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> | |
bdc8cb98 | 35 | #include <linux/memory_hotplug.h> |
1da177e4 | 36 | #include <linux/nodemask.h> |
a6cccdc3 | 37 | #include <linux/vmstat.h> |
933e312e | 38 | #include <linux/fault-inject.h> |
56de7263 | 39 | #include <linux/compaction.h> |
0d3d062a | 40 | #include <trace/events/kmem.h> |
d379f01d | 41 | #include <trace/events/oom.h> |
268bb0ce | 42 | #include <linux/prefetch.h> |
6e543d57 | 43 | #include <linux/mm_inline.h> |
f920e413 | 44 | #include <linux/mmu_notifier.h> |
041d3a8c | 45 | #include <linux/migrate.h> |
5b3cc15a | 46 | #include <linux/sched/mm.h> |
48c96a36 | 47 | #include <linux/page_owner.h> |
df4e817b | 48 | #include <linux/page_table_check.h> |
4949148a | 49 | #include <linux/memcontrol.h> |
42c269c8 | 50 | #include <linux/ftrace.h> |
d92a8cfc | 51 | #include <linux/lockdep.h> |
eb414681 | 52 | #include <linux/psi.h> |
4aab2be0 | 53 | #include <linux/khugepaged.h> |
5bf18281 | 54 | #include <linux/delayacct.h> |
362d37a1 | 55 | #include <linux/cacheinfo.h> |
ac924c60 | 56 | #include <asm/div64.h> |
1da177e4 | 57 | #include "internal.h" |
e900a918 | 58 | #include "shuffle.h" |
36e66c55 | 59 | #include "page_reporting.h" |
1da177e4 | 60 | |
f04a5d5d DH |
61 | /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */ |
62 | typedef int __bitwise fpi_t; | |
63 | ||
64 | /* No special request */ | |
65 | #define FPI_NONE ((__force fpi_t)0) | |
66 | ||
67 | /* | |
68 | * Skip free page reporting notification for the (possibly merged) page. | |
69 | * This does not hinder free page reporting from grabbing the page, | |
70 | * reporting it and marking it "reported" - it only skips notifying | |
71 | * the free page reporting infrastructure about a newly freed page. For | |
72 | * example, used when temporarily pulling a page from a freelist and | |
73 | * putting it back unmodified. | |
74 | */ | |
75 | #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0)) | |
76 | ||
47b6a24a DH |
77 | /* |
78 | * Place the (possibly merged) page to the tail of the freelist. Will ignore | |
79 | * page shuffling (relevant code - e.g., memory onlining - is expected to | |
80 | * shuffle the whole zone). | |
81 | * | |
82 | * Note: No code should rely on this flag for correctness - it's purely | |
83 | * to allow for optimizations when handing back either fresh pages | |
84 | * (memory onlining) or untouched pages (page isolation, free page | |
85 | * reporting). | |
86 | */ | |
87 | #define FPI_TO_TAIL ((__force fpi_t)BIT(1)) | |
88 | ||
c8e251fa CS |
89 | /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */ |
90 | static DEFINE_MUTEX(pcp_batch_high_lock); | |
74f44822 | 91 | #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8) |
c8e251fa | 92 | |
4b23a68f MG |
93 | #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT) |
94 | /* | |
95 | * On SMP, spin_trylock is sufficient protection. | |
96 | * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP. | |
97 | */ | |
98 | #define pcp_trylock_prepare(flags) do { } while (0) | |
99 | #define pcp_trylock_finish(flag) do { } while (0) | |
100 | #else | |
101 | ||
102 | /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */ | |
103 | #define pcp_trylock_prepare(flags) local_irq_save(flags) | |
104 | #define pcp_trylock_finish(flags) local_irq_restore(flags) | |
105 | #endif | |
106 | ||
01b44456 MG |
107 | /* |
108 | * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid | |
109 | * a migration causing the wrong PCP to be locked and remote memory being | |
110 | * potentially allocated, pin the task to the CPU for the lookup+lock. | |
111 | * preempt_disable is used on !RT because it is faster than migrate_disable. | |
112 | * migrate_disable is used on RT because otherwise RT spinlock usage is | |
113 | * interfered with and a high priority task cannot preempt the allocator. | |
114 | */ | |
115 | #ifndef CONFIG_PREEMPT_RT | |
116 | #define pcpu_task_pin() preempt_disable() | |
117 | #define pcpu_task_unpin() preempt_enable() | |
118 | #else | |
119 | #define pcpu_task_pin() migrate_disable() | |
120 | #define pcpu_task_unpin() migrate_enable() | |
121 | #endif | |
c8e251fa | 122 | |
01b44456 MG |
123 | /* |
124 | * Generic helper to lookup and a per-cpu variable with an embedded spinlock. | |
125 | * Return value should be used with equivalent unlock helper. | |
126 | */ | |
127 | #define pcpu_spin_lock(type, member, ptr) \ | |
128 | ({ \ | |
129 | type *_ret; \ | |
130 | pcpu_task_pin(); \ | |
131 | _ret = this_cpu_ptr(ptr); \ | |
132 | spin_lock(&_ret->member); \ | |
133 | _ret; \ | |
134 | }) | |
135 | ||
57490774 | 136 | #define pcpu_spin_trylock(type, member, ptr) \ |
01b44456 MG |
137 | ({ \ |
138 | type *_ret; \ | |
139 | pcpu_task_pin(); \ | |
140 | _ret = this_cpu_ptr(ptr); \ | |
57490774 | 141 | if (!spin_trylock(&_ret->member)) { \ |
01b44456 MG |
142 | pcpu_task_unpin(); \ |
143 | _ret = NULL; \ | |
144 | } \ | |
145 | _ret; \ | |
146 | }) | |
147 | ||
148 | #define pcpu_spin_unlock(member, ptr) \ | |
149 | ({ \ | |
150 | spin_unlock(&ptr->member); \ | |
151 | pcpu_task_unpin(); \ | |
152 | }) | |
153 | ||
01b44456 MG |
154 | /* struct per_cpu_pages specific helpers. */ |
155 | #define pcp_spin_lock(ptr) \ | |
156 | pcpu_spin_lock(struct per_cpu_pages, lock, ptr) | |
157 | ||
57490774 MG |
158 | #define pcp_spin_trylock(ptr) \ |
159 | pcpu_spin_trylock(struct per_cpu_pages, lock, ptr) | |
01b44456 MG |
160 | |
161 | #define pcp_spin_unlock(ptr) \ | |
162 | pcpu_spin_unlock(lock, ptr) | |
163 | ||
72812019 LS |
164 | #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID |
165 | DEFINE_PER_CPU(int, numa_node); | |
166 | EXPORT_PER_CPU_SYMBOL(numa_node); | |
167 | #endif | |
168 | ||
4518085e KW |
169 | DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key); |
170 | ||
7aac7898 LS |
171 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES |
172 | /* | |
173 | * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. | |
174 | * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. | |
175 | * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() | |
176 | * defined in <linux/topology.h>. | |
177 | */ | |
178 | DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ | |
179 | EXPORT_PER_CPU_SYMBOL(_numa_mem_); | |
180 | #endif | |
181 | ||
8b885f53 | 182 | static DEFINE_MUTEX(pcpu_drain_mutex); |
bd233f53 | 183 | |
38addce8 | 184 | #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY |
58bea414 | 185 | volatile unsigned long latent_entropy __latent_entropy; |
38addce8 ER |
186 | EXPORT_SYMBOL(latent_entropy); |
187 | #endif | |
188 | ||
1da177e4 | 189 | /* |
13808910 | 190 | * Array of node states. |
1da177e4 | 191 | */ |
13808910 CL |
192 | nodemask_t node_states[NR_NODE_STATES] __read_mostly = { |
193 | [N_POSSIBLE] = NODE_MASK_ALL, | |
194 | [N_ONLINE] = { { [0] = 1UL } }, | |
195 | #ifndef CONFIG_NUMA | |
196 | [N_NORMAL_MEMORY] = { { [0] = 1UL } }, | |
197 | #ifdef CONFIG_HIGHMEM | |
198 | [N_HIGH_MEMORY] = { { [0] = 1UL } }, | |
20b2f52b | 199 | #endif |
20b2f52b | 200 | [N_MEMORY] = { { [0] = 1UL } }, |
13808910 CL |
201 | [N_CPU] = { { [0] = 1UL } }, |
202 | #endif /* NUMA */ | |
203 | }; | |
204 | EXPORT_SYMBOL(node_states); | |
205 | ||
dcce284a | 206 | gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; |
6471384a | 207 | |
bb14c2c7 VB |
208 | /* |
209 | * A cached value of the page's pageblock's migratetype, used when the page is | |
210 | * put on a pcplist. Used to avoid the pageblock migratetype lookup when | |
211 | * freeing from pcplists in most cases, at the cost of possibly becoming stale. | |
212 | * Also the migratetype set in the page does not necessarily match the pcplist | |
213 | * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any | |
214 | * other index - this ensures that it will be put on the correct CMA freelist. | |
215 | */ | |
216 | static inline int get_pcppage_migratetype(struct page *page) | |
217 | { | |
218 | return page->index; | |
219 | } | |
220 | ||
221 | static inline void set_pcppage_migratetype(struct page *page, int migratetype) | |
222 | { | |
223 | page->index = migratetype; | |
224 | } | |
225 | ||
d9c23400 | 226 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE |
d00181b9 | 227 | unsigned int pageblock_order __read_mostly; |
d9c23400 MG |
228 | #endif |
229 | ||
7fef431b DH |
230 | static void __free_pages_ok(struct page *page, unsigned int order, |
231 | fpi_t fpi_flags); | |
a226f6c8 | 232 | |
1da177e4 LT |
233 | /* |
234 | * results with 256, 32 in the lowmem_reserve sysctl: | |
235 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) | |
236 | * 1G machine -> (16M dma, 784M normal, 224M high) | |
237 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA | |
238 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL | |
84109e15 | 239 | * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA |
a2f1b424 AK |
240 | * |
241 | * TBD: should special case ZONE_DMA32 machines here - in those we normally | |
242 | * don't need any ZONE_NORMAL reservation | |
1da177e4 | 243 | */ |
62069aac | 244 | static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = { |
4b51d669 | 245 | #ifdef CONFIG_ZONE_DMA |
d3cda233 | 246 | [ZONE_DMA] = 256, |
4b51d669 | 247 | #endif |
fb0e7942 | 248 | #ifdef CONFIG_ZONE_DMA32 |
d3cda233 | 249 | [ZONE_DMA32] = 256, |
fb0e7942 | 250 | #endif |
d3cda233 | 251 | [ZONE_NORMAL] = 32, |
e53ef38d | 252 | #ifdef CONFIG_HIGHMEM |
d3cda233 | 253 | [ZONE_HIGHMEM] = 0, |
e53ef38d | 254 | #endif |
d3cda233 | 255 | [ZONE_MOVABLE] = 0, |
2f1b6248 | 256 | }; |
1da177e4 | 257 | |
9420f89d | 258 | char * const zone_names[MAX_NR_ZONES] = { |
4b51d669 | 259 | #ifdef CONFIG_ZONE_DMA |
2f1b6248 | 260 | "DMA", |
4b51d669 | 261 | #endif |
fb0e7942 | 262 | #ifdef CONFIG_ZONE_DMA32 |
2f1b6248 | 263 | "DMA32", |
fb0e7942 | 264 | #endif |
2f1b6248 | 265 | "Normal", |
e53ef38d | 266 | #ifdef CONFIG_HIGHMEM |
2a1e274a | 267 | "HighMem", |
e53ef38d | 268 | #endif |
2a1e274a | 269 | "Movable", |
033fbae9 DW |
270 | #ifdef CONFIG_ZONE_DEVICE |
271 | "Device", | |
272 | #endif | |
2f1b6248 CL |
273 | }; |
274 | ||
c999fbd3 | 275 | const char * const migratetype_names[MIGRATE_TYPES] = { |
60f30350 VB |
276 | "Unmovable", |
277 | "Movable", | |
278 | "Reclaimable", | |
279 | "HighAtomic", | |
280 | #ifdef CONFIG_CMA | |
281 | "CMA", | |
282 | #endif | |
283 | #ifdef CONFIG_MEMORY_ISOLATION | |
284 | "Isolate", | |
285 | #endif | |
286 | }; | |
287 | ||
1da177e4 | 288 | int min_free_kbytes = 1024; |
42aa83cb | 289 | int user_min_free_kbytes = -1; |
e95d372c KW |
290 | static int watermark_boost_factor __read_mostly = 15000; |
291 | static int watermark_scale_factor = 10; | |
0ee332c1 TH |
292 | |
293 | /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ | |
294 | int movable_zone; | |
295 | EXPORT_SYMBOL(movable_zone); | |
c713216d | 296 | |
418508c1 | 297 | #if MAX_NUMNODES > 1 |
b9726c26 | 298 | unsigned int nr_node_ids __read_mostly = MAX_NUMNODES; |
ce0725f7 | 299 | unsigned int nr_online_nodes __read_mostly = 1; |
418508c1 | 300 | EXPORT_SYMBOL(nr_node_ids); |
62bc62a8 | 301 | EXPORT_SYMBOL(nr_online_nodes); |
418508c1 MS |
302 | #endif |
303 | ||
dcdfdd40 KS |
304 | static bool page_contains_unaccepted(struct page *page, unsigned int order); |
305 | static void accept_page(struct page *page, unsigned int order); | |
306 | static bool try_to_accept_memory(struct zone *zone, unsigned int order); | |
307 | static inline bool has_unaccepted_memory(void); | |
308 | static bool __free_unaccepted(struct page *page); | |
309 | ||
9ef9acb0 MG |
310 | int page_group_by_mobility_disabled __read_mostly; |
311 | ||
3a80a7fa | 312 | #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT |
3c0c12cc WL |
313 | /* |
314 | * During boot we initialize deferred pages on-demand, as needed, but once | |
315 | * page_alloc_init_late() has finished, the deferred pages are all initialized, | |
316 | * and we can permanently disable that path. | |
317 | */ | |
9420f89d | 318 | DEFINE_STATIC_KEY_TRUE(deferred_pages); |
3c0c12cc | 319 | |
94ae8b83 | 320 | static inline bool deferred_pages_enabled(void) |
3c0c12cc | 321 | { |
94ae8b83 | 322 | return static_branch_unlikely(&deferred_pages); |
3c0c12cc WL |
323 | } |
324 | ||
3a80a7fa | 325 | /* |
9420f89d MRI |
326 | * deferred_grow_zone() is __init, but it is called from |
327 | * get_page_from_freelist() during early boot until deferred_pages permanently | |
328 | * disables this call. This is why we have refdata wrapper to avoid warning, | |
329 | * and to ensure that the function body gets unloaded. | |
3a80a7fa | 330 | */ |
9420f89d MRI |
331 | static bool __ref |
332 | _deferred_grow_zone(struct zone *zone, unsigned int order) | |
3a80a7fa | 333 | { |
9420f89d | 334 | return deferred_grow_zone(zone, order); |
3a80a7fa MG |
335 | } |
336 | #else | |
94ae8b83 | 337 | static inline bool deferred_pages_enabled(void) |
2c335680 | 338 | { |
94ae8b83 | 339 | return false; |
2c335680 | 340 | } |
9420f89d | 341 | #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */ |
3a80a7fa | 342 | |
0b423ca2 | 343 | /* Return a pointer to the bitmap storing bits affecting a block of pages */ |
ca891f41 | 344 | static inline unsigned long *get_pageblock_bitmap(const struct page *page, |
0b423ca2 MG |
345 | unsigned long pfn) |
346 | { | |
347 | #ifdef CONFIG_SPARSEMEM | |
f1eca35a | 348 | return section_to_usemap(__pfn_to_section(pfn)); |
0b423ca2 MG |
349 | #else |
350 | return page_zone(page)->pageblock_flags; | |
351 | #endif /* CONFIG_SPARSEMEM */ | |
352 | } | |
353 | ||
ca891f41 | 354 | static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn) |
0b423ca2 MG |
355 | { |
356 | #ifdef CONFIG_SPARSEMEM | |
357 | pfn &= (PAGES_PER_SECTION-1); | |
0b423ca2 | 358 | #else |
4f9bc69a | 359 | pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn); |
0b423ca2 | 360 | #endif /* CONFIG_SPARSEMEM */ |
399b795b | 361 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; |
0b423ca2 MG |
362 | } |
363 | ||
a04d12c2 KS |
364 | /** |
365 | * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages | |
366 | * @page: The page within the block of interest | |
367 | * @pfn: The target page frame number | |
368 | * @mask: mask of bits that the caller is interested in | |
369 | * | |
370 | * Return: pageblock_bits flags | |
371 | */ | |
372 | unsigned long get_pfnblock_flags_mask(const struct page *page, | |
373 | unsigned long pfn, unsigned long mask) | |
0b423ca2 MG |
374 | { |
375 | unsigned long *bitmap; | |
376 | unsigned long bitidx, word_bitidx; | |
377 | unsigned long word; | |
378 | ||
379 | bitmap = get_pageblock_bitmap(page, pfn); | |
380 | bitidx = pfn_to_bitidx(page, pfn); | |
381 | word_bitidx = bitidx / BITS_PER_LONG; | |
382 | bitidx &= (BITS_PER_LONG-1); | |
1c563432 MK |
383 | /* |
384 | * This races, without locks, with set_pfnblock_flags_mask(). Ensure | |
385 | * a consistent read of the memory array, so that results, even though | |
386 | * racy, are not corrupted. | |
387 | */ | |
388 | word = READ_ONCE(bitmap[word_bitidx]); | |
d93d5ab9 | 389 | return (word >> bitidx) & mask; |
0b423ca2 MG |
390 | } |
391 | ||
ca891f41 MWO |
392 | static __always_inline int get_pfnblock_migratetype(const struct page *page, |
393 | unsigned long pfn) | |
0b423ca2 | 394 | { |
a04d12c2 | 395 | return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK); |
0b423ca2 MG |
396 | } |
397 | ||
398 | /** | |
399 | * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages | |
400 | * @page: The page within the block of interest | |
401 | * @flags: The flags to set | |
402 | * @pfn: The target page frame number | |
0b423ca2 MG |
403 | * @mask: mask of bits that the caller is interested in |
404 | */ | |
405 | void set_pfnblock_flags_mask(struct page *page, unsigned long flags, | |
406 | unsigned long pfn, | |
0b423ca2 MG |
407 | unsigned long mask) |
408 | { | |
409 | unsigned long *bitmap; | |
410 | unsigned long bitidx, word_bitidx; | |
04ec0061 | 411 | unsigned long word; |
0b423ca2 MG |
412 | |
413 | BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4); | |
125b860b | 414 | BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits)); |
0b423ca2 MG |
415 | |
416 | bitmap = get_pageblock_bitmap(page, pfn); | |
417 | bitidx = pfn_to_bitidx(page, pfn); | |
418 | word_bitidx = bitidx / BITS_PER_LONG; | |
419 | bitidx &= (BITS_PER_LONG-1); | |
420 | ||
421 | VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page); | |
422 | ||
d93d5ab9 WY |
423 | mask <<= bitidx; |
424 | flags <<= bitidx; | |
0b423ca2 MG |
425 | |
426 | word = READ_ONCE(bitmap[word_bitidx]); | |
04ec0061 UB |
427 | do { |
428 | } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags)); | |
0b423ca2 | 429 | } |
3a80a7fa | 430 | |
ee6f509c | 431 | void set_pageblock_migratetype(struct page *page, int migratetype) |
b2a0ac88 | 432 | { |
5d0f3f72 KM |
433 | if (unlikely(page_group_by_mobility_disabled && |
434 | migratetype < MIGRATE_PCPTYPES)) | |
49255c61 MG |
435 | migratetype = MIGRATE_UNMOVABLE; |
436 | ||
d93d5ab9 | 437 | set_pfnblock_flags_mask(page, (unsigned long)migratetype, |
535b81e2 | 438 | page_to_pfn(page), MIGRATETYPE_MASK); |
b2a0ac88 MG |
439 | } |
440 | ||
13e7444b | 441 | #ifdef CONFIG_DEBUG_VM |
c6a57e19 | 442 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page) |
1da177e4 | 443 | { |
82d9b8c8 | 444 | int ret; |
bdc8cb98 DH |
445 | unsigned seq; |
446 | unsigned long pfn = page_to_pfn(page); | |
b5e6a5a2 | 447 | unsigned long sp, start_pfn; |
c6a57e19 | 448 | |
bdc8cb98 DH |
449 | do { |
450 | seq = zone_span_seqbegin(zone); | |
b5e6a5a2 CS |
451 | start_pfn = zone->zone_start_pfn; |
452 | sp = zone->spanned_pages; | |
82d9b8c8 | 453 | ret = !zone_spans_pfn(zone, pfn); |
bdc8cb98 DH |
454 | } while (zone_span_seqretry(zone, seq)); |
455 | ||
b5e6a5a2 | 456 | if (ret) |
613813e8 DH |
457 | pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n", |
458 | pfn, zone_to_nid(zone), zone->name, | |
459 | start_pfn, start_pfn + sp); | |
b5e6a5a2 | 460 | |
bdc8cb98 | 461 | return ret; |
c6a57e19 DH |
462 | } |
463 | ||
c6a57e19 DH |
464 | /* |
465 | * Temporary debugging check for pages not lying within a given zone. | |
466 | */ | |
d73d3c9f | 467 | static int __maybe_unused bad_range(struct zone *zone, struct page *page) |
c6a57e19 DH |
468 | { |
469 | if (page_outside_zone_boundaries(zone, page)) | |
1da177e4 | 470 | return 1; |
5b855aa3 | 471 | if (zone != page_zone(page)) |
c6a57e19 DH |
472 | return 1; |
473 | ||
1da177e4 LT |
474 | return 0; |
475 | } | |
13e7444b | 476 | #else |
d73d3c9f | 477 | static inline int __maybe_unused bad_range(struct zone *zone, struct page *page) |
13e7444b NP |
478 | { |
479 | return 0; | |
480 | } | |
481 | #endif | |
482 | ||
82a3241a | 483 | static void bad_page(struct page *page, const char *reason) |
1da177e4 | 484 | { |
d936cf9b HD |
485 | static unsigned long resume; |
486 | static unsigned long nr_shown; | |
487 | static unsigned long nr_unshown; | |
488 | ||
489 | /* | |
490 | * Allow a burst of 60 reports, then keep quiet for that minute; | |
491 | * or allow a steady drip of one report per second. | |
492 | */ | |
493 | if (nr_shown == 60) { | |
494 | if (time_before(jiffies, resume)) { | |
495 | nr_unshown++; | |
496 | goto out; | |
497 | } | |
498 | if (nr_unshown) { | |
ff8e8116 | 499 | pr_alert( |
1e9e6365 | 500 | "BUG: Bad page state: %lu messages suppressed\n", |
d936cf9b HD |
501 | nr_unshown); |
502 | nr_unshown = 0; | |
503 | } | |
504 | nr_shown = 0; | |
505 | } | |
506 | if (nr_shown++ == 0) | |
507 | resume = jiffies + 60 * HZ; | |
508 | ||
ff8e8116 | 509 | pr_alert("BUG: Bad page state in process %s pfn:%05lx\n", |
3dc14741 | 510 | current->comm, page_to_pfn(page)); |
d2f07ec0 | 511 | dump_page(page, reason); |
3dc14741 | 512 | |
4f31888c | 513 | print_modules(); |
1da177e4 | 514 | dump_stack(); |
d936cf9b | 515 | out: |
8cc3b392 | 516 | /* Leave bad fields for debug, except PageBuddy could make trouble */ |
22b751c3 | 517 | page_mapcount_reset(page); /* remove PageBuddy */ |
373d4d09 | 518 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
1da177e4 LT |
519 | } |
520 | ||
44042b44 MG |
521 | static inline unsigned int order_to_pindex(int migratetype, int order) |
522 | { | |
44042b44 MG |
523 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
524 | if (order > PAGE_ALLOC_COSTLY_ORDER) { | |
525 | VM_BUG_ON(order != pageblock_order); | |
5d0a661d | 526 | return NR_LOWORDER_PCP_LISTS; |
44042b44 MG |
527 | } |
528 | #else | |
529 | VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER); | |
530 | #endif | |
531 | ||
c1dc69e6 | 532 | return (MIGRATE_PCPTYPES * order) + migratetype; |
44042b44 MG |
533 | } |
534 | ||
535 | static inline int pindex_to_order(unsigned int pindex) | |
536 | { | |
537 | int order = pindex / MIGRATE_PCPTYPES; | |
538 | ||
539 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
5d0a661d | 540 | if (pindex == NR_LOWORDER_PCP_LISTS) |
44042b44 | 541 | order = pageblock_order; |
44042b44 MG |
542 | #else |
543 | VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER); | |
544 | #endif | |
545 | ||
546 | return order; | |
547 | } | |
548 | ||
549 | static inline bool pcp_allowed_order(unsigned int order) | |
550 | { | |
551 | if (order <= PAGE_ALLOC_COSTLY_ORDER) | |
552 | return true; | |
553 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
554 | if (order == pageblock_order) | |
555 | return true; | |
556 | #endif | |
557 | return false; | |
558 | } | |
559 | ||
21d02f8f MG |
560 | static inline void free_the_page(struct page *page, unsigned int order) |
561 | { | |
44042b44 MG |
562 | if (pcp_allowed_order(order)) /* Via pcp? */ |
563 | free_unref_page(page, order); | |
21d02f8f MG |
564 | else |
565 | __free_pages_ok(page, order, FPI_NONE); | |
566 | } | |
567 | ||
1da177e4 LT |
568 | /* |
569 | * Higher-order pages are called "compound pages". They are structured thusly: | |
570 | * | |
1d798ca3 | 571 | * The first PAGE_SIZE page is called the "head page" and have PG_head set. |
1da177e4 | 572 | * |
1d798ca3 KS |
573 | * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded |
574 | * in bit 0 of page->compound_head. The rest of bits is pointer to head page. | |
1da177e4 | 575 | * |
1d798ca3 | 576 | * The first tail page's ->compound_order holds the order of allocation. |
41d78ba5 | 577 | * This usage means that zero-order pages may not be compound. |
1da177e4 | 578 | */ |
d98c7a09 | 579 | |
d00181b9 | 580 | void prep_compound_page(struct page *page, unsigned int order) |
18229df5 AW |
581 | { |
582 | int i; | |
583 | int nr_pages = 1 << order; | |
584 | ||
18229df5 | 585 | __SetPageHead(page); |
5b24eeef JM |
586 | for (i = 1; i < nr_pages; i++) |
587 | prep_compound_tail(page, i); | |
1378a5ee | 588 | |
5b24eeef | 589 | prep_compound_head(page, order); |
18229df5 AW |
590 | } |
591 | ||
5375336c MWO |
592 | void destroy_large_folio(struct folio *folio) |
593 | { | |
dd6fa0b6 | 594 | if (folio_test_hugetlb(folio)) { |
454a00c4 | 595 | free_huge_folio(folio); |
dd6fa0b6 MWO |
596 | return; |
597 | } | |
8dc4a8f1 | 598 | |
de53c05f | 599 | if (folio_test_large_rmappable(folio)) |
8dc4a8f1 | 600 | folio_undo_large_rmappable(folio); |
5375336c | 601 | |
0f2f43fa MWO |
602 | mem_cgroup_uncharge(folio); |
603 | free_the_page(&folio->page, folio_order(folio)); | |
5375336c MWO |
604 | } |
605 | ||
ab130f91 | 606 | static inline void set_buddy_order(struct page *page, unsigned int order) |
6aa3001b | 607 | { |
4c21e2f2 | 608 | set_page_private(page, order); |
676165a8 | 609 | __SetPageBuddy(page); |
1da177e4 LT |
610 | } |
611 | ||
5e1f0f09 MG |
612 | #ifdef CONFIG_COMPACTION |
613 | static inline struct capture_control *task_capc(struct zone *zone) | |
614 | { | |
615 | struct capture_control *capc = current->capture_control; | |
616 | ||
deba0487 | 617 | return unlikely(capc) && |
5e1f0f09 MG |
618 | !(current->flags & PF_KTHREAD) && |
619 | !capc->page && | |
deba0487 | 620 | capc->cc->zone == zone ? capc : NULL; |
5e1f0f09 MG |
621 | } |
622 | ||
623 | static inline bool | |
624 | compaction_capture(struct capture_control *capc, struct page *page, | |
625 | int order, int migratetype) | |
626 | { | |
627 | if (!capc || order != capc->cc->order) | |
628 | return false; | |
629 | ||
630 | /* Do not accidentally pollute CMA or isolated regions*/ | |
631 | if (is_migrate_cma(migratetype) || | |
632 | is_migrate_isolate(migratetype)) | |
633 | return false; | |
634 | ||
635 | /* | |
f0953a1b | 636 | * Do not let lower order allocations pollute a movable pageblock. |
5e1f0f09 MG |
637 | * This might let an unmovable request use a reclaimable pageblock |
638 | * and vice-versa but no more than normal fallback logic which can | |
639 | * have trouble finding a high-order free page. | |
640 | */ | |
641 | if (order < pageblock_order && migratetype == MIGRATE_MOVABLE) | |
642 | return false; | |
643 | ||
644 | capc->page = page; | |
645 | return true; | |
646 | } | |
647 | ||
648 | #else | |
649 | static inline struct capture_control *task_capc(struct zone *zone) | |
650 | { | |
651 | return NULL; | |
652 | } | |
653 | ||
654 | static inline bool | |
655 | compaction_capture(struct capture_control *capc, struct page *page, | |
656 | int order, int migratetype) | |
657 | { | |
658 | return false; | |
659 | } | |
660 | #endif /* CONFIG_COMPACTION */ | |
661 | ||
6ab01363 AD |
662 | /* Used for pages not on another list */ |
663 | static inline void add_to_free_list(struct page *page, struct zone *zone, | |
664 | unsigned int order, int migratetype) | |
665 | { | |
666 | struct free_area *area = &zone->free_area[order]; | |
667 | ||
bf75f200 | 668 | list_add(&page->buddy_list, &area->free_list[migratetype]); |
6ab01363 AD |
669 | area->nr_free++; |
670 | } | |
671 | ||
672 | /* Used for pages not on another list */ | |
673 | static inline void add_to_free_list_tail(struct page *page, struct zone *zone, | |
674 | unsigned int order, int migratetype) | |
675 | { | |
676 | struct free_area *area = &zone->free_area[order]; | |
677 | ||
bf75f200 | 678 | list_add_tail(&page->buddy_list, &area->free_list[migratetype]); |
6ab01363 AD |
679 | area->nr_free++; |
680 | } | |
681 | ||
293ffa5e DH |
682 | /* |
683 | * Used for pages which are on another list. Move the pages to the tail | |
684 | * of the list - so the moved pages won't immediately be considered for | |
685 | * allocation again (e.g., optimization for memory onlining). | |
686 | */ | |
6ab01363 AD |
687 | static inline void move_to_free_list(struct page *page, struct zone *zone, |
688 | unsigned int order, int migratetype) | |
689 | { | |
690 | struct free_area *area = &zone->free_area[order]; | |
691 | ||
bf75f200 | 692 | list_move_tail(&page->buddy_list, &area->free_list[migratetype]); |
6ab01363 AD |
693 | } |
694 | ||
695 | static inline void del_page_from_free_list(struct page *page, struct zone *zone, | |
696 | unsigned int order) | |
697 | { | |
36e66c55 AD |
698 | /* clear reported state and update reported page count */ |
699 | if (page_reported(page)) | |
700 | __ClearPageReported(page); | |
701 | ||
bf75f200 | 702 | list_del(&page->buddy_list); |
6ab01363 AD |
703 | __ClearPageBuddy(page); |
704 | set_page_private(page, 0); | |
705 | zone->free_area[order].nr_free--; | |
706 | } | |
707 | ||
5d671eb4 MRI |
708 | static inline struct page *get_page_from_free_area(struct free_area *area, |
709 | int migratetype) | |
710 | { | |
711 | return list_first_entry_or_null(&area->free_list[migratetype], | |
1bf61092 | 712 | struct page, buddy_list); |
5d671eb4 MRI |
713 | } |
714 | ||
a2129f24 AD |
715 | /* |
716 | * If this is not the largest possible page, check if the buddy | |
717 | * of the next-highest order is free. If it is, it's possible | |
718 | * that pages are being freed that will coalesce soon. In case, | |
719 | * that is happening, add the free page to the tail of the list | |
720 | * so it's less likely to be used soon and more likely to be merged | |
721 | * as a higher order page | |
722 | */ | |
723 | static inline bool | |
724 | buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn, | |
725 | struct page *page, unsigned int order) | |
726 | { | |
8170ac47 ZY |
727 | unsigned long higher_page_pfn; |
728 | struct page *higher_page; | |
a2129f24 | 729 | |
5e0a760b | 730 | if (order >= MAX_PAGE_ORDER - 1) |
a2129f24 AD |
731 | return false; |
732 | ||
8170ac47 ZY |
733 | higher_page_pfn = buddy_pfn & pfn; |
734 | higher_page = page + (higher_page_pfn - pfn); | |
a2129f24 | 735 | |
8170ac47 ZY |
736 | return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1, |
737 | NULL) != NULL; | |
a2129f24 AD |
738 | } |
739 | ||
1da177e4 LT |
740 | /* |
741 | * Freeing function for a buddy system allocator. | |
742 | * | |
743 | * The concept of a buddy system is to maintain direct-mapped table | |
744 | * (containing bit values) for memory blocks of various "orders". | |
745 | * The bottom level table contains the map for the smallest allocatable | |
746 | * units of memory (here, pages), and each level above it describes | |
747 | * pairs of units from the levels below, hence, "buddies". | |
748 | * At a high level, all that happens here is marking the table entry | |
749 | * at the bottom level available, and propagating the changes upward | |
750 | * as necessary, plus some accounting needed to play nicely with other | |
751 | * parts of the VM system. | |
752 | * At each level, we keep a list of pages, which are heads of continuous | |
6e292b9b MW |
753 | * free pages of length of (1 << order) and marked with PageBuddy. |
754 | * Page's order is recorded in page_private(page) field. | |
1da177e4 | 755 | * So when we are allocating or freeing one, we can derive the state of the |
5f63b720 MN |
756 | * other. That is, if we allocate a small block, and both were |
757 | * free, the remainder of the region must be split into blocks. | |
1da177e4 | 758 | * If a block is freed, and its buddy is also free, then this |
5f63b720 | 759 | * triggers coalescing into a block of larger size. |
1da177e4 | 760 | * |
6d49e352 | 761 | * -- nyc |
1da177e4 LT |
762 | */ |
763 | ||
48db57f8 | 764 | static inline void __free_one_page(struct page *page, |
dc4b0caf | 765 | unsigned long pfn, |
ed0ae21d | 766 | struct zone *zone, unsigned int order, |
f04a5d5d | 767 | int migratetype, fpi_t fpi_flags) |
1da177e4 | 768 | { |
a2129f24 | 769 | struct capture_control *capc = task_capc(zone); |
dae37a5d | 770 | unsigned long buddy_pfn = 0; |
a2129f24 | 771 | unsigned long combined_pfn; |
a2129f24 AD |
772 | struct page *buddy; |
773 | bool to_tail; | |
d9dddbf5 | 774 | |
d29bb978 | 775 | VM_BUG_ON(!zone_is_initialized(zone)); |
6e9f0d58 | 776 | VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page); |
1da177e4 | 777 | |
ed0ae21d | 778 | VM_BUG_ON(migratetype == -1); |
d9dddbf5 | 779 | if (likely(!is_migrate_isolate(migratetype))) |
8f82b55d | 780 | __mod_zone_freepage_state(zone, 1 << order, migratetype); |
ed0ae21d | 781 | |
76741e77 | 782 | VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page); |
309381fe | 783 | VM_BUG_ON_PAGE(bad_range(zone, page), page); |
1da177e4 | 784 | |
5e0a760b | 785 | while (order < MAX_PAGE_ORDER) { |
5e1f0f09 MG |
786 | if (compaction_capture(capc, page, order, migratetype)) { |
787 | __mod_zone_freepage_state(zone, -(1 << order), | |
788 | migratetype); | |
789 | return; | |
790 | } | |
13ad59df | 791 | |
8170ac47 ZY |
792 | buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn); |
793 | if (!buddy) | |
d9dddbf5 | 794 | goto done_merging; |
bb0e28eb ZY |
795 | |
796 | if (unlikely(order >= pageblock_order)) { | |
797 | /* | |
798 | * We want to prevent merge between freepages on pageblock | |
799 | * without fallbacks and normal pageblock. Without this, | |
800 | * pageblock isolation could cause incorrect freepage or CMA | |
801 | * accounting or HIGHATOMIC accounting. | |
802 | */ | |
b5ffd297 | 803 | int buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn); |
bb0e28eb ZY |
804 | |
805 | if (migratetype != buddy_mt | |
806 | && (!migratetype_is_mergeable(migratetype) || | |
807 | !migratetype_is_mergeable(buddy_mt))) | |
808 | goto done_merging; | |
809 | } | |
810 | ||
c0a32fc5 SG |
811 | /* |
812 | * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, | |
813 | * merge with it and move up one order. | |
814 | */ | |
b03641af | 815 | if (page_is_guard(buddy)) |
2847cf95 | 816 | clear_page_guard(zone, buddy, order, migratetype); |
b03641af | 817 | else |
6ab01363 | 818 | del_page_from_free_list(buddy, zone, order); |
76741e77 VB |
819 | combined_pfn = buddy_pfn & pfn; |
820 | page = page + (combined_pfn - pfn); | |
821 | pfn = combined_pfn; | |
1da177e4 LT |
822 | order++; |
823 | } | |
d9dddbf5 VB |
824 | |
825 | done_merging: | |
ab130f91 | 826 | set_buddy_order(page, order); |
6dda9d55 | 827 | |
47b6a24a DH |
828 | if (fpi_flags & FPI_TO_TAIL) |
829 | to_tail = true; | |
830 | else if (is_shuffle_order(order)) | |
a2129f24 | 831 | to_tail = shuffle_pick_tail(); |
97500a4a | 832 | else |
a2129f24 | 833 | to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order); |
97500a4a | 834 | |
a2129f24 | 835 | if (to_tail) |
6ab01363 | 836 | add_to_free_list_tail(page, zone, order, migratetype); |
a2129f24 | 837 | else |
6ab01363 | 838 | add_to_free_list(page, zone, order, migratetype); |
36e66c55 AD |
839 | |
840 | /* Notify page reporting subsystem of freed page */ | |
f04a5d5d | 841 | if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY)) |
36e66c55 | 842 | page_reporting_notify_free(order); |
1da177e4 LT |
843 | } |
844 | ||
b2c9e2fb ZY |
845 | /** |
846 | * split_free_page() -- split a free page at split_pfn_offset | |
847 | * @free_page: the original free page | |
848 | * @order: the order of the page | |
849 | * @split_pfn_offset: split offset within the page | |
850 | * | |
86d28b07 ZY |
851 | * Return -ENOENT if the free page is changed, otherwise 0 |
852 | * | |
b2c9e2fb ZY |
853 | * It is used when the free page crosses two pageblocks with different migratetypes |
854 | * at split_pfn_offset within the page. The split free page will be put into | |
855 | * separate migratetype lists afterwards. Otherwise, the function achieves | |
856 | * nothing. | |
857 | */ | |
86d28b07 ZY |
858 | int split_free_page(struct page *free_page, |
859 | unsigned int order, unsigned long split_pfn_offset) | |
b2c9e2fb ZY |
860 | { |
861 | struct zone *zone = page_zone(free_page); | |
862 | unsigned long free_page_pfn = page_to_pfn(free_page); | |
863 | unsigned long pfn; | |
864 | unsigned long flags; | |
865 | int free_page_order; | |
86d28b07 ZY |
866 | int mt; |
867 | int ret = 0; | |
b2c9e2fb | 868 | |
88ee1343 | 869 | if (split_pfn_offset == 0) |
86d28b07 | 870 | return ret; |
88ee1343 | 871 | |
b2c9e2fb | 872 | spin_lock_irqsave(&zone->lock, flags); |
86d28b07 ZY |
873 | |
874 | if (!PageBuddy(free_page) || buddy_order(free_page) != order) { | |
875 | ret = -ENOENT; | |
876 | goto out; | |
877 | } | |
878 | ||
b5ffd297 | 879 | mt = get_pfnblock_migratetype(free_page, free_page_pfn); |
86d28b07 ZY |
880 | if (likely(!is_migrate_isolate(mt))) |
881 | __mod_zone_freepage_state(zone, -(1UL << order), mt); | |
882 | ||
b2c9e2fb ZY |
883 | del_page_from_free_list(free_page, zone, order); |
884 | for (pfn = free_page_pfn; | |
885 | pfn < free_page_pfn + (1UL << order);) { | |
886 | int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn); | |
887 | ||
86d28b07 | 888 | free_page_order = min_t(unsigned int, |
88ee1343 ZY |
889 | pfn ? __ffs(pfn) : order, |
890 | __fls(split_pfn_offset)); | |
b2c9e2fb ZY |
891 | __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order, |
892 | mt, FPI_NONE); | |
893 | pfn += 1UL << free_page_order; | |
894 | split_pfn_offset -= (1UL << free_page_order); | |
895 | /* we have done the first part, now switch to second part */ | |
896 | if (split_pfn_offset == 0) | |
897 | split_pfn_offset = (1UL << order) - (pfn - free_page_pfn); | |
898 | } | |
86d28b07 | 899 | out: |
b2c9e2fb | 900 | spin_unlock_irqrestore(&zone->lock, flags); |
86d28b07 | 901 | return ret; |
b2c9e2fb | 902 | } |
7bfec6f4 MG |
903 | /* |
904 | * A bad page could be due to a number of fields. Instead of multiple branches, | |
905 | * try and check multiple fields with one check. The caller must do a detailed | |
906 | * check if necessary. | |
907 | */ | |
908 | static inline bool page_expected_state(struct page *page, | |
909 | unsigned long check_flags) | |
910 | { | |
911 | if (unlikely(atomic_read(&page->_mapcount) != -1)) | |
912 | return false; | |
913 | ||
914 | if (unlikely((unsigned long)page->mapping | | |
915 | page_ref_count(page) | | |
916 | #ifdef CONFIG_MEMCG | |
48060834 | 917 | page->memcg_data | |
dba1b8a7 JDB |
918 | #endif |
919 | #ifdef CONFIG_PAGE_POOL | |
920 | ((page->pp_magic & ~0x3UL) == PP_SIGNATURE) | | |
7bfec6f4 MG |
921 | #endif |
922 | (page->flags & check_flags))) | |
923 | return false; | |
924 | ||
925 | return true; | |
926 | } | |
927 | ||
58b7f119 | 928 | static const char *page_bad_reason(struct page *page, unsigned long flags) |
1da177e4 | 929 | { |
82a3241a | 930 | const char *bad_reason = NULL; |
f0b791a3 | 931 | |
53f9263b | 932 | if (unlikely(atomic_read(&page->_mapcount) != -1)) |
f0b791a3 DH |
933 | bad_reason = "nonzero mapcount"; |
934 | if (unlikely(page->mapping != NULL)) | |
935 | bad_reason = "non-NULL mapping"; | |
fe896d18 | 936 | if (unlikely(page_ref_count(page) != 0)) |
0139aa7b | 937 | bad_reason = "nonzero _refcount"; |
58b7f119 WY |
938 | if (unlikely(page->flags & flags)) { |
939 | if (flags == PAGE_FLAGS_CHECK_AT_PREP) | |
940 | bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set"; | |
941 | else | |
942 | bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set"; | |
f0b791a3 | 943 | } |
9edad6ea | 944 | #ifdef CONFIG_MEMCG |
48060834 | 945 | if (unlikely(page->memcg_data)) |
9edad6ea | 946 | bad_reason = "page still charged to cgroup"; |
dba1b8a7 JDB |
947 | #endif |
948 | #ifdef CONFIG_PAGE_POOL | |
949 | if (unlikely((page->pp_magic & ~0x3UL) == PP_SIGNATURE)) | |
950 | bad_reason = "page_pool leak"; | |
9edad6ea | 951 | #endif |
58b7f119 WY |
952 | return bad_reason; |
953 | } | |
954 | ||
a8368cd8 | 955 | static void free_page_is_bad_report(struct page *page) |
58b7f119 WY |
956 | { |
957 | bad_page(page, | |
958 | page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE)); | |
bb552ac6 MG |
959 | } |
960 | ||
a8368cd8 | 961 | static inline bool free_page_is_bad(struct page *page) |
bb552ac6 | 962 | { |
da838d4f | 963 | if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE))) |
a8368cd8 | 964 | return false; |
bb552ac6 MG |
965 | |
966 | /* Something has gone sideways, find it */ | |
a8368cd8 AM |
967 | free_page_is_bad_report(page); |
968 | return true; | |
1da177e4 LT |
969 | } |
970 | ||
ecbb490d KW |
971 | static inline bool is_check_pages_enabled(void) |
972 | { | |
973 | return static_branch_unlikely(&check_pages_enabled); | |
974 | } | |
975 | ||
8666925c | 976 | static int free_tail_page_prepare(struct page *head_page, struct page *page) |
4db7548c | 977 | { |
94688e8e | 978 | struct folio *folio = (struct folio *)head_page; |
4db7548c MG |
979 | int ret = 1; |
980 | ||
981 | /* | |
982 | * We rely page->lru.next never has bit 0 set, unless the page | |
983 | * is PageTail(). Let's make sure that's true even for poisoned ->lru. | |
984 | */ | |
985 | BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1); | |
986 | ||
ecbb490d | 987 | if (!is_check_pages_enabled()) { |
4db7548c MG |
988 | ret = 0; |
989 | goto out; | |
990 | } | |
991 | switch (page - head_page) { | |
992 | case 1: | |
cb67f428 | 993 | /* the first tail page: these may be in place of ->mapping */ |
65a689f3 MWO |
994 | if (unlikely(folio_entire_mapcount(folio))) { |
995 | bad_page(page, "nonzero entire_mapcount"); | |
4db7548c MG |
996 | goto out; |
997 | } | |
65a689f3 MWO |
998 | if (unlikely(atomic_read(&folio->_nr_pages_mapped))) { |
999 | bad_page(page, "nonzero nr_pages_mapped"); | |
cb67f428 HD |
1000 | goto out; |
1001 | } | |
94688e8e MWO |
1002 | if (unlikely(atomic_read(&folio->_pincount))) { |
1003 | bad_page(page, "nonzero pincount"); | |
cb67f428 HD |
1004 | goto out; |
1005 | } | |
4db7548c MG |
1006 | break; |
1007 | case 2: | |
1008 | /* | |
1009 | * the second tail page: ->mapping is | |
fa3015b7 | 1010 | * deferred_list.next -- ignore value. |
4db7548c MG |
1011 | */ |
1012 | break; | |
1013 | default: | |
1014 | if (page->mapping != TAIL_MAPPING) { | |
82a3241a | 1015 | bad_page(page, "corrupted mapping in tail page"); |
4db7548c MG |
1016 | goto out; |
1017 | } | |
1018 | break; | |
1019 | } | |
1020 | if (unlikely(!PageTail(page))) { | |
82a3241a | 1021 | bad_page(page, "PageTail not set"); |
4db7548c MG |
1022 | goto out; |
1023 | } | |
1024 | if (unlikely(compound_head(page) != head_page)) { | |
82a3241a | 1025 | bad_page(page, "compound_head not consistent"); |
4db7548c MG |
1026 | goto out; |
1027 | } | |
1028 | ret = 0; | |
1029 | out: | |
1030 | page->mapping = NULL; | |
1031 | clear_compound_head(page); | |
1032 | return ret; | |
1033 | } | |
1034 | ||
94ae8b83 AK |
1035 | /* |
1036 | * Skip KASAN memory poisoning when either: | |
1037 | * | |
0a54864f PC |
1038 | * 1. For generic KASAN: deferred memory initialization has not yet completed. |
1039 | * Tag-based KASAN modes skip pages freed via deferred memory initialization | |
1040 | * using page tags instead (see below). | |
1041 | * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating | |
1042 | * that error detection is disabled for accesses via the page address. | |
1043 | * | |
1044 | * Pages will have match-all tags in the following circumstances: | |
1045 | * | |
1046 | * 1. Pages are being initialized for the first time, including during deferred | |
1047 | * memory init; see the call to page_kasan_tag_reset in __init_single_page. | |
1048 | * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the | |
1049 | * exception of pages unpoisoned by kasan_unpoison_vmalloc. | |
1050 | * 3. The allocation was excluded from being checked due to sampling, | |
44383cef | 1051 | * see the call to kasan_unpoison_pages. |
94ae8b83 AK |
1052 | * |
1053 | * Poisoning pages during deferred memory init will greatly lengthen the | |
1054 | * process and cause problem in large memory systems as the deferred pages | |
1055 | * initialization is done with interrupt disabled. | |
1056 | * | |
1057 | * Assuming that there will be no reference to those newly initialized | |
1058 | * pages before they are ever allocated, this should have no effect on | |
1059 | * KASAN memory tracking as the poison will be properly inserted at page | |
1060 | * allocation time. The only corner case is when pages are allocated by | |
1061 | * on-demand allocation and then freed again before the deferred pages | |
1062 | * initialization is done, but this is not likely to happen. | |
1063 | */ | |
1064 | static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags) | |
1065 | { | |
0a54864f PC |
1066 | if (IS_ENABLED(CONFIG_KASAN_GENERIC)) |
1067 | return deferred_pages_enabled(); | |
1068 | ||
5cb6674b | 1069 | return page_kasan_tag(page) == KASAN_TAG_KERNEL; |
94ae8b83 AK |
1070 | } |
1071 | ||
aeaec8e2 | 1072 | static void kernel_init_pages(struct page *page, int numpages) |
6471384a AP |
1073 | { |
1074 | int i; | |
1075 | ||
9e15afa5 QC |
1076 | /* s390's use of memset() could override KASAN redzones. */ |
1077 | kasan_disable_current(); | |
d9da8f6c AK |
1078 | for (i = 0; i < numpages; i++) |
1079 | clear_highpage_kasan_tagged(page + i); | |
9e15afa5 | 1080 | kasan_enable_current(); |
6471384a AP |
1081 | } |
1082 | ||
e2769dbd | 1083 | static __always_inline bool free_pages_prepare(struct page *page, |
700d2e9a | 1084 | unsigned int order, fpi_t fpi_flags) |
4db7548c | 1085 | { |
e2769dbd | 1086 | int bad = 0; |
f446883d | 1087 | bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags); |
c3525330 | 1088 | bool init = want_init_on_free(); |
76f26535 | 1089 | bool compound = PageCompound(page); |
4db7548c | 1090 | |
4db7548c MG |
1091 | VM_BUG_ON_PAGE(PageTail(page), page); |
1092 | ||
e2769dbd | 1093 | trace_mm_page_free(page, order); |
b073d7f8 | 1094 | kmsan_free_page(page, order); |
e2769dbd | 1095 | |
17b46e7b BJ |
1096 | if (memcg_kmem_online() && PageMemcgKmem(page)) |
1097 | __memcg_kmem_uncharge_page(page, order); | |
1098 | ||
79f5f8fa | 1099 | if (unlikely(PageHWPoison(page)) && !order) { |
17b46e7b | 1100 | /* Do not let hwpoison pages hit pcplists/buddy */ |
79f5f8fa | 1101 | reset_page_owner(page, order); |
df4e817b | 1102 | page_table_check_free(page, order); |
79f5f8fa OS |
1103 | return false; |
1104 | } | |
1105 | ||
76f26535 HY |
1106 | VM_BUG_ON_PAGE(compound && compound_order(page) != order, page); |
1107 | ||
e2769dbd MG |
1108 | /* |
1109 | * Check tail pages before head page information is cleared to | |
1110 | * avoid checking PageCompound for order-0 pages. | |
1111 | */ | |
1112 | if (unlikely(order)) { | |
e2769dbd MG |
1113 | int i; |
1114 | ||
cb67f428 | 1115 | if (compound) |
9c5ccf2d | 1116 | page[1].flags &= ~PAGE_FLAGS_SECOND; |
e2769dbd MG |
1117 | for (i = 1; i < (1 << order); i++) { |
1118 | if (compound) | |
8666925c | 1119 | bad += free_tail_page_prepare(page, page + i); |
fce0b421 | 1120 | if (is_check_pages_enabled()) { |
8666925c | 1121 | if (free_page_is_bad(page + i)) { |
700d2e9a VB |
1122 | bad++; |
1123 | continue; | |
1124 | } | |
e2769dbd MG |
1125 | } |
1126 | (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
1127 | } | |
1128 | } | |
bda807d4 | 1129 | if (PageMappingFlags(page)) |
4db7548c | 1130 | page->mapping = NULL; |
fce0b421 | 1131 | if (is_check_pages_enabled()) { |
700d2e9a VB |
1132 | if (free_page_is_bad(page)) |
1133 | bad++; | |
1134 | if (bad) | |
1135 | return false; | |
1136 | } | |
4db7548c | 1137 | |
e2769dbd MG |
1138 | page_cpupid_reset_last(page); |
1139 | page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; | |
1140 | reset_page_owner(page, order); | |
df4e817b | 1141 | page_table_check_free(page, order); |
4db7548c MG |
1142 | |
1143 | if (!PageHighMem(page)) { | |
1144 | debug_check_no_locks_freed(page_address(page), | |
e2769dbd | 1145 | PAGE_SIZE << order); |
4db7548c | 1146 | debug_check_no_obj_freed(page_address(page), |
e2769dbd | 1147 | PAGE_SIZE << order); |
4db7548c | 1148 | } |
6471384a | 1149 | |
8db26a3d VB |
1150 | kernel_poison_pages(page, 1 << order); |
1151 | ||
f9d79e8d | 1152 | /* |
1bb5eab3 | 1153 | * As memory initialization might be integrated into KASAN, |
7c13c163 | 1154 | * KASAN poisoning and memory initialization code must be |
1bb5eab3 AK |
1155 | * kept together to avoid discrepancies in behavior. |
1156 | * | |
f9d79e8d AK |
1157 | * With hardware tag-based KASAN, memory tags must be set before the |
1158 | * page becomes unavailable via debug_pagealloc or arch_free_page. | |
1159 | */ | |
f446883d | 1160 | if (!skip_kasan_poison) { |
c3525330 | 1161 | kasan_poison_pages(page, order, init); |
f9d79e8d | 1162 | |
db8a0477 AK |
1163 | /* Memory is already initialized if KASAN did it internally. */ |
1164 | if (kasan_has_integrated_init()) | |
1165 | init = false; | |
1166 | } | |
1167 | if (init) | |
aeaec8e2 | 1168 | kernel_init_pages(page, 1 << order); |
db8a0477 | 1169 | |
234fdce8 QC |
1170 | /* |
1171 | * arch_free_page() can make the page's contents inaccessible. s390 | |
1172 | * does this. So nothing which can access the page's contents should | |
1173 | * happen after this. | |
1174 | */ | |
1175 | arch_free_page(page, order); | |
1176 | ||
77bc7fd6 | 1177 | debug_pagealloc_unmap_pages(page, 1 << order); |
d6332692 | 1178 | |
4db7548c MG |
1179 | return true; |
1180 | } | |
1181 | ||
1da177e4 | 1182 | /* |
5f8dcc21 | 1183 | * Frees a number of pages from the PCP lists |
7cba630b | 1184 | * Assumes all pages on list are in same zone. |
207f36ee | 1185 | * count is the number of pages to free. |
1da177e4 | 1186 | */ |
5f8dcc21 | 1187 | static void free_pcppages_bulk(struct zone *zone, int count, |
fd56eef2 MG |
1188 | struct per_cpu_pages *pcp, |
1189 | int pindex) | |
1da177e4 | 1190 | { |
57490774 | 1191 | unsigned long flags; |
44042b44 | 1192 | unsigned int order; |
3777999d | 1193 | bool isolated_pageblocks; |
8b10b465 | 1194 | struct page *page; |
f2260e6b | 1195 | |
88e8ac11 CTR |
1196 | /* |
1197 | * Ensure proper count is passed which otherwise would stuck in the | |
1198 | * below while (list_empty(list)) loop. | |
1199 | */ | |
1200 | count = min(pcp->count, count); | |
d61372bc MG |
1201 | |
1202 | /* Ensure requested pindex is drained first. */ | |
1203 | pindex = pindex - 1; | |
1204 | ||
57490774 | 1205 | spin_lock_irqsave(&zone->lock, flags); |
8b10b465 MG |
1206 | isolated_pageblocks = has_isolate_pageblock(zone); |
1207 | ||
44042b44 | 1208 | while (count > 0) { |
5f8dcc21 | 1209 | struct list_head *list; |
fd56eef2 | 1210 | int nr_pages; |
5f8dcc21 | 1211 | |
fd56eef2 | 1212 | /* Remove pages from lists in a round-robin fashion. */ |
5f8dcc21 | 1213 | do { |
f142b2c2 KS |
1214 | if (++pindex > NR_PCP_LISTS - 1) |
1215 | pindex = 0; | |
44042b44 | 1216 | list = &pcp->lists[pindex]; |
f142b2c2 | 1217 | } while (list_empty(list)); |
48db57f8 | 1218 | |
44042b44 | 1219 | order = pindex_to_order(pindex); |
fd56eef2 | 1220 | nr_pages = 1 << order; |
a6f9edd6 | 1221 | do { |
8b10b465 MG |
1222 | int mt; |
1223 | ||
bf75f200 | 1224 | page = list_last_entry(list, struct page, pcp_list); |
8b10b465 MG |
1225 | mt = get_pcppage_migratetype(page); |
1226 | ||
0a5f4e5b | 1227 | /* must delete to avoid corrupting pcp list */ |
bf75f200 | 1228 | list_del(&page->pcp_list); |
fd56eef2 MG |
1229 | count -= nr_pages; |
1230 | pcp->count -= nr_pages; | |
aa016d14 | 1231 | |
8b10b465 MG |
1232 | /* MIGRATE_ISOLATE page should not go to pcplists */ |
1233 | VM_BUG_ON_PAGE(is_migrate_isolate(mt), page); | |
1234 | /* Pageblock could have been isolated meanwhile */ | |
1235 | if (unlikely(isolated_pageblocks)) | |
1236 | mt = get_pageblock_migratetype(page); | |
0a5f4e5b | 1237 | |
8b10b465 MG |
1238 | __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE); |
1239 | trace_mm_page_pcpu_drain(page, order, mt); | |
1240 | } while (count > 0 && !list_empty(list)); | |
0a5f4e5b | 1241 | } |
8b10b465 | 1242 | |
57490774 | 1243 | spin_unlock_irqrestore(&zone->lock, flags); |
1da177e4 LT |
1244 | } |
1245 | ||
dc4b0caf MG |
1246 | static void free_one_page(struct zone *zone, |
1247 | struct page *page, unsigned long pfn, | |
7aeb09f9 | 1248 | unsigned int order, |
7fef431b | 1249 | int migratetype, fpi_t fpi_flags) |
1da177e4 | 1250 | { |
df1acc85 MG |
1251 | unsigned long flags; |
1252 | ||
1253 | spin_lock_irqsave(&zone->lock, flags); | |
ad53f92e JK |
1254 | if (unlikely(has_isolate_pageblock(zone) || |
1255 | is_migrate_isolate(migratetype))) { | |
1256 | migratetype = get_pfnblock_migratetype(page, pfn); | |
ad53f92e | 1257 | } |
7fef431b | 1258 | __free_one_page(page, pfn, zone, order, migratetype, fpi_flags); |
df1acc85 | 1259 | spin_unlock_irqrestore(&zone->lock, flags); |
48db57f8 NP |
1260 | } |
1261 | ||
7fef431b DH |
1262 | static void __free_pages_ok(struct page *page, unsigned int order, |
1263 | fpi_t fpi_flags) | |
ec95f53a | 1264 | { |
95e34412 | 1265 | int migratetype; |
dc4b0caf | 1266 | unsigned long pfn = page_to_pfn(page); |
56f0e661 | 1267 | struct zone *zone = page_zone(page); |
ec95f53a | 1268 | |
700d2e9a | 1269 | if (!free_pages_prepare(page, order, fpi_flags)) |
ec95f53a KM |
1270 | return; |
1271 | ||
ac4b2901 DW |
1272 | /* |
1273 | * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here | |
1274 | * is used to avoid calling get_pfnblock_migratetype() under the lock. | |
1275 | * This will reduce the lock holding time. | |
1276 | */ | |
cfc47a28 | 1277 | migratetype = get_pfnblock_migratetype(page, pfn); |
dbbee9d5 | 1278 | |
250ae189 | 1279 | free_one_page(zone, page, pfn, order, migratetype, fpi_flags); |
90249993 | 1280 | |
d34b0733 | 1281 | __count_vm_events(PGFREE, 1 << order); |
1da177e4 LT |
1282 | } |
1283 | ||
a9cd410a | 1284 | void __free_pages_core(struct page *page, unsigned int order) |
a226f6c8 | 1285 | { |
c3993076 | 1286 | unsigned int nr_pages = 1 << order; |
e2d0bd2b | 1287 | struct page *p = page; |
c3993076 | 1288 | unsigned int loop; |
a226f6c8 | 1289 | |
7fef431b DH |
1290 | /* |
1291 | * When initializing the memmap, __init_single_page() sets the refcount | |
1292 | * of all pages to 1 ("allocated"/"not free"). We have to set the | |
1293 | * refcount of all involved pages to 0. | |
1294 | */ | |
e2d0bd2b YL |
1295 | prefetchw(p); |
1296 | for (loop = 0; loop < (nr_pages - 1); loop++, p++) { | |
1297 | prefetchw(p + 1); | |
c3993076 JW |
1298 | __ClearPageReserved(p); |
1299 | set_page_count(p, 0); | |
a226f6c8 | 1300 | } |
e2d0bd2b YL |
1301 | __ClearPageReserved(p); |
1302 | set_page_count(p, 0); | |
c3993076 | 1303 | |
9705bea5 | 1304 | atomic_long_add(nr_pages, &page_zone(page)->managed_pages); |
7fef431b | 1305 | |
dcdfdd40 | 1306 | if (page_contains_unaccepted(page, order)) { |
5e0a760b | 1307 | if (order == MAX_PAGE_ORDER && __free_unaccepted(page)) |
dcdfdd40 KS |
1308 | return; |
1309 | ||
1310 | accept_page(page, order); | |
1311 | } | |
1312 | ||
7fef431b DH |
1313 | /* |
1314 | * Bypass PCP and place fresh pages right to the tail, primarily | |
1315 | * relevant for memory onlining. | |
1316 | */ | |
0a54864f | 1317 | __free_pages_ok(page, order, FPI_TO_TAIL); |
a226f6c8 DH |
1318 | } |
1319 | ||
7cf91a98 JK |
1320 | /* |
1321 | * Check that the whole (or subset of) a pageblock given by the interval of | |
1322 | * [start_pfn, end_pfn) is valid and within the same zone, before scanning it | |
859a85dd | 1323 | * with the migration of free compaction scanner. |
7cf91a98 JK |
1324 | * |
1325 | * Return struct page pointer of start_pfn, or NULL if checks were not passed. | |
1326 | * | |
1327 | * It's possible on some configurations to have a setup like node0 node1 node0 | |
1328 | * i.e. it's possible that all pages within a zones range of pages do not | |
1329 | * belong to a single zone. We assume that a border between node0 and node1 | |
1330 | * can occur within a single pageblock, but not a node0 node1 node0 | |
1331 | * interleaving within a single pageblock. It is therefore sufficient to check | |
1332 | * the first and last page of a pageblock and avoid checking each individual | |
1333 | * page in a pageblock. | |
65f67a3e BW |
1334 | * |
1335 | * Note: the function may return non-NULL struct page even for a page block | |
1336 | * which contains a memory hole (i.e. there is no physical memory for a subset | |
5e0a760b | 1337 | * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which |
65f67a3e BW |
1338 | * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole |
1339 | * even though the start pfn is online and valid. This should be safe most of | |
1340 | * the time because struct pages are still initialized via init_unavailable_range() | |
1341 | * and pfn walkers shouldn't touch any physical memory range for which they do | |
1342 | * not recognize any specific metadata in struct pages. | |
7cf91a98 JK |
1343 | */ |
1344 | struct page *__pageblock_pfn_to_page(unsigned long start_pfn, | |
1345 | unsigned long end_pfn, struct zone *zone) | |
1346 | { | |
1347 | struct page *start_page; | |
1348 | struct page *end_page; | |
1349 | ||
1350 | /* end_pfn is one past the range we are checking */ | |
1351 | end_pfn--; | |
1352 | ||
3c4322c9 | 1353 | if (!pfn_valid(end_pfn)) |
7cf91a98 JK |
1354 | return NULL; |
1355 | ||
2d070eab MH |
1356 | start_page = pfn_to_online_page(start_pfn); |
1357 | if (!start_page) | |
1358 | return NULL; | |
7cf91a98 JK |
1359 | |
1360 | if (page_zone(start_page) != zone) | |
1361 | return NULL; | |
1362 | ||
1363 | end_page = pfn_to_page(end_pfn); | |
1364 | ||
1365 | /* This gives a shorter code than deriving page_zone(end_page) */ | |
1366 | if (page_zone_id(start_page) != page_zone_id(end_page)) | |
1367 | return NULL; | |
1368 | ||
1369 | return start_page; | |
1370 | } | |
1371 | ||
2f47a91f | 1372 | /* |
9420f89d MRI |
1373 | * The order of subdivision here is critical for the IO subsystem. |
1374 | * Please do not alter this order without good reasons and regression | |
1375 | * testing. Specifically, as large blocks of memory are subdivided, | |
1376 | * the order in which smaller blocks are delivered depends on the order | |
1377 | * they're subdivided in this function. This is the primary factor | |
1378 | * influencing the order in which pages are delivered to the IO | |
1379 | * subsystem according to empirical testing, and this is also justified | |
1380 | * by considering the behavior of a buddy system containing a single | |
1381 | * large block of memory acted on by a series of small allocations. | |
1382 | * This behavior is a critical factor in sglist merging's success. | |
80b1f41c | 1383 | * |
9420f89d | 1384 | * -- nyc |
2f47a91f | 1385 | */ |
9420f89d MRI |
1386 | static inline void expand(struct zone *zone, struct page *page, |
1387 | int low, int high, int migratetype) | |
2f47a91f | 1388 | { |
9420f89d | 1389 | unsigned long size = 1 << high; |
2f47a91f | 1390 | |
9420f89d MRI |
1391 | while (high > low) { |
1392 | high--; | |
1393 | size >>= 1; | |
1394 | VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]); | |
2f47a91f | 1395 | |
9420f89d MRI |
1396 | /* |
1397 | * Mark as guard pages (or page), that will allow to | |
1398 | * merge back to allocator when buddy will be freed. | |
1399 | * Corresponding page table entries will not be touched, | |
1400 | * pages will stay not present in virtual address space | |
1401 | */ | |
1402 | if (set_page_guard(zone, &page[size], high, migratetype)) | |
2f47a91f | 1403 | continue; |
9420f89d MRI |
1404 | |
1405 | add_to_free_list(&page[size], zone, high, migratetype); | |
1406 | set_buddy_order(&page[size], high); | |
2f47a91f | 1407 | } |
2f47a91f PT |
1408 | } |
1409 | ||
9420f89d | 1410 | static void check_new_page_bad(struct page *page) |
0e56acae | 1411 | { |
9420f89d MRI |
1412 | if (unlikely(page->flags & __PG_HWPOISON)) { |
1413 | /* Don't complain about hwpoisoned pages */ | |
1414 | page_mapcount_reset(page); /* remove PageBuddy */ | |
1415 | return; | |
0e56acae AD |
1416 | } |
1417 | ||
9420f89d MRI |
1418 | bad_page(page, |
1419 | page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP)); | |
0e56acae AD |
1420 | } |
1421 | ||
1422 | /* | |
9420f89d | 1423 | * This page is about to be returned from the page allocator |
0e56acae | 1424 | */ |
9420f89d | 1425 | static int check_new_page(struct page *page) |
0e56acae | 1426 | { |
9420f89d MRI |
1427 | if (likely(page_expected_state(page, |
1428 | PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON))) | |
1429 | return 0; | |
0e56acae | 1430 | |
9420f89d MRI |
1431 | check_new_page_bad(page); |
1432 | return 1; | |
1433 | } | |
0e56acae | 1434 | |
9420f89d MRI |
1435 | static inline bool check_new_pages(struct page *page, unsigned int order) |
1436 | { | |
1437 | if (is_check_pages_enabled()) { | |
1438 | for (int i = 0; i < (1 << order); i++) { | |
1439 | struct page *p = page + i; | |
0e56acae | 1440 | |
8666925c | 1441 | if (check_new_page(p)) |
9420f89d | 1442 | return true; |
0e56acae AD |
1443 | } |
1444 | } | |
1445 | ||
9420f89d | 1446 | return false; |
0e56acae AD |
1447 | } |
1448 | ||
9420f89d | 1449 | static inline bool should_skip_kasan_unpoison(gfp_t flags) |
e4443149 | 1450 | { |
9420f89d MRI |
1451 | /* Don't skip if a software KASAN mode is enabled. */ |
1452 | if (IS_ENABLED(CONFIG_KASAN_GENERIC) || | |
1453 | IS_ENABLED(CONFIG_KASAN_SW_TAGS)) | |
1454 | return false; | |
e4443149 | 1455 | |
9420f89d MRI |
1456 | /* Skip, if hardware tag-based KASAN is not enabled. */ |
1457 | if (!kasan_hw_tags_enabled()) | |
1458 | return true; | |
e4443149 DJ |
1459 | |
1460 | /* | |
9420f89d MRI |
1461 | * With hardware tag-based KASAN enabled, skip if this has been |
1462 | * requested via __GFP_SKIP_KASAN. | |
e4443149 | 1463 | */ |
9420f89d | 1464 | return flags & __GFP_SKIP_KASAN; |
e4443149 DJ |
1465 | } |
1466 | ||
9420f89d | 1467 | static inline bool should_skip_init(gfp_t flags) |
ecd09650 | 1468 | { |
9420f89d MRI |
1469 | /* Don't skip, if hardware tag-based KASAN is not enabled. */ |
1470 | if (!kasan_hw_tags_enabled()) | |
1471 | return false; | |
1472 | ||
1473 | /* For hardware tag-based KASAN, skip if requested. */ | |
1474 | return (flags & __GFP_SKIP_ZERO); | |
ecd09650 DJ |
1475 | } |
1476 | ||
9420f89d MRI |
1477 | inline void post_alloc_hook(struct page *page, unsigned int order, |
1478 | gfp_t gfp_flags) | |
7e18adb4 | 1479 | { |
9420f89d MRI |
1480 | bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) && |
1481 | !should_skip_init(gfp_flags); | |
1482 | bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS); | |
1483 | int i; | |
1484 | ||
1485 | set_page_private(page, 0); | |
1486 | set_page_refcounted(page); | |
0e1cc95b | 1487 | |
9420f89d MRI |
1488 | arch_alloc_page(page, order); |
1489 | debug_pagealloc_map_pages(page, 1 << order); | |
7e18adb4 | 1490 | |
3d060856 | 1491 | /* |
9420f89d MRI |
1492 | * Page unpoisoning must happen before memory initialization. |
1493 | * Otherwise, the poison pattern will be overwritten for __GFP_ZERO | |
1494 | * allocations and the page unpoisoning code will complain. | |
3d060856 | 1495 | */ |
9420f89d | 1496 | kernel_unpoison_pages(page, 1 << order); |
862b6dee | 1497 | |
1bb5eab3 AK |
1498 | /* |
1499 | * As memory initialization might be integrated into KASAN, | |
b42090ae | 1500 | * KASAN unpoisoning and memory initializion code must be |
1bb5eab3 AK |
1501 | * kept together to avoid discrepancies in behavior. |
1502 | */ | |
9294b128 AK |
1503 | |
1504 | /* | |
44383cef AK |
1505 | * If memory tags should be zeroed |
1506 | * (which happens only when memory should be initialized as well). | |
9294b128 | 1507 | */ |
44383cef | 1508 | if (zero_tags) { |
420ef683 | 1509 | /* Initialize both memory and memory tags. */ |
9294b128 AK |
1510 | for (i = 0; i != 1 << order; ++i) |
1511 | tag_clear_highpage(page + i); | |
1512 | ||
44383cef | 1513 | /* Take note that memory was initialized by the loop above. */ |
9294b128 AK |
1514 | init = false; |
1515 | } | |
0a54864f PC |
1516 | if (!should_skip_kasan_unpoison(gfp_flags) && |
1517 | kasan_unpoison_pages(page, order, init)) { | |
1518 | /* Take note that memory was initialized by KASAN. */ | |
1519 | if (kasan_has_integrated_init()) | |
1520 | init = false; | |
1521 | } else { | |
1522 | /* | |
1523 | * If memory tags have not been set by KASAN, reset the page | |
1524 | * tags to ensure page_address() dereferencing does not fault. | |
1525 | */ | |
70c248ac CM |
1526 | for (i = 0; i != 1 << order; ++i) |
1527 | page_kasan_tag_reset(page + i); | |
7a3b8353 | 1528 | } |
44383cef | 1529 | /* If memory is still not initialized, initialize it now. */ |
7e3cbba6 | 1530 | if (init) |
aeaec8e2 | 1531 | kernel_init_pages(page, 1 << order); |
1bb5eab3 AK |
1532 | |
1533 | set_page_owner(page, order, gfp_flags); | |
df4e817b | 1534 | page_table_check_alloc(page, order); |
46f24fd8 JK |
1535 | } |
1536 | ||
479f854a | 1537 | static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags, |
c603844b | 1538 | unsigned int alloc_flags) |
2a7684a2 | 1539 | { |
46f24fd8 | 1540 | post_alloc_hook(page, order, gfp_flags); |
17cf4406 | 1541 | |
17cf4406 NP |
1542 | if (order && (gfp_flags & __GFP_COMP)) |
1543 | prep_compound_page(page, order); | |
1544 | ||
75379191 | 1545 | /* |
2f064f34 | 1546 | * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to |
75379191 VB |
1547 | * allocate the page. The expectation is that the caller is taking |
1548 | * steps that will free more memory. The caller should avoid the page | |
1549 | * being used for !PFMEMALLOC purposes. | |
1550 | */ | |
2f064f34 MH |
1551 | if (alloc_flags & ALLOC_NO_WATERMARKS) |
1552 | set_page_pfmemalloc(page); | |
1553 | else | |
1554 | clear_page_pfmemalloc(page); | |
1da177e4 LT |
1555 | } |
1556 | ||
56fd56b8 MG |
1557 | /* |
1558 | * Go through the free lists for the given migratetype and remove | |
1559 | * the smallest available page from the freelists | |
1560 | */ | |
85ccc8fa | 1561 | static __always_inline |
728ec980 | 1562 | struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, |
56fd56b8 MG |
1563 | int migratetype) |
1564 | { | |
1565 | unsigned int current_order; | |
b8af2941 | 1566 | struct free_area *area; |
56fd56b8 MG |
1567 | struct page *page; |
1568 | ||
1569 | /* Find a page of the appropriate size in the preferred list */ | |
fd377218 | 1570 | for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) { |
56fd56b8 | 1571 | area = &(zone->free_area[current_order]); |
b03641af | 1572 | page = get_page_from_free_area(area, migratetype); |
a16601c5 GT |
1573 | if (!page) |
1574 | continue; | |
6ab01363 AD |
1575 | del_page_from_free_list(page, zone, current_order); |
1576 | expand(zone, page, order, current_order, migratetype); | |
bb14c2c7 | 1577 | set_pcppage_migratetype(page, migratetype); |
10e0f753 WY |
1578 | trace_mm_page_alloc_zone_locked(page, order, migratetype, |
1579 | pcp_allowed_order(order) && | |
1580 | migratetype < MIGRATE_PCPTYPES); | |
56fd56b8 MG |
1581 | return page; |
1582 | } | |
1583 | ||
1584 | return NULL; | |
1585 | } | |
1586 | ||
1587 | ||
b2a0ac88 MG |
1588 | /* |
1589 | * This array describes the order lists are fallen back to when | |
1590 | * the free lists for the desirable migrate type are depleted | |
1dd214b8 ZY |
1591 | * |
1592 | * The other migratetypes do not have fallbacks. | |
b2a0ac88 | 1593 | */ |
aa02d3c1 YD |
1594 | static int fallbacks[MIGRATE_TYPES][MIGRATE_PCPTYPES - 1] = { |
1595 | [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE }, | |
1596 | [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE }, | |
1597 | [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE }, | |
b2a0ac88 MG |
1598 | }; |
1599 | ||
dc67647b | 1600 | #ifdef CONFIG_CMA |
85ccc8fa | 1601 | static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone, |
dc67647b JK |
1602 | unsigned int order) |
1603 | { | |
1604 | return __rmqueue_smallest(zone, order, MIGRATE_CMA); | |
1605 | } | |
1606 | #else | |
1607 | static inline struct page *__rmqueue_cma_fallback(struct zone *zone, | |
1608 | unsigned int order) { return NULL; } | |
1609 | #endif | |
1610 | ||
c361be55 | 1611 | /* |
293ffa5e | 1612 | * Move the free pages in a range to the freelist tail of the requested type. |
d9c23400 | 1613 | * Note that start_page and end_pages are not aligned on a pageblock |
c361be55 MG |
1614 | * boundary. If alignment is required, use move_freepages_block() |
1615 | */ | |
02aa0cdd | 1616 | static int move_freepages(struct zone *zone, |
39ddb991 | 1617 | unsigned long start_pfn, unsigned long end_pfn, |
02aa0cdd | 1618 | int migratetype, int *num_movable) |
c361be55 MG |
1619 | { |
1620 | struct page *page; | |
39ddb991 | 1621 | unsigned long pfn; |
d00181b9 | 1622 | unsigned int order; |
d100313f | 1623 | int pages_moved = 0; |
c361be55 | 1624 | |
39ddb991 | 1625 | for (pfn = start_pfn; pfn <= end_pfn;) { |
39ddb991 | 1626 | page = pfn_to_page(pfn); |
c361be55 | 1627 | if (!PageBuddy(page)) { |
02aa0cdd VB |
1628 | /* |
1629 | * We assume that pages that could be isolated for | |
1630 | * migration are movable. But we don't actually try | |
1631 | * isolating, as that would be expensive. | |
1632 | */ | |
1633 | if (num_movable && | |
1634 | (PageLRU(page) || __PageMovable(page))) | |
1635 | (*num_movable)++; | |
39ddb991 | 1636 | pfn++; |
c361be55 MG |
1637 | continue; |
1638 | } | |
1639 | ||
cd961038 DR |
1640 | /* Make sure we are not inadvertently changing nodes */ |
1641 | VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page); | |
1642 | VM_BUG_ON_PAGE(page_zone(page) != zone, page); | |
1643 | ||
ab130f91 | 1644 | order = buddy_order(page); |
6ab01363 | 1645 | move_to_free_list(page, zone, order, migratetype); |
39ddb991 | 1646 | pfn += 1 << order; |
d100313f | 1647 | pages_moved += 1 << order; |
c361be55 MG |
1648 | } |
1649 | ||
d100313f | 1650 | return pages_moved; |
c361be55 MG |
1651 | } |
1652 | ||
ee6f509c | 1653 | int move_freepages_block(struct zone *zone, struct page *page, |
02aa0cdd | 1654 | int migratetype, int *num_movable) |
c361be55 | 1655 | { |
39ddb991 | 1656 | unsigned long start_pfn, end_pfn, pfn; |
c361be55 | 1657 | |
4a222127 DR |
1658 | if (num_movable) |
1659 | *num_movable = 0; | |
1660 | ||
39ddb991 | 1661 | pfn = page_to_pfn(page); |
4f9bc69a KW |
1662 | start_pfn = pageblock_start_pfn(pfn); |
1663 | end_pfn = pageblock_end_pfn(pfn) - 1; | |
c361be55 MG |
1664 | |
1665 | /* Do not cross zone boundaries */ | |
108bcc96 | 1666 | if (!zone_spans_pfn(zone, start_pfn)) |
39ddb991 | 1667 | start_pfn = pfn; |
108bcc96 | 1668 | if (!zone_spans_pfn(zone, end_pfn)) |
c361be55 MG |
1669 | return 0; |
1670 | ||
39ddb991 | 1671 | return move_freepages(zone, start_pfn, end_pfn, migratetype, |
02aa0cdd | 1672 | num_movable); |
c361be55 MG |
1673 | } |
1674 | ||
2f66a68f MG |
1675 | static void change_pageblock_range(struct page *pageblock_page, |
1676 | int start_order, int migratetype) | |
1677 | { | |
1678 | int nr_pageblocks = 1 << (start_order - pageblock_order); | |
1679 | ||
1680 | while (nr_pageblocks--) { | |
1681 | set_pageblock_migratetype(pageblock_page, migratetype); | |
1682 | pageblock_page += pageblock_nr_pages; | |
1683 | } | |
1684 | } | |
1685 | ||
fef903ef | 1686 | /* |
9c0415eb VB |
1687 | * When we are falling back to another migratetype during allocation, try to |
1688 | * steal extra free pages from the same pageblocks to satisfy further | |
1689 | * allocations, instead of polluting multiple pageblocks. | |
1690 | * | |
1691 | * If we are stealing a relatively large buddy page, it is likely there will | |
1692 | * be more free pages in the pageblock, so try to steal them all. For | |
1693 | * reclaimable and unmovable allocations, we steal regardless of page size, | |
1694 | * as fragmentation caused by those allocations polluting movable pageblocks | |
1695 | * is worse than movable allocations stealing from unmovable and reclaimable | |
1696 | * pageblocks. | |
fef903ef | 1697 | */ |
4eb7dce6 JK |
1698 | static bool can_steal_fallback(unsigned int order, int start_mt) |
1699 | { | |
1700 | /* | |
1701 | * Leaving this order check is intended, although there is | |
1702 | * relaxed order check in next check. The reason is that | |
1703 | * we can actually steal whole pageblock if this condition met, | |
1704 | * but, below check doesn't guarantee it and that is just heuristic | |
1705 | * so could be changed anytime. | |
1706 | */ | |
1707 | if (order >= pageblock_order) | |
1708 | return true; | |
1709 | ||
1710 | if (order >= pageblock_order / 2 || | |
1711 | start_mt == MIGRATE_RECLAIMABLE || | |
1712 | start_mt == MIGRATE_UNMOVABLE || | |
1713 | page_group_by_mobility_disabled) | |
1714 | return true; | |
1715 | ||
1716 | return false; | |
1717 | } | |
1718 | ||
597c8920 | 1719 | static inline bool boost_watermark(struct zone *zone) |
1c30844d MG |
1720 | { |
1721 | unsigned long max_boost; | |
1722 | ||
1723 | if (!watermark_boost_factor) | |
597c8920 | 1724 | return false; |
14f69140 HW |
1725 | /* |
1726 | * Don't bother in zones that are unlikely to produce results. | |
1727 | * On small machines, including kdump capture kernels running | |
1728 | * in a small area, boosting the watermark can cause an out of | |
1729 | * memory situation immediately. | |
1730 | */ | |
1731 | if ((pageblock_nr_pages * 4) > zone_managed_pages(zone)) | |
597c8920 | 1732 | return false; |
1c30844d MG |
1733 | |
1734 | max_boost = mult_frac(zone->_watermark[WMARK_HIGH], | |
1735 | watermark_boost_factor, 10000); | |
94b3334c MG |
1736 | |
1737 | /* | |
1738 | * high watermark may be uninitialised if fragmentation occurs | |
1739 | * very early in boot so do not boost. We do not fall | |
1740 | * through and boost by pageblock_nr_pages as failing | |
1741 | * allocations that early means that reclaim is not going | |
1742 | * to help and it may even be impossible to reclaim the | |
1743 | * boosted watermark resulting in a hang. | |
1744 | */ | |
1745 | if (!max_boost) | |
597c8920 | 1746 | return false; |
94b3334c | 1747 | |
1c30844d MG |
1748 | max_boost = max(pageblock_nr_pages, max_boost); |
1749 | ||
1750 | zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages, | |
1751 | max_boost); | |
597c8920 JW |
1752 | |
1753 | return true; | |
1c30844d MG |
1754 | } |
1755 | ||
4eb7dce6 JK |
1756 | /* |
1757 | * This function implements actual steal behaviour. If order is large enough, | |
1758 | * we can steal whole pageblock. If not, we first move freepages in this | |
02aa0cdd VB |
1759 | * pageblock to our migratetype and determine how many already-allocated pages |
1760 | * are there in the pageblock with a compatible migratetype. If at least half | |
1761 | * of pages are free or compatible, we can change migratetype of the pageblock | |
1762 | * itself, so pages freed in the future will be put on the correct free list. | |
4eb7dce6 JK |
1763 | */ |
1764 | static void steal_suitable_fallback(struct zone *zone, struct page *page, | |
1c30844d | 1765 | unsigned int alloc_flags, int start_type, bool whole_block) |
fef903ef | 1766 | { |
ab130f91 | 1767 | unsigned int current_order = buddy_order(page); |
02aa0cdd VB |
1768 | int free_pages, movable_pages, alike_pages; |
1769 | int old_block_type; | |
1770 | ||
1771 | old_block_type = get_pageblock_migratetype(page); | |
fef903ef | 1772 | |
3bc48f96 VB |
1773 | /* |
1774 | * This can happen due to races and we want to prevent broken | |
1775 | * highatomic accounting. | |
1776 | */ | |
02aa0cdd | 1777 | if (is_migrate_highatomic(old_block_type)) |
3bc48f96 VB |
1778 | goto single_page; |
1779 | ||
fef903ef SB |
1780 | /* Take ownership for orders >= pageblock_order */ |
1781 | if (current_order >= pageblock_order) { | |
1782 | change_pageblock_range(page, current_order, start_type); | |
3bc48f96 | 1783 | goto single_page; |
fef903ef SB |
1784 | } |
1785 | ||
1c30844d MG |
1786 | /* |
1787 | * Boost watermarks to increase reclaim pressure to reduce the | |
1788 | * likelihood of future fallbacks. Wake kswapd now as the node | |
1789 | * may be balanced overall and kswapd will not wake naturally. | |
1790 | */ | |
597c8920 | 1791 | if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD)) |
73444bc4 | 1792 | set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags); |
1c30844d | 1793 | |
3bc48f96 VB |
1794 | /* We are not allowed to try stealing from the whole block */ |
1795 | if (!whole_block) | |
1796 | goto single_page; | |
1797 | ||
02aa0cdd VB |
1798 | free_pages = move_freepages_block(zone, page, start_type, |
1799 | &movable_pages); | |
ebddd111 ML |
1800 | /* moving whole block can fail due to zone boundary conditions */ |
1801 | if (!free_pages) | |
1802 | goto single_page; | |
1803 | ||
02aa0cdd VB |
1804 | /* |
1805 | * Determine how many pages are compatible with our allocation. | |
1806 | * For movable allocation, it's the number of movable pages which | |
1807 | * we just obtained. For other types it's a bit more tricky. | |
1808 | */ | |
1809 | if (start_type == MIGRATE_MOVABLE) { | |
1810 | alike_pages = movable_pages; | |
1811 | } else { | |
1812 | /* | |
1813 | * If we are falling back a RECLAIMABLE or UNMOVABLE allocation | |
1814 | * to MOVABLE pageblock, consider all non-movable pages as | |
1815 | * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or | |
1816 | * vice versa, be conservative since we can't distinguish the | |
1817 | * exact migratetype of non-movable pages. | |
1818 | */ | |
1819 | if (old_block_type == MIGRATE_MOVABLE) | |
1820 | alike_pages = pageblock_nr_pages | |
1821 | - (free_pages + movable_pages); | |
1822 | else | |
1823 | alike_pages = 0; | |
1824 | } | |
02aa0cdd VB |
1825 | /* |
1826 | * If a sufficient number of pages in the block are either free or of | |
ebddd111 | 1827 | * compatible migratability as our allocation, claim the whole block. |
02aa0cdd VB |
1828 | */ |
1829 | if (free_pages + alike_pages >= (1 << (pageblock_order-1)) || | |
4eb7dce6 JK |
1830 | page_group_by_mobility_disabled) |
1831 | set_pageblock_migratetype(page, start_type); | |
3bc48f96 VB |
1832 | |
1833 | return; | |
1834 | ||
1835 | single_page: | |
6ab01363 | 1836 | move_to_free_list(page, zone, current_order, start_type); |
4eb7dce6 JK |
1837 | } |
1838 | ||
2149cdae JK |
1839 | /* |
1840 | * Check whether there is a suitable fallback freepage with requested order. | |
1841 | * If only_stealable is true, this function returns fallback_mt only if | |
1842 | * we can steal other freepages all together. This would help to reduce | |
1843 | * fragmentation due to mixed migratetype pages in one pageblock. | |
1844 | */ | |
1845 | int find_suitable_fallback(struct free_area *area, unsigned int order, | |
1846 | int migratetype, bool only_stealable, bool *can_steal) | |
4eb7dce6 JK |
1847 | { |
1848 | int i; | |
1849 | int fallback_mt; | |
1850 | ||
1851 | if (area->nr_free == 0) | |
1852 | return -1; | |
1853 | ||
1854 | *can_steal = false; | |
aa02d3c1 | 1855 | for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) { |
4eb7dce6 | 1856 | fallback_mt = fallbacks[migratetype][i]; |
b03641af | 1857 | if (free_area_empty(area, fallback_mt)) |
4eb7dce6 | 1858 | continue; |
fef903ef | 1859 | |
4eb7dce6 JK |
1860 | if (can_steal_fallback(order, migratetype)) |
1861 | *can_steal = true; | |
1862 | ||
2149cdae JK |
1863 | if (!only_stealable) |
1864 | return fallback_mt; | |
1865 | ||
1866 | if (*can_steal) | |
1867 | return fallback_mt; | |
fef903ef | 1868 | } |
4eb7dce6 JK |
1869 | |
1870 | return -1; | |
fef903ef SB |
1871 | } |
1872 | ||
0aaa29a5 MG |
1873 | /* |
1874 | * Reserve a pageblock for exclusive use of high-order atomic allocations if | |
1875 | * there are no empty page blocks that contain a page with a suitable order | |
1876 | */ | |
368d983b | 1877 | static void reserve_highatomic_pageblock(struct page *page, struct zone *zone) |
0aaa29a5 MG |
1878 | { |
1879 | int mt; | |
1880 | unsigned long max_managed, flags; | |
1881 | ||
1882 | /* | |
d68e39fc | 1883 | * The number reserved as: minimum is 1 pageblock, maximum is |
9cd20f3f CTK |
1884 | * roughly 1% of a zone. But if 1% of a zone falls below a |
1885 | * pageblock size, then don't reserve any pageblocks. | |
0aaa29a5 MG |
1886 | * Check is race-prone but harmless. |
1887 | */ | |
9cd20f3f CTK |
1888 | if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages) |
1889 | return; | |
d68e39fc | 1890 | max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages); |
0aaa29a5 MG |
1891 | if (zone->nr_reserved_highatomic >= max_managed) |
1892 | return; | |
1893 | ||
1894 | spin_lock_irqsave(&zone->lock, flags); | |
1895 | ||
1896 | /* Recheck the nr_reserved_highatomic limit under the lock */ | |
1897 | if (zone->nr_reserved_highatomic >= max_managed) | |
1898 | goto out_unlock; | |
1899 | ||
1900 | /* Yoink! */ | |
1901 | mt = get_pageblock_migratetype(page); | |
1dd214b8 ZY |
1902 | /* Only reserve normal pageblocks (i.e., they can merge with others) */ |
1903 | if (migratetype_is_mergeable(mt)) { | |
0aaa29a5 MG |
1904 | zone->nr_reserved_highatomic += pageblock_nr_pages; |
1905 | set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC); | |
02aa0cdd | 1906 | move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL); |
0aaa29a5 MG |
1907 | } |
1908 | ||
1909 | out_unlock: | |
1910 | spin_unlock_irqrestore(&zone->lock, flags); | |
1911 | } | |
1912 | ||
1913 | /* | |
1914 | * Used when an allocation is about to fail under memory pressure. This | |
1915 | * potentially hurts the reliability of high-order allocations when under | |
1916 | * intense memory pressure but failed atomic allocations should be easier | |
1917 | * to recover from than an OOM. | |
29fac03b MK |
1918 | * |
1919 | * If @force is true, try to unreserve a pageblock even though highatomic | |
1920 | * pageblock is exhausted. | |
0aaa29a5 | 1921 | */ |
29fac03b MK |
1922 | static bool unreserve_highatomic_pageblock(const struct alloc_context *ac, |
1923 | bool force) | |
0aaa29a5 MG |
1924 | { |
1925 | struct zonelist *zonelist = ac->zonelist; | |
1926 | unsigned long flags; | |
1927 | struct zoneref *z; | |
1928 | struct zone *zone; | |
1929 | struct page *page; | |
1930 | int order; | |
04c8716f | 1931 | bool ret; |
0aaa29a5 | 1932 | |
97a225e6 | 1933 | for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx, |
0aaa29a5 | 1934 | ac->nodemask) { |
29fac03b MK |
1935 | /* |
1936 | * Preserve at least one pageblock unless memory pressure | |
1937 | * is really high. | |
1938 | */ | |
1939 | if (!force && zone->nr_reserved_highatomic <= | |
1940 | pageblock_nr_pages) | |
0aaa29a5 MG |
1941 | continue; |
1942 | ||
1943 | spin_lock_irqsave(&zone->lock, flags); | |
fd377218 | 1944 | for (order = 0; order < NR_PAGE_ORDERS; order++) { |
0aaa29a5 MG |
1945 | struct free_area *area = &(zone->free_area[order]); |
1946 | ||
b03641af | 1947 | page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC); |
a16601c5 | 1948 | if (!page) |
0aaa29a5 MG |
1949 | continue; |
1950 | ||
0aaa29a5 | 1951 | /* |
4855e4a7 MK |
1952 | * In page freeing path, migratetype change is racy so |
1953 | * we can counter several free pages in a pageblock | |
f0953a1b | 1954 | * in this loop although we changed the pageblock type |
4855e4a7 MK |
1955 | * from highatomic to ac->migratetype. So we should |
1956 | * adjust the count once. | |
0aaa29a5 | 1957 | */ |
a6ffdc07 | 1958 | if (is_migrate_highatomic_page(page)) { |
4855e4a7 MK |
1959 | /* |
1960 | * It should never happen but changes to | |
1961 | * locking could inadvertently allow a per-cpu | |
1962 | * drain to add pages to MIGRATE_HIGHATOMIC | |
1963 | * while unreserving so be safe and watch for | |
1964 | * underflows. | |
1965 | */ | |
1966 | zone->nr_reserved_highatomic -= min( | |
1967 | pageblock_nr_pages, | |
1968 | zone->nr_reserved_highatomic); | |
1969 | } | |
0aaa29a5 MG |
1970 | |
1971 | /* | |
1972 | * Convert to ac->migratetype and avoid the normal | |
1973 | * pageblock stealing heuristics. Minimally, the caller | |
1974 | * is doing the work and needs the pages. More | |
1975 | * importantly, if the block was always converted to | |
1976 | * MIGRATE_UNMOVABLE or another type then the number | |
1977 | * of pageblocks that cannot be completely freed | |
1978 | * may increase. | |
1979 | */ | |
1980 | set_pageblock_migratetype(page, ac->migratetype); | |
02aa0cdd VB |
1981 | ret = move_freepages_block(zone, page, ac->migratetype, |
1982 | NULL); | |
29fac03b MK |
1983 | if (ret) { |
1984 | spin_unlock_irqrestore(&zone->lock, flags); | |
1985 | return ret; | |
1986 | } | |
0aaa29a5 MG |
1987 | } |
1988 | spin_unlock_irqrestore(&zone->lock, flags); | |
1989 | } | |
04c8716f MK |
1990 | |
1991 | return false; | |
0aaa29a5 MG |
1992 | } |
1993 | ||
3bc48f96 VB |
1994 | /* |
1995 | * Try finding a free buddy page on the fallback list and put it on the free | |
1996 | * list of requested migratetype, possibly along with other pages from the same | |
1997 | * block, depending on fragmentation avoidance heuristics. Returns true if | |
1998 | * fallback was found so that __rmqueue_smallest() can grab it. | |
b002529d RV |
1999 | * |
2000 | * The use of signed ints for order and current_order is a deliberate | |
2001 | * deviation from the rest of this file, to make the for loop | |
2002 | * condition simpler. | |
3bc48f96 | 2003 | */ |
85ccc8fa | 2004 | static __always_inline bool |
6bb15450 MG |
2005 | __rmqueue_fallback(struct zone *zone, int order, int start_migratetype, |
2006 | unsigned int alloc_flags) | |
b2a0ac88 | 2007 | { |
b8af2941 | 2008 | struct free_area *area; |
b002529d | 2009 | int current_order; |
6bb15450 | 2010 | int min_order = order; |
b2a0ac88 | 2011 | struct page *page; |
4eb7dce6 JK |
2012 | int fallback_mt; |
2013 | bool can_steal; | |
b2a0ac88 | 2014 | |
6bb15450 MG |
2015 | /* |
2016 | * Do not steal pages from freelists belonging to other pageblocks | |
2017 | * i.e. orders < pageblock_order. If there are no local zones free, | |
2018 | * the zonelists will be reiterated without ALLOC_NOFRAGMENT. | |
2019 | */ | |
e933dc4a | 2020 | if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT) |
6bb15450 MG |
2021 | min_order = pageblock_order; |
2022 | ||
7a8f58f3 VB |
2023 | /* |
2024 | * Find the largest available free page in the other list. This roughly | |
2025 | * approximates finding the pageblock with the most free pages, which | |
2026 | * would be too costly to do exactly. | |
2027 | */ | |
5e0a760b | 2028 | for (current_order = MAX_PAGE_ORDER; current_order >= min_order; |
7aeb09f9 | 2029 | --current_order) { |
4eb7dce6 JK |
2030 | area = &(zone->free_area[current_order]); |
2031 | fallback_mt = find_suitable_fallback(area, current_order, | |
2149cdae | 2032 | start_migratetype, false, &can_steal); |
4eb7dce6 JK |
2033 | if (fallback_mt == -1) |
2034 | continue; | |
b2a0ac88 | 2035 | |
7a8f58f3 VB |
2036 | /* |
2037 | * We cannot steal all free pages from the pageblock and the | |
2038 | * requested migratetype is movable. In that case it's better to | |
2039 | * steal and split the smallest available page instead of the | |
2040 | * largest available page, because even if the next movable | |
2041 | * allocation falls back into a different pageblock than this | |
2042 | * one, it won't cause permanent fragmentation. | |
2043 | */ | |
2044 | if (!can_steal && start_migratetype == MIGRATE_MOVABLE | |
2045 | && current_order > order) | |
2046 | goto find_smallest; | |
b2a0ac88 | 2047 | |
7a8f58f3 VB |
2048 | goto do_steal; |
2049 | } | |
e0fff1bd | 2050 | |
7a8f58f3 | 2051 | return false; |
e0fff1bd | 2052 | |
7a8f58f3 | 2053 | find_smallest: |
fd377218 | 2054 | for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) { |
7a8f58f3 VB |
2055 | area = &(zone->free_area[current_order]); |
2056 | fallback_mt = find_suitable_fallback(area, current_order, | |
2057 | start_migratetype, false, &can_steal); | |
2058 | if (fallback_mt != -1) | |
2059 | break; | |
b2a0ac88 MG |
2060 | } |
2061 | ||
7a8f58f3 VB |
2062 | /* |
2063 | * This should not happen - we already found a suitable fallback | |
2064 | * when looking for the largest page. | |
2065 | */ | |
5e0a760b | 2066 | VM_BUG_ON(current_order > MAX_PAGE_ORDER); |
7a8f58f3 VB |
2067 | |
2068 | do_steal: | |
b03641af | 2069 | page = get_page_from_free_area(area, fallback_mt); |
7a8f58f3 | 2070 | |
1c30844d MG |
2071 | steal_suitable_fallback(zone, page, alloc_flags, start_migratetype, |
2072 | can_steal); | |
7a8f58f3 VB |
2073 | |
2074 | trace_mm_page_alloc_extfrag(page, order, current_order, | |
2075 | start_migratetype, fallback_mt); | |
2076 | ||
2077 | return true; | |
2078 | ||
b2a0ac88 MG |
2079 | } |
2080 | ||
56fd56b8 | 2081 | /* |
1da177e4 LT |
2082 | * Do the hard work of removing an element from the buddy allocator. |
2083 | * Call me with the zone->lock already held. | |
2084 | */ | |
85ccc8fa | 2085 | static __always_inline struct page * |
6bb15450 MG |
2086 | __rmqueue(struct zone *zone, unsigned int order, int migratetype, |
2087 | unsigned int alloc_flags) | |
1da177e4 | 2088 | { |
1da177e4 LT |
2089 | struct page *page; |
2090 | ||
ce8f86ee H |
2091 | if (IS_ENABLED(CONFIG_CMA)) { |
2092 | /* | |
2093 | * Balance movable allocations between regular and CMA areas by | |
2094 | * allocating from CMA when over half of the zone's free memory | |
2095 | * is in the CMA area. | |
2096 | */ | |
2097 | if (alloc_flags & ALLOC_CMA && | |
2098 | zone_page_state(zone, NR_FREE_CMA_PAGES) > | |
2099 | zone_page_state(zone, NR_FREE_PAGES) / 2) { | |
2100 | page = __rmqueue_cma_fallback(zone, order); | |
2101 | if (page) | |
10e0f753 | 2102 | return page; |
ce8f86ee | 2103 | } |
16867664 | 2104 | } |
3bc48f96 | 2105 | retry: |
56fd56b8 | 2106 | page = __rmqueue_smallest(zone, order, migratetype); |
974a786e | 2107 | if (unlikely(!page)) { |
8510e69c | 2108 | if (alloc_flags & ALLOC_CMA) |
dc67647b JK |
2109 | page = __rmqueue_cma_fallback(zone, order); |
2110 | ||
6bb15450 MG |
2111 | if (!page && __rmqueue_fallback(zone, order, migratetype, |
2112 | alloc_flags)) | |
3bc48f96 | 2113 | goto retry; |
728ec980 | 2114 | } |
b2a0ac88 | 2115 | return page; |
1da177e4 LT |
2116 | } |
2117 | ||
5f63b720 | 2118 | /* |
1da177e4 LT |
2119 | * Obtain a specified number of elements from the buddy allocator, all under |
2120 | * a single hold of the lock, for efficiency. Add them to the supplied list. | |
2121 | * Returns the number of new pages which were placed at *list. | |
2122 | */ | |
5f63b720 | 2123 | static int rmqueue_bulk(struct zone *zone, unsigned int order, |
b2a0ac88 | 2124 | unsigned long count, struct list_head *list, |
6bb15450 | 2125 | int migratetype, unsigned int alloc_flags) |
1da177e4 | 2126 | { |
57490774 | 2127 | unsigned long flags; |
700d2e9a | 2128 | int i; |
5f63b720 | 2129 | |
57490774 | 2130 | spin_lock_irqsave(&zone->lock, flags); |
1da177e4 | 2131 | for (i = 0; i < count; ++i) { |
6bb15450 MG |
2132 | struct page *page = __rmqueue(zone, order, migratetype, |
2133 | alloc_flags); | |
085cc7d5 | 2134 | if (unlikely(page == NULL)) |
1da177e4 | 2135 | break; |
81eabcbe MG |
2136 | |
2137 | /* | |
0fac3ba5 VB |
2138 | * Split buddy pages returned by expand() are received here in |
2139 | * physical page order. The page is added to the tail of | |
2140 | * caller's list. From the callers perspective, the linked list | |
2141 | * is ordered by page number under some conditions. This is | |
2142 | * useful for IO devices that can forward direction from the | |
2143 | * head, thus also in the physical page order. This is useful | |
2144 | * for IO devices that can merge IO requests if the physical | |
2145 | * pages are ordered properly. | |
81eabcbe | 2146 | */ |
bf75f200 | 2147 | list_add_tail(&page->pcp_list, list); |
bb14c2c7 | 2148 | if (is_migrate_cma(get_pcppage_migratetype(page))) |
d1ce749a BZ |
2149 | __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, |
2150 | -(1 << order)); | |
1da177e4 | 2151 | } |
a6de734b | 2152 | |
f2260e6b | 2153 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); |
57490774 | 2154 | spin_unlock_irqrestore(&zone->lock, flags); |
2ede3c13 | 2155 | |
700d2e9a | 2156 | return i; |
1da177e4 LT |
2157 | } |
2158 | ||
51a755c5 HY |
2159 | /* |
2160 | * Called from the vmstat counter updater to decay the PCP high. | |
2161 | * Return whether there are addition works to do. | |
2162 | */ | |
2163 | int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp) | |
2164 | { | |
2165 | int high_min, to_drain, batch; | |
2166 | int todo = 0; | |
2167 | ||
2168 | high_min = READ_ONCE(pcp->high_min); | |
2169 | batch = READ_ONCE(pcp->batch); | |
2170 | /* | |
2171 | * Decrease pcp->high periodically to try to free possible | |
2172 | * idle PCP pages. And, avoid to free too many pages to | |
2173 | * control latency. This caps pcp->high decrement too. | |
2174 | */ | |
2175 | if (pcp->high > high_min) { | |
2176 | pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX), | |
2177 | pcp->high - (pcp->high >> 3), high_min); | |
2178 | if (pcp->high > high_min) | |
2179 | todo++; | |
2180 | } | |
2181 | ||
2182 | to_drain = pcp->count - pcp->high; | |
2183 | if (to_drain > 0) { | |
2184 | spin_lock(&pcp->lock); | |
2185 | free_pcppages_bulk(zone, to_drain, pcp, 0); | |
2186 | spin_unlock(&pcp->lock); | |
2187 | todo++; | |
2188 | } | |
2189 | ||
2190 | return todo; | |
2191 | } | |
2192 | ||
4ae7c039 | 2193 | #ifdef CONFIG_NUMA |
8fce4d8e | 2194 | /* |
4037d452 CL |
2195 | * Called from the vmstat counter updater to drain pagesets of this |
2196 | * currently executing processor on remote nodes after they have | |
2197 | * expired. | |
8fce4d8e | 2198 | */ |
4037d452 | 2199 | void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) |
4ae7c039 | 2200 | { |
7be12fc9 | 2201 | int to_drain, batch; |
4ae7c039 | 2202 | |
4db0c3c2 | 2203 | batch = READ_ONCE(pcp->batch); |
7be12fc9 | 2204 | to_drain = min(pcp->count, batch); |
4b23a68f | 2205 | if (to_drain > 0) { |
57490774 | 2206 | spin_lock(&pcp->lock); |
fd56eef2 | 2207 | free_pcppages_bulk(zone, to_drain, pcp, 0); |
57490774 | 2208 | spin_unlock(&pcp->lock); |
4b23a68f | 2209 | } |
4ae7c039 CL |
2210 | } |
2211 | #endif | |
2212 | ||
9f8f2172 | 2213 | /* |
93481ff0 | 2214 | * Drain pcplists of the indicated processor and zone. |
9f8f2172 | 2215 | */ |
93481ff0 | 2216 | static void drain_pages_zone(unsigned int cpu, struct zone *zone) |
1da177e4 | 2217 | { |
93481ff0 | 2218 | struct per_cpu_pages *pcp; |
1da177e4 | 2219 | |
28f836b6 | 2220 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); |
4b23a68f | 2221 | if (pcp->count) { |
57490774 | 2222 | spin_lock(&pcp->lock); |
4b23a68f | 2223 | free_pcppages_bulk(zone, pcp->count, pcp, 0); |
57490774 | 2224 | spin_unlock(&pcp->lock); |
4b23a68f | 2225 | } |
93481ff0 | 2226 | } |
3dfa5721 | 2227 | |
93481ff0 VB |
2228 | /* |
2229 | * Drain pcplists of all zones on the indicated processor. | |
93481ff0 VB |
2230 | */ |
2231 | static void drain_pages(unsigned int cpu) | |
2232 | { | |
2233 | struct zone *zone; | |
2234 | ||
2235 | for_each_populated_zone(zone) { | |
2236 | drain_pages_zone(cpu, zone); | |
1da177e4 LT |
2237 | } |
2238 | } | |
1da177e4 | 2239 | |
9f8f2172 CL |
2240 | /* |
2241 | * Spill all of this CPU's per-cpu pages back into the buddy allocator. | |
2242 | */ | |
93481ff0 | 2243 | void drain_local_pages(struct zone *zone) |
9f8f2172 | 2244 | { |
93481ff0 VB |
2245 | int cpu = smp_processor_id(); |
2246 | ||
2247 | if (zone) | |
2248 | drain_pages_zone(cpu, zone); | |
2249 | else | |
2250 | drain_pages(cpu); | |
9f8f2172 CL |
2251 | } |
2252 | ||
2253 | /* | |
ec6e8c7e VB |
2254 | * The implementation of drain_all_pages(), exposing an extra parameter to |
2255 | * drain on all cpus. | |
93481ff0 | 2256 | * |
ec6e8c7e VB |
2257 | * drain_all_pages() is optimized to only execute on cpus where pcplists are |
2258 | * not empty. The check for non-emptiness can however race with a free to | |
2259 | * pcplist that has not yet increased the pcp->count from 0 to 1. Callers | |
2260 | * that need the guarantee that every CPU has drained can disable the | |
2261 | * optimizing racy check. | |
9f8f2172 | 2262 | */ |
3b1f3658 | 2263 | static void __drain_all_pages(struct zone *zone, bool force_all_cpus) |
9f8f2172 | 2264 | { |
74046494 | 2265 | int cpu; |
74046494 GBY |
2266 | |
2267 | /* | |
041711ce | 2268 | * Allocate in the BSS so we won't require allocation in |
74046494 GBY |
2269 | * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y |
2270 | */ | |
2271 | static cpumask_t cpus_with_pcps; | |
2272 | ||
bd233f53 MG |
2273 | /* |
2274 | * Do not drain if one is already in progress unless it's specific to | |
2275 | * a zone. Such callers are primarily CMA and memory hotplug and need | |
2276 | * the drain to be complete when the call returns. | |
2277 | */ | |
2278 | if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) { | |
2279 | if (!zone) | |
2280 | return; | |
2281 | mutex_lock(&pcpu_drain_mutex); | |
2282 | } | |
0ccce3b9 | 2283 | |
74046494 GBY |
2284 | /* |
2285 | * We don't care about racing with CPU hotplug event | |
2286 | * as offline notification will cause the notified | |
2287 | * cpu to drain that CPU pcps and on_each_cpu_mask | |
2288 | * disables preemption as part of its processing | |
2289 | */ | |
2290 | for_each_online_cpu(cpu) { | |
28f836b6 | 2291 | struct per_cpu_pages *pcp; |
93481ff0 | 2292 | struct zone *z; |
74046494 | 2293 | bool has_pcps = false; |
93481ff0 | 2294 | |
ec6e8c7e VB |
2295 | if (force_all_cpus) { |
2296 | /* | |
2297 | * The pcp.count check is racy, some callers need a | |
2298 | * guarantee that no cpu is missed. | |
2299 | */ | |
2300 | has_pcps = true; | |
2301 | } else if (zone) { | |
28f836b6 MG |
2302 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); |
2303 | if (pcp->count) | |
74046494 | 2304 | has_pcps = true; |
93481ff0 VB |
2305 | } else { |
2306 | for_each_populated_zone(z) { | |
28f836b6 MG |
2307 | pcp = per_cpu_ptr(z->per_cpu_pageset, cpu); |
2308 | if (pcp->count) { | |
93481ff0 VB |
2309 | has_pcps = true; |
2310 | break; | |
2311 | } | |
74046494 GBY |
2312 | } |
2313 | } | |
93481ff0 | 2314 | |
74046494 GBY |
2315 | if (has_pcps) |
2316 | cpumask_set_cpu(cpu, &cpus_with_pcps); | |
2317 | else | |
2318 | cpumask_clear_cpu(cpu, &cpus_with_pcps); | |
2319 | } | |
0ccce3b9 | 2320 | |
bd233f53 | 2321 | for_each_cpu(cpu, &cpus_with_pcps) { |
443c2acc NSJ |
2322 | if (zone) |
2323 | drain_pages_zone(cpu, zone); | |
2324 | else | |
2325 | drain_pages(cpu); | |
0ccce3b9 | 2326 | } |
bd233f53 MG |
2327 | |
2328 | mutex_unlock(&pcpu_drain_mutex); | |
9f8f2172 CL |
2329 | } |
2330 | ||
ec6e8c7e VB |
2331 | /* |
2332 | * Spill all the per-cpu pages from all CPUs back into the buddy allocator. | |
2333 | * | |
2334 | * When zone parameter is non-NULL, spill just the single zone's pages. | |
ec6e8c7e VB |
2335 | */ |
2336 | void drain_all_pages(struct zone *zone) | |
2337 | { | |
2338 | __drain_all_pages(zone, false); | |
2339 | } | |
2340 | ||
44042b44 MG |
2341 | static bool free_unref_page_prepare(struct page *page, unsigned long pfn, |
2342 | unsigned int order) | |
1da177e4 | 2343 | { |
5f8dcc21 | 2344 | int migratetype; |
1da177e4 | 2345 | |
700d2e9a | 2346 | if (!free_pages_prepare(page, order, FPI_NONE)) |
9cca35d4 | 2347 | return false; |
689bcebf | 2348 | |
dc4b0caf | 2349 | migratetype = get_pfnblock_migratetype(page, pfn); |
bb14c2c7 | 2350 | set_pcppage_migratetype(page, migratetype); |
9cca35d4 MG |
2351 | return true; |
2352 | } | |
2353 | ||
51a755c5 | 2354 | static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high) |
3b12e7e9 MG |
2355 | { |
2356 | int min_nr_free, max_nr_free; | |
2357 | ||
51a755c5 | 2358 | /* Free as much as possible if batch freeing high-order pages. */ |
f26b3fa0 | 2359 | if (unlikely(free_high)) |
51a755c5 | 2360 | return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX); |
f26b3fa0 | 2361 | |
3b12e7e9 MG |
2362 | /* Check for PCP disabled or boot pageset */ |
2363 | if (unlikely(high < batch)) | |
2364 | return 1; | |
2365 | ||
2366 | /* Leave at least pcp->batch pages on the list */ | |
2367 | min_nr_free = batch; | |
2368 | max_nr_free = high - batch; | |
2369 | ||
2370 | /* | |
6ccdcb6d HY |
2371 | * Increase the batch number to the number of the consecutive |
2372 | * freed pages to reduce zone lock contention. | |
3b12e7e9 | 2373 | */ |
6ccdcb6d | 2374 | batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free); |
3b12e7e9 MG |
2375 | |
2376 | return batch; | |
2377 | } | |
2378 | ||
f26b3fa0 | 2379 | static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone, |
51a755c5 | 2380 | int batch, bool free_high) |
c49c2c47 | 2381 | { |
51a755c5 | 2382 | int high, high_min, high_max; |
c49c2c47 | 2383 | |
51a755c5 HY |
2384 | high_min = READ_ONCE(pcp->high_min); |
2385 | high_max = READ_ONCE(pcp->high_max); | |
2386 | high = pcp->high = clamp(pcp->high, high_min, high_max); | |
2387 | ||
2388 | if (unlikely(!high)) | |
c49c2c47 MG |
2389 | return 0; |
2390 | ||
51a755c5 HY |
2391 | if (unlikely(free_high)) { |
2392 | pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX), | |
2393 | high_min); | |
2394 | return 0; | |
2395 | } | |
c49c2c47 MG |
2396 | |
2397 | /* | |
2398 | * If reclaim is active, limit the number of pages that can be | |
2399 | * stored on pcp lists | |
2400 | */ | |
51a755c5 | 2401 | if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) { |
6ccdcb6d HY |
2402 | int free_count = max_t(int, pcp->free_count, batch); |
2403 | ||
2404 | pcp->high = max(high - free_count, high_min); | |
51a755c5 HY |
2405 | return min(batch << 2, pcp->high); |
2406 | } | |
2407 | ||
57c0419c HY |
2408 | if (high_min == high_max) |
2409 | return high; | |
2410 | ||
2411 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) { | |
6ccdcb6d HY |
2412 | int free_count = max_t(int, pcp->free_count, batch); |
2413 | ||
2414 | pcp->high = max(high - free_count, high_min); | |
57c0419c HY |
2415 | high = max(pcp->count, high_min); |
2416 | } else if (pcp->count >= high) { | |
6ccdcb6d | 2417 | int need_high = pcp->free_count + batch; |
51a755c5 HY |
2418 | |
2419 | /* pcp->high should be large enough to hold batch freed pages */ | |
2420 | if (pcp->high < need_high) | |
2421 | pcp->high = clamp(need_high, high_min, high_max); | |
2422 | } | |
2423 | ||
2424 | return high; | |
c49c2c47 MG |
2425 | } |
2426 | ||
4b23a68f MG |
2427 | static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp, |
2428 | struct page *page, int migratetype, | |
56651377 | 2429 | unsigned int order) |
9cca35d4 | 2430 | { |
51a755c5 | 2431 | int high, batch; |
44042b44 | 2432 | int pindex; |
ca71fe1a | 2433 | bool free_high = false; |
9cca35d4 | 2434 | |
c0a24239 HY |
2435 | /* |
2436 | * On freeing, reduce the number of pages that are batch allocated. | |
2437 | * See nr_pcp_alloc() where alloc_factor is increased for subsequent | |
2438 | * allocations. | |
2439 | */ | |
2440 | pcp->alloc_factor >>= 1; | |
15cd9004 | 2441 | __count_vm_events(PGFREE, 1 << order); |
44042b44 | 2442 | pindex = order_to_pindex(migratetype, order); |
bf75f200 | 2443 | list_add(&page->pcp_list, &pcp->lists[pindex]); |
44042b44 | 2444 | pcp->count += 1 << order; |
f26b3fa0 | 2445 | |
51a755c5 | 2446 | batch = READ_ONCE(pcp->batch); |
f26b3fa0 MG |
2447 | /* |
2448 | * As high-order pages other than THP's stored on PCP can contribute | |
2449 | * to fragmentation, limit the number stored when PCP is heavily | |
2450 | * freeing without allocation. The remainder after bulk freeing | |
2451 | * stops will be drained from vmstat refresh context. | |
2452 | */ | |
ca71fe1a | 2453 | if (order && order <= PAGE_ALLOC_COSTLY_ORDER) { |
6ccdcb6d | 2454 | free_high = (pcp->free_count >= batch && |
362d37a1 HY |
2455 | (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) && |
2456 | (!(pcp->flags & PCPF_FREE_HIGH_BATCH) || | |
51a755c5 | 2457 | pcp->count >= READ_ONCE(batch))); |
ca71fe1a HY |
2458 | pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER; |
2459 | } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) { | |
2460 | pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER; | |
2461 | } | |
6ccdcb6d HY |
2462 | if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX)) |
2463 | pcp->free_count += (1 << order); | |
51a755c5 | 2464 | high = nr_pcp_high(pcp, zone, batch, free_high); |
3b12e7e9 | 2465 | if (pcp->count >= high) { |
51a755c5 HY |
2466 | free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high), |
2467 | pcp, pindex); | |
57c0419c HY |
2468 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags) && |
2469 | zone_watermark_ok(zone, 0, high_wmark_pages(zone), | |
2470 | ZONE_MOVABLE, 0)) | |
2471 | clear_bit(ZONE_BELOW_HIGH, &zone->flags); | |
3b12e7e9 | 2472 | } |
9cca35d4 | 2473 | } |
5f8dcc21 | 2474 | |
9cca35d4 | 2475 | /* |
44042b44 | 2476 | * Free a pcp page |
9cca35d4 | 2477 | */ |
44042b44 | 2478 | void free_unref_page(struct page *page, unsigned int order) |
9cca35d4 | 2479 | { |
4b23a68f MG |
2480 | unsigned long __maybe_unused UP_flags; |
2481 | struct per_cpu_pages *pcp; | |
2482 | struct zone *zone; | |
9cca35d4 | 2483 | unsigned long pfn = page_to_pfn(page); |
7b086755 | 2484 | int migratetype, pcpmigratetype; |
9cca35d4 | 2485 | |
44042b44 | 2486 | if (!free_unref_page_prepare(page, pfn, order)) |
9cca35d4 | 2487 | return; |
da456f14 | 2488 | |
5f8dcc21 MG |
2489 | /* |
2490 | * We only track unmovable, reclaimable and movable on pcp lists. | |
df1acc85 | 2491 | * Place ISOLATE pages on the isolated list because they are being |
7b086755 JW |
2492 | * offlined but treat HIGHATOMIC and CMA as movable pages so we can |
2493 | * get those areas back if necessary. Otherwise, we may have to free | |
5f8dcc21 MG |
2494 | * excessively into the page allocator |
2495 | */ | |
7b086755 | 2496 | migratetype = pcpmigratetype = get_pcppage_migratetype(page); |
df1acc85 | 2497 | if (unlikely(migratetype >= MIGRATE_PCPTYPES)) { |
194159fb | 2498 | if (unlikely(is_migrate_isolate(migratetype))) { |
44042b44 | 2499 | free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE); |
9cca35d4 | 2500 | return; |
5f8dcc21 | 2501 | } |
7b086755 | 2502 | pcpmigratetype = MIGRATE_MOVABLE; |
5f8dcc21 MG |
2503 | } |
2504 | ||
4b23a68f MG |
2505 | zone = page_zone(page); |
2506 | pcp_trylock_prepare(UP_flags); | |
57490774 | 2507 | pcp = pcp_spin_trylock(zone->per_cpu_pageset); |
01b44456 | 2508 | if (pcp) { |
7b086755 | 2509 | free_unref_page_commit(zone, pcp, page, pcpmigratetype, order); |
57490774 | 2510 | pcp_spin_unlock(pcp); |
4b23a68f MG |
2511 | } else { |
2512 | free_one_page(zone, page, pfn, order, migratetype, FPI_NONE); | |
2513 | } | |
2514 | pcp_trylock_finish(UP_flags); | |
1da177e4 LT |
2515 | } |
2516 | ||
cc59850e KK |
2517 | /* |
2518 | * Free a list of 0-order pages | |
2519 | */ | |
2d4894b5 | 2520 | void free_unref_page_list(struct list_head *list) |
cc59850e | 2521 | { |
57490774 | 2522 | unsigned long __maybe_unused UP_flags; |
cc59850e | 2523 | struct page *page, *next; |
4b23a68f MG |
2524 | struct per_cpu_pages *pcp = NULL; |
2525 | struct zone *locked_zone = NULL; | |
c24ad77d | 2526 | int batch_count = 0; |
df1acc85 | 2527 | int migratetype; |
9cca35d4 MG |
2528 | |
2529 | /* Prepare pages for freeing */ | |
2530 | list_for_each_entry_safe(page, next, list, lru) { | |
56651377 | 2531 | unsigned long pfn = page_to_pfn(page); |
053cfda1 | 2532 | if (!free_unref_page_prepare(page, pfn, 0)) { |
9cca35d4 | 2533 | list_del(&page->lru); |
053cfda1 ML |
2534 | continue; |
2535 | } | |
df1acc85 MG |
2536 | |
2537 | /* | |
2538 | * Free isolated pages directly to the allocator, see | |
2539 | * comment in free_unref_page. | |
2540 | */ | |
2541 | migratetype = get_pcppage_migratetype(page); | |
47aef601 DB |
2542 | if (unlikely(is_migrate_isolate(migratetype))) { |
2543 | list_del(&page->lru); | |
2544 | free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE); | |
2545 | continue; | |
df1acc85 | 2546 | } |
9cca35d4 | 2547 | } |
cc59850e KK |
2548 | |
2549 | list_for_each_entry_safe(page, next, list, lru) { | |
4b23a68f MG |
2550 | struct zone *zone = page_zone(page); |
2551 | ||
c3e58a70 | 2552 | list_del(&page->lru); |
57490774 | 2553 | migratetype = get_pcppage_migratetype(page); |
c3e58a70 | 2554 | |
a4bafffb MG |
2555 | /* |
2556 | * Either different zone requiring a different pcp lock or | |
2557 | * excessive lock hold times when freeing a large list of | |
2558 | * pages. | |
2559 | */ | |
2560 | if (zone != locked_zone || batch_count == SWAP_CLUSTER_MAX) { | |
57490774 MG |
2561 | if (pcp) { |
2562 | pcp_spin_unlock(pcp); | |
2563 | pcp_trylock_finish(UP_flags); | |
2564 | } | |
01b44456 | 2565 | |
a4bafffb MG |
2566 | batch_count = 0; |
2567 | ||
57490774 MG |
2568 | /* |
2569 | * trylock is necessary as pages may be getting freed | |
2570 | * from IRQ or SoftIRQ context after an IO completion. | |
2571 | */ | |
2572 | pcp_trylock_prepare(UP_flags); | |
2573 | pcp = pcp_spin_trylock(zone->per_cpu_pageset); | |
2574 | if (unlikely(!pcp)) { | |
2575 | pcp_trylock_finish(UP_flags); | |
2576 | free_one_page(zone, page, page_to_pfn(page), | |
2577 | 0, migratetype, FPI_NONE); | |
2578 | locked_zone = NULL; | |
2579 | continue; | |
2580 | } | |
4b23a68f | 2581 | locked_zone = zone; |
4b23a68f MG |
2582 | } |
2583 | ||
47aef601 DB |
2584 | /* |
2585 | * Non-isolated types over MIGRATE_PCPTYPES get added | |
2586 | * to the MIGRATE_MOVABLE pcp list. | |
2587 | */ | |
47aef601 DB |
2588 | if (unlikely(migratetype >= MIGRATE_PCPTYPES)) |
2589 | migratetype = MIGRATE_MOVABLE; | |
2590 | ||
2d4894b5 | 2591 | trace_mm_page_free_batched(page); |
4b23a68f | 2592 | free_unref_page_commit(zone, pcp, page, migratetype, 0); |
a4bafffb | 2593 | batch_count++; |
cc59850e | 2594 | } |
4b23a68f | 2595 | |
57490774 MG |
2596 | if (pcp) { |
2597 | pcp_spin_unlock(pcp); | |
2598 | pcp_trylock_finish(UP_flags); | |
2599 | } | |
cc59850e KK |
2600 | } |
2601 | ||
8dfcc9ba NP |
2602 | /* |
2603 | * split_page takes a non-compound higher-order page, and splits it into | |
2604 | * n (1<<order) sub-pages: page[0..n] | |
2605 | * Each sub-page must be freed individually. | |
2606 | * | |
2607 | * Note: this is probably too low level an operation for use in drivers. | |
2608 | * Please consult with lkml before using this in your driver. | |
2609 | */ | |
2610 | void split_page(struct page *page, unsigned int order) | |
2611 | { | |
2612 | int i; | |
2613 | ||
309381fe SL |
2614 | VM_BUG_ON_PAGE(PageCompound(page), page); |
2615 | VM_BUG_ON_PAGE(!page_count(page), page); | |
b1eeab67 | 2616 | |
a9627bc5 | 2617 | for (i = 1; i < (1 << order); i++) |
7835e98b | 2618 | set_page_refcounted(page + i); |
8fb156c9 | 2619 | split_page_owner(page, 1 << order); |
e1baddf8 | 2620 | split_page_memcg(page, 1 << order); |
8dfcc9ba | 2621 | } |
5853ff23 | 2622 | EXPORT_SYMBOL_GPL(split_page); |
8dfcc9ba | 2623 | |
3c605096 | 2624 | int __isolate_free_page(struct page *page, unsigned int order) |
748446bb | 2625 | { |
9a157dd8 KW |
2626 | struct zone *zone = page_zone(page); |
2627 | int mt = get_pageblock_migratetype(page); | |
748446bb | 2628 | |
194159fb | 2629 | if (!is_migrate_isolate(mt)) { |
9a157dd8 | 2630 | unsigned long watermark; |
8348faf9 VB |
2631 | /* |
2632 | * Obey watermarks as if the page was being allocated. We can | |
2633 | * emulate a high-order watermark check with a raised order-0 | |
2634 | * watermark, because we already know our high-order page | |
2635 | * exists. | |
2636 | */ | |
fd1444b2 | 2637 | watermark = zone->_watermark[WMARK_MIN] + (1UL << order); |
d883c6cf | 2638 | if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA)) |
2e30abd1 MS |
2639 | return 0; |
2640 | ||
8fb74b9f | 2641 | __mod_zone_freepage_state(zone, -(1UL << order), mt); |
2e30abd1 | 2642 | } |
748446bb | 2643 | |
6ab01363 | 2644 | del_page_from_free_list(page, zone, order); |
2139cbe6 | 2645 | |
400bc7fd | 2646 | /* |
2647 | * Set the pageblock if the isolated page is at least half of a | |
2648 | * pageblock | |
2649 | */ | |
748446bb MG |
2650 | if (order >= pageblock_order - 1) { |
2651 | struct page *endpage = page + (1 << order) - 1; | |
47118af0 MN |
2652 | for (; page < endpage; page += pageblock_nr_pages) { |
2653 | int mt = get_pageblock_migratetype(page); | |
1dd214b8 ZY |
2654 | /* |
2655 | * Only change normal pageblocks (i.e., they can merge | |
2656 | * with others) | |
2657 | */ | |
2658 | if (migratetype_is_mergeable(mt)) | |
47118af0 MN |
2659 | set_pageblock_migratetype(page, |
2660 | MIGRATE_MOVABLE); | |
2661 | } | |
748446bb MG |
2662 | } |
2663 | ||
8fb74b9f | 2664 | return 1UL << order; |
1fb3f8ca MG |
2665 | } |
2666 | ||
624f58d8 AD |
2667 | /** |
2668 | * __putback_isolated_page - Return a now-isolated page back where we got it | |
2669 | * @page: Page that was isolated | |
2670 | * @order: Order of the isolated page | |
e6a0a7ad | 2671 | * @mt: The page's pageblock's migratetype |
624f58d8 AD |
2672 | * |
2673 | * This function is meant to return a page pulled from the free lists via | |
2674 | * __isolate_free_page back to the free lists they were pulled from. | |
2675 | */ | |
2676 | void __putback_isolated_page(struct page *page, unsigned int order, int mt) | |
2677 | { | |
2678 | struct zone *zone = page_zone(page); | |
2679 | ||
2680 | /* zone lock should be held when this function is called */ | |
2681 | lockdep_assert_held(&zone->lock); | |
2682 | ||
2683 | /* Return isolated page to tail of freelist. */ | |
f04a5d5d | 2684 | __free_one_page(page, page_to_pfn(page), zone, order, mt, |
47b6a24a | 2685 | FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL); |
624f58d8 AD |
2686 | } |
2687 | ||
060e7417 MG |
2688 | /* |
2689 | * Update NUMA hit/miss statistics | |
060e7417 | 2690 | */ |
3e23060b MG |
2691 | static inline void zone_statistics(struct zone *preferred_zone, struct zone *z, |
2692 | long nr_account) | |
060e7417 MG |
2693 | { |
2694 | #ifdef CONFIG_NUMA | |
3a321d2a | 2695 | enum numa_stat_item local_stat = NUMA_LOCAL; |
060e7417 | 2696 | |
4518085e KW |
2697 | /* skip numa counters update if numa stats is disabled */ |
2698 | if (!static_branch_likely(&vm_numa_stat_key)) | |
2699 | return; | |
2700 | ||
c1093b74 | 2701 | if (zone_to_nid(z) != numa_node_id()) |
060e7417 | 2702 | local_stat = NUMA_OTHER; |
060e7417 | 2703 | |
c1093b74 | 2704 | if (zone_to_nid(z) == zone_to_nid(preferred_zone)) |
3e23060b | 2705 | __count_numa_events(z, NUMA_HIT, nr_account); |
2df26639 | 2706 | else { |
3e23060b MG |
2707 | __count_numa_events(z, NUMA_MISS, nr_account); |
2708 | __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account); | |
060e7417 | 2709 | } |
3e23060b | 2710 | __count_numa_events(z, local_stat, nr_account); |
060e7417 MG |
2711 | #endif |
2712 | } | |
2713 | ||
589d9973 MG |
2714 | static __always_inline |
2715 | struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone, | |
2716 | unsigned int order, unsigned int alloc_flags, | |
2717 | int migratetype) | |
2718 | { | |
2719 | struct page *page; | |
2720 | unsigned long flags; | |
2721 | ||
2722 | do { | |
2723 | page = NULL; | |
2724 | spin_lock_irqsave(&zone->lock, flags); | |
eb2e2b42 | 2725 | if (alloc_flags & ALLOC_HIGHATOMIC) |
589d9973 MG |
2726 | page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC); |
2727 | if (!page) { | |
2728 | page = __rmqueue(zone, order, migratetype, alloc_flags); | |
eb2e2b42 MG |
2729 | |
2730 | /* | |
2731 | * If the allocation fails, allow OOM handling access | |
2732 | * to HIGHATOMIC reserves as failing now is worse than | |
2733 | * failing a high-order atomic allocation in the | |
2734 | * future. | |
2735 | */ | |
2736 | if (!page && (alloc_flags & ALLOC_OOM)) | |
2737 | page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC); | |
2738 | ||
589d9973 MG |
2739 | if (!page) { |
2740 | spin_unlock_irqrestore(&zone->lock, flags); | |
2741 | return NULL; | |
2742 | } | |
2743 | } | |
2744 | __mod_zone_freepage_state(zone, -(1 << order), | |
2745 | get_pcppage_migratetype(page)); | |
2746 | spin_unlock_irqrestore(&zone->lock, flags); | |
2747 | } while (check_new_pages(page, order)); | |
2748 | ||
2749 | __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order); | |
2750 | zone_statistics(preferred_zone, zone, 1); | |
2751 | ||
2752 | return page; | |
2753 | } | |
2754 | ||
51a755c5 | 2755 | static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order) |
c0a24239 | 2756 | { |
51a755c5 HY |
2757 | int high, base_batch, batch, max_nr_alloc; |
2758 | int high_max, high_min; | |
c0a24239 | 2759 | |
51a755c5 HY |
2760 | base_batch = READ_ONCE(pcp->batch); |
2761 | high_min = READ_ONCE(pcp->high_min); | |
2762 | high_max = READ_ONCE(pcp->high_max); | |
2763 | high = pcp->high = clamp(pcp->high, high_min, high_max); | |
c0a24239 HY |
2764 | |
2765 | /* Check for PCP disabled or boot pageset */ | |
51a755c5 | 2766 | if (unlikely(high < base_batch)) |
c0a24239 HY |
2767 | return 1; |
2768 | ||
51a755c5 HY |
2769 | if (order) |
2770 | batch = base_batch; | |
2771 | else | |
2772 | batch = (base_batch << pcp->alloc_factor); | |
2773 | ||
c0a24239 | 2774 | /* |
51a755c5 HY |
2775 | * If we had larger pcp->high, we could avoid to allocate from |
2776 | * zone. | |
c0a24239 | 2777 | */ |
57c0419c | 2778 | if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags)) |
51a755c5 HY |
2779 | high = pcp->high = min(high + batch, high_max); |
2780 | ||
c0a24239 | 2781 | if (!order) { |
51a755c5 HY |
2782 | max_nr_alloc = max(high - pcp->count - base_batch, base_batch); |
2783 | /* | |
2784 | * Double the number of pages allocated each time there is | |
2785 | * subsequent allocation of order-0 pages without any freeing. | |
2786 | */ | |
c0a24239 HY |
2787 | if (batch <= max_nr_alloc && |
2788 | pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX) | |
2789 | pcp->alloc_factor++; | |
2790 | batch = min(batch, max_nr_alloc); | |
2791 | } | |
2792 | ||
2793 | /* | |
2794 | * Scale batch relative to order if batch implies free pages | |
2795 | * can be stored on the PCP. Batch can be 1 for small zones or | |
2796 | * for boot pagesets which should never store free pages as | |
2797 | * the pages may belong to arbitrary zones. | |
2798 | */ | |
2799 | if (batch > 1) | |
2800 | batch = max(batch >> order, 2); | |
2801 | ||
2802 | return batch; | |
2803 | } | |
2804 | ||
066b2393 | 2805 | /* Remove page from the per-cpu list, caller must protect the list */ |
3b822017 | 2806 | static inline |
44042b44 MG |
2807 | struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order, |
2808 | int migratetype, | |
6bb15450 | 2809 | unsigned int alloc_flags, |
453f85d4 | 2810 | struct per_cpu_pages *pcp, |
066b2393 MG |
2811 | struct list_head *list) |
2812 | { | |
2813 | struct page *page; | |
2814 | ||
2815 | do { | |
2816 | if (list_empty(list)) { | |
51a755c5 | 2817 | int batch = nr_pcp_alloc(pcp, zone, order); |
44042b44 MG |
2818 | int alloced; |
2819 | ||
44042b44 MG |
2820 | alloced = rmqueue_bulk(zone, order, |
2821 | batch, list, | |
6bb15450 | 2822 | migratetype, alloc_flags); |
44042b44 MG |
2823 | |
2824 | pcp->count += alloced << order; | |
066b2393 MG |
2825 | if (unlikely(list_empty(list))) |
2826 | return NULL; | |
2827 | } | |
2828 | ||
bf75f200 MG |
2829 | page = list_first_entry(list, struct page, pcp_list); |
2830 | list_del(&page->pcp_list); | |
44042b44 | 2831 | pcp->count -= 1 << order; |
700d2e9a | 2832 | } while (check_new_pages(page, order)); |
066b2393 MG |
2833 | |
2834 | return page; | |
2835 | } | |
2836 | ||
2837 | /* Lock and remove page from the per-cpu list */ | |
2838 | static struct page *rmqueue_pcplist(struct zone *preferred_zone, | |
44042b44 | 2839 | struct zone *zone, unsigned int order, |
663d0cfd | 2840 | int migratetype, unsigned int alloc_flags) |
066b2393 MG |
2841 | { |
2842 | struct per_cpu_pages *pcp; | |
2843 | struct list_head *list; | |
066b2393 | 2844 | struct page *page; |
4b23a68f | 2845 | unsigned long __maybe_unused UP_flags; |
066b2393 | 2846 | |
57490774 | 2847 | /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */ |
4b23a68f | 2848 | pcp_trylock_prepare(UP_flags); |
57490774 | 2849 | pcp = pcp_spin_trylock(zone->per_cpu_pageset); |
01b44456 | 2850 | if (!pcp) { |
4b23a68f | 2851 | pcp_trylock_finish(UP_flags); |
4b23a68f MG |
2852 | return NULL; |
2853 | } | |
3b12e7e9 MG |
2854 | |
2855 | /* | |
2856 | * On allocation, reduce the number of pages that are batch freed. | |
2857 | * See nr_pcp_free() where free_factor is increased for subsequent | |
2858 | * frees. | |
2859 | */ | |
6ccdcb6d | 2860 | pcp->free_count >>= 1; |
44042b44 MG |
2861 | list = &pcp->lists[order_to_pindex(migratetype, order)]; |
2862 | page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list); | |
57490774 | 2863 | pcp_spin_unlock(pcp); |
4b23a68f | 2864 | pcp_trylock_finish(UP_flags); |
066b2393 | 2865 | if (page) { |
15cd9004 | 2866 | __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order); |
3e23060b | 2867 | zone_statistics(preferred_zone, zone, 1); |
066b2393 | 2868 | } |
066b2393 MG |
2869 | return page; |
2870 | } | |
2871 | ||
1da177e4 | 2872 | /* |
a57ae9ef RX |
2873 | * Allocate a page from the given zone. |
2874 | * Use pcplists for THP or "cheap" high-order allocations. | |
1da177e4 | 2875 | */ |
b073d7f8 AP |
2876 | |
2877 | /* | |
2878 | * Do not instrument rmqueue() with KMSAN. This function may call | |
2879 | * __msan_poison_alloca() through a call to set_pfnblock_flags_mask(). | |
2880 | * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it | |
2881 | * may call rmqueue() again, which will result in a deadlock. | |
1da177e4 | 2882 | */ |
b073d7f8 | 2883 | __no_sanitize_memory |
0a15c3e9 | 2884 | static inline |
066b2393 | 2885 | struct page *rmqueue(struct zone *preferred_zone, |
7aeb09f9 | 2886 | struct zone *zone, unsigned int order, |
c603844b MG |
2887 | gfp_t gfp_flags, unsigned int alloc_flags, |
2888 | int migratetype) | |
1da177e4 | 2889 | { |
689bcebf | 2890 | struct page *page; |
1da177e4 | 2891 | |
589d9973 MG |
2892 | /* |
2893 | * We most definitely don't want callers attempting to | |
2894 | * allocate greater than order-1 page units with __GFP_NOFAIL. | |
2895 | */ | |
2896 | WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1)); | |
2897 | ||
44042b44 | 2898 | if (likely(pcp_allowed_order(order))) { |
f945116e JW |
2899 | page = rmqueue_pcplist(preferred_zone, zone, order, |
2900 | migratetype, alloc_flags); | |
2901 | if (likely(page)) | |
2902 | goto out; | |
066b2393 | 2903 | } |
83b9355b | 2904 | |
589d9973 MG |
2905 | page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags, |
2906 | migratetype); | |
1da177e4 | 2907 | |
066b2393 | 2908 | out: |
73444bc4 | 2909 | /* Separate test+clear to avoid unnecessary atomics */ |
3b11edf1 TH |
2910 | if ((alloc_flags & ALLOC_KSWAPD) && |
2911 | unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) { | |
73444bc4 MG |
2912 | clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags); |
2913 | wakeup_kswapd(zone, 0, 0, zone_idx(zone)); | |
2914 | } | |
2915 | ||
066b2393 | 2916 | VM_BUG_ON_PAGE(page && bad_range(zone, page), page); |
1da177e4 LT |
2917 | return page; |
2918 | } | |
2919 | ||
54aa3866 | 2920 | noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) |
af3b8544 BP |
2921 | { |
2922 | return __should_fail_alloc_page(gfp_mask, order); | |
2923 | } | |
2924 | ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE); | |
2925 | ||
f27ce0e1 JK |
2926 | static inline long __zone_watermark_unusable_free(struct zone *z, |
2927 | unsigned int order, unsigned int alloc_flags) | |
2928 | { | |
f27ce0e1 JK |
2929 | long unusable_free = (1 << order) - 1; |
2930 | ||
2931 | /* | |
ab350885 MG |
2932 | * If the caller does not have rights to reserves below the min |
2933 | * watermark then subtract the high-atomic reserves. This will | |
2934 | * over-estimate the size of the atomic reserve but it avoids a search. | |
f27ce0e1 | 2935 | */ |
ab350885 | 2936 | if (likely(!(alloc_flags & ALLOC_RESERVES))) |
f27ce0e1 JK |
2937 | unusable_free += z->nr_reserved_highatomic; |
2938 | ||
2939 | #ifdef CONFIG_CMA | |
2940 | /* If allocation can't use CMA areas don't use free CMA pages */ | |
2941 | if (!(alloc_flags & ALLOC_CMA)) | |
2942 | unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES); | |
2943 | #endif | |
dcdfdd40 KS |
2944 | #ifdef CONFIG_UNACCEPTED_MEMORY |
2945 | unusable_free += zone_page_state(z, NR_UNACCEPTED); | |
2946 | #endif | |
f27ce0e1 JK |
2947 | |
2948 | return unusable_free; | |
2949 | } | |
2950 | ||
1da177e4 | 2951 | /* |
97a16fc8 MG |
2952 | * Return true if free base pages are above 'mark'. For high-order checks it |
2953 | * will return true of the order-0 watermark is reached and there is at least | |
2954 | * one free page of a suitable size. Checking now avoids taking the zone lock | |
2955 | * to check in the allocation paths if no pages are free. | |
1da177e4 | 2956 | */ |
86a294a8 | 2957 | bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark, |
97a225e6 | 2958 | int highest_zoneidx, unsigned int alloc_flags, |
86a294a8 | 2959 | long free_pages) |
1da177e4 | 2960 | { |
d23ad423 | 2961 | long min = mark; |
1da177e4 LT |
2962 | int o; |
2963 | ||
0aaa29a5 | 2964 | /* free_pages may go negative - that's OK */ |
f27ce0e1 | 2965 | free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags); |
0aaa29a5 | 2966 | |
ab350885 MG |
2967 | if (unlikely(alloc_flags & ALLOC_RESERVES)) { |
2968 | /* | |
2969 | * __GFP_HIGH allows access to 50% of the min reserve as well | |
2970 | * as OOM. | |
2971 | */ | |
1ebbb218 | 2972 | if (alloc_flags & ALLOC_MIN_RESERVE) { |
ab350885 | 2973 | min -= min / 2; |
0aaa29a5 | 2974 | |
1ebbb218 MG |
2975 | /* |
2976 | * Non-blocking allocations (e.g. GFP_ATOMIC) can | |
2977 | * access more reserves than just __GFP_HIGH. Other | |
2978 | * non-blocking allocations requests such as GFP_NOWAIT | |
2979 | * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get | |
2980 | * access to the min reserve. | |
2981 | */ | |
2982 | if (alloc_flags & ALLOC_NON_BLOCK) | |
2983 | min -= min / 4; | |
2984 | } | |
0aaa29a5 | 2985 | |
cd04ae1e | 2986 | /* |
ab350885 | 2987 | * OOM victims can try even harder than the normal reserve |
cd04ae1e MH |
2988 | * users on the grounds that it's definitely going to be in |
2989 | * the exit path shortly and free memory. Any allocation it | |
2990 | * makes during the free path will be small and short-lived. | |
2991 | */ | |
2992 | if (alloc_flags & ALLOC_OOM) | |
2993 | min -= min / 2; | |
cd04ae1e MH |
2994 | } |
2995 | ||
97a16fc8 MG |
2996 | /* |
2997 | * Check watermarks for an order-0 allocation request. If these | |
2998 | * are not met, then a high-order request also cannot go ahead | |
2999 | * even if a suitable page happened to be free. | |
3000 | */ | |
97a225e6 | 3001 | if (free_pages <= min + z->lowmem_reserve[highest_zoneidx]) |
88f5acf8 | 3002 | return false; |
1da177e4 | 3003 | |
97a16fc8 MG |
3004 | /* If this is an order-0 request then the watermark is fine */ |
3005 | if (!order) | |
3006 | return true; | |
3007 | ||
3008 | /* For a high-order request, check at least one suitable page is free */ | |
fd377218 | 3009 | for (o = order; o < NR_PAGE_ORDERS; o++) { |
97a16fc8 MG |
3010 | struct free_area *area = &z->free_area[o]; |
3011 | int mt; | |
3012 | ||
3013 | if (!area->nr_free) | |
3014 | continue; | |
3015 | ||
97a16fc8 | 3016 | for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) { |
b03641af | 3017 | if (!free_area_empty(area, mt)) |
97a16fc8 MG |
3018 | return true; |
3019 | } | |
3020 | ||
3021 | #ifdef CONFIG_CMA | |
d883c6cf | 3022 | if ((alloc_flags & ALLOC_CMA) && |
b03641af | 3023 | !free_area_empty(area, MIGRATE_CMA)) { |
97a16fc8 | 3024 | return true; |
d883c6cf | 3025 | } |
97a16fc8 | 3026 | #endif |
eb2e2b42 MG |
3027 | if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) && |
3028 | !free_area_empty(area, MIGRATE_HIGHATOMIC)) { | |
b050e376 | 3029 | return true; |
eb2e2b42 | 3030 | } |
1da177e4 | 3031 | } |
97a16fc8 | 3032 | return false; |
88f5acf8 MG |
3033 | } |
3034 | ||
7aeb09f9 | 3035 | bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark, |
97a225e6 | 3036 | int highest_zoneidx, unsigned int alloc_flags) |
88f5acf8 | 3037 | { |
97a225e6 | 3038 | return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags, |
88f5acf8 MG |
3039 | zone_page_state(z, NR_FREE_PAGES)); |
3040 | } | |
3041 | ||
48ee5f36 | 3042 | static inline bool zone_watermark_fast(struct zone *z, unsigned int order, |
97a225e6 | 3043 | unsigned long mark, int highest_zoneidx, |
f80b08fc | 3044 | unsigned int alloc_flags, gfp_t gfp_mask) |
48ee5f36 | 3045 | { |
f27ce0e1 | 3046 | long free_pages; |
d883c6cf | 3047 | |
f27ce0e1 | 3048 | free_pages = zone_page_state(z, NR_FREE_PAGES); |
48ee5f36 MG |
3049 | |
3050 | /* | |
3051 | * Fast check for order-0 only. If this fails then the reserves | |
f27ce0e1 | 3052 | * need to be calculated. |
48ee5f36 | 3053 | */ |
f27ce0e1 | 3054 | if (!order) { |
9282012f JK |
3055 | long usable_free; |
3056 | long reserved; | |
f27ce0e1 | 3057 | |
9282012f JK |
3058 | usable_free = free_pages; |
3059 | reserved = __zone_watermark_unusable_free(z, 0, alloc_flags); | |
3060 | ||
3061 | /* reserved may over estimate high-atomic reserves. */ | |
3062 | usable_free -= min(usable_free, reserved); | |
3063 | if (usable_free > mark + z->lowmem_reserve[highest_zoneidx]) | |
f27ce0e1 JK |
3064 | return true; |
3065 | } | |
48ee5f36 | 3066 | |
f80b08fc CTR |
3067 | if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags, |
3068 | free_pages)) | |
3069 | return true; | |
2973d822 | 3070 | |
f80b08fc | 3071 | /* |
2973d822 | 3072 | * Ignore watermark boosting for __GFP_HIGH order-0 allocations |
f80b08fc CTR |
3073 | * when checking the min watermark. The min watermark is the |
3074 | * point where boosting is ignored so that kswapd is woken up | |
3075 | * when below the low watermark. | |
3076 | */ | |
2973d822 | 3077 | if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost |
f80b08fc CTR |
3078 | && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) { |
3079 | mark = z->_watermark[WMARK_MIN]; | |
3080 | return __zone_watermark_ok(z, order, mark, highest_zoneidx, | |
3081 | alloc_flags, free_pages); | |
3082 | } | |
3083 | ||
3084 | return false; | |
48ee5f36 MG |
3085 | } |
3086 | ||
7aeb09f9 | 3087 | bool zone_watermark_ok_safe(struct zone *z, unsigned int order, |
97a225e6 | 3088 | unsigned long mark, int highest_zoneidx) |
88f5acf8 MG |
3089 | { |
3090 | long free_pages = zone_page_state(z, NR_FREE_PAGES); | |
3091 | ||
3092 | if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) | |
3093 | free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); | |
3094 | ||
97a225e6 | 3095 | return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0, |
88f5acf8 | 3096 | free_pages); |
1da177e4 LT |
3097 | } |
3098 | ||
9276b1bc | 3099 | #ifdef CONFIG_NUMA |
61bb6cd2 GU |
3100 | int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE; |
3101 | ||
957f822a DR |
3102 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) |
3103 | { | |
e02dc017 | 3104 | return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <= |
a55c7454 | 3105 | node_reclaim_distance; |
957f822a | 3106 | } |
9276b1bc | 3107 | #else /* CONFIG_NUMA */ |
957f822a DR |
3108 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) |
3109 | { | |
3110 | return true; | |
3111 | } | |
9276b1bc PJ |
3112 | #endif /* CONFIG_NUMA */ |
3113 | ||
6bb15450 MG |
3114 | /* |
3115 | * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid | |
3116 | * fragmentation is subtle. If the preferred zone was HIGHMEM then | |
3117 | * premature use of a lower zone may cause lowmem pressure problems that | |
3118 | * are worse than fragmentation. If the next zone is ZONE_DMA then it is | |
3119 | * probably too small. It only makes sense to spread allocations to avoid | |
3120 | * fragmentation between the Normal and DMA32 zones. | |
3121 | */ | |
3122 | static inline unsigned int | |
0a79cdad | 3123 | alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask) |
6bb15450 | 3124 | { |
736838e9 | 3125 | unsigned int alloc_flags; |
0a79cdad | 3126 | |
736838e9 MN |
3127 | /* |
3128 | * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD | |
3129 | * to save a branch. | |
3130 | */ | |
3131 | alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM); | |
0a79cdad MG |
3132 | |
3133 | #ifdef CONFIG_ZONE_DMA32 | |
8139ad04 AR |
3134 | if (!zone) |
3135 | return alloc_flags; | |
3136 | ||
6bb15450 | 3137 | if (zone_idx(zone) != ZONE_NORMAL) |
8118b82e | 3138 | return alloc_flags; |
6bb15450 MG |
3139 | |
3140 | /* | |
3141 | * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and | |
3142 | * the pointer is within zone->zone_pgdat->node_zones[]. Also assume | |
3143 | * on UMA that if Normal is populated then so is DMA32. | |
3144 | */ | |
3145 | BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1); | |
3146 | if (nr_online_nodes > 1 && !populated_zone(--zone)) | |
8118b82e | 3147 | return alloc_flags; |
6bb15450 | 3148 | |
8118b82e | 3149 | alloc_flags |= ALLOC_NOFRAGMENT; |
0a79cdad MG |
3150 | #endif /* CONFIG_ZONE_DMA32 */ |
3151 | return alloc_flags; | |
6bb15450 | 3152 | } |
6bb15450 | 3153 | |
8e3560d9 PT |
3154 | /* Must be called after current_gfp_context() which can change gfp_mask */ |
3155 | static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask, | |
3156 | unsigned int alloc_flags) | |
8510e69c JK |
3157 | { |
3158 | #ifdef CONFIG_CMA | |
8e3560d9 | 3159 | if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE) |
8510e69c | 3160 | alloc_flags |= ALLOC_CMA; |
8510e69c JK |
3161 | #endif |
3162 | return alloc_flags; | |
3163 | } | |
3164 | ||
7fb1d9fc | 3165 | /* |
0798e519 | 3166 | * get_page_from_freelist goes through the zonelist trying to allocate |
7fb1d9fc RS |
3167 | * a page. |
3168 | */ | |
3169 | static struct page * | |
a9263751 VB |
3170 | get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags, |
3171 | const struct alloc_context *ac) | |
753ee728 | 3172 | { |
6bb15450 | 3173 | struct zoneref *z; |
5117f45d | 3174 | struct zone *zone; |
8a87d695 WY |
3175 | struct pglist_data *last_pgdat = NULL; |
3176 | bool last_pgdat_dirty_ok = false; | |
6bb15450 | 3177 | bool no_fallback; |
3b8c0be4 | 3178 | |
6bb15450 | 3179 | retry: |
7fb1d9fc | 3180 | /* |
9276b1bc | 3181 | * Scan zonelist, looking for a zone with enough free. |
8e464522 | 3182 | * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c. |
7fb1d9fc | 3183 | */ |
6bb15450 MG |
3184 | no_fallback = alloc_flags & ALLOC_NOFRAGMENT; |
3185 | z = ac->preferred_zoneref; | |
30d8ec73 MN |
3186 | for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx, |
3187 | ac->nodemask) { | |
be06af00 | 3188 | struct page *page; |
e085dbc5 JW |
3189 | unsigned long mark; |
3190 | ||
664eedde MG |
3191 | if (cpusets_enabled() && |
3192 | (alloc_flags & ALLOC_CPUSET) && | |
002f2906 | 3193 | !__cpuset_zone_allowed(zone, gfp_mask)) |
cd38b115 | 3194 | continue; |
a756cf59 JW |
3195 | /* |
3196 | * When allocating a page cache page for writing, we | |
281e3726 MG |
3197 | * want to get it from a node that is within its dirty |
3198 | * limit, such that no single node holds more than its | |
a756cf59 | 3199 | * proportional share of globally allowed dirty pages. |
281e3726 | 3200 | * The dirty limits take into account the node's |
a756cf59 JW |
3201 | * lowmem reserves and high watermark so that kswapd |
3202 | * should be able to balance it without having to | |
3203 | * write pages from its LRU list. | |
3204 | * | |
a756cf59 | 3205 | * XXX: For now, allow allocations to potentially |
281e3726 | 3206 | * exceed the per-node dirty limit in the slowpath |
c9ab0c4f | 3207 | * (spread_dirty_pages unset) before going into reclaim, |
a756cf59 | 3208 | * which is important when on a NUMA setup the allowed |
281e3726 | 3209 | * nodes are together not big enough to reach the |
a756cf59 | 3210 | * global limit. The proper fix for these situations |
281e3726 | 3211 | * will require awareness of nodes in the |
a756cf59 JW |
3212 | * dirty-throttling and the flusher threads. |
3213 | */ | |
3b8c0be4 | 3214 | if (ac->spread_dirty_pages) { |
8a87d695 WY |
3215 | if (last_pgdat != zone->zone_pgdat) { |
3216 | last_pgdat = zone->zone_pgdat; | |
3217 | last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat); | |
3218 | } | |
3b8c0be4 | 3219 | |
8a87d695 | 3220 | if (!last_pgdat_dirty_ok) |
3b8c0be4 | 3221 | continue; |
3b8c0be4 | 3222 | } |
7fb1d9fc | 3223 | |
6bb15450 MG |
3224 | if (no_fallback && nr_online_nodes > 1 && |
3225 | zone != ac->preferred_zoneref->zone) { | |
3226 | int local_nid; | |
3227 | ||
3228 | /* | |
3229 | * If moving to a remote node, retry but allow | |
3230 | * fragmenting fallbacks. Locality is more important | |
3231 | * than fragmentation avoidance. | |
3232 | */ | |
3233 | local_nid = zone_to_nid(ac->preferred_zoneref->zone); | |
3234 | if (zone_to_nid(zone) != local_nid) { | |
3235 | alloc_flags &= ~ALLOC_NOFRAGMENT; | |
3236 | goto retry; | |
3237 | } | |
3238 | } | |
3239 | ||
57c0419c HY |
3240 | /* |
3241 | * Detect whether the number of free pages is below high | |
3242 | * watermark. If so, we will decrease pcp->high and free | |
3243 | * PCP pages in free path to reduce the possibility of | |
3244 | * premature page reclaiming. Detection is done here to | |
3245 | * avoid to do that in hotter free path. | |
3246 | */ | |
3247 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) | |
3248 | goto check_alloc_wmark; | |
3249 | ||
3250 | mark = high_wmark_pages(zone); | |
3251 | if (zone_watermark_fast(zone, order, mark, | |
3252 | ac->highest_zoneidx, alloc_flags, | |
3253 | gfp_mask)) | |
3254 | goto try_this_zone; | |
3255 | else | |
3256 | set_bit(ZONE_BELOW_HIGH, &zone->flags); | |
3257 | ||
3258 | check_alloc_wmark: | |
a9214443 | 3259 | mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK); |
48ee5f36 | 3260 | if (!zone_watermark_fast(zone, order, mark, |
f80b08fc CTR |
3261 | ac->highest_zoneidx, alloc_flags, |
3262 | gfp_mask)) { | |
fa5e084e MG |
3263 | int ret; |
3264 | ||
dcdfdd40 KS |
3265 | if (has_unaccepted_memory()) { |
3266 | if (try_to_accept_memory(zone, order)) | |
3267 | goto try_this_zone; | |
3268 | } | |
3269 | ||
c9e97a19 PT |
3270 | #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT |
3271 | /* | |
3272 | * Watermark failed for this zone, but see if we can | |
3273 | * grow this zone if it contains deferred pages. | |
3274 | */ | |
076cf7ea | 3275 | if (deferred_pages_enabled()) { |
c9e97a19 PT |
3276 | if (_deferred_grow_zone(zone, order)) |
3277 | goto try_this_zone; | |
3278 | } | |
3279 | #endif | |
5dab2911 MG |
3280 | /* Checked here to keep the fast path fast */ |
3281 | BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); | |
3282 | if (alloc_flags & ALLOC_NO_WATERMARKS) | |
3283 | goto try_this_zone; | |
3284 | ||
202e35db | 3285 | if (!node_reclaim_enabled() || |
c33d6c06 | 3286 | !zone_allows_reclaim(ac->preferred_zoneref->zone, zone)) |
cd38b115 MG |
3287 | continue; |
3288 | ||
a5f5f91d | 3289 | ret = node_reclaim(zone->zone_pgdat, gfp_mask, order); |
fa5e084e | 3290 | switch (ret) { |
a5f5f91d | 3291 | case NODE_RECLAIM_NOSCAN: |
fa5e084e | 3292 | /* did not scan */ |
cd38b115 | 3293 | continue; |
a5f5f91d | 3294 | case NODE_RECLAIM_FULL: |
fa5e084e | 3295 | /* scanned but unreclaimable */ |
cd38b115 | 3296 | continue; |
fa5e084e MG |
3297 | default: |
3298 | /* did we reclaim enough */ | |
fed2719e | 3299 | if (zone_watermark_ok(zone, order, mark, |
97a225e6 | 3300 | ac->highest_zoneidx, alloc_flags)) |
fed2719e MG |
3301 | goto try_this_zone; |
3302 | ||
fed2719e | 3303 | continue; |
0798e519 | 3304 | } |
7fb1d9fc RS |
3305 | } |
3306 | ||
fa5e084e | 3307 | try_this_zone: |
066b2393 | 3308 | page = rmqueue(ac->preferred_zoneref->zone, zone, order, |
0aaa29a5 | 3309 | gfp_mask, alloc_flags, ac->migratetype); |
75379191 | 3310 | if (page) { |
479f854a | 3311 | prep_new_page(page, order, gfp_mask, alloc_flags); |
0aaa29a5 MG |
3312 | |
3313 | /* | |
3314 | * If this is a high-order atomic allocation then check | |
3315 | * if the pageblock should be reserved for the future | |
3316 | */ | |
eb2e2b42 | 3317 | if (unlikely(alloc_flags & ALLOC_HIGHATOMIC)) |
368d983b | 3318 | reserve_highatomic_pageblock(page, zone); |
0aaa29a5 | 3319 | |
75379191 | 3320 | return page; |
c9e97a19 | 3321 | } else { |
dcdfdd40 KS |
3322 | if (has_unaccepted_memory()) { |
3323 | if (try_to_accept_memory(zone, order)) | |
3324 | goto try_this_zone; | |
3325 | } | |
3326 | ||
c9e97a19 PT |
3327 | #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT |
3328 | /* Try again if zone has deferred pages */ | |
076cf7ea | 3329 | if (deferred_pages_enabled()) { |
c9e97a19 PT |
3330 | if (_deferred_grow_zone(zone, order)) |
3331 | goto try_this_zone; | |
3332 | } | |
3333 | #endif | |
75379191 | 3334 | } |
54a6eb5c | 3335 | } |
9276b1bc | 3336 | |
6bb15450 MG |
3337 | /* |
3338 | * It's possible on a UMA machine to get through all zones that are | |
3339 | * fragmented. If avoiding fragmentation, reset and try again. | |
3340 | */ | |
3341 | if (no_fallback) { | |
3342 | alloc_flags &= ~ALLOC_NOFRAGMENT; | |
3343 | goto retry; | |
3344 | } | |
3345 | ||
4ffeaf35 | 3346 | return NULL; |
753ee728 MH |
3347 | } |
3348 | ||
9af744d7 | 3349 | static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask) |
a238ab5b | 3350 | { |
a238ab5b | 3351 | unsigned int filter = SHOW_MEM_FILTER_NODES; |
a238ab5b DH |
3352 | |
3353 | /* | |
3354 | * This documents exceptions given to allocations in certain | |
3355 | * contexts that are allowed to allocate outside current's set | |
3356 | * of allowed nodes. | |
3357 | */ | |
3358 | if (!(gfp_mask & __GFP_NOMEMALLOC)) | |
cd04ae1e | 3359 | if (tsk_is_oom_victim(current) || |
a238ab5b DH |
3360 | (current->flags & (PF_MEMALLOC | PF_EXITING))) |
3361 | filter &= ~SHOW_MEM_FILTER_NODES; | |
88dc6f20 | 3362 | if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM)) |
a238ab5b DH |
3363 | filter &= ~SHOW_MEM_FILTER_NODES; |
3364 | ||
974f4367 | 3365 | __show_mem(filter, nodemask, gfp_zone(gfp_mask)); |
aa187507 MH |
3366 | } |
3367 | ||
a8e99259 | 3368 | void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...) |
aa187507 MH |
3369 | { |
3370 | struct va_format vaf; | |
3371 | va_list args; | |
1be334e5 | 3372 | static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1); |
aa187507 | 3373 | |
c4dc63f0 BH |
3374 | if ((gfp_mask & __GFP_NOWARN) || |
3375 | !__ratelimit(&nopage_rs) || | |
3376 | ((gfp_mask & __GFP_DMA) && !has_managed_dma())) | |
aa187507 MH |
3377 | return; |
3378 | ||
7877cdcc MH |
3379 | va_start(args, fmt); |
3380 | vaf.fmt = fmt; | |
3381 | vaf.va = &args; | |
ef8444ea | 3382 | pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl", |
0205f755 MH |
3383 | current->comm, &vaf, gfp_mask, &gfp_mask, |
3384 | nodemask_pr_args(nodemask)); | |
7877cdcc | 3385 | va_end(args); |
3ee9a4f0 | 3386 | |
a8e99259 | 3387 | cpuset_print_current_mems_allowed(); |
ef8444ea | 3388 | pr_cont("\n"); |
a238ab5b | 3389 | dump_stack(); |
685dbf6f | 3390 | warn_alloc_show_mem(gfp_mask, nodemask); |
a238ab5b DH |
3391 | } |
3392 | ||
6c18ba7a MH |
3393 | static inline struct page * |
3394 | __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order, | |
3395 | unsigned int alloc_flags, | |
3396 | const struct alloc_context *ac) | |
3397 | { | |
3398 | struct page *page; | |
3399 | ||
3400 | page = get_page_from_freelist(gfp_mask, order, | |
3401 | alloc_flags|ALLOC_CPUSET, ac); | |
3402 | /* | |
3403 | * fallback to ignore cpuset restriction if our nodes | |
3404 | * are depleted | |
3405 | */ | |
3406 | if (!page) | |
3407 | page = get_page_from_freelist(gfp_mask, order, | |
3408 | alloc_flags, ac); | |
3409 | ||
3410 | return page; | |
3411 | } | |
3412 | ||
11e33f6a MG |
3413 | static inline struct page * |
3414 | __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, | |
a9263751 | 3415 | const struct alloc_context *ac, unsigned long *did_some_progress) |
11e33f6a | 3416 | { |
6e0fc46d DR |
3417 | struct oom_control oc = { |
3418 | .zonelist = ac->zonelist, | |
3419 | .nodemask = ac->nodemask, | |
2a966b77 | 3420 | .memcg = NULL, |
6e0fc46d DR |
3421 | .gfp_mask = gfp_mask, |
3422 | .order = order, | |
6e0fc46d | 3423 | }; |
11e33f6a MG |
3424 | struct page *page; |
3425 | ||
9879de73 JW |
3426 | *did_some_progress = 0; |
3427 | ||
9879de73 | 3428 | /* |
dc56401f JW |
3429 | * Acquire the oom lock. If that fails, somebody else is |
3430 | * making progress for us. | |
9879de73 | 3431 | */ |
dc56401f | 3432 | if (!mutex_trylock(&oom_lock)) { |
9879de73 | 3433 | *did_some_progress = 1; |
11e33f6a | 3434 | schedule_timeout_uninterruptible(1); |
1da177e4 LT |
3435 | return NULL; |
3436 | } | |
6b1de916 | 3437 | |
11e33f6a MG |
3438 | /* |
3439 | * Go through the zonelist yet one more time, keep very high watermark | |
3440 | * here, this is only to catch a parallel oom killing, we must fail if | |
e746bf73 TH |
3441 | * we're still under heavy pressure. But make sure that this reclaim |
3442 | * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY | |
3443 | * allocation which will never fail due to oom_lock already held. | |
11e33f6a | 3444 | */ |
e746bf73 TH |
3445 | page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) & |
3446 | ~__GFP_DIRECT_RECLAIM, order, | |
3447 | ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac); | |
7fb1d9fc | 3448 | if (page) |
11e33f6a MG |
3449 | goto out; |
3450 | ||
06ad276a MH |
3451 | /* Coredumps can quickly deplete all memory reserves */ |
3452 | if (current->flags & PF_DUMPCORE) | |
3453 | goto out; | |
3454 | /* The OOM killer will not help higher order allocs */ | |
3455 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
3456 | goto out; | |
dcda9b04 MH |
3457 | /* |
3458 | * We have already exhausted all our reclaim opportunities without any | |
3459 | * success so it is time to admit defeat. We will skip the OOM killer | |
3460 | * because it is very likely that the caller has a more reasonable | |
3461 | * fallback than shooting a random task. | |
cfb4a541 MN |
3462 | * |
3463 | * The OOM killer may not free memory on a specific node. | |
dcda9b04 | 3464 | */ |
cfb4a541 | 3465 | if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE)) |
dcda9b04 | 3466 | goto out; |
06ad276a | 3467 | /* The OOM killer does not needlessly kill tasks for lowmem */ |
97a225e6 | 3468 | if (ac->highest_zoneidx < ZONE_NORMAL) |
06ad276a MH |
3469 | goto out; |
3470 | if (pm_suspended_storage()) | |
3471 | goto out; | |
3472 | /* | |
3473 | * XXX: GFP_NOFS allocations should rather fail than rely on | |
3474 | * other request to make a forward progress. | |
3475 | * We are in an unfortunate situation where out_of_memory cannot | |
3476 | * do much for this context but let's try it to at least get | |
3477 | * access to memory reserved if the current task is killed (see | |
3478 | * out_of_memory). Once filesystems are ready to handle allocation | |
3479 | * failures more gracefully we should just bail out here. | |
3480 | */ | |
3481 | ||
3c2c6488 | 3482 | /* Exhausted what can be done so it's blame time */ |
3f913fc5 QZ |
3483 | if (out_of_memory(&oc) || |
3484 | WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) { | |
c32b3cbe | 3485 | *did_some_progress = 1; |
5020e285 | 3486 | |
6c18ba7a MH |
3487 | /* |
3488 | * Help non-failing allocations by giving them access to memory | |
3489 | * reserves | |
3490 | */ | |
3491 | if (gfp_mask & __GFP_NOFAIL) | |
3492 | page = __alloc_pages_cpuset_fallback(gfp_mask, order, | |
5020e285 | 3493 | ALLOC_NO_WATERMARKS, ac); |
5020e285 | 3494 | } |
11e33f6a | 3495 | out: |
dc56401f | 3496 | mutex_unlock(&oom_lock); |
11e33f6a MG |
3497 | return page; |
3498 | } | |
3499 | ||
33c2d214 | 3500 | /* |
baf2f90b | 3501 | * Maximum number of compaction retries with a progress before OOM |
33c2d214 MH |
3502 | * killer is consider as the only way to move forward. |
3503 | */ | |
3504 | #define MAX_COMPACT_RETRIES 16 | |
3505 | ||
56de7263 MG |
3506 | #ifdef CONFIG_COMPACTION |
3507 | /* Try memory compaction for high-order allocations before reclaim */ | |
3508 | static struct page * | |
3509 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, | |
c603844b | 3510 | unsigned int alloc_flags, const struct alloc_context *ac, |
a5508cd8 | 3511 | enum compact_priority prio, enum compact_result *compact_result) |
56de7263 | 3512 | { |
5e1f0f09 | 3513 | struct page *page = NULL; |
eb414681 | 3514 | unsigned long pflags; |
499118e9 | 3515 | unsigned int noreclaim_flag; |
53853e2d VB |
3516 | |
3517 | if (!order) | |
66199712 | 3518 | return NULL; |
66199712 | 3519 | |
eb414681 | 3520 | psi_memstall_enter(&pflags); |
5bf18281 | 3521 | delayacct_compact_start(); |
499118e9 | 3522 | noreclaim_flag = memalloc_noreclaim_save(); |
eb414681 | 3523 | |
c5d01d0d | 3524 | *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac, |
5e1f0f09 | 3525 | prio, &page); |
eb414681 | 3526 | |
499118e9 | 3527 | memalloc_noreclaim_restore(noreclaim_flag); |
eb414681 | 3528 | psi_memstall_leave(&pflags); |
5bf18281 | 3529 | delayacct_compact_end(); |
56de7263 | 3530 | |
06dac2f4 CTR |
3531 | if (*compact_result == COMPACT_SKIPPED) |
3532 | return NULL; | |
98dd3b48 VB |
3533 | /* |
3534 | * At least in one zone compaction wasn't deferred or skipped, so let's | |
3535 | * count a compaction stall | |
3536 | */ | |
3537 | count_vm_event(COMPACTSTALL); | |
8fb74b9f | 3538 | |
5e1f0f09 MG |
3539 | /* Prep a captured page if available */ |
3540 | if (page) | |
3541 | prep_new_page(page, order, gfp_mask, alloc_flags); | |
3542 | ||
3543 | /* Try get a page from the freelist if available */ | |
3544 | if (!page) | |
3545 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); | |
53853e2d | 3546 | |
98dd3b48 VB |
3547 | if (page) { |
3548 | struct zone *zone = page_zone(page); | |
53853e2d | 3549 | |
98dd3b48 VB |
3550 | zone->compact_blockskip_flush = false; |
3551 | compaction_defer_reset(zone, order, true); | |
3552 | count_vm_event(COMPACTSUCCESS); | |
3553 | return page; | |
3554 | } | |
56de7263 | 3555 | |
98dd3b48 VB |
3556 | /* |
3557 | * It's bad if compaction run occurs and fails. The most likely reason | |
3558 | * is that pages exist, but not enough to satisfy watermarks. | |
3559 | */ | |
3560 | count_vm_event(COMPACTFAIL); | |
66199712 | 3561 | |
98dd3b48 | 3562 | cond_resched(); |
56de7263 MG |
3563 | |
3564 | return NULL; | |
3565 | } | |
33c2d214 | 3566 | |
3250845d VB |
3567 | static inline bool |
3568 | should_compact_retry(struct alloc_context *ac, int order, int alloc_flags, | |
3569 | enum compact_result compact_result, | |
3570 | enum compact_priority *compact_priority, | |
d9436498 | 3571 | int *compaction_retries) |
3250845d VB |
3572 | { |
3573 | int max_retries = MAX_COMPACT_RETRIES; | |
c2033b00 | 3574 | int min_priority; |
65190cff MH |
3575 | bool ret = false; |
3576 | int retries = *compaction_retries; | |
3577 | enum compact_priority priority = *compact_priority; | |
3250845d VB |
3578 | |
3579 | if (!order) | |
3580 | return false; | |
3581 | ||
691d9497 AT |
3582 | if (fatal_signal_pending(current)) |
3583 | return false; | |
3584 | ||
49433085 | 3585 | /* |
ecd8b292 JW |
3586 | * Compaction was skipped due to a lack of free order-0 |
3587 | * migration targets. Continue if reclaim can help. | |
49433085 | 3588 | */ |
ecd8b292 | 3589 | if (compact_result == COMPACT_SKIPPED) { |
49433085 VB |
3590 | ret = compaction_zonelist_suitable(ac, order, alloc_flags); |
3591 | goto out; | |
3592 | } | |
3593 | ||
3250845d | 3594 | /* |
511a69b2 JW |
3595 | * Compaction managed to coalesce some page blocks, but the |
3596 | * allocation failed presumably due to a race. Retry some. | |
3250845d | 3597 | */ |
511a69b2 JW |
3598 | if (compact_result == COMPACT_SUCCESS) { |
3599 | /* | |
3600 | * !costly requests are much more important than | |
3601 | * __GFP_RETRY_MAYFAIL costly ones because they are de | |
3602 | * facto nofail and invoke OOM killer to move on while | |
3603 | * costly can fail and users are ready to cope with | |
3604 | * that. 1/4 retries is rather arbitrary but we would | |
3605 | * need much more detailed feedback from compaction to | |
3606 | * make a better decision. | |
3607 | */ | |
3608 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
3609 | max_retries /= 4; | |
3250845d | 3610 | |
511a69b2 JW |
3611 | if (++(*compaction_retries) <= max_retries) { |
3612 | ret = true; | |
3613 | goto out; | |
3614 | } | |
65190cff | 3615 | } |
3250845d | 3616 | |
d9436498 | 3617 | /* |
511a69b2 | 3618 | * Compaction failed. Retry with increasing priority. |
d9436498 | 3619 | */ |
c2033b00 VB |
3620 | min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ? |
3621 | MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY; | |
65190cff | 3622 | |
c2033b00 | 3623 | if (*compact_priority > min_priority) { |
d9436498 VB |
3624 | (*compact_priority)--; |
3625 | *compaction_retries = 0; | |
65190cff | 3626 | ret = true; |
d9436498 | 3627 | } |
65190cff MH |
3628 | out: |
3629 | trace_compact_retry(order, priority, compact_result, retries, max_retries, ret); | |
3630 | return ret; | |
3250845d | 3631 | } |
56de7263 MG |
3632 | #else |
3633 | static inline struct page * | |
3634 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, | |
c603844b | 3635 | unsigned int alloc_flags, const struct alloc_context *ac, |
a5508cd8 | 3636 | enum compact_priority prio, enum compact_result *compact_result) |
56de7263 | 3637 | { |
33c2d214 | 3638 | *compact_result = COMPACT_SKIPPED; |
56de7263 MG |
3639 | return NULL; |
3640 | } | |
33c2d214 MH |
3641 | |
3642 | static inline bool | |
86a294a8 MH |
3643 | should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags, |
3644 | enum compact_result compact_result, | |
a5508cd8 | 3645 | enum compact_priority *compact_priority, |
d9436498 | 3646 | int *compaction_retries) |
33c2d214 | 3647 | { |
31e49bfd MH |
3648 | struct zone *zone; |
3649 | struct zoneref *z; | |
3650 | ||
3651 | if (!order || order > PAGE_ALLOC_COSTLY_ORDER) | |
3652 | return false; | |
3653 | ||
3654 | /* | |
3655 | * There are setups with compaction disabled which would prefer to loop | |
3656 | * inside the allocator rather than hit the oom killer prematurely. | |
3657 | * Let's give them a good hope and keep retrying while the order-0 | |
3658 | * watermarks are OK. | |
3659 | */ | |
97a225e6 JK |
3660 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
3661 | ac->highest_zoneidx, ac->nodemask) { | |
31e49bfd | 3662 | if (zone_watermark_ok(zone, 0, min_wmark_pages(zone), |
97a225e6 | 3663 | ac->highest_zoneidx, alloc_flags)) |
31e49bfd MH |
3664 | return true; |
3665 | } | |
33c2d214 MH |
3666 | return false; |
3667 | } | |
3250845d | 3668 | #endif /* CONFIG_COMPACTION */ |
56de7263 | 3669 | |
d92a8cfc | 3670 | #ifdef CONFIG_LOCKDEP |
93781325 | 3671 | static struct lockdep_map __fs_reclaim_map = |
d92a8cfc PZ |
3672 | STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map); |
3673 | ||
f920e413 | 3674 | static bool __need_reclaim(gfp_t gfp_mask) |
d92a8cfc | 3675 | { |
d92a8cfc PZ |
3676 | /* no reclaim without waiting on it */ |
3677 | if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) | |
3678 | return false; | |
3679 | ||
3680 | /* this guy won't enter reclaim */ | |
2e517d68 | 3681 | if (current->flags & PF_MEMALLOC) |
d92a8cfc PZ |
3682 | return false; |
3683 | ||
d92a8cfc PZ |
3684 | if (gfp_mask & __GFP_NOLOCKDEP) |
3685 | return false; | |
3686 | ||
3687 | return true; | |
3688 | } | |
3689 | ||
4f3eaf45 | 3690 | void __fs_reclaim_acquire(unsigned long ip) |
93781325 | 3691 | { |
4f3eaf45 | 3692 | lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip); |
93781325 OS |
3693 | } |
3694 | ||
4f3eaf45 | 3695 | void __fs_reclaim_release(unsigned long ip) |
93781325 | 3696 | { |
4f3eaf45 | 3697 | lock_release(&__fs_reclaim_map, ip); |
93781325 OS |
3698 | } |
3699 | ||
d92a8cfc PZ |
3700 | void fs_reclaim_acquire(gfp_t gfp_mask) |
3701 | { | |
f920e413 DV |
3702 | gfp_mask = current_gfp_context(gfp_mask); |
3703 | ||
3704 | if (__need_reclaim(gfp_mask)) { | |
3705 | if (gfp_mask & __GFP_FS) | |
4f3eaf45 | 3706 | __fs_reclaim_acquire(_RET_IP_); |
f920e413 DV |
3707 | |
3708 | #ifdef CONFIG_MMU_NOTIFIER | |
3709 | lock_map_acquire(&__mmu_notifier_invalidate_range_start_map); | |
3710 | lock_map_release(&__mmu_notifier_invalidate_range_start_map); | |
3711 | #endif | |
3712 | ||
3713 | } | |
d92a8cfc PZ |
3714 | } |
3715 | EXPORT_SYMBOL_GPL(fs_reclaim_acquire); | |
3716 | ||
3717 | void fs_reclaim_release(gfp_t gfp_mask) | |
3718 | { | |
f920e413 DV |
3719 | gfp_mask = current_gfp_context(gfp_mask); |
3720 | ||
3721 | if (__need_reclaim(gfp_mask)) { | |
3722 | if (gfp_mask & __GFP_FS) | |
4f3eaf45 | 3723 | __fs_reclaim_release(_RET_IP_); |
f920e413 | 3724 | } |
d92a8cfc PZ |
3725 | } |
3726 | EXPORT_SYMBOL_GPL(fs_reclaim_release); | |
3727 | #endif | |
3728 | ||
3d36424b MG |
3729 | /* |
3730 | * Zonelists may change due to hotplug during allocation. Detect when zonelists | |
3731 | * have been rebuilt so allocation retries. Reader side does not lock and | |
3732 | * retries the allocation if zonelist changes. Writer side is protected by the | |
3733 | * embedded spin_lock. | |
3734 | */ | |
3735 | static DEFINE_SEQLOCK(zonelist_update_seq); | |
3736 | ||
3737 | static unsigned int zonelist_iter_begin(void) | |
3738 | { | |
3739 | if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) | |
3740 | return read_seqbegin(&zonelist_update_seq); | |
3741 | ||
3742 | return 0; | |
3743 | } | |
3744 | ||
3745 | static unsigned int check_retry_zonelist(unsigned int seq) | |
3746 | { | |
3747 | if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE)) | |
3748 | return read_seqretry(&zonelist_update_seq, seq); | |
3749 | ||
3750 | return seq; | |
3751 | } | |
3752 | ||
bba90710 | 3753 | /* Perform direct synchronous page reclaim */ |
2187e17b | 3754 | static unsigned long |
a9263751 VB |
3755 | __perform_reclaim(gfp_t gfp_mask, unsigned int order, |
3756 | const struct alloc_context *ac) | |
11e33f6a | 3757 | { |
499118e9 | 3758 | unsigned int noreclaim_flag; |
fa7fc75f | 3759 | unsigned long progress; |
11e33f6a MG |
3760 | |
3761 | cond_resched(); | |
3762 | ||
3763 | /* We now go into synchronous reclaim */ | |
3764 | cpuset_memory_pressure_bump(); | |
d92a8cfc | 3765 | fs_reclaim_acquire(gfp_mask); |
93781325 | 3766 | noreclaim_flag = memalloc_noreclaim_save(); |
11e33f6a | 3767 | |
a9263751 VB |
3768 | progress = try_to_free_pages(ac->zonelist, order, gfp_mask, |
3769 | ac->nodemask); | |
11e33f6a | 3770 | |
499118e9 | 3771 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 3772 | fs_reclaim_release(gfp_mask); |
11e33f6a MG |
3773 | |
3774 | cond_resched(); | |
3775 | ||
bba90710 MS |
3776 | return progress; |
3777 | } | |
3778 | ||
3779 | /* The really slow allocator path where we enter direct reclaim */ | |
3780 | static inline struct page * | |
3781 | __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, | |
c603844b | 3782 | unsigned int alloc_flags, const struct alloc_context *ac, |
a9263751 | 3783 | unsigned long *did_some_progress) |
bba90710 MS |
3784 | { |
3785 | struct page *page = NULL; | |
fa7fc75f | 3786 | unsigned long pflags; |
bba90710 MS |
3787 | bool drained = false; |
3788 | ||
fa7fc75f | 3789 | psi_memstall_enter(&pflags); |
a9263751 | 3790 | *did_some_progress = __perform_reclaim(gfp_mask, order, ac); |
9ee493ce | 3791 | if (unlikely(!(*did_some_progress))) |
fa7fc75f | 3792 | goto out; |
11e33f6a | 3793 | |
9ee493ce | 3794 | retry: |
31a6c190 | 3795 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); |
9ee493ce MG |
3796 | |
3797 | /* | |
3798 | * If an allocation failed after direct reclaim, it could be because | |
0aaa29a5 | 3799 | * pages are pinned on the per-cpu lists or in high alloc reserves. |
047b9967 | 3800 | * Shrink them and try again |
9ee493ce MG |
3801 | */ |
3802 | if (!page && !drained) { | |
29fac03b | 3803 | unreserve_highatomic_pageblock(ac, false); |
93481ff0 | 3804 | drain_all_pages(NULL); |
9ee493ce MG |
3805 | drained = true; |
3806 | goto retry; | |
3807 | } | |
fa7fc75f SB |
3808 | out: |
3809 | psi_memstall_leave(&pflags); | |
9ee493ce | 3810 | |
11e33f6a MG |
3811 | return page; |
3812 | } | |
3813 | ||
5ecd9d40 DR |
3814 | static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask, |
3815 | const struct alloc_context *ac) | |
3a025760 JW |
3816 | { |
3817 | struct zoneref *z; | |
3818 | struct zone *zone; | |
e1a55637 | 3819 | pg_data_t *last_pgdat = NULL; |
97a225e6 | 3820 | enum zone_type highest_zoneidx = ac->highest_zoneidx; |
3a025760 | 3821 | |
97a225e6 | 3822 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx, |
5ecd9d40 | 3823 | ac->nodemask) { |
bc53008e WY |
3824 | if (!managed_zone(zone)) |
3825 | continue; | |
d137a7cb | 3826 | if (last_pgdat != zone->zone_pgdat) { |
97a225e6 | 3827 | wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx); |
d137a7cb CW |
3828 | last_pgdat = zone->zone_pgdat; |
3829 | } | |
e1a55637 | 3830 | } |
3a025760 JW |
3831 | } |
3832 | ||
c603844b | 3833 | static inline unsigned int |
eb2e2b42 | 3834 | gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order) |
341ce06f | 3835 | { |
c603844b | 3836 | unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; |
1da177e4 | 3837 | |
736838e9 | 3838 | /* |
524c4807 | 3839 | * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE |
736838e9 MN |
3840 | * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD |
3841 | * to save two branches. | |
3842 | */ | |
524c4807 | 3843 | BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE); |
736838e9 | 3844 | BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD); |
933e312e | 3845 | |
341ce06f PZ |
3846 | /* |
3847 | * The caller may dip into page reserves a bit more if the caller | |
3848 | * cannot run direct reclaim, or if the caller has realtime scheduling | |
3849 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will | |
1ebbb218 | 3850 | * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH). |
341ce06f | 3851 | */ |
736838e9 MN |
3852 | alloc_flags |= (__force int) |
3853 | (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM)); | |
1da177e4 | 3854 | |
1ebbb218 | 3855 | if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) { |
5c3240d9 | 3856 | /* |
b104a35d DR |
3857 | * Not worth trying to allocate harder for __GFP_NOMEMALLOC even |
3858 | * if it can't schedule. | |
5c3240d9 | 3859 | */ |
eb2e2b42 | 3860 | if (!(gfp_mask & __GFP_NOMEMALLOC)) { |
1ebbb218 | 3861 | alloc_flags |= ALLOC_NON_BLOCK; |
eb2e2b42 MG |
3862 | |
3863 | if (order > 0) | |
3864 | alloc_flags |= ALLOC_HIGHATOMIC; | |
3865 | } | |
3866 | ||
523b9458 | 3867 | /* |
1ebbb218 MG |
3868 | * Ignore cpuset mems for non-blocking __GFP_HIGH (probably |
3869 | * GFP_ATOMIC) rather than fail, see the comment for | |
8e464522 | 3870 | * cpuset_node_allowed(). |
523b9458 | 3871 | */ |
1ebbb218 MG |
3872 | if (alloc_flags & ALLOC_MIN_RESERVE) |
3873 | alloc_flags &= ~ALLOC_CPUSET; | |
88dc6f20 | 3874 | } else if (unlikely(rt_task(current)) && in_task()) |
c988dcbe | 3875 | alloc_flags |= ALLOC_MIN_RESERVE; |
341ce06f | 3876 | |
8e3560d9 | 3877 | alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags); |
8510e69c | 3878 | |
341ce06f PZ |
3879 | return alloc_flags; |
3880 | } | |
3881 | ||
cd04ae1e | 3882 | static bool oom_reserves_allowed(struct task_struct *tsk) |
072bb0aa | 3883 | { |
cd04ae1e MH |
3884 | if (!tsk_is_oom_victim(tsk)) |
3885 | return false; | |
3886 | ||
3887 | /* | |
3888 | * !MMU doesn't have oom reaper so give access to memory reserves | |
3889 | * only to the thread with TIF_MEMDIE set | |
3890 | */ | |
3891 | if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE)) | |
31a6c190 VB |
3892 | return false; |
3893 | ||
cd04ae1e MH |
3894 | return true; |
3895 | } | |
3896 | ||
3897 | /* | |
3898 | * Distinguish requests which really need access to full memory | |
3899 | * reserves from oom victims which can live with a portion of it | |
3900 | */ | |
3901 | static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask) | |
3902 | { | |
3903 | if (unlikely(gfp_mask & __GFP_NOMEMALLOC)) | |
3904 | return 0; | |
31a6c190 | 3905 | if (gfp_mask & __GFP_MEMALLOC) |
cd04ae1e | 3906 | return ALLOC_NO_WATERMARKS; |
31a6c190 | 3907 | if (in_serving_softirq() && (current->flags & PF_MEMALLOC)) |
cd04ae1e MH |
3908 | return ALLOC_NO_WATERMARKS; |
3909 | if (!in_interrupt()) { | |
3910 | if (current->flags & PF_MEMALLOC) | |
3911 | return ALLOC_NO_WATERMARKS; | |
3912 | else if (oom_reserves_allowed(current)) | |
3913 | return ALLOC_OOM; | |
3914 | } | |
31a6c190 | 3915 | |
cd04ae1e MH |
3916 | return 0; |
3917 | } | |
3918 | ||
3919 | bool gfp_pfmemalloc_allowed(gfp_t gfp_mask) | |
3920 | { | |
3921 | return !!__gfp_pfmemalloc_flags(gfp_mask); | |
072bb0aa MG |
3922 | } |
3923 | ||
0a0337e0 MH |
3924 | /* |
3925 | * Checks whether it makes sense to retry the reclaim to make a forward progress | |
3926 | * for the given allocation request. | |
491d79ae JW |
3927 | * |
3928 | * We give up when we either have tried MAX_RECLAIM_RETRIES in a row | |
3929 | * without success, or when we couldn't even meet the watermark if we | |
3930 | * reclaimed all remaining pages on the LRU lists. | |
0a0337e0 MH |
3931 | * |
3932 | * Returns true if a retry is viable or false to enter the oom path. | |
3933 | */ | |
3934 | static inline bool | |
3935 | should_reclaim_retry(gfp_t gfp_mask, unsigned order, | |
3936 | struct alloc_context *ac, int alloc_flags, | |
423b452e | 3937 | bool did_some_progress, int *no_progress_loops) |
0a0337e0 MH |
3938 | { |
3939 | struct zone *zone; | |
3940 | struct zoneref *z; | |
15f570bf | 3941 | bool ret = false; |
0a0337e0 | 3942 | |
423b452e VB |
3943 | /* |
3944 | * Costly allocations might have made a progress but this doesn't mean | |
3945 | * their order will become available due to high fragmentation so | |
3946 | * always increment the no progress counter for them | |
3947 | */ | |
3948 | if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) | |
3949 | *no_progress_loops = 0; | |
3950 | else | |
3951 | (*no_progress_loops)++; | |
3952 | ||
ac3f3b0a CTK |
3953 | if (*no_progress_loops > MAX_RECLAIM_RETRIES) |
3954 | goto out; | |
3955 | ||
0a0337e0 | 3956 | |
bca67592 MG |
3957 | /* |
3958 | * Keep reclaiming pages while there is a chance this will lead | |
3959 | * somewhere. If none of the target zones can satisfy our allocation | |
3960 | * request even if all reclaimable pages are considered then we are | |
3961 | * screwed and have to go OOM. | |
0a0337e0 | 3962 | */ |
97a225e6 JK |
3963 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, |
3964 | ac->highest_zoneidx, ac->nodemask) { | |
0a0337e0 | 3965 | unsigned long available; |
ede37713 | 3966 | unsigned long reclaimable; |
d379f01d MH |
3967 | unsigned long min_wmark = min_wmark_pages(zone); |
3968 | bool wmark; | |
0a0337e0 | 3969 | |
5a1c84b4 | 3970 | available = reclaimable = zone_reclaimable_pages(zone); |
5a1c84b4 | 3971 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES); |
0a0337e0 MH |
3972 | |
3973 | /* | |
491d79ae JW |
3974 | * Would the allocation succeed if we reclaimed all |
3975 | * reclaimable pages? | |
0a0337e0 | 3976 | */ |
d379f01d | 3977 | wmark = __zone_watermark_ok(zone, order, min_wmark, |
97a225e6 | 3978 | ac->highest_zoneidx, alloc_flags, available); |
d379f01d MH |
3979 | trace_reclaim_retry_zone(z, order, reclaimable, |
3980 | available, min_wmark, *no_progress_loops, wmark); | |
3981 | if (wmark) { | |
15f570bf | 3982 | ret = true; |
132b0d21 | 3983 | break; |
0a0337e0 MH |
3984 | } |
3985 | } | |
3986 | ||
15f570bf MH |
3987 | /* |
3988 | * Memory allocation/reclaim might be called from a WQ context and the | |
3989 | * current implementation of the WQ concurrency control doesn't | |
3990 | * recognize that a particular WQ is congested if the worker thread is | |
3991 | * looping without ever sleeping. Therefore we have to do a short sleep | |
3992 | * here rather than calling cond_resched(). | |
3993 | */ | |
3994 | if (current->flags & PF_WQ_WORKER) | |
3995 | schedule_timeout_uninterruptible(1); | |
3996 | else | |
3997 | cond_resched(); | |
ac3f3b0a CTK |
3998 | out: |
3999 | /* Before OOM, exhaust highatomic_reserve */ | |
4000 | if (!ret) | |
4001 | return unreserve_highatomic_pageblock(ac, true); | |
4002 | ||
15f570bf | 4003 | return ret; |
0a0337e0 MH |
4004 | } |
4005 | ||
902b6281 VB |
4006 | static inline bool |
4007 | check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac) | |
4008 | { | |
4009 | /* | |
4010 | * It's possible that cpuset's mems_allowed and the nodemask from | |
4011 | * mempolicy don't intersect. This should be normally dealt with by | |
4012 | * policy_nodemask(), but it's possible to race with cpuset update in | |
4013 | * such a way the check therein was true, and then it became false | |
4014 | * before we got our cpuset_mems_cookie here. | |
4015 | * This assumes that for all allocations, ac->nodemask can come only | |
4016 | * from MPOL_BIND mempolicy (whose documented semantics is to be ignored | |
4017 | * when it does not intersect with the cpuset restrictions) or the | |
4018 | * caller can deal with a violated nodemask. | |
4019 | */ | |
4020 | if (cpusets_enabled() && ac->nodemask && | |
4021 | !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) { | |
4022 | ac->nodemask = NULL; | |
4023 | return true; | |
4024 | } | |
4025 | ||
4026 | /* | |
4027 | * When updating a task's mems_allowed or mempolicy nodemask, it is | |
4028 | * possible to race with parallel threads in such a way that our | |
4029 | * allocation can fail while the mask is being updated. If we are about | |
4030 | * to fail, check if the cpuset changed during allocation and if so, | |
4031 | * retry. | |
4032 | */ | |
4033 | if (read_mems_allowed_retry(cpuset_mems_cookie)) | |
4034 | return true; | |
4035 | ||
4036 | return false; | |
4037 | } | |
4038 | ||
11e33f6a MG |
4039 | static inline struct page * |
4040 | __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, | |
a9263751 | 4041 | struct alloc_context *ac) |
11e33f6a | 4042 | { |
d0164adc | 4043 | bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM; |
803de900 | 4044 | bool can_compact = gfp_compaction_allowed(gfp_mask); |
282722b0 | 4045 | const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER; |
11e33f6a | 4046 | struct page *page = NULL; |
c603844b | 4047 | unsigned int alloc_flags; |
11e33f6a | 4048 | unsigned long did_some_progress; |
5ce9bfef | 4049 | enum compact_priority compact_priority; |
c5d01d0d | 4050 | enum compact_result compact_result; |
5ce9bfef VB |
4051 | int compaction_retries; |
4052 | int no_progress_loops; | |
5ce9bfef | 4053 | unsigned int cpuset_mems_cookie; |
3d36424b | 4054 | unsigned int zonelist_iter_cookie; |
cd04ae1e | 4055 | int reserve_flags; |
1da177e4 | 4056 | |
3d36424b | 4057 | restart: |
5ce9bfef VB |
4058 | compaction_retries = 0; |
4059 | no_progress_loops = 0; | |
4060 | compact_priority = DEF_COMPACT_PRIORITY; | |
4061 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
3d36424b | 4062 | zonelist_iter_cookie = zonelist_iter_begin(); |
9a67f648 MH |
4063 | |
4064 | /* | |
4065 | * The fast path uses conservative alloc_flags to succeed only until | |
4066 | * kswapd needs to be woken up, and to avoid the cost of setting up | |
4067 | * alloc_flags precisely. So we do that now. | |
4068 | */ | |
eb2e2b42 | 4069 | alloc_flags = gfp_to_alloc_flags(gfp_mask, order); |
9a67f648 | 4070 | |
e47483bc VB |
4071 | /* |
4072 | * We need to recalculate the starting point for the zonelist iterator | |
4073 | * because we might have used different nodemask in the fast path, or | |
4074 | * there was a cpuset modification and we are retrying - otherwise we | |
4075 | * could end up iterating over non-eligible zones endlessly. | |
4076 | */ | |
4077 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist, | |
97a225e6 | 4078 | ac->highest_zoneidx, ac->nodemask); |
e47483bc VB |
4079 | if (!ac->preferred_zoneref->zone) |
4080 | goto nopage; | |
4081 | ||
8ca1b5a4 FT |
4082 | /* |
4083 | * Check for insane configurations where the cpuset doesn't contain | |
4084 | * any suitable zone to satisfy the request - e.g. non-movable | |
4085 | * GFP_HIGHUSER allocations from MOVABLE nodes only. | |
4086 | */ | |
4087 | if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) { | |
4088 | struct zoneref *z = first_zones_zonelist(ac->zonelist, | |
4089 | ac->highest_zoneidx, | |
4090 | &cpuset_current_mems_allowed); | |
4091 | if (!z->zone) | |
4092 | goto nopage; | |
4093 | } | |
4094 | ||
0a79cdad | 4095 | if (alloc_flags & ALLOC_KSWAPD) |
5ecd9d40 | 4096 | wake_all_kswapds(order, gfp_mask, ac); |
23771235 VB |
4097 | |
4098 | /* | |
4099 | * The adjusted alloc_flags might result in immediate success, so try | |
4100 | * that first | |
4101 | */ | |
4102 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); | |
4103 | if (page) | |
4104 | goto got_pg; | |
4105 | ||
a8161d1e VB |
4106 | /* |
4107 | * For costly allocations, try direct compaction first, as it's likely | |
282722b0 VB |
4108 | * that we have enough base pages and don't need to reclaim. For non- |
4109 | * movable high-order allocations, do that as well, as compaction will | |
4110 | * try prevent permanent fragmentation by migrating from blocks of the | |
4111 | * same migratetype. | |
4112 | * Don't try this for allocations that are allowed to ignore | |
4113 | * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen. | |
a8161d1e | 4114 | */ |
803de900 | 4115 | if (can_direct_reclaim && can_compact && |
282722b0 VB |
4116 | (costly_order || |
4117 | (order > 0 && ac->migratetype != MIGRATE_MOVABLE)) | |
4118 | && !gfp_pfmemalloc_allowed(gfp_mask)) { | |
a8161d1e VB |
4119 | page = __alloc_pages_direct_compact(gfp_mask, order, |
4120 | alloc_flags, ac, | |
a5508cd8 | 4121 | INIT_COMPACT_PRIORITY, |
a8161d1e VB |
4122 | &compact_result); |
4123 | if (page) | |
4124 | goto got_pg; | |
4125 | ||
cc638f32 VB |
4126 | /* |
4127 | * Checks for costly allocations with __GFP_NORETRY, which | |
4128 | * includes some THP page fault allocations | |
4129 | */ | |
4130 | if (costly_order && (gfp_mask & __GFP_NORETRY)) { | |
b39d0ee2 DR |
4131 | /* |
4132 | * If allocating entire pageblock(s) and compaction | |
4133 | * failed because all zones are below low watermarks | |
4134 | * or is prohibited because it recently failed at this | |
3f36d866 DR |
4135 | * order, fail immediately unless the allocator has |
4136 | * requested compaction and reclaim retry. | |
b39d0ee2 DR |
4137 | * |
4138 | * Reclaim is | |
4139 | * - potentially very expensive because zones are far | |
4140 | * below their low watermarks or this is part of very | |
4141 | * bursty high order allocations, | |
4142 | * - not guaranteed to help because isolate_freepages() | |
4143 | * may not iterate over freed pages as part of its | |
4144 | * linear scan, and | |
4145 | * - unlikely to make entire pageblocks free on its | |
4146 | * own. | |
4147 | */ | |
4148 | if (compact_result == COMPACT_SKIPPED || | |
4149 | compact_result == COMPACT_DEFERRED) | |
4150 | goto nopage; | |
a8161d1e | 4151 | |
a8161d1e | 4152 | /* |
3eb2771b VB |
4153 | * Looks like reclaim/compaction is worth trying, but |
4154 | * sync compaction could be very expensive, so keep | |
25160354 | 4155 | * using async compaction. |
a8161d1e | 4156 | */ |
a5508cd8 | 4157 | compact_priority = INIT_COMPACT_PRIORITY; |
a8161d1e VB |
4158 | } |
4159 | } | |
23771235 | 4160 | |
31a6c190 | 4161 | retry: |
23771235 | 4162 | /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */ |
0a79cdad | 4163 | if (alloc_flags & ALLOC_KSWAPD) |
5ecd9d40 | 4164 | wake_all_kswapds(order, gfp_mask, ac); |
31a6c190 | 4165 | |
cd04ae1e MH |
4166 | reserve_flags = __gfp_pfmemalloc_flags(gfp_mask); |
4167 | if (reserve_flags) | |
ce96fa62 ML |
4168 | alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) | |
4169 | (alloc_flags & ALLOC_KSWAPD); | |
23771235 | 4170 | |
e46e7b77 | 4171 | /* |
d6a24df0 VB |
4172 | * Reset the nodemask and zonelist iterators if memory policies can be |
4173 | * ignored. These allocations are high priority and system rather than | |
4174 | * user oriented. | |
e46e7b77 | 4175 | */ |
cd04ae1e | 4176 | if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) { |
d6a24df0 | 4177 | ac->nodemask = NULL; |
e46e7b77 | 4178 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist, |
97a225e6 | 4179 | ac->highest_zoneidx, ac->nodemask); |
e46e7b77 MG |
4180 | } |
4181 | ||
23771235 | 4182 | /* Attempt with potentially adjusted zonelist and alloc_flags */ |
31a6c190 | 4183 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); |
7fb1d9fc RS |
4184 | if (page) |
4185 | goto got_pg; | |
1da177e4 | 4186 | |
d0164adc | 4187 | /* Caller is not willing to reclaim, we can't balance anything */ |
9a67f648 | 4188 | if (!can_direct_reclaim) |
1da177e4 LT |
4189 | goto nopage; |
4190 | ||
9a67f648 MH |
4191 | /* Avoid recursion of direct reclaim */ |
4192 | if (current->flags & PF_MEMALLOC) | |
6583bb64 DR |
4193 | goto nopage; |
4194 | ||
a8161d1e VB |
4195 | /* Try direct reclaim and then allocating */ |
4196 | page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac, | |
4197 | &did_some_progress); | |
4198 | if (page) | |
4199 | goto got_pg; | |
4200 | ||
4201 | /* Try direct compaction and then allocating */ | |
a9263751 | 4202 | page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac, |
a5508cd8 | 4203 | compact_priority, &compact_result); |
56de7263 MG |
4204 | if (page) |
4205 | goto got_pg; | |
75f30861 | 4206 | |
9083905a JW |
4207 | /* Do not loop if specifically requested */ |
4208 | if (gfp_mask & __GFP_NORETRY) | |
a8161d1e | 4209 | goto nopage; |
9083905a | 4210 | |
0a0337e0 MH |
4211 | /* |
4212 | * Do not retry costly high order allocations unless they are | |
803de900 | 4213 | * __GFP_RETRY_MAYFAIL and we can compact |
0a0337e0 | 4214 | */ |
803de900 VB |
4215 | if (costly_order && (!can_compact || |
4216 | !(gfp_mask & __GFP_RETRY_MAYFAIL))) | |
a8161d1e | 4217 | goto nopage; |
0a0337e0 | 4218 | |
0a0337e0 | 4219 | if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags, |
423b452e | 4220 | did_some_progress > 0, &no_progress_loops)) |
0a0337e0 MH |
4221 | goto retry; |
4222 | ||
33c2d214 MH |
4223 | /* |
4224 | * It doesn't make any sense to retry for the compaction if the order-0 | |
4225 | * reclaim is not able to make any progress because the current | |
4226 | * implementation of the compaction depends on the sufficient amount | |
4227 | * of free memory (see __compaction_suitable) | |
4228 | */ | |
803de900 | 4229 | if (did_some_progress > 0 && can_compact && |
86a294a8 | 4230 | should_compact_retry(ac, order, alloc_flags, |
a5508cd8 | 4231 | compact_result, &compact_priority, |
d9436498 | 4232 | &compaction_retries)) |
33c2d214 MH |
4233 | goto retry; |
4234 | ||
902b6281 | 4235 | |
3d36424b MG |
4236 | /* |
4237 | * Deal with possible cpuset update races or zonelist updates to avoid | |
4238 | * a unnecessary OOM kill. | |
4239 | */ | |
4240 | if (check_retry_cpuset(cpuset_mems_cookie, ac) || | |
4241 | check_retry_zonelist(zonelist_iter_cookie)) | |
4242 | goto restart; | |
e47483bc | 4243 | |
9083905a JW |
4244 | /* Reclaim has failed us, start killing things */ |
4245 | page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress); | |
4246 | if (page) | |
4247 | goto got_pg; | |
4248 | ||
9a67f648 | 4249 | /* Avoid allocations with no watermarks from looping endlessly */ |
cd04ae1e | 4250 | if (tsk_is_oom_victim(current) && |
8510e69c | 4251 | (alloc_flags & ALLOC_OOM || |
c288983d | 4252 | (gfp_mask & __GFP_NOMEMALLOC))) |
9a67f648 MH |
4253 | goto nopage; |
4254 | ||
9083905a | 4255 | /* Retry as long as the OOM killer is making progress */ |
0a0337e0 MH |
4256 | if (did_some_progress) { |
4257 | no_progress_loops = 0; | |
9083905a | 4258 | goto retry; |
0a0337e0 | 4259 | } |
9083905a | 4260 | |
1da177e4 | 4261 | nopage: |
3d36424b MG |
4262 | /* |
4263 | * Deal with possible cpuset update races or zonelist updates to avoid | |
4264 | * a unnecessary OOM kill. | |
4265 | */ | |
4266 | if (check_retry_cpuset(cpuset_mems_cookie, ac) || | |
4267 | check_retry_zonelist(zonelist_iter_cookie)) | |
4268 | goto restart; | |
5ce9bfef | 4269 | |
9a67f648 MH |
4270 | /* |
4271 | * Make sure that __GFP_NOFAIL request doesn't leak out and make sure | |
4272 | * we always retry | |
4273 | */ | |
4274 | if (gfp_mask & __GFP_NOFAIL) { | |
4275 | /* | |
4276 | * All existing users of the __GFP_NOFAIL are blockable, so warn | |
4277 | * of any new users that actually require GFP_NOWAIT | |
4278 | */ | |
3f913fc5 | 4279 | if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask)) |
9a67f648 MH |
4280 | goto fail; |
4281 | ||
4282 | /* | |
4283 | * PF_MEMALLOC request from this context is rather bizarre | |
4284 | * because we cannot reclaim anything and only can loop waiting | |
4285 | * for somebody to do a work for us | |
4286 | */ | |
3f913fc5 | 4287 | WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask); |
9a67f648 MH |
4288 | |
4289 | /* | |
4290 | * non failing costly orders are a hard requirement which we | |
4291 | * are not prepared for much so let's warn about these users | |
4292 | * so that we can identify them and convert them to something | |
4293 | * else. | |
4294 | */ | |
896c4d52 | 4295 | WARN_ON_ONCE_GFP(costly_order, gfp_mask); |
9a67f648 | 4296 | |
6c18ba7a | 4297 | /* |
1ebbb218 MG |
4298 | * Help non-failing allocations by giving some access to memory |
4299 | * reserves normally used for high priority non-blocking | |
4300 | * allocations but do not use ALLOC_NO_WATERMARKS because this | |
6c18ba7a | 4301 | * could deplete whole memory reserves which would just make |
1ebbb218 | 4302 | * the situation worse. |
6c18ba7a | 4303 | */ |
1ebbb218 | 4304 | page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac); |
6c18ba7a MH |
4305 | if (page) |
4306 | goto got_pg; | |
4307 | ||
9a67f648 MH |
4308 | cond_resched(); |
4309 | goto retry; | |
4310 | } | |
4311 | fail: | |
a8e99259 | 4312 | warn_alloc(gfp_mask, ac->nodemask, |
7877cdcc | 4313 | "page allocation failure: order:%u", order); |
1da177e4 | 4314 | got_pg: |
072bb0aa | 4315 | return page; |
1da177e4 | 4316 | } |
11e33f6a | 4317 | |
9cd75558 | 4318 | static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order, |
04ec6264 | 4319 | int preferred_nid, nodemask_t *nodemask, |
8e6a930b | 4320 | struct alloc_context *ac, gfp_t *alloc_gfp, |
9cd75558 | 4321 | unsigned int *alloc_flags) |
11e33f6a | 4322 | { |
97a225e6 | 4323 | ac->highest_zoneidx = gfp_zone(gfp_mask); |
04ec6264 | 4324 | ac->zonelist = node_zonelist(preferred_nid, gfp_mask); |
9cd75558 | 4325 | ac->nodemask = nodemask; |
01c0bfe0 | 4326 | ac->migratetype = gfp_migratetype(gfp_mask); |
11e33f6a | 4327 | |
682a3385 | 4328 | if (cpusets_enabled()) { |
8e6a930b | 4329 | *alloc_gfp |= __GFP_HARDWALL; |
182f3d7a MS |
4330 | /* |
4331 | * When we are in the interrupt context, it is irrelevant | |
4332 | * to the current task context. It means that any node ok. | |
4333 | */ | |
88dc6f20 | 4334 | if (in_task() && !ac->nodemask) |
9cd75558 | 4335 | ac->nodemask = &cpuset_current_mems_allowed; |
51047820 VB |
4336 | else |
4337 | *alloc_flags |= ALLOC_CPUSET; | |
682a3385 MG |
4338 | } |
4339 | ||
446ec838 | 4340 | might_alloc(gfp_mask); |
11e33f6a MG |
4341 | |
4342 | if (should_fail_alloc_page(gfp_mask, order)) | |
9cd75558 | 4343 | return false; |
11e33f6a | 4344 | |
8e3560d9 | 4345 | *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags); |
d883c6cf | 4346 | |
c9ab0c4f | 4347 | /* Dirty zone balancing only done in the fast path */ |
9cd75558 | 4348 | ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE); |
c9ab0c4f | 4349 | |
e46e7b77 MG |
4350 | /* |
4351 | * The preferred zone is used for statistics but crucially it is | |
4352 | * also used as the starting point for the zonelist iterator. It | |
4353 | * may get reset for allocations that ignore memory policies. | |
4354 | */ | |
9cd75558 | 4355 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist, |
97a225e6 | 4356 | ac->highest_zoneidx, ac->nodemask); |
a0622d05 MN |
4357 | |
4358 | return true; | |
9cd75558 MG |
4359 | } |
4360 | ||
387ba26f | 4361 | /* |
0f87d9d3 | 4362 | * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array |
387ba26f MG |
4363 | * @gfp: GFP flags for the allocation |
4364 | * @preferred_nid: The preferred NUMA node ID to allocate from | |
4365 | * @nodemask: Set of nodes to allocate from, may be NULL | |
0f87d9d3 MG |
4366 | * @nr_pages: The number of pages desired on the list or array |
4367 | * @page_list: Optional list to store the allocated pages | |
4368 | * @page_array: Optional array to store the pages | |
387ba26f MG |
4369 | * |
4370 | * This is a batched version of the page allocator that attempts to | |
0f87d9d3 MG |
4371 | * allocate nr_pages quickly. Pages are added to page_list if page_list |
4372 | * is not NULL, otherwise it is assumed that the page_array is valid. | |
387ba26f | 4373 | * |
0f87d9d3 MG |
4374 | * For lists, nr_pages is the number of pages that should be allocated. |
4375 | * | |
4376 | * For arrays, only NULL elements are populated with pages and nr_pages | |
4377 | * is the maximum number of pages that will be stored in the array. | |
4378 | * | |
4379 | * Returns the number of pages on the list or array. | |
387ba26f MG |
4380 | */ |
4381 | unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid, | |
4382 | nodemask_t *nodemask, int nr_pages, | |
0f87d9d3 MG |
4383 | struct list_head *page_list, |
4384 | struct page **page_array) | |
387ba26f MG |
4385 | { |
4386 | struct page *page; | |
4b23a68f | 4387 | unsigned long __maybe_unused UP_flags; |
387ba26f MG |
4388 | struct zone *zone; |
4389 | struct zoneref *z; | |
4390 | struct per_cpu_pages *pcp; | |
4391 | struct list_head *pcp_list; | |
4392 | struct alloc_context ac; | |
4393 | gfp_t alloc_gfp; | |
4394 | unsigned int alloc_flags = ALLOC_WMARK_LOW; | |
3e23060b | 4395 | int nr_populated = 0, nr_account = 0; |
387ba26f | 4396 | |
0f87d9d3 MG |
4397 | /* |
4398 | * Skip populated array elements to determine if any pages need | |
4399 | * to be allocated before disabling IRQs. | |
4400 | */ | |
b08e50dd | 4401 | while (page_array && nr_populated < nr_pages && page_array[nr_populated]) |
0f87d9d3 MG |
4402 | nr_populated++; |
4403 | ||
06147843 CL |
4404 | /* No pages requested? */ |
4405 | if (unlikely(nr_pages <= 0)) | |
4406 | goto out; | |
4407 | ||
b3b64ebd MG |
4408 | /* Already populated array? */ |
4409 | if (unlikely(page_array && nr_pages - nr_populated == 0)) | |
06147843 | 4410 | goto out; |
b3b64ebd | 4411 | |
8dcb3060 | 4412 | /* Bulk allocator does not support memcg accounting. */ |
f7a449f7 | 4413 | if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT)) |
8dcb3060 SB |
4414 | goto failed; |
4415 | ||
387ba26f | 4416 | /* Use the single page allocator for one page. */ |
0f87d9d3 | 4417 | if (nr_pages - nr_populated == 1) |
387ba26f MG |
4418 | goto failed; |
4419 | ||
187ad460 MG |
4420 | #ifdef CONFIG_PAGE_OWNER |
4421 | /* | |
4422 | * PAGE_OWNER may recurse into the allocator to allocate space to | |
4423 | * save the stack with pagesets.lock held. Releasing/reacquiring | |
4424 | * removes much of the performance benefit of bulk allocation so | |
4425 | * force the caller to allocate one page at a time as it'll have | |
4426 | * similar performance to added complexity to the bulk allocator. | |
4427 | */ | |
4428 | if (static_branch_unlikely(&page_owner_inited)) | |
4429 | goto failed; | |
4430 | #endif | |
4431 | ||
387ba26f MG |
4432 | /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */ |
4433 | gfp &= gfp_allowed_mask; | |
4434 | alloc_gfp = gfp; | |
4435 | if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags)) | |
06147843 | 4436 | goto out; |
387ba26f MG |
4437 | gfp = alloc_gfp; |
4438 | ||
4439 | /* Find an allowed local zone that meets the low watermark. */ | |
4440 | for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) { | |
4441 | unsigned long mark; | |
4442 | ||
4443 | if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) && | |
4444 | !__cpuset_zone_allowed(zone, gfp)) { | |
4445 | continue; | |
4446 | } | |
4447 | ||
4448 | if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone && | |
4449 | zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) { | |
4450 | goto failed; | |
4451 | } | |
4452 | ||
4453 | mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages; | |
4454 | if (zone_watermark_fast(zone, 0, mark, | |
4455 | zonelist_zone_idx(ac.preferred_zoneref), | |
4456 | alloc_flags, gfp)) { | |
4457 | break; | |
4458 | } | |
4459 | } | |
4460 | ||
4461 | /* | |
4462 | * If there are no allowed local zones that meets the watermarks then | |
4463 | * try to allocate a single page and reclaim if necessary. | |
4464 | */ | |
ce76f9a1 | 4465 | if (unlikely(!zone)) |
387ba26f MG |
4466 | goto failed; |
4467 | ||
57490774 | 4468 | /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */ |
4b23a68f | 4469 | pcp_trylock_prepare(UP_flags); |
57490774 | 4470 | pcp = pcp_spin_trylock(zone->per_cpu_pageset); |
01b44456 | 4471 | if (!pcp) |
4b23a68f | 4472 | goto failed_irq; |
387ba26f | 4473 | |
387ba26f | 4474 | /* Attempt the batch allocation */ |
44042b44 | 4475 | pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)]; |
0f87d9d3 MG |
4476 | while (nr_populated < nr_pages) { |
4477 | ||
4478 | /* Skip existing pages */ | |
4479 | if (page_array && page_array[nr_populated]) { | |
4480 | nr_populated++; | |
4481 | continue; | |
4482 | } | |
4483 | ||
44042b44 | 4484 | page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags, |
387ba26f | 4485 | pcp, pcp_list); |
ce76f9a1 | 4486 | if (unlikely(!page)) { |
c572e488 | 4487 | /* Try and allocate at least one page */ |
4b23a68f | 4488 | if (!nr_account) { |
57490774 | 4489 | pcp_spin_unlock(pcp); |
387ba26f | 4490 | goto failed_irq; |
4b23a68f | 4491 | } |
387ba26f MG |
4492 | break; |
4493 | } | |
3e23060b | 4494 | nr_account++; |
387ba26f MG |
4495 | |
4496 | prep_new_page(page, 0, gfp, 0); | |
0f87d9d3 MG |
4497 | if (page_list) |
4498 | list_add(&page->lru, page_list); | |
4499 | else | |
4500 | page_array[nr_populated] = page; | |
4501 | nr_populated++; | |
387ba26f MG |
4502 | } |
4503 | ||
57490774 | 4504 | pcp_spin_unlock(pcp); |
4b23a68f | 4505 | pcp_trylock_finish(UP_flags); |
43c95bcc | 4506 | |
3e23060b MG |
4507 | __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account); |
4508 | zone_statistics(ac.preferred_zoneref->zone, zone, nr_account); | |
387ba26f | 4509 | |
06147843 | 4510 | out: |
0f87d9d3 | 4511 | return nr_populated; |
387ba26f MG |
4512 | |
4513 | failed_irq: | |
4b23a68f | 4514 | pcp_trylock_finish(UP_flags); |
387ba26f MG |
4515 | |
4516 | failed: | |
4517 | page = __alloc_pages(gfp, 0, preferred_nid, nodemask); | |
4518 | if (page) { | |
0f87d9d3 MG |
4519 | if (page_list) |
4520 | list_add(&page->lru, page_list); | |
4521 | else | |
4522 | page_array[nr_populated] = page; | |
4523 | nr_populated++; | |
387ba26f MG |
4524 | } |
4525 | ||
06147843 | 4526 | goto out; |
387ba26f MG |
4527 | } |
4528 | EXPORT_SYMBOL_GPL(__alloc_pages_bulk); | |
4529 | ||
9cd75558 MG |
4530 | /* |
4531 | * This is the 'heart' of the zoned buddy allocator. | |
4532 | */ | |
84172f4b | 4533 | struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid, |
04ec6264 | 4534 | nodemask_t *nodemask) |
9cd75558 MG |
4535 | { |
4536 | struct page *page; | |
4537 | unsigned int alloc_flags = ALLOC_WMARK_LOW; | |
8e6a930b | 4538 | gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */ |
9cd75558 MG |
4539 | struct alloc_context ac = { }; |
4540 | ||
c63ae43b MH |
4541 | /* |
4542 | * There are several places where we assume that the order value is sane | |
4543 | * so bail out early if the request is out of bound. | |
4544 | */ | |
5e0a760b | 4545 | if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp)) |
c63ae43b | 4546 | return NULL; |
c63ae43b | 4547 | |
6e5e0f28 | 4548 | gfp &= gfp_allowed_mask; |
da6df1b0 PT |
4549 | /* |
4550 | * Apply scoped allocation constraints. This is mainly about GFP_NOFS | |
4551 | * resp. GFP_NOIO which has to be inherited for all allocation requests | |
4552 | * from a particular context which has been marked by | |
8e3560d9 PT |
4553 | * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures |
4554 | * movable zones are not used during allocation. | |
da6df1b0 PT |
4555 | */ |
4556 | gfp = current_gfp_context(gfp); | |
6e5e0f28 MWO |
4557 | alloc_gfp = gfp; |
4558 | if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac, | |
8e6a930b | 4559 | &alloc_gfp, &alloc_flags)) |
9cd75558 MG |
4560 | return NULL; |
4561 | ||
6bb15450 MG |
4562 | /* |
4563 | * Forbid the first pass from falling back to types that fragment | |
4564 | * memory until all local zones are considered. | |
4565 | */ | |
6e5e0f28 | 4566 | alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp); |
6bb15450 | 4567 | |
5117f45d | 4568 | /* First allocation attempt */ |
8e6a930b | 4569 | page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac); |
4fcb0971 MG |
4570 | if (likely(page)) |
4571 | goto out; | |
11e33f6a | 4572 | |
da6df1b0 | 4573 | alloc_gfp = gfp; |
4fcb0971 | 4574 | ac.spread_dirty_pages = false; |
23f086f9 | 4575 | |
4741526b MG |
4576 | /* |
4577 | * Restore the original nodemask if it was potentially replaced with | |
4578 | * &cpuset_current_mems_allowed to optimize the fast-path attempt. | |
4579 | */ | |
97ce86f9 | 4580 | ac.nodemask = nodemask; |
16096c25 | 4581 | |
8e6a930b | 4582 | page = __alloc_pages_slowpath(alloc_gfp, order, &ac); |
cc9a6c87 | 4583 | |
4fcb0971 | 4584 | out: |
f7a449f7 | 4585 | if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page && |
6e5e0f28 | 4586 | unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) { |
c4159a75 VD |
4587 | __free_pages(page, order); |
4588 | page = NULL; | |
4949148a VD |
4589 | } |
4590 | ||
8e6a930b | 4591 | trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype); |
b073d7f8 | 4592 | kmsan_alloc_page(page, order, alloc_gfp); |
4fcb0971 | 4593 | |
11e33f6a | 4594 | return page; |
1da177e4 | 4595 | } |
84172f4b | 4596 | EXPORT_SYMBOL(__alloc_pages); |
1da177e4 | 4597 | |
cc09cb13 MWO |
4598 | struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid, |
4599 | nodemask_t *nodemask) | |
4600 | { | |
4601 | struct page *page = __alloc_pages(gfp | __GFP_COMP, order, | |
23e48832 HD |
4602 | preferred_nid, nodemask); |
4603 | return page_rmappable_folio(page); | |
cc09cb13 MWO |
4604 | } |
4605 | EXPORT_SYMBOL(__folio_alloc); | |
4606 | ||
1da177e4 | 4607 | /* |
9ea9a680 MH |
4608 | * Common helper functions. Never use with __GFP_HIGHMEM because the returned |
4609 | * address cannot represent highmem pages. Use alloc_pages and then kmap if | |
4610 | * you need to access high mem. | |
1da177e4 | 4611 | */ |
920c7a5d | 4612 | unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) |
1da177e4 | 4613 | { |
945a1113 AM |
4614 | struct page *page; |
4615 | ||
9ea9a680 | 4616 | page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order); |
1da177e4 LT |
4617 | if (!page) |
4618 | return 0; | |
4619 | return (unsigned long) page_address(page); | |
4620 | } | |
1da177e4 LT |
4621 | EXPORT_SYMBOL(__get_free_pages); |
4622 | ||
920c7a5d | 4623 | unsigned long get_zeroed_page(gfp_t gfp_mask) |
1da177e4 | 4624 | { |
dcc1be11 | 4625 | return __get_free_page(gfp_mask | __GFP_ZERO); |
1da177e4 | 4626 | } |
1da177e4 LT |
4627 | EXPORT_SYMBOL(get_zeroed_page); |
4628 | ||
7f194fbb MWO |
4629 | /** |
4630 | * __free_pages - Free pages allocated with alloc_pages(). | |
4631 | * @page: The page pointer returned from alloc_pages(). | |
4632 | * @order: The order of the allocation. | |
4633 | * | |
4634 | * This function can free multi-page allocations that are not compound | |
4635 | * pages. It does not check that the @order passed in matches that of | |
4636 | * the allocation, so it is easy to leak memory. Freeing more memory | |
4637 | * than was allocated will probably emit a warning. | |
4638 | * | |
4639 | * If the last reference to this page is speculative, it will be released | |
4640 | * by put_page() which only frees the first page of a non-compound | |
4641 | * allocation. To prevent the remaining pages from being leaked, we free | |
4642 | * the subsequent pages here. If you want to use the page's reference | |
4643 | * count to decide when to free the allocation, you should allocate a | |
4644 | * compound page, and use put_page() instead of __free_pages(). | |
4645 | * | |
4646 | * Context: May be called in interrupt context or while holding a normal | |
4647 | * spinlock, but not in NMI context or while holding a raw spinlock. | |
4648 | */ | |
742aa7fb AL |
4649 | void __free_pages(struct page *page, unsigned int order) |
4650 | { | |
462a8e08 DC |
4651 | /* get PageHead before we drop reference */ |
4652 | int head = PageHead(page); | |
4653 | ||
742aa7fb AL |
4654 | if (put_page_testzero(page)) |
4655 | free_the_page(page, order); | |
462a8e08 | 4656 | else if (!head) |
e320d301 MWO |
4657 | while (order-- > 0) |
4658 | free_the_page(page + (1 << order), order); | |
742aa7fb | 4659 | } |
1da177e4 LT |
4660 | EXPORT_SYMBOL(__free_pages); |
4661 | ||
920c7a5d | 4662 | void free_pages(unsigned long addr, unsigned int order) |
1da177e4 LT |
4663 | { |
4664 | if (addr != 0) { | |
725d704e | 4665 | VM_BUG_ON(!virt_addr_valid((void *)addr)); |
1da177e4 LT |
4666 | __free_pages(virt_to_page((void *)addr), order); |
4667 | } | |
4668 | } | |
4669 | ||
4670 | EXPORT_SYMBOL(free_pages); | |
4671 | ||
b63ae8ca AD |
4672 | /* |
4673 | * Page Fragment: | |
4674 | * An arbitrary-length arbitrary-offset area of memory which resides | |
4675 | * within a 0 or higher order page. Multiple fragments within that page | |
4676 | * are individually refcounted, in the page's reference counter. | |
4677 | * | |
4678 | * The page_frag functions below provide a simple allocation framework for | |
4679 | * page fragments. This is used by the network stack and network device | |
4680 | * drivers to provide a backing region of memory for use as either an | |
4681 | * sk_buff->head, or to be used in the "frags" portion of skb_shared_info. | |
4682 | */ | |
2976db80 AD |
4683 | static struct page *__page_frag_cache_refill(struct page_frag_cache *nc, |
4684 | gfp_t gfp_mask) | |
b63ae8ca AD |
4685 | { |
4686 | struct page *page = NULL; | |
4687 | gfp_t gfp = gfp_mask; | |
4688 | ||
4689 | #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) | |
4690 | gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY | | |
4691 | __GFP_NOMEMALLOC; | |
4692 | page = alloc_pages_node(NUMA_NO_NODE, gfp_mask, | |
4693 | PAGE_FRAG_CACHE_MAX_ORDER); | |
4694 | nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE; | |
4695 | #endif | |
4696 | if (unlikely(!page)) | |
4697 | page = alloc_pages_node(NUMA_NO_NODE, gfp, 0); | |
4698 | ||
4699 | nc->va = page ? page_address(page) : NULL; | |
4700 | ||
4701 | return page; | |
4702 | } | |
4703 | ||
2976db80 | 4704 | void __page_frag_cache_drain(struct page *page, unsigned int count) |
44fdffd7 AD |
4705 | { |
4706 | VM_BUG_ON_PAGE(page_ref_count(page) == 0, page); | |
4707 | ||
742aa7fb AL |
4708 | if (page_ref_sub_and_test(page, count)) |
4709 | free_the_page(page, compound_order(page)); | |
44fdffd7 | 4710 | } |
2976db80 | 4711 | EXPORT_SYMBOL(__page_frag_cache_drain); |
44fdffd7 | 4712 | |
b358e212 KH |
4713 | void *page_frag_alloc_align(struct page_frag_cache *nc, |
4714 | unsigned int fragsz, gfp_t gfp_mask, | |
4715 | unsigned int align_mask) | |
b63ae8ca AD |
4716 | { |
4717 | unsigned int size = PAGE_SIZE; | |
4718 | struct page *page; | |
4719 | int offset; | |
4720 | ||
4721 | if (unlikely(!nc->va)) { | |
4722 | refill: | |
2976db80 | 4723 | page = __page_frag_cache_refill(nc, gfp_mask); |
b63ae8ca AD |
4724 | if (!page) |
4725 | return NULL; | |
4726 | ||
4727 | #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) | |
4728 | /* if size can vary use size else just use PAGE_SIZE */ | |
4729 | size = nc->size; | |
4730 | #endif | |
4731 | /* Even if we own the page, we do not use atomic_set(). | |
4732 | * This would break get_page_unless_zero() users. | |
4733 | */ | |
86447726 | 4734 | page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE); |
b63ae8ca AD |
4735 | |
4736 | /* reset page count bias and offset to start of new frag */ | |
2f064f34 | 4737 | nc->pfmemalloc = page_is_pfmemalloc(page); |
86447726 | 4738 | nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1; |
b63ae8ca AD |
4739 | nc->offset = size; |
4740 | } | |
4741 | ||
4742 | offset = nc->offset - fragsz; | |
4743 | if (unlikely(offset < 0)) { | |
4744 | page = virt_to_page(nc->va); | |
4745 | ||
fe896d18 | 4746 | if (!page_ref_sub_and_test(page, nc->pagecnt_bias)) |
b63ae8ca AD |
4747 | goto refill; |
4748 | ||
d8c19014 DZ |
4749 | if (unlikely(nc->pfmemalloc)) { |
4750 | free_the_page(page, compound_order(page)); | |
4751 | goto refill; | |
4752 | } | |
4753 | ||
b63ae8ca AD |
4754 | #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) |
4755 | /* if size can vary use size else just use PAGE_SIZE */ | |
4756 | size = nc->size; | |
4757 | #endif | |
4758 | /* OK, page count is 0, we can safely set it */ | |
86447726 | 4759 | set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1); |
b63ae8ca AD |
4760 | |
4761 | /* reset page count bias and offset to start of new frag */ | |
86447726 | 4762 | nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1; |
b63ae8ca | 4763 | offset = size - fragsz; |
dac22531 ML |
4764 | if (unlikely(offset < 0)) { |
4765 | /* | |
4766 | * The caller is trying to allocate a fragment | |
4767 | * with fragsz > PAGE_SIZE but the cache isn't big | |
4768 | * enough to satisfy the request, this may | |
4769 | * happen in low memory conditions. | |
4770 | * We don't release the cache page because | |
4771 | * it could make memory pressure worse | |
4772 | * so we simply return NULL here. | |
4773 | */ | |
4774 | return NULL; | |
4775 | } | |
b63ae8ca AD |
4776 | } |
4777 | ||
4778 | nc->pagecnt_bias--; | |
b358e212 | 4779 | offset &= align_mask; |
b63ae8ca AD |
4780 | nc->offset = offset; |
4781 | ||
4782 | return nc->va + offset; | |
4783 | } | |
b358e212 | 4784 | EXPORT_SYMBOL(page_frag_alloc_align); |
b63ae8ca AD |
4785 | |
4786 | /* | |
4787 | * Frees a page fragment allocated out of either a compound or order 0 page. | |
4788 | */ | |
8c2dd3e4 | 4789 | void page_frag_free(void *addr) |
b63ae8ca AD |
4790 | { |
4791 | struct page *page = virt_to_head_page(addr); | |
4792 | ||
742aa7fb AL |
4793 | if (unlikely(put_page_testzero(page))) |
4794 | free_the_page(page, compound_order(page)); | |
b63ae8ca | 4795 | } |
8c2dd3e4 | 4796 | EXPORT_SYMBOL(page_frag_free); |
b63ae8ca | 4797 | |
d00181b9 KS |
4798 | static void *make_alloc_exact(unsigned long addr, unsigned int order, |
4799 | size_t size) | |
ee85c2e1 AK |
4800 | { |
4801 | if (addr) { | |
df48a5f7 LH |
4802 | unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE); |
4803 | struct page *page = virt_to_page((void *)addr); | |
4804 | struct page *last = page + nr; | |
4805 | ||
4806 | split_page_owner(page, 1 << order); | |
4807 | split_page_memcg(page, 1 << order); | |
4808 | while (page < --last) | |
4809 | set_page_refcounted(last); | |
4810 | ||
4811 | last = page + (1UL << order); | |
4812 | for (page += nr; page < last; page++) | |
4813 | __free_pages_ok(page, 0, FPI_TO_TAIL); | |
ee85c2e1 AK |
4814 | } |
4815 | return (void *)addr; | |
4816 | } | |
4817 | ||
2be0ffe2 TT |
4818 | /** |
4819 | * alloc_pages_exact - allocate an exact number physically-contiguous pages. | |
4820 | * @size: the number of bytes to allocate | |
63931eb9 | 4821 | * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP |
2be0ffe2 TT |
4822 | * |
4823 | * This function is similar to alloc_pages(), except that it allocates the | |
4824 | * minimum number of pages to satisfy the request. alloc_pages() can only | |
4825 | * allocate memory in power-of-two pages. | |
4826 | * | |
5e0a760b | 4827 | * This function is also limited by MAX_PAGE_ORDER. |
2be0ffe2 TT |
4828 | * |
4829 | * Memory allocated by this function must be released by free_pages_exact(). | |
a862f68a MR |
4830 | * |
4831 | * Return: pointer to the allocated area or %NULL in case of error. | |
2be0ffe2 TT |
4832 | */ |
4833 | void *alloc_pages_exact(size_t size, gfp_t gfp_mask) | |
4834 | { | |
4835 | unsigned int order = get_order(size); | |
4836 | unsigned long addr; | |
4837 | ||
ba7f1b9e ML |
4838 | if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM))) |
4839 | gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM); | |
63931eb9 | 4840 | |
2be0ffe2 | 4841 | addr = __get_free_pages(gfp_mask, order); |
ee85c2e1 | 4842 | return make_alloc_exact(addr, order, size); |
2be0ffe2 TT |
4843 | } |
4844 | EXPORT_SYMBOL(alloc_pages_exact); | |
4845 | ||
ee85c2e1 AK |
4846 | /** |
4847 | * alloc_pages_exact_nid - allocate an exact number of physically-contiguous | |
4848 | * pages on a node. | |
b5e6ab58 | 4849 | * @nid: the preferred node ID where memory should be allocated |
ee85c2e1 | 4850 | * @size: the number of bytes to allocate |
63931eb9 | 4851 | * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP |
ee85c2e1 AK |
4852 | * |
4853 | * Like alloc_pages_exact(), but try to allocate on node nid first before falling | |
4854 | * back. | |
a862f68a MR |
4855 | * |
4856 | * Return: pointer to the allocated area or %NULL in case of error. | |
ee85c2e1 | 4857 | */ |
e1931811 | 4858 | void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) |
ee85c2e1 | 4859 | { |
d00181b9 | 4860 | unsigned int order = get_order(size); |
63931eb9 VB |
4861 | struct page *p; |
4862 | ||
ba7f1b9e ML |
4863 | if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM))) |
4864 | gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM); | |
63931eb9 VB |
4865 | |
4866 | p = alloc_pages_node(nid, gfp_mask, order); | |
ee85c2e1 AK |
4867 | if (!p) |
4868 | return NULL; | |
4869 | return make_alloc_exact((unsigned long)page_address(p), order, size); | |
4870 | } | |
ee85c2e1 | 4871 | |
2be0ffe2 TT |
4872 | /** |
4873 | * free_pages_exact - release memory allocated via alloc_pages_exact() | |
4874 | * @virt: the value returned by alloc_pages_exact. | |
4875 | * @size: size of allocation, same value as passed to alloc_pages_exact(). | |
4876 | * | |
4877 | * Release the memory allocated by a previous call to alloc_pages_exact. | |
4878 | */ | |
4879 | void free_pages_exact(void *virt, size_t size) | |
4880 | { | |
4881 | unsigned long addr = (unsigned long)virt; | |
4882 | unsigned long end = addr + PAGE_ALIGN(size); | |
4883 | ||
4884 | while (addr < end) { | |
4885 | free_page(addr); | |
4886 | addr += PAGE_SIZE; | |
4887 | } | |
4888 | } | |
4889 | EXPORT_SYMBOL(free_pages_exact); | |
4890 | ||
e0fb5815 ZY |
4891 | /** |
4892 | * nr_free_zone_pages - count number of pages beyond high watermark | |
4893 | * @offset: The zone index of the highest zone | |
4894 | * | |
a862f68a | 4895 | * nr_free_zone_pages() counts the number of pages which are beyond the |
e0fb5815 ZY |
4896 | * high watermark within all zones at or below a given zone index. For each |
4897 | * zone, the number of pages is calculated as: | |
0e056eb5 | 4898 | * |
4899 | * nr_free_zone_pages = managed_pages - high_pages | |
a862f68a MR |
4900 | * |
4901 | * Return: number of pages beyond high watermark. | |
e0fb5815 | 4902 | */ |
ebec3862 | 4903 | static unsigned long nr_free_zone_pages(int offset) |
1da177e4 | 4904 | { |
dd1a239f | 4905 | struct zoneref *z; |
54a6eb5c MG |
4906 | struct zone *zone; |
4907 | ||
e310fd43 | 4908 | /* Just pick one node, since fallback list is circular */ |
ebec3862 | 4909 | unsigned long sum = 0; |
1da177e4 | 4910 | |
0e88460d | 4911 | struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); |
1da177e4 | 4912 | |
54a6eb5c | 4913 | for_each_zone_zonelist(zone, z, zonelist, offset) { |
9705bea5 | 4914 | unsigned long size = zone_managed_pages(zone); |
41858966 | 4915 | unsigned long high = high_wmark_pages(zone); |
e310fd43 MB |
4916 | if (size > high) |
4917 | sum += size - high; | |
1da177e4 LT |
4918 | } |
4919 | ||
4920 | return sum; | |
4921 | } | |
4922 | ||
e0fb5815 ZY |
4923 | /** |
4924 | * nr_free_buffer_pages - count number of pages beyond high watermark | |
4925 | * | |
4926 | * nr_free_buffer_pages() counts the number of pages which are beyond the high | |
4927 | * watermark within ZONE_DMA and ZONE_NORMAL. | |
a862f68a MR |
4928 | * |
4929 | * Return: number of pages beyond high watermark within ZONE_DMA and | |
4930 | * ZONE_NORMAL. | |
1da177e4 | 4931 | */ |
ebec3862 | 4932 | unsigned long nr_free_buffer_pages(void) |
1da177e4 | 4933 | { |
af4ca457 | 4934 | return nr_free_zone_pages(gfp_zone(GFP_USER)); |
1da177e4 | 4935 | } |
c2f1a551 | 4936 | EXPORT_SYMBOL_GPL(nr_free_buffer_pages); |
1da177e4 | 4937 | |
19770b32 MG |
4938 | static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) |
4939 | { | |
4940 | zoneref->zone = zone; | |
4941 | zoneref->zone_idx = zone_idx(zone); | |
4942 | } | |
4943 | ||
1da177e4 LT |
4944 | /* |
4945 | * Builds allocation fallback zone lists. | |
1a93205b CL |
4946 | * |
4947 | * Add all populated zones of a node to the zonelist. | |
1da177e4 | 4948 | */ |
9d3be21b | 4949 | static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs) |
1da177e4 | 4950 | { |
1a93205b | 4951 | struct zone *zone; |
bc732f1d | 4952 | enum zone_type zone_type = MAX_NR_ZONES; |
9d3be21b | 4953 | int nr_zones = 0; |
02a68a5e CL |
4954 | |
4955 | do { | |
2f6726e5 | 4956 | zone_type--; |
070f8032 | 4957 | zone = pgdat->node_zones + zone_type; |
e553f62f | 4958 | if (populated_zone(zone)) { |
9d3be21b | 4959 | zoneref_set_zone(zone, &zonerefs[nr_zones++]); |
070f8032 | 4960 | check_highest_zone(zone_type); |
1da177e4 | 4961 | } |
2f6726e5 | 4962 | } while (zone_type); |
bc732f1d | 4963 | |
070f8032 | 4964 | return nr_zones; |
1da177e4 LT |
4965 | } |
4966 | ||
4967 | #ifdef CONFIG_NUMA | |
f0c0b2b8 KH |
4968 | |
4969 | static int __parse_numa_zonelist_order(char *s) | |
4970 | { | |
c9bff3ee | 4971 | /* |
f0953a1b | 4972 | * We used to support different zonelists modes but they turned |
c9bff3ee MH |
4973 | * out to be just not useful. Let's keep the warning in place |
4974 | * if somebody still use the cmd line parameter so that we do | |
4975 | * not fail it silently | |
4976 | */ | |
4977 | if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) { | |
4978 | pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s); | |
f0c0b2b8 KH |
4979 | return -EINVAL; |
4980 | } | |
4981 | return 0; | |
4982 | } | |
4983 | ||
e95d372c KW |
4984 | static char numa_zonelist_order[] = "Node"; |
4985 | #define NUMA_ZONELIST_ORDER_LEN 16 | |
f0c0b2b8 KH |
4986 | /* |
4987 | * sysctl handler for numa_zonelist_order | |
4988 | */ | |
e95d372c | 4989 | static int numa_zonelist_order_handler(struct ctl_table *table, int write, |
32927393 | 4990 | void *buffer, size_t *length, loff_t *ppos) |
f0c0b2b8 | 4991 | { |
32927393 CH |
4992 | if (write) |
4993 | return __parse_numa_zonelist_order(buffer); | |
4994 | return proc_dostring(table, write, buffer, length, ppos); | |
f0c0b2b8 KH |
4995 | } |
4996 | ||
f0c0b2b8 KH |
4997 | static int node_load[MAX_NUMNODES]; |
4998 | ||
1da177e4 | 4999 | /** |
4dc3b16b | 5000 | * find_next_best_node - find the next node that should appear in a given node's fallback list |
1da177e4 LT |
5001 | * @node: node whose fallback list we're appending |
5002 | * @used_node_mask: nodemask_t of already used nodes | |
5003 | * | |
5004 | * We use a number of factors to determine which is the next node that should | |
5005 | * appear on a given node's fallback list. The node should not have appeared | |
5006 | * already in @node's fallback list, and it should be the next closest node | |
5007 | * according to the distance array (which contains arbitrary distance values | |
5008 | * from each node to each node in the system), and should also prefer nodes | |
5009 | * with no CPUs, since presumably they'll have very little allocation pressure | |
5010 | * on them otherwise. | |
a862f68a MR |
5011 | * |
5012 | * Return: node id of the found node or %NUMA_NO_NODE if no node is found. | |
1da177e4 | 5013 | */ |
79c28a41 | 5014 | int find_next_best_node(int node, nodemask_t *used_node_mask) |
1da177e4 | 5015 | { |
4cf808eb | 5016 | int n, val; |
1da177e4 | 5017 | int min_val = INT_MAX; |
00ef2d2f | 5018 | int best_node = NUMA_NO_NODE; |
1da177e4 | 5019 | |
c2baef39 QZ |
5020 | /* |
5021 | * Use the local node if we haven't already, but for memoryless local | |
5022 | * node, we should skip it and fall back to other nodes. | |
5023 | */ | |
5024 | if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) { | |
4cf808eb LT |
5025 | node_set(node, *used_node_mask); |
5026 | return node; | |
5027 | } | |
1da177e4 | 5028 | |
4b0ef1fe | 5029 | for_each_node_state(n, N_MEMORY) { |
1da177e4 LT |
5030 | |
5031 | /* Don't want a node to appear more than once */ | |
5032 | if (node_isset(n, *used_node_mask)) | |
5033 | continue; | |
5034 | ||
1da177e4 LT |
5035 | /* Use the distance array to find the distance */ |
5036 | val = node_distance(node, n); | |
5037 | ||
4cf808eb LT |
5038 | /* Penalize nodes under us ("prefer the next node") */ |
5039 | val += (n < node); | |
5040 | ||
1da177e4 | 5041 | /* Give preference to headless and unused nodes */ |
b630749f | 5042 | if (!cpumask_empty(cpumask_of_node(n))) |
1da177e4 LT |
5043 | val += PENALTY_FOR_NODE_WITH_CPUS; |
5044 | ||
5045 | /* Slight preference for less loaded node */ | |
37931324 | 5046 | val *= MAX_NUMNODES; |
1da177e4 LT |
5047 | val += node_load[n]; |
5048 | ||
5049 | if (val < min_val) { | |
5050 | min_val = val; | |
5051 | best_node = n; | |
5052 | } | |
5053 | } | |
5054 | ||
5055 | if (best_node >= 0) | |
5056 | node_set(best_node, *used_node_mask); | |
5057 | ||
5058 | return best_node; | |
5059 | } | |
5060 | ||
f0c0b2b8 KH |
5061 | |
5062 | /* | |
5063 | * Build zonelists ordered by node and zones within node. | |
5064 | * This results in maximum locality--normal zone overflows into local | |
5065 | * DMA zone, if any--but risks exhausting DMA zone. | |
5066 | */ | |
9d3be21b MH |
5067 | static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order, |
5068 | unsigned nr_nodes) | |
1da177e4 | 5069 | { |
9d3be21b MH |
5070 | struct zoneref *zonerefs; |
5071 | int i; | |
5072 | ||
5073 | zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs; | |
5074 | ||
5075 | for (i = 0; i < nr_nodes; i++) { | |
5076 | int nr_zones; | |
5077 | ||
5078 | pg_data_t *node = NODE_DATA(node_order[i]); | |
f0c0b2b8 | 5079 | |
9d3be21b MH |
5080 | nr_zones = build_zonerefs_node(node, zonerefs); |
5081 | zonerefs += nr_zones; | |
5082 | } | |
5083 | zonerefs->zone = NULL; | |
5084 | zonerefs->zone_idx = 0; | |
f0c0b2b8 KH |
5085 | } |
5086 | ||
523b9458 CL |
5087 | /* |
5088 | * Build gfp_thisnode zonelists | |
5089 | */ | |
5090 | static void build_thisnode_zonelists(pg_data_t *pgdat) | |
5091 | { | |
9d3be21b MH |
5092 | struct zoneref *zonerefs; |
5093 | int nr_zones; | |
523b9458 | 5094 | |
9d3be21b MH |
5095 | zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs; |
5096 | nr_zones = build_zonerefs_node(pgdat, zonerefs); | |
5097 | zonerefs += nr_zones; | |
5098 | zonerefs->zone = NULL; | |
5099 | zonerefs->zone_idx = 0; | |
523b9458 CL |
5100 | } |
5101 | ||
f0c0b2b8 KH |
5102 | /* |
5103 | * Build zonelists ordered by zone and nodes within zones. | |
5104 | * This results in conserving DMA zone[s] until all Normal memory is | |
5105 | * exhausted, but results in overflowing to remote node while memory | |
5106 | * may still exist in local DMA zone. | |
5107 | */ | |
f0c0b2b8 | 5108 | |
f0c0b2b8 KH |
5109 | static void build_zonelists(pg_data_t *pgdat) |
5110 | { | |
9d3be21b | 5111 | static int node_order[MAX_NUMNODES]; |
37931324 | 5112 | int node, nr_nodes = 0; |
d0ddf49b | 5113 | nodemask_t used_mask = NODE_MASK_NONE; |
f0c0b2b8 | 5114 | int local_node, prev_node; |
1da177e4 LT |
5115 | |
5116 | /* NUMA-aware ordering of nodes */ | |
5117 | local_node = pgdat->node_id; | |
1da177e4 | 5118 | prev_node = local_node; |
f0c0b2b8 | 5119 | |
f0c0b2b8 | 5120 | memset(node_order, 0, sizeof(node_order)); |
1da177e4 LT |
5121 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { |
5122 | /* | |
5123 | * We don't want to pressure a particular node. | |
5124 | * So adding penalty to the first node in same | |
5125 | * distance group to make it round-robin. | |
5126 | */ | |
957f822a DR |
5127 | if (node_distance(local_node, node) != |
5128 | node_distance(local_node, prev_node)) | |
37931324 | 5129 | node_load[node] += 1; |
f0c0b2b8 | 5130 | |
9d3be21b | 5131 | node_order[nr_nodes++] = node; |
1da177e4 | 5132 | prev_node = node; |
1da177e4 | 5133 | } |
523b9458 | 5134 | |
9d3be21b | 5135 | build_zonelists_in_node_order(pgdat, node_order, nr_nodes); |
523b9458 | 5136 | build_thisnode_zonelists(pgdat); |
6cf25392 BR |
5137 | pr_info("Fallback order for Node %d: ", local_node); |
5138 | for (node = 0; node < nr_nodes; node++) | |
5139 | pr_cont("%d ", node_order[node]); | |
5140 | pr_cont("\n"); | |
1da177e4 LT |
5141 | } |
5142 | ||
7aac7898 LS |
5143 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES |
5144 | /* | |
5145 | * Return node id of node used for "local" allocations. | |
5146 | * I.e., first node id of first zone in arg node's generic zonelist. | |
5147 | * Used for initializing percpu 'numa_mem', which is used primarily | |
5148 | * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. | |
5149 | */ | |
5150 | int local_memory_node(int node) | |
5151 | { | |
c33d6c06 | 5152 | struct zoneref *z; |
7aac7898 | 5153 | |
c33d6c06 | 5154 | z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL), |
7aac7898 | 5155 | gfp_zone(GFP_KERNEL), |
c33d6c06 | 5156 | NULL); |
c1093b74 | 5157 | return zone_to_nid(z->zone); |
7aac7898 LS |
5158 | } |
5159 | #endif | |
f0c0b2b8 | 5160 | |
6423aa81 JK |
5161 | static void setup_min_unmapped_ratio(void); |
5162 | static void setup_min_slab_ratio(void); | |
1da177e4 LT |
5163 | #else /* CONFIG_NUMA */ |
5164 | ||
f0c0b2b8 | 5165 | static void build_zonelists(pg_data_t *pgdat) |
1da177e4 | 5166 | { |
19655d34 | 5167 | int node, local_node; |
9d3be21b MH |
5168 | struct zoneref *zonerefs; |
5169 | int nr_zones; | |
1da177e4 LT |
5170 | |
5171 | local_node = pgdat->node_id; | |
1da177e4 | 5172 | |
9d3be21b MH |
5173 | zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs; |
5174 | nr_zones = build_zonerefs_node(pgdat, zonerefs); | |
5175 | zonerefs += nr_zones; | |
1da177e4 | 5176 | |
54a6eb5c MG |
5177 | /* |
5178 | * Now we build the zonelist so that it contains the zones | |
5179 | * of all the other nodes. | |
5180 | * We don't want to pressure a particular node, so when | |
5181 | * building the zones for node N, we make sure that the | |
5182 | * zones coming right after the local ones are those from | |
5183 | * node N+1 (modulo N) | |
5184 | */ | |
5185 | for (node = local_node + 1; node < MAX_NUMNODES; node++) { | |
5186 | if (!node_online(node)) | |
5187 | continue; | |
9d3be21b MH |
5188 | nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs); |
5189 | zonerefs += nr_zones; | |
1da177e4 | 5190 | } |
54a6eb5c MG |
5191 | for (node = 0; node < local_node; node++) { |
5192 | if (!node_online(node)) | |
5193 | continue; | |
9d3be21b MH |
5194 | nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs); |
5195 | zonerefs += nr_zones; | |
54a6eb5c MG |
5196 | } |
5197 | ||
9d3be21b MH |
5198 | zonerefs->zone = NULL; |
5199 | zonerefs->zone_idx = 0; | |
1da177e4 LT |
5200 | } |
5201 | ||
5202 | #endif /* CONFIG_NUMA */ | |
5203 | ||
99dcc3e5 CL |
5204 | /* |
5205 | * Boot pageset table. One per cpu which is going to be used for all | |
5206 | * zones and all nodes. The parameters will be set in such a way | |
5207 | * that an item put on a list will immediately be handed over to | |
5208 | * the buddy list. This is safe since pageset manipulation is done | |
5209 | * with interrupts disabled. | |
5210 | * | |
5211 | * The boot_pagesets must be kept even after bootup is complete for | |
5212 | * unused processors and/or zones. They do play a role for bootstrapping | |
5213 | * hotplugged processors. | |
5214 | * | |
5215 | * zoneinfo_show() and maybe other functions do | |
5216 | * not check if the processor is online before following the pageset pointer. | |
5217 | * Other parts of the kernel may not check if the zone is available. | |
5218 | */ | |
28f836b6 | 5219 | static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats); |
952eaf81 VB |
5220 | /* These effectively disable the pcplists in the boot pageset completely */ |
5221 | #define BOOT_PAGESET_HIGH 0 | |
5222 | #define BOOT_PAGESET_BATCH 1 | |
28f836b6 MG |
5223 | static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset); |
5224 | static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats); | |
99dcc3e5 | 5225 | |
11cd8638 | 5226 | static void __build_all_zonelists(void *data) |
1da177e4 | 5227 | { |
6811378e | 5228 | int nid; |
afb6ebb3 | 5229 | int __maybe_unused cpu; |
9adb62a5 | 5230 | pg_data_t *self = data; |
1007843a | 5231 | unsigned long flags; |
b93e0f32 | 5232 | |
1007843a | 5233 | /* |
a2ebb515 SAS |
5234 | * The zonelist_update_seq must be acquired with irqsave because the |
5235 | * reader can be invoked from IRQ with GFP_ATOMIC. | |
1007843a | 5236 | */ |
a2ebb515 | 5237 | write_seqlock_irqsave(&zonelist_update_seq, flags); |
1007843a | 5238 | /* |
a2ebb515 SAS |
5239 | * Also disable synchronous printk() to prevent any printk() from |
5240 | * trying to hold port->lock, for | |
1007843a TH |
5241 | * tty_insert_flip_string_and_push_buffer() on other CPU might be |
5242 | * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held. | |
5243 | */ | |
5244 | printk_deferred_enter(); | |
9276b1bc | 5245 | |
7f9cfb31 BL |
5246 | #ifdef CONFIG_NUMA |
5247 | memset(node_load, 0, sizeof(node_load)); | |
5248 | #endif | |
9adb62a5 | 5249 | |
c1152583 WY |
5250 | /* |
5251 | * This node is hotadded and no memory is yet present. So just | |
5252 | * building zonelists is fine - no need to touch other nodes. | |
5253 | */ | |
9adb62a5 JL |
5254 | if (self && !node_online(self->node_id)) { |
5255 | build_zonelists(self); | |
c1152583 | 5256 | } else { |
09f49dca MH |
5257 | /* |
5258 | * All possible nodes have pgdat preallocated | |
5259 | * in free_area_init | |
5260 | */ | |
5261 | for_each_node(nid) { | |
c1152583 | 5262 | pg_data_t *pgdat = NODE_DATA(nid); |
7ea1530a | 5263 | |
c1152583 WY |
5264 | build_zonelists(pgdat); |
5265 | } | |
99dcc3e5 | 5266 | |
7aac7898 LS |
5267 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES |
5268 | /* | |
5269 | * We now know the "local memory node" for each node-- | |
5270 | * i.e., the node of the first zone in the generic zonelist. | |
5271 | * Set up numa_mem percpu variable for on-line cpus. During | |
5272 | * boot, only the boot cpu should be on-line; we'll init the | |
5273 | * secondary cpus' numa_mem as they come on-line. During | |
5274 | * node/memory hotplug, we'll fixup all on-line cpus. | |
5275 | */ | |
d9c9a0b9 | 5276 | for_each_online_cpu(cpu) |
7aac7898 | 5277 | set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); |
afb6ebb3 | 5278 | #endif |
d9c9a0b9 | 5279 | } |
b93e0f32 | 5280 | |
1007843a | 5281 | printk_deferred_exit(); |
a2ebb515 | 5282 | write_sequnlock_irqrestore(&zonelist_update_seq, flags); |
6811378e YG |
5283 | } |
5284 | ||
061f67bc RV |
5285 | static noinline void __init |
5286 | build_all_zonelists_init(void) | |
5287 | { | |
afb6ebb3 MH |
5288 | int cpu; |
5289 | ||
061f67bc | 5290 | __build_all_zonelists(NULL); |
afb6ebb3 MH |
5291 | |
5292 | /* | |
5293 | * Initialize the boot_pagesets that are going to be used | |
5294 | * for bootstrapping processors. The real pagesets for | |
5295 | * each zone will be allocated later when the per cpu | |
5296 | * allocator is available. | |
5297 | * | |
5298 | * boot_pagesets are used also for bootstrapping offline | |
5299 | * cpus if the system is already booted because the pagesets | |
5300 | * are needed to initialize allocators on a specific cpu too. | |
5301 | * F.e. the percpu allocator needs the page allocator which | |
5302 | * needs the percpu allocator in order to allocate its pagesets | |
5303 | * (a chicken-egg dilemma). | |
5304 | */ | |
5305 | for_each_possible_cpu(cpu) | |
28f836b6 | 5306 | per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu)); |
afb6ebb3 | 5307 | |
061f67bc RV |
5308 | mminit_verify_zonelist(); |
5309 | cpuset_init_current_mems_allowed(); | |
5310 | } | |
5311 | ||
4eaf3f64 | 5312 | /* |
4eaf3f64 | 5313 | * unless system_state == SYSTEM_BOOTING. |
061f67bc | 5314 | * |
72675e13 | 5315 | * __ref due to call of __init annotated helper build_all_zonelists_init |
061f67bc | 5316 | * [protected by SYSTEM_BOOTING]. |
4eaf3f64 | 5317 | */ |
72675e13 | 5318 | void __ref build_all_zonelists(pg_data_t *pgdat) |
6811378e | 5319 | { |
0a18e607 DH |
5320 | unsigned long vm_total_pages; |
5321 | ||
6811378e | 5322 | if (system_state == SYSTEM_BOOTING) { |
061f67bc | 5323 | build_all_zonelists_init(); |
6811378e | 5324 | } else { |
11cd8638 | 5325 | __build_all_zonelists(pgdat); |
6811378e YG |
5326 | /* cpuset refresh routine should be here */ |
5327 | } | |
56b9413b DH |
5328 | /* Get the number of free pages beyond high watermark in all zones. */ |
5329 | vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); | |
9ef9acb0 MG |
5330 | /* |
5331 | * Disable grouping by mobility if the number of pages in the | |
5332 | * system is too low to allow the mechanism to work. It would be | |
5333 | * more accurate, but expensive to check per-zone. This check is | |
5334 | * made on memory-hotadd so a system can start with mobility | |
5335 | * disabled and enable it later | |
5336 | */ | |
d9c23400 | 5337 | if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) |
9ef9acb0 MG |
5338 | page_group_by_mobility_disabled = 1; |
5339 | else | |
5340 | page_group_by_mobility_disabled = 0; | |
5341 | ||
ce0725f7 | 5342 | pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n", |
756a025f | 5343 | nr_online_nodes, |
756a025f JP |
5344 | page_group_by_mobility_disabled ? "off" : "on", |
5345 | vm_total_pages); | |
f0c0b2b8 | 5346 | #ifdef CONFIG_NUMA |
f88dfff5 | 5347 | pr_info("Policy zone: %s\n", zone_names[policy_zone]); |
f0c0b2b8 | 5348 | #endif |
1da177e4 LT |
5349 | } |
5350 | ||
9420f89d | 5351 | static int zone_batchsize(struct zone *zone) |
1da177e4 | 5352 | { |
9420f89d MRI |
5353 | #ifdef CONFIG_MMU |
5354 | int batch; | |
1da177e4 | 5355 | |
9420f89d MRI |
5356 | /* |
5357 | * The number of pages to batch allocate is either ~0.1% | |
5358 | * of the zone or 1MB, whichever is smaller. The batch | |
5359 | * size is striking a balance between allocation latency | |
5360 | * and zone lock contention. | |
5361 | */ | |
5362 | batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE); | |
5363 | batch /= 4; /* We effectively *= 4 below */ | |
5364 | if (batch < 1) | |
5365 | batch = 1; | |
22b31eec | 5366 | |
4b94ffdc | 5367 | /* |
9420f89d MRI |
5368 | * Clamp the batch to a 2^n - 1 value. Having a power |
5369 | * of 2 value was found to be more likely to have | |
5370 | * suboptimal cache aliasing properties in some cases. | |
5371 | * | |
5372 | * For example if 2 tasks are alternately allocating | |
5373 | * batches of pages, one task can end up with a lot | |
5374 | * of pages of one half of the possible page colors | |
5375 | * and the other with pages of the other colors. | |
4b94ffdc | 5376 | */ |
9420f89d | 5377 | batch = rounddown_pow_of_two(batch + batch/2) - 1; |
966cf44f | 5378 | |
9420f89d | 5379 | return batch; |
3a6be87f DH |
5380 | |
5381 | #else | |
5382 | /* The deferral and batching of frees should be suppressed under NOMMU | |
5383 | * conditions. | |
5384 | * | |
5385 | * The problem is that NOMMU needs to be able to allocate large chunks | |
5386 | * of contiguous memory as there's no hardware page translation to | |
5387 | * assemble apparent contiguous memory from discontiguous pages. | |
5388 | * | |
5389 | * Queueing large contiguous runs of pages for batching, however, | |
5390 | * causes the pages to actually be freed in smaller chunks. As there | |
5391 | * can be a significant delay between the individual batches being | |
5392 | * recycled, this leads to the once large chunks of space being | |
5393 | * fragmented and becoming unavailable for high-order allocations. | |
5394 | */ | |
5395 | return 0; | |
5396 | #endif | |
e7c8d5c9 CL |
5397 | } |
5398 | ||
e95d372c | 5399 | static int percpu_pagelist_high_fraction; |
90b41691 HY |
5400 | static int zone_highsize(struct zone *zone, int batch, int cpu_online, |
5401 | int high_fraction) | |
b92ca18e | 5402 | { |
9420f89d MRI |
5403 | #ifdef CONFIG_MMU |
5404 | int high; | |
5405 | int nr_split_cpus; | |
5406 | unsigned long total_pages; | |
c13291a5 | 5407 | |
90b41691 | 5408 | if (!high_fraction) { |
2a1e274a | 5409 | /* |
9420f89d MRI |
5410 | * By default, the high value of the pcp is based on the zone |
5411 | * low watermark so that if they are full then background | |
5412 | * reclaim will not be started prematurely. | |
2a1e274a | 5413 | */ |
9420f89d MRI |
5414 | total_pages = low_wmark_pages(zone); |
5415 | } else { | |
2a1e274a | 5416 | /* |
9420f89d MRI |
5417 | * If percpu_pagelist_high_fraction is configured, the high |
5418 | * value is based on a fraction of the managed pages in the | |
5419 | * zone. | |
2a1e274a | 5420 | */ |
90b41691 | 5421 | total_pages = zone_managed_pages(zone) / high_fraction; |
2a1e274a MG |
5422 | } |
5423 | ||
5424 | /* | |
9420f89d MRI |
5425 | * Split the high value across all online CPUs local to the zone. Note |
5426 | * that early in boot that CPUs may not be online yet and that during | |
5427 | * CPU hotplug that the cpumask is not yet updated when a CPU is being | |
90b41691 HY |
5428 | * onlined. For memory nodes that have no CPUs, split the high value |
5429 | * across all online CPUs to mitigate the risk that reclaim is triggered | |
9420f89d | 5430 | * prematurely due to pages stored on pcp lists. |
2a1e274a | 5431 | */ |
9420f89d MRI |
5432 | nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online; |
5433 | if (!nr_split_cpus) | |
5434 | nr_split_cpus = num_online_cpus(); | |
5435 | high = total_pages / nr_split_cpus; | |
2a1e274a | 5436 | |
9420f89d MRI |
5437 | /* |
5438 | * Ensure high is at least batch*4. The multiple is based on the | |
5439 | * historical relationship between high and batch. | |
5440 | */ | |
5441 | high = max(high, batch << 2); | |
37b07e41 | 5442 | |
9420f89d MRI |
5443 | return high; |
5444 | #else | |
5445 | return 0; | |
5446 | #endif | |
37b07e41 LS |
5447 | } |
5448 | ||
51930df5 | 5449 | /* |
9420f89d MRI |
5450 | * pcp->high and pcp->batch values are related and generally batch is lower |
5451 | * than high. They are also related to pcp->count such that count is lower | |
5452 | * than high, and as soon as it reaches high, the pcplist is flushed. | |
5453 | * | |
5454 | * However, guaranteeing these relations at all times would require e.g. write | |
5455 | * barriers here but also careful usage of read barriers at the read side, and | |
5456 | * thus be prone to error and bad for performance. Thus the update only prevents | |
90b41691 HY |
5457 | * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max |
5458 | * should ensure they can cope with those fields changing asynchronously, and | |
5459 | * fully trust only the pcp->count field on the local CPU with interrupts | |
5460 | * disabled. | |
9420f89d MRI |
5461 | * |
5462 | * mutex_is_locked(&pcp_batch_high_lock) required when calling this function | |
5463 | * outside of boot time (or some other assurance that no concurrent updaters | |
5464 | * exist). | |
51930df5 | 5465 | */ |
90b41691 HY |
5466 | static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min, |
5467 | unsigned long high_max, unsigned long batch) | |
51930df5 | 5468 | { |
9420f89d | 5469 | WRITE_ONCE(pcp->batch, batch); |
90b41691 HY |
5470 | WRITE_ONCE(pcp->high_min, high_min); |
5471 | WRITE_ONCE(pcp->high_max, high_max); | |
51930df5 MR |
5472 | } |
5473 | ||
9420f89d | 5474 | static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats) |
c713216d | 5475 | { |
9420f89d | 5476 | int pindex; |
90cae1fe | 5477 | |
9420f89d MRI |
5478 | memset(pcp, 0, sizeof(*pcp)); |
5479 | memset(pzstats, 0, sizeof(*pzstats)); | |
90cae1fe | 5480 | |
9420f89d MRI |
5481 | spin_lock_init(&pcp->lock); |
5482 | for (pindex = 0; pindex < NR_PCP_LISTS; pindex++) | |
5483 | INIT_LIST_HEAD(&pcp->lists[pindex]); | |
2a1e274a | 5484 | |
9420f89d MRI |
5485 | /* |
5486 | * Set batch and high values safe for a boot pageset. A true percpu | |
5487 | * pageset's initialization will update them subsequently. Here we don't | |
5488 | * need to be as careful as pageset_update() as nobody can access the | |
5489 | * pageset yet. | |
5490 | */ | |
90b41691 HY |
5491 | pcp->high_min = BOOT_PAGESET_HIGH; |
5492 | pcp->high_max = BOOT_PAGESET_HIGH; | |
9420f89d | 5493 | pcp->batch = BOOT_PAGESET_BATCH; |
6ccdcb6d | 5494 | pcp->free_count = 0; |
9420f89d | 5495 | } |
c713216d | 5496 | |
90b41691 HY |
5497 | static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min, |
5498 | unsigned long high_max, unsigned long batch) | |
9420f89d MRI |
5499 | { |
5500 | struct per_cpu_pages *pcp; | |
5501 | int cpu; | |
2a1e274a | 5502 | |
9420f89d MRI |
5503 | for_each_possible_cpu(cpu) { |
5504 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); | |
90b41691 | 5505 | pageset_update(pcp, high_min, high_max, batch); |
2a1e274a | 5506 | } |
9420f89d | 5507 | } |
c713216d | 5508 | |
9420f89d MRI |
5509 | /* |
5510 | * Calculate and set new high and batch values for all per-cpu pagesets of a | |
5511 | * zone based on the zone's size. | |
5512 | */ | |
5513 | static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online) | |
5514 | { | |
90b41691 | 5515 | int new_high_min, new_high_max, new_batch; |
09f49dca | 5516 | |
9420f89d | 5517 | new_batch = max(1, zone_batchsize(zone)); |
90b41691 HY |
5518 | if (percpu_pagelist_high_fraction) { |
5519 | new_high_min = zone_highsize(zone, new_batch, cpu_online, | |
5520 | percpu_pagelist_high_fraction); | |
5521 | /* | |
5522 | * PCP high is tuned manually, disable auto-tuning via | |
5523 | * setting high_min and high_max to the manual value. | |
5524 | */ | |
5525 | new_high_max = new_high_min; | |
5526 | } else { | |
5527 | new_high_min = zone_highsize(zone, new_batch, cpu_online, 0); | |
5528 | new_high_max = zone_highsize(zone, new_batch, cpu_online, | |
5529 | MIN_PERCPU_PAGELIST_HIGH_FRACTION); | |
5530 | } | |
09f49dca | 5531 | |
90b41691 HY |
5532 | if (zone->pageset_high_min == new_high_min && |
5533 | zone->pageset_high_max == new_high_max && | |
9420f89d MRI |
5534 | zone->pageset_batch == new_batch) |
5535 | return; | |
37b07e41 | 5536 | |
90b41691 HY |
5537 | zone->pageset_high_min = new_high_min; |
5538 | zone->pageset_high_max = new_high_max; | |
9420f89d | 5539 | zone->pageset_batch = new_batch; |
122e093c | 5540 | |
90b41691 HY |
5541 | __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max, |
5542 | new_batch); | |
c713216d | 5543 | } |
2a1e274a | 5544 | |
9420f89d | 5545 | void __meminit setup_zone_pageset(struct zone *zone) |
2a1e274a | 5546 | { |
9420f89d | 5547 | int cpu; |
2a1e274a | 5548 | |
9420f89d MRI |
5549 | /* Size may be 0 on !SMP && !NUMA */ |
5550 | if (sizeof(struct per_cpu_zonestat) > 0) | |
5551 | zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat); | |
2a1e274a | 5552 | |
9420f89d MRI |
5553 | zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages); |
5554 | for_each_possible_cpu(cpu) { | |
5555 | struct per_cpu_pages *pcp; | |
5556 | struct per_cpu_zonestat *pzstats; | |
2a1e274a | 5557 | |
9420f89d MRI |
5558 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); |
5559 | pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu); | |
5560 | per_cpu_pages_init(pcp, pzstats); | |
a5c6d650 | 5561 | } |
9420f89d MRI |
5562 | |
5563 | zone_set_pageset_high_and_batch(zone, 0); | |
2a1e274a | 5564 | } |
ed7ed365 | 5565 | |
7e63efef | 5566 | /* |
9420f89d MRI |
5567 | * The zone indicated has a new number of managed_pages; batch sizes and percpu |
5568 | * page high values need to be recalculated. | |
7e63efef | 5569 | */ |
9420f89d | 5570 | static void zone_pcp_update(struct zone *zone, int cpu_online) |
7e63efef | 5571 | { |
9420f89d MRI |
5572 | mutex_lock(&pcp_batch_high_lock); |
5573 | zone_set_pageset_high_and_batch(zone, cpu_online); | |
5574 | mutex_unlock(&pcp_batch_high_lock); | |
7e63efef MG |
5575 | } |
5576 | ||
362d37a1 HY |
5577 | static void zone_pcp_update_cacheinfo(struct zone *zone) |
5578 | { | |
5579 | int cpu; | |
5580 | struct per_cpu_pages *pcp; | |
5581 | struct cpu_cacheinfo *cci; | |
5582 | ||
5583 | for_each_online_cpu(cpu) { | |
5584 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu); | |
5585 | cci = get_cpu_cacheinfo(cpu); | |
5586 | /* | |
5587 | * If data cache slice of CPU is large enough, "pcp->batch" | |
5588 | * pages can be preserved in PCP before draining PCP for | |
5589 | * consecutive high-order pages freeing without allocation. | |
5590 | * This can reduce zone lock contention without hurting | |
5591 | * cache-hot pages sharing. | |
5592 | */ | |
5593 | spin_lock(&pcp->lock); | |
5594 | if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch) | |
5595 | pcp->flags |= PCPF_FREE_HIGH_BATCH; | |
5596 | else | |
5597 | pcp->flags &= ~PCPF_FREE_HIGH_BATCH; | |
5598 | spin_unlock(&pcp->lock); | |
5599 | } | |
5600 | } | |
5601 | ||
5602 | void setup_pcp_cacheinfo(void) | |
5603 | { | |
5604 | struct zone *zone; | |
5605 | ||
5606 | for_each_populated_zone(zone) | |
5607 | zone_pcp_update_cacheinfo(zone); | |
5608 | } | |
5609 | ||
7e63efef | 5610 | /* |
9420f89d MRI |
5611 | * Allocate per cpu pagesets and initialize them. |
5612 | * Before this call only boot pagesets were available. | |
7e63efef | 5613 | */ |
9420f89d | 5614 | void __init setup_per_cpu_pageset(void) |
7e63efef | 5615 | { |
9420f89d MRI |
5616 | struct pglist_data *pgdat; |
5617 | struct zone *zone; | |
5618 | int __maybe_unused cpu; | |
5619 | ||
5620 | for_each_populated_zone(zone) | |
5621 | setup_zone_pageset(zone); | |
5622 | ||
5623 | #ifdef CONFIG_NUMA | |
5624 | /* | |
5625 | * Unpopulated zones continue using the boot pagesets. | |
5626 | * The numa stats for these pagesets need to be reset. | |
5627 | * Otherwise, they will end up skewing the stats of | |
5628 | * the nodes these zones are associated with. | |
5629 | */ | |
5630 | for_each_possible_cpu(cpu) { | |
5631 | struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu); | |
5632 | memset(pzstats->vm_numa_event, 0, | |
5633 | sizeof(pzstats->vm_numa_event)); | |
5634 | } | |
5635 | #endif | |
5636 | ||
5637 | for_each_online_pgdat(pgdat) | |
5638 | pgdat->per_cpu_nodestats = | |
5639 | alloc_percpu(struct per_cpu_nodestat); | |
7e63efef MG |
5640 | } |
5641 | ||
9420f89d MRI |
5642 | __meminit void zone_pcp_init(struct zone *zone) |
5643 | { | |
5644 | /* | |
5645 | * per cpu subsystem is not up at this point. The following code | |
5646 | * relies on the ability of the linker to provide the | |
5647 | * offset of a (static) per cpu variable into the per cpu area. | |
5648 | */ | |
5649 | zone->per_cpu_pageset = &boot_pageset; | |
5650 | zone->per_cpu_zonestats = &boot_zonestats; | |
90b41691 HY |
5651 | zone->pageset_high_min = BOOT_PAGESET_HIGH; |
5652 | zone->pageset_high_max = BOOT_PAGESET_HIGH; | |
9420f89d MRI |
5653 | zone->pageset_batch = BOOT_PAGESET_BATCH; |
5654 | ||
5655 | if (populated_zone(zone)) | |
5656 | pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name, | |
5657 | zone->present_pages, zone_batchsize(zone)); | |
5658 | } | |
ed7ed365 | 5659 | |
c3d5f5f0 JL |
5660 | void adjust_managed_page_count(struct page *page, long count) |
5661 | { | |
9705bea5 | 5662 | atomic_long_add(count, &page_zone(page)->managed_pages); |
ca79b0c2 | 5663 | totalram_pages_add(count); |
3dcc0571 JL |
5664 | #ifdef CONFIG_HIGHMEM |
5665 | if (PageHighMem(page)) | |
ca79b0c2 | 5666 | totalhigh_pages_add(count); |
3dcc0571 | 5667 | #endif |
c3d5f5f0 | 5668 | } |
3dcc0571 | 5669 | EXPORT_SYMBOL(adjust_managed_page_count); |
c3d5f5f0 | 5670 | |
e5cb113f | 5671 | unsigned long free_reserved_area(void *start, void *end, int poison, const char *s) |
69afade7 | 5672 | { |
11199692 JL |
5673 | void *pos; |
5674 | unsigned long pages = 0; | |
69afade7 | 5675 | |
11199692 JL |
5676 | start = (void *)PAGE_ALIGN((unsigned long)start); |
5677 | end = (void *)((unsigned long)end & PAGE_MASK); | |
5678 | for (pos = start; pos < end; pos += PAGE_SIZE, pages++) { | |
0d834328 DH |
5679 | struct page *page = virt_to_page(pos); |
5680 | void *direct_map_addr; | |
5681 | ||
5682 | /* | |
5683 | * 'direct_map_addr' might be different from 'pos' | |
5684 | * because some architectures' virt_to_page() | |
5685 | * work with aliases. Getting the direct map | |
5686 | * address ensures that we get a _writeable_ | |
5687 | * alias for the memset(). | |
5688 | */ | |
5689 | direct_map_addr = page_address(page); | |
c746170d VF |
5690 | /* |
5691 | * Perform a kasan-unchecked memset() since this memory | |
5692 | * has not been initialized. | |
5693 | */ | |
5694 | direct_map_addr = kasan_reset_tag(direct_map_addr); | |
dbe67df4 | 5695 | if ((unsigned int)poison <= 0xFF) |
0d834328 DH |
5696 | memset(direct_map_addr, poison, PAGE_SIZE); |
5697 | ||
5698 | free_reserved_page(page); | |
69afade7 JL |
5699 | } |
5700 | ||
5701 | if (pages && s) | |
ff7ed9e4 | 5702 | pr_info("Freeing %s memory: %ldK\n", s, K(pages)); |
69afade7 JL |
5703 | |
5704 | return pages; | |
5705 | } | |
5706 | ||
005fd4bb | 5707 | static int page_alloc_cpu_dead(unsigned int cpu) |
1da177e4 | 5708 | { |
04f8cfea | 5709 | struct zone *zone; |
1da177e4 | 5710 | |
005fd4bb | 5711 | lru_add_drain_cpu(cpu); |
96f97c43 | 5712 | mlock_drain_remote(cpu); |
005fd4bb | 5713 | drain_pages(cpu); |
9f8f2172 | 5714 | |
005fd4bb SAS |
5715 | /* |
5716 | * Spill the event counters of the dead processor | |
5717 | * into the current processors event counters. | |
5718 | * This artificially elevates the count of the current | |
5719 | * processor. | |
5720 | */ | |
5721 | vm_events_fold_cpu(cpu); | |
9f8f2172 | 5722 | |
005fd4bb SAS |
5723 | /* |
5724 | * Zero the differential counters of the dead processor | |
5725 | * so that the vm statistics are consistent. | |
5726 | * | |
5727 | * This is only okay since the processor is dead and cannot | |
5728 | * race with what we are doing. | |
5729 | */ | |
5730 | cpu_vm_stats_fold(cpu); | |
04f8cfea MG |
5731 | |
5732 | for_each_populated_zone(zone) | |
5733 | zone_pcp_update(zone, 0); | |
5734 | ||
5735 | return 0; | |
5736 | } | |
5737 | ||
5738 | static int page_alloc_cpu_online(unsigned int cpu) | |
5739 | { | |
5740 | struct zone *zone; | |
5741 | ||
5742 | for_each_populated_zone(zone) | |
5743 | zone_pcp_update(zone, 1); | |
005fd4bb | 5744 | return 0; |
1da177e4 | 5745 | } |
1da177e4 | 5746 | |
c4fbed4b | 5747 | void __init page_alloc_init_cpuhp(void) |
1da177e4 | 5748 | { |
005fd4bb SAS |
5749 | int ret; |
5750 | ||
04f8cfea MG |
5751 | ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC, |
5752 | "mm/page_alloc:pcp", | |
5753 | page_alloc_cpu_online, | |
005fd4bb SAS |
5754 | page_alloc_cpu_dead); |
5755 | WARN_ON(ret < 0); | |
1da177e4 LT |
5756 | } |
5757 | ||
cb45b0e9 | 5758 | /* |
34b10060 | 5759 | * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio |
cb45b0e9 HA |
5760 | * or min_free_kbytes changes. |
5761 | */ | |
5762 | static void calculate_totalreserve_pages(void) | |
5763 | { | |
5764 | struct pglist_data *pgdat; | |
5765 | unsigned long reserve_pages = 0; | |
2f6726e5 | 5766 | enum zone_type i, j; |
cb45b0e9 HA |
5767 | |
5768 | for_each_online_pgdat(pgdat) { | |
281e3726 MG |
5769 | |
5770 | pgdat->totalreserve_pages = 0; | |
5771 | ||
cb45b0e9 HA |
5772 | for (i = 0; i < MAX_NR_ZONES; i++) { |
5773 | struct zone *zone = pgdat->node_zones + i; | |
3484b2de | 5774 | long max = 0; |
9705bea5 | 5775 | unsigned long managed_pages = zone_managed_pages(zone); |
cb45b0e9 HA |
5776 | |
5777 | /* Find valid and maximum lowmem_reserve in the zone */ | |
5778 | for (j = i; j < MAX_NR_ZONES; j++) { | |
5779 | if (zone->lowmem_reserve[j] > max) | |
5780 | max = zone->lowmem_reserve[j]; | |
5781 | } | |
5782 | ||
41858966 MG |
5783 | /* we treat the high watermark as reserved pages. */ |
5784 | max += high_wmark_pages(zone); | |
cb45b0e9 | 5785 | |
3d6357de AK |
5786 | if (max > managed_pages) |
5787 | max = managed_pages; | |
a8d01437 | 5788 | |
281e3726 | 5789 | pgdat->totalreserve_pages += max; |
a8d01437 | 5790 | |
cb45b0e9 HA |
5791 | reserve_pages += max; |
5792 | } | |
5793 | } | |
5794 | totalreserve_pages = reserve_pages; | |
5795 | } | |
5796 | ||
1da177e4 LT |
5797 | /* |
5798 | * setup_per_zone_lowmem_reserve - called whenever | |
34b10060 | 5799 | * sysctl_lowmem_reserve_ratio changes. Ensures that each zone |
1da177e4 LT |
5800 | * has a correct pages reserved value, so an adequate number of |
5801 | * pages are left in the zone after a successful __alloc_pages(). | |
5802 | */ | |
5803 | static void setup_per_zone_lowmem_reserve(void) | |
5804 | { | |
5805 | struct pglist_data *pgdat; | |
470c61d7 | 5806 | enum zone_type i, j; |
1da177e4 | 5807 | |
ec936fc5 | 5808 | for_each_online_pgdat(pgdat) { |
470c61d7 LS |
5809 | for (i = 0; i < MAX_NR_ZONES - 1; i++) { |
5810 | struct zone *zone = &pgdat->node_zones[i]; | |
5811 | int ratio = sysctl_lowmem_reserve_ratio[i]; | |
5812 | bool clear = !ratio || !zone_managed_pages(zone); | |
5813 | unsigned long managed_pages = 0; | |
5814 | ||
5815 | for (j = i + 1; j < MAX_NR_ZONES; j++) { | |
f7ec1044 LS |
5816 | struct zone *upper_zone = &pgdat->node_zones[j]; |
5817 | ||
5818 | managed_pages += zone_managed_pages(upper_zone); | |
470c61d7 | 5819 | |
f7ec1044 LS |
5820 | if (clear) |
5821 | zone->lowmem_reserve[j] = 0; | |
5822 | else | |
470c61d7 | 5823 | zone->lowmem_reserve[j] = managed_pages / ratio; |
1da177e4 LT |
5824 | } |
5825 | } | |
5826 | } | |
cb45b0e9 HA |
5827 | |
5828 | /* update totalreserve_pages */ | |
5829 | calculate_totalreserve_pages(); | |
1da177e4 LT |
5830 | } |
5831 | ||
cfd3da1e | 5832 | static void __setup_per_zone_wmarks(void) |
1da177e4 LT |
5833 | { |
5834 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); | |
5835 | unsigned long lowmem_pages = 0; | |
5836 | struct zone *zone; | |
5837 | unsigned long flags; | |
5838 | ||
416ef04f | 5839 | /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */ |
1da177e4 | 5840 | for_each_zone(zone) { |
416ef04f | 5841 | if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE) |
9705bea5 | 5842 | lowmem_pages += zone_managed_pages(zone); |
1da177e4 LT |
5843 | } |
5844 | ||
5845 | for_each_zone(zone) { | |
ac924c60 AM |
5846 | u64 tmp; |
5847 | ||
1125b4e3 | 5848 | spin_lock_irqsave(&zone->lock, flags); |
9705bea5 | 5849 | tmp = (u64)pages_min * zone_managed_pages(zone); |
ac924c60 | 5850 | do_div(tmp, lowmem_pages); |
416ef04f | 5851 | if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) { |
1da177e4 | 5852 | /* |
669ed175 | 5853 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't |
416ef04f | 5854 | * need highmem and movable zones pages, so cap pages_min |
5855 | * to a small value here. | |
669ed175 | 5856 | * |
41858966 | 5857 | * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) |
8bb4e7a2 | 5858 | * deltas control async page reclaim, and so should |
416ef04f | 5859 | * not be capped for highmem and movable zones. |
1da177e4 | 5860 | */ |
90ae8d67 | 5861 | unsigned long min_pages; |
1da177e4 | 5862 | |
9705bea5 | 5863 | min_pages = zone_managed_pages(zone) / 1024; |
90ae8d67 | 5864 | min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL); |
a9214443 | 5865 | zone->_watermark[WMARK_MIN] = min_pages; |
1da177e4 | 5866 | } else { |
669ed175 NP |
5867 | /* |
5868 | * If it's a lowmem zone, reserve a number of pages | |
1da177e4 LT |
5869 | * proportionate to the zone's size. |
5870 | */ | |
a9214443 | 5871 | zone->_watermark[WMARK_MIN] = tmp; |
1da177e4 LT |
5872 | } |
5873 | ||
795ae7a0 JW |
5874 | /* |
5875 | * Set the kswapd watermarks distance according to the | |
5876 | * scale factor in proportion to available memory, but | |
5877 | * ensure a minimum size on small systems. | |
5878 | */ | |
5879 | tmp = max_t(u64, tmp >> 2, | |
9705bea5 | 5880 | mult_frac(zone_managed_pages(zone), |
795ae7a0 JW |
5881 | watermark_scale_factor, 10000)); |
5882 | ||
aa092591 | 5883 | zone->watermark_boost = 0; |
a9214443 | 5884 | zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp; |
c574bbe9 HY |
5885 | zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp; |
5886 | zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp; | |
49f223a9 | 5887 | |
1125b4e3 | 5888 | spin_unlock_irqrestore(&zone->lock, flags); |
1da177e4 | 5889 | } |
cb45b0e9 HA |
5890 | |
5891 | /* update totalreserve_pages */ | |
5892 | calculate_totalreserve_pages(); | |
1da177e4 LT |
5893 | } |
5894 | ||
cfd3da1e MG |
5895 | /** |
5896 | * setup_per_zone_wmarks - called when min_free_kbytes changes | |
5897 | * or when memory is hot-{added|removed} | |
5898 | * | |
5899 | * Ensures that the watermark[min,low,high] values for each zone are set | |
5900 | * correctly with respect to min_free_kbytes. | |
5901 | */ | |
5902 | void setup_per_zone_wmarks(void) | |
5903 | { | |
b92ca18e | 5904 | struct zone *zone; |
b93e0f32 MH |
5905 | static DEFINE_SPINLOCK(lock); |
5906 | ||
5907 | spin_lock(&lock); | |
cfd3da1e | 5908 | __setup_per_zone_wmarks(); |
b93e0f32 | 5909 | spin_unlock(&lock); |
b92ca18e MG |
5910 | |
5911 | /* | |
5912 | * The watermark size have changed so update the pcpu batch | |
5913 | * and high limits or the limits may be inappropriate. | |
5914 | */ | |
5915 | for_each_zone(zone) | |
04f8cfea | 5916 | zone_pcp_update(zone, 0); |
cfd3da1e MG |
5917 | } |
5918 | ||
1da177e4 LT |
5919 | /* |
5920 | * Initialise min_free_kbytes. | |
5921 | * | |
5922 | * For small machines we want it small (128k min). For large machines | |
8beeae86 | 5923 | * we want it large (256MB max). But it is not linear, because network |
1da177e4 LT |
5924 | * bandwidth does not increase linearly with machine size. We use |
5925 | * | |
b8af2941 | 5926 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: |
1da177e4 LT |
5927 | * min_free_kbytes = sqrt(lowmem_kbytes * 16) |
5928 | * | |
5929 | * which yields | |
5930 | * | |
5931 | * 16MB: 512k | |
5932 | * 32MB: 724k | |
5933 | * 64MB: 1024k | |
5934 | * 128MB: 1448k | |
5935 | * 256MB: 2048k | |
5936 | * 512MB: 2896k | |
5937 | * 1024MB: 4096k | |
5938 | * 2048MB: 5792k | |
5939 | * 4096MB: 8192k | |
5940 | * 8192MB: 11584k | |
5941 | * 16384MB: 16384k | |
5942 | */ | |
bd3400ea | 5943 | void calculate_min_free_kbytes(void) |
1da177e4 LT |
5944 | { |
5945 | unsigned long lowmem_kbytes; | |
5f12733e | 5946 | int new_min_free_kbytes; |
1da177e4 LT |
5947 | |
5948 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); | |
5f12733e MH |
5949 | new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16); |
5950 | ||
59d336bd WS |
5951 | if (new_min_free_kbytes > user_min_free_kbytes) |
5952 | min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144); | |
5953 | else | |
5f12733e MH |
5954 | pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n", |
5955 | new_min_free_kbytes, user_min_free_kbytes); | |
59d336bd | 5956 | |
bd3400ea LF |
5957 | } |
5958 | ||
5959 | int __meminit init_per_zone_wmark_min(void) | |
5960 | { | |
5961 | calculate_min_free_kbytes(); | |
bc75d33f | 5962 | setup_per_zone_wmarks(); |
a6cccdc3 | 5963 | refresh_zone_stat_thresholds(); |
1da177e4 | 5964 | setup_per_zone_lowmem_reserve(); |
6423aa81 JK |
5965 | |
5966 | #ifdef CONFIG_NUMA | |
5967 | setup_min_unmapped_ratio(); | |
5968 | setup_min_slab_ratio(); | |
5969 | #endif | |
5970 | ||
4aab2be0 VB |
5971 | khugepaged_min_free_kbytes_update(); |
5972 | ||
1da177e4 LT |
5973 | return 0; |
5974 | } | |
e08d3fdf | 5975 | postcore_initcall(init_per_zone_wmark_min) |
1da177e4 LT |
5976 | |
5977 | /* | |
b8af2941 | 5978 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so |
1da177e4 LT |
5979 | * that we can call two helper functions whenever min_free_kbytes |
5980 | * changes. | |
5981 | */ | |
e95d372c | 5982 | static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 5983 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 5984 | { |
da8c757b HP |
5985 | int rc; |
5986 | ||
5987 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
5988 | if (rc) | |
5989 | return rc; | |
5990 | ||
5f12733e MH |
5991 | if (write) { |
5992 | user_min_free_kbytes = min_free_kbytes; | |
bc75d33f | 5993 | setup_per_zone_wmarks(); |
5f12733e | 5994 | } |
1da177e4 LT |
5995 | return 0; |
5996 | } | |
5997 | ||
e95d372c | 5998 | static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 5999 | void *buffer, size_t *length, loff_t *ppos) |
795ae7a0 JW |
6000 | { |
6001 | int rc; | |
6002 | ||
6003 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
6004 | if (rc) | |
6005 | return rc; | |
6006 | ||
6007 | if (write) | |
6008 | setup_per_zone_wmarks(); | |
6009 | ||
6010 | return 0; | |
6011 | } | |
6012 | ||
9614634f | 6013 | #ifdef CONFIG_NUMA |
6423aa81 | 6014 | static void setup_min_unmapped_ratio(void) |
9614634f | 6015 | { |
6423aa81 | 6016 | pg_data_t *pgdat; |
9614634f | 6017 | struct zone *zone; |
9614634f | 6018 | |
a5f5f91d | 6019 | for_each_online_pgdat(pgdat) |
81cbcbc2 | 6020 | pgdat->min_unmapped_pages = 0; |
a5f5f91d | 6021 | |
9614634f | 6022 | for_each_zone(zone) |
9705bea5 AK |
6023 | zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) * |
6024 | sysctl_min_unmapped_ratio) / 100; | |
9614634f | 6025 | } |
0ff38490 | 6026 | |
6423aa81 | 6027 | |
e95d372c | 6028 | static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 6029 | void *buffer, size_t *length, loff_t *ppos) |
0ff38490 | 6030 | { |
0ff38490 CL |
6031 | int rc; |
6032 | ||
8d65af78 | 6033 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); |
0ff38490 CL |
6034 | if (rc) |
6035 | return rc; | |
6036 | ||
6423aa81 JK |
6037 | setup_min_unmapped_ratio(); |
6038 | ||
6039 | return 0; | |
6040 | } | |
6041 | ||
6042 | static void setup_min_slab_ratio(void) | |
6043 | { | |
6044 | pg_data_t *pgdat; | |
6045 | struct zone *zone; | |
6046 | ||
a5f5f91d MG |
6047 | for_each_online_pgdat(pgdat) |
6048 | pgdat->min_slab_pages = 0; | |
6049 | ||
0ff38490 | 6050 | for_each_zone(zone) |
9705bea5 AK |
6051 | zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) * |
6052 | sysctl_min_slab_ratio) / 100; | |
6423aa81 JK |
6053 | } |
6054 | ||
e95d372c | 6055 | static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 6056 | void *buffer, size_t *length, loff_t *ppos) |
6423aa81 JK |
6057 | { |
6058 | int rc; | |
6059 | ||
6060 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); | |
6061 | if (rc) | |
6062 | return rc; | |
6063 | ||
6064 | setup_min_slab_ratio(); | |
6065 | ||
0ff38490 CL |
6066 | return 0; |
6067 | } | |
9614634f CL |
6068 | #endif |
6069 | ||
1da177e4 LT |
6070 | /* |
6071 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around | |
6072 | * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() | |
6073 | * whenever sysctl_lowmem_reserve_ratio changes. | |
6074 | * | |
6075 | * The reserve ratio obviously has absolutely no relation with the | |
41858966 | 6076 | * minimum watermarks. The lowmem reserve ratio can only make sense |
1da177e4 LT |
6077 | * if in function of the boot time zone sizes. |
6078 | */ | |
e95d372c KW |
6079 | static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, |
6080 | int write, void *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 6081 | { |
86aaf255 BH |
6082 | int i; |
6083 | ||
8d65af78 | 6084 | proc_dointvec_minmax(table, write, buffer, length, ppos); |
86aaf255 BH |
6085 | |
6086 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
6087 | if (sysctl_lowmem_reserve_ratio[i] < 1) | |
6088 | sysctl_lowmem_reserve_ratio[i] = 0; | |
6089 | } | |
6090 | ||
1da177e4 LT |
6091 | setup_per_zone_lowmem_reserve(); |
6092 | return 0; | |
6093 | } | |
6094 | ||
8ad4b1fb | 6095 | /* |
74f44822 MG |
6096 | * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each |
6097 | * cpu. It is the fraction of total pages in each zone that a hot per cpu | |
b8af2941 | 6098 | * pagelist can have before it gets flushed back to buddy allocator. |
8ad4b1fb | 6099 | */ |
e95d372c | 6100 | static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table, |
74f44822 | 6101 | int write, void *buffer, size_t *length, loff_t *ppos) |
8ad4b1fb RS |
6102 | { |
6103 | struct zone *zone; | |
74f44822 | 6104 | int old_percpu_pagelist_high_fraction; |
8ad4b1fb RS |
6105 | int ret; |
6106 | ||
7cd2b0a3 | 6107 | mutex_lock(&pcp_batch_high_lock); |
74f44822 | 6108 | old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction; |
7cd2b0a3 | 6109 | |
8d65af78 | 6110 | ret = proc_dointvec_minmax(table, write, buffer, length, ppos); |
7cd2b0a3 DR |
6111 | if (!write || ret < 0) |
6112 | goto out; | |
6113 | ||
6114 | /* Sanity checking to avoid pcp imbalance */ | |
74f44822 MG |
6115 | if (percpu_pagelist_high_fraction && |
6116 | percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) { | |
6117 | percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction; | |
7cd2b0a3 DR |
6118 | ret = -EINVAL; |
6119 | goto out; | |
6120 | } | |
6121 | ||
6122 | /* No change? */ | |
74f44822 | 6123 | if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction) |
7cd2b0a3 | 6124 | goto out; |
c8e251fa | 6125 | |
cb1ef534 | 6126 | for_each_populated_zone(zone) |
74f44822 | 6127 | zone_set_pageset_high_and_batch(zone, 0); |
7cd2b0a3 | 6128 | out: |
c8e251fa | 6129 | mutex_unlock(&pcp_batch_high_lock); |
7cd2b0a3 | 6130 | return ret; |
8ad4b1fb RS |
6131 | } |
6132 | ||
e95d372c KW |
6133 | static struct ctl_table page_alloc_sysctl_table[] = { |
6134 | { | |
6135 | .procname = "min_free_kbytes", | |
6136 | .data = &min_free_kbytes, | |
6137 | .maxlen = sizeof(min_free_kbytes), | |
6138 | .mode = 0644, | |
6139 | .proc_handler = min_free_kbytes_sysctl_handler, | |
6140 | .extra1 = SYSCTL_ZERO, | |
6141 | }, | |
6142 | { | |
6143 | .procname = "watermark_boost_factor", | |
6144 | .data = &watermark_boost_factor, | |
6145 | .maxlen = sizeof(watermark_boost_factor), | |
6146 | .mode = 0644, | |
6147 | .proc_handler = proc_dointvec_minmax, | |
6148 | .extra1 = SYSCTL_ZERO, | |
6149 | }, | |
6150 | { | |
6151 | .procname = "watermark_scale_factor", | |
6152 | .data = &watermark_scale_factor, | |
6153 | .maxlen = sizeof(watermark_scale_factor), | |
6154 | .mode = 0644, | |
6155 | .proc_handler = watermark_scale_factor_sysctl_handler, | |
6156 | .extra1 = SYSCTL_ONE, | |
6157 | .extra2 = SYSCTL_THREE_THOUSAND, | |
6158 | }, | |
6159 | { | |
6160 | .procname = "percpu_pagelist_high_fraction", | |
6161 | .data = &percpu_pagelist_high_fraction, | |
6162 | .maxlen = sizeof(percpu_pagelist_high_fraction), | |
6163 | .mode = 0644, | |
6164 | .proc_handler = percpu_pagelist_high_fraction_sysctl_handler, | |
6165 | .extra1 = SYSCTL_ZERO, | |
6166 | }, | |
6167 | { | |
6168 | .procname = "lowmem_reserve_ratio", | |
6169 | .data = &sysctl_lowmem_reserve_ratio, | |
6170 | .maxlen = sizeof(sysctl_lowmem_reserve_ratio), | |
6171 | .mode = 0644, | |
6172 | .proc_handler = lowmem_reserve_ratio_sysctl_handler, | |
6173 | }, | |
6174 | #ifdef CONFIG_NUMA | |
6175 | { | |
6176 | .procname = "numa_zonelist_order", | |
6177 | .data = &numa_zonelist_order, | |
6178 | .maxlen = NUMA_ZONELIST_ORDER_LEN, | |
6179 | .mode = 0644, | |
6180 | .proc_handler = numa_zonelist_order_handler, | |
6181 | }, | |
6182 | { | |
6183 | .procname = "min_unmapped_ratio", | |
6184 | .data = &sysctl_min_unmapped_ratio, | |
6185 | .maxlen = sizeof(sysctl_min_unmapped_ratio), | |
6186 | .mode = 0644, | |
6187 | .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler, | |
6188 | .extra1 = SYSCTL_ZERO, | |
6189 | .extra2 = SYSCTL_ONE_HUNDRED, | |
6190 | }, | |
6191 | { | |
6192 | .procname = "min_slab_ratio", | |
6193 | .data = &sysctl_min_slab_ratio, | |
6194 | .maxlen = sizeof(sysctl_min_slab_ratio), | |
6195 | .mode = 0644, | |
6196 | .proc_handler = sysctl_min_slab_ratio_sysctl_handler, | |
6197 | .extra1 = SYSCTL_ZERO, | |
6198 | .extra2 = SYSCTL_ONE_HUNDRED, | |
6199 | }, | |
6200 | #endif | |
6201 | {} | |
6202 | }; | |
6203 | ||
6204 | void __init page_alloc_sysctl_init(void) | |
6205 | { | |
6206 | register_sysctl_init("vm", page_alloc_sysctl_table); | |
6207 | } | |
6208 | ||
8df995f6 | 6209 | #ifdef CONFIG_CONTIG_ALLOC |
a1394bdd MK |
6210 | /* Usage: See admin-guide/dynamic-debug-howto.rst */ |
6211 | static void alloc_contig_dump_pages(struct list_head *page_list) | |
6212 | { | |
6213 | DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure"); | |
6214 | ||
6215 | if (DYNAMIC_DEBUG_BRANCH(descriptor)) { | |
6216 | struct page *page; | |
6217 | ||
6218 | dump_stack(); | |
6219 | list_for_each_entry(page, page_list, lru) | |
6220 | dump_page(page, "migration failure"); | |
6221 | } | |
6222 | } | |
a1394bdd | 6223 | |
041d3a8c | 6224 | /* [start, end) must belong to a single zone. */ |
b2c9e2fb | 6225 | int __alloc_contig_migrate_range(struct compact_control *cc, |
bb13ffeb | 6226 | unsigned long start, unsigned long end) |
041d3a8c MN |
6227 | { |
6228 | /* This function is based on compact_zone() from compaction.c. */ | |
730ec8c0 | 6229 | unsigned int nr_reclaimed; |
041d3a8c MN |
6230 | unsigned long pfn = start; |
6231 | unsigned int tries = 0; | |
6232 | int ret = 0; | |
8b94e0b8 JK |
6233 | struct migration_target_control mtc = { |
6234 | .nid = zone_to_nid(cc->zone), | |
6235 | .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL, | |
6236 | }; | |
041d3a8c | 6237 | |
361a2a22 | 6238 | lru_cache_disable(); |
041d3a8c | 6239 | |
bb13ffeb | 6240 | while (pfn < end || !list_empty(&cc->migratepages)) { |
041d3a8c MN |
6241 | if (fatal_signal_pending(current)) { |
6242 | ret = -EINTR; | |
6243 | break; | |
6244 | } | |
6245 | ||
bb13ffeb MG |
6246 | if (list_empty(&cc->migratepages)) { |
6247 | cc->nr_migratepages = 0; | |
c2ad7a1f OS |
6248 | ret = isolate_migratepages_range(cc, pfn, end); |
6249 | if (ret && ret != -EAGAIN) | |
041d3a8c | 6250 | break; |
c2ad7a1f | 6251 | pfn = cc->migrate_pfn; |
041d3a8c MN |
6252 | tries = 0; |
6253 | } else if (++tries == 5) { | |
c8e28b47 | 6254 | ret = -EBUSY; |
041d3a8c MN |
6255 | break; |
6256 | } | |
6257 | ||
beb51eaa MK |
6258 | nr_reclaimed = reclaim_clean_pages_from_list(cc->zone, |
6259 | &cc->migratepages); | |
6260 | cc->nr_migratepages -= nr_reclaimed; | |
02c6de8d | 6261 | |
8b94e0b8 | 6262 | ret = migrate_pages(&cc->migratepages, alloc_migration_target, |
5ac95884 | 6263 | NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL); |
c8e28b47 OS |
6264 | |
6265 | /* | |
6266 | * On -ENOMEM, migrate_pages() bails out right away. It is pointless | |
6267 | * to retry again over this error, so do the same here. | |
6268 | */ | |
6269 | if (ret == -ENOMEM) | |
6270 | break; | |
041d3a8c | 6271 | } |
d479960e | 6272 | |
361a2a22 | 6273 | lru_cache_enable(); |
2a6f5124 | 6274 | if (ret < 0) { |
3f913fc5 | 6275 | if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY) |
151e084a | 6276 | alloc_contig_dump_pages(&cc->migratepages); |
2a6f5124 SP |
6277 | putback_movable_pages(&cc->migratepages); |
6278 | return ret; | |
6279 | } | |
6280 | return 0; | |
041d3a8c MN |
6281 | } |
6282 | ||
6283 | /** | |
6284 | * alloc_contig_range() -- tries to allocate given range of pages | |
6285 | * @start: start PFN to allocate | |
6286 | * @end: one-past-the-last PFN to allocate | |
f0953a1b | 6287 | * @migratetype: migratetype of the underlying pageblocks (either |
0815f3d8 MN |
6288 | * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks |
6289 | * in range must have the same migratetype and it must | |
6290 | * be either of the two. | |
ca96b625 | 6291 | * @gfp_mask: GFP mask to use during compaction |
041d3a8c | 6292 | * |
11ac3e87 ZY |
6293 | * The PFN range does not have to be pageblock aligned. The PFN range must |
6294 | * belong to a single zone. | |
041d3a8c | 6295 | * |
2c7452a0 MK |
6296 | * The first thing this routine does is attempt to MIGRATE_ISOLATE all |
6297 | * pageblocks in the range. Once isolated, the pageblocks should not | |
6298 | * be modified by others. | |
041d3a8c | 6299 | * |
a862f68a | 6300 | * Return: zero on success or negative error code. On success all |
041d3a8c MN |
6301 | * pages which PFN is in [start, end) are allocated for the caller and |
6302 | * need to be freed with free_contig_range(). | |
6303 | */ | |
0815f3d8 | 6304 | int alloc_contig_range(unsigned long start, unsigned long end, |
ca96b625 | 6305 | unsigned migratetype, gfp_t gfp_mask) |
041d3a8c | 6306 | { |
041d3a8c | 6307 | unsigned long outer_start, outer_end; |
b2c9e2fb | 6308 | int order; |
d00181b9 | 6309 | int ret = 0; |
041d3a8c | 6310 | |
bb13ffeb MG |
6311 | struct compact_control cc = { |
6312 | .nr_migratepages = 0, | |
6313 | .order = -1, | |
6314 | .zone = page_zone(pfn_to_page(start)), | |
e0b9daeb | 6315 | .mode = MIGRATE_SYNC, |
bb13ffeb | 6316 | .ignore_skip_hint = true, |
2583d671 | 6317 | .no_set_skip_hint = true, |
7dea19f9 | 6318 | .gfp_mask = current_gfp_context(gfp_mask), |
b06eda09 | 6319 | .alloc_contig = true, |
bb13ffeb MG |
6320 | }; |
6321 | INIT_LIST_HEAD(&cc.migratepages); | |
6322 | ||
041d3a8c MN |
6323 | /* |
6324 | * What we do here is we mark all pageblocks in range as | |
6325 | * MIGRATE_ISOLATE. Because pageblock and max order pages may | |
6326 | * have different sizes, and due to the way page allocator | |
b2c9e2fb | 6327 | * work, start_isolate_page_range() has special handlings for this. |
041d3a8c MN |
6328 | * |
6329 | * Once the pageblocks are marked as MIGRATE_ISOLATE, we | |
6330 | * migrate the pages from an unaligned range (ie. pages that | |
b2c9e2fb | 6331 | * we are interested in). This will put all the pages in |
041d3a8c MN |
6332 | * range back to page allocator as MIGRATE_ISOLATE. |
6333 | * | |
6334 | * When this is done, we take the pages in range from page | |
6335 | * allocator removing them from the buddy system. This way | |
6336 | * page allocator will never consider using them. | |
6337 | * | |
6338 | * This lets us mark the pageblocks back as | |
6339 | * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the | |
6340 | * aligned range but not in the unaligned, original range are | |
6341 | * put back to page allocator so that buddy can use them. | |
6342 | */ | |
6343 | ||
6e263fff | 6344 | ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask); |
3fa0c7c7 | 6345 | if (ret) |
b2c9e2fb | 6346 | goto done; |
041d3a8c | 6347 | |
7612921f VB |
6348 | drain_all_pages(cc.zone); |
6349 | ||
8ef5849f JK |
6350 | /* |
6351 | * In case of -EBUSY, we'd like to know which page causes problem. | |
63cd4489 MK |
6352 | * So, just fall through. test_pages_isolated() has a tracepoint |
6353 | * which will report the busy page. | |
6354 | * | |
6355 | * It is possible that busy pages could become available before | |
6356 | * the call to test_pages_isolated, and the range will actually be | |
6357 | * allocated. So, if we fall through be sure to clear ret so that | |
6358 | * -EBUSY is not accidentally used or returned to caller. | |
8ef5849f | 6359 | */ |
bb13ffeb | 6360 | ret = __alloc_contig_migrate_range(&cc, start, end); |
8ef5849f | 6361 | if (ret && ret != -EBUSY) |
041d3a8c | 6362 | goto done; |
68d68ff6 | 6363 | ret = 0; |
041d3a8c MN |
6364 | |
6365 | /* | |
b2c9e2fb | 6366 | * Pages from [start, end) are within a pageblock_nr_pages |
041d3a8c MN |
6367 | * aligned blocks that are marked as MIGRATE_ISOLATE. What's |
6368 | * more, all pages in [start, end) are free in page allocator. | |
6369 | * What we are going to do is to allocate all pages from | |
6370 | * [start, end) (that is remove them from page allocator). | |
6371 | * | |
6372 | * The only problem is that pages at the beginning and at the | |
6373 | * end of interesting range may be not aligned with pages that | |
6374 | * page allocator holds, ie. they can be part of higher order | |
6375 | * pages. Because of this, we reserve the bigger range and | |
6376 | * once this is done free the pages we are not interested in. | |
6377 | * | |
6378 | * We don't have to hold zone->lock here because the pages are | |
6379 | * isolated thus they won't get removed from buddy. | |
6380 | */ | |
6381 | ||
041d3a8c MN |
6382 | order = 0; |
6383 | outer_start = start; | |
6384 | while (!PageBuddy(pfn_to_page(outer_start))) { | |
5e0a760b | 6385 | if (++order > MAX_PAGE_ORDER) { |
8ef5849f JK |
6386 | outer_start = start; |
6387 | break; | |
041d3a8c MN |
6388 | } |
6389 | outer_start &= ~0UL << order; | |
6390 | } | |
6391 | ||
8ef5849f | 6392 | if (outer_start != start) { |
ab130f91 | 6393 | order = buddy_order(pfn_to_page(outer_start)); |
8ef5849f JK |
6394 | |
6395 | /* | |
6396 | * outer_start page could be small order buddy page and | |
6397 | * it doesn't include start page. Adjust outer_start | |
6398 | * in this case to report failed page properly | |
6399 | * on tracepoint in test_pages_isolated() | |
6400 | */ | |
6401 | if (outer_start + (1UL << order) <= start) | |
6402 | outer_start = start; | |
6403 | } | |
6404 | ||
041d3a8c | 6405 | /* Make sure the range is really isolated. */ |
756d25be | 6406 | if (test_pages_isolated(outer_start, end, 0)) { |
041d3a8c MN |
6407 | ret = -EBUSY; |
6408 | goto done; | |
6409 | } | |
6410 | ||
49f223a9 | 6411 | /* Grab isolated pages from freelists. */ |
bb13ffeb | 6412 | outer_end = isolate_freepages_range(&cc, outer_start, end); |
041d3a8c MN |
6413 | if (!outer_end) { |
6414 | ret = -EBUSY; | |
6415 | goto done; | |
6416 | } | |
6417 | ||
6418 | /* Free head and tail (if any) */ | |
6419 | if (start != outer_start) | |
6420 | free_contig_range(outer_start, start - outer_start); | |
6421 | if (end != outer_end) | |
6422 | free_contig_range(end, outer_end - end); | |
6423 | ||
6424 | done: | |
6e263fff | 6425 | undo_isolate_page_range(start, end, migratetype); |
041d3a8c MN |
6426 | return ret; |
6427 | } | |
255f5985 | 6428 | EXPORT_SYMBOL(alloc_contig_range); |
5e27a2df AK |
6429 | |
6430 | static int __alloc_contig_pages(unsigned long start_pfn, | |
6431 | unsigned long nr_pages, gfp_t gfp_mask) | |
6432 | { | |
6433 | unsigned long end_pfn = start_pfn + nr_pages; | |
6434 | ||
6435 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, | |
6436 | gfp_mask); | |
6437 | } | |
6438 | ||
6439 | static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn, | |
6440 | unsigned long nr_pages) | |
6441 | { | |
6442 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
6443 | struct page *page; | |
6444 | ||
6445 | for (i = start_pfn; i < end_pfn; i++) { | |
6446 | page = pfn_to_online_page(i); | |
6447 | if (!page) | |
6448 | return false; | |
6449 | ||
6450 | if (page_zone(page) != z) | |
6451 | return false; | |
6452 | ||
6453 | if (PageReserved(page)) | |
4d73ba5f MG |
6454 | return false; |
6455 | ||
6456 | if (PageHuge(page)) | |
5e27a2df | 6457 | return false; |
5e27a2df AK |
6458 | } |
6459 | return true; | |
6460 | } | |
6461 | ||
6462 | static bool zone_spans_last_pfn(const struct zone *zone, | |
6463 | unsigned long start_pfn, unsigned long nr_pages) | |
6464 | { | |
6465 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
6466 | ||
6467 | return zone_spans_pfn(zone, last_pfn); | |
6468 | } | |
6469 | ||
6470 | /** | |
6471 | * alloc_contig_pages() -- tries to find and allocate contiguous range of pages | |
6472 | * @nr_pages: Number of contiguous pages to allocate | |
6473 | * @gfp_mask: GFP mask to limit search and used during compaction | |
6474 | * @nid: Target node | |
6475 | * @nodemask: Mask for other possible nodes | |
6476 | * | |
6477 | * This routine is a wrapper around alloc_contig_range(). It scans over zones | |
6478 | * on an applicable zonelist to find a contiguous pfn range which can then be | |
6479 | * tried for allocation with alloc_contig_range(). This routine is intended | |
6480 | * for allocation requests which can not be fulfilled with the buddy allocator. | |
6481 | * | |
6482 | * The allocated memory is always aligned to a page boundary. If nr_pages is a | |
eaab8e75 AK |
6483 | * power of two, then allocated range is also guaranteed to be aligned to same |
6484 | * nr_pages (e.g. 1GB request would be aligned to 1GB). | |
5e27a2df AK |
6485 | * |
6486 | * Allocated pages can be freed with free_contig_range() or by manually calling | |
6487 | * __free_page() on each allocated page. | |
6488 | * | |
6489 | * Return: pointer to contiguous pages on success, or NULL if not successful. | |
6490 | */ | |
6491 | struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask, | |
6492 | int nid, nodemask_t *nodemask) | |
6493 | { | |
6494 | unsigned long ret, pfn, flags; | |
6495 | struct zonelist *zonelist; | |
6496 | struct zone *zone; | |
6497 | struct zoneref *z; | |
6498 | ||
6499 | zonelist = node_zonelist(nid, gfp_mask); | |
6500 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
6501 | gfp_zone(gfp_mask), nodemask) { | |
6502 | spin_lock_irqsave(&zone->lock, flags); | |
6503 | ||
6504 | pfn = ALIGN(zone->zone_start_pfn, nr_pages); | |
6505 | while (zone_spans_last_pfn(zone, pfn, nr_pages)) { | |
6506 | if (pfn_range_valid_contig(zone, pfn, nr_pages)) { | |
6507 | /* | |
6508 | * We release the zone lock here because | |
6509 | * alloc_contig_range() will also lock the zone | |
6510 | * at some point. If there's an allocation | |
6511 | * spinning on this lock, it may win the race | |
6512 | * and cause alloc_contig_range() to fail... | |
6513 | */ | |
6514 | spin_unlock_irqrestore(&zone->lock, flags); | |
6515 | ret = __alloc_contig_pages(pfn, nr_pages, | |
6516 | gfp_mask); | |
6517 | if (!ret) | |
6518 | return pfn_to_page(pfn); | |
6519 | spin_lock_irqsave(&zone->lock, flags); | |
6520 | } | |
6521 | pfn += nr_pages; | |
6522 | } | |
6523 | spin_unlock_irqrestore(&zone->lock, flags); | |
6524 | } | |
6525 | return NULL; | |
6526 | } | |
4eb0716e | 6527 | #endif /* CONFIG_CONTIG_ALLOC */ |
041d3a8c | 6528 | |
78fa5150 | 6529 | void free_contig_range(unsigned long pfn, unsigned long nr_pages) |
041d3a8c | 6530 | { |
78fa5150 | 6531 | unsigned long count = 0; |
bcc2b02f MS |
6532 | |
6533 | for (; nr_pages--; pfn++) { | |
6534 | struct page *page = pfn_to_page(pfn); | |
6535 | ||
6536 | count += page_count(page) != 1; | |
6537 | __free_page(page); | |
6538 | } | |
78fa5150 | 6539 | WARN(count != 0, "%lu pages are still in use!\n", count); |
041d3a8c | 6540 | } |
255f5985 | 6541 | EXPORT_SYMBOL(free_contig_range); |
041d3a8c | 6542 | |
ec6e8c7e VB |
6543 | /* |
6544 | * Effectively disable pcplists for the zone by setting the high limit to 0 | |
6545 | * and draining all cpus. A concurrent page freeing on another CPU that's about | |
6546 | * to put the page on pcplist will either finish before the drain and the page | |
6547 | * will be drained, or observe the new high limit and skip the pcplist. | |
6548 | * | |
6549 | * Must be paired with a call to zone_pcp_enable(). | |
6550 | */ | |
6551 | void zone_pcp_disable(struct zone *zone) | |
6552 | { | |
6553 | mutex_lock(&pcp_batch_high_lock); | |
90b41691 | 6554 | __zone_set_pageset_high_and_batch(zone, 0, 0, 1); |
ec6e8c7e VB |
6555 | __drain_all_pages(zone, true); |
6556 | } | |
6557 | ||
6558 | void zone_pcp_enable(struct zone *zone) | |
6559 | { | |
90b41691 HY |
6560 | __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min, |
6561 | zone->pageset_high_max, zone->pageset_batch); | |
ec6e8c7e VB |
6562 | mutex_unlock(&pcp_batch_high_lock); |
6563 | } | |
6564 | ||
340175b7 JL |
6565 | void zone_pcp_reset(struct zone *zone) |
6566 | { | |
5a883813 | 6567 | int cpu; |
28f836b6 | 6568 | struct per_cpu_zonestat *pzstats; |
340175b7 | 6569 | |
28f836b6 | 6570 | if (zone->per_cpu_pageset != &boot_pageset) { |
5a883813 | 6571 | for_each_online_cpu(cpu) { |
28f836b6 MG |
6572 | pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu); |
6573 | drain_zonestat(zone, pzstats); | |
5a883813 | 6574 | } |
28f836b6 | 6575 | free_percpu(zone->per_cpu_pageset); |
28f836b6 | 6576 | zone->per_cpu_pageset = &boot_pageset; |
022e7fa0 ML |
6577 | if (zone->per_cpu_zonestats != &boot_zonestats) { |
6578 | free_percpu(zone->per_cpu_zonestats); | |
6579 | zone->per_cpu_zonestats = &boot_zonestats; | |
6580 | } | |
340175b7 | 6581 | } |
340175b7 JL |
6582 | } |
6583 | ||
6dcd73d7 | 6584 | #ifdef CONFIG_MEMORY_HOTREMOVE |
0c0e6195 | 6585 | /* |
257bea71 DH |
6586 | * All pages in the range must be in a single zone, must not contain holes, |
6587 | * must span full sections, and must be isolated before calling this function. | |
0c0e6195 | 6588 | */ |
257bea71 | 6589 | void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) |
0c0e6195 | 6590 | { |
257bea71 | 6591 | unsigned long pfn = start_pfn; |
0c0e6195 KH |
6592 | struct page *page; |
6593 | struct zone *zone; | |
0ee5f4f3 | 6594 | unsigned int order; |
0c0e6195 | 6595 | unsigned long flags; |
5557c766 | 6596 | |
2d070eab | 6597 | offline_mem_sections(pfn, end_pfn); |
0c0e6195 KH |
6598 | zone = page_zone(pfn_to_page(pfn)); |
6599 | spin_lock_irqsave(&zone->lock, flags); | |
0c0e6195 | 6600 | while (pfn < end_pfn) { |
0c0e6195 | 6601 | page = pfn_to_page(pfn); |
b023f468 WC |
6602 | /* |
6603 | * The HWPoisoned page may be not in buddy system, and | |
6604 | * page_count() is not 0. | |
6605 | */ | |
6606 | if (unlikely(!PageBuddy(page) && PageHWPoison(page))) { | |
6607 | pfn++; | |
b023f468 WC |
6608 | continue; |
6609 | } | |
aa218795 DH |
6610 | /* |
6611 | * At this point all remaining PageOffline() pages have a | |
6612 | * reference count of 0 and can simply be skipped. | |
6613 | */ | |
6614 | if (PageOffline(page)) { | |
6615 | BUG_ON(page_count(page)); | |
6616 | BUG_ON(PageBuddy(page)); | |
6617 | pfn++; | |
aa218795 DH |
6618 | continue; |
6619 | } | |
b023f468 | 6620 | |
0c0e6195 KH |
6621 | BUG_ON(page_count(page)); |
6622 | BUG_ON(!PageBuddy(page)); | |
ab130f91 | 6623 | order = buddy_order(page); |
6ab01363 | 6624 | del_page_from_free_list(page, zone, order); |
0c0e6195 KH |
6625 | pfn += (1 << order); |
6626 | } | |
6627 | spin_unlock_irqrestore(&zone->lock, flags); | |
6628 | } | |
6629 | #endif | |
8d22ba1b | 6630 | |
8446b59b ED |
6631 | /* |
6632 | * This function returns a stable result only if called under zone lock. | |
6633 | */ | |
8d22ba1b WF |
6634 | bool is_free_buddy_page(struct page *page) |
6635 | { | |
8d22ba1b | 6636 | unsigned long pfn = page_to_pfn(page); |
7aeb09f9 | 6637 | unsigned int order; |
8d22ba1b | 6638 | |
fd377218 | 6639 | for (order = 0; order < NR_PAGE_ORDERS; order++) { |
8d22ba1b WF |
6640 | struct page *page_head = page - (pfn & ((1 << order) - 1)); |
6641 | ||
8446b59b ED |
6642 | if (PageBuddy(page_head) && |
6643 | buddy_order_unsafe(page_head) >= order) | |
8d22ba1b WF |
6644 | break; |
6645 | } | |
8d22ba1b | 6646 | |
5e0a760b | 6647 | return order <= MAX_PAGE_ORDER; |
8d22ba1b | 6648 | } |
a581865e | 6649 | EXPORT_SYMBOL(is_free_buddy_page); |
d4ae9916 NH |
6650 | |
6651 | #ifdef CONFIG_MEMORY_FAILURE | |
6652 | /* | |
06be6ff3 OS |
6653 | * Break down a higher-order page in sub-pages, and keep our target out of |
6654 | * buddy allocator. | |
d4ae9916 | 6655 | */ |
06be6ff3 OS |
6656 | static void break_down_buddy_pages(struct zone *zone, struct page *page, |
6657 | struct page *target, int low, int high, | |
6658 | int migratetype) | |
6659 | { | |
6660 | unsigned long size = 1 << high; | |
0dfca313 | 6661 | struct page *current_buddy; |
06be6ff3 OS |
6662 | |
6663 | while (high > low) { | |
6664 | high--; | |
6665 | size >>= 1; | |
6666 | ||
6667 | if (target >= &page[size]) { | |
06be6ff3 | 6668 | current_buddy = page; |
0dfca313 | 6669 | page = page + size; |
06be6ff3 | 6670 | } else { |
06be6ff3 OS |
6671 | current_buddy = page + size; |
6672 | } | |
6673 | ||
6674 | if (set_page_guard(zone, current_buddy, high, migratetype)) | |
6675 | continue; | |
6676 | ||
27e0db3c KS |
6677 | add_to_free_list(current_buddy, zone, high, migratetype); |
6678 | set_buddy_order(current_buddy, high); | |
06be6ff3 OS |
6679 | } |
6680 | } | |
6681 | ||
6682 | /* | |
6683 | * Take a page that will be marked as poisoned off the buddy allocator. | |
6684 | */ | |
6685 | bool take_page_off_buddy(struct page *page) | |
d4ae9916 NH |
6686 | { |
6687 | struct zone *zone = page_zone(page); | |
6688 | unsigned long pfn = page_to_pfn(page); | |
6689 | unsigned long flags; | |
6690 | unsigned int order; | |
06be6ff3 | 6691 | bool ret = false; |
d4ae9916 NH |
6692 | |
6693 | spin_lock_irqsave(&zone->lock, flags); | |
fd377218 | 6694 | for (order = 0; order < NR_PAGE_ORDERS; order++) { |
d4ae9916 | 6695 | struct page *page_head = page - (pfn & ((1 << order) - 1)); |
ab130f91 | 6696 | int page_order = buddy_order(page_head); |
d4ae9916 | 6697 | |
ab130f91 | 6698 | if (PageBuddy(page_head) && page_order >= order) { |
06be6ff3 OS |
6699 | unsigned long pfn_head = page_to_pfn(page_head); |
6700 | int migratetype = get_pfnblock_migratetype(page_head, | |
6701 | pfn_head); | |
6702 | ||
ab130f91 | 6703 | del_page_from_free_list(page_head, zone, page_order); |
06be6ff3 | 6704 | break_down_buddy_pages(zone, page_head, page, 0, |
ab130f91 | 6705 | page_order, migratetype); |
bf181c58 | 6706 | SetPageHWPoisonTakenOff(page); |
bac9c6fa DH |
6707 | if (!is_migrate_isolate(migratetype)) |
6708 | __mod_zone_freepage_state(zone, -1, migratetype); | |
06be6ff3 | 6709 | ret = true; |
d4ae9916 NH |
6710 | break; |
6711 | } | |
06be6ff3 OS |
6712 | if (page_count(page_head) > 0) |
6713 | break; | |
d4ae9916 NH |
6714 | } |
6715 | spin_unlock_irqrestore(&zone->lock, flags); | |
06be6ff3 | 6716 | return ret; |
d4ae9916 | 6717 | } |
bf181c58 NH |
6718 | |
6719 | /* | |
6720 | * Cancel takeoff done by take_page_off_buddy(). | |
6721 | */ | |
6722 | bool put_page_back_buddy(struct page *page) | |
6723 | { | |
6724 | struct zone *zone = page_zone(page); | |
6725 | unsigned long pfn = page_to_pfn(page); | |
6726 | unsigned long flags; | |
6727 | int migratetype = get_pfnblock_migratetype(page, pfn); | |
6728 | bool ret = false; | |
6729 | ||
6730 | spin_lock_irqsave(&zone->lock, flags); | |
6731 | if (put_page_testzero(page)) { | |
6732 | ClearPageHWPoisonTakenOff(page); | |
6733 | __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE); | |
6734 | if (TestClearPageHWPoison(page)) { | |
bf181c58 NH |
6735 | ret = true; |
6736 | } | |
6737 | } | |
6738 | spin_unlock_irqrestore(&zone->lock, flags); | |
6739 | ||
6740 | return ret; | |
6741 | } | |
d4ae9916 | 6742 | #endif |
62b31070 BH |
6743 | |
6744 | #ifdef CONFIG_ZONE_DMA | |
6745 | bool has_managed_dma(void) | |
6746 | { | |
6747 | struct pglist_data *pgdat; | |
6748 | ||
6749 | for_each_online_pgdat(pgdat) { | |
6750 | struct zone *zone = &pgdat->node_zones[ZONE_DMA]; | |
6751 | ||
6752 | if (managed_zone(zone)) | |
6753 | return true; | |
6754 | } | |
6755 | return false; | |
6756 | } | |
6757 | #endif /* CONFIG_ZONE_DMA */ | |
dcdfdd40 KS |
6758 | |
6759 | #ifdef CONFIG_UNACCEPTED_MEMORY | |
6760 | ||
6761 | /* Counts number of zones with unaccepted pages. */ | |
6762 | static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages); | |
6763 | ||
6764 | static bool lazy_accept = true; | |
6765 | ||
6766 | static int __init accept_memory_parse(char *p) | |
6767 | { | |
6768 | if (!strcmp(p, "lazy")) { | |
6769 | lazy_accept = true; | |
6770 | return 0; | |
6771 | } else if (!strcmp(p, "eager")) { | |
6772 | lazy_accept = false; | |
6773 | return 0; | |
6774 | } else { | |
6775 | return -EINVAL; | |
6776 | } | |
6777 | } | |
6778 | early_param("accept_memory", accept_memory_parse); | |
6779 | ||
6780 | static bool page_contains_unaccepted(struct page *page, unsigned int order) | |
6781 | { | |
6782 | phys_addr_t start = page_to_phys(page); | |
6783 | phys_addr_t end = start + (PAGE_SIZE << order); | |
6784 | ||
6785 | return range_contains_unaccepted_memory(start, end); | |
6786 | } | |
6787 | ||
6788 | static void accept_page(struct page *page, unsigned int order) | |
6789 | { | |
6790 | phys_addr_t start = page_to_phys(page); | |
6791 | ||
6792 | accept_memory(start, start + (PAGE_SIZE << order)); | |
6793 | } | |
6794 | ||
6795 | static bool try_to_accept_memory_one(struct zone *zone) | |
6796 | { | |
6797 | unsigned long flags; | |
6798 | struct page *page; | |
6799 | bool last; | |
6800 | ||
6801 | if (list_empty(&zone->unaccepted_pages)) | |
6802 | return false; | |
6803 | ||
6804 | spin_lock_irqsave(&zone->lock, flags); | |
6805 | page = list_first_entry_or_null(&zone->unaccepted_pages, | |
6806 | struct page, lru); | |
6807 | if (!page) { | |
6808 | spin_unlock_irqrestore(&zone->lock, flags); | |
6809 | return false; | |
6810 | } | |
6811 | ||
6812 | list_del(&page->lru); | |
6813 | last = list_empty(&zone->unaccepted_pages); | |
6814 | ||
6815 | __mod_zone_freepage_state(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE); | |
6816 | __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES); | |
6817 | spin_unlock_irqrestore(&zone->lock, flags); | |
6818 | ||
5e0a760b | 6819 | accept_page(page, MAX_PAGE_ORDER); |
dcdfdd40 | 6820 | |
5e0a760b | 6821 | __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL); |
dcdfdd40 KS |
6822 | |
6823 | if (last) | |
6824 | static_branch_dec(&zones_with_unaccepted_pages); | |
6825 | ||
6826 | return true; | |
6827 | } | |
6828 | ||
6829 | static bool try_to_accept_memory(struct zone *zone, unsigned int order) | |
6830 | { | |
6831 | long to_accept; | |
6832 | int ret = false; | |
6833 | ||
6834 | /* How much to accept to get to high watermark? */ | |
6835 | to_accept = high_wmark_pages(zone) - | |
6836 | (zone_page_state(zone, NR_FREE_PAGES) - | |
6837 | __zone_watermark_unusable_free(zone, order, 0)); | |
6838 | ||
6839 | /* Accept at least one page */ | |
6840 | do { | |
6841 | if (!try_to_accept_memory_one(zone)) | |
6842 | break; | |
6843 | ret = true; | |
6844 | to_accept -= MAX_ORDER_NR_PAGES; | |
6845 | } while (to_accept > 0); | |
6846 | ||
6847 | return ret; | |
6848 | } | |
6849 | ||
6850 | static inline bool has_unaccepted_memory(void) | |
6851 | { | |
6852 | return static_branch_unlikely(&zones_with_unaccepted_pages); | |
6853 | } | |
6854 | ||
6855 | static bool __free_unaccepted(struct page *page) | |
6856 | { | |
6857 | struct zone *zone = page_zone(page); | |
6858 | unsigned long flags; | |
6859 | bool first = false; | |
6860 | ||
6861 | if (!lazy_accept) | |
6862 | return false; | |
6863 | ||
6864 | spin_lock_irqsave(&zone->lock, flags); | |
6865 | first = list_empty(&zone->unaccepted_pages); | |
6866 | list_add_tail(&page->lru, &zone->unaccepted_pages); | |
6867 | __mod_zone_freepage_state(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE); | |
6868 | __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES); | |
6869 | spin_unlock_irqrestore(&zone->lock, flags); | |
6870 | ||
6871 | if (first) | |
6872 | static_branch_inc(&zones_with_unaccepted_pages); | |
6873 | ||
6874 | return true; | |
6875 | } | |
6876 | ||
6877 | #else | |
6878 | ||
6879 | static bool page_contains_unaccepted(struct page *page, unsigned int order) | |
6880 | { | |
6881 | return false; | |
6882 | } | |
6883 | ||
6884 | static void accept_page(struct page *page, unsigned int order) | |
6885 | { | |
6886 | } | |
6887 | ||
6888 | static bool try_to_accept_memory(struct zone *zone, unsigned int order) | |
6889 | { | |
6890 | return false; | |
6891 | } | |
6892 | ||
6893 | static inline bool has_unaccepted_memory(void) | |
6894 | { | |
6895 | return false; | |
6896 | } | |
6897 | ||
6898 | static bool __free_unaccepted(struct page *page) | |
6899 | { | |
6900 | BUILD_BUG(); | |
6901 | return false; | |
6902 | } | |
6903 | ||
6904 | #endif /* CONFIG_UNACCEPTED_MEMORY */ |