pfn ? __ffs(pfn) : order, |
891 | __fls(split_pfn_offset)); | |
b2c9e2fb ZY |
892 | __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order, |
893 | mt, FPI_NONE); | |
894 | pfn += 1UL << free_page_order; | |
895 | split_pfn_offset -= (1UL << free_page_order); | |
896 | /* we have done the first part, now switch to second part */ | |
897 | if (split_pfn_offset == 0) | |
898 | split_pfn_offset = (1UL << order) - (pfn - free_page_pfn); | |
899 | } | |
86d28b07 | 900 | out: |
b2c9e2fb | 901 | spin_unlock_irqrestore(&zone->lock, flags); |
86d28b07 | 902 | return ret; |
b2c9e2fb | 903 | } |
7bfec6f4 MG |
904 | /* |
905 | * A bad page could be due to a number of fields. Instead of multiple branches, | |
906 | * try and check multiple fields with one check. The caller must do a detailed | |
907 | * check if necessary. | |
908 | */ | |
909 | static inline bool page_expected_state(struct page *page, | |
910 | unsigned long check_flags) | |
911 | { | |
912 | if (unlikely(atomic_read(&page->_mapcount) != -1)) | |
913 | return false; | |
914 | ||
915 | if (unlikely((unsigned long)page->mapping | | |
916 | page_ref_count(page) | | |
917 | #ifdef CONFIG_MEMCG | |
48060834 | 918 | page->memcg_data | |
dba1b8a7 JDB |
919 | #endif |
920 | #ifdef CONFIG_PAGE_POOL | |
921 | ((page->pp_magic & ~0x3UL) == PP_SIGNATURE) | | |
7bfec6f4 MG |
922 | #endif |
923 | (page->flags & check_flags))) | |
924 | return false; | |
925 | ||
926 | return true; | |
927 | } | |
928 | ||
58b7f119 | 929 | static const char *page_bad_reason(struct page *page, unsigned long flags) |
1da177e4 | 930 | { |
82a3241a | 931 | const char *bad_reason = NULL; |
f0b791a3 | 932 | |
53f9263b | 933 | if (unlikely(atomic_read(&page->_mapcount) != -1)) |
f0b791a3 DH |
934 | bad_reason = "nonzero mapcount"; |
935 | if (unlikely(page->mapping != NULL)) | |
936 | bad_reason = "non-NULL mapping"; | |
fe896d18 | 937 | if (unlikely(page_ref_count(page) != 0)) |
0139aa7b | 938 | bad_reason = "nonzero _refcount"; |
58b7f119 WY |
939 | if (unlikely(page->flags & flags)) { |
940 | if (flags == PAGE_FLAGS_CHECK_AT_PREP) | |
941 | bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set"; | |
942 | else | |
943 | bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set"; | |
f0b791a3 | 944 | } |
9edad6ea | 945 | #ifdef CONFIG_MEMCG |
48060834 | 946 | if (unlikely(page->memcg_data)) |
9edad6ea | 947 | bad_reason = "page still charged to cgroup"; |
dba1b8a7 JDB |
948 | #endif |
949 | #ifdef CONFIG_PAGE_POOL | |
950 | if (unlikely((page->pp_magic & ~0x3UL) == PP_SIGNATURE)) | |
951 | bad_reason = "page_pool leak"; | |
9edad6ea | 952 | #endif |
58b7f119 WY |
953 | return bad_reason; |
954 | } | |
955 | ||
a8368cd8 | 956 | static void free_page_is_bad_report(struct page *page) |
58b7f119 WY |
957 | { |
958 | bad_page(page, | |
959 | page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE)); | |
bb552ac6 MG |
960 | } |
961 | ||
a8368cd8 | 962 | static inline bool free_page_is_bad(struct page *page) |
bb552ac6 | 963 | { |
da838d4f | 964 | if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE))) |
a8368cd8 | 965 | return false; |
bb552ac6 MG |
966 | |
967 | /* Something has gone sideways, find it */ | |
a8368cd8 AM |
968 | free_page_is_bad_report(page); |
969 | return true; | |
1da177e4 LT |
970 | } |
971 | ||
ecbb490d KW |
972 | static inline bool is_check_pages_enabled(void) |
973 | { | |
974 | return static_branch_unlikely(&check_pages_enabled); | |
975 | } | |
976 | ||
8666925c | 977 | static int free_tail_page_prepare(struct page *head_page, struct page *page) |
4db7548c | 978 | { |
94688e8e | 979 | struct folio *folio = (struct folio *)head_page; |
4db7548c MG |
980 | int ret = 1; |
981 | ||
982 | /* | |
983 | * We rely page->lru.next never has bit 0 set, unless the page | |
984 | * is PageTail(). Let's make sure that's true even for poisoned ->lru. | |
985 | */ | |
986 | BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1); | |
987 | ||
ecbb490d | 988 | if (!is_check_pages_enabled()) { |
4db7548c MG |
989 | ret = 0; |
990 | goto out; | |
991 | } | |
992 | switch (page - head_page) { | |
993 | case 1: | |
cb67f428 | 994 | /* the first tail page: these may be in place of ->mapping */ |
65a689f3 MWO |
995 | if (unlikely(folio_entire_mapcount(folio))) { |
996 | bad_page(page, "nonzero entire_mapcount"); | |
4db7548c MG |
997 | goto out; |
998 | } | |
65a689f3 MWO |
999 | if (unlikely(atomic_read(&folio->_nr_pages_mapped))) { |
1000 | bad_page(page, "nonzero nr_pages_mapped"); | |
cb67f428 HD |
1001 | goto out; |
1002 | } | |
94688e8e MWO |
1003 | if (unlikely(atomic_read(&folio->_pincount))) { |
1004 | bad_page(page, "nonzero pincount"); | |
cb67f428 HD |
1005 | goto out; |
1006 | } | |
4db7548c MG |
1007 | break; |
1008 | case 2: | |
1009 | /* | |
1010 | * the second tail page: ->mapping is | |
fa3015b7 | 1011 | * deferred_list.next -- ignore value. |
4db7548c MG |
1012 | */ |
1013 | break; | |
1014 | default: | |
1015 | if (page->mapping != TAIL_MAPPING) { | |
82a3241a | 1016 | bad_page(page, "corrupted mapping in tail page"); |
4db7548c MG |
1017 | goto out; |
1018 | } | |
1019 | break; | |
1020 | } | |
1021 | if (unlikely(!PageTail(page))) { | |
82a3241a | 1022 | bad_page(page, "PageTail not set"); |
4db7548c MG |
1023 | goto out; |
1024 | } | |
1025 | if (unlikely(compound_head(page) != head_page)) { | |
82a3241a | 1026 | bad_page(page, "compound_head not consistent"); |
4db7548c MG |
1027 | goto out; |
1028 | } | |
1029 | ret = 0; | |
1030 | out: | |
1031 | page->mapping = NULL; | |
1032 | clear_compound_head(page); | |
1033 | return ret; | |
1034 | } | |
1035 | ||
94ae8b83 AK |
1036 | /* |
1037 | * Skip KASAN memory poisoning when either: | |
1038 | * | |
0a54864f PC |
1039 | * 1. For generic KASAN: deferred memory initialization has not yet completed. |
1040 | * Tag-based KASAN modes skip pages freed via deferred memory initialization | |
1041 | * using page tags instead (see below). | |
1042 | * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating | |
1043 | * that error detection is disabled for accesses via the page address. | |
1044 | * | |
1045 | * Pages will have match-all tags in the following circumstances: | |
1046 | * | |
1047 | * 1. Pages are being initialized for the first time, including during deferred | |
1048 | * memory init; see the call to page_kasan_tag_reset in __init_single_page. | |
1049 | * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the | |
1050 | * exception of pages unpoisoned by kasan_unpoison_vmalloc. | |
1051 | * 3. The allocation was excluded from being checked due to sampling, | |
44383cef | 1052 | * see the call to kasan_unpoison_pages. |
94ae8b83 AK |
1053 | * |
1054 | * Poisoning pages during deferred memory init will greatly lengthen the | |
1055 | * process and cause problem in large memory systems as the deferred pages | |
1056 | * initialization is done with interrupt disabled. | |
1057 | * | |
1058 | * Assuming that there will be no reference to those newly initialized | |
1059 | * pages before they are ever allocated, this should have no effect on | |
1060 | * KASAN memory tracking as the poison will be properly inserted at page | |
1061 | * allocation time. The only corner case is when pages are allocated by | |
1062 | * on-demand allocation and then freed again before the deferred pages | |
1063 | * initialization is done, but this is not likely to happen. | |
1064 | */ | |
5267fe5d | 1065 | static inline bool should_skip_kasan_poison(struct page *page) |
94ae8b83 | 1066 | { |
0a54864f PC |
1067 | if (IS_ENABLED(CONFIG_KASAN_GENERIC)) |
1068 | return deferred_pages_enabled(); | |
1069 | ||
5cb6674b | 1070 | return page_kasan_tag(page) == KASAN_TAG_KERNEL; |
94ae8b83 AK |
1071 | } |
1072 | ||
aeaec8e2 | 1073 | static void kernel_init_pages(struct page *page, int numpages) |
6471384a AP |
1074 | { |
1075 | int i; | |
1076 | ||
9e15afa5 QC |
1077 | /* s390's use of memset() could override KASAN redzones. */ |
1078 | kasan_disable_current(); | |
d9da8f6c AK |
1079 | for (i = 0; i < numpages; i++) |
1080 | clear_highpage_kasan_tagged(page + i); | |
9e15afa5 | 1081 | kasan_enable_current(); |
6471384a AP |
1082 | } |
1083 | ||
733aea0b | 1084 | __always_inline bool free_pages_prepare(struct page *page, |
5267fe5d | 1085 | unsigned int order) |
4db7548c | 1086 | { |
e2769dbd | 1087 | int bad = 0; |
5267fe5d | 1088 | bool skip_kasan_poison = should_skip_kasan_poison(page); |
c3525330 | 1089 | bool init = want_init_on_free(); |
76f26535 | 1090 | bool compound = PageCompound(page); |
4db7548c | 1091 | |
4db7548c MG |
1092 | VM_BUG_ON_PAGE(PageTail(page), page); |
1093 | ||
e2769dbd | 1094 | trace_mm_page_free(page, order); |
b073d7f8 | 1095 | kmsan_free_page(page, order); |
e2769dbd | 1096 | |
17b46e7b BJ |
1097 | if (memcg_kmem_online() && PageMemcgKmem(page)) |
1098 | __memcg_kmem_uncharge_page(page, order); | |
1099 | ||
79f5f8fa | 1100 | if (unlikely(PageHWPoison(page)) && !order) { |
17b46e7b | 1101 | /* Do not let hwpoison pages hit pcplists/buddy */ |
79f5f8fa | 1102 | reset_page_owner(page, order); |
df4e817b | 1103 | page_table_check_free(page, order); |
79f5f8fa OS |
1104 | return false; |
1105 | } | |
1106 | ||
76f26535 HY |
1107 | VM_BUG_ON_PAGE(compound && compound_order(page) != order, page); |
1108 | ||
e2769dbd MG |
1109 | /* |
1110 | * Check tail pages before head page information is cleared to | |
1111 | * avoid checking PageCompound for order-0 pages. | |
1112 | */ | |
1113 | if (unlikely(order)) { | |
e2769dbd MG |
1114 | int i; |
1115 | ||
cb67f428 | 1116 | if (compound) |
9c5ccf2d | 1117 | page[1].flags &= ~PAGE_FLAGS_SECOND; |
e2769dbd MG |
1118 | for (i = 1; i < (1 << order); i++) { |
1119 | if (compound) | |
8666925c | 1120 | bad += free_tail_page_prepare(page, page + i); |
fce0b421 | 1121 | if (is_check_pages_enabled()) { |
8666925c | 1122 | if (free_page_is_bad(page + i)) { |
700d2e9a VB |
1123 | bad++; |
1124 | continue; | |
1125 | } | |
e2769dbd MG |
1126 | } |
1127 | (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
1128 | } | |
1129 | } | |
bda807d4 | 1130 | if (PageMappingFlags(page)) |
4db7548c | 1131 | page->mapping = NULL; |
fce0b421 | 1132 | if (is_check_pages_enabled()) { |
700d2e9a VB |
1133 | if (free_page_is_bad(page)) |
1134 | bad++; | |
1135 | if (bad) | |
1136 | return false; | |
1137 | } | |
4db7548c | 1138 | |
e2769dbd MG |
1139 | page_cpupid_reset_last(page); |
1140 | page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
1141 | reset_page_owner(page, order); | |
df4e817b | 1142 | page_table_check_free(page, order); |
4db7548c MG |
1143 | |
1144 | if (!PageHighMem(page)) { | |
1145 | debug_check_no_locks_freed(page_address(page), | |
e2769dbd | 1146 | PAGE_SIZE << order); |
4db7548c | 1147 | debug_check_no_obj_freed(page_address(page), |
e2769dbd | 1148 | PAGE_SIZE << order); |
4db7548c | 1149 | } |
6471384a | 1150 | |
8db26a3d VB |
1151 | kernel_poison_pages(page, 1 << order); |
1152 | ||
f9d79e8d | 1153 | /* |
1bb5eab3 | 1154 | * As memory initialization might be integrated into KASAN, |
7c13c163 | 1155 | * KASAN poisoning and memory initialization code must be |
1bb5eab3 AK |
1156 | * kept together to avoid discrepancies in behavior. |
1157 | * | |
f9d79e8d AK |
1158 | * With hardware tag-based KASAN, memory tags must be set before the |
1159 | * page becomes unavailable via debug_pagealloc or arch_free_page. | |
1160 | */ | |
f446883d | 1161 | if (!skip_kasan_poison) { |
c3525330 | 1162 | kasan_poison_pages(page, order, init); |
f9d79e8d | 1163 | |
db8a0477 AK |
1164 | /* Memory is already initialized if KASAN did it internally. */ |
1165 | if (kasan_has_integrated_init()) | |
1166 | init = false; | |
1167 | } | |
1168 | if (init) | |
aeaec8e2 | 1169 | kernel_init_pages(page, 1 << order); |
db8a0477 | 1170 | |
234fdce8 QC |
1171 | /* |
1172 | * arch_free_page() can make the page's contents inaccessible. s390 | |
1173 | * does this. So nothing which can access the page's contents should | |
1174 | * happen after this. | |
1175 | */ | |
1176 | arch_free_page(page, order); | |
1177 | ||
77bc7fd6 | 1178 | debug_pagealloc_unmap_pages(page, 1 << order); |
d6332692 | 1179 | |
4db7548c MG |
1180 | return true; |
1181 | } | |
1182 | ||
1da177e4 | 1183 | /* |
5f8dcc21 | 1184 | * Frees a number of pages from the PCP lists |
7cba630b | 1185 | * Assumes all pages on list are in same zone. |
207f36ee | 1186 | * count is the number of pages to free. |
1da177e4 | 1187 | */ |
5f8dcc21 | 1188 | static void free_pcppages_bulk(struct zone *zone, int count, |
fd56eef2 MG |
1189 | struct per_cpu_pages *pcp, |
1190 | int pindex) | |
1da177e4 | 1191 | { |
57490774 | 1192 | unsigned long flags; |
44042b44 | 1193 | unsigned int order; |
3777999d | 1194 | bool isolated_pageblocks; |
8b10b465 | 1195 | struct page *page; |
f2260e6b | 1196 | |
88e8ac11 CTR |
1197 | /* |
1198 | * Ensure proper count is passed which otherwise would stuck in the | |
1199 | * below while (list_empty(list)) loop. | |
1200 | */ | |
1201 | count = min(pcp->count, count); | |
d61372bc MG |
1202 | |
1203 | /* Ensure requested pindex is drained first. */ | |
1204 | pindex = pindex - 1; | |
1205 | ||
57490774 | 1206 | spin_lock_irqsave(&zone->lock, flags); |
8b10b465 MG |
1207 | isolated_pageblocks = has_isolate_pageblock(zone); |
1208 | ||
44042b44 | 1209 | while (count > 0) { |
5f8dcc21 | 1210 | struct list_head *list; |
fd56eef2 | 1211 | int nr_pages; |
5f8dcc21 | 1212 | |
fd56eef2 | 1213 | /* Remove pages from lists in a round-robin fashion. */ |
5f8dcc21 | 1214 | do { |
f142b2c2 KS |
1215 | if (++pindex > NR_PCP_LISTS - 1) |
1216 | pindex = 0; | |
44042b44 | 1217 | list = &pcp->lists[pindex]; |
f142b2c2 | 1218 | } while (list_empty(list)); |
48db57f8 | 1219 | |
44042b44 | 1220 | order = pindex_to_order(pindex); |
fd56eef2 | 1221 | nr_pages = 1 << order; |
a6f9edd6 | 1222 | do { |
8b10b465 MG |
1223 | int mt; |
1224 | ||
bf75f200 | 1225 | page = list_last_entry(list, struct page, pcp_list); |
8b10b465 MG |
1226 | mt = get_pcppage_migratetype(page); |
1227 | ||
0a5f4e5b | 1228 | /* must delete to avoid corrupting pcp list */ |
bf75f200 | 1229 | list_del(&page->pcp_list); |
fd56eef2 MG |
1230 | count -= nr_pages; |
1231 | pcp->count -= nr_pages; | |
aa016d14 | 1232 | |
8b10b465 MG |
1233 | /* MIGRATE_ISOLATE page should not go to pcplists */ |
1234 | VM_BUG_ON_PAGE(is_migrate_isolate(mt), page); | |
1235 | /* Pageblock could have been isolated meanwhile */ | |
1236 | if (unlikely(isolated_pageblocks)) | |
1237 | mt = get_pageblock_migratetype(page); | |
0a5f4e5b | 1238 | |
8b10b465 MG |
1239 | __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE); |
1240 | trace_mm_page_pcpu_drain(page, order, mt); | |
1241 | } while (count > 0 && !list_empty(list)); | |
0a5f4e5b | 1242 | } |
8b10b465 | 1243 | |
57490774 | 1244 | spin_unlock_irqrestore(&zone->lock, flags); |
1da177e4 LT |
1245 | } |
1246 | ||
dc4b0caf MG |
1247 | static void free_one_page(struct zone *zone, |
1248 | struct page *page, unsigned long pfn, | |
7aeb09f9 | 1249 | unsigned int order, |
7fef431b | 1250 | int migratetype, fpi_t fpi_flags) |
1da177e4 | 1251 | { |
df1acc85 MG |
1252 | unsigned long flags; |
1253 | ||
1254 | spin_lock_irqsave(&zone->lock, flags); | |
ad53f92e JK |
1255 | if (unlikely(has_isolate_pageblock(zone) || |
1256 | is_migrate_isolate(migratetype))) { | |
1257 | migratetype = get_pfnblock_migratetype(page, pfn); | |
ad53f92e | 1258 | } |
7fef431b | 1259 | __free_one_page(page, pfn, zone, order, migratetype, fpi_flags); |
df1acc85 | 1260 | spin_unlock_irqrestore(&zone->lock, flags); |
48db57f8 NP |
1261 | } |
1262 | ||
7fef431b DH |
1263 | static void __free_pages_ok(struct page *page, unsigned int order, |
1264 | fpi_t fpi_flags) | |
ec95f53a | 1265 | { |
95e34412 | 1266 | int migratetype; |
dc4b0caf | 1267 | unsigned long pfn = page_to_pfn(page); |
56f0e661 | 1268 | struct zone *zone = page_zone(page); |
ec95f53a | 1269 | |
5267fe5d | 1270 | if (!free_pages_prepare(page, order)) |
ec95f53a KM |
1271 | return; |
1272 | ||
ac4b2901 DW |
1273 | /* |
1274 | * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here | |
1275 | * is used to avoid calling get_pfnblock_migratetype() under the lock. | |
1276 | * This will reduce the lock holding time. | |
1277 | */ | |
cfc47a28 | 1278 | migratetype = get_pfnblock_migratetype(page, pfn); |
dbbee9d5 | 1279 | |
250ae189 | 1280 | free_one_page(zone, page, pfn, order, migratetype, fpi_flags); |
90249993 | 1281 | |
d34b0733 | 1282 | __count_vm_events(PGFREE, 1 << order); |
1da177e4 LT |
1283 | } |
1284 | ||
a9cd410a | 1285 | void __free_pages_core(struct page *page, unsigned int order) |
a226f6c8 | 1286 | { |
c3993076 | 1287 | unsigned int nr_pages = 1 << order; |
e2d0bd2b | 1288 | struct page *p = page; |
c3993076 | 1289 | unsigned int loop; |
a226f6c8 | 1290 | |
7fef431b DH |
1291 | /* |
1292 | * When initializing the memmap, __init_single_page() sets the refcount | |
1293 | * of all pages to 1 ("allocated"/"not free"). We have to set the | |
1294 | * refcount of all involved pages to 0. | |
1295 | */ | |
e2d0bd2b YL |
1296 | prefetchw(p); |
1297 | for (loop = 0; loop < (nr_pages - 1); loop++, p++) { | |
1298 | prefetchw(p + 1); | |
c3993076 JW |
1299 | __ClearPageReserved(p); |
1300 | set_page_count(p, 0); | |
a226f6c8 | 1301 | } |
e2d0bd2b YL |
1302 | __ClearPageReserved(p); |
1303 | set_page_count(p, 0); | |
c3993076 | 1304 | |
9705bea5 | 1305 | atomic_long_add(nr_pages, &page_zone(page)->managed_pages); |
7fef431b | 1306 | |
dcdfdd40 | 1307 | if (page_contains_unaccepted(page, order)) { |
5e0a760b | 1308 | if (order == MAX_PAGE_ORDER && __free_unaccepted(page)) |
dcdfdd40 KS |
1309 | return; |
1310 | ||
1311 | accept_page(page, order); | |
1312 | } | |
1313 | ||
7fef431b DH |
1314 | /* |
1315 | * Bypass PCP and place fresh pages right to the tail, primarily | |
1316 | * relevant for memory onlining. | |
1317 | */ | |
0a54864f | 1318 | __free_pages_ok(page, order, FPI_TO_TAIL); |
a226f6c8 DH |
1319 | } |
1320 | ||
7cf91a98 JK |
1321 | /* |
1322 | * Check that the whole (or subset of) a pageblock given by the interval of | |
1323 | * [start_pfn, end_pfn) is valid and within the same zone, before scanning it | |
859a85dd | 1324 | * with the migration of free compaction scanner. |
7cf91a98 JK |
1325 | * |
1326 | * Return struct page pointer of start_pfn, or NULL if checks were not passed. | |
1327 | * | |
1328 | * It's possible on some configurations to have a setup like node0 node1 node0 | |
1329 | * i.e. it's possible that all pages within a zones range of pages do not | |
1330 | * belong to a single zone. We assume that a border between node0 and node1 | |
1331 | * can occur within a single pageblock, but not a node0 node1 node0 | |
1332 | * interleaving within a single pageblock. It is therefore sufficient to check | |
1333 | * the first and last page of a pageblock and avoid checking each individual | |
1334 | * page in a pageblock. | |
65f67a3e BW |
1335 | * |
1336 | * Note: the function may return non-NULL struct page even for a page block | |
1337 | * which contains a memory hole (i.e. there is no physical memory for a subset | |
5e0a760b | 1338 | * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which |
65f67a3e BW |
1339 | * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole |
1340 | * even though the start pfn is online and valid. This should be safe most of | |
1341 | * the time because struct pages are still initialized via init_unavailable_range() | |
1342 | * and pfn walkers shouldn't touch any physical memory range for which they do | |
1343 | * not recognize any specific metadata in struct pages. | |
7cf91a98 JK |
1344 | */ |
1345 | struct page *__pageblock_pfn_to_page(unsigned long start_pfn, | |
1346 | unsigned long end_pfn, struct zone *zone) | |
1347 | { | |
1348 | struct page *start_page; | |
1349 | struct page *end_page; | |
1350 | ||
1351 | /* end_pfn is one past the range we are checking */ | |
1352 | end_pfn--; | |
1353 | ||
3c4322c9 | 1354 | if (!pfn_valid(end_pfn)) |
7cf91a98 JK |
1355 | return NULL; |
1356 | ||
2d070eab MH |
1357 | start_page = pfn_to_online_page(start_pfn); |
1358 | if (!start_page) | |
1359 | return NULL; | |
7cf91a98 JK |
1360 | |
1361 | if (page_zone(start_page) != zone) | |
1362 | return NULL; | |
1363 | ||
1364 | end_page = pfn_to_page(end_pfn); | |
1365 | ||
1366 | /* This gives a shorter code than deriving page_zone(end_page) */ | |
1367 | if (page_zone_id(start_page) != page_zone_id(end_page)) | |
1368 | return NULL; | |
1369 | ||
1370 | return start_page; | |
1371 | } | |
1372 | ||
2f47a91f | 1373 | /* |
9420f89d MRI |
1374 | * The order of subdivision here is critical for the IO subsystem. |
1375 | * Please do not alter this order without good reasons and regression | |
1376 | * testing. Specifically, as large blocks of memory are subdivided, | |
1377 | * the order in which smaller blocks are delivered depends on the order | |
1378 | * they're subdivided in this function. This is the primary factor | |
1379 | * influencing the order in which pages are delivered to the IO | |
1380 | * subsystem according to empirical testing, and this is also justified | |
1381 | * by considering the behavior of a buddy system containing a single | |
1382 | * large block of memory acted on by a series of small allocations. | |
1383 | * This behavior is a critical factor in sglist merging's success. | |
80b1f41c | 1384 | * |
9420f89d | 1385 | * -- nyc |
2f47a91f | 1386 | */ |
9420f89d MRI |
1387 | static inline void expand(struct zone *zone, struct page *page, |
1388 | int low, int high, int migratetype) | |
2f47a91f | 1389 | { |
9420f89d | 1390 | unsigned long size = 1 << high; |
2f47a91f | 1391 | |
9420f89d MRI |
1392 | while (high > low) { |
1393 | high--; | |
1394 | size >>= 1; | |
1395 | VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]); | |
2f47a91f | 1396 | |
9420f89d MRI |
1397 | /* |
1398 | * Mark as guard pages (or page), that will allow to | |
1399 | * merge back to allocator when buddy will be freed. | |
1400 | * Corresponding page table entries will not be touched, | |
1401 | * pages will stay not present in virtual address space | |
1402 | */ | |
1403 | if (set_page_guard(zone, &page[size], high, migratetype)) | |
2f47a91f | 1404 | continue; |
9420f89d MRI |
1405 | |
1406 | add_to_free_list(&page[size], zone, high, migratetype); | |
1407 | set_buddy_order(&page[size], high); | |
2f47a91f | 1408 | } |
2f47a91f PT |
1409 | } |
1410 | ||
9420f89d | 1411 | static void check_new_page_bad(struct page *page) |
0e56acae | 1412 | { |
9420f89d MRI |
1413 | if (unlikely(page->flags & __PG_HWPOISON)) { |
1414 | /* Don't complain about hwpoisoned pages */ | |
1415 | page_mapcount_reset(page); /* remove PageBuddy */ | |
1416 | return; | |
0e56acae AD |
1417 | } |
1418 | ||
9420f89d MRI |
1419 | bad_page(page, |
1420 | page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP)); | |
0e56acae AD |
1421 | } |
1422 | ||
1423 | /* | |
9420f89d | 1424 | * This page is about to be returned from the page allocator |
0e56acae | 1425 | */ |
77c7a095 | 1426 | static bool check_new_page(struct page *page) |
0e56acae | 1427 | { |
9420f89d MRI |
1428 | if (likely(page_expected_state(page, |
1429 | PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON))) | |
77c7a095 | 1430 | return false; |
0e56acae | 1431 | |
9420f89d | 1432 | check_new_page_bad(page); |
77c7a095 | 1433 | return true; |
9420f89d | 1434 | } |
0e56acae | 1435 | |
9420f89d MRI |
1436 | static inline bool check_new_pages(struct page *page, unsigned int order) |
1437 | { | |
1438 | if (is_check_pages_enabled()) { | |
1439 | for (int i = 0; i < (1 << order); i++) { | |
1440 | struct page *p = page + i; | |
0e56acae | 1441 | |
8666925c | 1442 | if (check_new_page(p)) |
9420f89d | 1443 | return true; |
0e56acae AD |
1444 | } |
1445 | } | |
1446 | ||
9420f89d | 1447 | return false; |
0e56acae AD |
1448 | } |
1449 | ||
9420f89d | 1450 | static inline bool should_skip_kasan_unpoison(gfp_t flags) |
e4443149 | 1451 | { |
9420f89d MRI |
1452 | /* Don't skip if a software KASAN mode is enabled. */ |
1453 | if (IS_ENABLED(CONFIG_KASAN_GENERIC) || | |
1454 | IS_ENABLED(CONFIG_KASAN_SW_TAGS)) | |
1455 | return false; | |
e4443149 | 1456 | |
9420f89d MRI |
1457 | /* Skip, if hardware tag-based KASAN is not enabled. */ |
1458 | if (!kasan_hw_tags_enabled()) | |
1459 | return true; | |
e4443149 DJ |
1460 | |
1461 | /* | |
9420f89d MRI |
1462 | * With hardware tag-based KASAN enabled, skip if this has been |
1463 | * requested via __GFP_SKIP_KASAN. | |
e4443149 | 1464 | */ |
9420f89d | 1465 | return flags & __GFP_SKIP_KASAN; |
e4443149 DJ |
1466 | } |
1467 | ||
9420f89d | 1468 | static inline bool should_skip_init(gfp_t flags) |
ecd09650 | 1469 | { |
9420f89d MRI |
1470 | /* Don't skip, if hardware tag-based KASAN is not enabled. */ |
1471 | if (!kasan_hw_tags_enabled()) | |
1472 | return false; | |
1473 | ||
1474 | /* For hardware tag-based KASAN, skip if requested. */ | |
1475 | return (flags & __GFP_SKIP_ZERO); | |
ecd09650 DJ |
1476 | } |
1477 | ||
9420f89d MRI |
1478 | inline void post_alloc_hook(struct page *page, unsigned int order, |
1479 | gfp_t gfp_flags) | |
7e18adb4 | 1480 | { |
9420f89d MRI |
1481 | bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) && |
1482 | !should_skip_init(gfp_flags); | |
1483 | bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS); | |
1484 | int i; | |
1485 | ||
1486 | set_page_private(page, 0); | |
1487 | set_page_refcounted(page); | |
0e1cc95b | 1488 | |
9420f89d MRI |
1489 | arch_alloc_page(page, order); |
1490 | debug_pagealloc_map_pages(page, 1 << order); | |
7e18adb4 | 1491 | |
3d060856 | 1492 | /* |
9420f89d MRI |
1493 | * Page unpoisoning must happen before memory initialization. |
1494 | * Otherwise, the poison pattern will be overwritten for __GFP_ZERO | |
1495 | * allocations and the page unpoisoning code will complain. | |
3d060856 | 1496 | */ |
9420f89d | 1497 | kernel_unpoison_pages(page, 1 << order); |
862b6dee | 1498 | |
1bb5eab3 AK |
1499 | /* |
1500 | * As memory initialization might be integrated into KASAN, | |
b42090ae | 1501 | * KASAN unpoisoning and memory initializion code must be |
1bb5eab3 AK |
1502 | * kept together to avoid discrepancies in behavior. |
1503 | */ | |
9294b128 AK |
1504 | |
1505 | /* | |
44383cef AK |
1506 | * If memory tags should be zeroed |
1507 | * (which happens only when memory should be initialized as well). | |
9294b128 | 1508 | */ |
44383cef | 1509 | if (zero_tags) { |
420ef683 | 1510 | /* Initialize both memory and memory tags. */ |
9294b128 AK |
1511 | for (i = 0; i != 1 << order; ++i) |
1512 | tag_clear_highpage(page + i); | |
1513 | ||
44383cef | 1514 | /* Take note that memory was initialized by the loop above. */ |
9294b128 AK |
1515 | init = false; |
1516 | } | |
0a54864f PC |
1517 | if (!should_skip_kasan_unpoison(gfp_flags) && |
1518 | kasan_unpoison_pages(page, order, init)) { | |
1519 | /* Take note that memory was initialized by KASAN. */ | |
1520 | if (kasan_has_integrated_init()) | |
1521 | init = false; | |
1522 | } else { | |
1523 | /* | |
1524 | * If memory tags have not been set by KASAN, reset the page | |
1525 | * tags to ensure page_address() dereferencing does not fault. | |
1526 | */ | |
70c248ac CM |
1527 | for (i = 0; i != 1 << order; ++i) |
1528 | page_kasan_tag_reset(page + i); | |
7a3b8353 | 1529 | } |
44383cef | 1530 | /* If memory is still not initialized, initialize it now. */ |
7e3cbba6 | 1531 | if (init) |
aeaec8e2 | 1532 | kernel_init_pages(page, 1 << order); |
1bb5eab3 AK |
1533 | |
1534 | set_page_owner(page, order, gfp_flags); | |
df4e817b | 1535 | page_table_check_alloc(page, order); |
46f24fd8 JK |
1536 | } |
1537 | ||
479f854a | 1538 | static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags, |
c603844b | 1539 | unsigned int alloc_flags) |
2a7684a2 | 1540 | { |
46f24fd8 | 1541 | post_alloc_hook(page, order, gfp_flags); |
17cf4406 | 1542 | |
17cf4406 NP |
1543 | if (order && (gfp_flags & __GFP_COMP)) |
1544 | prep_compound_page(page, order); | |
1545 | ||
75379191 | 1546 | /* |
2f064f34 | 1547 | * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to |
75379191 VB |
1548 | * allocate the page. The expectation is that the caller is taking |
1549 | * steps that will free more memory. The caller should avoid the page | |
1550 | * being used for !PFMEMALLOC purposes. | |
1551 | */ | |
2f064f34 MH |
1552 | if (alloc_flags & ALLOC_NO_WATERMARKS) |
1553 | set_page_pfmemalloc(page); | |
1554 | else | |
1555 | clear_page_pfmemalloc(page); | |
1da177e4 LT |
1556 | } |
1557 | ||
56fd56b8 MG |
1558 | /* |
1559 | * Go through the free lists for the given migratetype and remove | |
1560 | * the smallest available page from the freelists | |
1561 | */ | |
85ccc8fa | 1562 | static __always_inline |
728ec980 | 1563 | struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, |
56fd56b8 MG |
1564 | int migratetype) |
1565 | { | |
1566 | unsigned int current_order; | |
b8af2941 | 1567 | struct free_area *area; |
56fd56b8 MG |
1568 | struct page *page; |
1569 | ||
1570 | /* Find a page of the appropriate size in the preferred list */ | |
fd377218 | 1571 | for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) { |
56fd56b8 | 1572 | area = &(zone->free_area[current_order]); |
b03641af | 1573 | page = get_page_from_free_area(area, migratetype); |
a16601c5 GT |
1574 | if (!page) |
1575 | continue; | |
6ab01363 AD |
1576 | del_page_from_free_list(page, zone, current_order); |
1577 | expand(zone, page, order, current_order, migratetype); | |
bb14c2c7 | 1578 | set_pcppage_migratetype(page, migratetype); |
10e0f753 WY |
1579 | trace_mm_page_alloc_zone_locked(page, order, migratetype, |
1580 | pcp_allowed_order(order) && | |
1581 | migratetype < MIGRATE_PCPTYPES); | |
56fd56b8 MG |
1582 | return page; |
1583 | } | |
1584 | ||
1585 | return NULL; | |
1586 | } | |
1587 | ||
1588 | ||
b2a0ac88 MG |
1589 | /* |
1590 | * This array describes the order lists are fallen back to when | |
1591 | * the free lists for the desirable migrate type are depleted | |
1dd214b8 ZY |
1592 | * |
1593 | * The other migratetypes do not have fallbacks. | |
b2a0ac88 | 1594 | */ |
aa02d3c1 YD |
1595 | static int fallbacks[MIGRATE_TYPES][MIGRATE_PCPTYPES - 1] = { |
1596 | [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE }, | |
1597 | [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE }, | |
1598 | [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE }, | |
b2a0ac88 MG |
1599 | }; |
1600 | ||
dc67647b | 1601 | #ifdef CONFIG_CMA |
85ccc8fa | 1602 | static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone, |
dc67647b JK |
1603 | unsigned int order) |
1604 | { | |
1605 | return __rmqueue_smallest(zone, order, MIGRATE_CMA); | |
1606 | } | |
1607 | #else | |
1608 | static inline struct page *__rmqueue_cma_fallback(struct zone *zone, | |
1609 | unsigned int order) { return NULL; } | |
1610 | #endif | |
1611 | ||
c361be55 | 1612 | /* |
293ffa5e | 1613 | * Move the free pages in a range to the freelist tail of the requested type. |
d9c23400 | 1614 | * Note that start_page and end_pages are not aligned on a pageblock |
c361be55 MG |
1615 | * boundary. If alignment is required, use move_freepages_block() |
1616 | */ | |
02aa0cdd | 1617 | static int move_freepages(struct zone *zone, |
39ddb991 | 1618 | unsigned long start_pfn, unsigned long end_pfn, |
02aa0cdd | 1619 | int migratetype, int *num_movable) |
c361be55 MG |
1620 | { |
1621 | struct page *page; | |
39ddb991 | 1622 | unsigned long pfn; |
d00181b9 | 1623 | unsigned int order; |
d100313f | 1624 | int pages_moved = 0; |
c361be55 | 1625 | |
39ddb991 | 1626 | for (pfn = start_pfn; pfn <= end_pfn;) { |
39ddb991 | 1627 | page = pfn_to_page(pfn); |
c361be55 | 1628 | if (!PageBuddy(page)) { |
02aa0cdd VB |
1629 | /* |
1630 | * We assume that pages that could be isolated for | |
1631 | * migration are movable. But we don't actually try | |
1632 | * isolating, as that would be expensive. | |
1633 | */ | |
1634 | if (num_movable && | |
1635 | (PageLRU(page) || __PageMovable(page))) | |
1636 | (*num_movable)++; | |
39ddb991 | 1637 | pfn++; |
c361be55 MG |
1638 | continue; |
1639 | } | |
1640 | ||
cd961038 DR |
1641 | /* Make sure we are not inadvertently changing nodes */ |
1642 | VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page); | |
1643 | VM_BUG_ON_PAGE(page_zone(page) != zone, page); | |
1644 | ||
ab130f91 | 1645 | order = buddy_order(page); |
6ab01363 | 1646 | move_to_free_list(page, zone, order, migratetype); |
39ddb991 | 1647 | pfn += 1 << order; |
d100313f | 1648 | pages_moved += 1 << order; |
c361be55 MG |
1649 | } |
1650 | ||
d100313f | 1651 | return pages_moved; |
c361be55 MG |
1652 | } |
1653 | ||
ee6f509c | 1654 | int move_freepages_block(struct zone *zone, struct page *page, |
02aa0cdd | 1655 | int migratetype, int *num_movable) |
c361be55 | 1656 | { |
39ddb991 | 1657 | unsigned long start_pfn, end_pfn, pfn; |
c361be55 | 1658 | |
4a222127 DR |
1659 | if (num_movable) |
1660 | *num_movable = 0; | |
1661 | ||
39ddb991 | 1662 | pfn = page_to_pfn(page); |
4f9bc69a KW |
1663 | start_pfn = pageblock_start_pfn(pfn); |
1664 | end_pfn = pageblock_end_pfn(pfn) - 1; | |
c361be55 MG |
1665 | |
1666 | /* Do not cross zone boundaries */ | |
108bcc96 | 1667 | if (!zone_spans_pfn(zone, start_pfn)) |
39ddb991 | 1668 | start_pfn = pfn; |
108bcc96 | 1669 | if (!zone_spans_pfn(zone, end_pfn)) |
c361be55 MG |
1670 | return 0; |
1671 | ||
39ddb991 | 1672 | return move_freepages(zone, start_pfn, end_pfn, migratetype, |
02aa0cdd | 1673 | num_movable); |
c361be55 MG |
1674 | } |
1675 | ||
2f66a68f MG |
1676 | static void change_pageblock_range(struct page *pageblock_page, |
1677 | int start_order, int migratetype) | |
1678 | { | |
1679 | int nr_pageblocks = 1 << (start_order - pageblock_order); | |
1680 | ||
1681 | while (nr_pageblocks--) { | |
1682 | set_pageblock_migratetype(pageblock_page, migratetype); | |
1683 | pageblock_page += pageblock_nr_pages; | |
1684 | } | |
1685 | } | |
1686 | ||
fef903ef | 1687 | /* |
9c0415eb VB |
1688 | * When we are falling back to another migratetype during allocation, try to |
1689 | * steal extra free pages from the same pageblocks to satisfy further | |
1690 | * allocations, instead of polluting multiple pageblocks. | |
1691 | * | |
1692 | * If we are stealing a relatively large buddy page, it is likely there will | |
1693 | * be more free pages in the pageblock, so try to steal them all. For | |
1694 | * reclaimable and unmovable allocations, we steal regardless of page size, | |
1695 | * as fragmentation caused by those allocations polluting movable pageblocks | |
1696 | * is worse than movable allocations stealing from unmovable and reclaimable | |
1697 | * pageblocks. | |
fef903ef | 1698 | */ |
4eb7dce6 JK |
1699 | static bool can_steal_fallback(unsigned int order, int start_mt) |
1700 | { | |
1701 | /* | |
1702 | * Leaving this order check is intended, although there is | |
1703 | * relaxed order check in next check. The reason is that | |
1704 | * we can actually steal whole pageblock if this condition met, | |
1705 | * but, below check doesn't guarantee it and that is just heuristic | |
1706 | * so could be changed anytime. | |
1707 | */ | |
1708 | if (order >= pageblock_order) | |
1709 | return true; | |
1710 | ||
1711 | if (order >= pageblock_order / 2 || | |
1712 | start_mt == MIGRATE_RECLAIMABLE || | |
1713 | start_mt == MIGRATE_UNMOVABLE || | |
1714 | page_group_by_mobility_disabled) | |
1715 | return true; | |
1716 | ||
1717 | return false; | |
1718 | } | |
1719 | ||
597c8920 | 1720 | static inline bool boost_watermark(struct zone *zone) |
1c30844d MG |
1721 | { |
1722 | unsigned long max_boost; | |
1723 | ||
1724 | if (!watermark_boost_factor) | |
597c8920 | 1725 | return false; |
14f69140 HW |
1726 | /* |
1727 | * Don't bother in zones that are unlikely to produce results. | |
1728 | * On small machines, including kdump capture kernels running | |
1729 | * in a small area, boosting the watermark can cause an out of | |
1730 | * memory situation immediately. | |
1731 | */ | |
1732 | if ((pageblock_nr_pages * 4) > zone_managed_pages(zone)) | |
597c8920 | 1733 | return false; |
1c30844d MG |
1734 | |
1735 | max_boost = mult_frac(zone->_watermark[WMARK_HIGH], | |
1736 | watermark_boost_factor, 10000); | |
94b3334c MG |
1737 | |
1738 | /* | |
1739 | * high watermark may be uninitialised if fragmentation occurs | |
1740 | * very early in boot so do not boost. We do not fall | |
1741 | * through and boost by pageblock_nr_pages as failing | |
1742 | * allocations that early means that reclaim is not going | |
1743 | * to help and it may even be impossible to reclaim the | |
1744 | * boosted watermark resulting in a hang. | |
1745 | */ | |
1746 | if (!max_boost) | |
597c8920 | 1747 | return false; |
94b3334c | 1748 | |
1c30844d MG |
1749 | max_boost = max(pageblock_nr_pages, max_boost); |
1750 | ||
1751 | zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages, | |
1752 | max_boost); | |
597c8920 JW |
1753 | |
1754 | return true; | |
1c30844d MG |
1755 | } |
1756 | ||
4eb7dce6 JK |
1757 | /* |
1758 | * This function implements actual steal behaviour. If order is large enough, | |
1759 | * we can steal whole pageblock. If not, we first move freepages in this | |
02aa0cdd VB |
1760 | * pageblock to our migratetype and determine how many already-allocated pages |
1761 | * are there in the pageblock with a compatible migratetype. If at least half | |
1762 | * of pages are free or compatible, we can change migratetype of the pageblock | |
1763 | * itself, so pages freed in the future will be put on the correct free list. | |
4eb7dce6 JK |
1764 | */ |
1765 | static void steal_suitable_fallback(struct zone *zone, struct page *page, | |
1c30844d | 1766 | unsigned int alloc_flags, int start_type, bool whole_block) |
fef903ef | 1767 | { |
ab130f91 | 1768 | unsigned int current_order = buddy_order(page); |
02aa0cdd VB |
1769 | int free_pages, movable_pages, alike_pages; |
1770 | int old_block_type; | |
1771 | ||
1772 | old_block_type = get_pageblock_migratetype(page); | |
fef903ef | 1773 | |
3bc48f96 VB |
1774 | /* |
1775 | * This can happen due to races and we want to prevent broken | |
1776 | * highatomic accounting. | |
1777 | */ | |
02aa0cdd | 1778 | if (is_migrate_highatomic(old_block_type)) |
3bc48f96 VB |
1779 | goto single_page; |
1780 | ||
fef903ef SB |
1781 | /* Take ownership for orders >= pageblock_order */ |
1782 | if (current_order >= pageblock_order) { | |
1783 | change_pageblock_range(page, current_order, start_type); | |
3bc48f96 | 1784 | goto single_page; |
fef903ef SB |
1785 | } |
1786 | ||
1c30844d MG |
1787 | /* |
1788 | * Boost watermarks to increase reclaim pressure to reduce the | |
1789 | * likelihood of future fallbacks. Wake kswapd now as the node | |
1790 | * may be balanced overall and kswapd will not wake naturally. | |
1791 | */ | |
597c8920 | 1792 | if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD)) |
73444bc4 | 1793 | set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags); |
1c30844d | 1794 | |
3bc48f96 VB |
1795 | /* We are not allowed to try stealing from the whole block */ |
1796 | if (!whole_block) | |
1797 | goto single_page; | |
1798 | ||
02aa0cdd VB |
1799 | free_pages = move_freepages_block(zone, page, start_type, |
1800 | &movable_pages); | |
ebddd111 ML |
1801 | /* moving whole block can fail due to zone boundary conditions */ |
1802 | if (!free_pages) | |
1803 | goto single_page; | |
1804 | ||
02aa0cdd VB |
1805 | /* |
1806 | * Determine how many pages are compatible with our allocation. | |
1807 | * For movable allocation, it's the number of movable pages which | |
1808 | * we just obtained. For other types it's a bit more tricky. | |
1809 | */ | |
1810 | if (start_type == MIGRATE_MOVABLE) { | |
1811 | alike_pages = movable_pages; | |
1812 | } else { | |
1813 | /* | |
1814 | * If we are falling back a RECLAIMABLE or UNMOVABLE allocation | |
1815 | * to MOVABLE pageblock, consider all non-movable pages as | |
1816 | * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or | |
1817 | * vice versa, be conservative since we can't distinguish the | |
1818 | * exact migratetype of non-movable pages. | |
1819 | */ | |
1820 | if (old_block_type == MIGRATE_MOVABLE) | |
1821 | alike_pages = pageblock_nr_pages | |
1822 | - (free_pages + movable_pages); | |
1823 | else | |
1824 | alike_pages = 0; | |
1825 | } | |
02aa0cdd VB |
1826 | /* |
1827 | * If a sufficient number of pages in the block are either free or of | |
ebddd111 | 1828 | * compatible migratability as our allocation, claim the whole block. |
02aa0cdd VB |
1829 | */ |
1830 | if (free_pages + alike_pages >= (1 << (pageblock_order-1)) || | |
4eb7dce6 JK |
1831 | page_group_by_mobility_disabled) |
1832 | set_pageblock_migratetype(page, start_type); | |
3bc48f96 VB |
1833 | |
1834 | return; | |
1835 | ||
1836 | single_page: | |
6ab01363 | 1837 | move_to_free_list(page, zone, current_order, start_type); |
4eb7dce6 JK |
1838 | } |
1839 | ||
2149cdae JK |
1840 | /* |
1841 | * Check whether there is a suitable fallback freepage with requested order. | |
1842 | * If only_stealable is true, this function returns fallback_mt only if | |
1843 | * we can steal other freepages all together. This would help to reduce | |
1844 | * fragmentation due to mixed migratetype pages in one pageblock. | |
1845 | */ | |
1846 | int find_suitable_fallback(struct free_area *area, unsigned int order, | |
1847 | int migratetype, bool only_stealable, bool *can_steal) | |
4eb7dce6 JK |
1848 | { |
1849 | int i; | |
1850 | int fallback_mt; | |
1851 | ||
1852 | if (area->nr_free == 0) | |
1853 | return -1; | |
1854 | ||
1855 | *can_steal = false; | |
aa02d3c1 | 1856 | for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) { |
4eb7dce6 | 1857 | fallback_mt = fallbacks[migratetype][i]; |
b03641af | 1858 | if (free_area_empty(area, fallback_mt)) |
4eb7dce6 | 1859 | continue; |
fef903ef | 1860 | |
4eb7dce6 JK |
1861 | if (can_steal_fallback(order, migratetype)) |
1862 | *can_steal = true; | |
1863 | ||
2149cdae JK |
1864 | if (!only_stealable) |
1865 | return fallback_mt; | |
1866 | ||
1867 | if (*can_steal) | |
1868 | return fallback_mt; | |
fef903ef | 1869 | } |
4eb7dce6 JK |
1870 | |
1871 | return -1; | |
fef903ef SB |
1872 | } |
1873 | ||
0aaa29a5 MG |
1874 | /* |
1875 | * Reserve a pageblock for exclusive use of high-order atomic allocations if | |
1876 | * there are no empty page blocks that contain a page with a suitable order | |
1877 | */ | |
368d983b | 1878 | static void reserve_highatomic_pageblock(struct page *page, struct zone *zone) |
0aaa29a5 MG |
1879 | { |
1880 | int mt; | |
1881 | unsigned long max_managed, flags; | |
1882 | ||
1883 | /* | |
d68e39fc | 1884 | * The number reserved as: minimum is 1 pageblock, maximum is |
9cd20f3f CTK |
1885 | * roughly 1% of a zone. But if 1% of a zone falls below a |
1886 | * pageblock size, then don't reserve any pageblocks. | |
0aaa29a5 MG |
1887 | * Check is race-prone but harmless. |
1888 | */ | |
9cd20f3f CTK |
1889 | if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages) |
1890 | return; | |
d68e39fc | 1891 | max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages); |
0aaa29a5 MG |
1892 | if (zone->nr_reserved_highatomic >= max_managed) |
1893 | return; | |
1894 | ||
1895 | spin_lock_irqsave(&zone->lock, flags); | |
1896 | ||
1897 | /* Recheck the nr_reserved_highatomic limit under the lock */ | |
1898 | if (zone->nr_reserved_highatomic >= max_managed) | |
1899 | goto out_unlock; | |
1900 | ||
1901 | /* Yoink! */ | |
1902 | mt = get_pageblock_migratetype(page); | |
1dd214b8 ZY |
1903 | /* Only reserve normal pageblocks (i.e., they can merge with others) */ |
1904 | if (migratetype_is_mergeable(mt)) { | |
0aaa29a5 MG |
1905 | zone->nr_reserved_highatomic += pageblock_nr_pages; |
1906 | set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC); | |
02aa0cdd | 1907 | move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL); |
0aaa29a5 MG |
1908 | } |
1909 | ||
1910 | out_unlock: | |
1911 | spin_unlock_irqrestore(&zone->lock, flags); | |
1912 | } | |
1913 | ||
1914 | /* | |
1915 | * Used when an allocation is about to fail under memory pressure. This | |
1916 | * potentially hurts the reliability of high-order allocations when under | |
1917 | * intense memory pressure but failed atomic allocations should be easier | |
1918 | * to recover from than an OOM. | |
29fac03b MK |
1919 | * |
1920 | * If @force is true, try to unreserve a pageblock even though highatomic | |
1921 | * pageblock is exhausted. | |
0aaa29a5 | 1922 | */ |
29fac03b MK |
1923 | static bool unreserve_highatomic_pageblock(const struct alloc_context *ac, |
1924 | bool force) | |
0aaa29a5 MG |
1925 | { |
1926 | struct zonelist *zonelist = ac->zonelist; | |
1927 | unsigned long flags; | |
1928 | struct zoneref *z; | |
1929 | struct zone *zone; | |
1930 | struct page *page; | |
1931 | int order; | |
04c8716f | 1932 | bool ret; |
0aaa29a5 | 1933 | |
97a225e6 | 1934 | for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx, |
0aaa29a5 | 1935 | ac->nodemask) { |
29fac03b MK |
1936 | /* |
1937 | * Preserve at least one pageblock unless memory pressure | |
1938 | * is really high. | |
1939 | */ | |
1940 | if (!force && zone->nr_reserved_highatomic <= | |
1941 | pageblock_nr_pages) | |
0aaa29a5 MG |
1942 | continue; |
1943 | ||
1944 | spin_lock_irqsave(&zone->lock, flags); | |
fd377218 | 1945 | for (order = 0; order < NR_PAGE_ORDERS; order++) { |
0aaa29a5 MG |
1946 | struct free_area *area = &(zone->free_area[order]); |
1947 | ||
b03641af | 1948 | page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC); |
a16601c5 | 1949 | if (!page) |
0aaa29a5 MG |
1950 | continue; |
1951 | ||
0aaa29a5 | 1952 | /* |
4855e4a7 MK |
1953 | * In page freeing path, migratetype change is racy so |
1954 | * we can counter several free pages in a pageblock | |
f0953a1b | 1955 | * in this loop although we changed the pageblock type |
4855e4a7 MK |
1956 | * from highatomic to ac->migratetype. So we should |
1957 | * adjust the count once. | |
0aaa29a5 | 1958 | */ |
a6ffdc07 | 1959 | if (is_migrate_highatomic_page(page)) { |
4855e4a7 MK |
1960 | /* |
1961 | * It should never happen but changes to | |
1962 | * locking could inadvertently allow a per-cpu | |
1963 | * drain to add pages to MIGRATE_HIGHATOMIC | |
1964 | * while unreserving so be safe and watch for | |
1965 | * underflows. | |
1966 | */ | |
1967 | zone->nr_reserved_highatomic -= min( | |
1968 | pageblock_nr_pages, | |
1969 | zone->nr_reserved_highatomic); | |
1970 | } | |
0aaa29a5 MG |
1971 | |
1972 | /* | |
1973 | * Convert to ac->migratetype and avoid the normal | |
1974 | * pageblock stealing heuristics. Minimally, the caller | |
1975 | * is doing the work and needs the pages. More | |
1976 | * importantly, if the block was always converted to | |
1977 | * MIGRATE_UNMOVABLE or another type then the number | |
1978 | * of pageblocks that cannot be completely freed | |
1979 | * may increase. | |
1980 | */ | |
1981 | set_pageblock_migratetype(page, ac->migratetype); | |
02aa0cdd VB |
1982 | ret = move_freepages_block(zone, page, ac->migratetype, |
1983 | NULL); | |
29fac03b MK |
1984 | if (ret) { |
1985 | spin_unlock_irqrestore(&zone->lock, flags); | |
1986 | return ret; | |
1987 | } | |
0aaa29a5 MG |
1988 | } |
1989 | spin_unlock_irqrestore(&zone->lock, flags); | |
1990 | } | |
04c8716f MK |
1991 | |
1992 | return false; | |
0aaa29a5 MG |
1993 | } |
1994 | ||
3bc48f96 VB |
1995 | /* |
1996 | * Try finding a free buddy page on the fallback list and put it on the free | |
1997 | * list of requested migratetype, possibly along with other pages from the same | |
1998 | * block, depending on fragmentation avoidance heuristics. Returns true if | |
1999 | * fallback was found so that __rmqueue_smallest() can grab it. | |
b002529d RV |
2000 | * |
2001 | * The use of signed ints for order and current_order is a deliberate | |
2002 | * deviation from the rest of this file, to make the for loop | |
2003 | * condition simpler. | |
3bc48f96 | 2004 | */ |
85ccc8fa | 2005 | static __always_inline bool |
6bb15450 MG |
2006 | __rmqueue_fallback(struct zone *zone, int order, int start_migratetype, |
2007 | unsigned int alloc_flags) | |
b2a0ac88 | 2008 | { |
b8af2941 | 2009 | struct free_area *area; |
b002529d | 2010 | int current_order; |
6bb15450 | 2011 | int min_order = order; |
b2a0ac88 | 2012 | struct page *page; |
4eb7dce6 JK |
2013 | int fallback_mt; |
2014 | bool can_steal; | |
b2a0ac88 | 2015 | |
6bb15450 MG |
2016 | /* |
2017 | * Do not steal pages from freelists belonging to other pageblocks | |
2018 | * i.e. orders < pageblock_order. If there are no local zones free, | |
2019 | * the zonelists will be reiterated without ALLOC_NOFRAGMENT. | |
2020 | */ | |
e933dc4a | 2021 | if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT) |
6bb15450 MG |
2022 | min_order = pageblock_order; |
2023 | ||
7a8f58f3 VB |
2024 | /* |
2025 | * Find the largest available free page in the other list. This roughly | |
2026 | * approximates finding the pageblock with the most free pages, which | |
2027 | * would be too costly to do exactly. | |
2028 | */ | |
5e0a760b | 2029 | for (current_order = MAX_PAGE_ORDER; current_order >= min_order; |
7aeb09f9 | 2030 | --current_order) { |
4eb7dce6 JK |
2031 | area = &(zone->free_area[current_order]); |
2032 | fallback_mt = find_suitable_fallback(area, current_order, | |
2149cdae | 2033 | start_migratetype, false, &can_steal); |
4eb7dce6 JK |
2034 | if (fallback_mt == -1) |
2035 | continue; | |
b2a0ac88 | 2036 | |
7a8f58f3 VB |
2037 | /* |
2038 | * We cannot steal all free pages from the pageblock and the | |
2039 | * requested migratetype is movable. In that case it's better to | |
2040 | * steal and split the smallest available page instead of the | |
2041 | * largest available page, because even if the next movable | |
2042 | * allocation falls back into a different pageblock than this | |
2043 | * one, it won't cause permanent fragmentation. | |
2044 | */ | |
2045 | if (!can_steal && start_migratetype == MIGRATE_MOVABLE | |
2046 | && current_order > order) | |
2047 | goto find_smallest; | |
b2a0ac88 | 2048 | |
7a8f58f3 VB |
2049 | goto do_steal; |
2050 | } | |
e0fff1bd | 2051 | |
7a8f58f3 | 2052 | return false; |
e0fff1bd | 2053 | |
7a8f58f3 | 2054 | find_smallest: |
fd377218 | 2055 | for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) { |
7a8f58f3 VB |
2056 | area = &(zone->free_area[current_order]); |
2057 | fallback_mt = find_suitable_fallback(area, current_order, | |
2058 | start_migratetype, false, &can_steal); | |
2059 | if (fallback_mt != -1) | |
2060 | break; | |
b2a0ac88 MG |
2061 | } |
2062 | ||
7a8f58f3 VB |
2063 | /* |
2064 | * This should not happen - we already found a suitable fallback | |
2065 | * when looking for the largest page. | |
2066 | */ | |
5e0a760b | 2067 | VM_BUG_ON(current_order > MAX_PAGE_ORDER); |
7a8f58f3 VB |
2068 | |
2069 | do_steal: | |
b03641af | 2070 | page = get_page_from_free_area(area, fallback_mt); |
7a8f58f3 | 2071 | |
1c30844d MG |
2072 | steal_suitable_fallback(zone, page, alloc_flags, start_migratetype, |
2073 | can_steal); | |
7a8f58f3 VB |
2074 | |
2075 | trace_mm_page_alloc_extfrag(page, order, current_order, | |
2076 | start_migratetype, fallback_mt); | |
2077 | ||
2078 | return true; | |
2079 | ||
b2a0ac88 MG |
2080 | } |
2081 | ||
56fd56b8 | 2082 | /* |
1da177e4 LT |
2083 | * Do the hard work of removing an element from the buddy allocator. |
2084 | * Call me with the zone->lock already held. | |
2085 | */ | |
85ccc8fa | 2086 | static __always_inline struct page * |
6bb15450 MG |
2087 | __rmqueue(struct zone *zone, unsigned int order, int migratetype, |
2088 | unsigned int alloc_flags) | |
1da177e4 | 2089 | { |
1da177e4 LT |
2090 | struct page *page; |
2091 | ||
ce8f86ee H |
2092 | if (IS_ENABLED(CONFIG_CMA)) { |
2093 | /* | |
2094 | * Balance movable allocations between regular and CMA areas by | |
2095 | * allocating from CMA when over half of the zone's free memory | |
2096 | * is in the CMA area. | |
2097 | */ | |
2098 | if (alloc_flags & ALLOC_CMA && | |
2099 | zone_page_state(zone, NR_FREE_CMA_PAGES) > | |
2100 | zone_page_state(zone, NR_FREE_PAGES) / 2) { | |
2101 | page = __rmqueue_cma_fallback(zone, order); | |
2102 | if (page) | |
10e0f753 | 2103 | return page; |
ce8f86ee | 2104 | } |
16867664 | 2105 | } |
3bc48f96 | 2106 | retry: |
56fd56b8 | 2107 | page = __rmqueue_smallest(zone, order, migratetype); |
974a786e | 2108 | if (unlikely(!page)) { |
8510e69c | 2109 | if (alloc_flags & ALLOC_CMA) |
dc67647b JK |
2110 | page = __rmqueue_cma_fallback(zone, order); |
2111 | ||
6bb15450 MG |
2112 | if (!page && __rmqueue_fallback(zone, order, migratetype, |
2113 | alloc_flags)) | |
3bc48f96 | 2114 | goto retry; |
728ec980 | 2115 | } |
b2a0ac88 | 2116 | return page; |
1da177e4 LT |
2117 | } |
2118 | ||
5f63b720 | 2119 | /* |
1da177e4 LT |
2120 | * Obtain a specified number of elements from the buddy allocator, all under |
2121 | * a single hold of the lock, for efficiency. Add them to the supplied list. | |
2122 | * Returns the number of new pages which were placed at *list. | |
2123 | */ | |
5f63b720 | 2124 | static int rmqueue_bulk(struct zone *zone, unsigned int order, |
b2a0ac88 | 2125 | unsigned long count, struct list_head *list, |
6bb15450 | 2126 | int migratetype, unsigned int alloc_flags) |
1da177e4 | 2127 | { |
57490774 | 2128 | unsigned long flags; |
700d2e9a | 2129 | int i; |
5f63b720 | 2130 | |
57490774 | 2131 | spin_lock_irqsave(&zone->lock, flags); |
1da177e4 | 2132 | for (i = 0; i < count; ++i) { |
6bb15450 MG |
2133 | struct page *page = __rmqueue(zone, order, migratetype, |
2134 | alloc_flags); | |
085cc7d5 | 2135 | if (unlikely(page == NULL)) |
1da177e4 | 2136 | break; |
81eabcbe MG |
2137 | |
2138 | /* | |
0fac3ba5 VB |
2139 | * Split buddy pages returned by expand() are received here in |
2140 | * physical page order. The page is added to the tail of | |
2141 | * caller's list. From the callers perspective, the linked list | |
2142 | * is ordered by page number under some conditions. This is | |
2143 | * useful for IO devices that can forward direction from the | |
2144 | * head, thus also in the physical page order. This is useful | |
2145 | * for IO devices that can merge IO requests if the physical | |
2146 | * pages are ordered properly. | |
81eabcbe | 2147 | */ |
bf75f200 | 2148 | list_add_tail(&page->pcp_list, list); |
bb14c2c7 | 2149 | if (is_migrate_cma(get_pcppage_migratetype(page))) |
d1ce749a BZ |
2150 | __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, |
2151 | -(1 << order)); | |
1da177e4 | 2152 | } |
a6de734b | 2153 | |
f2260e6b | 2154 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); |
57490774 | 2155 | spin_unlock_irqrestore(&zone->lock, flags); |
2ede3c13 | 2156 | |
700d2e9a | 2157 | return i; |
1da177e4 LT |
2158 | } |
2159 | ||
51a755c5 HY |
2160 | /* |
2161 | * Called from the vmstat counter updater to decay the PCP high. | |
2162 | * Return whether there are addition works to do. | |
2163 | */ | |
2164 | int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp) | |
2165 | { | |
2166 | int high_min, to_drain, batch; | |
2167 | int todo = 0; | |
2168 | ||
2169 | high_min = READ_ONCE(pcp->high_min); | |
2170 | batch = READ_ONCE(pcp->batch); | |
2171 | /* | |
2172 | * Decrease pcp->high periodically to try to free possible | |
2173 | * idle PCP pages. And, avoid to free too many pages to | |
2174 | * control latency. This caps pcp->high decrement too. | |
2175 | */ | |
2176 | if (pcp->high > high_min) { | |
2177 | pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX), | |
2178 | pcp->high - (pcp->high >> 3), high_min); | |
2179 | if (pcp->high > high_min) | |
2180 | todo++; | |
2181 | } | |
2182 | ||
2183 | to_drain = pcp->count - pcp->high; | |
2184 | if (to_drain > 0) { | |
2185 | spin_lock(&pcp->lock); | |
2186 | free_pcppages_bulk(zone, to_drain, pcp, 0); | |
2187 | spin_unlock(&pcp->lock); | |
2188 | todo++; | |
2189 | } | |
2190 | ||
2191 | return todo; | |
2192 | } | |
2193 | ||
4ae7c039 | 2194 | #ifdef CONFIG_NUMA |
8fce4d8e | 2195 | /* |
4037d452 CL |
2196 | * Called from the vmstat counter updater to drain pagesets of this |
2197 | * currently executing processor on remote nodes after they have | |
2198 | * expired. | |
8fce4d8e | 2199 | */ |
4037d452 | 2200 | void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) |
4ae7c039 | 2201 | { |
7be12fc9 | 2202 | int to_drain, batch; |
4ae7c039 | 2203 | |
4db0c3c2 | 2204 | batch = READ_ONCE(pcp->batch); |
7be12fc9 | 2205 | to_drain = min(pcp->count, batch); |
4b23a68f | 2206 | if (to_drain > 0) { |
57490774 | 2207 | spin_lock(&pcp->lock); |
fd56eef2 | 2208 | free_pcppages_bulk(zone, to_drain, pcp, 0); |
57490774 | 2209 | spin_unlock(&pcp->lock); |
4b23a68f | 2210 | } |
4ae7c039 CL |
2211 | } |
2212 | #endif | |
2213 | ||
9f8f2172 | 2214 | /* |
93481ff0 | 2215 | * Drain pcplists of the indicated processor and zone. |
9f8f2172 | 2216 | */ |
93481ff0 | 2217 | static void drain_pages_zone(unsigned int cpu, struct zone *zone) |
1da177e4 | 2218 | { |
93481ff0 | 2219 | struct per_cpu_pages *pcp; |
1da177e4 | 2220 | |
28f836b6 | 2221 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); |
4b23a68f | 2222 | if (pcp->count) { |
57490774 | 2223 | spin_lock(&pcp->lock); |
4b23a68f | 2224 | free_pcppages_bulk(zone, pcp->count, pcp, 0); |
57490774 | 2225 | spin_unlock(&pcp->lock); |
4b23a68f | 2226 | } |
93481ff0 | 2227 | } |
3dfa5721 | 2228 | |
93481ff0 VB |
2229 | /* |
2230 | * Drain pcplists of all zones on the indicated processor. | |
93481ff0 VB |
2231 | */ |
2232 | static void drain_pages(unsigned int cpu) | |
2233 | { | |
2234 | struct zone *zone; | |
2235 | ||
2236 | for_each_populated_zone(zone) { | |
2237 | drain_pages_zone(cpu, zone); | |
1da177e4 LT |
2238 | } |
2239 | } | |
1da177e4 | 2240 | |
9f8f2172 CL |
2241 | /* |
2242 | * Spill all of this CPU's per-cpu pages back into the buddy allocator. | |
2243 | */ | |
93481ff0 | 2244 | void drain_local_pages(struct zone *zone) |
9f8f2172 | 2245 | { |
93481ff0 VB |
2246 | int cpu = smp_processor_id(); |
2247 | ||
2248 | if (zone) | |
2249 | drain_pages_zone(cpu, zone); | |
2250 | else | |
2251 | drain_pages(cpu); | |
9f8f2172 CL |
2252 | } |
2253 | ||
2254 | /* | |
ec6e8c7e VB |
2255 | * The implementation of drain_all_pages(), exposing an extra parameter to |
2256 | * drain on all cpus. | |
93481ff0 | 2257 | * |
ec6e8c7e VB |
2258 | * drain_all_pages() is optimized to only execute on cpus where pcplists are |
2259 | * not empty. The check for non-emptiness can however race with a free to | |
2260 | * pcplist that has not yet increased the pcp->count from 0 to 1. Callers | |
2261 | * that need the guarantee that every CPU has drained can disable the | |
2262 | * optimizing racy check. | |
9f8f2172 | 2263 | */ |
3b1f3658 | 2264 | static void __drain_all_pages(struct zone *zone, bool force_all_cpus) |
9f8f2172 | 2265 | { |
74046494 | 2266 | int cpu; |
74046494 GBY |
2267 | |
2268 | /* | |
041711ce | 2269 | * Allocate in the BSS so we won't require allocation in |
74046494 GBY |
2270 | * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y |
2271 | */ | |
2272 | static cpumask_t cpus_with_pcps; | |
2273 | ||
bd233f53 MG |
2274 | /* |
2275 | * Do not drain if one is already in progress unless it's specific to | |
2276 | * a zone. Such callers are primarily CMA and memory hotplug and need | |
2277 | * the drain to be complete when the call returns. | |
2278 | */ | |
2279 | if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) { | |
2280 | if (!zone) | |
2281 | return; | |
2282 | mutex_lock(&pcpu_drain_mutex); | |
2283 | } | |
0ccce3b9 | 2284 | |
74046494 GBY |
2285 | /* |
2286 | * We don't care about racing with CPU hotplug event | |
2287 | * as offline notification will cause the notified | |
2288 | * cpu to drain that CPU pcps and on_each_cpu_mask | |
2289 | * disables preemption as part of its processing | |
2290 | */ | |
2291 | for_each_online_cpu(cpu) { | |
28f836b6 | 2292 | struct per_cpu_pages *pcp; |
93481ff0 | 2293 | struct zone *z; |
74046494 | 2294 | bool has_pcps = false; |
93481ff0 | 2295 | |
ec6e8c7e VB |
2296 | if (force_all_cpus) { |
2297 | /* | |
2298 | * The pcp.count check is racy, some callers need a | |
2299 | * guarantee that no cpu is missed. | |
2300 | */ | |
2301 | has_pcps = true; | |
2302 | } else if (zone) { | |
28f836b6 MG |
2303 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); |
2304 | if (pcp->count) | |
74046494 | 2305 | has_pcps = true; |
93481ff0 VB |
2306 | } else { |
2307 | for_each_populated_zone(z) { | |
28f836b6 MG |
2308 | pcp = per_cpu_ptr(z->per_cpu_pageset, cpu); |
2309 | if (pcp->count) { | |
93481ff0 VB |
2310 | has_pcps = true; |
2311 | break; | |
2312 | } | |
74046494 GBY |
2313 | } |
2314 | } | |
93481ff0 | 2315 | |
74046494 GBY |
2316 | if (has_pcps) |
2317 | cpumask_set_cpu(cpu, &cpus_with_pcps); | |
2318 | else | |
2319 | cpumask_clear_cpu(cpu, &cpus_with_pcps); | |
2320 | } | |
0ccce3b9 | 2321 | |
bd233f53 | 2322 | for_each_cpu(cpu, &cpus_with_pcps) { |
443c2acc NSJ |
2323 | if (zone) |
2324 | drain_pages_zone(cpu, zone); | |
2325 | else | |
2326 | drain_pages(cpu); | |
0ccce3b9 | 2327 | } |
bd233f53 MG |
2328 | |
2329 | mutex_unlock(&pcpu_drain_mutex); | |
9f8f2172 CL |
2330 | } |
2331 | ||
ec6e8c7e VB |
2332 | /* |
2333 | * Spill all the per-cpu pages from all CPUs back into the buddy allocator. | |
2334 | * | |
2335 | * When zone parameter is non-NULL, spill just the single zone's pages. | |
ec6e8c7e VB |
2336 | */ |
2337 | void drain_all_pages(struct zone *zone) | |
2338 | { | |
2339 | __drain_all_pages(zone, false); | |
2340 | } | |
2341 | ||
44042b44 MG |
2342 | static bool free_unref_page_prepare(struct page *page, unsigned long pfn, |
2343 | unsigned int order) | |
1da177e4 | 2344 | { |
5f8dcc21 | 2345 | int migratetype; |
1da177e4 | 2346 | |
5267fe5d | 2347 | if (!free_pages_prepare(page, order)) |
9cca35d4 | 2348 | return false; |
689bcebf | 2349 | |
dc4b0caf | 2350 | migratetype = get_pfnblock_migratetype(page, pfn); |
bb14c2c7 | 2351 | set_pcppage_migratetype(page, migratetype); |
9cca35d4 MG |
2352 | return true; |
2353 | } | |
2354 | ||
51a755c5 | 2355 | static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high) |
3b12e7e9 MG |
2356 | { |
2357 | int min_nr_free, max_nr_free; | |
2358 | ||
51a755c5 | 2359 | /* Free as much as possible if batch freeing high-order pages. */ |
f26b3fa0 | 2360 | if (unlikely(free_high)) |
51a755c5 | 2361 | return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX); |
f26b3fa0 | 2362 | |
3b12e7e9 MG |
2363 | /* Check for PCP disabled or boot pageset */ |
2364 | if (unlikely(high < batch)) | |
2365 | return 1; | |
2366 | ||
2367 | /* Leave at least pcp->batch pages on the list */ | |
2368 | min_nr_free = batch; | |
2369 | max_nr_free = high - batch; | |
2370 | ||
2371 | /* | |
6ccdcb6d HY |
2372 | * Increase the batch number to the number of the consecutive |
2373 | * freed pages to reduce zone lock contention. | |
3b12e7e9 | 2374 | */ |
6ccdcb6d | 2375 | batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free); |
3b12e7e9 MG |
2376 | |
2377 | return batch; | |
2378 | } | |
2379 | ||
f26b3fa0 | 2380 | static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone, |
51a755c5 | 2381 | int batch, bool free_high) |
c49c2c47 | 2382 | { |
51a755c5 | 2383 | int high, high_min, high_max; |
c49c2c47 | 2384 | |
51a755c5 HY |
2385 | high_min = READ_ONCE(pcp->high_min); |
2386 | high_max = READ_ONCE(pcp->high_max); | |
2387 | high = pcp->high = clamp(pcp->high, high_min, high_max); | |
2388 | ||
2389 | if (unlikely(!high)) | |
c49c2c47 MG |
2390 | return 0; |
2391 | ||
51a755c5 HY |
2392 | if (unlikely(free_high)) { |
2393 | pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX), | |
2394 | high_min); | |
2395 | return 0; | |
2396 | } | |
c49c2c47 MG |
2397 | |
2398 | /* | |
2399 | * If reclaim is active, limit the number of pages that can be | |
2400 | * stored on pcp lists | |
2401 | */ | |
51a755c5 | 2402 | if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) { |
6ccdcb6d HY |
2403 | int free_count = max_t(int, pcp->free_count, batch); |
2404 | ||
2405 | pcp->high = max(high - free_count, high_min); | |
51a755c5 HY |
2406 | return min(batch << 2, pcp->high); |
2407 | } | |
2408 | ||
57c0419c HY |
2409 | if (high_min == high_max) |
2410 | return high; | |
2411 | ||
2412 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) { | |
6ccdcb6d HY |
2413 | int free_count = max_t(int, pcp->free_count, batch); |
2414 | ||
2415 | pcp->high = max(high - free_count, high_min); | |
57c0419c HY |
2416 | high = max(pcp->count, high_min); |
2417 | } else if (pcp->count >= high) { | |
6ccdcb6d | 2418 | int need_high = pcp->free_count + batch; |
51a755c5 HY |
2419 | |
2420 | /* pcp->high should be large enough to hold batch freed pages */ | |
2421 | if (pcp->high < need_high) | |
2422 | pcp->high = clamp(need_high, high_min, high_max); | |
2423 | } | |
2424 | ||
2425 | return high; | |
c49c2c47 MG |
2426 | } |
2427 | ||
4b23a68f MG |
2428 | static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp, |
2429 | struct page *page, int migratetype, | |
56651377 | 2430 | unsigned int order) |
9cca35d4 | 2431 | { |
51a755c5 | 2432 | int high, batch; |
44042b44 | 2433 | int pindex; |
ca71fe1a | 2434 | bool free_high = false; |
9cca35d4 | 2435 | |
c0a24239 HY |
2436 | /* |
2437 | * On freeing, reduce the number of pages that are batch allocated. | |
2438 | * See nr_pcp_alloc() where alloc_factor is increased for subsequent | |
2439 | * allocations. | |
2440 | */ | |
2441 | pcp->alloc_factor >>= 1; | |
15cd9004 | 2442 | __count_vm_events(PGFREE, 1 << order); |
44042b44 | 2443 | pindex = order_to_pindex(migratetype, order); |
bf75f200 | 2444 | list_add(&page->pcp_list, &pcp->lists[pindex]); |
44042b44 | 2445 | pcp->count += 1 << order; |
f26b3fa0 | 2446 | |
51a755c5 | 2447 | batch = READ_ONCE(pcp->batch); |
f26b3fa0 MG |
2448 | /* |
2449 | * As high-order pages other than THP's stored on PCP can contribute | |
2450 | * to fragmentation, limit the number stored when PCP is heavily | |
2451 | * freeing without allocation. The remainder after bulk freeing | |
2452 | * stops will be drained from vmstat refresh context. | |
2453 | */ | |
ca71fe1a | 2454 | if (order && order <= PAGE_ALLOC_COSTLY_ORDER) { |
6ccdcb6d | 2455 | free_high = (pcp->free_count >= batch && |
362d37a1 HY |
2456 | (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) && |
2457 | (!(pcp->flags & PCPF_FREE_HIGH_BATCH) || | |
51a755c5 | 2458 | pcp->count >= READ_ONCE(batch))); |
ca71fe1a HY |
2459 | pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER; |
2460 | } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) { | |
2461 | pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER; | |
2462 | } | |
6ccdcb6d HY |
2463 | if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX)) |
2464 | pcp->free_count += (1 << order); | |
51a755c5 | 2465 | high = nr_pcp_high(pcp, zone, batch, free_high); |
3b12e7e9 | 2466 | if (pcp->count >= high) { |
51a755c5 HY |
2467 | free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high), |
2468 | pcp, pindex); | |
57c0419c HY |
2469 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags) && |
2470 | zone_watermark_ok(zone, 0, high_wmark_pages(zone), | |
2471 | ZONE_MOVABLE, 0)) | |
2472 | clear_bit(ZONE_BELOW_HIGH, &zone->flags); | |
3b12e7e9 | 2473 | } |
9cca35d4 | 2474 | } |
5f8dcc21 | 2475 | |
9cca35d4 | 2476 | /* |
44042b44 | 2477 | * Free a pcp page |
9cca35d4 | 2478 | */ |
44042b44 | 2479 | void free_unref_page(struct page *page, unsigned int order) |
9cca35d4 | 2480 | { |
4b23a68f MG |
2481 | unsigned long __maybe_unused UP_flags; |
2482 | struct per_cpu_pages *pcp; | |
2483 | struct zone *zone; | |
9cca35d4 | 2484 | unsigned long pfn = page_to_pfn(page); |
7b086755 | 2485 | int migratetype, pcpmigratetype; |
9cca35d4 | 2486 | |
44042b44 | 2487 | if (!free_unref_page_prepare(page, pfn, order)) |
9cca35d4 | 2488 | return; |
da456f14 | 2489 | |
5f8dcc21 MG |
2490 | /* |
2491 | * We only track unmovable, reclaimable and movable on pcp lists. | |
df1acc85 | 2492 | * Place ISOLATE pages on the isolated list because they are being |
7b086755 JW |
2493 | * offlined but treat HIGHATOMIC and CMA as movable pages so we can |
2494 | * get those areas back if necessary. Otherwise, we may have to free | |
5f8dcc21 MG |
2495 | * excessively into the page allocator |
2496 | */ | |
7b086755 | 2497 | migratetype = pcpmigratetype = get_pcppage_migratetype(page); |
df1acc85 | 2498 | if (unlikely(migratetype >= MIGRATE_PCPTYPES)) { |
194159fb | 2499 | if (unlikely(is_migrate_isolate(migratetype))) { |
44042b44 | 2500 | free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE); |
9cca35d4 | 2501 | return; |
5f8dcc21 | 2502 | } |
7b086755 | 2503 | pcpmigratetype = MIGRATE_MOVABLE; |
5f8dcc21 MG |
2504 | } |
2505 | ||
4b23a68f MG |
2506 | zone = page_zone(page); |
2507 | pcp_trylock_prepare(UP_flags); | |
57490774 | 2508 | pcp = pcp_spin_trylock(zone->per_cpu_pageset); |
01b44456 | 2509 | if (pcp) { |
7b086755 | 2510 | free_unref_page_commit(zone, pcp, page, pcpmigratetype, order); |
57490774 | 2511 | pcp_spin_unlock(pcp); |
4b23a68f MG |
2512 | } else { |
2513 | free_one_page(zone, page, pfn, order, migratetype, FPI_NONE); | |
2514 | } | |
2515 | pcp_trylock_finish(UP_flags); | |
1da177e4 LT |
2516 | } |
2517 | ||
cc59850e | 2518 | /* |
31b2ff82 | 2519 | * Free a batch of folios |
cc59850e | 2520 | */ |
90491d87 | 2521 | void free_unref_folios(struct folio_batch *folios) |
cc59850e | 2522 | { |
57490774 | 2523 | unsigned long __maybe_unused UP_flags; |
4b23a68f MG |
2524 | struct per_cpu_pages *pcp = NULL; |
2525 | struct zone *locked_zone = NULL; | |
90491d87 | 2526 | int i, j, migratetype; |
9cca35d4 | 2527 | |
90491d87 MWO |
2528 | /* Prepare folios for freeing */ |
2529 | for (i = 0, j = 0; i < folios->nr; i++) { | |
2530 | struct folio *folio = folios->folios[i]; | |
7c76d922 | 2531 | unsigned long pfn = folio_pfn(folio); |
31b2ff82 | 2532 | unsigned int order = folio_order(folio); |
9cca35d4 | 2533 | |
31b2ff82 MWO |
2534 | if (order > 0 && folio_test_large_rmappable(folio)) |
2535 | folio_undo_large_rmappable(folio); | |
2536 | if (!free_unref_page_prepare(&folio->page, pfn, order)) | |
053cfda1 | 2537 | continue; |
df1acc85 MG |
2538 | |
2539 | /* | |
31b2ff82 MWO |
2540 | * Free isolated folios and orders not handled on the PCP |
2541 | * directly to the allocator, see comment in free_unref_page. | |
df1acc85 | 2542 | */ |
7c76d922 | 2543 | migratetype = get_pcppage_migratetype(&folio->page); |
31b2ff82 MWO |
2544 | if (!pcp_allowed_order(order) || |
2545 | is_migrate_isolate(migratetype)) { | |
7c76d922 | 2546 | free_one_page(folio_zone(folio), &folio->page, pfn, |
31b2ff82 | 2547 | order, migratetype, FPI_NONE); |
47aef601 | 2548 | continue; |
df1acc85 | 2549 | } |
31b2ff82 | 2550 | folio->private = (void *)(unsigned long)order; |
90491d87 MWO |
2551 | if (j != i) |
2552 | folios->folios[j] = folio; | |
2553 | j++; | |
9cca35d4 | 2554 | } |
90491d87 | 2555 | folios->nr = j; |
cc59850e | 2556 | |
90491d87 MWO |
2557 | for (i = 0; i < folios->nr; i++) { |
2558 | struct folio *folio = folios->folios[i]; | |
7c76d922 | 2559 | struct zone *zone = folio_zone(folio); |
31b2ff82 | 2560 | unsigned int order = (unsigned long)folio->private; |
4b23a68f | 2561 | |
31b2ff82 | 2562 | folio->private = NULL; |
7c76d922 | 2563 | migratetype = get_pcppage_migratetype(&folio->page); |
c3e58a70 | 2564 | |
90491d87 MWO |
2565 | /* Different zone requires a different pcp lock */ |
2566 | if (zone != locked_zone) { | |
57490774 MG |
2567 | if (pcp) { |
2568 | pcp_spin_unlock(pcp); | |
2569 | pcp_trylock_finish(UP_flags); | |
2570 | } | |
01b44456 | 2571 | |
57490774 | 2572 | /* |
7c76d922 | 2573 | * trylock is necessary as folios may be getting freed |
57490774 MG |
2574 | * from IRQ or SoftIRQ context after an IO completion. |
2575 | */ | |
2576 | pcp_trylock_prepare(UP_flags); | |
2577 | pcp = pcp_spin_trylock(zone->per_cpu_pageset); | |
2578 | if (unlikely(!pcp)) { | |
2579 | pcp_trylock_finish(UP_flags); | |
7c76d922 | 2580 | free_one_page(zone, &folio->page, |
31b2ff82 | 2581 | folio_pfn(folio), order, |
7c76d922 | 2582 | migratetype, FPI_NONE); |
57490774 MG |
2583 | locked_zone = NULL; |
2584 | continue; | |
2585 | } | |
4b23a68f | 2586 | locked_zone = zone; |
4b23a68f MG |
2587 | } |
2588 | ||
47aef601 DB |
2589 | /* |
2590 | * Non-isolated types over MIGRATE_PCPTYPES get added | |
2591 | * to the MIGRATE_MOVABLE pcp list. | |
2592 | */ | |
47aef601 DB |
2593 | if (unlikely(migratetype >= MIGRATE_PCPTYPES)) |
2594 | migratetype = MIGRATE_MOVABLE; | |
2595 | ||
7c76d922 | 2596 | trace_mm_page_free_batched(&folio->page); |
31b2ff82 MWO |
2597 | free_unref_page_commit(zone, pcp, &folio->page, migratetype, |
2598 | order); | |
cc59850e | 2599 | } |
4b23a68f | 2600 | |
57490774 MG |
2601 | if (pcp) { |
2602 | pcp_spin_unlock(pcp); | |
2603 | pcp_trylock_finish(UP_flags); | |
2604 | } | |
90491d87 | 2605 | folio_batch_reinit(folios); |
cc59850e KK |
2606 | } |
2607 | ||
8dfcc9ba NP |
2608 | /* |
2609 | * split_page takes a non-compound higher-order page, and splits it into | |
2610 | * n (1<<order) sub-pages: page[0..n] | |
2611 | * Each sub-page must be freed individually. | |
2612 | * | |
2613 | * Note: this is probably too low level an operation for use in drivers. | |
2614 | * Please consult with lkml before using this in your driver. | |
2615 | */ | |
2616 | void split_page(struct page *page, unsigned int order) | |
2617 | { | |
2618 | int i; | |
2619 | ||
309381fe SL |
2620 | VM_BUG_ON_PAGE(PageCompound(page), page); |
2621 | VM_BUG_ON_PAGE(!page_count(page), page); | |
b1eeab67 | 2622 | |
a9627bc5 | 2623 | for (i = 1; i < (1 << order); i++) |
7835e98b | 2624 | set_page_refcounted(page + i); |
46d44d09 | 2625 | split_page_owner(page, order, 0); |
b8791381 | 2626 | split_page_memcg(page, order, 0); |
8dfcc9ba | 2627 | } |
5853ff23 | 2628 | EXPORT_SYMBOL_GPL(split_page); |
8dfcc9ba | 2629 | |
3c605096 | 2630 | int __isolate_free_page(struct page *page, unsigned int order) |
748446bb | 2631 | { |
9a157dd8 KW |
2632 | struct zone *zone = page_zone(page); |
2633 | int mt = get_pageblock_migratetype(page); | |
748446bb | 2634 | |
194159fb | 2635 | if (!is_migrate_isolate(mt)) { |
9a157dd8 | 2636 | unsigned long watermark; |
8348faf9 VB |
2637 | /* |
2638 | * Obey watermarks as if the page was being allocated. We can | |
2639 | * emulate a high-order watermark check with a raised order-0 | |
2640 | * watermark, because we already know our high-order page | |
2641 | * exists. | |
2642 | */ | |
fd1444b2 | 2643 | watermark = zone->_watermark[WMARK_MIN] + (1UL << order); |
d883c6cf | 2644 | if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA)) |
2e30abd1 MS |
2645 | return 0; |
2646 | ||
8fb74b9f | 2647 | __mod_zone_freepage_state(zone, -(1UL << order), mt); |
2e30abd1 | 2648 | } |
748446bb | 2649 | |
6ab01363 | 2650 | del_page_from_free_list(page, zone, order); |
2139cbe6 | 2651 | |
400bc7fd | 2652 | /* |
2653 | * Set the pageblock if the isolated page is at least half of a | |
2654 | * pageblock | |
2655 | */ | |
748446bb MG |
2656 | if (order >= pageblock_order - 1) { |
2657 | struct page *endpage = page + (1 << order) - 1; | |
47118af0 MN |
2658 | for (; page < endpage; page += pageblock_nr_pages) { |
2659 | int mt = get_pageblock_migratetype(page); | |
1dd214b8 ZY |
2660 | /* |
2661 | * Only change normal pageblocks (i.e., they can merge | |
2662 | * with others) | |
2663 | */ | |
2664 | if (migratetype_is_mergeable(mt)) | |
47118af0 MN |
2665 | set_pageblock_migratetype(page, |
2666 | MIGRATE_MOVABLE); | |
2667 | } | |
748446bb MG |
2668 | } |
2669 | ||
8fb74b9f | 2670 | return 1UL << order; |
1fb3f8ca MG |
2671 | } |
2672 | ||
624f58d8 AD |
2673 | /** |
2674 | * __putback_isolated_page - Return a now-isolated page back where we got it | |
2675 | * @page: Page that was isolated | |
2676 | * @order: Order of the isolated page | |
e6a0a7ad | 2677 | * @mt: The page's pageblock's migratetype |
624f58d8 AD |
2678 | * |
2679 | * This function is meant to return a page pulled from the free lists via | |
2680 | * __isolate_free_page back to the free lists they were pulled from. | |
2681 | */ | |
2682 | void __putback_isolated_page(struct page *page, unsigned int order, int mt) | |
2683 | { | |
2684 | struct zone *zone = page_zone(page); | |
2685 | ||
2686 | /* zone lock should be held when this function is called */ | |
2687 | lockdep_assert_held(&zone->lock); | |
2688 | ||
2689 | /* Return isolated page to tail of freelist. */ | |
f04a5d5d | 2690 | __free_one_page(page, page_to_pfn(page), zone, order, mt, |
47b6a24a | 2691 | FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL); |
624f58d8 AD |
2692 | } |
2693 | ||
060e7417 MG |
2694 | /* |
2695 | * Update NUMA hit/miss statistics | |
060e7417 | 2696 | */ |
3e23060b MG |
2697 | static inline void zone_statistics(struct zone *preferred_zone, struct zone *z, |
2698 | long nr_account) | |
060e7417 MG |
2699 | { |
2700 | #ifdef CONFIG_NUMA | |
3a321d2a | 2701 | enum numa_stat_item local_stat = NUMA_LOCAL; |
060e7417 | 2702 | |
4518085e KW |
2703 | /* skip numa counters update if numa stats is disabled */ |
2704 | if (!static_branch_likely(&vm_numa_stat_key)) | |
2705 | return; | |
2706 | ||
c1093b74 | 2707 | if (zone_to_nid(z) != numa_node_id()) |
060e7417 | 2708 | local_stat = NUMA_OTHER; |
060e7417 | 2709 | |
c1093b74 | 2710 | if (zone_to_nid(z) == zone_to_nid(preferred_zone)) |
3e23060b | 2711 | __count_numa_events(z, NUMA_HIT, nr_account); |
2df26639 | 2712 | else { |
3e23060b MG |
2713 | __count_numa_events(z, NUMA_MISS, nr_account); |
2714 | __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account); | |
060e7417 | 2715 | } |
3e23060b | 2716 | __count_numa_events(z, local_stat, nr_account); |
060e7417 MG |
2717 | #endif |
2718 | } | |
2719 | ||
589d9973 MG |
2720 | static __always_inline |
2721 | struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone, | |
2722 | unsigned int order, unsigned int alloc_flags, | |
2723 | int migratetype) | |
2724 | { | |
2725 | struct page *page; | |
2726 | unsigned long flags; | |
2727 | ||
2728 | do { | |
2729 | page = NULL; | |
2730 | spin_lock_irqsave(&zone->lock, flags); | |
eb2e2b42 | 2731 | if (alloc_flags & ALLOC_HIGHATOMIC) |
589d9973 MG |
2732 | page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC); |
2733 | if (!page) { | |
2734 | page = __rmqueue(zone, order, migratetype, alloc_flags); | |
eb2e2b42 MG |
2735 | |
2736 | /* | |
2737 | * If the allocation fails, allow OOM handling access | |
2738 | * to HIGHATOMIC reserves as failing now is worse than | |
2739 | * failing a high-order atomic allocation in the | |
2740 | * future. | |
2741 | */ | |
2742 | if (!page && (alloc_flags & ALLOC_OOM)) | |
2743 | page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC); | |
2744 | ||
589d9973 MG |
2745 | if (!page) { |
2746 | spin_unlock_irqrestore(&zone->lock, flags); | |
2747 | return NULL; | |
2748 | } | |
2749 | } | |
2750 | __mod_zone_freepage_state(zone, -(1 << order), | |
2751 | get_pcppage_migratetype(page)); | |
2752 | spin_unlock_irqrestore(&zone->lock, flags); | |
2753 | } while (check_new_pages(page, order)); | |
2754 | ||
2755 | __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order); | |
2756 | zone_statistics(preferred_zone, zone, 1); | |
2757 | ||
2758 | return page; | |
2759 | } | |
2760 | ||
51a755c5 | 2761 | static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order) |
c0a24239 | 2762 | { |
51a755c5 HY |
2763 | int high, base_batch, batch, max_nr_alloc; |
2764 | int high_max, high_min; | |
c0a24239 | 2765 | |
51a755c5 HY |
2766 | base_batch = READ_ONCE(pcp->batch); |
2767 | high_min = READ_ONCE(pcp->high_min); | |
2768 | high_max = READ_ONCE(pcp->high_max); | |
2769 | high = pcp->high = clamp(pcp->high, high_min, high_max); | |
c0a24239 HY |
2770 | |
2771 | /* Check for PCP disabled or boot pageset */ | |
51a755c5 | 2772 | if (unlikely(high < base_batch)) |
c0a24239 HY |
2773 | return 1; |
2774 | ||
51a755c5 HY |
2775 | if (order) |
2776 | batch = base_batch; | |
2777 | else | |
2778 | batch = (base_batch << pcp->alloc_factor); | |
2779 | ||
c0a24239 | 2780 | /* |
51a755c5 HY |
2781 | * If we had larger pcp->high, we could avoid to allocate from |
2782 | * zone. | |
c0a24239 | 2783 | */ |
57c0419c | 2784 | if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags)) |
51a755c5 HY |
2785 | high = pcp->high = min(high + batch, high_max); |
2786 | ||
c0a24239 | 2787 | if (!order) { |
51a755c5 HY |
2788 | max_nr_alloc = max(high - pcp->count - base_batch, base_batch); |
2789 | /* | |
2790 | * Double the number of pages allocated each time there is | |
2791 | * subsequent allocation of order-0 pages without any freeing. | |
2792 | */ | |
c0a24239 HY |
2793 | if (batch <= max_nr_alloc && |
2794 | pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX) | |
2795 | pcp->alloc_factor++; | |
2796 | batch = min(batch, max_nr_alloc); | |
2797 | } | |
2798 | ||
2799 | /* | |
2800 | * Scale batch relative to order if batch implies free pages | |
2801 | * can be stored on the PCP. Batch can be 1 for small zones or | |
2802 | * for boot pagesets which should never store free pages as | |
2803 | * the pages may belong to arbitrary zones. | |
2804 | */ | |
2805 | if (batch > 1) | |
2806 | batch = max(batch >> order, 2); | |
2807 | ||
2808 | return batch; | |
2809 | } | |
2810 | ||
066b2393 | 2811 | /* Remove page from the per-cpu list, caller must protect the list */ |
3b822017 | 2812 | static inline |
44042b44 MG |
2813 | struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order, |
2814 | int migratetype, | |
6bb15450 | 2815 | unsigned int alloc_flags, |
453f85d4 | 2816 | struct per_cpu_pages *pcp, |
066b2393 MG |
2817 | struct list_head *list) |
2818 | { | |
2819 | struct page *page; | |
2820 | ||
2821 | do { | |
2822 | if (list_empty(list)) { | |
51a755c5 | 2823 | int batch = nr_pcp_alloc(pcp, zone, order); |
44042b44 MG |
2824 | int alloced; |
2825 | ||
44042b44 MG |
2826 | alloced = rmqueue_bulk(zone, order, |
2827 | batch, list, | |
6bb15450 | 2828 | migratetype, alloc_flags); |
44042b44 MG |
2829 | |
2830 | pcp->count += alloced << order; | |
066b2393 MG |
2831 | if (unlikely(list_empty(list))) |
2832 | return NULL; | |
2833 | } | |
2834 | ||
bf75f200 MG |
2835 | page = list_first_entry(list, struct page, pcp_list); |
2836 | list_del(&page->pcp_list); | |
44042b44 | 2837 | pcp->count -= 1 << order; |
700d2e9a | 2838 | } while (check_new_pages(page, order)); |
066b2393 MG |
2839 | |
2840 | return page; | |
2841 | } | |
2842 | ||
2843 | /* Lock and remove page from the per-cpu list */ | |
2844 | static struct page *rmqueue_pcplist(struct zone *preferred_zone, | |
44042b44 | 2845 | struct zone *zone, unsigned int order, |
663d0cfd | 2846 | int migratetype, unsigned int alloc_flags) |
066b2393 MG |
2847 | { |
2848 | struct per_cpu_pages *pcp; | |
2849 | struct list_head *list; | |
066b2393 | 2850 | struct page *page; |
4b23a68f | 2851 | unsigned long __maybe_unused UP_flags; |
066b2393 | 2852 | |
57490774 | 2853 | /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */ |
4b23a68f | 2854 | pcp_trylock_prepare(UP_flags); |
57490774 | 2855 | pcp = pcp_spin_trylock(zone->per_cpu_pageset); |
01b44456 | 2856 | if (!pcp) { |
4b23a68f | 2857 | pcp_trylock_finish(UP_flags); |
4b23a68f MG |
2858 | return NULL; |
2859 | } | |
3b12e7e9 MG |
2860 | |
2861 | /* | |
2862 | * On allocation, reduce the number of pages that are batch freed. | |
2863 | * See nr_pcp_free() where free_factor is increased for subsequent | |
2864 | * frees. | |
2865 | */ | |
6ccdcb6d | 2866 | pcp->free_count >>= 1; |
44042b44 MG |
2867 | list = &pcp->lists[order_to_pindex(migratetype, order)]; |
2868 | page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list); | |
57490774 | 2869 | pcp_spin_unlock(pcp); |
4b23a68f | 2870 | pcp_trylock_finish(UP_flags); |
066b2393 | 2871 | if (page) { |
15cd9004 | 2872 | __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order); |
3e23060b | 2873 | zone_statistics(preferred_zone, zone, 1); |
066b2393 | 2874 | } |
066b2393 MG |
2875 | return page; |
2876 | } | |
2877 | ||
1da177e4 | 2878 | /* |
a57ae9ef RX |
2879 | * Allocate a page from the given zone. |
2880 | * Use pcplists for THP or "cheap" high-order allocations. | |
1da177e4 | 2881 | */ |
b073d7f8 AP |
2882 | |
2883 | /* | |
2884 | * Do not instrument rmqueue() with KMSAN. This function may call | |
2885 | * __msan_poison_alloca() through a call to set_pfnblock_flags_mask(). | |
2886 | * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it | |
2887 | * may call rmqueue() again, which will result in a deadlock. | |
1da177e4 | 2888 | */ |
b073d7f8 | 2889 | __no_sanitize_memory |
0a15c3e9 | 2890 | static inline |
066b2393 | 2891 | struct page *rmqueue(struct zone *preferred_zone, |
7aeb09f9 | 2892 | struct zone *zone, unsigned int order, |
c603844b MG |
2893 | gfp_t gfp_flags, unsigned int alloc_flags, |
2894 | int migratetype) | |
1da177e4 | 2895 | { |
689bcebf | 2896 | struct page *page; |
1da177e4 | 2897 | |
589d9973 MG |
2898 | /* |
2899 | * We most definitely don't want callers attempting to | |
2900 | * allocate greater than order-1 page units with __GFP_NOFAIL. | |
2901 | */ | |
2902 | WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1)); | |
2903 | ||
44042b44 | 2904 | if (likely(pcp_allowed_order(order))) { |
f945116e JW |
2905 | page = rmqueue_pcplist(preferred_zone, zone, order, |
2906 | migratetype, alloc_flags); | |
2907 | if (likely(page)) | |
2908 | goto out; | |
066b2393 | 2909 | } |
83b9355b | 2910 | |
589d9973 MG |
2911 | page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags, |
2912 | migratetype); | |
1da177e4 | 2913 | |
066b2393 | 2914 | out: |
73444bc4 | 2915 | /* Separate test+clear to avoid unnecessary atomics */ |
3b11edf1 TH |
2916 | if ((alloc_flags & ALLOC_KSWAPD) && |
2917 | unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) { | |
73444bc4 MG |
2918 | clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags); |
2919 | wakeup_kswapd(zone, 0, 0, zone_idx(zone)); | |
2920 | } | |
2921 | ||
066b2393 | 2922 | VM_BUG_ON_PAGE(page && bad_range(zone, page), page); |
1da177e4 LT |
2923 | return page; |
2924 | } | |
2925 | ||
54aa3866 | 2926 | noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) |
af3b8544 BP |
2927 | { |
2928 | return __should_fail_alloc_page(gfp_mask, order); | |
2929 | } | |
2930 | ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE); | |
2931 | ||
f27ce0e1 JK |
2932 | static inline long __zone_watermark_unusable_free(struct zone *z, |
2933 | unsigned int order, unsigned int alloc_flags) | |
2934 | { | |
f27ce0e1 JK |
2935 | long unusable_free = (1 << order) - 1; |
2936 | ||
2937 | /* | |
ab350885 MG |
2938 | * If the caller does not have rights to reserves below the min |
2939 | * watermark then subtract the high-atomic reserves. This will | |
2940 | * over-estimate the size of the atomic reserve but it avoids a search. | |
f27ce0e1 | 2941 | */ |
ab350885 | 2942 | if (likely(!(alloc_flags & ALLOC_RESERVES))) |
f27ce0e1 JK |
2943 | unusable_free += z->nr_reserved_highatomic; |
2944 | ||
2945 | #ifdef CONFIG_CMA | |
2946 | /* If allocation can't use CMA areas don't use free CMA pages */ | |
2947 | if (!(alloc_flags & ALLOC_CMA)) | |
2948 | unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES); | |
2949 | #endif | |
dcdfdd40 KS |
2950 | #ifdef CONFIG_UNACCEPTED_MEMORY |
2951 | unusable_free += zone_page_state(z, NR_UNACCEPTED); | |
2952 | #endif | |
f27ce0e1 JK |
2953 | |
2954 | return unusable_free; | |
2955 | } | |
2956 | ||
1da177e4 | 2957 | /* |
97a16fc8 MG |
2958 | * Return true if free base pages are above 'mark'. For high-order checks it |
2959 | * will return true of the order-0 watermark is reached and there is at least | |
2960 | * one free page of a suitable size. Checking now avoids taking the zone lock | |
2961 | * to check in the allocation paths if no pages are free. | |
1da177e4 | 2962 | */ |
86a294a8 | 2963 | bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark, |
97a225e6 | 2964 | int highest_zoneidx, unsigned int alloc_flags, |
86a294a8 | 2965 | long free_pages) |
1da177e4 | 2966 | { |
d23ad423 | 2967 | long min = mark; |
1da177e4 LT |
2968 | int o; |
2969 | ||
0aaa29a5 | 2970 | /* free_pages may go negative - that's OK */ |
f27ce0e1 | 2971 | free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags); |
0aaa29a5 | 2972 | |
ab350885 MG |
2973 | if (unlikely(alloc_flags & ALLOC_RESERVES)) { |
2974 | /* | |
2975 | * __GFP_HIGH allows access to 50% of the min reserve as well | |
2976 | * as OOM. | |
2977 | */ | |
1ebbb218 | 2978 | if (alloc_flags & ALLOC_MIN_RESERVE) { |
ab350885 | 2979 | min -= min / 2; |
0aaa29a5 | 2980 | |
1ebbb218 MG |
2981 | /* |
2982 | * Non-blocking allocations (e.g. GFP_ATOMIC) can | |
2983 | * access more reserves than just __GFP_HIGH. Other | |
2984 | * non-blocking allocations requests such as GFP_NOWAIT | |
2985 | * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get | |
2986 | * access to the min reserve. | |
2987 | */ | |
2988 | if (alloc_flags & ALLOC_NON_BLOCK) | |
2989 | min -= min / 4; | |
2990 | } | |
0aaa29a5 | 2991 | |
cd04ae1e | 2992 | /* |
ab350885 | 2993 | * OOM victims can try even harder than the normal reserve |
cd04ae1e MH |
2994 | * users on the grounds that it's definitely going to be in |
2995 | * the exit path shortly and free memory. Any allocation it | |
2996 | * makes during the free path will be small and short-lived. | |
2997 | */ | |
2998 | if (alloc_flags & ALLOC_OOM) | |
2999 | min -= min / 2; | |
cd04ae1e MH |
3000 | } |
3001 | ||
97a16fc8 MG |
3002 | /* |
3003 | * Check watermarks for an order-0 allocation request. If these | |
3004 | * are not met, then a high-order request also cannot go ahead | |
3005 | * even if a suitable page happened to be free. | |
3006 | */ | |
97a225e6 | 3007 | if (free_pages <= min + z->lowmem_reserve[highest_zoneidx]) |
88f5acf8 | 3008 | return false; |
1da177e4 | 3009 | |
97a16fc8 MG |
3010 | /* If this is an order-0 request then the watermark is fine */ |
3011 | if (!order) | |
3012 | return true; | |
3013 | ||
3014 | /* For a high-order request, check at least one suitable page is free */ | |
fd377218 | 3015 | for (o = order; o < NR_PAGE_ORDERS; o++) { |
97a16fc8 MG |
3016 | struct free_area *area = &z->free_area[o]; |
3017 | int mt; | |
3018 | ||
3019 | if (!area->nr_free) | |
3020 | continue; | |
3021 | ||
97a16fc8 | 3022 | for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) { |
b03641af | 3023 | if (!free_area_empty(area, mt)) |
97a16fc8 MG |
3024 | return true; |
3025 | } | |
3026 | ||
3027 | #ifdef CONFIG_CMA | |
d883c6cf | 3028 | if ((alloc_flags & ALLOC_CMA) && |
b03641af | 3029 | !free_area_empty(area, MIGRATE_CMA)) { |
97a16fc8 | 3030 | return true; |
d883c6cf | 3031 | } |
97a16fc8 | 3032 | #endif |
eb2e2b42 MG |
3033 | if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) && |
3034 | !free_area_empty(area, MIGRATE_HIGHATOMIC)) { | |
b050e376 | 3035 | return true; |
eb2e2b42 | 3036 | } |
1da177e4 | 3037 | } |
97a16fc8 | 3038 | return false; |
88f5acf8 MG |
3039 | } |
3040 | ||
7aeb09f9 | 3041 | bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark, |
97a225e6 | 3042 | int highest_zoneidx, unsigned int alloc_flags) |
88f5acf8 | 3043 | { |
97a225e6 | 3044 | return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags, |
88f5acf8 MG |
3045 | zone_page_state(z, NR_FREE_PAGES)); |
3046 | } | |
3047 | ||
48ee5f36 | 3048 | static inline bool zone_watermark_fast(struct zone *z, unsigned int order, |
97a225e6 | 3049 | unsigned long mark, int highest_zoneidx, |
f80b08fc | 3050 | unsigned int alloc_flags, gfp_t gfp_mask) |
48ee5f36 | 3051 | { |
f27ce0e1 | 3052 | long free_pages; |
d883c6cf | 3053 | |
f27ce0e1 | 3054 | free_pages = zone_page_state(z, NR_FREE_PAGES); |
48ee5f36 MG |
3055 | |
3056 | /* | |
3057 | * Fast check for order-0 only. If this fails then the reserves | |
f27ce0e1 | 3058 | * need to be calculated. |
48ee5f36 | 3059 | */ |
f27ce0e1 | 3060 | if (!order) { |
9282012f JK |
3061 | long usable_free; |
3062 | long reserved; | |
f27ce0e1 | 3063 | |
9282012f JK |
3064 | usable_free = free_pages; |
3065 | reserved = __zone_watermark_unusable_free(z, 0, alloc_flags); | |
3066 | ||
3067 | /* reserved may over estimate high-atomic reserves. */ | |
3068 | usable_free -= min(usable_free, reserved); | |
3069 | if (usable_free > mark + z->lowmem_reserve[highest_zoneidx]) | |
f27ce0e1 JK |
3070 | return true; |
3071 | } | |
48ee5f36 | 3072 | |
f80b08fc CTR |
3073 | if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags, |
3074 | free_pages)) | |
3075 | return true; | |
2973d822 | 3076 | |
f80b08fc | 3077 | /* |
2973d822 | 3078 | * Ignore watermark boosting for __GFP_HIGH order-0 allocations |
f80b08fc CTR |
3079 | * when checking the min watermark. The min watermark is the |
3080 | * point where boosting is ignored so that kswapd is woken up | |
3081 | * when below the low watermark. | |
3082 | */ | |
2973d822 | 3083 | if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost |
f80b08fc CTR |
3084 | && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) { |
3085 | mark = z->_watermark[WMARK_MIN]; | |
3086 | return __zone_watermark_ok(z, order, mark, highest_zoneidx, | |
3087 | alloc_flags, free_pages); | |
3088 | } | |
3089 | ||
3090 | return false; | |
48ee5f36 MG |
3091 | } |
3092 | ||
7aeb09f9 | 3093 | bool zone_watermark_ok_safe(struct zone *z, unsigned int order, |
97a225e6 | 3094 | unsigned long mark, int highest_zoneidx) |
88f5acf8 MG |
3095 | { |
3096 | long free_pages = zone_page_state(z, NR_FREE_PAGES); | |
3097 | ||
3098 | if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) | |
3099 | free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); | |
3100 | ||
97a225e6 | 3101 | return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0, |
88f5acf8 | 3102 | free_pages); |
1da177e4 LT |
3103 | } |
3104 | ||
9276b1bc | 3105 | #ifdef CONFIG_NUMA |
61bb6cd2 GU |
3106 | int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE; |
3107 | ||
957f822a DR |
3108 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) |
3109 | { | |
e02dc017 | 3110 | return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <= |
a55c7454 | 3111 | node_reclaim_distance; |
957f822a | 3112 | } |
9276b1bc | 3113 | #else /* CONFIG_NUMA */ |
957f822a DR |
3114 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) |
3115 | { | |
3116 | return true; | |
3117 | } | |
9276b1bc PJ |
3118 | #endif /* CONFIG_NUMA */ |
3119 | ||
6bb15450 MG |
3120 | /* |
3121 | * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid | |
3122 | * fragmentation is subtle. If the preferred zone was HIGHMEM then | |
3123 | * premature use of a lower zone may cause lowmem pressure problems that | |
3124 | * are worse than fragmentation. If the next zone is ZONE_DMA then it is | |
3125 | * probably too small. It only makes sense to spread allocations to avoid | |
3126 | * fragmentation between the Normal and DMA32 zones. | |
3127 | */ | |
3128 | static inline unsigned int | |
0a79cdad | 3129 | alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask) |
6bb15450 | 3130 | { |
736838e9 | 3131 | unsigned int alloc_flags; |
0a79cdad | 3132 | |
736838e9 MN |
3133 | /* |
3134 | * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD | |
3135 | * to save a branch. | |
3136 | */ | |
3137 | alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM); | |
0a79cdad MG |
3138 | |
3139 | #ifdef CONFIG_ZONE_DMA32 | |
8139ad04 AR |
3140 | if (!zone) |
3141 | return alloc_flags; | |
3142 | ||
6bb15450 | 3143 | if (zone_idx(zone) != ZONE_NORMAL) |
8118b82e | 3144 | return alloc_flags; |
6bb15450 MG |
3145 | |
3146 | /* | |
3147 | * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and | |
3148 | * the pointer is within zone->zone_pgdat->node_zones[]. Also assume | |
3149 | * on UMA that if Normal is populated then so is DMA32. | |
3150 | */ | |
3151 | BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1); | |
3152 | if (nr_online_nodes > 1 && !populated_zone(--zone)) | |
8118b82e | 3153 | return alloc_flags; |
6bb15450 | 3154 | |
8118b82e | 3155 | alloc_flags |= ALLOC_NOFRAGMENT; |
0a79cdad MG |
3156 | #endif /* CONFIG_ZONE_DMA32 */ |
3157 | return alloc_flags; | |
6bb15450 | 3158 | } |
6bb15450 | 3159 | |
8e3560d9 PT |
3160 | /* Must be called after current_gfp_context() which can change gfp_mask */ |
3161 | static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask, | |
3162 | unsigned int alloc_flags) | |
8510e69c JK |
3163 | { |
3164 | #ifdef CONFIG_CMA | |
8e3560d9 | 3165 | if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE) |
8510e69c | 3166 | alloc_flags |= ALLOC_CMA; |
8510e69c JK |
3167 | #endif |
3168 | return alloc_flags; | |
3169 | } | |
3170 | ||
7fb1d9fc | 3171 | /* |
0798e519 | 3172 | * get_page_from_freelist goes through the zonelist trying to allocate |
7fb1d9fc RS |
3173 | * a page. |
3174 | */ | |
3175 | static struct page * | |
a9263751 VB |
3176 | get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags, |
3177 | const struct alloc_context *ac) | |
753ee728 | 3178 | { |
6bb15450 | 3179 | struct zoneref *z; |
5117f45d | 3180 | struct zone *zone; |
8a87d695 WY |
3181 | struct pglist_data *last_pgdat = NULL; |
3182 | bool last_pgdat_dirty_ok = false; | |
6bb15450 | 3183 | bool no_fallback; |
3b8c0be4 | 3184 | |
6bb15450 | 3185 | retry: |
7fb1d9fc | 3186 | /* |
9276b1bc | 3187 | * Scan zonelist, looking for a zone with enough free. |
8e464522 | 3188 | * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c. |
7fb1d9fc | 3189 | */ |
6bb15450 MG |
3190 | no_fallback = alloc_flags & ALLOC_NOFRAGMENT; |
3191 | z = ac->preferred_zoneref; | |
30d8ec73 MN |
3192 | for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx, |
3193 | ac->nodemask) { | |
be06af00 | 3194 | struct page *page; |
e085dbc5 JW |
3195 | unsigned long mark; |
3196 | ||
664eedde MG |
3197 | if (cpusets_enabled() && |
3198 | (alloc_flags & ALLOC_CPUSET) && | |
002f2906 | 3199 | !__cpuset_zone_allowed(zone, gfp_mask)) |
cd38b115 | 3200 | continue; |
a756cf59 JW |
3201 | /* |
3202 | * When allocating a page cache page for writing, we | |
281e3726 MG |
3203 | * want to get it from a node that is within its dirty |
3204 | * limit, such that no single node holds more than its | |
a756cf59 | 3205 | * proportional share of globally allowed dirty pages. |
281e3726 | 3206 | * The dirty limits take into account the node's |
a756cf59 JW |
3207 | * lowmem reserves and high watermark so that kswapd |
3208 | * should be able to balance it without having to | |
3209 | * write pages from its LRU list. | |
3210 | * | |
a756cf59 | 3211 | * XXX: For now, allow allocations to potentially |
281e3726 | 3212 | * exceed the per-node dirty limit in the slowpath |
c9ab0c4f | 3213 | * (spread_dirty_pages unset) before going into reclaim, |
a756cf59 | 3214 | * which is important when on a NUMA setup the allowed |
281e3726 | 3215 | * nodes are together not big enough to reach the |
a756cf59 | 3216 | * global limit. The proper fix for these situations |
281e3726 | 3217 | * will require awareness of nodes in the |
a756cf59 JW |
3218 | * dirty-throttling and the flusher threads. |
3219 | */ | |
3b8c0be4 | 3220 | if (ac->spread_dirty_pages) { |
8a87d695 WY |
3221 | if (last_pgdat != zone->zone_pgdat) { |
3222 | last_pgdat = zone->zone_pgdat; | |
3223 | last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat); | |
3224 | } | |
3b8c0be4 | 3225 | |
8a87d695 | 3226 | if (!last_pgdat_dirty_ok) |
3b8c0be4 | 3227 | continue; |
3b8c0be4 | 3228 | } |
7fb1d9fc | 3229 | |
6bb15450 MG |
3230 | if (no_fallback && nr_online_nodes > 1 && |
3231 | zone != ac->preferred_zoneref->zone) { | |
3232 | int local_nid; | |
3233 | ||
3234 | /* | |
3235 | * If moving to a remote node, retry but allow | |
3236 | * fragmenting fallbacks. Locality is more important | |
3237 | * than fragmentation avoidance. | |
3238 | */ | |
3239 | local_nid = zone_to_nid(ac->preferred_zoneref->zone); | |
3240 | if (zone_to_nid(zone) != local_nid) { | |
3241 | alloc_flags &= ~ALLOC_NOFRAGMENT; | |
3242 | goto retry; | |
3243 | } | |
3244 | } | |
3245 | ||
57c0419c HY |
3246 | /* |
3247 | * Detect whether the number of free pages is below high | |
3248 | * watermark. If so, we will decrease pcp->high and free | |
3249 | * PCP pages in free path to reduce the possibility of | |
3250 | * premature page reclaiming. Detection is done here to | |
3251 | * avoid to do that in hotter free path. | |
3252 | */ | |
3253 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) | |
3254 | goto check_alloc_wmark; | |
3255 | ||
3256 | mark = high_wmark_pages(zone); | |
3257 | if (zone_watermark_fast(zone, order, mark, | |
3258 | ac->highest_zoneidx, alloc_flags, | |
3259 | gfp_mask)) | |
3260 | goto try_this_zone; | |
3261 | else | |
3262 | set_bit(ZONE_BELOW_HIGH, &zone->flags); | |
3263 | ||
3264 | check_alloc_wmark: | |
a9214443 | 3265 | mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); |
48ee5f36 | 3266 | if (!zone_watermark_fast(zone, order, mark, |
f80b08fc CTR |
3267 | ac->highest_zoneidx, alloc_flags, |
3268 | gfp_mask)) { | |
fa5e084e MG |
3269 | int ret; |
3270 | ||
dcdfdd40 KS |
3271 | if (has_unaccepted_memory()) { |
3272 | if (try_to_accept_memory(zone, order)) | |
3273 | goto try_this_zone; | |
3274 | } | |
3275 | ||
c9e97a19 PT |
3276 | #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT |
3277 | /* | |
3278 | * Watermark failed for this zone, but see if we can | |
3279 | * grow this zone if it contains deferred pages. | |
3280 | */ | |
076cf7ea | 3281 | if (deferred_pages_enabled()) { |
c9e97a19 PT |
3282 | if (_deferred_grow_zone(zone, order)) |
3283 | goto try_this_zone; | |
3284 | } | |
3285 | #endif | |
5dab2911 MG |
3286 | /* Checked here to keep the fast path fast */ |
3287 | BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); | |
3288 | if (alloc_flags & ALLOC_NO_WATERMARKS) | |
3289 | goto try_this_zone; | |
3290 | ||
202e35db | 3291 | if (!node_reclaim_enabled() || |
c33d6c06 | 3292 | !zone_allows_reclaim(ac->preferred_zoneref->zone, zone)) |
cd38b115 MG |
3293 | continue; |
3294 | ||
a5f5f91d | 3295 | ret = node_reclaim(zone->zone_pgdat, gfp_mask, order); |
fa5e084e | 3296 | switch (ret) { |
a5f5f91d | 3297 | case NODE_RECLAIM_NOSCAN: |
fa5e084e | 3298 | /* did not scan */ |
cd38b115 | 3299 | continue; |
a5f5f91d | 3300 | case NODE_RECLAIM_FULL: |
fa5e084e | 3301 | /* scanned but unreclaimable */ |
cd38b115 | 3302 | continue; |
fa5e084e MG |
3303 | default: |
3304 | /* did we reclaim enough */ | |
fed2719e | 3305 | if (zone_watermark_ok(zone, order, mark, |
97a225e6 | 3306 | ac->highest_zoneidx, alloc_flags)) |
fed2719e MG |
3307 | goto try_this_zone; |
3308 | ||
fed2719e | 3309 | continue; |
0798e519 | 3310 | } |
7fb1d9fc RS |
3311 | } |
3312 | ||
fa5e084e | 3313 | try_this_zone: |
066b2393 | 3314 | page = rmqueue(ac->preferred_zoneref->zone, zone, order, |
0aaa29a5 | 3315 | gfp_mask, alloc_flags, ac->migratetype); |
75379191 | 3316 | if (page) { |
479f854a | 3317 | prep_new_page(page, order, gfp_mask, alloc_flags); |
0aaa29a5 MG |
3318 | |
3319 | /* | |
3320 | * If this is a high-order atomic allocation then check | |
3321 | * if the pageblock should be reserved for the future | |
3322 | */ | |
eb2e2b42 | 3323 | if (unlikely(alloc_flags & ALLOC_HIGHATOMIC)) |
368d983b | 3324 | reserve_highatomic_pageblock(page, zone); |
0aaa29a5 | 3325 | |
75379191 | 3326 | return page; |
c9e97a19 | 3327 | } else { |
dcdfdd40 KS |
3328 | if (has_unaccepted_memory()) { |
3329 | if (try_to_accept_memory(zone, order)) | |
3330 | goto try_this_zone; | |
3331 | } | |
3332 | ||
c9e97a19 PT |
3333 | #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT |
3334 | /* Try again if zone has deferred pages */ | |
076cf7ea | 3335 | if (deferred_pages_enabled()) { |
c9e97a19 PT |
3336 | if (_deferred_grow_zone(zone, order)) |
3337 | goto try_this_zone; | |
3338 | } | |
3339 | #endif | |
75379191 | 3340 | } |
54a6eb5c | 3341 | } |
9276b1bc | 3342 | |
6bb15450 MG |
3343 | /* |
3344 | * It's possible on a UMA machine to get through all zones that are | |
3345 | * fragmented. If avoiding fragmentation, reset and try again. | |
3346 | */ | |
3347 | if (no_fallback) { | |
3348 | alloc_flags &= ~ALLOC_NOFRAGMENT; | |
3349 | goto retry; | |
3350 | } | |
3351 | ||
4ffeaf35 | 3352 | return NULL; |
753ee728 MH |
3353 | } |
3354 | ||
9af744d7 | 3355 | static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask) |
a238ab5b | 3356 | { |
a238ab5b | 3357 | unsigned int filter = SHOW_MEM_FILTER_NODES; |
a238ab5b DH |
3358 | |
3359 | /* | |
3360 | * This documents exceptions given to allocations in certain | |
3361 | * contexts that are allowed to allocate outside current's set | |
3362 | * of allowed nodes. | |
3363 | */ | |
3364 | if (!(gfp_mask & __GFP_NOMEMALLOC)) | |
cd04ae1e | 3365 | if (tsk_is_oom_victim(current) || |
a238ab5b DH |
3366 | (current->flags & (PF_MEMALLOC | PF_EXITING))) |
3367 | filter &= ~SHOW_MEM_FILTER_NODES; | |
88dc6f20 | 3368 | if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM)) |
a238ab5b DH |
3369 | filter &= ~SHOW_MEM_FILTER_NODES; |
3370 | ||
974f4367 | 3371 | __show_mem(filter, nodemask, gfp_zone(gfp_mask)); |
aa187507 MH |
3372 | } |
3373 | ||
a8e99259 | 3374 | void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...) |
aa187507 MH |
3375 | { |
3376 | struct va_format vaf; | |
3377 | va_list args; | |
1be334e5 | 3378 | static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1); |
aa187507 | 3379 | |
c4dc63f0 BH |
3380 | if ((gfp_mask & __GFP_NOWARN) || |
3381 | !__ratelimit(&nopage_rs) || | |
3382 | ((gfp_mask & __GFP_DMA) && !has_managed_dma())) | |
aa187507 MH |
3383 | return; |
3384 | ||
7877cdcc MH |
3385 | va_start(args, fmt); |
3386 | vaf.fmt = fmt; | |
3387 | vaf.va = &args; | |
ef8444ea | 3388 | pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl", |
0205f755 MH |
3389 | current->comm, &vaf, gfp_mask, &gfp_mask, |
3390 | nodemask_pr_args(nodemask)); | |
7877cdcc | 3391 | va_end(args); |
3ee9a4f0 | 3392 | |
a8e99259 | 3393 | cpuset_print_current_mems_allowed(); |
ef8444ea | 3394 | pr_cont("\n"); |
a238ab5b | 3395 | dump_stack(); |
685dbf6f | 3396 | warn_alloc_show_mem(gfp_mask, nodemask); |
a238ab5b DH |
3397 | } |
3398 | ||
6c18ba7a MH |
3399 | static inline struct page * |
3400 | __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order, | |
3401 | unsigned int alloc_flags, | |
3402 | const struct alloc_context *ac) | |
3403 | { | |
3404 | struct page *page; | |
3405 | ||
3406 | page = get_page_from_freelist(gfp_mask, order, | |
3407 | alloc_flags|ALLOC_CPUSET, ac); | |
3408 | /* | |
3409 | * fallback to ignore cpuset restriction if our nodes | |
3410 | * are depleted | |
3411 | */ | |
3412 | if (!page) | |
3413 | page = get_page_from_freelist(gfp_mask, order, | |
3414 | alloc_flags, ac); | |
3415 | ||
3416 | return page; | |
3417 | } | |
3418 | ||
11e33f6a MG |
3419 | static inline struct page * |
3420 | __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, | |
a9263751 | 3421 | const struct alloc_context *ac, unsigned long *did_some_progress) |
11e33f6a | 3422 | { |
6e0fc46d DR |
3423 | struct oom_control oc = { |
3424 | .zonelist = ac->zonelist, | |
3425 | .nodemask = ac->nodemask, | |
2a966b77 | 3426 | .memcg = NULL, |
6e0fc46d DR |
3427 | .gfp_mask = gfp_mask, |
3428 | .order = order, | |
6e0fc46d | 3429 | }; |
11e33f6a MG |
3430 | struct page *page; |
3431 | ||
9879de73 JW |
3432 | *did_some_progress = 0; |
3433 | ||
9879de73 | 3434 | /* |
dc56401f JW |
3435 | * Acquire the oom lock. If that fails, somebody else is |
3436 | * making progress for us. | |
9879de73 | 3437 | */ |
dc56401f | 3438 | if (!mutex_trylock(&oom_lock)) { |
9879de73 | 3439 | *did_some_progress = 1; |
11e33f6a | 3440 | schedule_timeout_uninterruptible(1); |
1da177e4 LT |
3441 | return NULL; |
3442 | } | |
6b1de916 | 3443 | |
11e33f6a MG |
3444 | /* |
3445 | * Go through the zonelist yet one more time, keep very high watermark | |
3446 | * here, this is only to catch a parallel oom killing, we must fail if | |
e746bf73 TH |
3447 | * we're still under heavy pressure. But make sure that this reclaim |
3448 | * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY | |
3449 | * allocation which will never fail due to oom_lock already held. | |
11e33f6a | 3450 | */ |
e746bf73 TH |
3451 | page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) & |
3452 | ~__GFP_DIRECT_RECLAIM, order, | |
3453 | ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac); | |
7fb1d9fc | 3454 | if (page) |
11e33f6a MG |
3455 | goto out; |
3456 | ||
06ad276a MH |
3457 | /* Coredumps can quickly deplete all memory reserves */ |
3458 | if (current->flags & PF_DUMPCORE) | |
3459 | goto out; | |
3460 | /* The OOM killer will not help higher order allocs */ | |
3461 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
3462 | goto out; | |
dcda9b04 MH |
3463 | /* |
3464 | * We have already exhausted all our reclaim opportunities without any | |
3465 | * success so it is time to admit defeat. We will skip the OOM killer | |
3466 | * because it is very likely that the caller has a more reasonable | |
3467 | * fallback than shooting a random task. | |
cfb4a541 MN |
3468 | * |
3469 | * The OOM killer may not free memory on a specific node. | |
dcda9b04 | 3470 | */ |
cfb4a541 | 3471 | if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE)) |
dcda9b04 | 3472 | goto out; |
06ad276a | 3473 | /* The OOM killer does not needlessly kill tasks for lowmem */ |
97a225e6 | 3474 | if (ac->highest_zoneidx < ZONE_NORMAL) |
06ad276a MH |
3475 | goto out; |
3476 | if (pm_suspended_storage()) | |
3477 | goto out; | |
3478 | /* | |
3479 | * XXX: GFP_NOFS allocations should rather fail than rely on | |
3480 | * other request to make a forward progress. | |
3481 | * We are in an unfortunate situation where out_of_memory cannot | |
3482 | * do much for this context but let's try it to at least get | |
3483 | * access to memory reserved if the current task is killed (see | |
3484 | * out_of_memory). Once filesystems are ready to handle allocation | |
3485 | * failures more gracefully we should just bail out here. | |
3486 | */ | |
3487 | ||
3c2c6488 | 3488 | /* Exhausted what can be done so it's blame time */ |
3f913fc5 QZ |
3489 | if (out_of_memory(&oc) || |
3490 | WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) { | |
c32b3cbe | 3491 | *did_some_progress = 1; |
5020e285 | 3492 | |
6c18ba7a MH |
3493 | /* |
3494 | * Help non-failing allocations by giving them access to memory | |
3495 | * reserves | |
3496 | */ | |
3497 | if (gfp_mask & __GFP_NOFAIL) | |
3498 | page = __alloc_pages_cpuset_fallback(gfp_mask, order, | |
5020e285 | 3499 | ALLOC_NO_WATERMARKS, ac); |
5020e285 | 3500 | } |
11e33f6a | 3501 | out: |
dc56401f | 3502 | mutex_unlock(&oom_lock); |
11e33f6a MG |
3503 | return page; |
3504 | } | |
3505 | ||
33c2d214 | 3506 | /* |
baf2f90b | 3507 | * Maximum number of compaction retries with a progress before OOM |
33c2d214 MH |
3508 | * killer is consider as the only way to move forward. |
3509 | */ | |
3510 | #define MAX_COMPACT_RETRIES 16 | |
3511 | ||
56de7263 MG |
3512 | #ifdef CONFIG_COMPACTION |
3513 | /* Try memory compaction for high-order allocations before reclaim */ | |
3514 | static struct page * | |
3515 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, | |
c603844b | 3516 | unsigned int alloc_flags, const struct alloc_context *ac, |
a5508cd8 | 3517 | enum compact_priority prio, enum compact_result *compact_result) |
56de7263 | 3518 | { |
5e1f0f09 | 3519 | struct page *page = NULL; |
eb414681 | 3520 | unsigned long pflags; |
499118e9 | 3521 | unsigned int noreclaim_flag; |
53853e2d VB |
3522 | |
3523 | if (!order) | |
66199712 | 3524 | return NULL; |
66199712 | 3525 | |
eb414681 | 3526 | psi_memstall_enter(&pflags); |
5bf18281 | 3527 | delayacct_compact_start(); |
499118e9 | 3528 | noreclaim_flag = memalloc_noreclaim_save(); |
eb414681 | 3529 | |
c5d01d0d | 3530 | *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac, |
5e1f0f09 | 3531 | prio, &page); |
eb414681 | 3532 | |
499118e9 | 3533 | memalloc_noreclaim_restore(noreclaim_flag); |
eb414681 | 3534 | psi_memstall_leave(&pflags); |
5bf18281 | 3535 | delayacct_compact_end(); |
56de7263 | 3536 | |
06dac2f4 CTR |
3537 | if (*compact_result == COMPACT_SKIPPED) |
3538 | return NULL; | |
98dd3b48 VB |
3539 | /* |
3540 | * At least in one zone compaction wasn't deferred or skipped, so let's | |
3541 | * count a compaction stall | |
3542 | */ | |
3543 | count_vm_event(COMPACTSTALL); | |
8fb74b9f | 3544 | |
5e1f0f09 MG |
3545 | /* Prep a captured page if available */ |
3546 | if (page) | |
3547 | prep_new_page(page, order, gfp_mask, alloc_flags); | |
3548 | ||
3549 | /* Try get a page from the freelist if available */ | |
3550 | if (!page) | |
3551 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); | |
53853e2d | 3552 | |
98dd3b48 VB |
3553 | if (page) { |
3554 | struct zone *zone = page_zone(page); | |
53853e2d | 3555 | |
98dd3b48 VB |
3556 | zone->compact_blockskip_flush = false; |
3557 | compaction_defer_reset(zone, order, true); | |
3558 | count_vm_event(COMPACTSUCCESS); | |
3559 | return page; | |
3560 | } | |
56de7263 | 3561 | |
98dd3b48 VB |
3562 | /* |
3563 | * It's bad if compaction run occurs and fails. The most likely reason | |
3564 | * is that pages exist, but not enough to satisfy watermarks. | |
3565 | */ | |
3566 | count_vm_event(COMPACTFAIL); | |
66199712 | 3567 | |
98dd3b48 | 3568 | cond_resched(); |
56de7263 MG |
3569 | |
3570 | return NULL; | |
3571 | } | |
33c2d214 | 3572 | |
3250845d VB |
3573 | static inline bool |
3574 | should_compact_retry(struct alloc_context *ac, int order, int alloc_flags, | |
3575 | enum compact_result compact_result, | |
3576 | enum compact_priority *compact_priority, | |
d9436498 | 3577 | int *compaction_retries) |
3250845d VB |
3578 | { |
3579 | int max_retries = MAX_COMPACT_RETRIES; | |
c2033b00 | 3580 | int min_priority; |
65190cff MH |
3581 | bool ret = false; |
3582 | int retries = *compaction_retries; | |
3583 | enum compact_priority priority = *compact_priority; | |
3250845d VB |
3584 | |
3585 | if (!order) | |
3586 | return false; | |
3587 | ||
691d9497 AT |
3588 | if (fatal_signal_pending(current)) |
3589 | return false; | |
3590 | ||
49433085 | 3591 | /* |
ecd8b292 JW |
3592 | * Compaction was skipped due to a lack of free order-0 |
3593 | * migration targets. Continue if reclaim can help. | |
49433085 | 3594 | */ |
ecd8b292 | 3595 | if (compact_result == COMPACT_SKIPPED) { |
49433085 VB |
3596 | ret = compaction_zonelist_suitable(ac, order, alloc_flags); |
3597 | goto out; | |
3598 | } | |
3599 | ||
3250845d | 3600 | /* |
511a69b2 JW |
3601 | * Compaction managed to coalesce some page blocks, but the |
3602 | * allocation failed presumably due to a race. Retry some. | |
3250845d | 3603 | */ |
511a69b2 JW |
3604 | if (compact_result == COMPACT_SUCCESS) { |
3605 | /* | |
3606 | * !costly requests are much more important than | |
3607 | * __GFP_RETRY_MAYFAIL costly ones because they are de | |
3608 | * facto nofail and invoke OOM killer to move on while | |
3609 | * costly can fail and users are ready to cope with | |
3610 | * that. 1/4 retries is rather arbitrary but we would | |
3611 | * need much more detailed feedback from compaction to | |
3612 | * make a better decision. | |
3613 | */ | |
3614 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
3615 | max_retries /= 4; | |
3250845d | 3616 | |
511a69b2 JW |
3617 | if (++(*compaction_retries) <= max_retries) { |
3618 | ret = true; | |
3619 | goto out; | |
3620 | } | |
65190cff | 3621 | } |
3250845d | 3622 | |
d9436498 | 3623 | /* |
511a69b2 | 3624 | * Compaction failed. Retry with increasing priority. |
d9436498 | 3625 | */ |
c2033b00 VB |
3626 | min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
3627 | MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY; | |
65190cff | 3628 | |
c2033b00 | 3629 | if (*compact_priority > min_priority) { |
d9436498 VB |
3630 | (*compact_priority)--; |
3631 | *compaction_retries = 0; | |
65190cff | 3632 | ret = true; |
d9436498 | 3633 | } |
65190cff MH |
3634 | out: |
3635 | trace_compact_retry(order, priority, compact_result, retries, max_retries, ret); | |
3636 | return ret; | |
3250845d | 3637 | } |
56de7263 MG |
3638 | #else |
3639 | static inline struct page * | |
3640 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, | |
c603844b | 3641 | unsigned int alloc_flags, const struct alloc_context *ac, |
a5508cd8 | 3642 | enum compact_priority prio, enum compact_result *compact_result) |
56de7263 | 3643 | { |
33c2d214 | 3644 | *compact_result = COMPACT_SKIPPED; |
56de7263 MG |
3645 | return NULL; |
3646 | } | |
33c2d214 MH |
3647 | |
3648 | static inline bool | |
86a294a8 MH |
3649 | should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags, |
3650 | enum compact_result compact_result, | |
a5508cd8 | 3651 | enum compact_priority *compact_priority, |
d9436498 | 3652 | int *compaction_retries) |
33c2d214 | 3653 | { |
31e49bfd MH |
3654 | struct zone *zone; |
3655 | struct zoneref *z; | |
3656 | ||
3657 | if (!order || order > PAGE_ALLOC_COSTLY_ORDER) | |
3658 | return false; | |
3659 | ||
3660 | /* | |
3661 | * There are setups with compaction disabled which would prefer to loop | |
3662 | * inside the allocator rather than hit the oom killer prematurely. | |
3663 | * Let's give them a good hope and keep retrying while the order-0 | |
3664 | * watermarks are OK. | |
3665 | */ | |
97a225e6 JK |
3666 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
3667 | ac->highest_zoneidx, ac->nodemask) { | |
31e49bfd | 3668 | if (zone_watermark_ok(zone, 0, min_wmark_pages(zone), |
97a225e6 | 3669 | ac->highest_zoneidx, alloc_flags)) |
31e49bfd MH |
3670 | return true; |
3671 | } | |
33c2d214 MH |
3672 | return false; |
3673 | } | |
3250845d | 3674 | #endif /* CONFIG_COMPACTION */ |
56de7263 | 3675 | |
d92a8cfc | 3676 | #ifdef CONFIG_LOCKDEP |
93781325 | 3677 | static struct lockdep_map __fs_reclaim_map = |
d92a8cfc PZ |
3678 | STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map); |
3679 | ||
f920e413 | 3680 | static bool __need_reclaim(gfp_t gfp_mask) |
d92a8cfc | 3681 | { |
d92a8cfc PZ |
3682 | /* no reclaim without waiting on it */ |
3683 | if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) | |
3684 | return false; | |
3685 | ||
3686 | /* this guy won't enter reclaim */ | |
2e517d68 | 3687 | if (current->flags & PF_MEMALLOC) |
d92a8cfc PZ |
3688 | return false; |
3689 | ||
d92a8cfc PZ |
3690 | if (gfp_mask & __GFP_NOLOCKDEP) |
3691 | return false; | |
3692 | ||
3693 | return true; | |
3694 | } | |
3695 | ||
4f3eaf45 | 3696 | void __fs_reclaim_acquire(unsigned long ip) |
93781325 | 3697 | { |
4f3eaf45 | 3698 | lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip); |
93781325 OS |
3699 | } |
3700 | ||
4f3eaf45 | 3701 | void __fs_reclaim_release(unsigned long ip) |
93781325 | 3702 | { |
4f3eaf45 | 3703 | lock_release(&__fs_reclaim_map, ip); |
93781325 OS |
3704 | } |
3705 | ||
d92a8cfc PZ |
3706 | void fs_reclaim_acquire(gfp_t gfp_mask) |
3707 | { | |
f920e413 DV |
3708 | gfp_mask = current_gfp_context(gfp_mask); |
3709 | ||
3710 | if (__need_reclaim(gfp_mask)) { | |
3711 | if (gfp_mask & __GFP_FS) | |
4f3eaf45 | 3712 | __fs_reclaim_acquire(_RET_IP_); |
f920e413 DV |
3713 | |
3714 | #ifdef CONFIG_MMU_NOTIFIER | |
3715 | lock_map_acquire(&__mmu_notifier_invalidate_range_start_map); | |
3716 | lock_map_release(&__mmu_notifier_invalidate_range_start_map); | |
3717 | #endif | |
3718 | ||
3719 | } | |
d92a8cfc PZ |
3720 | } |
3721 | EXPORT_SYMBOL_GPL(fs_reclaim_acquire); | |
3722 | ||
3723 | void fs_reclaim_release(gfp_t gfp_mask) | |
3724 | { | |
f920e413 DV |
3725 | gfp_mask = current_gfp_context(gfp_mask); |
3726 | ||
3727 | if (__need_reclaim(gfp_mask)) { | |
3728 | if (gfp_mask & __GFP_FS) | |
4f3eaf45 | 3729 | __fs_reclaim_release(_RET_IP_); |
f920e413 | 3730 | } |
d92a8cfc PZ |
3731 | } |
3732 | EXPORT_SYMBOL_GPL(fs_reclaim_release); | |
3733 | #endif | |
3734 | ||
3d36424b MG |
3735 | /* |
3736 | * Zonelists may change due to hotplug during allocation. Detect when zonelists | |
3737 | * have been rebuilt so allocation retries. Reader side does not lock and | |
3738 | * retries the allocation if zonelist changes. Writer side is protected by the | |
3739 | * embedded spin_lock. | |
3740 | */ | |
3741 | static DEFINE_SEQLOCK(zonelist_update_seq); | |
3742 | ||
3743 | static unsigned int zonelist_iter_begin(void) | |
3744 | { | |
3745 | if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) | |
3746 | return read_seqbegin(&zonelist_update_seq); | |
3747 | ||
3748 | return 0; | |
3749 | } | |
3750 | ||
3751 | static unsigned int check_retry_zonelist(unsigned int seq) | |
3752 | { | |
3753 | if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) | |
3754 | return read_seqretry(&zonelist_update_seq, seq); | |
3755 | ||
3756 | return seq; | |
3757 | } | |
3758 | ||
bba90710 | 3759 | /* Perform direct synchronous page reclaim */ |
2187e17b | 3760 | static unsigned long |
a9263751 VB |
3761 | __perform_reclaim(gfp_t gfp_mask, unsigned int order, |
3762 | const struct alloc_context *ac) | |
11e33f6a | 3763 | { |
499118e9 | 3764 | unsigned int noreclaim_flag; |
fa7fc75f | 3765 | unsigned long progress; |
11e33f6a MG |
3766 | |
3767 | cond_resched(); | |
3768 | ||
3769 | /* We now go into synchronous reclaim */ | |
3770 | cpuset_memory_pressure_bump(); | |
d92a8cfc | 3771 | fs_reclaim_acquire(gfp_mask); |
93781325 | 3772 | noreclaim_flag = memalloc_noreclaim_save(); |
11e33f6a | 3773 | |
a9263751 VB |
3774 | progress = try_to_free_pages(ac->zonelist, order, gfp_mask, |
3775 | ac->nodemask); | |
11e33f6a | 3776 | |
499118e9 | 3777 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 3778 | fs_reclaim_release(gfp_mask); |
11e33f6a MG |
3779 | |
3780 | cond_resched(); | |
3781 | ||
bba90710 MS |
3782 | return progress; |
3783 | } | |
3784 | ||
3785 | /* The really slow allocator path where we enter direct reclaim */ | |
3786 | static inline struct page * | |
3787 | __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, | |
c603844b | 3788 | unsigned int alloc_flags, const struct alloc_context *ac, |
a9263751 | 3789 | unsigned long *did_some_progress) |
bba90710 MS |
3790 | { |
3791 | struct page *page = NULL; | |
fa7fc75f | 3792 | unsigned long pflags; |
bba90710 MS |
3793 | bool drained = false; |
3794 | ||
fa7fc75f | 3795 | psi_memstall_enter(&pflags); |
a9263751 | 3796 | *did_some_progress = __perform_reclaim(gfp_mask, order, ac); |
9ee493ce | 3797 | if (unlikely(!(*did_some_progress))) |
fa7fc75f | 3798 | goto out; |
11e33f6a | 3799 | |
9ee493ce | 3800 | retry: |
31a6c190 | 3801 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); |
9ee493ce MG |
3802 | |
3803 | /* | |
3804 | * If an allocation failed after direct reclaim, it could be because | |
0aaa29a5 | 3805 | * pages are pinned on the per-cpu lists or in high alloc reserves. |
047b9967 | 3806 | * Shrink them and try again |
9ee493ce MG |
3807 | */ |
3808 | if (!page && !drained) { | |
29fac03b | 3809 | unreserve_highatomic_pageblock(ac, false); |
93481ff0 | 3810 | drain_all_pages(NULL); |
9ee493ce MG |
3811 | drained = true; |
3812 | goto retry; | |
3813 | } | |
fa7fc75f SB |
3814 | out: |
3815 | psi_memstall_leave(&pflags); | |
9ee493ce | 3816 | |
11e33f6a MG |
3817 | return page; |
3818 | } | |
3819 | ||
5ecd9d40 DR |
3820 | static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask, |
3821 | const struct alloc_context *ac) | |
3a025760 JW |
3822 | { |
3823 | struct zoneref *z; | |
3824 | struct zone *zone; | |
e1a55637 | 3825 | pg_data_t *last_pgdat = NULL; |
97a225e6 | 3826 | enum zone_type highest_zoneidx = ac->highest_zoneidx; |
3a025760 | 3827 | |
97a225e6 | 3828 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx, |
5ecd9d40 | 3829 | ac->nodemask) { |
bc53008e WY |
3830 | if (!managed_zone(zone)) |
3831 | continue; | |
d137a7cb | 3832 | if (last_pgdat != zone->zone_pgdat) { |
97a225e6 | 3833 | wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx); |
d137a7cb CW |
3834 | last_pgdat = zone->zone_pgdat; |
3835 | } | |
e1a55637 | 3836 | } |
3a025760 JW |
3837 | } |
3838 | ||
c603844b | 3839 | static inline unsigned int |
eb2e2b42 | 3840 | gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order) |
341ce06f | 3841 | { |
c603844b | 3842 | unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; |
1da177e4 | 3843 | |
736838e9 | 3844 | /* |
524c4807 | 3845 | * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE |
736838e9 MN |
3846 | * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD |
3847 | * to save two branches. | |
3848 | */ | |
524c4807 | 3849 | BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE); |
736838e9 | 3850 | BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD); |
933e312e | 3851 | |
341ce06f PZ |
3852 | /* |
3853 | * The caller may dip into page reserves a bit more if the caller | |
3854 | * cannot run direct reclaim, or if the caller has realtime scheduling | |
3855 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will | |
1ebbb218 | 3856 | * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH). |
341ce06f | 3857 | */ |
736838e9 MN |
3858 | alloc_flags |= (__force int) |
3859 | (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM)); | |
1da177e4 | 3860 | |
1ebbb218 | 3861 | if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) { |
5c3240d9 | 3862 | /* |
b104a35d DR |
3863 | * Not worth trying to allocate harder for __GFP_NOMEMALLOC even |
3864 | * if it can't schedule. | |
5c3240d9 | 3865 | */ |
eb2e2b42 | 3866 | if (!(gfp_mask & __GFP_NOMEMALLOC)) { |
1ebbb218 | 3867 | alloc_flags |= ALLOC_NON_BLOCK; |
eb2e2b42 MG |
3868 | |
3869 | if (order > 0) | |
3870 | alloc_flags |= ALLOC_HIGHATOMIC; | |
3871 | } | |
3872 | ||
523b9458 | 3873 | /* |
1ebbb218 MG |
3874 | * Ignore cpuset mems for non-blocking __GFP_HIGH (probably |
3875 | * GFP_ATOMIC) rather than fail, see the comment for | |
8e464522 | 3876 | * cpuset_node_allowed(). |
523b9458 | 3877 | */ |
1ebbb218 MG |
3878 | if (alloc_flags & ALLOC_MIN_RESERVE) |
3879 | alloc_flags &= ~ALLOC_CPUSET; | |
88dc6f20 | 3880 | } else if (unlikely(rt_task(current)) && in_task()) |
c988dcbe | 3881 | alloc_flags |= ALLOC_MIN_RESERVE; |
341ce06f | 3882 | |
8e3560d9 | 3883 | alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags); |
8510e69c | 3884 | |
341ce06f PZ |
3885 | return alloc_flags; |
3886 | } | |
3887 | ||
cd04ae1e | 3888 | static bool oom_reserves_allowed(struct task_struct *tsk) |
072bb0aa | 3889 | { |
cd04ae1e MH |
3890 | if (!tsk_is_oom_victim(tsk)) |
3891 | return false; | |
3892 | ||
3893 | /* | |
3894 | * !MMU doesn't have oom reaper so give access to memory reserves | |
3895 | * only to the thread with TIF_MEMDIE set | |
3896 | */ | |
3897 | if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE)) | |
31a6c190 VB |
3898 | return false; |
3899 | ||
cd04ae1e MH |
3900 | return true; |
3901 | } | |
3902 | ||
3903 | /* | |
3904 | * Distinguish requests which really need access to full memory | |
3905 | * reserves from oom victims which can live with a portion of it | |
3906 | */ | |
3907 | static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask) | |
3908 | { | |
3909 | if (unlikely(gfp_mask & __GFP_NOMEMALLOC)) | |
3910 | return 0; | |
31a6c190 | 3911 | if (gfp_mask & __GFP_MEMALLOC) |
cd04ae1e | 3912 | return ALLOC_NO_WATERMARKS; |
31a6c190 | 3913 | if (in_serving_softirq() && (current->flags & PF_MEMALLOC)) |
cd04ae1e MH |
3914 | return ALLOC_NO_WATERMARKS; |
3915 | if (!in_interrupt()) { | |
3916 | if (current->flags & PF_MEMALLOC) | |
3917 | return ALLOC_NO_WATERMARKS; | |
3918 | else if (oom_reserves_allowed(current)) | |
3919 | return ALLOC_OOM; | |
3920 | } | |
31a6c190 | 3921 | |
cd04ae1e MH |
3922 | return 0; |
3923 | } | |
3924 | ||
3925 | bool gfp_pfmemalloc_allowed(gfp_t gfp_mask) | |
3926 | { | |
3927 | return !!__gfp_pfmemalloc_flags(gfp_mask); | |
072bb0aa MG |
3928 | } |
3929 | ||
0a0337e0 MH |
3930 | /* |
3931 | * Checks whether it makes sense to retry the reclaim to make a forward progress | |
3932 | * for the given allocation request. | |
491d79ae JW |
3933 | * |
3934 | * We give up when we either have tried MAX_RECLAIM_RETRIES in a row | |
3935 | * without success, or when we couldn't even meet the watermark if we | |
3936 | * reclaimed all remaining pages on the LRU lists. | |
0a0337e0 MH |
3937 | * |
3938 | * Returns true if a retry is viable or false to enter the oom path. | |
3939 | */ | |
3940 | static inline bool | |
3941 | should_reclaim_retry(gfp_t gfp_mask, unsigned order, | |
3942 | struct alloc_context *ac, int alloc_flags, | |
423b452e | 3943 | bool did_some_progress, int *no_progress_loops) |
0a0337e0 MH |
3944 | { |
3945 | struct zone *zone; | |
3946 | struct zoneref *z; | |
15f570bf | 3947 | bool ret = false; |
0a0337e0 | 3948 | |
423b452e VB |
3949 | /* |
3950 | * Costly allocations might have made a progress but this doesn't mean | |
3951 | * their order will become available due to high fragmentation so | |
3952 | * always increment the no progress counter for them | |
3953 | */ | |
3954 | if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) | |
3955 | *no_progress_loops = 0; | |
3956 | else | |
3957 | (*no_progress_loops)++; | |
3958 | ||
ac3f3b0a CTK |
3959 | if (*no_progress_loops > MAX_RECLAIM_RETRIES) |
3960 | goto out; | |
3961 | ||
0a0337e0 | 3962 | |
bca67592 MG |
3963 | /* |
3964 | * Keep reclaiming pages while there is a chance this will lead | |
3965 | * somewhere. If none of the target zones can satisfy our allocation | |
3966 | * request even if all reclaimable pages are considered then we are | |
3967 | * screwed and have to go OOM. | |
0a0337e0 | 3968 | */ |
97a225e6 JK |
3969 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
3970 | ac->highest_zoneidx, ac->nodemask) { | |
0a0337e0 | 3971 | unsigned long available; |
ede37713 | 3972 | unsigned long reclaimable; |
d379f01d MH |
3973 | unsigned long min_wmark = min_wmark_pages(zone); |
3974 | bool wmark; | |
0a0337e0 | 3975 | |
5a1c84b4 | 3976 | available = reclaimable = zone_reclaimable_pages(zone); |
5a1c84b4 | 3977 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
0a0337e0 MH |
3978 | |
3979 | /* | |
491d79ae JW |
3980 | * Would the allocation succeed if we reclaimed all |
3981 | * reclaimable pages? | |
0a0337e0 | 3982 | */ |
d379f01d | 3983 | wmark = __zone_watermark_ok(zone, order, min_wmark, |
97a225e6 | 3984 | ac->highest_zoneidx, alloc_flags, available); |
d379f01d MH |
3985 | trace_reclaim_retry_zone(z, order, reclaimable, |
3986 | available, min_wmark, *no_progress_loops, wmark); | |
3987 | if (wmark) { | |
15f570bf | 3988 | ret = true; |
132b0d21 | 3989 | break; |
0a0337e0 MH |
3990 | } |
3991 | } | |
3992 | ||
15f570bf MH |
3993 | /* |
3994 | * Memory allocation/reclaim might be called from a WQ context and the | |
3995 | * current implementation of the WQ concurrency control doesn't | |
3996 | * recognize that a particular WQ is congested if the worker thread is | |
3997 | * looping without ever sleeping. Therefore we have to do a short sleep | |
3998 | * here rather than calling cond_resched(). | |
3999 | */ | |
4000 | if (current->flags & PF_WQ_WORKER) | |
4001 | schedule_timeout_uninterruptible(1); | |
4002 | else | |
4003 | cond_resched(); | |
ac3f3b0a CTK |
4004 | out: |
4005 | /* Before OOM, exhaust highatomic_reserve */ | |
4006 | if (!ret) | |
4007 | return unreserve_highatomic_pageblock(ac, true); | |
4008 | ||
15f570bf | 4009 | return ret; |
0a0337e0 MH |
4010 | } |
4011 | ||
902b6281 VB |
4012 | static inline bool |
4013 | check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac) | |
4014 | { | |
4015 | /* | |
4016 | * It's possible that cpuset's mems_allowed and the nodemask from | |
4017 | * mempolicy don't intersect. This should be normally dealt with by | |
4018 | * policy_nodemask(), but it's possible to race with cpuset update in | |
4019 | * such a way the check therein was true, and then it became false | |
4020 | * before we got our cpuset_mems_cookie here. | |
4021 | * This assumes that for all allocations, ac->nodemask can come only | |
4022 | * from MPOL_BIND mempolicy (whose documented semantics is to be ignored | |
4023 | * when it does not intersect with the cpuset restrictions) or the | |
4024 | * caller can deal with a violated nodemask. | |
4025 | */ | |
4026 | if (cpusets_enabled() && ac->nodemask && | |
4027 | !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) { | |
4028 | ac->nodemask = NULL; | |
4029 | return true; | |
4030 | } | |
4031 | ||
4032 | /* | |
4033 | * When updating a task's mems_allowed or mempolicy nodemask, it is | |
4034 | * possible to race with parallel threads in such a way that our | |
4035 | * allocation can fail while the mask is being updated. If we are about | |
4036 | * to fail, check if the cpuset changed during allocation and if so, | |
4037 | * retry. | |
4038 | */ | |
4039 | if (read_mems_allowed_retry(cpuset_mems_cookie)) | |
4040 | return true; | |
4041 | ||
4042 | return false; | |
4043 | } | |
4044 | ||
11e33f6a MG |
4045 | static inline struct page * |
4046 | __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, | |
a9263751 | 4047 | struct alloc_context *ac) |
11e33f6a | 4048 | { |
d0164adc | 4049 | bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM; |
803de900 | 4050 | bool can_compact = gfp_compaction_allowed(gfp_mask); |
282722b0 | 4051 | const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER; |
11e33f6a | 4052 | struct page *page = NULL; |
c603844b | 4053 | unsigned int alloc_flags; |
11e33f6a | 4054 | unsigned long did_some_progress; |
5ce9bfef | 4055 | enum compact_priority compact_priority; |
c5d01d0d | 4056 | enum compact_result compact_result; |
5ce9bfef VB |
4057 | int compaction_retries; |
4058 | int no_progress_loops; | |
5ce9bfef | 4059 | unsigned int cpuset_mems_cookie; |
3d36424b | 4060 | unsigned int zonelist_iter_cookie; |
cd04ae1e | 4061 | int reserve_flags; |
1da177e4 | 4062 | |
3d36424b | 4063 | restart: |
5ce9bfef VB |
4064 | compaction_retries = 0; |
4065 | no_progress_loops = 0; | |
4066 | compact_priority = DEF_COMPACT_PRIORITY; | |
4067 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
3d36424b | 4068 | zonelist_iter_cookie = zonelist_iter_begin(); |
9a67f648 MH |
4069 | |
4070 | /* | |
4071 | * The fast path uses conservative alloc_flags to succeed only until | |
4072 | * kswapd needs to be woken up, and to avoid the cost of setting up | |
4073 | * alloc_flags precisely. So we do that now. | |
4074 | */ | |
eb2e2b42 | 4075 | alloc_flags = gfp_to_alloc_flags(gfp_mask, order); |
9a67f648 | 4076 | |
e47483bc VB |
4077 | /* |
4078 | * We need to recalculate the starting point for the zonelist iterator | |
4079 | * because we might have used different nodemask in the fast path, or | |
4080 | * there was a cpuset modification and we are retrying - otherwise we | |
4081 | * could end up iterating over non-eligible zones endlessly. | |
4082 | */ | |
4083 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist, | |
97a225e6 | 4084 | ac->highest_zoneidx, ac->nodemask); |
e47483bc VB |
4085 | if (!ac->preferred_zoneref->zone) |
4086 | goto nopage; | |
4087 | ||
8ca1b5a4 FT |
4088 | /* |
4089 | * Check for insane configurations where the cpuset doesn't contain | |
4090 | * any suitable zone to satisfy the request - e.g. non-movable | |
4091 | * GFP_HIGHUSER allocations from MOVABLE nodes only. | |
4092 | */ | |
4093 | if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) { | |
4094 | struct zoneref *z = first_zones_zonelist(ac->zonelist, | |
4095 | ac->highest_zoneidx, | |
4096 | &cpuset_current_mems_allowed); | |
4097 | if (!z->zone) | |
4098 | goto nopage; | |
4099 | } | |
4100 | ||
0a79cdad | 4101 | if (alloc_flags & ALLOC_KSWAPD) |
5ecd9d40 | 4102 | wake_all_kswapds(order, gfp_mask, ac); |
23771235 VB |
4103 | |
4104 | /* | |
4105 | * The adjusted alloc_flags might result in immediate success, so try | |
4106 | * that first | |
4107 | */ | |
4108 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); | |
4109 | if (page) | |
4110 | goto got_pg; | |
4111 | ||
a8161d1e VB |
4112 | /* |
4113 | * For costly allocations, try direct compaction first, as it's likely | |
282722b0 VB |
4114 | * that we have enough base pages and don't need to reclaim. For non- |
4115 | * movable high-order allocations, do that as well, as compaction will | |
4116 | * try prevent permanent fragmentation by migrating from blocks of the | |
4117 | * same migratetype. | |
4118 | * Don't try this for allocations that are allowed to ignore | |
4119 | * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen. | |
a8161d1e | 4120 | */ |
803de900 | 4121 | if (can_direct_reclaim && can_compact && |
282722b0 VB |
4122 | (costly_order || |
4123 | (order > 0 && ac->migratetype != MIGRATE_MOVABLE)) | |
4124 | && !gfp_pfmemalloc_allowed(gfp_mask)) { | |
a8161d1e VB |
4125 | page = __alloc_pages_direct_compact(gfp_mask, order, |
4126 | alloc_flags, ac, | |
a5508cd8 | 4127 | INIT_COMPACT_PRIORITY, |
a8161d1e VB |
4128 | &compact_result); |
4129 | if (page) | |
4130 | goto got_pg; | |
4131 | ||
cc638f32 VB |
4132 | /* |
4133 | * Checks for costly allocations with __GFP_NORETRY, which | |
4134 | * includes some THP page fault allocations | |
4135 | */ | |
4136 | if (costly_order && (gfp_mask & __GFP_NORETRY)) { | |
b39d0ee2 DR |
4137 | /* |
4138 | * If allocating entire pageblock(s) and compaction | |
4139 | * failed because all zones are below low watermarks | |
4140 | * or is prohibited because it recently failed at this | |
3f36d866 DR |
4141 | * order, fail immediately unless the allocator has |
4142 | * requested compaction and reclaim retry. | |
b39d0ee2 DR |
4143 | * |
4144 | * Reclaim is | |
4145 | * - potentially very expensive because zones are far | |
4146 | * below their low watermarks or this is part of very | |
4147 | * bursty high order allocations, | |
4148 | * - not guaranteed to help because isolate_freepages() | |
4149 | * may not iterate over freed pages as part of its | |
4150 | * linear scan, and | |
4151 | * - unlikely to make entire pageblocks free on its | |
4152 | * own. | |
4153 | */ | |
4154 | if (compact_result == COMPACT_SKIPPED || | |
4155 | compact_result == COMPACT_DEFERRED) | |
4156 | goto nopage; | |
a8161d1e | 4157 | |
a8161d1e | 4158 | /* |
3eb2771b VB |
4159 | * Looks like reclaim/compaction is worth trying, but |
4160 | * sync compaction could be very expensive, so keep | |
25160354 | 4161 | * using async compaction. |
a8161d1e | 4162 | */ |
a5508cd8 | 4163 | compact_priority = INIT_COMPACT_PRIORITY; |
a8161d1e VB |
4164 | } |
4165 | } | |
23771235 | 4166 | |
31a6c190 | 4167 | retry: |
23771235 | 4168 | /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */ |
0a79cdad | 4169 | if (alloc_flags & ALLOC_KSWAPD) |
5ecd9d40 | 4170 | wake_all_kswapds(order, gfp_mask, ac); |
31a6c190 | 4171 | |
cd04ae1e MH |
4172 | reserve_flags = __gfp_pfmemalloc_flags(gfp_mask); |
4173 | if (reserve_flags) | |
ce96fa62 ML |
4174 | alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) | |
4175 | (alloc_flags & ALLOC_KSWAPD); | |
23771235 | 4176 | |
e46e7b77 | 4177 | /* |
d6a24df0 VB |
4178 | * Reset the nodemask and zonelist iterators if memory policies can be |
4179 | * ignored. These allocations are high priority and system rather than | |
4180 | * user oriented. | |
e46e7b77 | 4181 | */ |
cd04ae1e | 4182 | if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) { |
d6a24df0 | 4183 | ac->nodemask = NULL; |
e46e7b77 | 4184 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist, |
97a225e6 | 4185 | ac->highest_zoneidx, ac->nodemask); |
e46e7b77 MG |
4186 | } |
4187 | ||
23771235 | 4188 | /* Attempt with potentially adjusted zonelist and alloc_flags */ |
31a6c190 | 4189 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); |
7fb1d9fc RS |
4190 | if (page) |
4191 | goto got_pg; | |
1da177e4 | 4192 | |
d0164adc | 4193 | /* Caller is not willing to reclaim, we can't balance anything */ |
9a67f648 | 4194 | if (!can_direct_reclaim) |
1da177e4 LT |
4195 | goto nopage; |
4196 | ||
9a67f648 MH |
4197 | /* Avoid recursion of direct reclaim */ |
4198 | if (current->flags & PF_MEMALLOC) | |
6583bb64 DR |
4199 | goto nopage; |
4200 | ||
a8161d1e VB |
4201 | /* Try direct reclaim and then allocating */ |
4202 | page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac, | |
4203 | &did_some_progress); | |
4204 | if (page) | |
4205 | goto got_pg; | |
4206 | ||
4207 | /* Try direct compaction and then allocating */ | |
a9263751 | 4208 | page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac, |
a5508cd8 | 4209 | compact_priority, &compact_result); |
56de7263 MG |
4210 | if (page) |
4211 | goto got_pg; | |
75f30861 | 4212 | |
9083905a JW |
4213 | /* Do not loop if specifically requested */ |
4214 | if (gfp_mask & __GFP_NORETRY) | |
a8161d1e | 4215 | goto nopage; |
9083905a | 4216 | |
0a0337e0 MH |
4217 | /* |
4218 | * Do not retry costly high order allocations unless they are | |
803de900 | 4219 | * __GFP_RETRY_MAYFAIL and we can compact |
0a0337e0 | 4220 | */ |
803de900 VB |
4221 | if (costly_order && (!can_compact || |
4222 | !(gfp_mask & __GFP_RETRY_MAYFAIL))) | |
a8161d1e | 4223 | goto nopage; |
0a0337e0 | 4224 | |
0a0337e0 | 4225 | if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags, |
423b452e | 4226 | did_some_progress > 0, &no_progress_loops)) |
0a0337e0 MH |
4227 | goto retry; |
4228 | ||
33c2d214 MH |
4229 | /* |
4230 | * It doesn't make any sense to retry for the compaction if the order-0 | |
4231 | * reclaim is not able to make any progress because the current | |
4232 | * implementation of the compaction depends on the sufficient amount | |
4233 | * of free memory (see __compaction_suitable) | |
4234 | */ | |
803de900 | 4235 | if (did_some_progress > 0 && can_compact && |
86a294a8 | 4236 | should_compact_retry(ac, order, alloc_flags, |
a5508cd8 | 4237 | compact_result, &compact_priority, |
d9436498 | 4238 | &compaction_retries)) |
33c2d214 MH |
4239 | goto retry; |
4240 | ||
902b6281 | 4241 | |
3d36424b MG |
4242 | /* |
4243 | * Deal with possible cpuset update races or zonelist updates to avoid | |
4244 | * a unnecessary OOM kill. | |
4245 | */ | |
4246 | if (check_retry_cpuset(cpuset_mems_cookie, ac) || | |
4247 | check_retry_zonelist(zonelist_iter_cookie)) | |
4248 | goto restart; | |
e47483bc | 4249 | |
9083905a JW |
4250 | /* Reclaim has failed us, start killing things */ |
4251 | page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress); | |
4252 | if (page) | |
4253 | goto got_pg; | |
4254 | ||
9a67f648 | 4255 | /* Avoid allocations with no watermarks from looping endlessly */ |
cd04ae1e | 4256 | if (tsk_is_oom_victim(current) && |
8510e69c | 4257 | (alloc_flags & ALLOC_OOM || |
c288983d | 4258 | (gfp_mask & __GFP_NOMEMALLOC))) |
9a67f648 MH |
4259 | goto nopage; |
4260 | ||
9083905a | 4261 | /* Retry as long as the OOM killer is making progress */ |
0a0337e0 MH |
4262 | if (did_some_progress) { |
4263 | no_progress_loops = 0; | |
9083905a | 4264 | goto retry; |
0a0337e0 | 4265 | } |
9083905a | 4266 | |
1da177e4 | 4267 | nopage: |
3d36424b MG |
4268 | /* |
4269 | * Deal with possible cpuset update races or zonelist updates to avoid | |
4270 | * a unnecessary OOM kill. | |
4271 | */ | |
4272 | if (check_retry_cpuset(cpuset_mems_cookie, ac) || | |
4273 | check_retry_zonelist(zonelist_iter_cookie)) | |
4274 | goto restart; | |
5ce9bfef | 4275 | |
9a67f648 MH |
4276 | /* |
4277 | * Make sure that __GFP_NOFAIL request doesn't leak out and make sure | |
4278 | * we always retry | |
4279 | */ | |
4280 | if (gfp_mask & __GFP_NOFAIL) { | |
4281 | /* | |
4282 | * All existing users of the __GFP_NOFAIL are blockable, so warn | |
4283 | * of any new users that actually require GFP_NOWAIT | |
4284 | */ | |
3f913fc5 | 4285 | if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask)) |
9a67f648 MH |
4286 | goto fail; |
4287 | ||
4288 | /* | |
4289 | * PF_MEMALLOC request from this context is rather bizarre | |
4290 | * because we cannot reclaim anything and only can loop waiting | |
4291 | * for somebody to do a work for us | |
4292 | */ | |
3f913fc5 | 4293 | WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask); |
9a67f648 MH |
4294 | |
4295 | /* | |
4296 | * non failing costly orders are a hard requirement which we | |
4297 | * are not prepared for much so let's warn about these users | |
4298 | * so that we can identify them and convert them to something | |
4299 | * else. | |
4300 | */ | |
896c4d52 | 4301 | WARN_ON_ONCE_GFP(costly_order, gfp_mask); |
9a67f648 | 4302 | |
6c18ba7a | 4303 | /* |
1ebbb218 MG |
4304 | * Help non-failing allocations by giving some access to memory |
4305 | * reserves normally used for high priority non-blocking | |
4306 | * allocations but do not use ALLOC_NO_WATERMARKS because this | |
6c18ba7a | 4307 | * could deplete whole memory reserves which would just make |
1ebbb218 | 4308 | * the situation worse. |
6c18ba7a | 4309 | */ |
1ebbb218 | 4310 | page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac); |
6c18ba7a MH |
4311 | if (page) |
4312 | goto got_pg; | |
4313 | ||
9a67f648 MH |
4314 | cond_resched(); |
4315 | goto retry; | |
4316 | } | |
4317 | fail: | |
a8e99259 | 4318 | warn_alloc(gfp_mask, ac->nodemask, |
7877cdcc | 4319 | "page allocation failure: order:%u", order); |
1da177e4 | 4320 | got_pg: |
072bb0aa | 4321 | return page; |
1da177e4 | 4322 | } |
11e33f6a | 4323 | |
9cd75558 | 4324 | static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order, |
04ec6264 | 4325 | int preferred_nid, nodemask_t *nodemask, |
8e6a930b | 4326 | struct alloc_context *ac, gfp_t *alloc_gfp, |
9cd75558 | 4327 | unsigned int *alloc_flags) |
11e33f6a | 4328 | { |
97a225e6 | 4329 | ac->highest_zoneidx = gfp_zone(gfp_mask); |
04ec6264 | 4330 | ac->zonelist = node_zonelist(preferred_nid, gfp_mask); |
9cd75558 | 4331 | ac->nodemask = nodemask; |
01c0bfe0 | 4332 | ac->migratetype = gfp_migratetype(gfp_mask); |
11e33f6a | 4333 | |
682a3385 | 4334 | if (cpusets_enabled()) { |
8e6a930b | 4335 | *alloc_gfp |= __GFP_HARDWALL; |
182f3d7a MS |
4336 | /* |
4337 | * When we are in the interrupt context, it is irrelevant | |
4338 | * to the current task context. It means that any node ok. | |
4339 | */ | |
88dc6f20 | 4340 | if (in_task() && !ac->nodemask) |
9cd75558 | 4341 | ac->nodemask = &cpuset_current_mems_allowed; |
51047820 VB |
4342 | else |
4343 | *alloc_flags |= ALLOC_CPUSET; | |
682a3385 MG |
4344 | } |
4345 | ||
446ec838 | 4346 | might_alloc(gfp_mask); |
11e33f6a MG |
4347 | |
4348 | if (should_fail_alloc_page(gfp_mask, order)) | |
9cd75558 | 4349 | return false; |
11e33f6a | 4350 | |
8e3560d9 | 4351 | *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags); |
d883c6cf | 4352 | |
c9ab0c4f | 4353 | /* Dirty zone balancing only done in the fast path */ |
9cd75558 | 4354 | ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE); |
c9ab0c4f | 4355 | |
e46e7b77 MG |
4356 | /* |
4357 | * The preferred zone is used for statistics but crucially it is | |
4358 | * also used as the starting point for the zonelist iterator. It | |
4359 | * may get reset for allocations that ignore memory policies. | |
4360 | */ | |
9cd75558 | 4361 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist, |
97a225e6 | 4362 | ac->highest_zoneidx, ac->nodemask); |
a0622d05 MN |
4363 | |
4364 | return true; | |
9cd75558 MG |
4365 | } |
4366 | ||
387ba26f | 4367 | /* |
0f87d9d3 | 4368 | * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array |
387ba26f MG |
4369 | * @gfp: GFP flags for the allocation |
4370 | * @preferred_nid: The preferred NUMA node ID to allocate from | |
4371 | * @nodemask: Set of nodes to allocate from, may be NULL | |
0f87d9d3 MG |
4372 | * @nr_pages: The number of pages desired on the list or array |
4373 | * @page_list: Optional list to store the allocated pages | |
4374 | * @page_array: Optional array to store the pages | |
387ba26f MG |
4375 | * |
4376 | * This is a batched version of the page allocator that attempts to | |
0f87d9d3 MG |
4377 | * allocate nr_pages quickly. Pages are added to page_list if page_list |
4378 | * is not NULL, otherwise it is assumed that the page_array is valid. | |
387ba26f | 4379 | * |
0f87d9d3 MG |
4380 | * For lists, nr_pages is the number of pages that should be allocated. |
4381 | * | |
4382 | * For arrays, only NULL elements are populated with pages and nr_pages | |
4383 | * is the maximum number of pages that will be stored in the array. | |
4384 | * | |
4385 | * Returns the number of pages on the list or array. | |
387ba26f MG |
4386 | */ |
4387 | unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid, | |
4388 | nodemask_t *nodemask, int nr_pages, | |
0f87d9d3 MG |
4389 | struct list_head *page_list, |
4390 | struct page **page_array) | |
387ba26f MG |
4391 | { |
4392 | struct page *page; | |
4b23a68f | 4393 | unsigned long __maybe_unused UP_flags; |
387ba26f MG |
4394 | struct zone *zone; |
4395 | struct zoneref *z; | |
4396 | struct per_cpu_pages *pcp; | |
4397 | struct list_head *pcp_list; | |
4398 | struct alloc_context ac; | |
4399 | gfp_t alloc_gfp; | |
4400 | unsigned int alloc_flags = ALLOC_WMARK_LOW; | |
3e23060b | 4401 | int nr_populated = 0, nr_account = 0; |
387ba26f | 4402 | |
0f87d9d3 MG |
4403 | /* |
4404 | * Skip populated array elements to determine if any pages need | |
4405 | * to be allocated before disabling IRQs. | |
4406 | */ | |
b08e50dd | 4407 | while (page_array && nr_populated < nr_pages && page_array[nr_populated]) |
0f87d9d3 MG |
4408 | nr_populated++; |
4409 | ||
06147843 CL |
4410 | /* No pages requested? */ |
4411 | if (unlikely(nr_pages <= 0)) | |
4412 | goto out; | |
4413 | ||
b3b64ebd MG |
4414 | /* Already populated array? */ |
4415 | if (unlikely(page_array && nr_pages - nr_populated == 0)) | |
06147843 | 4416 | goto out; |
b3b64ebd | 4417 | |
8dcb3060 | 4418 | /* Bulk allocator does not support memcg accounting. */ |
f7a449f7 | 4419 | if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT)) |
8dcb3060 SB |
4420 | goto failed; |
4421 | ||
387ba26f | 4422 | /* Use the single page allocator for one page. */ |
0f87d9d3 | 4423 | if (nr_pages - nr_populated == 1) |
387ba26f MG |
4424 | goto failed; |
4425 | ||
187ad460 MG |
4426 | #ifdef CONFIG_PAGE_OWNER |
4427 | /* | |
4428 | * PAGE_OWNER may recurse into the allocator to allocate space to | |
4429 | * save the stack with pagesets.lock held. Releasing/reacquiring | |
4430 | * removes much of the performance benefit of bulk allocation so | |
4431 | * force the caller to allocate one page at a time as it'll have | |
4432 | * similar performance to added complexity to the bulk allocator. | |
4433 | */ | |
4434 | if (static_branch_unlikely(&page_owner_inited)) | |
4435 | goto failed; | |
4436 | #endif | |
4437 | ||
387ba26f MG |
4438 | /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */ |
4439 | gfp &= gfp_allowed_mask; | |
4440 | alloc_gfp = gfp; | |
4441 | if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags)) | |
06147843 | 4442 | goto out; |
387ba26f MG |
4443 | gfp = alloc_gfp; |
4444 | ||
4445 | /* Find an allowed local zone that meets the low watermark. */ | |
4446 | for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) { | |
4447 | unsigned long mark; | |
4448 | ||
4449 | if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) && | |
4450 | !__cpuset_zone_allowed(zone, gfp)) { | |
4451 | continue; | |
4452 | } | |
4453 | ||
4454 | if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone && | |
4455 | zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) { | |
4456 | goto failed; | |
4457 | } | |
4458 | ||
4459 | mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages; | |
4460 | if (zone_watermark_fast(zone, 0, mark, | |
4461 | zonelist_zone_idx(ac.preferred_zoneref), | |
4462 | alloc_flags, gfp)) { | |
4463 | break; | |
4464 | } | |
4465 | } | |
4466 | ||
4467 | /* | |
4468 | * If there are no allowed local zones that meets the watermarks then | |
4469 | * try to allocate a single page and reclaim if necessary. | |
4470 | */ | |
ce76f9a1 | 4471 | if (unlikely(!zone)) |
387ba26f MG |
4472 | goto failed; |
4473 | ||
57490774 | 4474 | /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */ |
4b23a68f | 4475 | pcp_trylock_prepare(UP_flags); |
57490774 | 4476 | pcp = pcp_spin_trylock(zone->per_cpu_pageset); |
01b44456 | 4477 | if (!pcp) |
4b23a68f | 4478 | goto failed_irq; |
387ba26f | 4479 | |
387ba26f | 4480 | /* Attempt the batch allocation */ |
44042b44 | 4481 | pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)]; |
0f87d9d3 MG |
4482 | while (nr_populated < nr_pages) { |
4483 | ||
4484 | /* Skip existing pages */ | |
4485 | if (page_array && page_array[nr_populated]) { | |
4486 | nr_populated++; | |
4487 | continue; | |
4488 | } | |
4489 | ||
44042b44 | 4490 | page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags, |
387ba26f | 4491 | pcp, pcp_list); |
ce76f9a1 | 4492 | if (unlikely(!page)) { |
c572e488 | 4493 | /* Try and allocate at least one page */ |
4b23a68f | 4494 | if (!nr_account) { |
57490774 | 4495 | pcp_spin_unlock(pcp); |
387ba26f | 4496 | goto failed_irq; |
4b23a68f | 4497 | } |
387ba26f MG |
4498 | break; |
4499 | } | |
3e23060b | 4500 | nr_account++; |
387ba26f MG |
4501 | |
4502 | prep_new_page(page, 0, gfp, 0); | |
0f87d9d3 MG |
4503 | if (page_list) |
4504 | list_add(&page->lru, page_list); | |
4505 | else | |
4506 | page_array[nr_populated] = page; | |
4507 | nr_populated++; | |
387ba26f MG |
4508 | } |
4509 | ||
57490774 | 4510 | pcp_spin_unlock(pcp); |
4b23a68f | 4511 | pcp_trylock_finish(UP_flags); |
43c95bcc | 4512 | |
3e23060b MG |
4513 | __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account); |
4514 | zone_statistics(ac.preferred_zoneref->zone, zone, nr_account); | |
387ba26f | 4515 | |
06147843 | 4516 | out: |
0f87d9d3 | 4517 | return nr_populated; |
387ba26f MG |
4518 | |
4519 | failed_irq: | |
4b23a68f | 4520 | pcp_trylock_finish(UP_flags); |
387ba26f MG |
4521 | |
4522 | failed: | |
4523 | page = __alloc_pages(gfp, 0, preferred_nid, nodemask); | |
4524 | if (page) { | |
0f87d9d3 MG |
4525 | if (page_list) |
4526 | list_add(&page->lru, page_list); | |
4527 | else | |
4528 | page_array[nr_populated] = page; | |
4529 | nr_populated++; | |
387ba26f MG |
4530 | } |
4531 | ||
06147843 | 4532 | goto out; |
387ba26f MG |
4533 | } |
4534 | EXPORT_SYMBOL_GPL(__alloc_pages_bulk); | |
4535 | ||
9cd75558 MG |
4536 | /* |
4537 | * This is the 'heart' of the zoned buddy allocator. | |
4538 | */ | |
84172f4b | 4539 | struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid, |
04ec6264 | 4540 | nodemask_t *nodemask) |
9cd75558 MG |
4541 | { |
4542 | struct page *page; | |
4543 | unsigned int alloc_flags = ALLOC_WMARK_LOW; | |
8e6a930b | 4544 | gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */ |
9cd75558 MG |
4545 | struct alloc_context ac = { }; |
4546 | ||
c63ae43b MH |
4547 | /* |
4548 | * There are several places where we assume that the order value is sane | |
4549 | * so bail out early if the request is out of bound. | |
4550 | */ | |
5e0a760b | 4551 | if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp)) |
c63ae43b | 4552 | return NULL; |
c63ae43b | 4553 | |
6e5e0f28 | 4554 | gfp &= gfp_allowed_mask; |
da6df1b0 PT |
4555 | /* |
4556 | * Apply scoped allocation constraints. This is mainly about GFP_NOFS | |
4557 | * resp. GFP_NOIO which has to be inherited for all allocation requests | |
4558 | * from a particular context which has been marked by | |
8e3560d9 PT |
4559 | * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures |
4560 | * movable zones are not used during allocation. | |
da6df1b0 PT |
4561 | */ |
4562 | gfp = current_gfp_context(gfp); | |
6e5e0f28 MWO |
4563 | alloc_gfp = gfp; |
4564 | if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac, | |
8e6a930b | 4565 | &alloc_gfp, &alloc_flags)) |
9cd75558 MG |
4566 | return NULL; |
4567 | ||
6bb15450 MG |
4568 | /* |
4569 | * Forbid the first pass from falling back to types that fragment | |
4570 | * memory until all local zones are considered. | |
4571 | */ | |
6e5e0f28 | 4572 | alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp); |
6bb15450 | 4573 | |
5117f45d | 4574 | /* First allocation attempt */ |
8e6a930b | 4575 | page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac); |
4fcb0971 MG |
4576 | if (likely(page)) |
4577 | goto out; | |
11e33f6a | 4578 | |
da6df1b0 | 4579 | alloc_gfp = gfp; |
4fcb0971 | 4580 | ac.spread_dirty_pages = false; |
23f086f9 | 4581 | |
4741526b MG |
4582 | /* |
4583 | * Restore the original nodemask if it was potentially replaced with | |
4584 | * &cpuset_current_mems_allowed to optimize the fast-path attempt. | |
4585 | */ | |
97ce86f9 | 4586 | ac.nodemask = nodemask; |
16096c25 | 4587 | |
8e6a930b | 4588 | page = __alloc_pages_slowpath(alloc_gfp, order, &ac); |
cc9a6c87 | 4589 | |
4fcb0971 | 4590 | out: |
f7a449f7 | 4591 | if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page && |
6e5e0f28 | 4592 | unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) { |
c4159a75 VD |
4593 | __free_pages(page, order); |
4594 | page = NULL; | |
4949148a VD |
4595 | } |
4596 | ||
8e6a930b | 4597 | trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype); |
b073d7f8 | 4598 | kmsan_alloc_page(page, order, alloc_gfp); |
4fcb0971 | 4599 | |
11e33f6a | 4600 | return page; |
1da177e4 | 4601 | } |
84172f4b | 4602 | EXPORT_SYMBOL(__alloc_pages); |
1da177e4 | 4603 | |
cc09cb13 MWO |
4604 | struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid, |
4605 | nodemask_t *nodemask) | |
4606 | { | |
4607 | struct page *page = __alloc_pages(gfp | __GFP_COMP, order, | |
23e48832 HD |
4608 | preferred_nid, nodemask); |
4609 | return page_rmappable_folio(page); | |
cc09cb13 MWO |
4610 | } |
4611 | EXPORT_SYMBOL(__folio_alloc); | |
4612 | ||
1da177e4 | 4613 | /* |
9ea9a680 MH |
4614 | * Common helper functions. Never use with __GFP_HIGHMEM because the returned |
4615 | * address cannot represent highmem pages. Use alloc_pages and then kmap if | |
4616 | * you need to access high mem. | |
1da177e4 | 4617 | */ |
920c7a5d | 4618 | unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) |
1da177e4 | 4619 | { |
945a1113 AM |
4620 | struct page *page; |
4621 | ||
9ea9a680 | 4622 | page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order); |
1da177e4 LT |
4623 | if (!page) |
4624 | return 0; | |
4625 | return (unsigned long) page_address(page); | |
4626 | } | |
1da177e4 LT |
4627 | EXPORT_SYMBOL(__get_free_pages); |
4628 | ||
920c7a5d | 4629 | unsigned long get_zeroed_page(gfp_t gfp_mask) |
1da177e4 | 4630 | { |
dcc1be11 | 4631 | return __get_free_page(gfp_mask | __GFP_ZERO); |
1da177e4 | 4632 | } |
1da177e4 LT |
4633 | EXPORT_SYMBOL(get_zeroed_page); |
4634 | ||
7f194fbb MWO |
4635 | /** |
4636 | * __free_pages - Free pages allocated with alloc_pages(). | |
4637 | * @page: The page pointer returned from alloc_pages(). | |
4638 | * @order: The order of the allocation. | |
4639 | * | |
4640 | * This function can free multi-page allocations that are not compound | |
4641 | * pages. It does not check that the @order passed in matches that of | |
4642 | * the allocation, so it is easy to leak memory. Freeing more memory | |
4643 | * than was allocated will probably emit a warning. | |
4644 | * | |
4645 | * If the last reference to this page is speculative, it will be released | |
4646 | * by put_page() which only frees the first page of a non-compound | |
4647 | * allocation. To prevent the remaining pages from being leaked, we free | |
4648 | * the subsequent pages here. If you want to use the page's reference | |
4649 | * count to decide when to free the allocation, you should allocate a | |
4650 | * compound page, and use put_page() instead of __free_pages(). | |
4651 | * | |
4652 | * Context: May be called in interrupt context or while holding a normal | |
4653 | * spinlock, but not in NMI context or while holding a raw spinlock. | |
4654 | */ | |
742aa7fb AL |
4655 | void __free_pages(struct page *page, unsigned int order) |
4656 | { | |
462a8e08 DC |
4657 | /* get PageHead before we drop reference */ |
4658 | int head = PageHead(page); | |
4659 | ||
742aa7fb AL |
4660 | if (put_page_testzero(page)) |
4661 | free_the_page(page, order); | |
462a8e08 | 4662 | else if (!head) |
e320d301 MWO |
4663 | while (order-- > 0) |
4664 | free_the_page(page + (1 << order), order); | |
742aa7fb | 4665 | } |
1da177e4 LT |
4666 | EXPORT_SYMBOL(__free_pages); |
4667 | ||
920c7a5d | 4668 | void free_pages(unsigned long addr, unsigned int order) |
1da177e4 LT |
4669 | { |
4670 | if (addr != 0) { | |
725d704e | 4671 | VM_BUG_ON(!virt_addr_valid((void *)addr)); |
1da177e4 LT |
4672 | __free_pages(virt_to_page((void *)addr), order); |
4673 | } | |
4674 | } | |
4675 | ||
4676 | EXPORT_SYMBOL(free_pages); | |
4677 | ||
b63ae8ca AD |
4678 | /* |
4679 | * Page Fragment: | |
4680 | * An arbitrary-length arbitrary-offset area of memory which resides | |
4681 | * within a 0 or higher order page. Multiple fragments within that page | |
4682 | * are individually refcounted, in the page's reference counter. | |
4683 | * | |
4684 | * The page_frag functions below provide a simple allocation framework for | |
4685 | * page fragments. This is used by the network stack and network device | |
4686 | * drivers to provide a backing region of memory for use as either an | |
4687 | * sk_buff->head, or to be used in the "frags" portion of skb_shared_info. | |
4688 | */ | |
2976db80 AD |
4689 | static struct page *__page_frag_cache_refill(struct page_frag_cache *nc, |
4690 | gfp_t gfp_mask) | |
b63ae8ca AD |
4691 | { |
4692 | struct page *page = NULL; | |
4693 | gfp_t gfp = gfp_mask; | |
4694 | ||
4695 | #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) | |
4bc0d63a YL |
4696 | gfp_mask = (gfp_mask & ~__GFP_DIRECT_RECLAIM) | __GFP_COMP | |
4697 | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC; | |
b63ae8ca AD |
4698 | page = alloc_pages_node(NUMA_NO_NODE, gfp_mask, |
4699 | PAGE_FRAG_CACHE_MAX_ORDER); | |
4700 | nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE; | |
4701 | #endif | |
4702 | if (unlikely(!page)) | |
4703 | page = alloc_pages_node(NUMA_NO_NODE, gfp, 0); | |
4704 | ||
4705 | nc->va = page ? page_address(page) : NULL; | |
4706 | ||
4707 | return page; | |
4708 | } | |
4709 | ||
a0727489 YL |
4710 | void page_frag_cache_drain(struct page_frag_cache *nc) |
4711 | { | |
4712 | if (!nc->va) | |
4713 | return; | |
4714 | ||
4715 | __page_frag_cache_drain(virt_to_head_page(nc->va), nc->pagecnt_bias); | |
4716 | nc->va = NULL; | |
4717 | } | |
4718 | EXPORT_SYMBOL(page_frag_cache_drain); | |
4719 | ||
2976db80 | 4720 | void __page_frag_cache_drain(struct page *page, unsigned int count) |
44fdffd7 AD |
4721 | { |
4722 | VM_BUG_ON_PAGE(page_ref_count(page) == 0, page); | |
4723 | ||
742aa7fb AL |
4724 | if (page_ref_sub_and_test(page, count)) |
4725 | free_the_page(page, compound_order(page)); | |
44fdffd7 | 4726 | } |
2976db80 | 4727 | EXPORT_SYMBOL(__page_frag_cache_drain); |
44fdffd7 | 4728 | |
411c5f36 YL |
4729 | void *__page_frag_alloc_align(struct page_frag_cache *nc, |
4730 | unsigned int fragsz, gfp_t gfp_mask, | |
4731 | unsigned int align_mask) | |
b63ae8ca AD |
4732 | { |
4733 | unsigned int size = PAGE_SIZE; | |
4734 | struct page *page; | |
4735 | int offset; | |
4736 | ||
4737 | if (unlikely(!nc->va)) { | |
4738 | refill: | |
2976db80 | 4739 | page = __page_frag_cache_refill(nc, gfp_mask); |
b63ae8ca AD |
4740 | if (!page) |
4741 | return NULL; | |
4742 | ||
4743 | #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) | |
4744 | /* if size can vary use size else just use PAGE_SIZE */ | |
4745 | size = nc->size; | |
4746 | #endif | |
4747 | /* Even if we own the page, we do not use atomic_set(). | |
4748 | * This would break get_page_unless_zero() users. | |
4749 | */ | |
86447726 | 4750 | page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE); |
b63ae8ca AD |
4751 | |
4752 | /* reset page count bias and offset to start of new frag */ | |
2f064f34 | 4753 | nc->pfmemalloc = page_is_pfmemalloc(page); |
86447726 | 4754 | nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1; |
b63ae8ca AD |
4755 | nc->offset = size; |
4756 | } | |
4757 | ||
4758 | offset = nc->offset - fragsz; | |
4759 | if (unlikely(offset < 0)) { | |
4760 | page = virt_to_page(nc->va); | |
4761 | ||
fe896d18 | 4762 | if (!page_ref_sub_and_test(page, nc->pagecnt_bias)) |
b63ae8ca AD |
4763 | goto refill; |
4764 | ||
d8c19014 DZ |
4765 | if (unlikely(nc->pfmemalloc)) { |
4766 | free_the_page(page, compound_order(page)); | |
4767 | goto refill; | |
4768 | } | |
4769 | ||
b63ae8ca AD |
4770 | #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) |
4771 | /* if size can vary use size else just use PAGE_SIZE */ | |
4772 | size = nc->size; | |
4773 | #endif | |
4774 | /* OK, page count is 0, we can safely set it */ | |
86447726 | 4775 | set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1); |
b63ae8ca AD |
4776 | |
4777 | /* reset page count bias and offset to start of new frag */ | |
86447726 | 4778 | nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1; |
b63ae8ca | 4779 | offset = size - fragsz; |
dac22531 ML |
4780 | if (unlikely(offset < 0)) { |
4781 | /* | |
4782 | * The caller is trying to allocate a fragment | |
4783 | * with fragsz > PAGE_SIZE but the cache isn't big | |
4784 | * enough to satisfy the request, this may | |
4785 | * happen in low memory conditions. | |
4786 | * We don't release the cache page because | |
4787 | * it could make memory pressure worse | |
4788 | * so we simply return NULL here. | |
4789 | */ | |
4790 | return NULL; | |
4791 | } | |
b63ae8ca AD |
4792 | } |
4793 | ||
4794 | nc->pagecnt_bias--; | |
b358e212 | 4795 | offset &= align_mask; |
b63ae8ca AD |
4796 | nc->offset = offset; |
4797 | ||
4798 | return nc->va + offset; | |
4799 | } | |
411c5f36 | 4800 | EXPORT_SYMBOL(__page_frag_alloc_align); |
b63ae8ca AD |
4801 | |
4802 | /* | |
4803 | * Frees a page fragment allocated out of either a compound or order 0 page. | |
4804 | */ | |
8c2dd3e4 | 4805 | void page_frag_free(void *addr) |
b63ae8ca AD |
4806 | { |
4807 | struct page *page = virt_to_head_page(addr); | |
4808 | ||
742aa7fb AL |
4809 | if (unlikely(put_page_testzero(page))) |
4810 | free_the_page(page, compound_order(page)); | |
b63ae8ca | 4811 | } |
8c2dd3e4 | 4812 | EXPORT_SYMBOL(page_frag_free); |
b63ae8ca | 4813 | |
d00181b9 KS |
4814 | static void *make_alloc_exact(unsigned long addr, unsigned int order, |
4815 | size_t size) | |
ee85c2e1 AK |
4816 | { |
4817 | if (addr) { | |
df48a5f7 LH |
4818 | unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE); |
4819 | struct page *page = virt_to_page((void *)addr); | |
4820 | struct page *last = page + nr; | |
4821 | ||
46d44d09 | 4822 | split_page_owner(page, order, 0); |
b8791381 | 4823 | split_page_memcg(page, order, 0); |
df48a5f7 LH |
4824 | while (page < --last) |
4825 | set_page_refcounted(last); | |
4826 | ||
4827 | last = page + (1UL << order); | |
4828 | for (page += nr; page < last; page++) | |
4829 | __free_pages_ok(page, 0, FPI_TO_TAIL); | |
ee85c2e1 AK |
4830 | } |
4831 | return (void *)addr; | |
4832 | } | |
4833 | ||
2be0ffe2 TT |
4834 | /** |
4835 | * alloc_pages_exact - allocate an exact number physically-contiguous pages. | |
4836 | * @size: the number of bytes to allocate | |
63931eb9 | 4837 | * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP |
2be0ffe2 TT |
4838 | * |
4839 | * This function is similar to alloc_pages(), except that it allocates the | |
4840 | * minimum number of pages to satisfy the request. alloc_pages() can only | |
4841 | * allocate memory in power-of-two pages. | |
4842 | * | |
5e0a760b | 4843 | * This function is also limited by MAX_PAGE_ORDER. |
2be0ffe2 TT |
4844 | * |
4845 | * Memory allocated by this function must be released by free_pages_exact(). | |
a862f68a MR |
4846 | * |
4847 | * Return: pointer to the allocated area or %NULL in case of error. | |
2be0ffe2 TT |
4848 | */ |
4849 | void *alloc_pages_exact(size_t size, gfp_t gfp_mask) | |
4850 | { | |
4851 | unsigned int order = get_order(size); | |
4852 | unsigned long addr; | |
4853 | ||
ba7f1b9e ML |
4854 | if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM))) |
4855 | gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM); | |
63931eb9 | 4856 | |
2be0ffe2 | 4857 | addr = __get_free_pages(gfp_mask, order); |
ee85c2e1 | 4858 | return make_alloc_exact(addr, order, size); |
2be0ffe2 TT |
4859 | } |
4860 | EXPORT_SYMBOL(alloc_pages_exact); | |
4861 | ||
ee85c2e1 AK |
4862 | /** |
4863 | * alloc_pages_exact_nid - allocate an exact number of physically-contiguous | |
4864 | * pages on a node. | |
b5e6ab58 | 4865 | * @nid: the preferred node ID where memory should be allocated |
ee85c2e1 | 4866 | * @size: the number of bytes to allocate |
63931eb9 | 4867 | * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP |
ee85c2e1 AK |
4868 | * |
4869 | * Like alloc_pages_exact(), but try to allocate on node nid first before falling | |
4870 | * back. | |
a862f68a MR |
4871 | * |
4872 | * Return: pointer to the allocated area or %NULL in case of error. | |
ee85c2e1 | 4873 | */ |
e1931811 | 4874 | void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) |
ee85c2e1 | 4875 | { |
d00181b9 | 4876 | unsigned int order = get_order(size); |
63931eb9 VB |
4877 | struct page *p; |
4878 | ||
ba7f1b9e ML |
4879 | if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM))) |
4880 | gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM); | |
63931eb9 VB |
4881 | |
4882 | p = alloc_pages_node(nid, gfp_mask, order); | |
ee85c2e1 AK |
4883 | if (!p) |
4884 | return NULL; | |
4885 | return make_alloc_exact((unsigned long)page_address(p), order, size); | |
4886 | } | |
ee85c2e1 | 4887 | |
2be0ffe2 TT |
4888 | /** |
4889 | * free_pages_exact - release memory allocated via alloc_pages_exact() | |
4890 | * @virt: the value returned by alloc_pages_exact. | |
4891 | * @size: size of allocation, same value as passed to alloc_pages_exact(). | |
4892 | * | |
4893 | * Release the memory allocated by a previous call to alloc_pages_exact. | |
4894 | */ | |
4895 | void free_pages_exact(void *virt, size_t size) | |
4896 | { | |
4897 | unsigned long addr = (unsigned long)virt; | |
4898 | unsigned long end = addr + PAGE_ALIGN(size); | |
4899 | ||
4900 | while (addr < end) { | |
4901 | free_page(addr); | |
4902 | addr += PAGE_SIZE; | |
4903 | } | |
4904 | } | |
4905 | EXPORT_SYMBOL(free_pages_exact); | |
4906 | ||
e0fb5815 ZY |
4907 | /** |
4908 | * nr_free_zone_pages - count number of pages beyond high watermark | |
4909 | * @offset: The zone index of the highest zone | |
4910 | * | |
a862f68a | 4911 | * nr_free_zone_pages() counts the number of pages which are beyond the |
e0fb5815 ZY |
4912 | * high watermark within all zones at or below a given zone index. For each |
4913 | * zone, the number of pages is calculated as: | |
0e056eb5 MCC |
4914 | * |
4915 | * nr_free_zone_pages = managed_pages - high_pages | |
a862f68a MR |
4916 | * |
4917 | * Return: number of pages beyond high watermark. | |
e0fb5815 | 4918 | */ |
ebec3862 | 4919 | static unsigned long nr_free_zone_pages(int offset) |
1da177e4 | 4920 | { |
dd1a239f | 4921 | struct zoneref *z; |
54a6eb5c MG |
4922 | struct zone *zone; |
4923 | ||
e310fd43 | 4924 | /* Just pick one node, since fallback list is circular */ |
ebec3862 | 4925 | unsigned long sum = 0; |
1da177e4 | 4926 | |
0e88460d | 4927 | struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); |
1da177e4 | 4928 | |
54a6eb5c | 4929 | for_each_zone_zonelist(zone, z, zonelist, offset) { |
9705bea5 | 4930 | unsigned long size = zone_managed_pages(zone); |
41858966 | 4931 | unsigned long high = high_wmark_pages(zone); |
e310fd43 MB |
4932 | if (size > high) |
4933 | sum += size - high; | |
1da177e4 LT |
4934 | } |
4935 | ||
4936 | return sum; | |
4937 | } | |
4938 | ||
e0fb5815 ZY |
4939 | /** |
4940 | * nr_free_buffer_pages - count number of pages beyond high watermark | |
4941 | * | |
4942 | * nr_free_buffer_pages() counts the number of pages which are beyond the high | |
4943 | * watermark within ZONE_DMA and ZONE_NORMAL. | |
a862f68a MR |
4944 | * |
4945 | * Return: number of pages beyond high watermark within ZONE_DMA and | |
4946 | * ZONE_NORMAL. | |
1da177e4 | 4947 | */ |
ebec3862 | 4948 | unsigned long nr_free_buffer_pages(void) |
1da177e4 | 4949 | { |
af4ca457 | 4950 | return nr_free_zone_pages(gfp_zone(GFP_USER)); |
1da177e4 | 4951 | } |
c2f1a551 | 4952 | EXPORT_SYMBOL_GPL(nr_free_buffer_pages); |
1da177e4 | 4953 | |
19770b32 MG |
4954 | static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) |
4955 | { | |
4956 | zoneref->zone = zone; | |
4957 | zoneref->zone_idx = zone_idx(zone); | |
4958 | } | |
4959 | ||
1da177e4 LT |
4960 | /* |
4961 | * Builds allocation fallback zone lists. | |
1a93205b CL |
4962 | * |
4963 | * Add all populated zones of a node to the zonelist. | |
1da177e4 | 4964 | */ |
9d3be21b | 4965 | static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs) |
1da177e4 | 4966 | { |
1a93205b | 4967 | struct zone *zone; |
bc732f1d | 4968 | enum zone_type zone_type = MAX_NR_ZONES; |
9d3be21b | 4969 | int nr_zones = 0; |
02a68a5e CL |
4970 | |
4971 | do { | |
2f6726e5 | 4972 | zone_type--; |
070f8032 | 4973 | zone = pgdat->node_zones + zone_type; |
e553f62f | 4974 | if (populated_zone(zone)) { |
9d3be21b | 4975 | zoneref_set_zone(zone, &zonerefs[nr_zones++]); |
070f8032 | 4976 | check_highest_zone(zone_type); |
1da177e4 | 4977 | } |
2f6726e5 | 4978 | } while (zone_type); |
bc732f1d | 4979 | |
070f8032 | 4980 | return nr_zones; |
1da177e4 LT |
4981 | } |
4982 | ||
4983 | #ifdef CONFIG_NUMA | |
f0c0b2b8 KH |
4984 | |
4985 | static int __parse_numa_zonelist_order(char *s) | |
4986 | { | |
c9bff3ee | 4987 | /* |
f0953a1b | 4988 | * We used to support different zonelists modes but they turned |
c9bff3ee MH |
4989 | * out to be just not useful. Let's keep the warning in place |
4990 | * if somebody still use the cmd line parameter so that we do | |
4991 | * not fail it silently | |
4992 | */ | |
4993 | if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) { | |
4994 | pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s); | |
f0c0b2b8 KH |
4995 | return -EINVAL; |
4996 | } | |
4997 | return 0; | |
4998 | } | |
4999 | ||
e95d372c KW |
5000 | static char numa_zonelist_order[] = "Node"; |
5001 | #define NUMA_ZONELIST_ORDER_LEN 16 | |
f0c0b2b8 KH |
5002 | /* |
5003 | * sysctl handler for numa_zonelist_order | |
5004 | */ | |
e95d372c | 5005 | static int numa_zonelist_order_handler(struct ctl_table *table, int write, |
32927393 | 5006 | void *buffer, size_t *length, loff_t *ppos) |
f0c0b2b8 | 5007 | { |
32927393 CH |
5008 | if (write) |
5009 | return __parse_numa_zonelist_order(buffer); | |
5010 | return proc_dostring(table, write, buffer, length, ppos); | |
f0c0b2b8 KH |
5011 | } |
5012 | ||
f0c0b2b8 KH |
5013 | static int node_load[MAX_NUMNODES]; |
5014 | ||
1da177e4 | 5015 | /** |
4dc3b16b | 5016 | * find_next_best_node - find the next node that should appear in a given node's fallback list |
1da177e4 LT |
5017 | * @node: node whose fallback list we're appending |
5018 | * @used_node_mask: nodemask_t of already used nodes | |
5019 | * | |
5020 | * We use a number of factors to determine which is the next node that should | |
5021 | * appear on a given node's fallback list. The node should not have appeared | |
5022 | * already in @node's fallback list, and it should be the next closest node | |
5023 | * according to the distance array (which contains arbitrary distance values | |
5024 | * from each node to each node in the system), and should also prefer nodes | |
5025 | * with no CPUs, since presumably they'll have very little allocation pressure | |
5026 | * on them otherwise. | |
a862f68a MR |
5027 | * |
5028 | * Return: node id of the found node or %NUMA_NO_NODE if no node is found. | |
1da177e4 | 5029 | */ |
79c28a41 | 5030 | int find_next_best_node(int node, nodemask_t *used_node_mask) |
1da177e4 | 5031 | { |
4cf808eb | 5032 | int n, val; |
1da177e4 | 5033 | int min_val = INT_MAX; |
00ef2d2f | 5034 | int best_node = NUMA_NO_NODE; |
1da177e4 | 5035 | |
c2baef39 QZ |
5036 | /* |
5037 | * Use the local node if we haven't already, but for memoryless local | |
5038 | * node, we should skip it and fall back to other nodes. | |
5039 | */ | |
5040 | if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) { | |
4cf808eb LT |
5041 | node_set(node, *used_node_mask); |
5042 | return node; | |
5043 | } | |
1da177e4 | 5044 | |
4b0ef1fe | 5045 | for_each_node_state(n, N_MEMORY) { |
1da177e4 LT |
5046 | |
5047 | /* Don't want a node to appear more than once */ | |
5048 | if (node_isset(n, *used_node_mask)) | |
5049 | continue; | |
5050 | ||
1da177e4 LT |
5051 | /* Use the distance array to find the distance */ |
5052 | val = node_distance(node, n); | |
5053 | ||
4cf808eb LT |
5054 | /* Penalize nodes under us ("prefer the next node") */ |
5055 | val += (n < node); | |
5056 | ||
1da177e4 | 5057 | /* Give preference to headless and unused nodes */ |
b630749f | 5058 | if (!cpumask_empty(cpumask_of_node(n))) |
1da177e4 LT |
5059 | val += PENALTY_FOR_NODE_WITH_CPUS; |
5060 | ||
5061 | /* Slight preference for less loaded node */ | |
37931324 | 5062 | val *= MAX_NUMNODES; |
1da177e4 LT |
5063 | val += node_load[n]; |
5064 | ||
5065 | if (val < min_val) { | |
5066 | min_val = val; | |
5067 | best_node = n; | |
5068 | } | |
5069 | } | |
5070 | ||
5071 | if (best_node >= 0) | |
5072 | node_set(best_node, *used_node_mask); | |
5073 | ||
5074 | return best_node; | |
5075 | } | |
5076 | ||
f0c0b2b8 KH |
5077 | |
5078 | /* | |
5079 | * Build zonelists ordered by node and zones within node. | |
5080 | * This results in maximum locality--normal zone overflows into local | |
5081 | * DMA zone, if any--but risks exhausting DMA zone. | |
5082 | */ | |
9d3be21b MH |
5083 | static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order, |
5084 | unsigned nr_nodes) | |
1da177e4 | 5085 | { |
9d3be21b MH |
5086 | struct zoneref *zonerefs; |
5087 | int i; | |
5088 | ||
5089 | zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs; | |
5090 | ||
5091 | for (i = 0; i < nr_nodes; i++) { | |
5092 | int nr_zones; | |
5093 | ||
5094 | pg_data_t *node = NODE_DATA(node_order[i]); | |
f0c0b2b8 | 5095 | |
9d3be21b MH |
5096 | nr_zones = build_zonerefs_node(node, zonerefs); |
5097 | zonerefs += nr_zones; | |
5098 | } | |
5099 | zonerefs->zone = NULL; | |
5100 | zonerefs->zone_idx = 0; | |
f0c0b2b8 KH |
5101 | } |
5102 | ||
523b9458 CL |
5103 | /* |
5104 | * Build gfp_thisnode zonelists | |
5105 | */ | |
5106 | static void build_thisnode_zonelists(pg_data_t *pgdat) | |
5107 | { | |
9d3be21b MH |
5108 | struct zoneref *zonerefs; |
5109 | int nr_zones; | |
523b9458 | 5110 | |
9d3be21b MH |
5111 | zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs; |
5112 | nr_zones = build_zonerefs_node(pgdat, zonerefs); | |
5113 | zonerefs += nr_zones; | |
5114 | zonerefs->zone = NULL; | |
5115 | zonerefs->zone_idx = 0; | |
523b9458 CL |
5116 | } |
5117 | ||
f0c0b2b8 KH |
5118 | /* |
5119 | * Build zonelists ordered by zone and nodes within zones. | |
5120 | * This results in conserving DMA zone[s] until all Normal memory is | |
5121 | * exhausted, but results in overflowing to remote node while memory | |
5122 | * may still exist in local DMA zone. | |
5123 | */ | |
f0c0b2b8 | 5124 | |
f0c0b2b8 KH |
5125 | static void build_zonelists(pg_data_t *pgdat) |
5126 | { | |
9d3be21b | 5127 | static int node_order[MAX_NUMNODES]; |
37931324 | 5128 | int node, nr_nodes = 0; |
d0ddf49b | 5129 | nodemask_t used_mask = NODE_MASK_NONE; |
f0c0b2b8 | 5130 | int local_node, prev_node; |
1da177e4 LT |
5131 | |
5132 | /* NUMA-aware ordering of nodes */ | |
5133 | local_node = pgdat->node_id; | |
1da177e4 | 5134 | prev_node = local_node; |
f0c0b2b8 | 5135 | |
f0c0b2b8 | 5136 | memset(node_order, 0, sizeof(node_order)); |
1da177e4 LT |
5137 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { |
5138 | /* | |
5139 | * We don't want to pressure a particular node. | |
5140 | * So adding penalty to the first node in same | |
5141 | * distance group to make it round-robin. | |
5142 | */ | |
957f822a DR |
5143 | if (node_distance(local_node, node) != |
5144 | node_distance(local_node, prev_node)) | |
37931324 | 5145 | node_load[node] += 1; |
f0c0b2b8 | 5146 | |
9d3be21b | 5147 | node_order[nr_nodes++] = node; |
1da177e4 | 5148 | prev_node = node; |
1da177e4 | 5149 | } |
523b9458 | 5150 | |
9d3be21b | 5151 | build_zonelists_in_node_order(pgdat, node_order, nr_nodes); |
523b9458 | 5152 | build_thisnode_zonelists(pgdat); |
6cf25392 BR |
5153 | pr_info("Fallback order for Node %d: ", local_node); |
5154 | for (node = 0; node < nr_nodes; node++) | |
5155 | pr_cont("%d ", node_order[node]); | |
5156 | pr_cont("\n"); | |
1da177e4 LT |
5157 | } |
5158 | ||
7aac7898 LS |
5159 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES |
5160 | /* | |
5161 | * Return node id of node used for "local" allocations. | |
5162 | * I.e., first node id of first zone in arg node's generic zonelist. | |
5163 | * Used for initializing percpu 'numa_mem', which is used primarily | |
5164 | * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. | |
5165 | */ | |
5166 | int local_memory_node(int node) | |
5167 | { | |
c33d6c06 | 5168 | struct zoneref *z; |
7aac7898 | 5169 | |
c33d6c06 | 5170 | z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL), |
7aac7898 | 5171 | gfp_zone(GFP_KERNEL), |
c33d6c06 | 5172 | NULL); |
c1093b74 | 5173 | return zone_to_nid(z->zone); |
7aac7898 LS |
5174 | } |
5175 | #endif | |
f0c0b2b8 | 5176 | |
6423aa81 JK |
5177 | static void setup_min_unmapped_ratio(void); |
5178 | static void setup_min_slab_ratio(void); | |
1da177e4 LT |
5179 | #else /* CONFIG_NUMA */ |
5180 | ||
f0c0b2b8 | 5181 | static void build_zonelists(pg_data_t *pgdat) |
1da177e4 | 5182 | { |
19655d34 | 5183 | int node, local_node; |
9d3be21b MH |
5184 | struct zoneref *zonerefs; |
5185 | int nr_zones; | |
1da177e4 LT |
5186 | |
5187 | local_node = pgdat->node_id; | |
1da177e4 | 5188 | |
9d3be21b MH |
5189 | zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs; |
5190 | nr_zones = build_zonerefs_node(pgdat, zonerefs); | |
5191 | zonerefs += nr_zones; | |
1da177e4 | 5192 | |
54a6eb5c MG |
5193 | /* |
5194 | * Now we build the zonelist so that it contains the zones | |
5195 | * of all the other nodes. | |
5196 | * We don't want to pressure a particular node, so when | |
5197 | * building the zones for node N, we make sure that the | |
5198 | * zones coming right after the local ones are those from | |
5199 | * node N+1 (modulo N) | |
5200 | */ | |
5201 | for (node = local_node + 1; node < MAX_NUMNODES; node++) { | |
5202 | if (!node_online(node)) | |
5203 | continue; | |
9d3be21b MH |
5204 | nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs); |
5205 | zonerefs += nr_zones; | |
1da177e4 | 5206 | } |
54a6eb5c MG |
5207 | for (node = 0; node < local_node; node++) { |
5208 | if (!node_online(node)) | |
5209 | continue; | |
9d3be21b MH |
5210 | nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs); |
5211 | zonerefs += nr_zones; | |
54a6eb5c MG |
5212 | } |
5213 | ||
9d3be21b MH |
5214 | zonerefs->zone = NULL; |
5215 | zonerefs->zone_idx = 0; | |
1da177e4 LT |
5216 | } |
5217 | ||
5218 | #endif /* CONFIG_NUMA */ | |
5219 | ||
99dcc3e5 CL |
5220 | /* |
5221 | * Boot pageset table. One per cpu which is going to be used for all | |
5222 | * zones and all nodes. The parameters will be set in such a way | |
5223 | * that an item put on a list will immediately be handed over to | |
5224 | * the buddy list. This is safe since pageset manipulation is done | |
5225 | * with interrupts disabled. | |
5226 | * | |
5227 | * The boot_pagesets must be kept even after bootup is complete for | |
5228 | * unused processors and/or zones. They do play a role for bootstrapping | |
5229 | * hotplugged processors. | |
5230 | * | |
5231 | * zoneinfo_show() and maybe other functions do | |
5232 | * not check if the processor is online before following the pageset pointer. | |
5233 | * Other parts of the kernel may not check if the zone is available. | |
5234 | */ | |
28f836b6 | 5235 | static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats); |
952eaf81 VB |
5236 | /* These effectively disable the pcplists in the boot pageset completely */ |
5237 | #define BOOT_PAGESET_HIGH 0 | |
5238 | #define BOOT_PAGESET_BATCH 1 | |
28f836b6 MG |
5239 | static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset); |
5240 | static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats); | |
99dcc3e5 | 5241 | |
11cd8638 | 5242 | static void __build_all_zonelists(void *data) |
1da177e4 | 5243 | { |
6811378e | 5244 | int nid; |
afb6ebb3 | 5245 | int __maybe_unused cpu; |
9adb62a5 | 5246 | pg_data_t *self = data; |
1007843a | 5247 | unsigned long flags; |
b93e0f32 | 5248 | |
1007843a | 5249 | /* |
a2ebb515 SAS |
5250 | * The zonelist_update_seq must be acquired with irqsave because the |
5251 | * reader can be invoked from IRQ with GFP_ATOMIC. | |
1007843a | 5252 | */ |
a2ebb515 | 5253 | write_seqlock_irqsave(&zonelist_update_seq, flags); |
1007843a | 5254 | /* |
a2ebb515 SAS |
5255 | * Also disable synchronous printk() to prevent any printk() from |
5256 | * trying to hold port->lock, for | |
1007843a TH |
5257 | * tty_insert_flip_string_and_push_buffer() on other CPU might be |
5258 | * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held. | |
5259 | */ | |
5260 | printk_deferred_enter(); | |
9276b1bc | 5261 | |
7f9cfb31 BL |
5262 | #ifdef CONFIG_NUMA |
5263 | memset(node_load, 0, sizeof(node_load)); | |
5264 | #endif | |
9adb62a5 | 5265 | |
c1152583 WY |
5266 | /* |
5267 | * This node is hotadded and no memory is yet present. So just | |
5268 | * building zonelists is fine - no need to touch other nodes. | |
5269 | */ | |
9adb62a5 JL |
5270 | if (self && !node_online(self->node_id)) { |
5271 | build_zonelists(self); | |
c1152583 | 5272 | } else { |
09f49dca MH |
5273 | /* |
5274 | * All possible nodes have pgdat preallocated | |
5275 | * in free_area_init | |
5276 | */ | |
5277 | for_each_node(nid) { | |
c1152583 | 5278 | pg_data_t *pgdat = NODE_DATA(nid); |
7ea1530a | 5279 | |
c1152583 WY |
5280 | build_zonelists(pgdat); |
5281 | } | |
99dcc3e5 | 5282 | |
7aac7898 LS |
5283 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES |
5284 | /* | |
5285 | * We now know the "local memory node" for each node-- | |
5286 | * i.e., the node of the first zone in the generic zonelist. | |
5287 | * Set up numa_mem percpu variable for on-line cpus. During | |
5288 | * boot, only the boot cpu should be on-line; we'll init the | |
5289 | * secondary cpus' numa_mem as they come on-line. During | |
5290 | * node/memory hotplug, we'll fixup all on-line cpus. | |
5291 | */ | |
d9c9a0b9 | 5292 | for_each_online_cpu(cpu) |
7aac7898 | 5293 | set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); |
afb6ebb3 | 5294 | #endif |
d9c9a0b9 | 5295 | } |
b93e0f32 | 5296 | |
1007843a | 5297 | printk_deferred_exit(); |
a2ebb515 | 5298 | write_sequnlock_irqrestore(&zonelist_update_seq, flags); |
6811378e YG |
5299 | } |
5300 | ||
061f67bc RV |
5301 | static noinline void __init |
5302 | build_all_zonelists_init(void) | |
5303 | { | |
afb6ebb3 MH |
5304 | int cpu; |
5305 | ||
061f67bc | 5306 | __build_all_zonelists(NULL); |
afb6ebb3 MH |
5307 | |
5308 | /* | |
5309 | * Initialize the boot_pagesets that are going to be used | |
5310 | * for bootstrapping processors. The real pagesets for | |
5311 | * each zone will be allocated later when the per cpu | |
5312 | * allocator is available. | |
5313 | * | |
5314 | * boot_pagesets are used also for bootstrapping offline | |
5315 | * cpus if the system is already booted because the pagesets | |
5316 | * are needed to initialize allocators on a specific cpu too. | |
5317 | * F.e. the percpu allocator needs the page allocator which | |
5318 | * needs the percpu allocator in order to allocate its pagesets | |
5319 | * (a chicken-egg dilemma). | |
5320 | */ | |
5321 | for_each_possible_cpu(cpu) | |
28f836b6 | 5322 | per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu)); |
afb6ebb3 | 5323 | |
061f67bc RV |
5324 | mminit_verify_zonelist(); |
5325 | cpuset_init_current_mems_allowed(); | |
5326 | } | |
5327 | ||
4eaf3f64 | 5328 | /* |
4eaf3f64 | 5329 | * unless system_state == SYSTEM_BOOTING. |
061f67bc | 5330 | * |
72675e13 | 5331 | * __ref due to call of __init annotated helper build_all_zonelists_init |
061f67bc | 5332 | * [protected by SYSTEM_BOOTING]. |
4eaf3f64 | 5333 | */ |
72675e13 | 5334 | void __ref build_all_zonelists(pg_data_t *pgdat) |
6811378e | 5335 | { |
0a18e607 DH |
5336 | unsigned long vm_total_pages; |
5337 | ||
6811378e | 5338 | if (system_state == SYSTEM_BOOTING) { |
061f67bc | 5339 | build_all_zonelists_init(); |
6811378e | 5340 | } else { |
11cd8638 | 5341 | __build_all_zonelists(pgdat); |
6811378e YG |
5342 | /* cpuset refresh routine should be here */ |
5343 | } | |
56b9413b DH |
5344 | /* Get the number of free pages beyond high watermark in all zones. */ |
5345 | vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); | |
9ef9acb0 MG |
5346 | /* |
5347 | * Disable grouping by mobility if the number of pages in the | |
5348 | * system is too low to allow the mechanism to work. It would be | |
5349 | * more accurate, but expensive to check per-zone. This check is | |
5350 | * made on memory-hotadd so a system can start with mobility | |
5351 | * disabled and enable it later | |
5352 | */ | |
d9c23400 | 5353 | if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) |
9ef9acb0 MG |
5354 | page_group_by_mobility_disabled = 1; |
5355 | else | |
5356 | page_group_by_mobility_disabled = 0; | |
5357 | ||
ce0725f7 | 5358 | pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n", |
756a025f | 5359 | nr_online_nodes, |
756a025f JP |
5360 | page_group_by_mobility_disabled ? "off" : "on", |
5361 | vm_total_pages); | |
f0c0b2b8 | 5362 | #ifdef CONFIG_NUMA |
f88dfff5 | 5363 | pr_info("Policy zone: %s\n", zone_names[policy_zone]); |
f0c0b2b8 | 5364 | #endif |
1da177e4 LT |
5365 | } |
5366 | ||
9420f89d | 5367 | static int zone_batchsize(struct zone *zone) |
1da177e4 | 5368 | { |
9420f89d MRI |
5369 | #ifdef CONFIG_MMU |
5370 | int batch; | |
1da177e4 | 5371 | |
9420f89d MRI |
5372 | /* |
5373 | * The number of pages to batch allocate is either ~0.1% | |
5374 | * of the zone or 1MB, whichever is smaller. The batch | |
5375 | * size is striking a balance between allocation latency | |
5376 | * and zone lock contention. | |
5377 | */ | |
5378 | batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE); | |
5379 | batch /= 4; /* We effectively *= 4 below */ | |
5380 | if (batch < 1) | |
5381 | batch = 1; | |
22b31eec | 5382 | |
4b94ffdc | 5383 | /* |
9420f89d MRI |
5384 | * Clamp the batch to a 2^n - 1 value. Having a power |
5385 | * of 2 value was found to be more likely to have | |
5386 | * suboptimal cache aliasing properties in some cases. | |
5387 | * | |
5388 | * For example if 2 tasks are alternately allocating | |
5389 | * batches of pages, one task can end up with a lot | |
5390 | * of pages of one half of the possible page colors | |
5391 | * and the other with pages of the other colors. | |
4b94ffdc | 5392 | */ |
9420f89d | 5393 | batch = rounddown_pow_of_two(batch + batch/2) - 1; |
966cf44f | 5394 | |
9420f89d | 5395 | return batch; |
3a6be87f DH |
5396 | |
5397 | #else | |
5398 | /* The deferral and batching of frees should be suppressed under NOMMU | |
5399 | * conditions. | |
5400 | * | |
5401 | * The problem is that NOMMU needs to be able to allocate large chunks | |
5402 | * of contiguous memory as there's no hardware page translation to | |
5403 | * assemble apparent contiguous memory from discontiguous pages. | |
5404 | * | |
5405 | * Queueing large contiguous runs of pages for batching, however, | |
5406 | * causes the pages to actually be freed in smaller chunks. As there | |
5407 | * can be a significant delay between the individual batches being | |
5408 | * recycled, this leads to the once large chunks of space being | |
5409 | * fragmented and becoming unavailable for high-order allocations. | |
5410 | */ | |
5411 | return 0; | |
5412 | #endif | |
e7c8d5c9 CL |
5413 | } |
5414 | ||
e95d372c | 5415 | static int percpu_pagelist_high_fraction; |
90b41691 HY |
5416 | static int zone_highsize(struct zone *zone, int batch, int cpu_online, |
5417 | int high_fraction) | |
b92ca18e | 5418 | { |
9420f89d MRI |
5419 | #ifdef CONFIG_MMU |
5420 | int high; | |
5421 | int nr_split_cpus; | |
5422 | unsigned long total_pages; | |
c13291a5 | 5423 | |
90b41691 | 5424 | if (!high_fraction) { |
2a1e274a | 5425 | /* |
9420f89d MRI |
5426 | * By default, the high value of the pcp is based on the zone |
5427 | * low watermark so that if they are full then background | |
5428 | * reclaim will not be started prematurely. | |
2a1e274a | 5429 | */ |
9420f89d MRI |
5430 | total_pages = low_wmark_pages(zone); |
5431 | } else { | |
2a1e274a | 5432 | /* |
9420f89d MRI |
5433 | * If percpu_pagelist_high_fraction is configured, the high |
5434 | * value is based on a fraction of the managed pages in the | |
5435 | * zone. | |
2a1e274a | 5436 | */ |
90b41691 | 5437 | total_pages = zone_managed_pages(zone) / high_fraction; |
2a1e274a MG |
5438 | } |
5439 | ||
5440 | /* | |
9420f89d MRI |
5441 | * Split the high value across all online CPUs local to the zone. Note |
5442 | * that early in boot that CPUs may not be online yet and that during | |
5443 | * CPU hotplug that the cpumask is not yet updated when a CPU is being | |
90b41691 HY |
5444 | * onlined. For memory nodes that have no CPUs, split the high value |
5445 | * across all online CPUs to mitigate the risk that reclaim is triggered | |
9420f89d | 5446 | * prematurely due to pages stored on pcp lists. |
2a1e274a | 5447 | */ |
9420f89d MRI |
5448 | nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online; |
5449 | if (!nr_split_cpus) | |
5450 | nr_split_cpus = num_online_cpus(); | |
5451 | high = total_pages / nr_split_cpus; | |
2a1e274a | 5452 | |
9420f89d MRI |
5453 | /* |
5454 | * Ensure high is at least batch*4. The multiple is based on the | |
5455 | * historical relationship between high and batch. | |
5456 | */ | |
5457 | high = max(high, batch << 2); | |
37b07e41 | 5458 | |
9420f89d MRI |
5459 | return high; |
5460 | #else | |
5461 | return 0; | |
5462 | #endif | |
37b07e41 LS |
5463 | } |
5464 | ||
51930df5 | 5465 | /* |
9420f89d MRI |
5466 | * pcp->high and pcp->batch values are related and generally batch is lower |
5467 | * than high. They are also related to pcp->count such that count is lower | |
5468 | * than high, and as soon as it reaches high, the pcplist is flushed. | |
5469 | * | |
5470 | * However, guaranteeing these relations at all times would require e.g. write | |
5471 | * barriers here but also careful usage of read barriers at the read side, and | |
5472 | * thus be prone to error and bad for performance. Thus the update only prevents | |
90b41691 HY |
5473 | * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max |
5474 | * should ensure they can cope with those fields changing asynchronously, and | |
5475 | * fully trust only the pcp->count field on the local CPU with interrupts | |
5476 | * disabled. | |
9420f89d MRI |
5477 | * |
5478 | * mutex_is_locked(&pcp_batch_high_lock) required when calling this function | |
5479 | * outside of boot time (or some other assurance that no concurrent updaters | |
5480 | * exist). | |
51930df5 | 5481 | */ |
90b41691 HY |
5482 | static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min, |
5483 | unsigned long high_max, unsigned long batch) | |
51930df5 | 5484 | { |
9420f89d | 5485 | WRITE_ONCE(pcp->batch, batch); |
90b41691 HY |
5486 | WRITE_ONCE(pcp->high_min, high_min); |
5487 | WRITE_ONCE(pcp->high_max, high_max); | |
51930df5 MR |
5488 | } |
5489 | ||
9420f89d | 5490 | static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats) |
c713216d | 5491 | { |
9420f89d | 5492 | int pindex; |
90cae1fe | 5493 | |
9420f89d MRI |
5494 | memset(pcp, 0, sizeof(*pcp)); |
5495 | memset(pzstats, 0, sizeof(*pzstats)); | |
90cae1fe | 5496 | |
9420f89d MRI |
5497 | spin_lock_init(&pcp->lock); |
5498 | for (pindex = 0; pindex < NR_PCP_LISTS; pindex++) | |
5499 | INIT_LIST_HEAD(&pcp->lists[pindex]); | |
2a1e274a | 5500 | |
9420f89d MRI |
5501 | /* |
5502 | * Set batch and high values safe for a boot pageset. A true percpu | |
5503 | * pageset's initialization will update them subsequently. Here we don't | |
5504 | * need to be as careful as pageset_update() as nobody can access the | |
5505 | * pageset yet. | |
5506 | */ | |
90b41691 HY |
5507 | pcp->high_min = BOOT_PAGESET_HIGH; |
5508 | pcp->high_max = BOOT_PAGESET_HIGH; | |
9420f89d | 5509 | pcp->batch = BOOT_PAGESET_BATCH; |
6ccdcb6d | 5510 | pcp->free_count = 0; |
9420f89d | 5511 | } |
c713216d | 5512 | |
90b41691 HY |
5513 | static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min, |
5514 | unsigned long high_max, unsigned long batch) | |
9420f89d MRI |
5515 | { |
5516 | struct per_cpu_pages *pcp; | |
5517 | int cpu; | |
2a1e274a | 5518 | |
9420f89d MRI |
5519 | for_each_possible_cpu(cpu) { |
5520 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); | |
90b41691 | 5521 | pageset_update(pcp, high_min, high_max, batch); |
2a1e274a | 5522 | } |
9420f89d | 5523 | } |
c713216d | 5524 | |
9420f89d MRI |
5525 | /* |
5526 | * Calculate and set new high and batch values for all per-cpu pagesets of a | |
5527 | * zone based on the zone's size. | |
5528 | */ | |
5529 | static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online) | |
5530 | { | |
90b41691 | 5531 | int new_high_min, new_high_max, new_batch; |
09f49dca | 5532 | |
9420f89d | 5533 | new_batch = max(1, zone_batchsize(zone)); |
90b41691 HY |
5534 | if (percpu_pagelist_high_fraction) { |
5535 | new_high_min = zone_highsize(zone, new_batch, cpu_online, | |
5536 | percpu_pagelist_high_fraction); | |
5537 | /* | |
5538 | * PCP high is tuned manually, disable auto-tuning via | |
5539 | * setting high_min and high_max to the manual value. | |
5540 | */ | |
5541 | new_high_max = new_high_min; | |
5542 | } else { | |
5543 | new_high_min = zone_highsize(zone, new_batch, cpu_online, 0); | |
5544 | new_high_max = zone_highsize(zone, new_batch, cpu_online, | |
5545 | MIN_PERCPU_PAGELIST_HIGH_FRACTION); | |
5546 | } | |
09f49dca | 5547 | |
90b41691 HY |
5548 | if (zone->pageset_high_min == new_high_min && |
5549 | zone->pageset_high_max == new_high_max && | |
9420f89d MRI |
5550 | zone->pageset_batch == new_batch) |
5551 | return; | |
37b07e41 | 5552 | |
90b41691 HY |
5553 | zone->pageset_high_min = new_high_min; |
5554 | zone->pageset_high_max = new_high_max; | |
9420f89d | 5555 | zone->pageset_batch = new_batch; |
122e093c | 5556 | |
90b41691 HY |
5557 | __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max, |
5558 | new_batch); | |
c713216d | 5559 | } |
2a1e274a | 5560 | |
9420f89d | 5561 | void __meminit setup_zone_pageset(struct zone *zone) |
2a1e274a | 5562 | { |
9420f89d | 5563 | int cpu; |
2a1e274a | 5564 | |
9420f89d MRI |
5565 | /* Size may be 0 on !SMP && !NUMA */ |
5566 | if (sizeof(struct per_cpu_zonestat) > 0) | |
5567 | zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat); | |
2a1e274a | 5568 | |
9420f89d MRI |
5569 | zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages); |
5570 | for_each_possible_cpu(cpu) { | |
5571 | struct per_cpu_pages *pcp; | |
5572 | struct per_cpu_zonestat *pzstats; | |
2a1e274a | 5573 | |
9420f89d MRI |
5574 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); |
5575 | pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu); | |
5576 | per_cpu_pages_init(pcp, pzstats); | |
a5c6d650 | 5577 | } |
9420f89d MRI |
5578 | |
5579 | zone_set_pageset_high_and_batch(zone, 0); | |
2a1e274a | 5580 | } |
ed7ed365 | 5581 | |
7e63efef | 5582 | /* |
9420f89d MRI |
5583 | * The zone indicated has a new number of managed_pages; batch sizes and percpu |
5584 | * page high values need to be recalculated. | |
7e63efef | 5585 | */ |
9420f89d | 5586 | static void zone_pcp_update(struct zone *zone, int cpu_online) |
7e63efef | 5587 | { |
9420f89d MRI |
5588 | mutex_lock(&pcp_batch_high_lock); |
5589 | zone_set_pageset_high_and_batch(zone, cpu_online); | |
5590 | mutex_unlock(&pcp_batch_high_lock); | |
7e63efef MG |
5591 | } |
5592 | ||
5cec4eb7 | 5593 | static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu) |
362d37a1 | 5594 | { |
362d37a1 HY |
5595 | struct per_cpu_pages *pcp; |
5596 | struct cpu_cacheinfo *cci; | |
5597 | ||
5cec4eb7 HY |
5598 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); |
5599 | cci = get_cpu_cacheinfo(cpu); | |
5600 | /* | |
5601 | * If data cache slice of CPU is large enough, "pcp->batch" | |
5602 | * pages can be preserved in PCP before draining PCP for | |
5603 | * consecutive high-order pages freeing without allocation. | |
5604 | * This can reduce zone lock contention without hurting | |
5605 | * cache-hot pages sharing. | |
5606 | */ | |
5607 | spin_lock(&pcp->lock); | |
5608 | if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch) | |
5609 | pcp->flags |= PCPF_FREE_HIGH_BATCH; | |
5610 | else | |
5611 | pcp->flags &= ~PCPF_FREE_HIGH_BATCH; | |
5612 | spin_unlock(&pcp->lock); | |
362d37a1 HY |
5613 | } |
5614 | ||
5cec4eb7 | 5615 | void setup_pcp_cacheinfo(unsigned int cpu) |
362d37a1 HY |
5616 | { |
5617 | struct zone *zone; | |
5618 | ||
5619 | for_each_populated_zone(zone) | |
5cec4eb7 | 5620 | zone_pcp_update_cacheinfo(zone, cpu); |
362d37a1 HY |
5621 | } |
5622 | ||
7e63efef | 5623 | /* |
9420f89d MRI |
5624 | * Allocate per cpu pagesets and initialize them. |
5625 | * Before this call only boot pagesets were available. | |
7e63efef | 5626 | */ |
9420f89d | 5627 | void __init setup_per_cpu_pageset(void) |
7e63efef | 5628 | { |
9420f89d MRI |
5629 | struct pglist_data *pgdat; |
5630 | struct zone *zone; | |
5631 | int __maybe_unused cpu; | |
5632 | ||
5633 | for_each_populated_zone(zone) | |
5634 | setup_zone_pageset(zone); | |
5635 | ||
5636 | #ifdef CONFIG_NUMA | |
5637 | /* | |
5638 | * Unpopulated zones continue using the boot pagesets. | |
5639 | * The numa stats for these pagesets need to be reset. | |
5640 | * Otherwise, they will end up skewing the stats of | |
5641 | * the nodes these zones are associated with. | |
5642 | */ | |
5643 | for_each_possible_cpu(cpu) { | |
5644 | struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu); | |
5645 | memset(pzstats->vm_numa_event, 0, | |
5646 | sizeof(pzstats->vm_numa_event)); | |
5647 | } | |
5648 | #endif | |
5649 | ||
5650 | for_each_online_pgdat(pgdat) | |
5651 | pgdat->per_cpu_nodestats = | |
5652 | alloc_percpu(struct per_cpu_nodestat); | |
7e63efef MG |
5653 | } |
5654 | ||
9420f89d MRI |
5655 | __meminit void zone_pcp_init(struct zone *zone) |
5656 | { | |
5657 | /* | |
5658 | * per cpu subsystem is not up at this point. The following code | |
5659 | * relies on the ability of the linker to provide the | |
5660 | * offset of a (static) per cpu variable into the per cpu area. | |
5661 | */ | |
5662 | zone->per_cpu_pageset = &boot_pageset; | |
5663 | zone->per_cpu_zonestats = &boot_zonestats; | |
90b41691 HY |
5664 | zone->pageset_high_min = BOOT_PAGESET_HIGH; |
5665 | zone->pageset_high_max = BOOT_PAGESET_HIGH; | |
9420f89d MRI |
5666 | zone->pageset_batch = BOOT_PAGESET_BATCH; |
5667 | ||
5668 | if (populated_zone(zone)) | |
5669 | pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name, | |
5670 | zone->present_pages, zone_batchsize(zone)); | |
5671 | } | |
ed7ed365 | 5672 | |
c3d5f5f0 JL |
5673 | void adjust_managed_page_count(struct page *page, long count) |
5674 | { | |
9705bea5 | 5675 | atomic_long_add(count, &page_zone(page)->managed_pages); |
ca79b0c2 | 5676 | totalram_pages_add(count); |
3dcc0571 JL |
5677 | #ifdef CONFIG_HIGHMEM |
5678 | if (PageHighMem(page)) | |
ca79b0c2 | 5679 | totalhigh_pages_add(count); |
3dcc0571 | 5680 | #endif |
c3d5f5f0 | 5681 | } |
3dcc0571 | 5682 | EXPORT_SYMBOL(adjust_managed_page_count); |
c3d5f5f0 | 5683 | |
e5cb113f | 5684 | unsigned long free_reserved_area(void *start, void *end, int poison, const char *s) |
69afade7 | 5685 | { |
11199692 JL |
5686 | void *pos; |
5687 | unsigned long pages = 0; | |
69afade7 | 5688 | |
11199692 JL |
5689 | start = (void *)PAGE_ALIGN((unsigned long)start); |
5690 | end = (void *)((unsigned long)end & PAGE_MASK); | |
5691 | for (pos = start; pos < end; pos += PAGE_SIZE, pages++) { | |
0d834328 DH |
5692 | struct page *page = virt_to_page(pos); |
5693 | void *direct_map_addr; | |
5694 | ||
5695 | /* | |
5696 | * 'direct_map_addr' might be different from 'pos' | |
5697 | * because some architectures' virt_to_page() | |
5698 | * work with aliases. Getting the direct map | |
5699 | * address ensures that we get a _writeable_ | |
5700 | * alias for the memset(). | |
5701 | */ | |
5702 | direct_map_addr = page_address(page); | |
c746170d VF |
5703 | /* |
5704 | * Perform a kasan-unchecked memset() since this memory | |
5705 | * has not been initialized. | |
5706 | */ | |
5707 | direct_map_addr = kasan_reset_tag(direct_map_addr); | |
dbe67df4 | 5708 | if ((unsigned int)poison <= 0xFF) |
0d834328 DH |
5709 | memset(direct_map_addr, poison, PAGE_SIZE); |
5710 | ||
5711 | free_reserved_page(page); | |
69afade7 JL |
5712 | } |
5713 | ||
5714 | if (pages && s) | |
ff7ed9e4 | 5715 | pr_info("Freeing %s memory: %ldK\n", s, K(pages)); |
69afade7 JL |
5716 | |
5717 | return pages; | |
5718 | } | |
5719 | ||
005fd4bb | 5720 | static int page_alloc_cpu_dead(unsigned int cpu) |
1da177e4 | 5721 | { |
04f8cfea | 5722 | struct zone *zone; |
1da177e4 | 5723 | |
005fd4bb | 5724 | lru_add_drain_cpu(cpu); |
96f97c43 | 5725 | mlock_drain_remote(cpu); |
005fd4bb | 5726 | drain_pages(cpu); |
9f8f2172 | 5727 | |
005fd4bb SAS |
5728 | /* |
5729 | * Spill the event counters of the dead processor | |
5730 | * into the current processors event counters. | |
5731 | * This artificially elevates the count of the current | |
5732 | * processor. | |
5733 | */ | |
5734 | vm_events_fold_cpu(cpu); | |
9f8f2172 | 5735 | |
005fd4bb SAS |
5736 | /* |
5737 | * Zero the differential counters of the dead processor | |
5738 | * so that the vm statistics are consistent. | |
5739 | * | |
5740 | * This is only okay since the processor is dead and cannot | |
5741 | * race with what we are doing. | |
5742 | */ | |
5743 | cpu_vm_stats_fold(cpu); | |
04f8cfea MG |
5744 | |
5745 | for_each_populated_zone(zone) | |
5746 | zone_pcp_update(zone, 0); | |
5747 | ||
5748 | return 0; | |
5749 | } | |
5750 | ||
5751 | static int page_alloc_cpu_online(unsigned int cpu) | |
5752 | { | |
5753 | struct zone *zone; | |
5754 | ||
5755 | for_each_populated_zone(zone) | |
5756 | zone_pcp_update(zone, 1); | |
005fd4bb | 5757 | return 0; |
1da177e4 | 5758 | } |
1da177e4 | 5759 | |
c4fbed4b | 5760 | void __init page_alloc_init_cpuhp(void) |
1da177e4 | 5761 | { |
005fd4bb SAS |
5762 | int ret; |
5763 | ||
04f8cfea MG |
5764 | ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC, |
5765 | "mm/page_alloc:pcp", | |
5766 | page_alloc_cpu_online, | |
005fd4bb SAS |
5767 | page_alloc_cpu_dead); |
5768 | WARN_ON(ret < 0); | |
1da177e4 LT |
5769 | } |
5770 | ||
cb45b0e9 | 5771 | /* |
34b10060 | 5772 | * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio |
cb45b0e9 HA |
5773 | * or min_free_kbytes changes. |
5774 | */ | |
5775 | static void calculate_totalreserve_pages(void) | |
5776 | { | |
5777 | struct pglist_data *pgdat; | |
5778 | unsigned long reserve_pages = 0; | |
2f6726e5 | 5779 | enum zone_type i, j; |
cb45b0e9 HA |
5780 | |
5781 | for_each_online_pgdat(pgdat) { | |
281e3726 MG |
5782 | |
5783 | pgdat->totalreserve_pages = 0; | |
5784 | ||
cb45b0e9 HA |
5785 | for (i = 0; i < MAX_NR_ZONES; i++) { |
5786 | struct zone *zone = pgdat->node_zones + i; | |
3484b2de | 5787 | long max = 0; |
9705bea5 | 5788 | unsigned long managed_pages = zone_managed_pages(zone); |
cb45b0e9 HA |
5789 | |
5790 | /* Find valid and maximum lowmem_reserve in the zone */ | |
5791 | for (j = i; j < MAX_NR_ZONES; j++) { | |
5792 | if (zone->lowmem_reserve[j] > max) | |
5793 | max = zone->lowmem_reserve[j]; | |
5794 | } | |
5795 | ||
41858966 MG |
5796 | /* we treat the high watermark as reserved pages. */ |
5797 | max += high_wmark_pages(zone); | |
cb45b0e9 | 5798 | |
3d6357de AK |
5799 | if (max > managed_pages) |
5800 | max = managed_pages; | |
a8d01437 | 5801 | |
281e3726 | 5802 | pgdat->totalreserve_pages += max; |
a8d01437 | 5803 | |
cb45b0e9 HA |
5804 | reserve_pages += max; |
5805 | } | |
5806 | } | |
5807 | totalreserve_pages = reserve_pages; | |
5808 | } | |
5809 | ||
1da177e4 LT |
5810 | /* |
5811 | * setup_per_zone_lowmem_reserve - called whenever | |
34b10060 | 5812 | * sysctl_lowmem_reserve_ratio changes. Ensures that each zone |
1da177e4 LT |
5813 | * has a correct pages reserved value, so an adequate number of |
5814 | * pages are left in the zone after a successful __alloc_pages(). | |
5815 | */ | |
5816 | static void setup_per_zone_lowmem_reserve(void) | |
5817 | { | |
5818 | struct pglist_data *pgdat; | |
470c61d7 | 5819 | enum zone_type i, j; |
1da177e4 | 5820 | |
ec936fc5 | 5821 | for_each_online_pgdat(pgdat) { |
470c61d7 LS |
5822 | for (i = 0; i < MAX_NR_ZONES - 1; i++) { |
5823 | struct zone *zone = &pgdat->node_zones[i]; | |
5824 | int ratio = sysctl_lowmem_reserve_ratio[i]; | |
5825 | bool clear = !ratio || !zone_managed_pages(zone); | |
5826 | unsigned long managed_pages = 0; | |
5827 | ||
5828 | for (j = i + 1; j < MAX_NR_ZONES; j++) { | |
f7ec1044 LS |
5829 | struct zone *upper_zone = &pgdat->node_zones[j]; |
5830 | ||
5831 | managed_pages += zone_managed_pages(upper_zone); | |
470c61d7 | 5832 | |
f7ec1044 LS |
5833 | if (clear) |
5834 | zone->lowmem_reserve[j] = 0; | |
5835 | else | |
470c61d7 | 5836 | zone->lowmem_reserve[j] = managed_pages / ratio; |
1da177e4 LT |
5837 | } |
5838 | } | |
5839 | } | |
cb45b0e9 HA |
5840 | |
5841 | /* update totalreserve_pages */ | |
5842 | calculate_totalreserve_pages(); | |
1da177e4 LT |
5843 | } |
5844 | ||
cfd3da1e | 5845 | static void __setup_per_zone_wmarks(void) |
1da177e4 LT |
5846 | { |
5847 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); | |
5848 | unsigned long lowmem_pages = 0; | |
5849 | struct zone *zone; | |
5850 | unsigned long flags; | |
5851 | ||
416ef04f | 5852 | /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */ |
1da177e4 | 5853 | for_each_zone(zone) { |
416ef04f | 5854 | if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE) |
9705bea5 | 5855 | lowmem_pages += zone_managed_pages(zone); |
1da177e4 LT |
5856 | } |
5857 | ||
5858 | for_each_zone(zone) { | |
ac924c60 AM |
5859 | u64 tmp; |
5860 | ||
1125b4e3 | 5861 | spin_lock_irqsave(&zone->lock, flags); |
9705bea5 | 5862 | tmp = (u64)pages_min * zone_managed_pages(zone); |
72741db6 | 5863 | tmp = div64_ul(tmp, lowmem_pages); |
416ef04f | 5864 | if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) { |
1da177e4 | 5865 | /* |
669ed175 | 5866 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't |
416ef04f | 5867 | * need highmem and movable zones pages, so cap pages_min |
5868 | * to a small value here. | |
669ed175 | 5869 | * |
41858966 | 5870 | * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) |
8bb4e7a2 | 5871 | * deltas control async page reclaim, and so should |
416ef04f | 5872 | * not be capped for highmem and movable zones. |
1da177e4 | 5873 | */ |
90ae8d67 | 5874 | unsigned long min_pages; |
1da177e4 | 5875 | |
9705bea5 | 5876 | min_pages = zone_managed_pages(zone) / 1024; |
90ae8d67 | 5877 | min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL); |
a9214443 | 5878 | zone->_watermark[WMARK_MIN] = min_pages; |
1da177e4 | 5879 | } else { |
669ed175 NP |
5880 | /* |
5881 | * If it's a lowmem zone, reserve a number of pages | |
1da177e4 LT |
5882 | * proportionate to the zone's size. |
5883 | */ | |
a9214443 | 5884 | zone->_watermark[WMARK_MIN] = tmp; |
1da177e4 LT |
5885 | } |
5886 | ||
795ae7a0 JW |
5887 | /* |
5888 | * Set the kswapd watermarks distance according to the | |
5889 | * scale factor in proportion to available memory, but | |
5890 | * ensure a minimum size on small systems. | |
5891 | */ | |
5892 | tmp = max_t(u64, tmp >> 2, | |
9705bea5 | 5893 | mult_frac(zone_managed_pages(zone), |
795ae7a0 JW |
5894 | watermark_scale_factor, 10000)); |
5895 | ||
aa092591 | 5896 | zone->watermark_boost = 0; |
a9214443 | 5897 | zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp; |
c574bbe9 HY |
5898 | zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp; |
5899 | zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp; | |
49f223a9 | 5900 | |
1125b4e3 | 5901 | spin_unlock_irqrestore(&zone->lock, flags); |
1da177e4 | 5902 | } |
cb45b0e9 HA |
5903 | |
5904 | /* update totalreserve_pages */ | |
5905 | calculate_totalreserve_pages(); | |
1da177e4 LT |
5906 | } |
5907 | ||
cfd3da1e MG |
5908 | /** |
5909 | * setup_per_zone_wmarks - called when min_free_kbytes changes | |
5910 | * or when memory is hot-{added|removed} | |
5911 | * | |
5912 | * Ensures that the watermark[min,low,high] values for each zone are set | |
5913 | * correctly with respect to min_free_kbytes. | |
5914 | */ | |
5915 | void setup_per_zone_wmarks(void) | |
5916 | { | |
b92ca18e | 5917 | struct zone *zone; |
b93e0f32 MH |
5918 | static DEFINE_SPINLOCK(lock); |
5919 | ||
5920 | spin_lock(&lock); | |
cfd3da1e | 5921 | __setup_per_zone_wmarks(); |
b93e0f32 | 5922 | spin_unlock(&lock); |
b92ca18e MG |
5923 | |
5924 | /* | |
5925 | * The watermark size have changed so update the pcpu batch | |
5926 | * and high limits or the limits may be inappropriate. | |
5927 | */ | |
5928 | for_each_zone(zone) | |
04f8cfea | 5929 | zone_pcp_update(zone, 0); |
cfd3da1e MG |
5930 | } |
5931 | ||
1da177e4 LT |
5932 | /* |
5933 | * Initialise min_free_kbytes. | |
5934 | * | |
5935 | * For small machines we want it small (128k min). For large machines | |
8beeae86 | 5936 | * we want it large (256MB max). But it is not linear, because network |
1da177e4 LT |
5937 | * bandwidth does not increase linearly with machine size. We use |
5938 | * | |
b8af2941 | 5939 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: |
1da177e4 LT |
5940 | * min_free_kbytes = sqrt(lowmem_kbytes * 16) |
5941 | * | |
5942 | * which yields | |
5943 | * | |
5944 | * 16MB: 512k | |
5945 | * 32MB: 724k | |
5946 | * 64MB: 1024k | |
5947 | * 128MB: 1448k | |
5948 | * 256MB: 2048k | |
5949 | * 512MB: 2896k | |
5950 | * 1024MB: 4096k | |
5951 | * 2048MB: 5792k | |
5952 | * 4096MB: 8192k | |
5953 | * 8192MB: 11584k | |
5954 | * 16384MB: 16384k | |
5955 | */ | |
bd3400ea | 5956 | void calculate_min_free_kbytes(void) |
1da177e4 LT |
5957 | { |
5958 | unsigned long lowmem_kbytes; | |
5f12733e | 5959 | int new_min_free_kbytes; |
1da177e4 LT |
5960 | |
5961 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); | |
5f12733e MH |
5962 | new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16); |
5963 | ||
59d336bd WS |
5964 | if (new_min_free_kbytes > user_min_free_kbytes) |
5965 | min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144); | |
5966 | else | |
5f12733e MH |
5967 | pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n", |
5968 | new_min_free_kbytes, user_min_free_kbytes); | |
59d336bd | 5969 | |
bd3400ea LF |
5970 | } |
5971 | ||
5972 | int __meminit init_per_zone_wmark_min(void) | |
5973 | { | |
5974 | calculate_min_free_kbytes(); | |
bc75d33f | 5975 | setup_per_zone_wmarks(); |
a6cccdc3 | 5976 | refresh_zone_stat_thresholds(); |
1da177e4 | 5977 | setup_per_zone_lowmem_reserve(); |
6423aa81 JK |
5978 | |
5979 | #ifdef CONFIG_NUMA | |
5980 | setup_min_unmapped_ratio(); | |
5981 | setup_min_slab_ratio(); | |
5982 | #endif | |
5983 | ||
4aab2be0 VB |
5984 | khugepaged_min_free_kbytes_update(); |
5985 | ||
1da177e4 LT |
5986 | return 0; |
5987 | } | |
e08d3fdf | 5988 | postcore_initcall(init_per_zone_wmark_min) |
1da177e4 LT |
5989 | |
5990 | /* | |
b8af2941 | 5991 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so |
1da177e4 LT |
5992 | * that we can call two helper functions whenever min_free_kbytes |
5993 | * changes. | |
5994 | */ | |
e95d372c | 5995 | static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 5996 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 5997 | { |
da8c757b HP |
5998 | int rc; |
5999 | ||
6000 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
6001 | if (rc) | |
6002 | return rc; | |
6003 | ||
5f12733e MH |
6004 | if (write) { |
6005 | user_min_free_kbytes = min_free_kbytes; | |
bc75d33f | 6006 | setup_per_zone_wmarks(); |
5f12733e | 6007 | } |
1da177e4 LT |
6008 | return 0; |
6009 | } | |
6010 | ||
e95d372c | 6011 | static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 6012 | void *buffer, size_t *length, loff_t *ppos) |
795ae7a0 JW |
6013 | { |
6014 | int rc; | |
6015 | ||
6016 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
6017 | if (rc) | |
6018 | return rc; | |
6019 | ||
6020 | if (write) | |
6021 | setup_per_zone_wmarks(); | |
6022 | ||
6023 | return 0; | |
6024 | } | |
6025 | ||
9614634f | 6026 | #ifdef CONFIG_NUMA |
6423aa81 | 6027 | static void setup_min_unmapped_ratio(void) |
9614634f | 6028 | { |
6423aa81 | 6029 | pg_data_t *pgdat; |
9614634f | 6030 | struct zone *zone; |
9614634f | 6031 | |
a5f5f91d | 6032 | for_each_online_pgdat(pgdat) |
81cbcbc2 | 6033 | pgdat->min_unmapped_pages = 0; |
a5f5f91d | 6034 | |
9614634f | 6035 | for_each_zone(zone) |
9705bea5 AK |
6036 | zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) * |
6037 | sysctl_min_unmapped_ratio) / 100; | |
9614634f | 6038 | } |
0ff38490 | 6039 | |
6423aa81 | 6040 | |
e95d372c | 6041 | static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 6042 | void *buffer, size_t *length, loff_t *ppos) |
0ff38490 | 6043 | { |
0ff38490 CL |
6044 | int rc; |
6045 | ||
8d65af78 | 6046 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); |
0ff38490 CL |
6047 | if (rc) |
6048 | return rc; | |
6049 | ||
6423aa81 JK |
6050 | setup_min_unmapped_ratio(); |
6051 | ||
6052 | return 0; | |
6053 | } | |
6054 | ||
6055 | static void setup_min_slab_ratio(void) | |
6056 | { | |
6057 | pg_data_t *pgdat; | |
6058 | struct zone *zone; | |
6059 | ||
a5f5f91d MG |
6060 | for_each_online_pgdat(pgdat) |
6061 | pgdat->min_slab_pages = 0; | |
6062 | ||
0ff38490 | 6063 | for_each_zone(zone) |
9705bea5 AK |
6064 | zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) * |
6065 | sysctl_min_slab_ratio) / 100; | |
6423aa81 JK |
6066 | } |
6067 | ||
e95d372c | 6068 | static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 6069 | void *buffer, size_t *length, loff_t *ppos) |
6423aa81 JK |
6070 | { |
6071 | int rc; | |
6072 | ||
6073 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
6074 | if (rc) | |
6075 | return rc; | |
6076 | ||
6077 | setup_min_slab_ratio(); | |
6078 | ||
0ff38490 CL |
6079 | return 0; |
6080 | } | |
9614634f CL |
6081 | #endif |
6082 | ||
1da177e4 LT |
6083 | /* |
6084 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around | |
6085 | * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() | |
6086 | * whenever sysctl_lowmem_reserve_ratio changes. | |
6087 | * | |
6088 | * The reserve ratio obviously has absolutely no relation with the | |
41858966 | 6089 | * minimum watermarks. The lowmem reserve ratio can only make sense |
1da177e4 LT |
6090 | * if in function of the boot time zone sizes. |
6091 | */ | |
e95d372c KW |
6092 | static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, |
6093 | int write, void *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 6094 | { |
86aaf255 BH |
6095 | int i; |
6096 | ||
8d65af78 | 6097 | proc_dointvec_minmax(table, write, buffer, length, ppos); |
86aaf255 BH |
6098 | |
6099 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
6100 | if (sysctl_lowmem_reserve_ratio[i] < 1) | |
6101 | sysctl_lowmem_reserve_ratio[i] = 0; | |
6102 | } | |
6103 | ||
1da177e4 LT |
6104 | setup_per_zone_lowmem_reserve(); |
6105 | return 0; | |
6106 | } | |
6107 | ||
8ad4b1fb | 6108 | /* |
74f44822 MG |
6109 | * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each |
6110 | * cpu. It is the fraction of total pages in each zone that a hot per cpu | |
b8af2941 | 6111 | * pagelist can have before it gets flushed back to buddy allocator. |
8ad4b1fb | 6112 | */ |
e95d372c | 6113 | static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table, |
74f44822 | 6114 | int write, void *buffer, size_t *length, loff_t *ppos) |
8ad4b1fb RS |
6115 | { |
6116 | struct zone *zone; | |
74f44822 | 6117 | int old_percpu_pagelist_high_fraction; |
8ad4b1fb RS |
6118 | int ret; |
6119 | ||
7cd2b0a3 | 6120 | mutex_lock(&pcp_batch_high_lock); |
74f44822 | 6121 | old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction; |
7cd2b0a3 | 6122 | |
8d65af78 | 6123 | ret = proc_dointvec_minmax(table, write, buffer, length, ppos); |
7cd2b0a3 DR |
6124 | if (!write || ret < 0) |
6125 | goto out; | |
6126 | ||
6127 | /* Sanity checking to avoid pcp imbalance */ | |
74f44822 MG |
6128 | if (percpu_pagelist_high_fraction && |
6129 | percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) { | |
6130 | percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction; | |
7cd2b0a3 DR |
6131 | ret = -EINVAL; |
6132 | goto out; | |
6133 | } | |
6134 | ||
6135 | /* No change? */ | |
74f44822 | 6136 | if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction) |
7cd2b0a3 | 6137 | goto out; |
c8e251fa | 6138 | |
cb1ef534 | 6139 | for_each_populated_zone(zone) |
74f44822 | 6140 | zone_set_pageset_high_and_batch(zone, 0); |
7cd2b0a3 | 6141 | out: |
c8e251fa | 6142 | mutex_unlock(&pcp_batch_high_lock); |
7cd2b0a3 | 6143 | return ret; |
8ad4b1fb RS |
6144 | } |
6145 | ||
e95d372c KW |
6146 | static struct ctl_table page_alloc_sysctl_table[] = { |
6147 | { | |
6148 | .procname = "min_free_kbytes", | |
6149 | .data = &min_free_kbytes, | |
6150 | .maxlen = sizeof(min_free_kbytes), | |
6151 | .mode = 0644, | |
6152 | .proc_handler = min_free_kbytes_sysctl_handler, | |
6153 | .extra1 = SYSCTL_ZERO, | |
6154 | }, | |
6155 | { | |
6156 | .procname = "watermark_boost_factor", | |
6157 | .data = &watermark_boost_factor, | |
6158 | .maxlen = sizeof(watermark_boost_factor), | |
6159 | .mode = 0644, | |
6160 | .proc_handler = proc_dointvec_minmax, | |
6161 | .extra1 = SYSCTL_ZERO, | |
6162 | }, | |
6163 | { | |
6164 | .procname = "watermark_scale_factor", | |
6165 | .data = &watermark_scale_factor, | |
6166 | .maxlen = sizeof(watermark_scale_factor), | |
6167 | .mode = 0644, | |
6168 | .proc_handler = watermark_scale_factor_sysctl_handler, | |
6169 | .extra1 = SYSCTL_ONE, | |
6170 | .extra2 = SYSCTL_THREE_THOUSAND, | |
6171 | }, | |
6172 | { | |
6173 | .procname = "percpu_pagelist_high_fraction", | |
6174 | .data = &percpu_pagelist_high_fraction, | |
6175 | .maxlen = sizeof(percpu_pagelist_high_fraction), | |
6176 | .mode = 0644, | |
6177 | .proc_handler = percpu_pagelist_high_fraction_sysctl_handler, | |
6178 | .extra1 = SYSCTL_ZERO, | |
6179 | }, | |
6180 | { | |
6181 | .procname = "lowmem_reserve_ratio", | |
6182 | .data = &sysctl_lowmem_reserve_ratio, | |
6183 | .maxlen = sizeof(sysctl_lowmem_reserve_ratio), | |
6184 | .mode = 0644, | |
6185 | .proc_handler = lowmem_reserve_ratio_sysctl_handler, | |
6186 | }, | |
6187 | #ifdef CONFIG_NUMA | |
6188 | { | |
6189 | .procname = "numa_zonelist_order", | |
6190 | .data = &numa_zonelist_order, | |
6191 | .maxlen = NUMA_ZONELIST_ORDER_LEN, | |
6192 | .mode = 0644, | |
6193 | .proc_handler = numa_zonelist_order_handler, | |
6194 | }, | |
6195 | { | |
6196 | .procname = "min_unmapped_ratio", | |
6197 | .data = &sysctl_min_unmapped_ratio, | |
6198 | .maxlen = sizeof(sysctl_min_unmapped_ratio), | |
6199 | .mode = 0644, | |
6200 | .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler, | |
6201 | .extra1 = SYSCTL_ZERO, | |
6202 | .extra2 = SYSCTL_ONE_HUNDRED, | |
6203 | }, | |
6204 | { | |
6205 | .procname = "min_slab_ratio", | |
6206 | .data = &sysctl_min_slab_ratio, | |
6207 | .maxlen = sizeof(sysctl_min_slab_ratio), | |
6208 | .mode = 0644, | |
6209 | .proc_handler = sysctl_min_slab_ratio_sysctl_handler, | |
6210 | .extra1 = SYSCTL_ZERO, | |
6211 | .extra2 = SYSCTL_ONE_HUNDRED, | |
6212 | }, | |
6213 | #endif | |
6214 | {} | |
6215 | }; | |
6216 | ||
6217 | void __init page_alloc_sysctl_init(void) | |
6218 | { | |
6219 | register_sysctl_init("vm", page_alloc_sysctl_table); | |
6220 | } | |
6221 | ||
8df995f6 | 6222 | #ifdef CONFIG_CONTIG_ALLOC |
a1394bdd MK |
6223 | /* Usage: See admin-guide/dynamic-debug-howto.rst */ |
6224 | static void alloc_contig_dump_pages(struct list_head *page_list) | |
6225 | { | |
6226 | DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure"); | |
6227 | ||
6228 | if (DYNAMIC_DEBUG_BRANCH(descriptor)) { | |
6229 | struct page *page; | |
6230 | ||
6231 | dump_stack(); | |
6232 | list_for_each_entry(page, page_list, lru) | |
6233 | dump_page(page, "migration failure"); | |
6234 | } | |
6235 | } | |
a1394bdd | 6236 | |
c8b36003 RC |
6237 | /* |
6238 | * [start, end) must belong to a single zone. | |
6239 | * @migratetype: using migratetype to filter the type of migration in | |
6240 | * trace_mm_alloc_contig_migrate_range_info. | |
6241 | */ | |
b2c9e2fb | 6242 | int __alloc_contig_migrate_range(struct compact_control *cc, |
c8b36003 RC |
6243 | unsigned long start, unsigned long end, |
6244 | int migratetype) | |
041d3a8c MN |
6245 | { |
6246 | /* This function is based on compact_zone() from compaction.c. */ | |
730ec8c0 | 6247 | unsigned int nr_reclaimed; |
041d3a8c MN |
6248 | unsigned long pfn = start; |
6249 | unsigned int tries = 0; | |
6250 | int ret = 0; | |
8b94e0b8 JK |
6251 | struct migration_target_control mtc = { |
6252 | .nid = zone_to_nid(cc->zone), | |
6253 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, | |
6254 | }; | |
c8b36003 RC |
6255 | struct page *page; |
6256 | unsigned long total_mapped = 0; | |
6257 | unsigned long total_migrated = 0; | |
6258 | unsigned long total_reclaimed = 0; | |
041d3a8c | 6259 | |
361a2a22 | 6260 | lru_cache_disable(); |
041d3a8c | 6261 | |
bb13ffeb | 6262 | while (pfn < end || !list_empty(&cc->migratepages)) { |
041d3a8c MN |
6263 | if (fatal_signal_pending(current)) { |
6264 | ret = -EINTR; | |
6265 | break; | |
6266 | } | |
6267 | ||
bb13ffeb MG |
6268 | if (list_empty(&cc->migratepages)) { |
6269 | cc->nr_migratepages = 0; | |
c2ad7a1f OS |
6270 | ret = isolate_migratepages_range(cc, pfn, end); |
6271 | if (ret && ret != -EAGAIN) | |
041d3a8c | 6272 | break; |
c2ad7a1f | 6273 | pfn = cc->migrate_pfn; |
041d3a8c MN |
6274 | tries = 0; |
6275 | } else if (++tries == 5) { | |
c8e28b47 | 6276 | ret = -EBUSY; |
041d3a8c MN |
6277 | break; |
6278 | } | |
6279 | ||
beb51eaa MK |
6280 | nr_reclaimed = reclaim_clean_pages_from_list(cc->zone, |
6281 | &cc->migratepages); | |
6282 | cc->nr_migratepages -= nr_reclaimed; | |
02c6de8d | 6283 | |
c8b36003 RC |
6284 | if (trace_mm_alloc_contig_migrate_range_info_enabled()) { |
6285 | total_reclaimed += nr_reclaimed; | |
6286 | list_for_each_entry(page, &cc->migratepages, lru) | |
6287 | total_mapped += page_mapcount(page); | |
6288 | } | |
6289 | ||
8b94e0b8 | 6290 | ret = migrate_pages(&cc->migratepages, alloc_migration_target, |
5ac95884 | 6291 | NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL); |
c8e28b47 | 6292 | |
c8b36003 RC |
6293 | if (trace_mm_alloc_contig_migrate_range_info_enabled() && !ret) |
6294 | total_migrated += cc->nr_migratepages; | |
6295 | ||
c8e28b47 OS |
6296 | /* |
6297 | * On -ENOMEM, migrate_pages() bails out right away. It is pointless | |
6298 | * to retry again over this error, so do the same here. | |
6299 | */ | |
6300 | if (ret == -ENOMEM) | |
6301 | break; | |
041d3a8c | 6302 | } |
d479960e | 6303 | |
361a2a22 | 6304 | lru_cache_enable(); |
2a6f5124 | 6305 | if (ret < 0) { |
3f913fc5 | 6306 | if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY) |
151e084a | 6307 | alloc_contig_dump_pages(&cc->migratepages); |
2a6f5124 | 6308 | putback_movable_pages(&cc->migratepages); |
2a6f5124 | 6309 | } |
c8b36003 RC |
6310 | |
6311 | trace_mm_alloc_contig_migrate_range_info(start, end, migratetype, | |
6312 | total_migrated, | |
6313 | total_reclaimed, | |
6314 | total_mapped); | |
6315 | return (ret < 0) ? ret : 0; | |
041d3a8c MN |
6316 | } |
6317 | ||
6318 | /** | |
6319 | * alloc_contig_range() -- tries to allocate given range of pages | |
6320 | * @start: start PFN to allocate | |
6321 | * @end: one-past-the-last PFN to allocate | |
f0953a1b | 6322 | * @migratetype: migratetype of the underlying pageblocks (either |
0815f3d8 MN |
6323 | * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks |
6324 | * in range must have the same migratetype and it must | |
6325 | * be either of the two. | |
ca96b625 | 6326 | * @gfp_mask: GFP mask to use during compaction |
041d3a8c | 6327 | * |
11ac3e87 ZY |
6328 | * The PFN range does not have to be pageblock aligned. The PFN range must |
6329 | * belong to a single zone. | |
041d3a8c | 6330 | * |
2c7452a0 MK |
6331 | * The first thing this routine does is attempt to MIGRATE_ISOLATE all |
6332 | * pageblocks in the range. Once isolated, the pageblocks should not | |
6333 | * be modified by others. | |
041d3a8c | 6334 | * |
a862f68a | 6335 | * Return: zero on success or negative error code. On success all |
041d3a8c MN |
6336 | * pages which PFN is in [start, end) are allocated for the caller and |
6337 | * need to be freed with free_contig_range(). | |
6338 | */ | |
0815f3d8 | 6339 | int alloc_contig_range(unsigned long start, unsigned long end, |
ca96b625 | 6340 | unsigned migratetype, gfp_t gfp_mask) |
041d3a8c | 6341 | { |
041d3a8c | 6342 | unsigned long outer_start, outer_end; |
b2c9e2fb | 6343 | int order; |
d00181b9 | 6344 | int ret = 0; |
041d3a8c | 6345 | |
bb13ffeb MG |
6346 | struct compact_control cc = { |
6347 | .nr_migratepages = 0, | |
6348 | .order = -1, | |
6349 | .zone = page_zone(pfn_to_page(start)), | |
e0b9daeb | 6350 | .mode = MIGRATE_SYNC, |
bb13ffeb | 6351 | .ignore_skip_hint = true, |
2583d671 | 6352 | .no_set_skip_hint = true, |
7dea19f9 | 6353 | .gfp_mask = current_gfp_context(gfp_mask), |
b06eda09 | 6354 | .alloc_contig = true, |
bb13ffeb MG |
6355 | }; |
6356 | INIT_LIST_HEAD(&cc.migratepages); | |
6357 | ||
041d3a8c MN |
6358 | /* |
6359 | * What we do here is we mark all pageblocks in range as | |
6360 | * MIGRATE_ISOLATE. Because pageblock and max order pages may | |
6361 | * have different sizes, and due to the way page allocator | |
b2c9e2fb | 6362 | * work, start_isolate_page_range() has special handlings for this. |
041d3a8c MN |
6363 | * |
6364 | * Once the pageblocks are marked as MIGRATE_ISOLATE, we | |
6365 | * migrate the pages from an unaligned range (ie. pages that | |
b2c9e2fb | 6366 | * we are interested in). This will put all the pages in |
041d3a8c MN |
6367 | * range back to page allocator as MIGRATE_ISOLATE. |
6368 | * | |
6369 | * When this is done, we take the pages in range from page | |
6370 | * allocator removing them from the buddy system. This way | |
6371 | * page allocator will never consider using them. | |
6372 | * | |
6373 | * This lets us mark the pageblocks back as | |
6374 | * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the | |
6375 | * aligned range but not in the unaligned, original range are | |
6376 | * put back to page allocator so that buddy can use them. | |
6377 | */ | |
6378 | ||
6e263fff | 6379 | ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask); |
3fa0c7c7 | 6380 | if (ret) |
b2c9e2fb | 6381 | goto done; |
041d3a8c | 6382 | |
7612921f VB |
6383 | drain_all_pages(cc.zone); |
6384 | ||
8ef5849f JK |
6385 | /* |
6386 | * In case of -EBUSY, we'd like to know which page causes problem. | |
63cd4489 MK |
6387 | * So, just fall through. test_pages_isolated() has a tracepoint |
6388 | * which will report the busy page. | |
6389 | * | |
6390 | * It is possible that busy pages could become available before | |
6391 | * the call to test_pages_isolated, and the range will actually be | |
6392 | * allocated. So, if we fall through be sure to clear ret so that | |
6393 | * -EBUSY is not accidentally used or returned to caller. | |
8ef5849f | 6394 | */ |
c8b36003 | 6395 | ret = __alloc_contig_migrate_range(&cc, start, end, migratetype); |
8ef5849f | 6396 | if (ret && ret != -EBUSY) |
041d3a8c | 6397 | goto done; |
68d68ff6 | 6398 | ret = 0; |
041d3a8c MN |
6399 | |
6400 | /* | |
b2c9e2fb | 6401 | * Pages from [start, end) are within a pageblock_nr_pages |
041d3a8c MN |
6402 | * aligned blocks that are marked as MIGRATE_ISOLATE. What's |
6403 | * more, all pages in [start, end) are free in page allocator. | |
6404 | * What we are going to do is to allocate all pages from | |
6405 | * [start, end) (that is remove them from page allocator). | |
6406 | * | |
6407 | * The only problem is that pages at the beginning and at the | |
6408 | * end of interesting range may be not aligned with pages that | |
6409 | * page allocator holds, ie. they can be part of higher order | |
6410 | * pages. Because of this, we reserve the bigger range and | |
6411 | * once this is done free the pages we are not interested in. | |
6412 | * | |
6413 | * We don't have to hold zone->lock here because the pages are | |
6414 | * isolated thus they won't get removed from buddy. | |
6415 | */ | |
6416 | ||
041d3a8c MN |
6417 | order = 0; |
6418 | outer_start = start; | |
6419 | while (!PageBuddy(pfn_to_page(outer_start))) { | |
5e0a760b | 6420 | if (++order > MAX_PAGE_ORDER) { |
8ef5849f JK |
6421 | outer_start = start; |
6422 | break; | |
041d3a8c MN |
6423 | } |
6424 | outer_start &= ~0UL << order; | |
6425 | } | |
6426 | ||
8ef5849f | 6427 | if (outer_start != start) { |
ab130f91 | 6428 | order = buddy_order(pfn_to_page(outer_start)); |
8ef5849f JK |
6429 | |
6430 | /* | |
6431 | * outer_start page could be small order buddy page and | |
6432 | * it doesn't include start page. Adjust outer_start | |
6433 | * in this case to report failed page properly | |
6434 | * on tracepoint in test_pages_isolated() | |
6435 | */ | |
6436 | if (outer_start + (1UL << order) <= start) | |
6437 | outer_start = start; | |
6438 | } | |
6439 | ||
041d3a8c | 6440 | /* Make sure the range is really isolated. */ |
756d25be | 6441 | if (test_pages_isolated(outer_start, end, 0)) { |
041d3a8c MN |
6442 | ret = -EBUSY; |
6443 | goto done; | |
6444 | } | |
6445 | ||
49f223a9 | 6446 | /* Grab isolated pages from freelists. */ |
bb13ffeb | 6447 | outer_end = isolate_freepages_range(&cc, outer_start, end); |
041d3a8c MN |
6448 | if (!outer_end) { |
6449 | ret = -EBUSY; | |
6450 | goto done; | |
6451 | } | |
6452 | ||
6453 | /* Free head and tail (if any) */ | |
6454 | if (start != outer_start) | |
6455 | free_contig_range(outer_start, start - outer_start); | |
6456 | if (end != outer_end) | |
6457 | free_contig_range(end, outer_end - end); | |
6458 | ||
6459 | done: | |
6e263fff | 6460 | undo_isolate_page_range(start, end, migratetype); |
041d3a8c MN |
6461 | return ret; |
6462 | } | |
255f5985 | 6463 | EXPORT_SYMBOL(alloc_contig_range); |
5e27a2df AK |
6464 | |
6465 | static int __alloc_contig_pages(unsigned long start_pfn, | |
6466 | unsigned long nr_pages, gfp_t gfp_mask) | |
6467 | { | |
6468 | unsigned long end_pfn = start_pfn + nr_pages; | |
6469 | ||
6470 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, | |
6471 | gfp_mask); | |
6472 | } | |
6473 | ||
6474 | static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn, | |
6475 | unsigned long nr_pages) | |
6476 | { | |
6477 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
6478 | struct page *page; | |
6479 | ||
6480 | for (i = start_pfn; i < end_pfn; i++) { | |
6481 | page = pfn_to_online_page(i); | |
6482 | if (!page) | |
6483 | return false; | |
6484 | ||
6485 | if (page_zone(page) != z) | |
6486 | return false; | |
6487 | ||
6488 | if (PageReserved(page)) | |
4d73ba5f MG |
6489 | return false; |
6490 | ||
6491 | if (PageHuge(page)) | |
5e27a2df | 6492 | return false; |
5e27a2df AK |
6493 | } |
6494 | return true; | |
6495 | } | |
6496 | ||
6497 | static bool zone_spans_last_pfn(const struct zone *zone, | |
6498 | unsigned long start_pfn, unsigned long nr_pages) | |
6499 | { | |
6500 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
6501 | ||
6502 | return zone_spans_pfn(zone, last_pfn); | |
6503 | } | |
6504 | ||
6505 | /** | |
6506 | * alloc_contig_pages() -- tries to find and allocate contiguous range of pages | |
6507 | * @nr_pages: Number of contiguous pages to allocate | |
6508 | * @gfp_mask: GFP mask to limit search and used during compaction | |
6509 | * @nid: Target node | |
6510 | * @nodemask: Mask for other possible nodes | |
6511 | * | |
6512 | * This routine is a wrapper around alloc_contig_range(). It scans over zones | |
6513 | * on an applicable zonelist to find a contiguous pfn range which can then be | |
6514 | * tried for allocation with alloc_contig_range(). This routine is intended | |
6515 | * for allocation requests which can not be fulfilled with the buddy allocator. | |
6516 | * | |
6517 | * The allocated memory is always aligned to a page boundary. If nr_pages is a | |
eaab8e75 AK |
6518 | * power of two, then allocated range is also guaranteed to be aligned to same |
6519 | * nr_pages (e.g. 1GB request would be aligned to 1GB). | |
5e27a2df AK |
6520 | * |
6521 | * Allocated pages can be freed with free_contig_range() or by manually calling | |
6522 | * __free_page() on each allocated page. | |
6523 | * | |
6524 | * Return: pointer to contiguous pages on success, or NULL if not successful. | |
6525 | */ | |
6526 | struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask, | |
6527 | int nid, nodemask_t *nodemask) | |
6528 | { | |
6529 | unsigned long ret, pfn, flags; | |
6530 | struct zonelist *zonelist; | |
6531 | struct zone *zone; | |
6532 | struct zoneref *z; | |
6533 | ||
6534 | zonelist = node_zonelist(nid, gfp_mask); | |
6535 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
6536 | gfp_zone(gfp_mask), nodemask) { | |
6537 | spin_lock_irqsave(&zone->lock, flags); | |
6538 | ||
6539 | pfn = ALIGN(zone->zone_start_pfn, nr_pages); | |
6540 | while (zone_spans_last_pfn(zone, pfn, nr_pages)) { | |
6541 | if (pfn_range_valid_contig(zone, pfn, nr_pages)) { | |
6542 | /* | |
6543 | * We release the zone lock here because | |
6544 | * alloc_contig_range() will also lock the zone | |
6545 | * at some point. If there's an allocation | |
6546 | * spinning on this lock, it may win the race | |
6547 | * and cause alloc_contig_range() to fail... | |
6548 | */ | |
6549 | spin_unlock_irqrestore(&zone->lock, flags); | |
6550 | ret = __alloc_contig_pages(pfn, nr_pages, | |
6551 | gfp_mask); | |
6552 | if (!ret) | |
6553 | return pfn_to_page(pfn); | |
6554 | spin_lock_irqsave(&zone->lock, flags); | |
6555 | } | |
6556 | pfn += nr_pages; | |
6557 | } | |
6558 | spin_unlock_irqrestore(&zone->lock, flags); | |
6559 | } | |
6560 | return NULL; | |
6561 | } | |
4eb0716e | 6562 | #endif /* CONFIG_CONTIG_ALLOC */ |
041d3a8c | 6563 | |
78fa5150 | 6564 | void free_contig_range(unsigned long pfn, unsigned long nr_pages) |
041d3a8c | 6565 | { |
78fa5150 | 6566 | unsigned long count = 0; |
bcc2b02f MS |
6567 | |
6568 | for (; nr_pages--; pfn++) { | |
6569 | struct page *page = pfn_to_page(pfn); | |
6570 | ||
6571 | count += page_count(page) != 1; | |
6572 | __free_page(page); | |
6573 | } | |
78fa5150 | 6574 | WARN(count != 0, "%lu pages are still in use!\n", count); |
041d3a8c | 6575 | } |
255f5985 | 6576 | EXPORT_SYMBOL(free_contig_range); |
041d3a8c | 6577 | |
ec6e8c7e VB |
6578 | /* |
6579 | * Effectively disable pcplists for the zone by setting the high limit to 0 | |
6580 | * and draining all cpus. A concurrent page freeing on another CPU that's about | |
6581 | * to put the page on pcplist will either finish before the drain and the page | |
6582 | * will be drained, or observe the new high limit and skip the pcplist. | |
6583 | * | |
6584 | * Must be paired with a call to zone_pcp_enable(). | |
6585 | */ | |
6586 | void zone_pcp_disable(struct zone *zone) | |
6587 | { | |
6588 | mutex_lock(&pcp_batch_high_lock); | |
90b41691 | 6589 | __zone_set_pageset_high_and_batch(zone, 0, 0, 1); |
ec6e8c7e VB |
6590 | __drain_all_pages(zone, true); |
6591 | } | |
6592 | ||
6593 | void zone_pcp_enable(struct zone *zone) | |
6594 | { | |
90b41691 HY |
6595 | __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min, |
6596 | zone->pageset_high_max, zone->pageset_batch); | |
ec6e8c7e VB |
6597 | mutex_unlock(&pcp_batch_high_lock); |
6598 | } | |
6599 | ||
340175b7 JL |
6600 | void zone_pcp_reset(struct zone *zone) |
6601 | { | |
5a883813 | 6602 | int cpu; |
28f836b6 | 6603 | struct per_cpu_zonestat *pzstats; |
340175b7 | 6604 | |
28f836b6 | 6605 | if (zone->per_cpu_pageset != &boot_pageset) { |
5a883813 | 6606 | for_each_online_cpu(cpu) { |
28f836b6 MG |
6607 | pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu); |
6608 | drain_zonestat(zone, pzstats); | |
5a883813 | 6609 | } |
28f836b6 | 6610 | free_percpu(zone->per_cpu_pageset); |
28f836b6 | 6611 | zone->per_cpu_pageset = &boot_pageset; |
022e7fa0 ML |
6612 | if (zone->per_cpu_zonestats != &boot_zonestats) { |
6613 | free_percpu(zone->per_cpu_zonestats); | |
6614 | zone->per_cpu_zonestats = &boot_zonestats; | |
6615 | } | |
340175b7 | 6616 | } |
340175b7 JL |
6617 | } |
6618 | ||
6dcd73d7 | 6619 | #ifdef CONFIG_MEMORY_HOTREMOVE |
0c0e6195 | 6620 | /* |
257bea71 DH |
6621 | * All pages in the range must be in a single zone, must not contain holes, |
6622 | * must span full sections, and must be isolated before calling this function. | |
0c0e6195 | 6623 | */ |
257bea71 | 6624 | void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) |
0c0e6195 | 6625 | { |
257bea71 | 6626 | unsigned long pfn = start_pfn; |
0c0e6195 KH |
6627 | struct page *page; |
6628 | struct zone *zone; | |
0ee5f4f3 | 6629 | unsigned int order; |
0c0e6195 | 6630 | unsigned long flags; |
5557c766 | 6631 | |
2d070eab | 6632 | offline_mem_sections(pfn, end_pfn); |
0c0e6195 KH |
6633 | zone = page_zone(pfn_to_page(pfn)); |
6634 | spin_lock_irqsave(&zone->lock, flags); | |
0c0e6195 | 6635 | while (pfn < end_pfn) { |
0c0e6195 | 6636 | page = pfn_to_page(pfn); |
b023f468 WC |
6637 | /* |
6638 | * The HWPoisoned page may be not in buddy system, and | |
6639 | * page_count() is not 0. | |
6640 | */ | |
6641 | if (unlikely(!PageBuddy(page) && PageHWPoison(page))) { | |
6642 | pfn++; | |
b023f468 WC |
6643 | continue; |
6644 | } | |
aa218795 DH |
6645 | /* |
6646 | * At this point all remaining PageOffline() pages have a | |
6647 | * reference count of 0 and can simply be skipped. | |
6648 | */ | |
6649 | if (PageOffline(page)) { | |
6650 | BUG_ON(page_count(page)); | |
6651 | BUG_ON(PageBuddy(page)); | |
6652 | pfn++; | |
aa218795 DH |
6653 | continue; |
6654 | } | |
b023f468 | 6655 | |
0c0e6195 KH |
6656 | BUG_ON(page_count(page)); |
6657 | BUG_ON(!PageBuddy(page)); | |
ab130f91 | 6658 | order = buddy_order(page); |
6ab01363 | 6659 | del_page_from_free_list(page, zone, order); |
0c0e6195 KH |
6660 | pfn += (1 << order); |
6661 | } | |
6662 | spin_unlock_irqrestore(&zone->lock, flags); | |
6663 | } | |
6664 | #endif | |
8d22ba1b | 6665 | |
8446b59b ED |
6666 | /* |
6667 | * This function returns a stable result only if called under zone lock. | |
6668 | */ | |
8d22ba1b WF |
6669 | bool is_free_buddy_page(struct page *page) |
6670 | { | |
8d22ba1b | 6671 | unsigned long pfn = page_to_pfn(page); |
7aeb09f9 | 6672 | unsigned int order; |
8d22ba1b | 6673 | |
fd377218 | 6674 | for (order = 0; order < NR_PAGE_ORDERS; order++) { |
8d22ba1b WF |
6675 | struct page *page_head = page - (pfn & ((1 << order) - 1)); |
6676 | ||
8446b59b ED |
6677 | if (PageBuddy(page_head) && |
6678 | buddy_order_unsafe(page_head) >= order) | |
8d22ba1b WF |
6679 | break; |
6680 | } | |
8d22ba1b | 6681 | |
5e0a760b | 6682 | return order <= MAX_PAGE_ORDER; |
8d22ba1b | 6683 | } |
a581865e | 6684 | EXPORT_SYMBOL(is_free_buddy_page); |
d4ae9916 NH |
6685 | |
6686 | #ifdef CONFIG_MEMORY_FAILURE | |
6687 | /* | |
06be6ff3 OS |
6688 | * Break down a higher-order page in sub-pages, and keep our target out of |
6689 | * buddy allocator. | |
d4ae9916 | 6690 | */ |
06be6ff3 OS |
6691 | static void break_down_buddy_pages(struct zone *zone, struct page *page, |
6692 | struct page *target, int low, int high, | |
6693 | int migratetype) | |
6694 | { | |
6695 | unsigned long size = 1 << high; | |
0dfca313 | 6696 | struct page *current_buddy; |
06be6ff3 OS |
6697 | |
6698 | while (high > low) { | |
6699 | high--; | |
6700 | size >>= 1; | |
6701 | ||
6702 | if (target >= &page[size]) { | |
06be6ff3 | 6703 | current_buddy = page; |
0dfca313 | 6704 | page = page + size; |
06be6ff3 | 6705 | } else { |
06be6ff3 OS |
6706 | current_buddy = page + size; |
6707 | } | |
6708 | ||
6709 | if (set_page_guard(zone, current_buddy, high, migratetype)) | |
6710 | continue; | |
6711 | ||
27e0db3c KS |
6712 | add_to_free_list(current_buddy, zone, high, migratetype); |
6713 | set_buddy_order(current_buddy, high); | |
06be6ff3 OS |
6714 | } |
6715 | } | |
6716 | ||
6717 | /* | |
6718 | * Take a page that will be marked as poisoned off the buddy allocator. | |
6719 | */ | |
6720 | bool take_page_off_buddy(struct page *page) | |
d4ae9916 NH |
6721 | { |
6722 | struct zone *zone = page_zone(page); | |
6723 | unsigned long pfn = page_to_pfn(page); | |
6724 | unsigned long flags; | |
6725 | unsigned int order; | |
06be6ff3 | 6726 | bool ret = false; |
d4ae9916 NH |
6727 | |
6728 | spin_lock_irqsave(&zone->lock, flags); | |
fd377218 | 6729 | for (order = 0; order < NR_PAGE_ORDERS; order++) { |
d4ae9916 | 6730 | struct page *page_head = page - (pfn & ((1 << order) - 1)); |
ab130f91 | 6731 | int page_order = buddy_order(page_head); |
d4ae9916 | 6732 | |
ab130f91 | 6733 | if (PageBuddy(page_head) && page_order >= order) { |
06be6ff3 OS |
6734 | unsigned long pfn_head = page_to_pfn(page_head); |
6735 | int migratetype = get_pfnblock_migratetype(page_head, | |
6736 | pfn_head); | |
6737 | ||
ab130f91 | 6738 | del_page_from_free_list(page_head, zone, page_order); |
06be6ff3 | 6739 | break_down_buddy_pages(zone, page_head, page, 0, |
ab130f91 | 6740 | page_order, migratetype); |
bf181c58 | 6741 | SetPageHWPoisonTakenOff(page); |
bac9c6fa DH |
6742 | if (!is_migrate_isolate(migratetype)) |
6743 | __mod_zone_freepage_state(zone, -1, migratetype); | |
06be6ff3 | 6744 | ret = true; |
d4ae9916 NH |
6745 | break; |
6746 | } | |
06be6ff3 OS |
6747 | if (page_count(page_head) > 0) |
6748 | break; | |
d4ae9916 NH |
6749 | } |
6750 | spin_unlock_irqrestore(&zone->lock, flags); | |
06be6ff3 | 6751 | return ret; |
d4ae9916 | 6752 | } |
bf181c58 NH |
6753 | |
6754 | /* | |
6755 | * Cancel takeoff done by take_page_off_buddy(). | |
6756 | */ | |
6757 | bool put_page_back_buddy(struct page *page) | |
6758 | { | |
6759 | struct zone *zone = page_zone(page); | |
6760 | unsigned long pfn = page_to_pfn(page); | |
6761 | unsigned long flags; | |
6762 | int migratetype = get_pfnblock_migratetype(page, pfn); | |
6763 | bool ret = false; | |
6764 | ||
6765 | spin_lock_irqsave(&zone->lock, flags); | |
6766 | if (put_page_testzero(page)) { | |
6767 | ClearPageHWPoisonTakenOff(page); | |
6768 | __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE); | |
6769 | if (TestClearPageHWPoison(page)) { | |
bf181c58 NH |
6770 | ret = true; |
6771 | } | |
6772 | } | |
6773 | spin_unlock_irqrestore(&zone->lock, flags); | |
6774 | ||
6775 | return ret; | |
6776 | } | |
d4ae9916 | 6777 | #endif |
62b31070 BH |
6778 | |
6779 | #ifdef CONFIG_ZONE_DMA | |
6780 | bool has_managed_dma(void) | |
6781 | { | |
6782 | struct pglist_data *pgdat; | |
6783 | ||
6784 | for_each_online_pgdat(pgdat) { | |
6785 | struct zone *zone = &pgdat->node_zones[ZONE_DMA]; | |
6786 | ||
6787 | if (managed_zone(zone)) | |
6788 | return true; | |
6789 | } | |
6790 | return false; | |
6791 | } | |
6792 | #endif /* CONFIG_ZONE_DMA */ | |
dcdfdd40 KS |
6793 | |
6794 | #ifdef CONFIG_UNACCEPTED_MEMORY | |
6795 | ||
6796 | /* Counts number of zones with unaccepted pages. */ | |
6797 | static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages); | |
6798 | ||
6799 | static bool lazy_accept = true; | |
6800 | ||
6801 | static int __init accept_memory_parse(char *p) | |
6802 | { | |
6803 | if (!strcmp(p, "lazy")) { | |
6804 | lazy_accept = true; | |
6805 | return 0; | |
6806 | } else if (!strcmp(p, "eager")) { | |
6807 | lazy_accept = false; | |
6808 | return 0; | |
6809 | } else { | |
6810 | return -EINVAL; | |
6811 | } | |
6812 | } | |
6813 | early_param("accept_memory", accept_memory_parse); | |
6814 | ||
6815 | static bool page_contains_unaccepted(struct page *page, unsigned int order) | |
6816 | { | |
6817 | phys_addr_t start = page_to_phys(page); | |
6818 | phys_addr_t end = start + (PAGE_SIZE << order); | |
6819 | ||
6820 | return range_contains_unaccepted_memory(start, end); | |
6821 | } | |
6822 | ||
6823 | static void accept_page(struct page *page, unsigned int order) | |
6824 | { | |
6825 | phys_addr_t start = page_to_phys(page); | |
6826 | ||
6827 | accept_memory(start, start + (PAGE_SIZE << order)); | |
6828 | } | |
6829 | ||
6830 | static bool try_to_accept_memory_one(struct zone *zone) | |
6831 | { | |
6832 | unsigned long flags; | |
6833 | struct page *page; | |
6834 | bool last; | |
6835 | ||
6836 | if (list_empty(&zone->unaccepted_pages)) | |
6837 | return false; | |
6838 | ||
6839 | spin_lock_irqsave(&zone->lock, flags); | |
6840 | page = list_first_entry_or_null(&zone->unaccepted_pages, | |
6841 | struct page, lru); | |
6842 | if (!page) { | |
6843 | spin_unlock_irqrestore(&zone->lock, flags); | |
6844 | return false; | |
6845 | } | |
6846 | ||
6847 | list_del(&page->lru); | |
6848 | last = list_empty(&zone->unaccepted_pages); | |
6849 | ||
6850 | __mod_zone_freepage_state(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE); | |
6851 | __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES); | |
6852 | spin_unlock_irqrestore(&zone->lock, flags); | |
6853 | ||
5e0a760b | 6854 | accept_page(page, MAX_PAGE_ORDER); |
dcdfdd40 | 6855 | |
5e0a760b | 6856 | __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL); |
dcdfdd40 KS |
6857 | |
6858 | if (last) | |
6859 | static_branch_dec(&zones_with_unaccepted_pages); | |
6860 | ||
6861 | return true; | |
6862 | } | |
6863 | ||
6864 | static bool try_to_accept_memory(struct zone *zone, unsigned int order) | |
6865 | { | |
6866 | long to_accept; | |
6867 | int ret = false; | |
6868 | ||
6869 | /* How much to accept to get to high watermark? */ | |
6870 | to_accept = high_wmark_pages(zone) - | |
6871 | (zone_page_state(zone, NR_FREE_PAGES) - | |
6872 | __zone_watermark_unusable_free(zone, order, 0)); | |
6873 | ||
6874 | /* Accept at least one page */ | |
6875 | do { | |
6876 | if (!try_to_accept_memory_one(zone)) | |
6877 | break; | |
6878 | ret = true; | |
6879 | to_accept -= MAX_ORDER_NR_PAGES; | |
6880 | } while (to_accept > 0); | |
6881 | ||
6882 | return ret; | |
6883 | } | |
6884 | ||
6885 | static inline bool has_unaccepted_memory(void) | |
6886 | { | |
6887 | return static_branch_unlikely(&zones_with_unaccepted_pages); | |
6888 | } | |
6889 | ||
6890 | static bool __free_unaccepted(struct page *page) | |
6891 | { | |
6892 | struct zone *zone = page_zone(page); | |
6893 | unsigned long flags; | |
6894 | bool first = false; | |
6895 | ||
6896 | if (!lazy_accept) | |
6897 | return false; | |
6898 | ||
6899 | spin_lock_irqsave(&zone->lock, flags); | |
6900 | first = list_empty(&zone->unaccepted_pages); | |
6901 | list_add_tail(&page->lru, &zone->unaccepted_pages); | |
6902 | __mod_zone_freepage_state(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE); | |
6903 | __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES); | |
6904 | spin_unlock_irqrestore(&zone->lock, flags); | |
6905 | ||
6906 | if (first) | |
6907 | static_branch_inc(&zones_with_unaccepted_pages); | |
6908 | ||
6909 | return true; | |
6910 | } | |
6911 | ||
6912 | #else | |
6913 | ||
6914 | static bool page_contains_unaccepted(struct page *page, unsigned int order) | |
6915 | { | |
6916 | return false; | |
6917 | } | |
6918 | ||
6919 | static void accept_page(struct page *page, unsigned int order) | |
6920 | { | |
6921 | } | |
6922 | ||
6923 | static bool try_to_accept_memory(struct zone *zone, unsigned int order) | |
6924 | { | |
6925 | return false; | |
6926 | } | |
6927 | ||
6928 | static inline bool has_unaccepted_memory(void) | |
6929 | { | |
6930 | return false; | |
6931 | } | |
6932 | ||
6933 | static bool __free_unaccepted(struct page *page) | |
6934 | { | |
6935 | BUILD_BUG(); | |
6936 | return false; | |
6937 | } | |
6938 | ||
6939 | #endif /* CONFIG_UNACCEPTED_MEMORY */ |