| 1 | |
| 2 | |
| 3 | *
|
| 4 | * Manages the free list, the system allocates free pages here.
|
| 5 | * Note that kmalloc() lives in slab.c
|
| 6 | *
|
| 7 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
|
| 8 | * Swap reorganised 29.12.95, Stephen Tweedie
|
| 9 | * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
|
| 10 | * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
|
| 11 | * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
|
| 12 | * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
|
| 13 | * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
|
| 14 | * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
|
| 15 |
|
| 16 |
|
| 17 | #include <linux/stddef.h>
|
| 18 | #include <linux/mm.h>
|
| 19 | #include <linux/highmem.h>
|
| 20 | #include <linux/interrupt.h>
|
| 21 | #include <linux/jiffies.h>
|
| 22 | #include <linux/compiler.h>
|
| 23 | #include <linux/kernel.h>
|
| 24 | #include <linux/kasan.h>
|
| 25 | #include <linux/kmsan.h>
|
| 26 | #include <linux/module.h>
|
| 27 | #include <linux/suspend.h>
|
| 28 | #include <linux/ratelimit.h>
|
| 29 | #include <linux/oom.h>
|
| 30 | #include <linux/topology.h>
|
| 31 | #include <linux/sysctl.h>
|
| 32 | #include <linux/cpu.h>
|
| 33 | #include <linux/cpuset.h>
|
| 34 | #include <linux/pagevec.h>
|
| 35 | #include <linux/memory_hotplug.h>
|
| 36 | #include <linux/nodemask.h>
|
| 37 | #include <linux/vmstat.h>
|
| 38 | #include <linux/fault-inject.h>
|
| 39 | #include <linux/compaction.h>
|
| 40 | #include <trace/events/kmem.h>
|
| 41 | #include <trace/events/oom.h>
|
| 42 | #include <linux/prefetch.h>
|
| 43 | #include <linux/mm_inline.h>
|
| 44 | #include <linux/mmu_notifier.h>
|
| 45 | #include <linux/migrate.h>
|
| 46 | #include <linux/sched/mm.h>
|
| 47 | #include <linux/page_owner.h>
|
| 48 | #include <linux/page_table_check.h>
|
| 49 | #include <linux/memcontrol.h>
|
| 50 | #include <linux/ftrace.h>
|
| 51 | #include <linux/lockdep.h>
|
| 52 | #include <linux/psi.h>
|
| 53 | #include <linux/khugepaged.h>
|
| 54 | #include <linux/delayacct.h>
|
| 55 | #include <linux/cacheinfo.h>
|
| 56 | #include <linux/pgalloc_tag.h>
|
| 57 | #include <asm/div64.h>
|
| 58 | #include "internal.h"
|
| 59 | #include "shuffle.h"
|
| 60 | #include "page_reporting.h"
|
| 61 |
|
| 62 |
|
| 63 | typedef int __bitwise fpi_t;
|
| 64 |
|
| 65 |
|
| 66 | #define FPI_NONE ((__force fpi_t)0)
|
| 67 |
|
| 68 | |
| 69 | * Skip free page reporting notification for the (possibly merged) page.
|
| 70 | * This does not hinder free page reporting from grabbing the page,
|
| 71 | * reporting it and marking it "reported" - it only skips notifying
|
| 72 | * the free page reporting infrastructure about a newly freed page. For
|
| 73 | * example, used when temporarily pulling a page from a freelist and
|
| 74 | * putting it back unmodified.
|
| 75 |
|
| 76 | #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
|
| 77 |
|
| 78 | |
| 79 | * Place the (possibly merged) page to the tail of the freelist. Will ignore
|
| 80 | * page shuffling (relevant code - e.g., memory onlining - is expected to
|
| 81 | * shuffle the whole zone).
|
| 82 | *
|
| 83 | * Note: No code should rely on this flag for correctness - it's purely
|
| 84 | * to allow for optimizations when handing back either fresh pages
|
| 85 | * (memory onlining) or untouched pages (page isolation, free page
|
| 86 | * reporting).
|
| 87 |
|
| 88 | #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
|
| 89 |
|
| 90 |
|
| 91 | #define FPI_TRYLOCK ((__force fpi_t)BIT(2))
|
| 92 |
|
| 93 |
|
| 94 | static DEFINE_MUTEX(pcp_batch_high_lock);
|
| 95 | #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
|
| 96 |
|
| 97 | #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
|
| 98 | |
| 99 | * On SMP, spin_trylock is sufficient protection.
|
| 100 | * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
|
| 101 | * Pass flags to a no-op inline function to typecheck and silence the unused
|
| 102 | * variable warning.
|
| 103 |
|
| 104 | static inline void __pcp_trylock_noop(unsigned long *flags) { }
|
| 105 | #define pcp_trylock_prepare(flags) __pcp_trylock_noop(&(flags))
|
| 106 | #define pcp_trylock_finish(flags) __pcp_trylock_noop(&(flags))
|
| 107 | #else
|
| 108 |
|
| 109 |
|
| 110 | #define pcp_trylock_prepare(flags) local_irq_save(flags)
|
| 111 | #define pcp_trylock_finish(flags) local_irq_restore(flags)
|
| 112 | #endif
|
| 113 |
|
| 114 | |
| 115 | * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
|
| 116 | * a migration causing the wrong PCP to be locked and remote memory being
|
| 117 | * potentially allocated, pin the task to the CPU for the lookup+lock.
|
| 118 | * preempt_disable is used on !RT because it is faster than migrate_disable.
|
| 119 | * migrate_disable is used on RT because otherwise RT spinlock usage is
|
| 120 | * interfered with and a high priority task cannot preempt the allocator.
|
| 121 |
|
| 122 | #ifndef CONFIG_PREEMPT_RT
|
| 123 | #define pcpu_task_pin() preempt_disable()
|
| 124 | #define pcpu_task_unpin() preempt_enable()
|
| 125 | #else
|
| 126 | #define pcpu_task_pin() migrate_disable()
|
| 127 | #define pcpu_task_unpin() migrate_enable()
|
| 128 | #endif
|
| 129 |
|
| 130 | |
| 131 | * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
|
| 132 | * Return value should be used with equivalent unlock helper.
|
| 133 |
|
| 134 | #define pcpu_spin_trylock(type, member, ptr) \
|
| 135 | ({ \
|
| 136 | type *_ret; \
|
| 137 | pcpu_task_pin(); \
|
| 138 | _ret = this_cpu_ptr(ptr); \
|
| 139 | if (!spin_trylock(&_ret->member)) { \
|
| 140 | pcpu_task_unpin(); \
|
| 141 | _ret = NULL; \
|
| 142 | } \
|
| 143 | _ret; \
|
| 144 | })
|
| 145 |
|
| 146 | #define pcpu_spin_unlock(member, ptr) \
|
| 147 | ({ \
|
| 148 | spin_unlock(&ptr->member); \
|
| 149 | pcpu_task_unpin(); \
|
| 150 | })
|
| 151 |
|
| 152 |
|
| 153 | #define pcp_spin_trylock(ptr, UP_flags) \
|
| 154 | ({ \
|
| 155 | struct per_cpu_pages *__ret; \
|
| 156 | pcp_trylock_prepare(UP_flags); \
|
| 157 | __ret = pcpu_spin_trylock(struct per_cpu_pages, lock, ptr); \
|
| 158 | if (!__ret) \
|
| 159 | pcp_trylock_finish(UP_flags); \
|
| 160 | __ret; \
|
| 161 | })
|
| 162 |
|
| 163 | #define pcp_spin_unlock(ptr, UP_flags) \
|
| 164 | ({ \
|
| 165 | pcpu_spin_unlock(lock, ptr); \
|
| 166 | pcp_trylock_finish(UP_flags); \
|
| 167 | })
|
| 168 |
|
| 169 | |
| 170 | * With the UP spinlock implementation, when we spin_lock(&pcp->lock) (for i.e.
|
| 171 | * a potentially remote cpu drain) and get interrupted by an operation that
|
| 172 | * attempts pcp_spin_trylock(), we can't rely on the trylock failure due to UP
|
| 173 | * spinlock assumptions making the trylock a no-op. So we have to turn that
|
| 174 | * spin_lock() to a spin_lock_irqsave(). This works because on UP there are no
|
| 175 | * remote cpu's so we can only be locking the only existing local one.
|
| 176 |
|
| 177 | #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
|
| 178 | static inline void __flags_noop(unsigned long *flags) { }
|
| 179 | #define pcp_spin_lock_maybe_irqsave(ptr, flags) \
|
| 180 | ({ \
|
| 181 | __flags_noop(&(flags)); \
|
| 182 | spin_lock(&(ptr)->lock); \
|
| 183 | })
|
| 184 | #define pcp_spin_unlock_maybe_irqrestore(ptr, flags) \
|
| 185 | ({ \
|
| 186 | spin_unlock(&(ptr)->lock); \
|
| 187 | __flags_noop(&(flags)); \
|
| 188 | })
|
| 189 | #else
|
| 190 | #define pcp_spin_lock_maybe_irqsave(ptr, flags) \
|
| 191 | spin_lock_irqsave(&(ptr)->lock, flags)
|
| 192 | #define pcp_spin_unlock_maybe_irqrestore(ptr, flags) \
|
| 193 | spin_unlock_irqrestore(&(ptr)->lock, flags)
|
| 194 | #endif
|
| 195 |
|
| 196 | #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
|
| 197 | DEFINE_PER_CPU(int, numa_node);
|
| 198 | EXPORT_PER_CPU_SYMBOL(numa_node);
|
| 199 | #endif
|
| 200 |
|
| 201 | DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
|
| 202 |
|
| 203 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES
|
| 204 | |
| 205 | * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
|
| 206 | * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
|
| 207 | * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
|
| 208 | * defined in <linux/topology.h>.
|
| 209 |
|
| 210 | DEFINE_PER_CPU(int, _numa_mem_);
|
| 211 | EXPORT_PER_CPU_SYMBOL(_numa_mem_);
|
| 212 | #endif
|
| 213 |
|
| 214 | static DEFINE_MUTEX(pcpu_drain_mutex);
|
| 215 |
|
| 216 | #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
|
| 217 | volatile unsigned long latent_entropy __latent_entropy;
|
| 218 | EXPORT_SYMBOL(latent_entropy);
|
| 219 | #endif
|
| 220 |
|
| 221 | |
| 222 | * Array of node states.
|
| 223 |
|
| 224 | nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
|
| 225 | [N_POSSIBLE] = NODE_MASK_ALL,
|
| 226 | [N_ONLINE] = { { [0] = 1UL } },
|
| 227 | #ifndef CONFIG_NUMA
|
| 228 | [N_NORMAL_MEMORY] = { { [0] = 1UL } },
|
| 229 | #ifdef CONFIG_HIGHMEM
|
| 230 | [N_HIGH_MEMORY] = { { [0] = 1UL } },
|
| 231 | #endif
|
| 232 | [N_MEMORY] = { { [0] = 1UL } },
|
| 233 | [N_CPU] = { { [0] = 1UL } },
|
| 234 | #endif
|
| 235 | };
|
| 236 | EXPORT_SYMBOL(node_states);
|
| 237 |
|
| 238 | gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
|
| 239 |
|
| 240 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
|
| 241 | unsigned int pageblock_order __read_mostly;
|
| 242 | #endif
|
| 243 |
|
| 244 | static void __free_pages_ok(struct page *page, unsigned int order,
|
| 245 | fpi_t fpi_flags);
|
| 246 |
|
| 247 | |
| 248 | * results with 256, 32 in the lowmem_reserve sysctl:
|
| 249 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
|
| 250 | * 1G machine -> (16M dma, 784M normal, 224M high)
|
| 251 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
|
| 252 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
|
| 253 | * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
|
| 254 | *
|
| 255 | * TBD: should special case ZONE_DMA32 machines here - in those we normally
|
| 256 | * don't need any ZONE_NORMAL reservation
|
| 257 |
|
| 258 | static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
|
| 259 | #ifdef CONFIG_ZONE_DMA
|
| 260 | [ZONE_DMA] = 256,
|
| 261 | #endif
|
| 262 | #ifdef CONFIG_ZONE_DMA32
|
| 263 | [ZONE_DMA32] = 256,
|
| 264 | #endif
|
| 265 | [ZONE_NORMAL] = 32,
|
| 266 | #ifdef CONFIG_HIGHMEM
|
| 267 | [ZONE_HIGHMEM] = 0,
|
| 268 | #endif
|
| 269 | [ZONE_MOVABLE] = 0,
|
| 270 | };
|
| 271 |
|
| 272 | char * const zone_names[MAX_NR_ZONES] = {
|
| 273 | #ifdef CONFIG_ZONE_DMA
|
| 274 | "DMA",
|
| 275 | #endif
|
| 276 | #ifdef CONFIG_ZONE_DMA32
|
| 277 | "DMA32",
|
| 278 | #endif
|
| 279 | "Normal",
|
| 280 | #ifdef CONFIG_HIGHMEM
|
| 281 | "HighMem",
|
| 282 | #endif
|
| 283 | "Movable",
|
| 284 | #ifdef CONFIG_ZONE_DEVICE
|
| 285 | "Device",
|
| 286 | #endif
|
| 287 | };
|
| 288 |
|
| 289 | const char * const migratetype_names[MIGRATE_TYPES] = {
|
| 290 | "Unmovable",
|
| 291 | "Movable",
|
| 292 | "Reclaimable",
|
| 293 | "HighAtomic",
|
| 294 | #ifdef CONFIG_CMA
|
| 295 | "CMA",
|
| 296 | #endif
|
| 297 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 298 | "Isolate",
|
| 299 | #endif
|
| 300 | };
|
| 301 |
|
| 302 | int min_free_kbytes = 1024;
|
| 303 | int user_min_free_kbytes = -1;
|
| 304 | static int watermark_boost_factor __read_mostly = 15000;
|
| 305 | static int watermark_scale_factor = 10;
|
| 306 | int defrag_mode;
|
| 307 |
|
| 308 |
|
| 309 | int movable_zone;
|
| 310 | EXPORT_SYMBOL(movable_zone);
|
| 311 |
|
| 312 | #if MAX_NUMNODES > 1
|
| 313 | unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
|
| 314 | unsigned int nr_online_nodes __read_mostly = 1;
|
| 315 | EXPORT_SYMBOL(nr_node_ids);
|
| 316 | EXPORT_SYMBOL(nr_online_nodes);
|
| 317 | #endif
|
| 318 |
|
| 319 | static bool page_contains_unaccepted(struct page *page, unsigned int order);
|
| 320 | static bool cond_accept_memory(struct zone *zone, unsigned int order,
|
| 321 | int alloc_flags);
|
| 322 | static bool __free_unaccepted(struct page *page);
|
| 323 |
|
| 324 | int page_group_by_mobility_disabled __read_mostly;
|
| 325 |
|
| 326 | #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
|
| 327 | |
| 328 | * During boot we initialize deferred pages on-demand, as needed, but once
|
| 329 | * page_alloc_init_late() has finished, the deferred pages are all initialized,
|
| 330 | * and we can permanently disable that path.
|
| 331 |
|
| 332 | DEFINE_STATIC_KEY_TRUE(deferred_pages);
|
| 333 |
|
| 334 | static inline bool deferred_pages_enabled(void)
|
| 335 | {
|
| 336 | return static_branch_unlikely(&deferred_pages);
|
| 337 | }
|
| 338 |
|
| 339 | |
| 340 | * deferred_grow_zone() is __init, but it is called from
|
| 341 | * get_page_from_freelist() during early boot until deferred_pages permanently
|
| 342 | * disables this call. This is why we have refdata wrapper to avoid warning,
|
| 343 | * and to ensure that the function body gets unloaded.
|
| 344 |
|
| 345 | static bool __ref
|
| 346 | _deferred_grow_zone(struct zone *zone, unsigned int order)
|
| 347 | {
|
| 348 | return deferred_grow_zone(zone, order);
|
| 349 | }
|
| 350 | #else
|
| 351 | static inline bool deferred_pages_enabled(void)
|
| 352 | {
|
| 353 | return false;
|
| 354 | }
|
| 355 |
|
| 356 | static inline bool _deferred_grow_zone(struct zone *zone, unsigned int order)
|
| 357 | {
|
| 358 | return false;
|
| 359 | }
|
| 360 | #endif
|
| 361 |
|
| 362 |
|
| 363 | static inline unsigned long *get_pageblock_bitmap(const struct page *page,
|
| 364 | unsigned long pfn)
|
| 365 | {
|
| 366 | #ifdef CONFIG_SPARSEMEM
|
| 367 | return section_to_usemap(__pfn_to_section(pfn));
|
| 368 | #else
|
| 369 | return page_zone(page)->pageblock_flags;
|
| 370 | #endif
|
| 371 | }
|
| 372 |
|
| 373 | static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
|
| 374 | {
|
| 375 | #ifdef CONFIG_SPARSEMEM
|
| 376 | pfn &= (PAGES_PER_SECTION-1);
|
| 377 | #else
|
| 378 | pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
|
| 379 | #endif
|
| 380 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
|
| 381 | }
|
| 382 |
|
| 383 | static __always_inline bool is_standalone_pb_bit(enum pageblock_bits pb_bit)
|
| 384 | {
|
| 385 | return pb_bit >= PB_compact_skip && pb_bit < __NR_PAGEBLOCK_BITS;
|
| 386 | }
|
| 387 |
|
| 388 | static __always_inline void
|
| 389 | get_pfnblock_bitmap_bitidx(const struct page *page, unsigned long pfn,
|
| 390 | unsigned long **bitmap_word, unsigned long *bitidx)
|
| 391 | {
|
| 392 | unsigned long *bitmap;
|
| 393 | unsigned long word_bitidx;
|
| 394 |
|
| 395 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 396 | BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 8);
|
| 397 | #else
|
| 398 | BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
|
| 399 | #endif
|
| 400 | BUILD_BUG_ON(__MIGRATE_TYPE_END > MIGRATETYPE_MASK);
|
| 401 | VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
|
| 402 |
|
| 403 | bitmap = get_pageblock_bitmap(page, pfn);
|
| 404 | *bitidx = pfn_to_bitidx(page, pfn);
|
| 405 | word_bitidx = *bitidx / BITS_PER_LONG;
|
| 406 | *bitidx &= (BITS_PER_LONG - 1);
|
| 407 | *bitmap_word = &bitmap[word_bitidx];
|
| 408 | }
|
| 409 |
|
| 410 |
|
| 411 | |
| 412 | * __get_pfnblock_flags_mask - Return the requested group of flags for
|
| 413 | * a pageblock_nr_pages block of pages
|
| 414 | * @page: The page within the block of interest
|
| 415 | * @pfn: The target page frame number
|
| 416 | * @mask: mask of bits that the caller is interested in
|
| 417 | *
|
| 418 | * Return: pageblock_bits flags
|
| 419 |
|
| 420 | static unsigned long __get_pfnblock_flags_mask(const struct page *page,
|
| 421 | unsigned long pfn,
|
| 422 | unsigned long mask)
|
| 423 | {
|
| 424 | unsigned long *bitmap_word;
|
| 425 | unsigned long bitidx;
|
| 426 | unsigned long word;
|
| 427 |
|
| 428 | get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
|
| 429 | |
| 430 | * This races, without locks, with set_pfnblock_migratetype(). Ensure
|
| 431 | * a consistent read of the memory array, so that results, even though
|
| 432 | * racy, are not corrupted.
|
| 433 |
|
| 434 | word = READ_ONCE(*bitmap_word);
|
| 435 | return (word >> bitidx) & mask;
|
| 436 | }
|
| 437 |
|
| 438 | |
| 439 | * get_pfnblock_bit - Check if a standalone bit of a pageblock is set
|
| 440 | * @page: The page within the block of interest
|
| 441 | * @pfn: The target page frame number
|
| 442 | * @pb_bit: pageblock bit to check
|
| 443 | *
|
| 444 | * Return: true if the bit is set, otherwise false
|
| 445 |
|
| 446 | bool get_pfnblock_bit(const struct page *page, unsigned long pfn,
|
| 447 | enum pageblock_bits pb_bit)
|
| 448 | {
|
| 449 | unsigned long *bitmap_word;
|
| 450 | unsigned long bitidx;
|
| 451 |
|
| 452 | if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
|
| 453 | return false;
|
| 454 |
|
| 455 | get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
|
| 456 |
|
| 457 | return test_bit(bitidx + pb_bit, bitmap_word);
|
| 458 | }
|
| 459 |
|
| 460 | |
| 461 | * get_pfnblock_migratetype - Return the migratetype of a pageblock
|
| 462 | * @page: The page within the block of interest
|
| 463 | * @pfn: The target page frame number
|
| 464 | *
|
| 465 | * Return: The migratetype of the pageblock
|
| 466 | *
|
| 467 | * Use get_pfnblock_migratetype() if caller already has both @page and @pfn
|
| 468 | * to save a call to page_to_pfn().
|
| 469 |
|
| 470 | __always_inline enum migratetype
|
| 471 | get_pfnblock_migratetype(const struct page *page, unsigned long pfn)
|
| 472 | {
|
| 473 | unsigned long mask = MIGRATETYPE_AND_ISO_MASK;
|
| 474 | unsigned long flags;
|
| 475 |
|
| 476 | flags = __get_pfnblock_flags_mask(page, pfn, mask);
|
| 477 |
|
| 478 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 479 | if (flags & BIT(PB_migrate_isolate))
|
| 480 | return MIGRATE_ISOLATE;
|
| 481 | #endif
|
| 482 | return flags & MIGRATETYPE_MASK;
|
| 483 | }
|
| 484 |
|
| 485 | |
| 486 | * __set_pfnblock_flags_mask - Set the requested group of flags for
|
| 487 | * a pageblock_nr_pages block of pages
|
| 488 | * @page: The page within the block of interest
|
| 489 | * @pfn: The target page frame number
|
| 490 | * @flags: The flags to set
|
| 491 | * @mask: mask of bits that the caller is interested in
|
| 492 |
|
| 493 | static void __set_pfnblock_flags_mask(struct page *page, unsigned long pfn,
|
| 494 | unsigned long flags, unsigned long mask)
|
| 495 | {
|
| 496 | unsigned long *bitmap_word;
|
| 497 | unsigned long bitidx;
|
| 498 | unsigned long word;
|
| 499 |
|
| 500 | get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
|
| 501 |
|
| 502 | mask <<= bitidx;
|
| 503 | flags <<= bitidx;
|
| 504 |
|
| 505 | word = READ_ONCE(*bitmap_word);
|
| 506 | do {
|
| 507 | } while (!try_cmpxchg(bitmap_word, &word, (word & ~mask) | flags));
|
| 508 | }
|
| 509 |
|
| 510 | |
| 511 | * set_pfnblock_bit - Set a standalone bit of a pageblock
|
| 512 | * @page: The page within the block of interest
|
| 513 | * @pfn: The target page frame number
|
| 514 | * @pb_bit: pageblock bit to set
|
| 515 |
|
| 516 | void set_pfnblock_bit(const struct page *page, unsigned long pfn,
|
| 517 | enum pageblock_bits pb_bit)
|
| 518 | {
|
| 519 | unsigned long *bitmap_word;
|
| 520 | unsigned long bitidx;
|
| 521 |
|
| 522 | if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
|
| 523 | return;
|
| 524 |
|
| 525 | get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
|
| 526 |
|
| 527 | set_bit(bitidx + pb_bit, bitmap_word);
|
| 528 | }
|
| 529 |
|
| 530 | |
| 531 | * clear_pfnblock_bit - Clear a standalone bit of a pageblock
|
| 532 | * @page: The page within the block of interest
|
| 533 | * @pfn: The target page frame number
|
| 534 | * @pb_bit: pageblock bit to clear
|
| 535 |
|
| 536 | void clear_pfnblock_bit(const struct page *page, unsigned long pfn,
|
| 537 | enum pageblock_bits pb_bit)
|
| 538 | {
|
| 539 | unsigned long *bitmap_word;
|
| 540 | unsigned long bitidx;
|
| 541 |
|
| 542 | if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
|
| 543 | return;
|
| 544 |
|
| 545 | get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
|
| 546 |
|
| 547 | clear_bit(bitidx + pb_bit, bitmap_word);
|
| 548 | }
|
| 549 |
|
| 550 | |
| 551 | * set_pageblock_migratetype - Set the migratetype of a pageblock
|
| 552 | * @page: The page within the block of interest
|
| 553 | * @migratetype: migratetype to set
|
| 554 |
|
| 555 | static void set_pageblock_migratetype(struct page *page,
|
| 556 | enum migratetype migratetype)
|
| 557 | {
|
| 558 | if (unlikely(page_group_by_mobility_disabled &&
|
| 559 | migratetype < MIGRATE_PCPTYPES))
|
| 560 | migratetype = MIGRATE_UNMOVABLE;
|
| 561 |
|
| 562 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 563 | if (migratetype == MIGRATE_ISOLATE) {
|
| 564 | VM_WARN_ONCE(1,
|
| 565 | "Use set_pageblock_isolate() for pageblock isolation");
|
| 566 | return;
|
| 567 | }
|
| 568 | VM_WARN_ONCE(get_pageblock_isolate(page),
|
| 569 | "Use clear_pageblock_isolate() to unisolate pageblock");
|
| 570 |
|
| 571 | #endif
|
| 572 | __set_pfnblock_flags_mask(page, page_to_pfn(page),
|
| 573 | (unsigned long)migratetype,
|
| 574 | MIGRATETYPE_AND_ISO_MASK);
|
| 575 | }
|
| 576 |
|
| 577 | void __meminit init_pageblock_migratetype(struct page *page,
|
| 578 | enum migratetype migratetype,
|
| 579 | bool isolate)
|
| 580 | {
|
| 581 | unsigned long flags;
|
| 582 |
|
| 583 | if (unlikely(page_group_by_mobility_disabled &&
|
| 584 | migratetype < MIGRATE_PCPTYPES))
|
| 585 | migratetype = MIGRATE_UNMOVABLE;
|
| 586 |
|
| 587 | flags = migratetype;
|
| 588 |
|
| 589 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 590 | if (migratetype == MIGRATE_ISOLATE) {
|
| 591 | VM_WARN_ONCE(
|
| 592 | 1,
|
| 593 | "Set isolate=true to isolate pageblock with a migratetype");
|
| 594 | return;
|
| 595 | }
|
| 596 | if (isolate)
|
| 597 | flags |= BIT(PB_migrate_isolate);
|
| 598 | #endif
|
| 599 | __set_pfnblock_flags_mask(page, page_to_pfn(page), flags,
|
| 600 | MIGRATETYPE_AND_ISO_MASK);
|
| 601 | }
|
| 602 |
|
| 603 | #ifdef CONFIG_DEBUG_VM
|
| 604 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
|
| 605 | {
|
| 606 | int ret;
|
| 607 | unsigned seq;
|
| 608 | unsigned long pfn = page_to_pfn(page);
|
| 609 | unsigned long sp, start_pfn;
|
| 610 |
|
| 611 | do {
|
| 612 | seq = zone_span_seqbegin(zone);
|
| 613 | start_pfn = zone->zone_start_pfn;
|
| 614 | sp = zone->spanned_pages;
|
| 615 | ret = !zone_spans_pfn(zone, pfn);
|
| 616 | } while (zone_span_seqretry(zone, seq));
|
| 617 |
|
| 618 | if (ret)
|
| 619 | pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
|
| 620 | pfn, zone_to_nid(zone), zone->name,
|
| 621 | start_pfn, start_pfn + sp);
|
| 622 |
|
| 623 | return ret;
|
| 624 | }
|
| 625 |
|
| 626 | |
| 627 | * Temporary debugging check for pages not lying within a given zone.
|
| 628 |
|
| 629 | static bool __maybe_unused bad_range(struct zone *zone, struct page *page)
|
| 630 | {
|
| 631 | if (page_outside_zone_boundaries(zone, page))
|
| 632 | return true;
|
| 633 | if (zone != page_zone(page))
|
| 634 | return true;
|
| 635 |
|
| 636 | return false;
|
| 637 | }
|
| 638 | #else
|
| 639 | static inline bool __maybe_unused bad_range(struct zone *zone, struct page *page)
|
| 640 | {
|
| 641 | return false;
|
| 642 | }
|
| 643 | #endif
|
| 644 |
|
| 645 | static void bad_page(struct page *page, const char *reason)
|
| 646 | {
|
| 647 | static unsigned long resume;
|
| 648 | static unsigned long nr_shown;
|
| 649 | static unsigned long nr_unshown;
|
| 650 |
|
| 651 | |
| 652 | * Allow a burst of 60 reports, then keep quiet for that minute;
|
| 653 | * or allow a steady drip of one report per second.
|
| 654 |
|
| 655 | if (nr_shown == 60) {
|
| 656 | if (time_before(jiffies, resume)) {
|
| 657 | nr_unshown++;
|
| 658 | goto out;
|
| 659 | }
|
| 660 | if (nr_unshown) {
|
| 661 | pr_alert(
|
| 662 | "BUG: Bad page state: %lu messages suppressed\n",
|
| 663 | nr_unshown);
|
| 664 | nr_unshown = 0;
|
| 665 | }
|
| 666 | nr_shown = 0;
|
| 667 | }
|
| 668 | if (nr_shown++ == 0)
|
| 669 | resume = jiffies + 60 * HZ;
|
| 670 |
|
| 671 | pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
|
| 672 | current->comm, page_to_pfn(page));
|
| 673 | dump_page(page, reason);
|
| 674 |
|
| 675 | print_modules();
|
| 676 | dump_stack();
|
| 677 | out:
|
| 678 |
|
| 679 | if (PageBuddy(page))
|
| 680 | __ClearPageBuddy(page);
|
| 681 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
|
| 682 | }
|
| 683 |
|
| 684 | static inline unsigned int order_to_pindex(int migratetype, int order)
|
| 685 | {
|
| 686 |
|
| 687 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE
|
| 688 | bool movable;
|
| 689 | if (order > PAGE_ALLOC_COSTLY_ORDER) {
|
| 690 | VM_BUG_ON(order != HPAGE_PMD_ORDER);
|
| 691 |
|
| 692 | movable = migratetype == MIGRATE_MOVABLE;
|
| 693 |
|
| 694 | return NR_LOWORDER_PCP_LISTS + movable;
|
| 695 | }
|
| 696 | #else
|
| 697 | VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
|
| 698 | #endif
|
| 699 |
|
| 700 | return (MIGRATE_PCPTYPES * order) + migratetype;
|
| 701 | }
|
| 702 |
|
| 703 | static inline int pindex_to_order(unsigned int pindex)
|
| 704 | {
|
| 705 | int order = pindex / MIGRATE_PCPTYPES;
|
| 706 |
|
| 707 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE
|
| 708 | if (pindex >= NR_LOWORDER_PCP_LISTS)
|
| 709 | order = HPAGE_PMD_ORDER;
|
| 710 | #else
|
| 711 | VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
|
| 712 | #endif
|
| 713 |
|
| 714 | return order;
|
| 715 | }
|
| 716 |
|
| 717 | static inline bool pcp_allowed_order(unsigned int order)
|
| 718 | {
|
| 719 | if (order <= PAGE_ALLOC_COSTLY_ORDER)
|
| 720 | return true;
|
| 721 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE
|
| 722 | if (order == HPAGE_PMD_ORDER)
|
| 723 | return true;
|
| 724 | #endif
|
| 725 | return false;
|
| 726 | }
|
| 727 |
|
| 728 | |
| 729 | * Higher-order pages are called "compound pages". They are structured thusly:
|
| 730 | *
|
| 731 | * The first PAGE_SIZE page is called the "head page" and have PG_head set.
|
| 732 | *
|
| 733 | * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
|
| 734 | * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
|
| 735 | *
|
| 736 | * The first tail page's ->compound_order holds the order of allocation.
|
| 737 | * This usage means that zero-order pages may not be compound.
|
| 738 |
|
| 739 |
|
| 740 | void prep_compound_page(struct page *page, unsigned int order)
|
| 741 | {
|
| 742 | int i;
|
| 743 | int nr_pages = 1 << order;
|
| 744 |
|
| 745 | __SetPageHead(page);
|
| 746 | for (i = 1; i < nr_pages; i++)
|
| 747 | prep_compound_tail(page, i);
|
| 748 |
|
| 749 | prep_compound_head(page, order);
|
| 750 | }
|
| 751 |
|
| 752 | static inline void set_buddy_order(struct page *page, unsigned int order)
|
| 753 | {
|
| 754 | set_page_private(page, order);
|
| 755 | __SetPageBuddy(page);
|
| 756 | }
|
| 757 |
|
| 758 | #ifdef CONFIG_COMPACTION
|
| 759 | static inline struct capture_control *task_capc(struct zone *zone)
|
| 760 | {
|
| 761 | struct capture_control *capc = current->capture_control;
|
| 762 |
|
| 763 | return unlikely(capc) &&
|
| 764 | !(current->flags & PF_KTHREAD) &&
|
| 765 | !capc->page &&
|
| 766 | capc->cc->zone == zone ? capc : NULL;
|
| 767 | }
|
| 768 |
|
| 769 | static inline bool
|
| 770 | compaction_capture(struct capture_control *capc, struct page *page,
|
| 771 | int order, int migratetype)
|
| 772 | {
|
| 773 | if (!capc || order != capc->cc->order)
|
| 774 | return false;
|
| 775 |
|
| 776 |
|
| 777 | if (is_migrate_cma(migratetype) ||
|
| 778 | is_migrate_isolate(migratetype))
|
| 779 | return false;
|
| 780 |
|
| 781 | |
| 782 | * Do not let lower order allocations pollute a movable pageblock
|
| 783 | * unless compaction is also requesting movable pages.
|
| 784 | * This might let an unmovable request use a reclaimable pageblock
|
| 785 | * and vice-versa but no more than normal fallback logic which can
|
| 786 | * have trouble finding a high-order free page.
|
| 787 |
|
| 788 | if (order < pageblock_order && migratetype == MIGRATE_MOVABLE &&
|
| 789 | capc->cc->migratetype != MIGRATE_MOVABLE)
|
| 790 | return false;
|
| 791 |
|
| 792 | if (migratetype != capc->cc->migratetype)
|
| 793 | trace_mm_page_alloc_extfrag(page, capc->cc->order, order,
|
| 794 | capc->cc->migratetype, migratetype);
|
| 795 |
|
| 796 | capc->page = page;
|
| 797 | return true;
|
| 798 | }
|
| 799 |
|
| 800 | #else
|
| 801 | static inline struct capture_control *task_capc(struct zone *zone)
|
| 802 | {
|
| 803 | return NULL;
|
| 804 | }
|
| 805 |
|
| 806 | static inline bool
|
| 807 | compaction_capture(struct capture_control *capc, struct page *page,
|
| 808 | int order, int migratetype)
|
| 809 | {
|
| 810 | return false;
|
| 811 | }
|
| 812 | #endif
|
| 813 |
|
| 814 | static inline void account_freepages(struct zone *zone, int nr_pages,
|
| 815 | int migratetype)
|
| 816 | {
|
| 817 | lockdep_assert_held(&zone->lock);
|
| 818 |
|
| 819 | if (is_migrate_isolate(migratetype))
|
| 820 | return;
|
| 821 |
|
| 822 | __mod_zone_page_state(zone, NR_FREE_PAGES, nr_pages);
|
| 823 |
|
| 824 | if (is_migrate_cma(migratetype))
|
| 825 | __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, nr_pages);
|
| 826 | else if (migratetype == MIGRATE_HIGHATOMIC)
|
| 827 | WRITE_ONCE(zone->nr_free_highatomic,
|
| 828 | zone->nr_free_highatomic + nr_pages);
|
| 829 | }
|
| 830 |
|
| 831 |
|
| 832 | static inline void __add_to_free_list(struct page *page, struct zone *zone,
|
| 833 | unsigned int order, int migratetype,
|
| 834 | bool tail)
|
| 835 | {
|
| 836 | struct free_area *area = &zone->free_area[order];
|
| 837 | int nr_pages = 1 << order;
|
| 838 |
|
| 839 | VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
|
| 840 | "page type is %d, passed migratetype is %d (nr=%d)\n",
|
| 841 | get_pageblock_migratetype(page), migratetype, nr_pages);
|
| 842 |
|
| 843 | if (tail)
|
| 844 | list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
|
| 845 | else
|
| 846 | list_add(&page->buddy_list, &area->free_list[migratetype]);
|
| 847 | area->nr_free++;
|
| 848 |
|
| 849 | if (order >= pageblock_order && !is_migrate_isolate(migratetype))
|
| 850 | __mod_zone_page_state(zone, NR_FREE_PAGES_BLOCKS, nr_pages);
|
| 851 | }
|
| 852 |
|
| 853 | |
| 854 | * Used for pages which are on another list. Move the pages to the tail
|
| 855 | * of the list - so the moved pages won't immediately be considered for
|
| 856 | * allocation again (e.g., optimization for memory onlining).
|
| 857 |
|
| 858 | static inline void move_to_free_list(struct page *page, struct zone *zone,
|
| 859 | unsigned int order, int old_mt, int new_mt)
|
| 860 | {
|
| 861 | struct free_area *area = &zone->free_area[order];
|
| 862 | int nr_pages = 1 << order;
|
| 863 |
|
| 864 |
|
| 865 | VM_WARN_ONCE(get_pageblock_migratetype(page) != old_mt,
|
| 866 | "page type is %d, passed migratetype is %d (nr=%d)\n",
|
| 867 | get_pageblock_migratetype(page), old_mt, nr_pages);
|
| 868 |
|
| 869 | list_move_tail(&page->buddy_list, &area->free_list[new_mt]);
|
| 870 |
|
| 871 | account_freepages(zone, -nr_pages, old_mt);
|
| 872 | account_freepages(zone, nr_pages, new_mt);
|
| 873 |
|
| 874 | if (order >= pageblock_order &&
|
| 875 | is_migrate_isolate(old_mt) != is_migrate_isolate(new_mt)) {
|
| 876 | if (!is_migrate_isolate(old_mt))
|
| 877 | nr_pages = -nr_pages;
|
| 878 | __mod_zone_page_state(zone, NR_FREE_PAGES_BLOCKS, nr_pages);
|
| 879 | }
|
| 880 | }
|
| 881 |
|
| 882 | static inline void __del_page_from_free_list(struct page *page, struct zone *zone,
|
| 883 | unsigned int order, int migratetype)
|
| 884 | {
|
| 885 | int nr_pages = 1 << order;
|
| 886 |
|
| 887 | VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
|
| 888 | "page type is %d, passed migratetype is %d (nr=%d)\n",
|
| 889 | get_pageblock_migratetype(page), migratetype, nr_pages);
|
| 890 |
|
| 891 |
|
| 892 | if (page_reported(page))
|
| 893 | __ClearPageReported(page);
|
| 894 |
|
| 895 | list_del(&page->buddy_list);
|
| 896 | __ClearPageBuddy(page);
|
| 897 | set_page_private(page, 0);
|
| 898 | zone->free_area[order].nr_free--;
|
| 899 |
|
| 900 | if (order >= pageblock_order && !is_migrate_isolate(migratetype))
|
| 901 | __mod_zone_page_state(zone, NR_FREE_PAGES_BLOCKS, -nr_pages);
|
| 902 | }
|
| 903 |
|
| 904 | static inline void del_page_from_free_list(struct page *page, struct zone *zone,
|
| 905 | unsigned int order, int migratetype)
|
| 906 | {
|
| 907 | __del_page_from_free_list(page, zone, order, migratetype);
|
| 908 | account_freepages(zone, -(1 << order), migratetype);
|
| 909 | }
|
| 910 |
|
| 911 | static inline struct page *get_page_from_free_area(struct free_area *area,
|
| 912 | int migratetype)
|
| 913 | {
|
| 914 | return list_first_entry_or_null(&area->free_list[migratetype],
|
| 915 | struct page, buddy_list);
|
| 916 | }
|
| 917 |
|
| 918 | |
| 919 | * If this is less than the 2nd largest possible page, check if the buddy
|
| 920 | * of the next-higher order is free. If it is, it's possible
|
| 921 | * that pages are being freed that will coalesce soon. In case,
|
| 922 | * that is happening, add the free page to the tail of the list
|
| 923 | * so it's less likely to be used soon and more likely to be merged
|
| 924 | * as a 2-level higher order page
|
| 925 |
|
| 926 | static inline bool
|
| 927 | buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
|
| 928 | struct page *page, unsigned int order)
|
| 929 | {
|
| 930 | unsigned long higher_page_pfn;
|
| 931 | struct page *higher_page;
|
| 932 |
|
| 933 | if (order >= MAX_PAGE_ORDER - 1)
|
| 934 | return false;
|
| 935 |
|
| 936 | higher_page_pfn = buddy_pfn & pfn;
|
| 937 | higher_page = page + (higher_page_pfn - pfn);
|
| 938 |
|
| 939 | return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
|
| 940 | NULL) != NULL;
|
| 941 | }
|
| 942 |
|
| 943 | static void change_pageblock_range(struct page *pageblock_page,
|
| 944 | int start_order, int migratetype)
|
| 945 | {
|
| 946 | int nr_pageblocks = 1 << (start_order - pageblock_order);
|
| 947 |
|
| 948 | while (nr_pageblocks--) {
|
| 949 | set_pageblock_migratetype(pageblock_page, migratetype);
|
| 950 | pageblock_page += pageblock_nr_pages;
|
| 951 | }
|
| 952 | }
|
| 953 |
|
| 954 | |
| 955 | * Freeing function for a buddy system allocator.
|
| 956 | *
|
| 957 | * The concept of a buddy system is to maintain direct-mapped table
|
| 958 | * (containing bit values) for memory blocks of various "orders".
|
| 959 | * The bottom level table contains the map for the smallest allocatable
|
| 960 | * units of memory (here, pages), and each level above it describes
|
| 961 | * pairs of units from the levels below, hence, "buddies".
|
| 962 | * At a high level, all that happens here is marking the table entry
|
| 963 | * at the bottom level available, and propagating the changes upward
|
| 964 | * as necessary, plus some accounting needed to play nicely with other
|
| 965 | * parts of the VM system.
|
| 966 | * At each level, we keep a list of pages, which are heads of continuous
|
| 967 | * free pages of length of (1 << order) and marked with PageBuddy.
|
| 968 | * Page's order is recorded in page_private(page) field.
|
| 969 | * So when we are allocating or freeing one, we can derive the state of the
|
| 970 | * other. That is, if we allocate a small block, and both were
|
| 971 | * free, the remainder of the region must be split into blocks.
|
| 972 | * If a block is freed, and its buddy is also free, then this
|
| 973 | * triggers coalescing into a block of larger size.
|
| 974 | *
|
| 975 | * -- nyc
|
| 976 |
|
| 977 |
|
| 978 | static inline void __free_one_page(struct page *page,
|
| 979 | unsigned long pfn,
|
| 980 | struct zone *zone, unsigned int order,
|
| 981 | int migratetype, fpi_t fpi_flags)
|
| 982 | {
|
| 983 | struct capture_control *capc = task_capc(zone);
|
| 984 | unsigned long buddy_pfn = 0;
|
| 985 | unsigned long combined_pfn;
|
| 986 | struct page *buddy;
|
| 987 | bool to_tail;
|
| 988 |
|
| 989 | VM_BUG_ON(!zone_is_initialized(zone));
|
| 990 | VM_BUG_ON_PAGE(page->flags.f & PAGE_FLAGS_CHECK_AT_PREP, page);
|
| 991 |
|
| 992 | VM_BUG_ON(migratetype == -1);
|
| 993 | VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
|
| 994 | VM_BUG_ON_PAGE(bad_range(zone, page), page);
|
| 995 |
|
| 996 | account_freepages(zone, 1 << order, migratetype);
|
| 997 |
|
| 998 | while (order < MAX_PAGE_ORDER) {
|
| 999 | int buddy_mt = migratetype;
|
| 1000 |
|
| 1001 | if (compaction_capture(capc, page, order, migratetype)) {
|
| 1002 | account_freepages(zone, -(1 << order), migratetype);
|
| 1003 | return;
|
| 1004 | }
|
| 1005 |
|
| 1006 | buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
|
| 1007 | if (!buddy)
|
| 1008 | goto done_merging;
|
| 1009 |
|
| 1010 | if (unlikely(order >= pageblock_order)) {
|
| 1011 | |
| 1012 | * We want to prevent merge between freepages on pageblock
|
| 1013 | * without fallbacks and normal pageblock. Without this,
|
| 1014 | * pageblock isolation could cause incorrect freepage or CMA
|
| 1015 | * accounting or HIGHATOMIC accounting.
|
| 1016 |
|
| 1017 | buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
|
| 1018 |
|
| 1019 | if (migratetype != buddy_mt &&
|
| 1020 | (!migratetype_is_mergeable(migratetype) ||
|
| 1021 | !migratetype_is_mergeable(buddy_mt)))
|
| 1022 | goto done_merging;
|
| 1023 | }
|
| 1024 |
|
| 1025 | |
| 1026 | * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
|
| 1027 | * merge with it and move up one order.
|
| 1028 |
|
| 1029 | if (page_is_guard(buddy))
|
| 1030 | clear_page_guard(zone, buddy, order);
|
| 1031 | else
|
| 1032 | __del_page_from_free_list(buddy, zone, order, buddy_mt);
|
| 1033 |
|
| 1034 | if (unlikely(buddy_mt != migratetype)) {
|
| 1035 | |
| 1036 | * Match buddy type. This ensures that an
|
| 1037 | * expand() down the line puts the sub-blocks
|
| 1038 | * on the right freelists.
|
| 1039 |
|
| 1040 | change_pageblock_range(buddy, order, migratetype);
|
| 1041 | }
|
| 1042 |
|
| 1043 | combined_pfn = buddy_pfn & pfn;
|
| 1044 | page = page + (combined_pfn - pfn);
|
| 1045 | pfn = combined_pfn;
|
| 1046 | order++;
|
| 1047 | }
|
| 1048 |
|
| 1049 | done_merging:
|
| 1050 | set_buddy_order(page, order);
|
| 1051 |
|
| 1052 | if (fpi_flags & FPI_TO_TAIL)
|
| 1053 | to_tail = true;
|
| 1054 | else if (is_shuffle_order(order))
|
| 1055 | to_tail = shuffle_pick_tail();
|
| 1056 | else
|
| 1057 | to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
|
| 1058 |
|
| 1059 | __add_to_free_list(page, zone, order, migratetype, to_tail);
|
| 1060 |
|
| 1061 |
|
| 1062 | if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
|
| 1063 | page_reporting_notify_free(order);
|
| 1064 | }
|
| 1065 |
|
| 1066 | |
| 1067 | * A bad page could be due to a number of fields. Instead of multiple branches,
|
| 1068 | * try and check multiple fields with one check. The caller must do a detailed
|
| 1069 | * check if necessary.
|
| 1070 |
|
| 1071 | static inline bool page_expected_state(struct page *page,
|
| 1072 | unsigned long check_flags)
|
| 1073 | {
|
| 1074 | if (unlikely(atomic_read(&page->_mapcount) != -1))
|
| 1075 | return false;
|
| 1076 |
|
| 1077 | if (unlikely((unsigned long)page->mapping |
|
| 1078 | page_ref_count(page) |
|
| 1079 | #ifdef CONFIG_MEMCG
|
| 1080 | page->memcg_data |
|
| 1081 | #endif
|
| 1082 | page_pool_page_is_pp(page) |
|
| 1083 | (page->flags.f & check_flags)))
|
| 1084 | return false;
|
| 1085 |
|
| 1086 | return true;
|
| 1087 | }
|
| 1088 |
|
| 1089 | static const char *page_bad_reason(struct page *page, unsigned long flags)
|
| 1090 | {
|
| 1091 | const char *bad_reason = NULL;
|
| 1092 |
|
| 1093 | if (unlikely(atomic_read(&page->_mapcount) != -1))
|
| 1094 | bad_reason = "nonzero mapcount";
|
| 1095 | if (unlikely(page->mapping != NULL))
|
| 1096 | bad_reason = "non-NULL mapping";
|
| 1097 | if (unlikely(page_ref_count(page) != 0))
|
| 1098 | bad_reason = "nonzero _refcount";
|
| 1099 | if (unlikely(page->flags.f & flags)) {
|
| 1100 | if (flags == PAGE_FLAGS_CHECK_AT_PREP)
|
| 1101 | bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
|
| 1102 | else
|
| 1103 | bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
|
| 1104 | }
|
| 1105 | #ifdef CONFIG_MEMCG
|
| 1106 | if (unlikely(page->memcg_data))
|
| 1107 | bad_reason = "page still charged to cgroup";
|
| 1108 | #endif
|
| 1109 | if (unlikely(page_pool_page_is_pp(page)))
|
| 1110 | bad_reason = "page_pool leak";
|
| 1111 | return bad_reason;
|
| 1112 | }
|
| 1113 |
|
| 1114 | static inline bool free_page_is_bad(struct page *page)
|
| 1115 | {
|
| 1116 | if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
|
| 1117 | return false;
|
| 1118 |
|
| 1119 |
|
| 1120 | bad_page(page, page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
|
| 1121 | return true;
|
| 1122 | }
|
| 1123 |
|
| 1124 | static inline bool is_check_pages_enabled(void)
|
| 1125 | {
|
| 1126 | return static_branch_unlikely(&check_pages_enabled);
|
| 1127 | }
|
| 1128 |
|
| 1129 | static int free_tail_page_prepare(struct page *head_page, struct page *page)
|
| 1130 | {
|
| 1131 | struct folio *folio = (struct folio *)head_page;
|
| 1132 | int ret = 1;
|
| 1133 |
|
| 1134 | |
| 1135 | * We rely page->lru.next never has bit 0 set, unless the page
|
| 1136 | * is PageTail(). Let's make sure that's true even for poisoned ->lru.
|
| 1137 |
|
| 1138 | BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
|
| 1139 |
|
| 1140 | if (!is_check_pages_enabled()) {
|
| 1141 | ret = 0;
|
| 1142 | goto out;
|
| 1143 | }
|
| 1144 | switch (page - head_page) {
|
| 1145 | case 1:
|
| 1146 |
|
| 1147 | if (unlikely(folio_large_mapcount(folio))) {
|
| 1148 | bad_page(page, "nonzero large_mapcount");
|
| 1149 | goto out;
|
| 1150 | }
|
| 1151 | if (IS_ENABLED(CONFIG_PAGE_MAPCOUNT) &&
|
| 1152 | unlikely(atomic_read(&folio->_nr_pages_mapped))) {
|
| 1153 | bad_page(page, "nonzero nr_pages_mapped");
|
| 1154 | goto out;
|
| 1155 | }
|
| 1156 | if (IS_ENABLED(CONFIG_MM_ID)) {
|
| 1157 | if (unlikely(folio->_mm_id_mapcount[0] != -1)) {
|
| 1158 | bad_page(page, "nonzero mm mapcount 0");
|
| 1159 | goto out;
|
| 1160 | }
|
| 1161 | if (unlikely(folio->_mm_id_mapcount[1] != -1)) {
|
| 1162 | bad_page(page, "nonzero mm mapcount 1");
|
| 1163 | goto out;
|
| 1164 | }
|
| 1165 | }
|
| 1166 | if (IS_ENABLED(CONFIG_64BIT)) {
|
| 1167 | if (unlikely(atomic_read(&folio->_entire_mapcount) + 1)) {
|
| 1168 | bad_page(page, "nonzero entire_mapcount");
|
| 1169 | goto out;
|
| 1170 | }
|
| 1171 | if (unlikely(atomic_read(&folio->_pincount))) {
|
| 1172 | bad_page(page, "nonzero pincount");
|
| 1173 | goto out;
|
| 1174 | }
|
| 1175 | }
|
| 1176 | break;
|
| 1177 | case 2:
|
| 1178 |
|
| 1179 | if (unlikely(!list_empty(&folio->_deferred_list))) {
|
| 1180 | bad_page(page, "on deferred list");
|
| 1181 | goto out;
|
| 1182 | }
|
| 1183 | if (!IS_ENABLED(CONFIG_64BIT)) {
|
| 1184 | if (unlikely(atomic_read(&folio->_entire_mapcount) + 1)) {
|
| 1185 | bad_page(page, "nonzero entire_mapcount");
|
| 1186 | goto out;
|
| 1187 | }
|
| 1188 | if (unlikely(atomic_read(&folio->_pincount))) {
|
| 1189 | bad_page(page, "nonzero pincount");
|
| 1190 | goto out;
|
| 1191 | }
|
| 1192 | }
|
| 1193 | break;
|
| 1194 | case 3:
|
| 1195 |
|
| 1196 | if (IS_ENABLED(CONFIG_HUGETLB_PAGE))
|
| 1197 | break;
|
| 1198 | fallthrough;
|
| 1199 | default:
|
| 1200 | if (page->mapping != TAIL_MAPPING) {
|
| 1201 | bad_page(page, "corrupted mapping in tail page");
|
| 1202 | goto out;
|
| 1203 | }
|
| 1204 | break;
|
| 1205 | }
|
| 1206 | if (unlikely(!PageTail(page))) {
|
| 1207 | bad_page(page, "PageTail not set");
|
| 1208 | goto out;
|
| 1209 | }
|
| 1210 | if (unlikely(compound_head(page) != head_page)) {
|
| 1211 | bad_page(page, "compound_head not consistent");
|
| 1212 | goto out;
|
| 1213 | }
|
| 1214 | ret = 0;
|
| 1215 | out:
|
| 1216 | page->mapping = NULL;
|
| 1217 | clear_compound_head(page);
|
| 1218 | return ret;
|
| 1219 | }
|
| 1220 |
|
| 1221 | |
| 1222 | * Skip KASAN memory poisoning when either:
|
| 1223 | *
|
| 1224 | * 1. For generic KASAN: deferred memory initialization has not yet completed.
|
| 1225 | * Tag-based KASAN modes skip pages freed via deferred memory initialization
|
| 1226 | * using page tags instead (see below).
|
| 1227 | * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
|
| 1228 | * that error detection is disabled for accesses via the page address.
|
| 1229 | *
|
| 1230 | * Pages will have match-all tags in the following circumstances:
|
| 1231 | *
|
| 1232 | * 1. Pages are being initialized for the first time, including during deferred
|
| 1233 | * memory init; see the call to page_kasan_tag_reset in __init_single_page.
|
| 1234 | * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
|
| 1235 | * exception of pages unpoisoned by kasan_unpoison_vmalloc.
|
| 1236 | * 3. The allocation was excluded from being checked due to sampling,
|
| 1237 | * see the call to kasan_unpoison_pages.
|
| 1238 | *
|
| 1239 | * Poisoning pages during deferred memory init will greatly lengthen the
|
| 1240 | * process and cause problem in large memory systems as the deferred pages
|
| 1241 | * initialization is done with interrupt disabled.
|
| 1242 | *
|
| 1243 | * Assuming that there will be no reference to those newly initialized
|
| 1244 | * pages before they are ever allocated, this should have no effect on
|
| 1245 | * KASAN memory tracking as the poison will be properly inserted at page
|
| 1246 | * allocation time. The only corner case is when pages are allocated by
|
| 1247 | * on-demand allocation and then freed again before the deferred pages
|
| 1248 | * initialization is done, but this is not likely to happen.
|
| 1249 |
|
| 1250 | static inline bool should_skip_kasan_poison(struct page *page)
|
| 1251 | {
|
| 1252 | if (IS_ENABLED(CONFIG_KASAN_GENERIC))
|
| 1253 | return deferred_pages_enabled();
|
| 1254 |
|
| 1255 | return page_kasan_tag(page) == KASAN_TAG_KERNEL;
|
| 1256 | }
|
| 1257 |
|
| 1258 | static void kernel_init_pages(struct page *page, int numpages)
|
| 1259 | {
|
| 1260 | int i;
|
| 1261 |
|
| 1262 |
|
| 1263 | kasan_disable_current();
|
| 1264 | for (i = 0; i < numpages; i++)
|
| 1265 | clear_highpage_kasan_tagged(page + i);
|
| 1266 | kasan_enable_current();
|
| 1267 | }
|
| 1268 |
|
| 1269 | #ifdef CONFIG_MEM_ALLOC_PROFILING
|
| 1270 |
|
| 1271 |
|
| 1272 | void __clear_page_tag_ref(struct page *page)
|
| 1273 | {
|
| 1274 | union pgtag_ref_handle handle;
|
| 1275 | union codetag_ref ref;
|
| 1276 |
|
| 1277 | if (get_page_tag_ref(page, &ref, &handle)) {
|
| 1278 | set_codetag_empty(&ref);
|
| 1279 | update_page_tag_ref(handle, &ref);
|
| 1280 | put_page_tag_ref(handle);
|
| 1281 | }
|
| 1282 | }
|
| 1283 |
|
| 1284 |
|
| 1285 | static noinline
|
| 1286 | void __pgalloc_tag_add(struct page *page, struct task_struct *task,
|
| 1287 | unsigned int nr)
|
| 1288 | {
|
| 1289 | union pgtag_ref_handle handle;
|
| 1290 | union codetag_ref ref;
|
| 1291 |
|
| 1292 | if (get_page_tag_ref(page, &ref, &handle)) {
|
| 1293 | alloc_tag_add(&ref, task->alloc_tag, PAGE_SIZE * nr);
|
| 1294 | update_page_tag_ref(handle, &ref);
|
| 1295 | put_page_tag_ref(handle);
|
| 1296 | }
|
| 1297 | }
|
| 1298 |
|
| 1299 | static inline void pgalloc_tag_add(struct page *page, struct task_struct *task,
|
| 1300 | unsigned int nr)
|
| 1301 | {
|
| 1302 | if (mem_alloc_profiling_enabled())
|
| 1303 | __pgalloc_tag_add(page, task, nr);
|
| 1304 | }
|
| 1305 |
|
| 1306 |
|
| 1307 | static noinline
|
| 1308 | void __pgalloc_tag_sub(struct page *page, unsigned int nr)
|
| 1309 | {
|
| 1310 | union pgtag_ref_handle handle;
|
| 1311 | union codetag_ref ref;
|
| 1312 |
|
| 1313 | if (get_page_tag_ref(page, &ref, &handle)) {
|
| 1314 | alloc_tag_sub(&ref, PAGE_SIZE * nr);
|
| 1315 | update_page_tag_ref(handle, &ref);
|
| 1316 | put_page_tag_ref(handle);
|
| 1317 | }
|
| 1318 | }
|
| 1319 |
|
| 1320 | static inline void pgalloc_tag_sub(struct page *page, unsigned int nr)
|
| 1321 | {
|
| 1322 | if (mem_alloc_profiling_enabled())
|
| 1323 | __pgalloc_tag_sub(page, nr);
|
| 1324 | }
|
| 1325 |
|
| 1326 |
|
| 1327 | static inline void pgalloc_tag_sub_pages(struct alloc_tag *tag, unsigned int nr)
|
| 1328 | {
|
| 1329 | if (tag)
|
| 1330 | this_cpu_sub(tag->counters->bytes, PAGE_SIZE * nr);
|
| 1331 | }
|
| 1332 |
|
| 1333 | #else
|
| 1334 |
|
| 1335 | static inline void pgalloc_tag_add(struct page *page, struct task_struct *task,
|
| 1336 | unsigned int nr) {}
|
| 1337 | static inline void pgalloc_tag_sub(struct page *page, unsigned int nr) {}
|
| 1338 | static inline void pgalloc_tag_sub_pages(struct alloc_tag *tag, unsigned int nr) {}
|
| 1339 |
|
| 1340 | #endif
|
| 1341 |
|
| 1342 | __always_inline bool __free_pages_prepare(struct page *page,
|
| 1343 | unsigned int order, fpi_t fpi_flags)
|
| 1344 | {
|
| 1345 | int bad = 0;
|
| 1346 | bool skip_kasan_poison = should_skip_kasan_poison(page);
|
| 1347 | bool init = want_init_on_free();
|
| 1348 | bool compound = PageCompound(page);
|
| 1349 | struct folio *folio = page_folio(page);
|
| 1350 |
|
| 1351 | VM_BUG_ON_PAGE(PageTail(page), page);
|
| 1352 |
|
| 1353 | trace_mm_page_free(page, order);
|
| 1354 | kmsan_free_page(page, order);
|
| 1355 |
|
| 1356 | if (memcg_kmem_online() && PageMemcgKmem(page))
|
| 1357 | __memcg_kmem_uncharge_page(page, order);
|
| 1358 |
|
| 1359 | |
| 1360 | * In rare cases, when truncation or holepunching raced with
|
| 1361 | * munlock after VM_LOCKED was cleared, Mlocked may still be
|
| 1362 | * found set here. This does not indicate a problem, unless
|
| 1363 | * "unevictable_pgs_cleared" appears worryingly large.
|
| 1364 |
|
| 1365 | if (unlikely(folio_test_mlocked(folio))) {
|
| 1366 | long nr_pages = folio_nr_pages(folio);
|
| 1367 |
|
| 1368 | __folio_clear_mlocked(folio);
|
| 1369 | zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
|
| 1370 | count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
|
| 1371 | }
|
| 1372 |
|
| 1373 | if (unlikely(PageHWPoison(page)) && !order) {
|
| 1374 |
|
| 1375 | reset_page_owner(page, order);
|
| 1376 | page_table_check_free(page, order);
|
| 1377 | pgalloc_tag_sub(page, 1 << order);
|
| 1378 |
|
| 1379 | |
| 1380 | * The page is isolated and accounted for.
|
| 1381 | * Mark the codetag as empty to avoid accounting error
|
| 1382 | * when the page is freed by unpoison_memory().
|
| 1383 |
|
| 1384 | clear_page_tag_ref(page);
|
| 1385 | return false;
|
| 1386 | }
|
| 1387 |
|
| 1388 | VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
|
| 1389 |
|
| 1390 | |
| 1391 | * Check tail pages before head page information is cleared to
|
| 1392 | * avoid checking PageCompound for order-0 pages.
|
| 1393 |
|
| 1394 | if (unlikely(order)) {
|
| 1395 | int i;
|
| 1396 |
|
| 1397 | if (compound) {
|
| 1398 | page[1].flags.f &= ~PAGE_FLAGS_SECOND;
|
| 1399 | #ifdef NR_PAGES_IN_LARGE_FOLIO
|
| 1400 | folio->_nr_pages = 0;
|
| 1401 | #endif
|
| 1402 | }
|
| 1403 | for (i = 1; i < (1 << order); i++) {
|
| 1404 | if (compound)
|
| 1405 | bad += free_tail_page_prepare(page, page + i);
|
| 1406 | if (is_check_pages_enabled()) {
|
| 1407 | if (free_page_is_bad(page + i)) {
|
| 1408 | bad++;
|
| 1409 | continue;
|
| 1410 | }
|
| 1411 | }
|
| 1412 | (page + i)->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
|
| 1413 | }
|
| 1414 | }
|
| 1415 | if (folio_test_anon(folio)) {
|
| 1416 | mod_mthp_stat(order, MTHP_STAT_NR_ANON, -1);
|
| 1417 | folio->mapping = NULL;
|
| 1418 | }
|
| 1419 | if (unlikely(page_has_type(page)))
|
| 1420 |
|
| 1421 | page->page_type = UINT_MAX;
|
| 1422 |
|
| 1423 | if (is_check_pages_enabled()) {
|
| 1424 | if (free_page_is_bad(page))
|
| 1425 | bad++;
|
| 1426 | if (bad)
|
| 1427 | return false;
|
| 1428 | }
|
| 1429 |
|
| 1430 | page_cpupid_reset_last(page);
|
| 1431 | page->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
|
| 1432 | page->private = 0;
|
| 1433 | reset_page_owner(page, order);
|
| 1434 | page_table_check_free(page, order);
|
| 1435 | pgalloc_tag_sub(page, 1 << order);
|
| 1436 |
|
| 1437 | if (!PageHighMem(page) && !(fpi_flags & FPI_TRYLOCK)) {
|
| 1438 | debug_check_no_locks_freed(page_address(page),
|
| 1439 | PAGE_SIZE << order);
|
| 1440 | debug_check_no_obj_freed(page_address(page),
|
| 1441 | PAGE_SIZE << order);
|
| 1442 | }
|
| 1443 |
|
| 1444 | kernel_poison_pages(page, 1 << order);
|
| 1445 |
|
| 1446 | |
| 1447 | * As memory initialization might be integrated into KASAN,
|
| 1448 | * KASAN poisoning and memory initialization code must be
|
| 1449 | * kept together to avoid discrepancies in behavior.
|
| 1450 | *
|
| 1451 | * With hardware tag-based KASAN, memory tags must be set before the
|
| 1452 | * page becomes unavailable via debug_pagealloc or arch_free_page.
|
| 1453 |
|
| 1454 | if (!skip_kasan_poison) {
|
| 1455 | kasan_poison_pages(page, order, init);
|
| 1456 |
|
| 1457 |
|
| 1458 | if (kasan_has_integrated_init())
|
| 1459 | init = false;
|
| 1460 | }
|
| 1461 | if (init)
|
| 1462 | kernel_init_pages(page, 1 << order);
|
| 1463 |
|
| 1464 | |
| 1465 | * arch_free_page() can make the page's contents inaccessible. s390
|
| 1466 | * does this. So nothing which can access the page's contents should
|
| 1467 | * happen after this.
|
| 1468 |
|
| 1469 | arch_free_page(page, order);
|
| 1470 |
|
| 1471 | debug_pagealloc_unmap_pages(page, 1 << order);
|
| 1472 |
|
| 1473 | return true;
|
| 1474 | }
|
| 1475 |
|
| 1476 | bool free_pages_prepare(struct page *page, unsigned int order)
|
| 1477 | {
|
| 1478 | return __free_pages_prepare(page, order, FPI_NONE);
|
| 1479 | }
|
| 1480 |
|
| 1481 | |
| 1482 | * Frees a number of pages from the PCP lists
|
| 1483 | * Assumes all pages on list are in same zone.
|
| 1484 | * count is the number of pages to free.
|
| 1485 |
|
| 1486 | static void free_pcppages_bulk(struct zone *zone, int count,
|
| 1487 | struct per_cpu_pages *pcp,
|
| 1488 | int pindex)
|
| 1489 | {
|
| 1490 | unsigned long flags;
|
| 1491 | unsigned int order;
|
| 1492 | struct page *page;
|
| 1493 |
|
| 1494 | |
| 1495 | * Ensure proper count is passed which otherwise would stuck in the
|
| 1496 | * below while (list_empty(list)) loop.
|
| 1497 |
|
| 1498 | count = min(pcp->count, count);
|
| 1499 |
|
| 1500 |
|
| 1501 | pindex = pindex - 1;
|
| 1502 |
|
| 1503 | spin_lock_irqsave(&zone->lock, flags);
|
| 1504 |
|
| 1505 | while (count > 0) {
|
| 1506 | struct list_head *list;
|
| 1507 | int nr_pages;
|
| 1508 |
|
| 1509 |
|
| 1510 | do {
|
| 1511 | if (++pindex > NR_PCP_LISTS - 1)
|
| 1512 | pindex = 0;
|
| 1513 | list = &pcp->lists[pindex];
|
| 1514 | } while (list_empty(list));
|
| 1515 |
|
| 1516 | order = pindex_to_order(pindex);
|
| 1517 | nr_pages = 1 << order;
|
| 1518 | do {
|
| 1519 | unsigned long pfn;
|
| 1520 | int mt;
|
| 1521 |
|
| 1522 | page = list_last_entry(list, struct page, pcp_list);
|
| 1523 | pfn = page_to_pfn(page);
|
| 1524 | mt = get_pfnblock_migratetype(page, pfn);
|
| 1525 |
|
| 1526 |
|
| 1527 | list_del(&page->pcp_list);
|
| 1528 | count -= nr_pages;
|
| 1529 | pcp->count -= nr_pages;
|
| 1530 |
|
| 1531 | __free_one_page(page, pfn, zone, order, mt, FPI_NONE);
|
| 1532 | trace_mm_page_pcpu_drain(page, order, mt);
|
| 1533 | } while (count > 0 && !list_empty(list));
|
| 1534 | }
|
| 1535 |
|
| 1536 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 1537 | }
|
| 1538 |
|
| 1539 |
|
| 1540 | static void split_large_buddy(struct zone *zone, struct page *page,
|
| 1541 | unsigned long pfn, int order, fpi_t fpi)
|
| 1542 | {
|
| 1543 | unsigned long end = pfn + (1 << order);
|
| 1544 |
|
| 1545 | VM_WARN_ON_ONCE(!IS_ALIGNED(pfn, 1 << order));
|
| 1546 |
|
| 1547 | VM_WARN_ON_ONCE(PageBuddy(page));
|
| 1548 |
|
| 1549 | if (order > pageblock_order)
|
| 1550 | order = pageblock_order;
|
| 1551 |
|
| 1552 | do {
|
| 1553 | int mt = get_pfnblock_migratetype(page, pfn);
|
| 1554 |
|
| 1555 | __free_one_page(page, pfn, zone, order, mt, fpi);
|
| 1556 | pfn += 1 << order;
|
| 1557 | if (pfn == end)
|
| 1558 | break;
|
| 1559 | page = pfn_to_page(pfn);
|
| 1560 | } while (1);
|
| 1561 | }
|
| 1562 |
|
| 1563 | static void add_page_to_zone_llist(struct zone *zone, struct page *page,
|
| 1564 | unsigned int order)
|
| 1565 | {
|
| 1566 |
|
| 1567 | page->private = order;
|
| 1568 |
|
| 1569 | llist_add(&page->pcp_llist, &zone->trylock_free_pages);
|
| 1570 | }
|
| 1571 |
|
| 1572 | static void free_one_page(struct zone *zone, struct page *page,
|
| 1573 | unsigned long pfn, unsigned int order,
|
| 1574 | fpi_t fpi_flags)
|
| 1575 | {
|
| 1576 | struct llist_head *llhead;
|
| 1577 | unsigned long flags;
|
| 1578 |
|
| 1579 | if (unlikely(fpi_flags & FPI_TRYLOCK)) {
|
| 1580 | if (!spin_trylock_irqsave(&zone->lock, flags)) {
|
| 1581 | add_page_to_zone_llist(zone, page, order);
|
| 1582 | return;
|
| 1583 | }
|
| 1584 | } else {
|
| 1585 | spin_lock_irqsave(&zone->lock, flags);
|
| 1586 | }
|
| 1587 |
|
| 1588 |
|
| 1589 | llhead = &zone->trylock_free_pages;
|
| 1590 | if (unlikely(!llist_empty(llhead) && !(fpi_flags & FPI_TRYLOCK))) {
|
| 1591 | struct llist_node *llnode;
|
| 1592 | struct page *p, *tmp;
|
| 1593 |
|
| 1594 | llnode = llist_del_all(llhead);
|
| 1595 | llist_for_each_entry_safe(p, tmp, llnode, pcp_llist) {
|
| 1596 | unsigned int p_order = p->private;
|
| 1597 |
|
| 1598 | split_large_buddy(zone, p, page_to_pfn(p), p_order, fpi_flags);
|
| 1599 | __count_vm_events(PGFREE, 1 << p_order);
|
| 1600 | }
|
| 1601 | }
|
| 1602 | split_large_buddy(zone, page, pfn, order, fpi_flags);
|
| 1603 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 1604 |
|
| 1605 | __count_vm_events(PGFREE, 1 << order);
|
| 1606 | }
|
| 1607 |
|
| 1608 | static void __free_pages_ok(struct page *page, unsigned int order,
|
| 1609 | fpi_t fpi_flags)
|
| 1610 | {
|
| 1611 | unsigned long pfn = page_to_pfn(page);
|
| 1612 | struct zone *zone = page_zone(page);
|
| 1613 |
|
| 1614 | if (__free_pages_prepare(page, order, fpi_flags))
|
| 1615 | free_one_page(zone, page, pfn, order, fpi_flags);
|
| 1616 | }
|
| 1617 |
|
| 1618 | void __meminit __free_pages_core(struct page *page, unsigned int order,
|
| 1619 | enum meminit_context context)
|
| 1620 | {
|
| 1621 | unsigned int nr_pages = 1 << order;
|
| 1622 | struct page *p = page;
|
| 1623 | unsigned int loop;
|
| 1624 |
|
| 1625 | |
| 1626 | * When initializing the memmap, __init_single_page() sets the refcount
|
| 1627 | * of all pages to 1 ("allocated"/"not free"). We have to set the
|
| 1628 | * refcount of all involved pages to 0.
|
| 1629 | *
|
| 1630 | * Note that hotplugged memory pages are initialized to PageOffline().
|
| 1631 | * Pages freed from memblock might be marked as reserved.
|
| 1632 |
|
| 1633 | if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG) &&
|
| 1634 | unlikely(context == MEMINIT_HOTPLUG)) {
|
| 1635 | for (loop = 0; loop < nr_pages; loop++, p++) {
|
| 1636 | VM_WARN_ON_ONCE(PageReserved(p));
|
| 1637 | __ClearPageOffline(p);
|
| 1638 | set_page_count(p, 0);
|
| 1639 | }
|
| 1640 |
|
| 1641 | adjust_managed_page_count(page, nr_pages);
|
| 1642 | } else {
|
| 1643 | for (loop = 0; loop < nr_pages; loop++, p++) {
|
| 1644 | __ClearPageReserved(p);
|
| 1645 | set_page_count(p, 0);
|
| 1646 | }
|
| 1647 |
|
| 1648 |
|
| 1649 | atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
|
| 1650 | }
|
| 1651 |
|
| 1652 | if (page_contains_unaccepted(page, order)) {
|
| 1653 | if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
|
| 1654 | return;
|
| 1655 |
|
| 1656 | accept_memory(page_to_phys(page), PAGE_SIZE << order);
|
| 1657 | }
|
| 1658 |
|
| 1659 | |
| 1660 | * Bypass PCP and place fresh pages right to the tail, primarily
|
| 1661 | * relevant for memory onlining.
|
| 1662 |
|
| 1663 | __free_pages_ok(page, order, FPI_TO_TAIL);
|
| 1664 | }
|
| 1665 |
|
| 1666 | |
| 1667 | * Check that the whole (or subset of) a pageblock given by the interval of
|
| 1668 | * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
|
| 1669 | * with the migration of free compaction scanner.
|
| 1670 | *
|
| 1671 | * Return struct page pointer of start_pfn, or NULL if checks were not passed.
|
| 1672 | *
|
| 1673 | * It's possible on some configurations to have a setup like node0 node1 node0
|
| 1674 | * i.e. it's possible that all pages within a zones range of pages do not
|
| 1675 | * belong to a single zone. We assume that a border between node0 and node1
|
| 1676 | * can occur within a single pageblock, but not a node0 node1 node0
|
| 1677 | * interleaving within a single pageblock. It is therefore sufficient to check
|
| 1678 | * the first and last page of a pageblock and avoid checking each individual
|
| 1679 | * page in a pageblock.
|
| 1680 | *
|
| 1681 | * Note: the function may return non-NULL struct page even for a page block
|
| 1682 | * which contains a memory hole (i.e. there is no physical memory for a subset
|
| 1683 | * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
|
| 1684 | * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
|
| 1685 | * even though the start pfn is online and valid. This should be safe most of
|
| 1686 | * the time because struct pages are still initialized via init_unavailable_range()
|
| 1687 | * and pfn walkers shouldn't touch any physical memory range for which they do
|
| 1688 | * not recognize any specific metadata in struct pages.
|
| 1689 |
|
| 1690 | struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
|
| 1691 | unsigned long end_pfn, struct zone *zone)
|
| 1692 | {
|
| 1693 | struct page *start_page;
|
| 1694 | struct page *end_page;
|
| 1695 |
|
| 1696 |
|
| 1697 | end_pfn--;
|
| 1698 |
|
| 1699 | if (!pfn_valid(end_pfn))
|
| 1700 | return NULL;
|
| 1701 |
|
| 1702 | start_page = pfn_to_online_page(start_pfn);
|
| 1703 | if (!start_page)
|
| 1704 | return NULL;
|
| 1705 |
|
| 1706 | if (page_zone(start_page) != zone)
|
| 1707 | return NULL;
|
| 1708 |
|
| 1709 | end_page = pfn_to_page(end_pfn);
|
| 1710 |
|
| 1711 |
|
| 1712 | if (page_zone_id(start_page) != page_zone_id(end_page))
|
| 1713 | return NULL;
|
| 1714 |
|
| 1715 | return start_page;
|
| 1716 | }
|
| 1717 |
|
| 1718 | |
| 1719 | * The order of subdivision here is critical for the IO subsystem.
|
| 1720 | * Please do not alter this order without good reasons and regression
|
| 1721 | * testing. Specifically, as large blocks of memory are subdivided,
|
| 1722 | * the order in which smaller blocks are delivered depends on the order
|
| 1723 | * they're subdivided in this function. This is the primary factor
|
| 1724 | * influencing the order in which pages are delivered to the IO
|
| 1725 | * subsystem according to empirical testing, and this is also justified
|
| 1726 | * by considering the behavior of a buddy system containing a single
|
| 1727 | * large block of memory acted on by a series of small allocations.
|
| 1728 | * This behavior is a critical factor in sglist merging's success.
|
| 1729 | *
|
| 1730 | * -- nyc
|
| 1731 |
|
| 1732 | static inline unsigned int expand(struct zone *zone, struct page *page, int low,
|
| 1733 | int high, int migratetype)
|
| 1734 | {
|
| 1735 | unsigned int size = 1 << high;
|
| 1736 | unsigned int nr_added = 0;
|
| 1737 |
|
| 1738 | while (high > low) {
|
| 1739 | high--;
|
| 1740 | size >>= 1;
|
| 1741 | VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
|
| 1742 |
|
| 1743 | |
| 1744 | * Mark as guard pages (or page), that will allow to
|
| 1745 | * merge back to allocator when buddy will be freed.
|
| 1746 | * Corresponding page table entries will not be touched,
|
| 1747 | * pages will stay not present in virtual address space
|
| 1748 |
|
| 1749 | if (set_page_guard(zone, &page[size], high))
|
| 1750 | continue;
|
| 1751 |
|
| 1752 | __add_to_free_list(&page[size], zone, high, migratetype, false);
|
| 1753 | set_buddy_order(&page[size], high);
|
| 1754 | nr_added += size;
|
| 1755 | }
|
| 1756 |
|
| 1757 | return nr_added;
|
| 1758 | }
|
| 1759 |
|
| 1760 | static __always_inline void page_del_and_expand(struct zone *zone,
|
| 1761 | struct page *page, int low,
|
| 1762 | int high, int migratetype)
|
| 1763 | {
|
| 1764 | int nr_pages = 1 << high;
|
| 1765 |
|
| 1766 | __del_page_from_free_list(page, zone, high, migratetype);
|
| 1767 | nr_pages -= expand(zone, page, low, high, migratetype);
|
| 1768 | account_freepages(zone, -nr_pages, migratetype);
|
| 1769 | }
|
| 1770 |
|
| 1771 | static void check_new_page_bad(struct page *page)
|
| 1772 | {
|
| 1773 | if (unlikely(PageHWPoison(page))) {
|
| 1774 |
|
| 1775 | if (PageBuddy(page))
|
| 1776 | __ClearPageBuddy(page);
|
| 1777 | return;
|
| 1778 | }
|
| 1779 |
|
| 1780 | bad_page(page,
|
| 1781 | page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
|
| 1782 | }
|
| 1783 |
|
| 1784 | |
| 1785 | * This page is about to be returned from the page allocator
|
| 1786 |
|
| 1787 | static bool check_new_page(struct page *page)
|
| 1788 | {
|
| 1789 | if (likely(page_expected_state(page,
|
| 1790 | PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
|
| 1791 | return false;
|
| 1792 |
|
| 1793 | check_new_page_bad(page);
|
| 1794 | return true;
|
| 1795 | }
|
| 1796 |
|
| 1797 | static inline bool check_new_pages(struct page *page, unsigned int order)
|
| 1798 | {
|
| 1799 | if (is_check_pages_enabled()) {
|
| 1800 | for (int i = 0; i < (1 << order); i++) {
|
| 1801 | struct page *p = page + i;
|
| 1802 |
|
| 1803 | if (check_new_page(p))
|
| 1804 | return true;
|
| 1805 | }
|
| 1806 | }
|
| 1807 |
|
| 1808 | return false;
|
| 1809 | }
|
| 1810 |
|
| 1811 | static inline bool should_skip_kasan_unpoison(gfp_t flags)
|
| 1812 | {
|
| 1813 |
|
| 1814 | if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
|
| 1815 | IS_ENABLED(CONFIG_KASAN_SW_TAGS))
|
| 1816 | return false;
|
| 1817 |
|
| 1818 |
|
| 1819 | if (!kasan_hw_tags_enabled())
|
| 1820 | return true;
|
| 1821 |
|
| 1822 | |
| 1823 | * With hardware tag-based KASAN enabled, skip if this has been
|
| 1824 | * requested via __GFP_SKIP_KASAN.
|
| 1825 |
|
| 1826 | return flags & __GFP_SKIP_KASAN;
|
| 1827 | }
|
| 1828 |
|
| 1829 | static inline bool should_skip_init(gfp_t flags)
|
| 1830 | {
|
| 1831 |
|
| 1832 | if (!kasan_hw_tags_enabled())
|
| 1833 | return false;
|
| 1834 |
|
| 1835 |
|
| 1836 | return (flags & __GFP_SKIP_ZERO);
|
| 1837 | }
|
| 1838 |
|
| 1839 | inline void post_alloc_hook(struct page *page, unsigned int order,
|
| 1840 | gfp_t gfp_flags)
|
| 1841 | {
|
| 1842 | bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
|
| 1843 | !should_skip_init(gfp_flags);
|
| 1844 | bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
|
| 1845 | int i;
|
| 1846 |
|
| 1847 | set_page_private(page, 0);
|
| 1848 |
|
| 1849 | arch_alloc_page(page, order);
|
| 1850 | debug_pagealloc_map_pages(page, 1 << order);
|
| 1851 |
|
| 1852 | |
| 1853 | * Page unpoisoning must happen before memory initialization.
|
| 1854 | * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
|
| 1855 | * allocations and the page unpoisoning code will complain.
|
| 1856 |
|
| 1857 | kernel_unpoison_pages(page, 1 << order);
|
| 1858 |
|
| 1859 | |
| 1860 | * As memory initialization might be integrated into KASAN,
|
| 1861 | * KASAN unpoisoning and memory initialization code must be
|
| 1862 | * kept together to avoid discrepancies in behavior.
|
| 1863 |
|
| 1864 |
|
| 1865 | |
| 1866 | * If memory tags should be zeroed
|
| 1867 | * (which happens only when memory should be initialized as well).
|
| 1868 |
|
| 1869 | if (zero_tags)
|
| 1870 | init = !tag_clear_highpages(page, 1 << order);
|
| 1871 |
|
| 1872 | if (!should_skip_kasan_unpoison(gfp_flags) &&
|
| 1873 | kasan_unpoison_pages(page, order, init)) {
|
| 1874 |
|
| 1875 | if (kasan_has_integrated_init())
|
| 1876 | init = false;
|
| 1877 | } else {
|
| 1878 | |
| 1879 | * If memory tags have not been set by KASAN, reset the page
|
| 1880 | * tags to ensure page_address() dereferencing does not fault.
|
| 1881 |
|
| 1882 | for (i = 0; i != 1 << order; ++i)
|
| 1883 | page_kasan_tag_reset(page + i);
|
| 1884 | }
|
| 1885 |
|
| 1886 | if (init)
|
| 1887 | kernel_init_pages(page, 1 << order);
|
| 1888 |
|
| 1889 | set_page_owner(page, order, gfp_flags);
|
| 1890 | page_table_check_alloc(page, order);
|
| 1891 | pgalloc_tag_add(page, current, 1 << order);
|
| 1892 | }
|
| 1893 |
|
| 1894 | static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
|
| 1895 | unsigned int alloc_flags)
|
| 1896 | {
|
| 1897 | post_alloc_hook(page, order, gfp_flags);
|
| 1898 |
|
| 1899 | if (order && (gfp_flags & __GFP_COMP))
|
| 1900 | prep_compound_page(page, order);
|
| 1901 |
|
| 1902 | |
| 1903 | * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
|
| 1904 | * allocate the page. The expectation is that the caller is taking
|
| 1905 | * steps that will free more memory. The caller should avoid the page
|
| 1906 | * being used for !PFMEMALLOC purposes.
|
| 1907 |
|
| 1908 | if (alloc_flags & ALLOC_NO_WATERMARKS)
|
| 1909 | set_page_pfmemalloc(page);
|
| 1910 | else
|
| 1911 | clear_page_pfmemalloc(page);
|
| 1912 | }
|
| 1913 |
|
| 1914 | |
| 1915 | * Go through the free lists for the given migratetype and remove
|
| 1916 | * the smallest available page from the freelists
|
| 1917 |
|
| 1918 | static __always_inline
|
| 1919 | struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
|
| 1920 | int migratetype)
|
| 1921 | {
|
| 1922 | unsigned int current_order;
|
| 1923 | struct free_area *area;
|
| 1924 | struct page *page;
|
| 1925 |
|
| 1926 |
|
| 1927 | for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
|
| 1928 | area = &(zone->free_area[current_order]);
|
| 1929 | page = get_page_from_free_area(area, migratetype);
|
| 1930 | if (!page)
|
| 1931 | continue;
|
| 1932 |
|
| 1933 | page_del_and_expand(zone, page, order, current_order,
|
| 1934 | migratetype);
|
| 1935 | trace_mm_page_alloc_zone_locked(page, order, migratetype,
|
| 1936 | pcp_allowed_order(order) &&
|
| 1937 | migratetype < MIGRATE_PCPTYPES);
|
| 1938 | return page;
|
| 1939 | }
|
| 1940 |
|
| 1941 | return NULL;
|
| 1942 | }
|
| 1943 |
|
| 1944 |
|
| 1945 | |
| 1946 | * This array describes the order lists are fallen back to when
|
| 1947 | * the free lists for the desirable migrate type are depleted
|
| 1948 | *
|
| 1949 | * The other migratetypes do not have fallbacks.
|
| 1950 |
|
| 1951 | static int fallbacks[MIGRATE_PCPTYPES][MIGRATE_PCPTYPES - 1] = {
|
| 1952 | [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
|
| 1953 | [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
|
| 1954 | [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
|
| 1955 | };
|
| 1956 |
|
| 1957 | #ifdef CONFIG_CMA
|
| 1958 | static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
|
| 1959 | unsigned int order)
|
| 1960 | {
|
| 1961 | return __rmqueue_smallest(zone, order, MIGRATE_CMA);
|
| 1962 | }
|
| 1963 | #else
|
| 1964 | static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
|
| 1965 | unsigned int order) { return NULL; }
|
| 1966 | #endif
|
| 1967 |
|
| 1968 | |
| 1969 | * Move all free pages of a block to new type's freelist. Caller needs to
|
| 1970 | * change the block type.
|
| 1971 |
|
| 1972 | static int __move_freepages_block(struct zone *zone, unsigned long start_pfn,
|
| 1973 | int old_mt, int new_mt)
|
| 1974 | {
|
| 1975 | struct page *page;
|
| 1976 | unsigned long pfn, end_pfn;
|
| 1977 | unsigned int order;
|
| 1978 | int pages_moved = 0;
|
| 1979 |
|
| 1980 | VM_WARN_ON(start_pfn & (pageblock_nr_pages - 1));
|
| 1981 | end_pfn = pageblock_end_pfn(start_pfn);
|
| 1982 |
|
| 1983 | for (pfn = start_pfn; pfn < end_pfn;) {
|
| 1984 | page = pfn_to_page(pfn);
|
| 1985 | if (!PageBuddy(page)) {
|
| 1986 | pfn++;
|
| 1987 | continue;
|
| 1988 | }
|
| 1989 |
|
| 1990 |
|
| 1991 | VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
|
| 1992 | VM_BUG_ON_PAGE(page_zone(page) != zone, page);
|
| 1993 |
|
| 1994 | order = buddy_order(page);
|
| 1995 |
|
| 1996 | move_to_free_list(page, zone, order, old_mt, new_mt);
|
| 1997 |
|
| 1998 | pfn += 1 << order;
|
| 1999 | pages_moved += 1 << order;
|
| 2000 | }
|
| 2001 |
|
| 2002 | return pages_moved;
|
| 2003 | }
|
| 2004 |
|
| 2005 | static bool prep_move_freepages_block(struct zone *zone, struct page *page,
|
| 2006 | unsigned long *start_pfn,
|
| 2007 | int *num_free, int *num_movable)
|
| 2008 | {
|
| 2009 | unsigned long pfn, start, end;
|
| 2010 |
|
| 2011 | pfn = page_to_pfn(page);
|
| 2012 | start = pageblock_start_pfn(pfn);
|
| 2013 | end = pageblock_end_pfn(pfn);
|
| 2014 |
|
| 2015 | |
| 2016 | * The caller only has the lock for @zone, don't touch ranges
|
| 2017 | * that straddle into other zones. While we could move part of
|
| 2018 | * the range that's inside the zone, this call is usually
|
| 2019 | * accompanied by other operations such as migratetype updates
|
| 2020 | * which also should be locked.
|
| 2021 |
|
| 2022 | if (!zone_spans_pfn(zone, start))
|
| 2023 | return false;
|
| 2024 | if (!zone_spans_pfn(zone, end - 1))
|
| 2025 | return false;
|
| 2026 |
|
| 2027 | *start_pfn = start;
|
| 2028 |
|
| 2029 | if (num_free) {
|
| 2030 | *num_free = 0;
|
| 2031 | *num_movable = 0;
|
| 2032 | for (pfn = start; pfn < end;) {
|
| 2033 | page = pfn_to_page(pfn);
|
| 2034 | if (PageBuddy(page)) {
|
| 2035 | int nr = 1 << buddy_order(page);
|
| 2036 |
|
| 2037 | *num_free += nr;
|
| 2038 | pfn += nr;
|
| 2039 | continue;
|
| 2040 | }
|
| 2041 | |
| 2042 | * We assume that pages that could be isolated for
|
| 2043 | * migration are movable. But we don't actually try
|
| 2044 | * isolating, as that would be expensive.
|
| 2045 |
|
| 2046 | if (PageLRU(page) || page_has_movable_ops(page))
|
| 2047 | (*num_movable)++;
|
| 2048 | pfn++;
|
| 2049 | }
|
| 2050 | }
|
| 2051 |
|
| 2052 | return true;
|
| 2053 | }
|
| 2054 |
|
| 2055 | static int move_freepages_block(struct zone *zone, struct page *page,
|
| 2056 | int old_mt, int new_mt)
|
| 2057 | {
|
| 2058 | unsigned long start_pfn;
|
| 2059 | int res;
|
| 2060 |
|
| 2061 | if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
|
| 2062 | return -1;
|
| 2063 |
|
| 2064 | res = __move_freepages_block(zone, start_pfn, old_mt, new_mt);
|
| 2065 | set_pageblock_migratetype(pfn_to_page(start_pfn), new_mt);
|
| 2066 |
|
| 2067 | return res;
|
| 2068 |
|
| 2069 | }
|
| 2070 |
|
| 2071 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 2072 |
|
| 2073 | static unsigned long find_large_buddy(unsigned long start_pfn)
|
| 2074 | {
|
| 2075 | |
| 2076 | * If start_pfn is not an order-0 PageBuddy, next PageBuddy containing
|
| 2077 | * start_pfn has minimal order of __ffs(start_pfn) + 1. Start checking
|
| 2078 | * the order with __ffs(start_pfn). If start_pfn is order-0 PageBuddy,
|
| 2079 | * the starting order does not matter.
|
| 2080 |
|
| 2081 | int order = start_pfn ? __ffs(start_pfn) : MAX_PAGE_ORDER;
|
| 2082 | struct page *page;
|
| 2083 | unsigned long pfn = start_pfn;
|
| 2084 |
|
| 2085 | while (!PageBuddy(page = pfn_to_page(pfn))) {
|
| 2086 |
|
| 2087 | if (++order > MAX_PAGE_ORDER)
|
| 2088 | return start_pfn;
|
| 2089 | pfn &= ~0UL << order;
|
| 2090 | }
|
| 2091 |
|
| 2092 | |
| 2093 | * Found a preceding buddy, but does it straddle?
|
| 2094 |
|
| 2095 | if (pfn + (1 << buddy_order(page)) > start_pfn)
|
| 2096 | return pfn;
|
| 2097 |
|
| 2098 |
|
| 2099 | return start_pfn;
|
| 2100 | }
|
| 2101 |
|
| 2102 | static inline void toggle_pageblock_isolate(struct page *page, bool isolate)
|
| 2103 | {
|
| 2104 | if (isolate)
|
| 2105 | set_pageblock_isolate(page);
|
| 2106 | else
|
| 2107 | clear_pageblock_isolate(page);
|
| 2108 | }
|
| 2109 |
|
| 2110 | |
| 2111 | * __move_freepages_block_isolate - move free pages in block for page isolation
|
| 2112 | * @zone: the zone
|
| 2113 | * @page: the pageblock page
|
| 2114 | * @isolate: to isolate the given pageblock or unisolate it
|
| 2115 | *
|
| 2116 | * This is similar to move_freepages_block(), but handles the special
|
| 2117 | * case encountered in page isolation, where the block of interest
|
| 2118 | * might be part of a larger buddy spanning multiple pageblocks.
|
| 2119 | *
|
| 2120 | * Unlike the regular page allocator path, which moves pages while
|
| 2121 | * stealing buddies off the freelist, page isolation is interested in
|
| 2122 | * arbitrary pfn ranges that may have overlapping buddies on both ends.
|
| 2123 | *
|
| 2124 | * This function handles that. Straddling buddies are split into
|
| 2125 | * individual pageblocks. Only the block of interest is moved.
|
| 2126 | *
|
| 2127 | * Returns %true if pages could be moved, %false otherwise.
|
| 2128 |
|
| 2129 | static bool __move_freepages_block_isolate(struct zone *zone,
|
| 2130 | struct page *page, bool isolate)
|
| 2131 | {
|
| 2132 | unsigned long start_pfn, buddy_pfn;
|
| 2133 | int from_mt;
|
| 2134 | int to_mt;
|
| 2135 | struct page *buddy;
|
| 2136 |
|
| 2137 | if (isolate == get_pageblock_isolate(page)) {
|
| 2138 | VM_WARN_ONCE(1, "%s a pageblock that is already in that state",
|
| 2139 | isolate ? "Isolate" : "Unisolate");
|
| 2140 | return false;
|
| 2141 | }
|
| 2142 |
|
| 2143 | if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
|
| 2144 | return false;
|
| 2145 |
|
| 2146 |
|
| 2147 | if (pageblock_order == MAX_PAGE_ORDER)
|
| 2148 | goto move;
|
| 2149 |
|
| 2150 | buddy_pfn = find_large_buddy(start_pfn);
|
| 2151 | buddy = pfn_to_page(buddy_pfn);
|
| 2152 |
|
| 2153 | if (PageBuddy(buddy) && buddy_order(buddy) > pageblock_order) {
|
| 2154 | int order = buddy_order(buddy);
|
| 2155 |
|
| 2156 | del_page_from_free_list(buddy, zone, order,
|
| 2157 | get_pfnblock_migratetype(buddy, buddy_pfn));
|
| 2158 | toggle_pageblock_isolate(page, isolate);
|
| 2159 | split_large_buddy(zone, buddy, buddy_pfn, order, FPI_NONE);
|
| 2160 | return true;
|
| 2161 | }
|
| 2162 |
|
| 2163 | move:
|
| 2164 |
|
| 2165 | if (isolate) {
|
| 2166 | from_mt = __get_pfnblock_flags_mask(page, page_to_pfn(page),
|
| 2167 | MIGRATETYPE_MASK);
|
| 2168 | to_mt = MIGRATE_ISOLATE;
|
| 2169 | } else {
|
| 2170 | from_mt = MIGRATE_ISOLATE;
|
| 2171 | to_mt = __get_pfnblock_flags_mask(page, page_to_pfn(page),
|
| 2172 | MIGRATETYPE_MASK);
|
| 2173 | }
|
| 2174 |
|
| 2175 | __move_freepages_block(zone, start_pfn, from_mt, to_mt);
|
| 2176 | toggle_pageblock_isolate(pfn_to_page(start_pfn), isolate);
|
| 2177 |
|
| 2178 | return true;
|
| 2179 | }
|
| 2180 |
|
| 2181 | bool pageblock_isolate_and_move_free_pages(struct zone *zone, struct page *page)
|
| 2182 | {
|
| 2183 | return __move_freepages_block_isolate(zone, page, true);
|
| 2184 | }
|
| 2185 |
|
| 2186 | bool pageblock_unisolate_and_move_free_pages(struct zone *zone, struct page *page)
|
| 2187 | {
|
| 2188 | return __move_freepages_block_isolate(zone, page, false);
|
| 2189 | }
|
| 2190 |
|
| 2191 | #endif
|
| 2192 |
|
| 2193 | static inline bool boost_watermark(struct zone *zone)
|
| 2194 | {
|
| 2195 | unsigned long max_boost;
|
| 2196 |
|
| 2197 | if (!watermark_boost_factor)
|
| 2198 | return false;
|
| 2199 | |
| 2200 | * Don't bother in zones that are unlikely to produce results.
|
| 2201 | * On small machines, including kdump capture kernels running
|
| 2202 | * in a small area, boosting the watermark can cause an out of
|
| 2203 | * memory situation immediately.
|
| 2204 |
|
| 2205 | if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
|
| 2206 | return false;
|
| 2207 |
|
| 2208 | max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
|
| 2209 | watermark_boost_factor, 10000);
|
| 2210 |
|
| 2211 | |
| 2212 | * high watermark may be uninitialised if fragmentation occurs
|
| 2213 | * very early in boot so do not boost. We do not fall
|
| 2214 | * through and boost by pageblock_nr_pages as failing
|
| 2215 | * allocations that early means that reclaim is not going
|
| 2216 | * to help and it may even be impossible to reclaim the
|
| 2217 | * boosted watermark resulting in a hang.
|
| 2218 |
|
| 2219 | if (!max_boost)
|
| 2220 | return false;
|
| 2221 |
|
| 2222 | max_boost = max(pageblock_nr_pages, max_boost);
|
| 2223 |
|
| 2224 | zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
|
| 2225 | max_boost);
|
| 2226 |
|
| 2227 | return true;
|
| 2228 | }
|
| 2229 |
|
| 2230 | |
| 2231 | * When we are falling back to another migratetype during allocation, should we
|
| 2232 | * try to claim an entire block to satisfy further allocations, instead of
|
| 2233 | * polluting multiple pageblocks?
|
| 2234 |
|
| 2235 | static bool should_try_claim_block(unsigned int order, int start_mt)
|
| 2236 | {
|
| 2237 | |
| 2238 | * Leaving this order check is intended, although there is
|
| 2239 | * relaxed order check in next check. The reason is that
|
| 2240 | * we can actually claim the whole pageblock if this condition met,
|
| 2241 | * but, below check doesn't guarantee it and that is just heuristic
|
| 2242 | * so could be changed anytime.
|
| 2243 |
|
| 2244 | if (order >= pageblock_order)
|
| 2245 | return true;
|
| 2246 |
|
| 2247 | |
| 2248 | * Above a certain threshold, always try to claim, as it's likely there
|
| 2249 | * will be more free pages in the pageblock.
|
| 2250 |
|
| 2251 | if (order >= pageblock_order / 2)
|
| 2252 | return true;
|
| 2253 |
|
| 2254 | |
| 2255 | * Unmovable/reclaimable allocations would cause permanent
|
| 2256 | * fragmentations if they fell back to allocating from a movable block
|
| 2257 | * (polluting it), so we try to claim the whole block regardless of the
|
| 2258 | * allocation size. Later movable allocations can always steal from this
|
| 2259 | * block, which is less problematic.
|
| 2260 |
|
| 2261 | if (start_mt == MIGRATE_RECLAIMABLE || start_mt == MIGRATE_UNMOVABLE)
|
| 2262 | return true;
|
| 2263 |
|
| 2264 | if (page_group_by_mobility_disabled)
|
| 2265 | return true;
|
| 2266 |
|
| 2267 | |
| 2268 | * Movable pages won't cause permanent fragmentation, so when you alloc
|
| 2269 | * small pages, we just need to temporarily steal unmovable or
|
| 2270 | * reclaimable pages that are closest to the request size. After a
|
| 2271 | * while, memory compaction may occur to form large contiguous pages,
|
| 2272 | * and the next movable allocation may not need to steal.
|
| 2273 |
|
| 2274 | return false;
|
| 2275 | }
|
| 2276 |
|
| 2277 | |
| 2278 | * Check whether there is a suitable fallback freepage with requested order.
|
| 2279 | * If claimable is true, this function returns fallback_mt only if
|
| 2280 | * we would do this whole-block claiming. This would help to reduce
|
| 2281 | * fragmentation due to mixed migratetype pages in one pageblock.
|
| 2282 |
|
| 2283 | int find_suitable_fallback(struct free_area *area, unsigned int order,
|
| 2284 | int migratetype, bool claimable)
|
| 2285 | {
|
| 2286 | int i;
|
| 2287 |
|
| 2288 | if (claimable && !should_try_claim_block(order, migratetype))
|
| 2289 | return -2;
|
| 2290 |
|
| 2291 | if (area->nr_free == 0)
|
| 2292 | return -1;
|
| 2293 |
|
| 2294 | for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
|
| 2295 | int fallback_mt = fallbacks[migratetype][i];
|
| 2296 |
|
| 2297 | if (!free_area_empty(area, fallback_mt))
|
| 2298 | return fallback_mt;
|
| 2299 | }
|
| 2300 |
|
| 2301 | return -1;
|
| 2302 | }
|
| 2303 |
|
| 2304 | |
| 2305 | * This function implements actual block claiming behaviour. If order is large
|
| 2306 | * enough, we can claim the whole pageblock for the requested migratetype. If
|
| 2307 | * not, we check the pageblock for constituent pages; if at least half of the
|
| 2308 | * pages are free or compatible, we can still claim the whole block, so pages
|
| 2309 | * freed in the future will be put on the correct free list.
|
| 2310 |
|
| 2311 | static struct page *
|
| 2312 | try_to_claim_block(struct zone *zone, struct page *page,
|
| 2313 | int current_order, int order, int start_type,
|
| 2314 | int block_type, unsigned int alloc_flags)
|
| 2315 | {
|
| 2316 | int free_pages, movable_pages, alike_pages;
|
| 2317 | unsigned long start_pfn;
|
| 2318 |
|
| 2319 |
|
| 2320 | if (current_order >= pageblock_order) {
|
| 2321 | unsigned int nr_added;
|
| 2322 |
|
| 2323 | del_page_from_free_list(page, zone, current_order, block_type);
|
| 2324 | change_pageblock_range(page, current_order, start_type);
|
| 2325 | nr_added = expand(zone, page, order, current_order, start_type);
|
| 2326 | account_freepages(zone, nr_added, start_type);
|
| 2327 | return page;
|
| 2328 | }
|
| 2329 |
|
| 2330 | |
| 2331 | * Boost watermarks to increase reclaim pressure to reduce the
|
| 2332 | * likelihood of future fallbacks. Wake kswapd now as the node
|
| 2333 | * may be balanced overall and kswapd will not wake naturally.
|
| 2334 |
|
| 2335 | if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
|
| 2336 | set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
|
| 2337 |
|
| 2338 |
|
| 2339 | if (!prep_move_freepages_block(zone, page, &start_pfn, &free_pages,
|
| 2340 | &movable_pages))
|
| 2341 | return NULL;
|
| 2342 |
|
| 2343 | |
| 2344 | * Determine how many pages are compatible with our allocation.
|
| 2345 | * For movable allocation, it's the number of movable pages which
|
| 2346 | * we just obtained. For other types it's a bit more tricky.
|
| 2347 |
|
| 2348 | if (start_type == MIGRATE_MOVABLE) {
|
| 2349 | alike_pages = movable_pages;
|
| 2350 | } else {
|
| 2351 | |
| 2352 | * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
|
| 2353 | * to MOVABLE pageblock, consider all non-movable pages as
|
| 2354 | * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
|
| 2355 | * vice versa, be conservative since we can't distinguish the
|
| 2356 | * exact migratetype of non-movable pages.
|
| 2357 |
|
| 2358 | if (block_type == MIGRATE_MOVABLE)
|
| 2359 | alike_pages = pageblock_nr_pages
|
| 2360 | - (free_pages + movable_pages);
|
| 2361 | else
|
| 2362 | alike_pages = 0;
|
| 2363 | }
|
| 2364 | |
| 2365 | * If a sufficient number of pages in the block are either free or of
|
| 2366 | * compatible migratability as our allocation, claim the whole block.
|
| 2367 |
|
| 2368 | if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
|
| 2369 | page_group_by_mobility_disabled) {
|
| 2370 | __move_freepages_block(zone, start_pfn, block_type, start_type);
|
| 2371 | set_pageblock_migratetype(pfn_to_page(start_pfn), start_type);
|
| 2372 | return __rmqueue_smallest(zone, order, start_type);
|
| 2373 | }
|
| 2374 |
|
| 2375 | return NULL;
|
| 2376 | }
|
| 2377 |
|
| 2378 | |
| 2379 | * Try to allocate from some fallback migratetype by claiming the entire block,
|
| 2380 | * i.e. converting it to the allocation's start migratetype.
|
| 2381 | *
|
| 2382 | * The use of signed ints for order and current_order is a deliberate
|
| 2383 | * deviation from the rest of this file, to make the for loop
|
| 2384 | * condition simpler.
|
| 2385 |
|
| 2386 | static __always_inline struct page *
|
| 2387 | __rmqueue_claim(struct zone *zone, int order, int start_migratetype,
|
| 2388 | unsigned int alloc_flags)
|
| 2389 | {
|
| 2390 | struct free_area *area;
|
| 2391 | int current_order;
|
| 2392 | int min_order = order;
|
| 2393 | struct page *page;
|
| 2394 | int fallback_mt;
|
| 2395 |
|
| 2396 | |
| 2397 | * Do not steal pages from freelists belonging to other pageblocks
|
| 2398 | * i.e. orders < pageblock_order. If there are no local zones free,
|
| 2399 | * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
|
| 2400 |
|
| 2401 | if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
|
| 2402 | min_order = pageblock_order;
|
| 2403 |
|
| 2404 | |
| 2405 | * Find the largest available free page in the other list. This roughly
|
| 2406 | * approximates finding the pageblock with the most free pages, which
|
| 2407 | * would be too costly to do exactly.
|
| 2408 |
|
| 2409 | for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
|
| 2410 | --current_order) {
|
| 2411 | area = &(zone->free_area[current_order]);
|
| 2412 | fallback_mt = find_suitable_fallback(area, current_order,
|
| 2413 | start_migratetype, true);
|
| 2414 |
|
| 2415 |
|
| 2416 | if (fallback_mt == -1)
|
| 2417 | continue;
|
| 2418 |
|
| 2419 |
|
| 2420 | if (fallback_mt == -2)
|
| 2421 | break;
|
| 2422 |
|
| 2423 | page = get_page_from_free_area(area, fallback_mt);
|
| 2424 | page = try_to_claim_block(zone, page, current_order, order,
|
| 2425 | start_migratetype, fallback_mt,
|
| 2426 | alloc_flags);
|
| 2427 | if (page) {
|
| 2428 | trace_mm_page_alloc_extfrag(page, order, current_order,
|
| 2429 | start_migratetype, fallback_mt);
|
| 2430 | return page;
|
| 2431 | }
|
| 2432 | }
|
| 2433 |
|
| 2434 | return NULL;
|
| 2435 | }
|
| 2436 |
|
| 2437 | |
| 2438 | * Try to steal a single page from some fallback migratetype. Leave the rest of
|
| 2439 | * the block as its current migratetype, potentially causing fragmentation.
|
| 2440 |
|
| 2441 | static __always_inline struct page *
|
| 2442 | __rmqueue_steal(struct zone *zone, int order, int start_migratetype)
|
| 2443 | {
|
| 2444 | struct free_area *area;
|
| 2445 | int current_order;
|
| 2446 | struct page *page;
|
| 2447 | int fallback_mt;
|
| 2448 |
|
| 2449 | for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
|
| 2450 | area = &(zone->free_area[current_order]);
|
| 2451 | fallback_mt = find_suitable_fallback(area, current_order,
|
| 2452 | start_migratetype, false);
|
| 2453 | if (fallback_mt == -1)
|
| 2454 | continue;
|
| 2455 |
|
| 2456 | page = get_page_from_free_area(area, fallback_mt);
|
| 2457 | page_del_and_expand(zone, page, order, current_order, fallback_mt);
|
| 2458 | trace_mm_page_alloc_extfrag(page, order, current_order,
|
| 2459 | start_migratetype, fallback_mt);
|
| 2460 | return page;
|
| 2461 | }
|
| 2462 |
|
| 2463 | return NULL;
|
| 2464 | }
|
| 2465 |
|
| 2466 | enum rmqueue_mode {
|
| 2467 | RMQUEUE_NORMAL,
|
| 2468 | RMQUEUE_CMA,
|
| 2469 | RMQUEUE_CLAIM,
|
| 2470 | RMQUEUE_STEAL,
|
| 2471 | };
|
| 2472 |
|
| 2473 | |
| 2474 | * Do the hard work of removing an element from the buddy allocator.
|
| 2475 | * Call me with the zone->lock already held.
|
| 2476 |
|
| 2477 | static __always_inline struct page *
|
| 2478 | __rmqueue(struct zone *zone, unsigned int order, int migratetype,
|
| 2479 | unsigned int alloc_flags, enum rmqueue_mode *mode)
|
| 2480 | {
|
| 2481 | struct page *page;
|
| 2482 |
|
| 2483 | if (IS_ENABLED(CONFIG_CMA)) {
|
| 2484 | |
| 2485 | * Balance movable allocations between regular and CMA areas by
|
| 2486 | * allocating from CMA when over half of the zone's free memory
|
| 2487 | * is in the CMA area.
|
| 2488 |
|
| 2489 | if (alloc_flags & ALLOC_CMA &&
|
| 2490 | zone_page_state(zone, NR_FREE_CMA_PAGES) >
|
| 2491 | zone_page_state(zone, NR_FREE_PAGES) / 2) {
|
| 2492 | page = __rmqueue_cma_fallback(zone, order);
|
| 2493 | if (page)
|
| 2494 | return page;
|
| 2495 | }
|
| 2496 | }
|
| 2497 |
|
| 2498 | |
| 2499 | * First try the freelists of the requested migratetype, then try
|
| 2500 | * fallbacks modes with increasing levels of fragmentation risk.
|
| 2501 | *
|
| 2502 | * The fallback logic is expensive and rmqueue_bulk() calls in
|
| 2503 | * a loop with the zone->lock held, meaning the freelists are
|
| 2504 | * not subject to any outside changes. Remember in *mode where
|
| 2505 | * we found pay dirt, to save us the search on the next call.
|
| 2506 |
|
| 2507 | switch (*mode) {
|
| 2508 | case RMQUEUE_NORMAL:
|
| 2509 | page = __rmqueue_smallest(zone, order, migratetype);
|
| 2510 | if (page)
|
| 2511 | return page;
|
| 2512 | fallthrough;
|
| 2513 | case RMQUEUE_CMA:
|
| 2514 | if (alloc_flags & ALLOC_CMA) {
|
| 2515 | page = __rmqueue_cma_fallback(zone, order);
|
| 2516 | if (page) {
|
| 2517 | *mode = RMQUEUE_CMA;
|
| 2518 | return page;
|
| 2519 | }
|
| 2520 | }
|
| 2521 | fallthrough;
|
| 2522 | case RMQUEUE_CLAIM:
|
| 2523 | page = __rmqueue_claim(zone, order, migratetype, alloc_flags);
|
| 2524 | if (page) {
|
| 2525 |
|
| 2526 | *mode = RMQUEUE_NORMAL;
|
| 2527 | return page;
|
| 2528 | }
|
| 2529 | fallthrough;
|
| 2530 | case RMQUEUE_STEAL:
|
| 2531 | if (!(alloc_flags & ALLOC_NOFRAGMENT)) {
|
| 2532 | page = __rmqueue_steal(zone, order, migratetype);
|
| 2533 | if (page) {
|
| 2534 | *mode = RMQUEUE_STEAL;
|
| 2535 | return page;
|
| 2536 | }
|
| 2537 | }
|
| 2538 | }
|
| 2539 | return NULL;
|
| 2540 | }
|
| 2541 |
|
| 2542 | |
| 2543 | * Obtain a specified number of elements from the buddy allocator, all under
|
| 2544 | * a single hold of the lock, for efficiency. Add them to the supplied list.
|
| 2545 | * Returns the number of new pages which were placed at *list.
|
| 2546 |
|
| 2547 | static int rmqueue_bulk(struct zone *zone, unsigned int order,
|
| 2548 | unsigned long count, struct list_head *list,
|
| 2549 | int migratetype, unsigned int alloc_flags)
|
| 2550 | {
|
| 2551 | enum rmqueue_mode rmqm = RMQUEUE_NORMAL;
|
| 2552 | unsigned long flags;
|
| 2553 | int i;
|
| 2554 |
|
| 2555 | if (unlikely(alloc_flags & ALLOC_TRYLOCK)) {
|
| 2556 | if (!spin_trylock_irqsave(&zone->lock, flags))
|
| 2557 | return 0;
|
| 2558 | } else {
|
| 2559 | spin_lock_irqsave(&zone->lock, flags);
|
| 2560 | }
|
| 2561 | for (i = 0; i < count; ++i) {
|
| 2562 | struct page *page = __rmqueue(zone, order, migratetype,
|
| 2563 | alloc_flags, &rmqm);
|
| 2564 | if (unlikely(page == NULL))
|
| 2565 | break;
|
| 2566 |
|
| 2567 | |
| 2568 | * Split buddy pages returned by expand() are received here in
|
| 2569 | * physical page order. The page is added to the tail of
|
| 2570 | * caller's list. From the callers perspective, the linked list
|
| 2571 | * is ordered by page number under some conditions. This is
|
| 2572 | * useful for IO devices that can forward direction from the
|
| 2573 | * head, thus also in the physical page order. This is useful
|
| 2574 | * for IO devices that can merge IO requests if the physical
|
| 2575 | * pages are ordered properly.
|
| 2576 |
|
| 2577 | list_add_tail(&page->pcp_list, list);
|
| 2578 | }
|
| 2579 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 2580 |
|
| 2581 | return i;
|
| 2582 | }
|
| 2583 |
|
| 2584 | |
| 2585 | * Called from the vmstat counter updater to decay the PCP high.
|
| 2586 | * Return whether there are addition works to do.
|
| 2587 |
|
| 2588 | bool decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
|
| 2589 | {
|
| 2590 | int high_min, to_drain, to_drain_batched, batch;
|
| 2591 | unsigned long UP_flags;
|
| 2592 | bool todo = false;
|
| 2593 |
|
| 2594 | high_min = READ_ONCE(pcp->high_min);
|
| 2595 | batch = READ_ONCE(pcp->batch);
|
| 2596 | |
| 2597 | * Decrease pcp->high periodically to try to free possible
|
| 2598 | * idle PCP pages. And, avoid to free too many pages to
|
| 2599 | * control latency. This caps pcp->high decrement too.
|
| 2600 |
|
| 2601 | if (pcp->high > high_min) {
|
| 2602 | pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
|
| 2603 | pcp->high - (pcp->high >> 3), high_min);
|
| 2604 | if (pcp->high > high_min)
|
| 2605 | todo = true;
|
| 2606 | }
|
| 2607 |
|
| 2608 | to_drain = pcp->count - pcp->high;
|
| 2609 | while (to_drain > 0) {
|
| 2610 | to_drain_batched = min(to_drain, batch);
|
| 2611 | pcp_spin_lock_maybe_irqsave(pcp, UP_flags);
|
| 2612 | free_pcppages_bulk(zone, to_drain_batched, pcp, 0);
|
| 2613 | pcp_spin_unlock_maybe_irqrestore(pcp, UP_flags);
|
| 2614 | todo = true;
|
| 2615 |
|
| 2616 | to_drain -= to_drain_batched;
|
| 2617 | }
|
| 2618 |
|
| 2619 | return todo;
|
| 2620 | }
|
| 2621 |
|
| 2622 | #ifdef CONFIG_NUMA
|
| 2623 | |
| 2624 | * Called from the vmstat counter updater to drain pagesets of this
|
| 2625 | * currently executing processor on remote nodes after they have
|
| 2626 | * expired.
|
| 2627 |
|
| 2628 | void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
|
| 2629 | {
|
| 2630 | unsigned long UP_flags;
|
| 2631 | int to_drain, batch;
|
| 2632 |
|
| 2633 | batch = READ_ONCE(pcp->batch);
|
| 2634 | to_drain = min(pcp->count, batch);
|
| 2635 | if (to_drain > 0) {
|
| 2636 | pcp_spin_lock_maybe_irqsave(pcp, UP_flags);
|
| 2637 | free_pcppages_bulk(zone, to_drain, pcp, 0);
|
| 2638 | pcp_spin_unlock_maybe_irqrestore(pcp, UP_flags);
|
| 2639 | }
|
| 2640 | }
|
| 2641 | #endif
|
| 2642 |
|
| 2643 | |
| 2644 | * Drain pcplists of the indicated processor and zone.
|
| 2645 |
|
| 2646 | static void drain_pages_zone(unsigned int cpu, struct zone *zone)
|
| 2647 | {
|
| 2648 | struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
|
| 2649 | unsigned long UP_flags;
|
| 2650 | int count;
|
| 2651 |
|
| 2652 | do {
|
| 2653 | pcp_spin_lock_maybe_irqsave(pcp, UP_flags);
|
| 2654 | count = pcp->count;
|
| 2655 | if (count) {
|
| 2656 | int to_drain = min(count,
|
| 2657 | pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
|
| 2658 |
|
| 2659 | free_pcppages_bulk(zone, to_drain, pcp, 0);
|
| 2660 | count -= to_drain;
|
| 2661 | }
|
| 2662 | pcp_spin_unlock_maybe_irqrestore(pcp, UP_flags);
|
| 2663 | } while (count);
|
| 2664 | }
|
| 2665 |
|
| 2666 | |
| 2667 | * Drain pcplists of all zones on the indicated processor.
|
| 2668 |
|
| 2669 | static void drain_pages(unsigned int cpu)
|
| 2670 | {
|
| 2671 | struct zone *zone;
|
| 2672 |
|
| 2673 | for_each_populated_zone(zone) {
|
| 2674 | drain_pages_zone(cpu, zone);
|
| 2675 | }
|
| 2676 | }
|
| 2677 |
|
| 2678 | |
| 2679 | * Spill all of this CPU's per-cpu pages back into the buddy allocator.
|
| 2680 |
|
| 2681 | void drain_local_pages(struct zone *zone)
|
| 2682 | {
|
| 2683 | int cpu = smp_processor_id();
|
| 2684 |
|
| 2685 | if (zone)
|
| 2686 | drain_pages_zone(cpu, zone);
|
| 2687 | else
|
| 2688 | drain_pages(cpu);
|
| 2689 | }
|
| 2690 |
|
| 2691 | |
| 2692 | * The implementation of drain_all_pages(), exposing an extra parameter to
|
| 2693 | * drain on all cpus.
|
| 2694 | *
|
| 2695 | * drain_all_pages() is optimized to only execute on cpus where pcplists are
|
| 2696 | * not empty. The check for non-emptiness can however race with a free to
|
| 2697 | * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
|
| 2698 | * that need the guarantee that every CPU has drained can disable the
|
| 2699 | * optimizing racy check.
|
| 2700 |
|
| 2701 | static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
|
| 2702 | {
|
| 2703 | int cpu;
|
| 2704 |
|
| 2705 | |
| 2706 | * Allocate in the BSS so we won't require allocation in
|
| 2707 | * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
|
| 2708 |
|
| 2709 | static cpumask_t cpus_with_pcps;
|
| 2710 |
|
| 2711 | |
| 2712 | * Do not drain if one is already in progress unless it's specific to
|
| 2713 | * a zone. Such callers are primarily CMA and memory hotplug and need
|
| 2714 | * the drain to be complete when the call returns.
|
| 2715 |
|
| 2716 | if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
|
| 2717 | if (!zone)
|
| 2718 | return;
|
| 2719 | mutex_lock(&pcpu_drain_mutex);
|
| 2720 | }
|
| 2721 |
|
| 2722 | |
| 2723 | * We don't care about racing with CPU hotplug event
|
| 2724 | * as offline notification will cause the notified
|
| 2725 | * cpu to drain that CPU pcps and on_each_cpu_mask
|
| 2726 | * disables preemption as part of its processing
|
| 2727 |
|
| 2728 | for_each_online_cpu(cpu) {
|
| 2729 | struct per_cpu_pages *pcp;
|
| 2730 | struct zone *z;
|
| 2731 | bool has_pcps = false;
|
| 2732 |
|
| 2733 | if (force_all_cpus) {
|
| 2734 | |
| 2735 | * The pcp.count check is racy, some callers need a
|
| 2736 | * guarantee that no cpu is missed.
|
| 2737 |
|
| 2738 | has_pcps = true;
|
| 2739 | } else if (zone) {
|
| 2740 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
|
| 2741 | if (pcp->count)
|
| 2742 | has_pcps = true;
|
| 2743 | } else {
|
| 2744 | for_each_populated_zone(z) {
|
| 2745 | pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
|
| 2746 | if (pcp->count) {
|
| 2747 | has_pcps = true;
|
| 2748 | break;
|
| 2749 | }
|
| 2750 | }
|
| 2751 | }
|
| 2752 |
|
| 2753 | if (has_pcps)
|
| 2754 | cpumask_set_cpu(cpu, &cpus_with_pcps);
|
| 2755 | else
|
| 2756 | cpumask_clear_cpu(cpu, &cpus_with_pcps);
|
| 2757 | }
|
| 2758 |
|
| 2759 | for_each_cpu(cpu, &cpus_with_pcps) {
|
| 2760 | if (zone)
|
| 2761 | drain_pages_zone(cpu, zone);
|
| 2762 | else
|
| 2763 | drain_pages(cpu);
|
| 2764 | }
|
| 2765 |
|
| 2766 | mutex_unlock(&pcpu_drain_mutex);
|
| 2767 | }
|
| 2768 |
|
| 2769 | |
| 2770 | * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
|
| 2771 | *
|
| 2772 | * When zone parameter is non-NULL, spill just the single zone's pages.
|
| 2773 |
|
| 2774 | void drain_all_pages(struct zone *zone)
|
| 2775 | {
|
| 2776 | __drain_all_pages(zone, false);
|
| 2777 | }
|
| 2778 |
|
| 2779 | static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
|
| 2780 | {
|
| 2781 | int min_nr_free, max_nr_free;
|
| 2782 |
|
| 2783 |
|
| 2784 | if (unlikely(free_high))
|
| 2785 | return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
|
| 2786 |
|
| 2787 |
|
| 2788 | if (unlikely(high < batch))
|
| 2789 | return 1;
|
| 2790 |
|
| 2791 |
|
| 2792 | min_nr_free = batch;
|
| 2793 | max_nr_free = high - batch;
|
| 2794 |
|
| 2795 | |
| 2796 | * Increase the batch number to the number of the consecutive
|
| 2797 | * freed pages to reduce zone lock contention.
|
| 2798 |
|
| 2799 | batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
|
| 2800 |
|
| 2801 | return batch;
|
| 2802 | }
|
| 2803 |
|
| 2804 | static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
|
| 2805 | int batch, bool free_high)
|
| 2806 | {
|
| 2807 | int high, high_min, high_max;
|
| 2808 |
|
| 2809 | high_min = READ_ONCE(pcp->high_min);
|
| 2810 | high_max = READ_ONCE(pcp->high_max);
|
| 2811 | high = pcp->high = clamp(pcp->high, high_min, high_max);
|
| 2812 |
|
| 2813 | if (unlikely(!high))
|
| 2814 | return 0;
|
| 2815 |
|
| 2816 | if (unlikely(free_high)) {
|
| 2817 | pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
|
| 2818 | high_min);
|
| 2819 | return 0;
|
| 2820 | }
|
| 2821 |
|
| 2822 | |
| 2823 | * If reclaim is active, limit the number of pages that can be
|
| 2824 | * stored on pcp lists
|
| 2825 |
|
| 2826 | if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
|
| 2827 | int free_count = max_t(int, pcp->free_count, batch);
|
| 2828 |
|
| 2829 | pcp->high = max(high - free_count, high_min);
|
| 2830 | return min(batch << 2, pcp->high);
|
| 2831 | }
|
| 2832 |
|
| 2833 | if (high_min == high_max)
|
| 2834 | return high;
|
| 2835 |
|
| 2836 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
|
| 2837 | int free_count = max_t(int, pcp->free_count, batch);
|
| 2838 |
|
| 2839 | pcp->high = max(high - free_count, high_min);
|
| 2840 | high = max(pcp->count, high_min);
|
| 2841 | } else if (pcp->count >= high) {
|
| 2842 | int need_high = pcp->free_count + batch;
|
| 2843 |
|
| 2844 |
|
| 2845 | if (pcp->high < need_high)
|
| 2846 | pcp->high = clamp(need_high, high_min, high_max);
|
| 2847 | }
|
| 2848 |
|
| 2849 | return high;
|
| 2850 | }
|
| 2851 |
|
| 2852 | |
| 2853 | * Tune pcp alloc factor and adjust count & free_count. Free pages to bring the
|
| 2854 | * pcp's watermarks below high.
|
| 2855 | *
|
| 2856 | * May return a freed pcp, if during page freeing the pcp spinlock cannot be
|
| 2857 | * reacquired. Return true if pcp is locked, false otherwise.
|
| 2858 |
|
| 2859 | static bool free_frozen_page_commit(struct zone *zone,
|
| 2860 | struct per_cpu_pages *pcp, struct page *page, int migratetype,
|
| 2861 | unsigned int order, fpi_t fpi_flags, unsigned long *UP_flags)
|
| 2862 | {
|
| 2863 | int high, batch;
|
| 2864 | int to_free, to_free_batched;
|
| 2865 | int pindex;
|
| 2866 | int cpu = smp_processor_id();
|
| 2867 | int ret = true;
|
| 2868 | bool free_high = false;
|
| 2869 |
|
| 2870 | |
| 2871 | * On freeing, reduce the number of pages that are batch allocated.
|
| 2872 | * See nr_pcp_alloc() where alloc_factor is increased for subsequent
|
| 2873 | * allocations.
|
| 2874 |
|
| 2875 | pcp->alloc_factor >>= 1;
|
| 2876 | __count_vm_events(PGFREE, 1 << order);
|
| 2877 | pindex = order_to_pindex(migratetype, order);
|
| 2878 | list_add(&page->pcp_list, &pcp->lists[pindex]);
|
| 2879 | pcp->count += 1 << order;
|
| 2880 |
|
| 2881 | batch = READ_ONCE(pcp->batch);
|
| 2882 | |
| 2883 | * As high-order pages other than THP's stored on PCP can contribute
|
| 2884 | * to fragmentation, limit the number stored when PCP is heavily
|
| 2885 | * freeing without allocation. The remainder after bulk freeing
|
| 2886 | * stops will be drained from vmstat refresh context.
|
| 2887 |
|
| 2888 | if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
|
| 2889 | free_high = (pcp->free_count >= (batch + pcp->high_min / 2) &&
|
| 2890 | (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
|
| 2891 | (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
|
| 2892 | pcp->count >= batch));
|
| 2893 | pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
|
| 2894 | } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
|
| 2895 | pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
|
| 2896 | }
|
| 2897 | if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
|
| 2898 | pcp->free_count += (1 << order);
|
| 2899 |
|
| 2900 | if (unlikely(fpi_flags & FPI_TRYLOCK)) {
|
| 2901 | |
| 2902 | * Do not attempt to take a zone lock. Let pcp->count get
|
| 2903 | * over high mark temporarily.
|
| 2904 |
|
| 2905 | return true;
|
| 2906 | }
|
| 2907 |
|
| 2908 | high = nr_pcp_high(pcp, zone, batch, free_high);
|
| 2909 | if (pcp->count < high)
|
| 2910 | return true;
|
| 2911 |
|
| 2912 | to_free = nr_pcp_free(pcp, batch, high, free_high);
|
| 2913 | while (to_free > 0 && pcp->count > 0) {
|
| 2914 | to_free_batched = min(to_free, batch);
|
| 2915 | free_pcppages_bulk(zone, to_free_batched, pcp, pindex);
|
| 2916 | to_free -= to_free_batched;
|
| 2917 |
|
| 2918 | if (to_free == 0 || pcp->count == 0)
|
| 2919 | break;
|
| 2920 |
|
| 2921 | pcp_spin_unlock(pcp, *UP_flags);
|
| 2922 |
|
| 2923 | pcp = pcp_spin_trylock(zone->per_cpu_pageset, *UP_flags);
|
| 2924 | if (!pcp) {
|
| 2925 | ret = false;
|
| 2926 | break;
|
| 2927 | }
|
| 2928 |
|
| 2929 | |
| 2930 | * Check if this thread has been migrated to a different CPU.
|
| 2931 | * If that is the case, give up and indicate that the pcp is
|
| 2932 | * returned in an unlocked state.
|
| 2933 |
|
| 2934 | if (smp_processor_id() != cpu) {
|
| 2935 | pcp_spin_unlock(pcp, *UP_flags);
|
| 2936 | ret = false;
|
| 2937 | break;
|
| 2938 | }
|
| 2939 | }
|
| 2940 |
|
| 2941 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
|
| 2942 | zone_watermark_ok(zone, 0, high_wmark_pages(zone),
|
| 2943 | ZONE_MOVABLE, 0)) {
|
| 2944 | struct pglist_data *pgdat = zone->zone_pgdat;
|
| 2945 | clear_bit(ZONE_BELOW_HIGH, &zone->flags);
|
| 2946 |
|
| 2947 | |
| 2948 | * Assume that memory pressure on this node is gone and may be
|
| 2949 | * in a reclaimable state. If a memory fallback node exists,
|
| 2950 | * direct reclaim may not have been triggered, causing a
|
| 2951 | * 'hopeless node' to stay in that state for a while. Let
|
| 2952 | * kswapd work again by resetting kswapd_failures.
|
| 2953 |
|
| 2954 | if (kswapd_test_hopeless(pgdat) &&
|
| 2955 | next_memory_node(pgdat->node_id) < MAX_NUMNODES)
|
| 2956 | kswapd_clear_hopeless(pgdat, KSWAPD_CLEAR_HOPELESS_PCP);
|
| 2957 | }
|
| 2958 | return ret;
|
| 2959 | }
|
| 2960 |
|
| 2961 | |
| 2962 | * Free a pcp page
|
| 2963 |
|
| 2964 | static void __free_frozen_pages(struct page *page, unsigned int order,
|
| 2965 | fpi_t fpi_flags)
|
| 2966 | {
|
| 2967 | unsigned long UP_flags;
|
| 2968 | struct per_cpu_pages *pcp;
|
| 2969 | struct zone *zone;
|
| 2970 | unsigned long pfn = page_to_pfn(page);
|
| 2971 | int migratetype;
|
| 2972 |
|
| 2973 | if (!pcp_allowed_order(order)) {
|
| 2974 | __free_pages_ok(page, order, fpi_flags);
|
| 2975 | return;
|
| 2976 | }
|
| 2977 |
|
| 2978 | if (!__free_pages_prepare(page, order, fpi_flags))
|
| 2979 | return;
|
| 2980 |
|
| 2981 | |
| 2982 | * We only track unmovable, reclaimable and movable on pcp lists.
|
| 2983 | * Place ISOLATE pages on the isolated list because they are being
|
| 2984 | * offlined but treat HIGHATOMIC and CMA as movable pages so we can
|
| 2985 | * get those areas back if necessary. Otherwise, we may have to free
|
| 2986 | * excessively into the page allocator
|
| 2987 |
|
| 2988 | zone = page_zone(page);
|
| 2989 | migratetype = get_pfnblock_migratetype(page, pfn);
|
| 2990 | if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
|
| 2991 | if (unlikely(is_migrate_isolate(migratetype))) {
|
| 2992 | free_one_page(zone, page, pfn, order, fpi_flags);
|
| 2993 | return;
|
| 2994 | }
|
| 2995 | migratetype = MIGRATE_MOVABLE;
|
| 2996 | }
|
| 2997 |
|
| 2998 | if (unlikely((fpi_flags & FPI_TRYLOCK) && IS_ENABLED(CONFIG_PREEMPT_RT)
|
| 2999 | && (in_nmi() || in_hardirq()))) {
|
| 3000 | add_page_to_zone_llist(zone, page, order);
|
| 3001 | return;
|
| 3002 | }
|
| 3003 | pcp = pcp_spin_trylock(zone->per_cpu_pageset, UP_flags);
|
| 3004 | if (pcp) {
|
| 3005 | if (!free_frozen_page_commit(zone, pcp, page, migratetype,
|
| 3006 | order, fpi_flags, &UP_flags))
|
| 3007 | return;
|
| 3008 | pcp_spin_unlock(pcp, UP_flags);
|
| 3009 | } else {
|
| 3010 | free_one_page(zone, page, pfn, order, fpi_flags);
|
| 3011 | }
|
| 3012 | }
|
| 3013 |
|
| 3014 | void free_frozen_pages(struct page *page, unsigned int order)
|
| 3015 | {
|
| 3016 | __free_frozen_pages(page, order, FPI_NONE);
|
| 3017 | }
|
| 3018 |
|
| 3019 | void free_frozen_pages_nolock(struct page *page, unsigned int order)
|
| 3020 | {
|
| 3021 | __free_frozen_pages(page, order, FPI_TRYLOCK);
|
| 3022 | }
|
| 3023 |
|
| 3024 | |
| 3025 | * Free a batch of folios
|
| 3026 |
|
| 3027 | void free_unref_folios(struct folio_batch *folios)
|
| 3028 | {
|
| 3029 | unsigned long UP_flags;
|
| 3030 | struct per_cpu_pages *pcp = NULL;
|
| 3031 | struct zone *locked_zone = NULL;
|
| 3032 | int i, j;
|
| 3033 |
|
| 3034 |
|
| 3035 | for (i = 0, j = 0; i < folios->nr; i++) {
|
| 3036 | struct folio *folio = folios->folios[i];
|
| 3037 | unsigned long pfn = folio_pfn(folio);
|
| 3038 | unsigned int order = folio_order(folio);
|
| 3039 |
|
| 3040 | if (!__free_pages_prepare(&folio->page, order, FPI_NONE))
|
| 3041 | continue;
|
| 3042 | |
| 3043 | * Free orders not handled on the PCP directly to the
|
| 3044 | * allocator.
|
| 3045 |
|
| 3046 | if (!pcp_allowed_order(order)) {
|
| 3047 | free_one_page(folio_zone(folio), &folio->page,
|
| 3048 | pfn, order, FPI_NONE);
|
| 3049 | continue;
|
| 3050 | }
|
| 3051 | folio->private = (void *)(unsigned long)order;
|
| 3052 | if (j != i)
|
| 3053 | folios->folios[j] = folio;
|
| 3054 | j++;
|
| 3055 | }
|
| 3056 | folios->nr = j;
|
| 3057 |
|
| 3058 | for (i = 0; i < folios->nr; i++) {
|
| 3059 | struct folio *folio = folios->folios[i];
|
| 3060 | struct zone *zone = folio_zone(folio);
|
| 3061 | unsigned long pfn = folio_pfn(folio);
|
| 3062 | unsigned int order = (unsigned long)folio->private;
|
| 3063 | int migratetype;
|
| 3064 |
|
| 3065 | folio->private = NULL;
|
| 3066 | migratetype = get_pfnblock_migratetype(&folio->page, pfn);
|
| 3067 |
|
| 3068 |
|
| 3069 | if (zone != locked_zone ||
|
| 3070 | is_migrate_isolate(migratetype)) {
|
| 3071 | if (pcp) {
|
| 3072 | pcp_spin_unlock(pcp, UP_flags);
|
| 3073 | locked_zone = NULL;
|
| 3074 | pcp = NULL;
|
| 3075 | }
|
| 3076 |
|
| 3077 | |
| 3078 | * Free isolated pages directly to the
|
| 3079 | * allocator, see comment in free_frozen_pages.
|
| 3080 |
|
| 3081 | if (is_migrate_isolate(migratetype)) {
|
| 3082 | free_one_page(zone, &folio->page, pfn,
|
| 3083 | order, FPI_NONE);
|
| 3084 | continue;
|
| 3085 | }
|
| 3086 |
|
| 3087 | |
| 3088 | * trylock is necessary as folios may be getting freed
|
| 3089 | * from IRQ or SoftIRQ context after an IO completion.
|
| 3090 |
|
| 3091 | pcp = pcp_spin_trylock(zone->per_cpu_pageset, UP_flags);
|
| 3092 | if (unlikely(!pcp)) {
|
| 3093 | free_one_page(zone, &folio->page, pfn,
|
| 3094 | order, FPI_NONE);
|
| 3095 | continue;
|
| 3096 | }
|
| 3097 | locked_zone = zone;
|
| 3098 | }
|
| 3099 |
|
| 3100 | |
| 3101 | * Non-isolated types over MIGRATE_PCPTYPES get added
|
| 3102 | * to the MIGRATE_MOVABLE pcp list.
|
| 3103 |
|
| 3104 | if (unlikely(migratetype >= MIGRATE_PCPTYPES))
|
| 3105 | migratetype = MIGRATE_MOVABLE;
|
| 3106 |
|
| 3107 | trace_mm_page_free_batched(&folio->page);
|
| 3108 | if (!free_frozen_page_commit(zone, pcp, &folio->page,
|
| 3109 | migratetype, order, FPI_NONE, &UP_flags)) {
|
| 3110 | pcp = NULL;
|
| 3111 | locked_zone = NULL;
|
| 3112 | }
|
| 3113 | }
|
| 3114 |
|
| 3115 | if (pcp)
|
| 3116 | pcp_spin_unlock(pcp, UP_flags);
|
| 3117 | folio_batch_reinit(folios);
|
| 3118 | }
|
| 3119 |
|
| 3120 | static void __split_page(struct page *page, unsigned int order)
|
| 3121 | {
|
| 3122 | VM_WARN_ON_PAGE(PageCompound(page), page);
|
| 3123 |
|
| 3124 | split_page_owner(page, order, 0);
|
| 3125 | pgalloc_tag_split(page_folio(page), order, 0);
|
| 3126 | split_page_memcg(page, order);
|
| 3127 | }
|
| 3128 |
|
| 3129 | |
| 3130 | * split_page takes a non-compound higher-order page, and splits it into
|
| 3131 | * n (1<<order) sub-pages: page[0..n]
|
| 3132 | * Each sub-page must be freed individually.
|
| 3133 | *
|
| 3134 | * Note: this is probably too low level an operation for use in drivers.
|
| 3135 | * Please consult with lkml before using this in your driver.
|
| 3136 |
|
| 3137 | void split_page(struct page *page, unsigned int order)
|
| 3138 | {
|
| 3139 | int i;
|
| 3140 |
|
| 3141 | VM_WARN_ON_PAGE(!page_count(page), page);
|
| 3142 |
|
| 3143 | for (i = 1; i < (1 << order); i++)
|
| 3144 | set_page_refcounted(page + i);
|
| 3145 |
|
| 3146 | __split_page(page, order);
|
| 3147 | }
|
| 3148 | EXPORT_SYMBOL_GPL(split_page);
|
| 3149 |
|
| 3150 | int __isolate_free_page(struct page *page, unsigned int order)
|
| 3151 | {
|
| 3152 | struct zone *zone = page_zone(page);
|
| 3153 | int mt = get_pageblock_migratetype(page);
|
| 3154 |
|
| 3155 | if (!is_migrate_isolate(mt)) {
|
| 3156 | unsigned long watermark;
|
| 3157 | |
| 3158 | * Obey watermarks as if the page was being allocated. We can
|
| 3159 | * emulate a high-order watermark check with a raised order-0
|
| 3160 | * watermark, because we already know our high-order page
|
| 3161 | * exists.
|
| 3162 |
|
| 3163 | watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
|
| 3164 | if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
|
| 3165 | return 0;
|
| 3166 | }
|
| 3167 |
|
| 3168 | del_page_from_free_list(page, zone, order, mt);
|
| 3169 |
|
| 3170 | |
| 3171 | * Set the pageblock if the isolated page is at least half of a
|
| 3172 | * pageblock
|
| 3173 |
|
| 3174 | if (order >= pageblock_order - 1) {
|
| 3175 | struct page *endpage = page + (1 << order) - 1;
|
| 3176 | for (; page < endpage; page += pageblock_nr_pages) {
|
| 3177 | int mt = get_pageblock_migratetype(page);
|
| 3178 | |
| 3179 | * Only change normal pageblocks (i.e., they can merge
|
| 3180 | * with others)
|
| 3181 |
|
| 3182 | if (migratetype_is_mergeable(mt))
|
| 3183 | move_freepages_block(zone, page, mt,
|
| 3184 | MIGRATE_MOVABLE);
|
| 3185 | }
|
| 3186 | }
|
| 3187 |
|
| 3188 | return 1UL << order;
|
| 3189 | }
|
| 3190 |
|
| 3191 | |
| 3192 | * __putback_isolated_page - Return a now-isolated page back where we got it
|
| 3193 | * @page: Page that was isolated
|
| 3194 | * @order: Order of the isolated page
|
| 3195 | * @mt: The page's pageblock's migratetype
|
| 3196 | *
|
| 3197 | * This function is meant to return a page pulled from the free lists via
|
| 3198 | * __isolate_free_page back to the free lists they were pulled from.
|
| 3199 |
|
| 3200 | void __putback_isolated_page(struct page *page, unsigned int order, int mt)
|
| 3201 | {
|
| 3202 | struct zone *zone = page_zone(page);
|
| 3203 |
|
| 3204 |
|
| 3205 | lockdep_assert_held(&zone->lock);
|
| 3206 |
|
| 3207 |
|
| 3208 | __free_one_page(page, page_to_pfn(page), zone, order, mt,
|
| 3209 | FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
|
| 3210 | }
|
| 3211 |
|
| 3212 | |
| 3213 | * Update NUMA hit/miss statistics
|
| 3214 |
|
| 3215 | static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
|
| 3216 | long nr_account)
|
| 3217 | {
|
| 3218 | #ifdef CONFIG_NUMA
|
| 3219 | enum numa_stat_item local_stat = NUMA_LOCAL;
|
| 3220 |
|
| 3221 |
|
| 3222 | if (!static_branch_likely(&vm_numa_stat_key))
|
| 3223 | return;
|
| 3224 |
|
| 3225 | if (zone_to_nid(z) != numa_node_id())
|
| 3226 | local_stat = NUMA_OTHER;
|
| 3227 |
|
| 3228 | if (zone_to_nid(z) == zone_to_nid(preferred_zone))
|
| 3229 | __count_numa_events(z, NUMA_HIT, nr_account);
|
| 3230 | else {
|
| 3231 | __count_numa_events(z, NUMA_MISS, nr_account);
|
| 3232 | __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
|
| 3233 | }
|
| 3234 | __count_numa_events(z, local_stat, nr_account);
|
| 3235 | #endif
|
| 3236 | }
|
| 3237 |
|
| 3238 | static __always_inline
|
| 3239 | struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
|
| 3240 | unsigned int order, unsigned int alloc_flags,
|
| 3241 | int migratetype)
|
| 3242 | {
|
| 3243 | struct page *page;
|
| 3244 | unsigned long flags;
|
| 3245 |
|
| 3246 | do {
|
| 3247 | page = NULL;
|
| 3248 | if (unlikely(alloc_flags & ALLOC_TRYLOCK)) {
|
| 3249 | if (!spin_trylock_irqsave(&zone->lock, flags))
|
| 3250 | return NULL;
|
| 3251 | } else {
|
| 3252 | spin_lock_irqsave(&zone->lock, flags);
|
| 3253 | }
|
| 3254 | if (alloc_flags & ALLOC_HIGHATOMIC)
|
| 3255 | page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
|
| 3256 | if (!page) {
|
| 3257 | enum rmqueue_mode rmqm = RMQUEUE_NORMAL;
|
| 3258 |
|
| 3259 | page = __rmqueue(zone, order, migratetype, alloc_flags, &rmqm);
|
| 3260 |
|
| 3261 | |
| 3262 | * If the allocation fails, allow OOM handling and
|
| 3263 | * order-0 (atomic) allocs access to HIGHATOMIC
|
| 3264 | * reserves as failing now is worse than failing a
|
| 3265 | * high-order atomic allocation in the future.
|
| 3266 |
|
| 3267 | if (!page && (alloc_flags & (ALLOC_OOM|ALLOC_NON_BLOCK)))
|
| 3268 | page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
|
| 3269 |
|
| 3270 | if (!page) {
|
| 3271 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 3272 | return NULL;
|
| 3273 | }
|
| 3274 | }
|
| 3275 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 3276 | } while (check_new_pages(page, order));
|
| 3277 |
|
| 3278 | __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
|
| 3279 | zone_statistics(preferred_zone, zone, 1);
|
| 3280 |
|
| 3281 | return page;
|
| 3282 | }
|
| 3283 |
|
| 3284 | static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
|
| 3285 | {
|
| 3286 | int high, base_batch, batch, max_nr_alloc;
|
| 3287 | int high_max, high_min;
|
| 3288 |
|
| 3289 | base_batch = READ_ONCE(pcp->batch);
|
| 3290 | high_min = READ_ONCE(pcp->high_min);
|
| 3291 | high_max = READ_ONCE(pcp->high_max);
|
| 3292 | high = pcp->high = clamp(pcp->high, high_min, high_max);
|
| 3293 |
|
| 3294 |
|
| 3295 | if (unlikely(high < base_batch))
|
| 3296 | return 1;
|
| 3297 |
|
| 3298 | if (order)
|
| 3299 | batch = base_batch;
|
| 3300 | else
|
| 3301 | batch = (base_batch << pcp->alloc_factor);
|
| 3302 |
|
| 3303 | |
| 3304 | * If we had larger pcp->high, we could avoid to allocate from
|
| 3305 | * zone.
|
| 3306 |
|
| 3307 | if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
|
| 3308 | high = pcp->high = min(high + batch, high_max);
|
| 3309 |
|
| 3310 | if (!order) {
|
| 3311 | max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
|
| 3312 | |
| 3313 | * Double the number of pages allocated each time there is
|
| 3314 | * subsequent allocation of order-0 pages without any freeing.
|
| 3315 |
|
| 3316 | if (batch <= max_nr_alloc &&
|
| 3317 | pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
|
| 3318 | pcp->alloc_factor++;
|
| 3319 | batch = min(batch, max_nr_alloc);
|
| 3320 | }
|
| 3321 |
|
| 3322 | |
| 3323 | * Scale batch relative to order if batch implies free pages
|
| 3324 | * can be stored on the PCP. Batch can be 1 for small zones or
|
| 3325 | * for boot pagesets which should never store free pages as
|
| 3326 | * the pages may belong to arbitrary zones.
|
| 3327 |
|
| 3328 | if (batch > 1)
|
| 3329 | batch = max(batch >> order, 2);
|
| 3330 |
|
| 3331 | return batch;
|
| 3332 | }
|
| 3333 |
|
| 3334 |
|
| 3335 | static inline
|
| 3336 | struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
|
| 3337 | int migratetype,
|
| 3338 | unsigned int alloc_flags,
|
| 3339 | struct per_cpu_pages *pcp,
|
| 3340 | struct list_head *list)
|
| 3341 | {
|
| 3342 | struct page *page;
|
| 3343 |
|
| 3344 | do {
|
| 3345 | if (list_empty(list)) {
|
| 3346 | int batch = nr_pcp_alloc(pcp, zone, order);
|
| 3347 | int alloced;
|
| 3348 |
|
| 3349 | alloced = rmqueue_bulk(zone, order,
|
| 3350 | batch, list,
|
| 3351 | migratetype, alloc_flags);
|
| 3352 |
|
| 3353 | pcp->count += alloced << order;
|
| 3354 | if (unlikely(list_empty(list)))
|
| 3355 | return NULL;
|
| 3356 | }
|
| 3357 |
|
| 3358 | page = list_first_entry(list, struct page, pcp_list);
|
| 3359 | list_del(&page->pcp_list);
|
| 3360 | pcp->count -= 1 << order;
|
| 3361 | } while (check_new_pages(page, order));
|
| 3362 |
|
| 3363 | return page;
|
| 3364 | }
|
| 3365 |
|
| 3366 |
|
| 3367 | static struct page *rmqueue_pcplist(struct zone *preferred_zone,
|
| 3368 | struct zone *zone, unsigned int order,
|
| 3369 | int migratetype, unsigned int alloc_flags)
|
| 3370 | {
|
| 3371 | struct per_cpu_pages *pcp;
|
| 3372 | struct list_head *list;
|
| 3373 | struct page *page;
|
| 3374 | unsigned long UP_flags;
|
| 3375 |
|
| 3376 |
|
| 3377 | pcp = pcp_spin_trylock(zone->per_cpu_pageset, UP_flags);
|
| 3378 | if (!pcp)
|
| 3379 | return NULL;
|
| 3380 |
|
| 3381 | |
| 3382 | * On allocation, reduce the number of pages that are batch freed.
|
| 3383 | * See nr_pcp_free() where free_factor is increased for subsequent
|
| 3384 | * frees.
|
| 3385 |
|
| 3386 | pcp->free_count >>= 1;
|
| 3387 | list = &pcp->lists[order_to_pindex(migratetype, order)];
|
| 3388 | page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
|
| 3389 | pcp_spin_unlock(pcp, UP_flags);
|
| 3390 | if (page) {
|
| 3391 | __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
|
| 3392 | zone_statistics(preferred_zone, zone, 1);
|
| 3393 | }
|
| 3394 | return page;
|
| 3395 | }
|
| 3396 |
|
| 3397 | |
| 3398 | * Allocate a page from the given zone.
|
| 3399 | * Use pcplists for THP or "cheap" high-order allocations.
|
| 3400 |
|
| 3401 |
|
| 3402 | |
| 3403 | * Do not instrument rmqueue() with KMSAN. This function may call
|
| 3404 | * __msan_poison_alloca() through a call to set_pfnblock_migratetype().
|
| 3405 | * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
|
| 3406 | * may call rmqueue() again, which will result in a deadlock.
|
| 3407 |
|
| 3408 | __no_sanitize_memory
|
| 3409 | static inline
|
| 3410 | struct page *rmqueue(struct zone *preferred_zone,
|
| 3411 | struct zone *zone, unsigned int order,
|
| 3412 | gfp_t gfp_flags, unsigned int alloc_flags,
|
| 3413 | int migratetype)
|
| 3414 | {
|
| 3415 | struct page *page;
|
| 3416 |
|
| 3417 | if (likely(pcp_allowed_order(order))) {
|
| 3418 | page = rmqueue_pcplist(preferred_zone, zone, order,
|
| 3419 | migratetype, alloc_flags);
|
| 3420 | if (likely(page))
|
| 3421 | goto out;
|
| 3422 | }
|
| 3423 |
|
| 3424 | page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
|
| 3425 | migratetype);
|
| 3426 |
|
| 3427 | out:
|
| 3428 |
|
| 3429 | if ((alloc_flags & ALLOC_KSWAPD) &&
|
| 3430 | unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
|
| 3431 | clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
|
| 3432 | wakeup_kswapd(zone, 0, 0, zone_idx(zone));
|
| 3433 | }
|
| 3434 |
|
| 3435 | VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
|
| 3436 | return page;
|
| 3437 | }
|
| 3438 |
|
| 3439 | |
| 3440 | * Reserve the pageblock(s) surrounding an allocation request for
|
| 3441 | * exclusive use of high-order atomic allocations if there are no
|
| 3442 | * empty page blocks that contain a page with a suitable order
|
| 3443 |
|
| 3444 | static void reserve_highatomic_pageblock(struct page *page, int order,
|
| 3445 | struct zone *zone)
|
| 3446 | {
|
| 3447 | int mt;
|
| 3448 | unsigned long max_managed, flags;
|
| 3449 |
|
| 3450 | |
| 3451 | * The number reserved as: minimum is 1 pageblock, maximum is
|
| 3452 | * roughly 1% of a zone. But if 1% of a zone falls below a
|
| 3453 | * pageblock size, then don't reserve any pageblocks.
|
| 3454 | * Check is race-prone but harmless.
|
| 3455 |
|
| 3456 | if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
|
| 3457 | return;
|
| 3458 | max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
|
| 3459 | if (zone->nr_reserved_highatomic >= max_managed)
|
| 3460 | return;
|
| 3461 |
|
| 3462 | spin_lock_irqsave(&zone->lock, flags);
|
| 3463 |
|
| 3464 |
|
| 3465 | if (zone->nr_reserved_highatomic >= max_managed)
|
| 3466 | goto out_unlock;
|
| 3467 |
|
| 3468 |
|
| 3469 | mt = get_pageblock_migratetype(page);
|
| 3470 |
|
| 3471 | if (!migratetype_is_mergeable(mt))
|
| 3472 | goto out_unlock;
|
| 3473 |
|
| 3474 | if (order < pageblock_order) {
|
| 3475 | if (move_freepages_block(zone, page, mt, MIGRATE_HIGHATOMIC) == -1)
|
| 3476 | goto out_unlock;
|
| 3477 | zone->nr_reserved_highatomic += pageblock_nr_pages;
|
| 3478 | } else {
|
| 3479 | change_pageblock_range(page, order, MIGRATE_HIGHATOMIC);
|
| 3480 | zone->nr_reserved_highatomic += 1 << order;
|
| 3481 | }
|
| 3482 |
|
| 3483 | out_unlock:
|
| 3484 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 3485 | }
|
| 3486 |
|
| 3487 | |
| 3488 | * Used when an allocation is about to fail under memory pressure. This
|
| 3489 | * potentially hurts the reliability of high-order allocations when under
|
| 3490 | * intense memory pressure but failed atomic allocations should be easier
|
| 3491 | * to recover from than an OOM.
|
| 3492 | *
|
| 3493 | * If @force is true, try to unreserve pageblocks even though highatomic
|
| 3494 | * pageblock is exhausted.
|
| 3495 |
|
| 3496 | static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
|
| 3497 | bool force)
|
| 3498 | {
|
| 3499 | struct zonelist *zonelist = ac->zonelist;
|
| 3500 | unsigned long flags;
|
| 3501 | struct zoneref *z;
|
| 3502 | struct zone *zone;
|
| 3503 | struct page *page;
|
| 3504 | int order;
|
| 3505 | int ret;
|
| 3506 |
|
| 3507 | for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
|
| 3508 | ac->nodemask) {
|
| 3509 | |
| 3510 | * Preserve at least one pageblock unless memory pressure
|
| 3511 | * is really high.
|
| 3512 |
|
| 3513 | if (!force && zone->nr_reserved_highatomic <=
|
| 3514 | pageblock_nr_pages)
|
| 3515 | continue;
|
| 3516 |
|
| 3517 | spin_lock_irqsave(&zone->lock, flags);
|
| 3518 | for (order = 0; order < NR_PAGE_ORDERS; order++) {
|
| 3519 | struct free_area *area = &(zone->free_area[order]);
|
| 3520 | unsigned long size;
|
| 3521 |
|
| 3522 | page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
|
| 3523 | if (!page)
|
| 3524 | continue;
|
| 3525 |
|
| 3526 | size = max(pageblock_nr_pages, 1UL << order);
|
| 3527 | |
| 3528 | * It should never happen but changes to
|
| 3529 | * locking could inadvertently allow a per-cpu
|
| 3530 | * drain to add pages to MIGRATE_HIGHATOMIC
|
| 3531 | * while unreserving so be safe and watch for
|
| 3532 | * underflows.
|
| 3533 |
|
| 3534 | if (WARN_ON_ONCE(size > zone->nr_reserved_highatomic))
|
| 3535 | size = zone->nr_reserved_highatomic;
|
| 3536 | zone->nr_reserved_highatomic -= size;
|
| 3537 |
|
| 3538 | |
| 3539 | * Convert to ac->migratetype and avoid the normal
|
| 3540 | * pageblock stealing heuristics. Minimally, the caller
|
| 3541 | * is doing the work and needs the pages. More
|
| 3542 | * importantly, if the block was always converted to
|
| 3543 | * MIGRATE_UNMOVABLE or another type then the number
|
| 3544 | * of pageblocks that cannot be completely freed
|
| 3545 | * may increase.
|
| 3546 |
|
| 3547 | if (order < pageblock_order)
|
| 3548 | ret = move_freepages_block(zone, page,
|
| 3549 | MIGRATE_HIGHATOMIC,
|
| 3550 | ac->migratetype);
|
| 3551 | else {
|
| 3552 | move_to_free_list(page, zone, order,
|
| 3553 | MIGRATE_HIGHATOMIC,
|
| 3554 | ac->migratetype);
|
| 3555 | change_pageblock_range(page, order,
|
| 3556 | ac->migratetype);
|
| 3557 | ret = 1;
|
| 3558 | }
|
| 3559 | |
| 3560 | * Reserving the block(s) already succeeded,
|
| 3561 | * so this should not fail on zone boundaries.
|
| 3562 |
|
| 3563 | WARN_ON_ONCE(ret == -1);
|
| 3564 | if (ret > 0) {
|
| 3565 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 3566 | return ret;
|
| 3567 | }
|
| 3568 | }
|
| 3569 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 3570 | }
|
| 3571 |
|
| 3572 | return false;
|
| 3573 | }
|
| 3574 |
|
| 3575 | static inline long __zone_watermark_unusable_free(struct zone *z,
|
| 3576 | unsigned int order, unsigned int alloc_flags)
|
| 3577 | {
|
| 3578 | long unusable_free = (1 << order) - 1;
|
| 3579 |
|
| 3580 | |
| 3581 | * If the caller does not have rights to reserves below the min
|
| 3582 | * watermark then subtract the free pages reserved for highatomic.
|
| 3583 |
|
| 3584 | if (likely(!(alloc_flags & ALLOC_RESERVES)))
|
| 3585 | unusable_free += READ_ONCE(z->nr_free_highatomic);
|
| 3586 |
|
| 3587 | #ifdef CONFIG_CMA
|
| 3588 |
|
| 3589 | if (!(alloc_flags & ALLOC_CMA))
|
| 3590 | unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
|
| 3591 | #endif
|
| 3592 |
|
| 3593 | return unusable_free;
|
| 3594 | }
|
| 3595 |
|
| 3596 | |
| 3597 | * Return true if free base pages are above 'mark'. For high-order checks it
|
| 3598 | * will return true of the order-0 watermark is reached and there is at least
|
| 3599 | * one free page of a suitable size. Checking now avoids taking the zone lock
|
| 3600 | * to check in the allocation paths if no pages are free.
|
| 3601 |
|
| 3602 | bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
|
| 3603 | int highest_zoneidx, unsigned int alloc_flags,
|
| 3604 | long free_pages)
|
| 3605 | {
|
| 3606 | long min = mark;
|
| 3607 | int o;
|
| 3608 |
|
| 3609 |
|
| 3610 | free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
|
| 3611 |
|
| 3612 | if (unlikely(alloc_flags & ALLOC_RESERVES)) {
|
| 3613 | |
| 3614 | * __GFP_HIGH allows access to 50% of the min reserve as well
|
| 3615 | * as OOM.
|
| 3616 |
|
| 3617 | if (alloc_flags & ALLOC_MIN_RESERVE) {
|
| 3618 | min -= min / 2;
|
| 3619 |
|
| 3620 | |
| 3621 | * Non-blocking allocations (e.g. GFP_ATOMIC) can
|
| 3622 | * access more reserves than just __GFP_HIGH. Other
|
| 3623 | * non-blocking allocations requests such as GFP_NOWAIT
|
| 3624 | * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
|
| 3625 | * access to the min reserve.
|
| 3626 |
|
| 3627 | if (alloc_flags & ALLOC_NON_BLOCK)
|
| 3628 | min -= min / 4;
|
| 3629 | }
|
| 3630 |
|
| 3631 | |
| 3632 | * OOM victims can try even harder than the normal reserve
|
| 3633 | * users on the grounds that it's definitely going to be in
|
| 3634 | * the exit path shortly and free memory. Any allocation it
|
| 3635 | * makes during the free path will be small and short-lived.
|
| 3636 |
|
| 3637 | if (alloc_flags & ALLOC_OOM)
|
| 3638 | min -= min / 2;
|
| 3639 | }
|
| 3640 |
|
| 3641 | |
| 3642 | * Check watermarks for an order-0 allocation request. If these
|
| 3643 | * are not met, then a high-order request also cannot go ahead
|
| 3644 | * even if a suitable page happened to be free.
|
| 3645 |
|
| 3646 | if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
|
| 3647 | return false;
|
| 3648 |
|
| 3649 |
|
| 3650 | if (!order)
|
| 3651 | return true;
|
| 3652 |
|
| 3653 |
|
| 3654 | for (o = order; o < NR_PAGE_ORDERS; o++) {
|
| 3655 | struct free_area *area = &z->free_area[o];
|
| 3656 | int mt;
|
| 3657 |
|
| 3658 | if (!area->nr_free)
|
| 3659 | continue;
|
| 3660 |
|
| 3661 | for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
|
| 3662 | if (!free_area_empty(area, mt))
|
| 3663 | return true;
|
| 3664 | }
|
| 3665 |
|
| 3666 | #ifdef CONFIG_CMA
|
| 3667 | if ((alloc_flags & ALLOC_CMA) &&
|
| 3668 | !free_area_empty(area, MIGRATE_CMA)) {
|
| 3669 | return true;
|
| 3670 | }
|
| 3671 | #endif
|
| 3672 | if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
|
| 3673 | !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
|
| 3674 | return true;
|
| 3675 | }
|
| 3676 | }
|
| 3677 | return false;
|
| 3678 | }
|
| 3679 |
|
| 3680 | bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
|
| 3681 | int highest_zoneidx, unsigned int alloc_flags)
|
| 3682 | {
|
| 3683 | return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
|
| 3684 | zone_page_state(z, NR_FREE_PAGES));
|
| 3685 | }
|
| 3686 |
|
| 3687 | static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
|
| 3688 | unsigned long mark, int highest_zoneidx,
|
| 3689 | unsigned int alloc_flags, gfp_t gfp_mask)
|
| 3690 | {
|
| 3691 | long free_pages;
|
| 3692 |
|
| 3693 | free_pages = zone_page_state(z, NR_FREE_PAGES);
|
| 3694 |
|
| 3695 | |
| 3696 | * Fast check for order-0 only. If this fails then the reserves
|
| 3697 | * need to be calculated.
|
| 3698 |
|
| 3699 | if (!order) {
|
| 3700 | long usable_free;
|
| 3701 | long reserved;
|
| 3702 |
|
| 3703 | usable_free = free_pages;
|
| 3704 | reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
|
| 3705 |
|
| 3706 |
|
| 3707 | usable_free -= min(usable_free, reserved);
|
| 3708 | if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
|
| 3709 | return true;
|
| 3710 | }
|
| 3711 |
|
| 3712 | if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
|
| 3713 | free_pages))
|
| 3714 | return true;
|
| 3715 |
|
| 3716 | |
| 3717 | * Ignore watermark boosting for __GFP_HIGH order-0 allocations
|
| 3718 | * when checking the min watermark. The min watermark is the
|
| 3719 | * point where boosting is ignored so that kswapd is woken up
|
| 3720 | * when below the low watermark.
|
| 3721 |
|
| 3722 | if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
|
| 3723 | && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
|
| 3724 | mark = z->_watermark[WMARK_MIN];
|
| 3725 | return __zone_watermark_ok(z, order, mark, highest_zoneidx,
|
| 3726 | alloc_flags, free_pages);
|
| 3727 | }
|
| 3728 |
|
| 3729 | return false;
|
| 3730 | }
|
| 3731 |
|
| 3732 | #ifdef CONFIG_NUMA
|
| 3733 | int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
|
| 3734 |
|
| 3735 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
|
| 3736 | {
|
| 3737 | return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
|
| 3738 | node_reclaim_distance;
|
| 3739 | }
|
| 3740 | #else
|
| 3741 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
|
| 3742 | {
|
| 3743 | return true;
|
| 3744 | }
|
| 3745 | #endif
|
| 3746 |
|
| 3747 | |
| 3748 | * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
|
| 3749 | * fragmentation is subtle. If the preferred zone was HIGHMEM then
|
| 3750 | * premature use of a lower zone may cause lowmem pressure problems that
|
| 3751 | * are worse than fragmentation. If the next zone is ZONE_DMA then it is
|
| 3752 | * probably too small. It only makes sense to spread allocations to avoid
|
| 3753 | * fragmentation between the Normal and DMA32 zones.
|
| 3754 |
|
| 3755 | static inline unsigned int
|
| 3756 | alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
|
| 3757 | {
|
| 3758 | unsigned int alloc_flags;
|
| 3759 |
|
| 3760 | |
| 3761 | * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
|
| 3762 | * to save a branch.
|
| 3763 |
|
| 3764 | alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
|
| 3765 |
|
| 3766 | if (defrag_mode) {
|
| 3767 | alloc_flags |= ALLOC_NOFRAGMENT;
|
| 3768 | return alloc_flags;
|
| 3769 | }
|
| 3770 |
|
| 3771 | #ifdef CONFIG_ZONE_DMA32
|
| 3772 | if (!zone)
|
| 3773 | return alloc_flags;
|
| 3774 |
|
| 3775 | if (zone_idx(zone) != ZONE_NORMAL)
|
| 3776 | return alloc_flags;
|
| 3777 |
|
| 3778 | |
| 3779 | * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
|
| 3780 | * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
|
| 3781 | * on UMA that if Normal is populated then so is DMA32.
|
| 3782 |
|
| 3783 | BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
|
| 3784 | if (nr_online_nodes > 1 && !populated_zone(--zone))
|
| 3785 | return alloc_flags;
|
| 3786 |
|
| 3787 | alloc_flags |= ALLOC_NOFRAGMENT;
|
| 3788 | #endif
|
| 3789 | return alloc_flags;
|
| 3790 | }
|
| 3791 |
|
| 3792 |
|
| 3793 | static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
|
| 3794 | unsigned int alloc_flags)
|
| 3795 | {
|
| 3796 | #ifdef CONFIG_CMA
|
| 3797 | if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
|
| 3798 | alloc_flags |= ALLOC_CMA;
|
| 3799 | #endif
|
| 3800 | return alloc_flags;
|
| 3801 | }
|
| 3802 |
|
| 3803 | |
| 3804 | * get_page_from_freelist goes through the zonelist trying to allocate
|
| 3805 | * a page.
|
| 3806 |
|
| 3807 | static struct page *
|
| 3808 | get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
|
| 3809 | const struct alloc_context *ac)
|
| 3810 | {
|
| 3811 | struct zoneref *z;
|
| 3812 | struct zone *zone;
|
| 3813 | struct pglist_data *last_pgdat = NULL;
|
| 3814 | bool last_pgdat_dirty_ok = false;
|
| 3815 | bool no_fallback;
|
| 3816 | bool skip_kswapd_nodes = nr_online_nodes > 1;
|
| 3817 | bool skipped_kswapd_nodes = false;
|
| 3818 |
|
| 3819 | retry:
|
| 3820 | |
| 3821 | * Scan zonelist, looking for a zone with enough free.
|
| 3822 | * See also cpuset_current_node_allowed() comment in kernel/cgroup/cpuset.c.
|
| 3823 |
|
| 3824 | no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
|
| 3825 | z = ac->preferred_zoneref;
|
| 3826 | for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
|
| 3827 | ac->nodemask) {
|
| 3828 | struct page *page;
|
| 3829 | unsigned long mark;
|
| 3830 |
|
| 3831 | if (cpusets_enabled() &&
|
| 3832 | (alloc_flags & ALLOC_CPUSET) &&
|
| 3833 | !__cpuset_zone_allowed(zone, gfp_mask))
|
| 3834 | continue;
|
| 3835 | |
| 3836 | * When allocating a page cache page for writing, we
|
| 3837 | * want to get it from a node that is within its dirty
|
| 3838 | * limit, such that no single node holds more than its
|
| 3839 | * proportional share of globally allowed dirty pages.
|
| 3840 | * The dirty limits take into account the node's
|
| 3841 | * lowmem reserves and high watermark so that kswapd
|
| 3842 | * should be able to balance it without having to
|
| 3843 | * write pages from its LRU list.
|
| 3844 | *
|
| 3845 | * XXX: For now, allow allocations to potentially
|
| 3846 | * exceed the per-node dirty limit in the slowpath
|
| 3847 | * (spread_dirty_pages unset) before going into reclaim,
|
| 3848 | * which is important when on a NUMA setup the allowed
|
| 3849 | * nodes are together not big enough to reach the
|
| 3850 | * global limit. The proper fix for these situations
|
| 3851 | * will require awareness of nodes in the
|
| 3852 | * dirty-throttling and the flusher threads.
|
| 3853 |
|
| 3854 | if (ac->spread_dirty_pages) {
|
| 3855 | if (last_pgdat != zone->zone_pgdat) {
|
| 3856 | last_pgdat = zone->zone_pgdat;
|
| 3857 | last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
|
| 3858 | }
|
| 3859 |
|
| 3860 | if (!last_pgdat_dirty_ok)
|
| 3861 | continue;
|
| 3862 | }
|
| 3863 |
|
| 3864 | if (no_fallback && !defrag_mode && nr_online_nodes > 1 &&
|
| 3865 | zone != zonelist_zone(ac->preferred_zoneref)) {
|
| 3866 | int local_nid;
|
| 3867 |
|
| 3868 | |
| 3869 | * If moving to a remote node, retry but allow
|
| 3870 | * fragmenting fallbacks. Locality is more important
|
| 3871 | * than fragmentation avoidance.
|
| 3872 |
|
| 3873 | local_nid = zonelist_node_idx(ac->preferred_zoneref);
|
| 3874 | if (zone_to_nid(zone) != local_nid) {
|
| 3875 | alloc_flags &= ~ALLOC_NOFRAGMENT;
|
| 3876 | goto retry;
|
| 3877 | }
|
| 3878 | }
|
| 3879 |
|
| 3880 | |
| 3881 | * If kswapd is already active on a node, keep looking
|
| 3882 | * for other nodes that might be idle. This can happen
|
| 3883 | * if another process has NUMA bindings and is causing
|
| 3884 | * kswapd wakeups on only some nodes. Avoid accidental
|
| 3885 | * "node_reclaim_mode"-like behavior in this case.
|
| 3886 |
|
| 3887 | if (skip_kswapd_nodes &&
|
| 3888 | !waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
|
| 3889 | skipped_kswapd_nodes = true;
|
| 3890 | continue;
|
| 3891 | }
|
| 3892 |
|
| 3893 | cond_accept_memory(zone, order, alloc_flags);
|
| 3894 |
|
| 3895 | |
| 3896 | * Detect whether the number of free pages is below high
|
| 3897 | * watermark. If so, we will decrease pcp->high and free
|
| 3898 | * PCP pages in free path to reduce the possibility of
|
| 3899 | * premature page reclaiming. Detection is done here to
|
| 3900 | * avoid to do that in hotter free path.
|
| 3901 |
|
| 3902 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
|
| 3903 | goto check_alloc_wmark;
|
| 3904 |
|
| 3905 | mark = high_wmark_pages(zone);
|
| 3906 | if (zone_watermark_fast(zone, order, mark,
|
| 3907 | ac->highest_zoneidx, alloc_flags,
|
| 3908 | gfp_mask))
|
| 3909 | goto try_this_zone;
|
| 3910 | else
|
| 3911 | set_bit(ZONE_BELOW_HIGH, &zone->flags);
|
| 3912 |
|
| 3913 | check_alloc_wmark:
|
| 3914 | mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
|
| 3915 | if (!zone_watermark_fast(zone, order, mark,
|
| 3916 | ac->highest_zoneidx, alloc_flags,
|
| 3917 | gfp_mask)) {
|
| 3918 | int ret;
|
| 3919 |
|
| 3920 | if (cond_accept_memory(zone, order, alloc_flags))
|
| 3921 | goto try_this_zone;
|
| 3922 |
|
| 3923 | |
| 3924 | * Watermark failed for this zone, but see if we can
|
| 3925 | * grow this zone if it contains deferred pages.
|
| 3926 |
|
| 3927 | if (deferred_pages_enabled()) {
|
| 3928 | if (_deferred_grow_zone(zone, order))
|
| 3929 | goto try_this_zone;
|
| 3930 | }
|
| 3931 |
|
| 3932 | BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
|
| 3933 | if (alloc_flags & ALLOC_NO_WATERMARKS)
|
| 3934 | goto try_this_zone;
|
| 3935 |
|
| 3936 | if (!node_reclaim_enabled() ||
|
| 3937 | !zone_allows_reclaim(zonelist_zone(ac->preferred_zoneref), zone))
|
| 3938 | continue;
|
| 3939 |
|
| 3940 | ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
|
| 3941 | switch (ret) {
|
| 3942 | case NODE_RECLAIM_NOSCAN:
|
| 3943 |
|
| 3944 | continue;
|
| 3945 | case NODE_RECLAIM_FULL:
|
| 3946 |
|
| 3947 | continue;
|
| 3948 | default:
|
| 3949 |
|
| 3950 | if (zone_watermark_ok(zone, order, mark,
|
| 3951 | ac->highest_zoneidx, alloc_flags))
|
| 3952 | goto try_this_zone;
|
| 3953 |
|
| 3954 | continue;
|
| 3955 | }
|
| 3956 | }
|
| 3957 |
|
| 3958 | try_this_zone:
|
| 3959 | page = rmqueue(zonelist_zone(ac->preferred_zoneref), zone, order,
|
| 3960 | gfp_mask, alloc_flags, ac->migratetype);
|
| 3961 | if (page) {
|
| 3962 | prep_new_page(page, order, gfp_mask, alloc_flags);
|
| 3963 |
|
| 3964 | |
| 3965 | * If this is a high-order atomic allocation then check
|
| 3966 | * if the pageblock should be reserved for the future
|
| 3967 |
|
| 3968 | if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
|
| 3969 | reserve_highatomic_pageblock(page, order, zone);
|
| 3970 |
|
| 3971 | return page;
|
| 3972 | } else {
|
| 3973 | if (cond_accept_memory(zone, order, alloc_flags))
|
| 3974 | goto try_this_zone;
|
| 3975 |
|
| 3976 |
|
| 3977 | if (deferred_pages_enabled()) {
|
| 3978 | if (_deferred_grow_zone(zone, order))
|
| 3979 | goto try_this_zone;
|
| 3980 | }
|
| 3981 | }
|
| 3982 | }
|
| 3983 |
|
| 3984 | |
| 3985 | * If we skipped over nodes with active kswapds and found no
|
| 3986 | * idle nodes, retry and place anywhere the watermarks permit.
|
| 3987 |
|
| 3988 | if (skip_kswapd_nodes && skipped_kswapd_nodes) {
|
| 3989 | skip_kswapd_nodes = false;
|
| 3990 | goto retry;
|
| 3991 | }
|
| 3992 |
|
| 3993 | |
| 3994 | * It's possible on a UMA machine to get through all zones that are
|
| 3995 | * fragmented. If avoiding fragmentation, reset and try again.
|
| 3996 |
|
| 3997 | if (no_fallback && !defrag_mode) {
|
| 3998 | alloc_flags &= ~ALLOC_NOFRAGMENT;
|
| 3999 | goto retry;
|
| 4000 | }
|
| 4001 |
|
| 4002 | return NULL;
|
| 4003 | }
|
| 4004 |
|
| 4005 | static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
|
| 4006 | {
|
| 4007 | unsigned int filter = SHOW_MEM_FILTER_NODES;
|
| 4008 |
|
| 4009 | |
| 4010 | * This documents exceptions given to allocations in certain
|
| 4011 | * contexts that are allowed to allocate outside current's set
|
| 4012 | * of allowed nodes.
|
| 4013 |
|
| 4014 | if (!(gfp_mask & __GFP_NOMEMALLOC))
|
| 4015 | if (tsk_is_oom_victim(current) ||
|
| 4016 | (current->flags & (PF_MEMALLOC | PF_EXITING)))
|
| 4017 | filter &= ~SHOW_MEM_FILTER_NODES;
|
| 4018 | if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
|
| 4019 | filter &= ~SHOW_MEM_FILTER_NODES;
|
| 4020 |
|
| 4021 | __show_mem(filter, nodemask, gfp_zone(gfp_mask));
|
| 4022 | mem_cgroup_show_protected_memory(NULL);
|
| 4023 | }
|
| 4024 |
|
| 4025 | void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
|
| 4026 | {
|
| 4027 | struct va_format vaf;
|
| 4028 | va_list args;
|
| 4029 | static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
|
| 4030 |
|
| 4031 | if ((gfp_mask & __GFP_NOWARN) ||
|
| 4032 | !__ratelimit(&nopage_rs) ||
|
| 4033 | ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
|
| 4034 | return;
|
| 4035 |
|
| 4036 | va_start(args, fmt);
|
| 4037 | vaf.fmt = fmt;
|
| 4038 | vaf.va = &args;
|
| 4039 | pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
|
| 4040 | current->comm, &vaf, gfp_mask, &gfp_mask,
|
| 4041 | nodemask_pr_args(nodemask));
|
| 4042 | va_end(args);
|
| 4043 |
|
| 4044 | cpuset_print_current_mems_allowed();
|
| 4045 | pr_cont("\n");
|
| 4046 | dump_stack();
|
| 4047 | warn_alloc_show_mem(gfp_mask, nodemask);
|
| 4048 | }
|
| 4049 |
|
| 4050 | static inline struct page *
|
| 4051 | __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
|
| 4052 | unsigned int alloc_flags,
|
| 4053 | const struct alloc_context *ac)
|
| 4054 | {
|
| 4055 | struct page *page;
|
| 4056 |
|
| 4057 | page = get_page_from_freelist(gfp_mask, order,
|
| 4058 | alloc_flags|ALLOC_CPUSET, ac);
|
| 4059 | |
| 4060 | * fallback to ignore cpuset restriction if our nodes
|
| 4061 | * are depleted
|
| 4062 |
|
| 4063 | if (!page)
|
| 4064 | page = get_page_from_freelist(gfp_mask, order,
|
| 4065 | alloc_flags, ac);
|
| 4066 | return page;
|
| 4067 | }
|
| 4068 |
|
| 4069 | static inline struct page *
|
| 4070 | __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
|
| 4071 | const struct alloc_context *ac, unsigned long *did_some_progress)
|
| 4072 | {
|
| 4073 | struct oom_control oc = {
|
| 4074 | .zonelist = ac->zonelist,
|
| 4075 | .nodemask = ac->nodemask,
|
| 4076 | .memcg = NULL,
|
| 4077 | .gfp_mask = gfp_mask,
|
| 4078 | .order = order,
|
| 4079 | };
|
| 4080 | struct page *page;
|
| 4081 |
|
| 4082 | *did_some_progress = 0;
|
| 4083 |
|
| 4084 | |
| 4085 | * Acquire the oom lock. If that fails, somebody else is
|
| 4086 | * making progress for us.
|
| 4087 |
|
| 4088 | if (!mutex_trylock(&oom_lock)) {
|
| 4089 | *did_some_progress = 1;
|
| 4090 | schedule_timeout_uninterruptible(1);
|
| 4091 | return NULL;
|
| 4092 | }
|
| 4093 |
|
| 4094 | |
| 4095 | * Go through the zonelist yet one more time, keep very high watermark
|
| 4096 | * here, this is only to catch a parallel oom killing, we must fail if
|
| 4097 | * we're still under heavy pressure. But make sure that this reclaim
|
| 4098 | * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
|
| 4099 | * allocation which will never fail due to oom_lock already held.
|
| 4100 |
|
| 4101 | page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
|
| 4102 | ~__GFP_DIRECT_RECLAIM, order,
|
| 4103 | ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
|
| 4104 | if (page)
|
| 4105 | goto out;
|
| 4106 |
|
| 4107 |
|
| 4108 | if (current->flags & PF_DUMPCORE)
|
| 4109 | goto out;
|
| 4110 |
|
| 4111 | if (order > PAGE_ALLOC_COSTLY_ORDER)
|
| 4112 | goto out;
|
| 4113 | |
| 4114 | * We have already exhausted all our reclaim opportunities without any
|
| 4115 | * success so it is time to admit defeat. We will skip the OOM killer
|
| 4116 | * because it is very likely that the caller has a more reasonable
|
| 4117 | * fallback than shooting a random task.
|
| 4118 | *
|
| 4119 | * The OOM killer may not free memory on a specific node.
|
| 4120 |
|
| 4121 | if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
|
| 4122 | goto out;
|
| 4123 |
|
| 4124 | if (ac->highest_zoneidx < ZONE_NORMAL)
|
| 4125 | goto out;
|
| 4126 | if (pm_suspended_storage())
|
| 4127 | goto out;
|
| 4128 | |
| 4129 | * XXX: GFP_NOFS allocations should rather fail than rely on
|
| 4130 | * other request to make a forward progress.
|
| 4131 | * We are in an unfortunate situation where out_of_memory cannot
|
| 4132 | * do much for this context but let's try it to at least get
|
| 4133 | * access to memory reserved if the current task is killed (see
|
| 4134 | * out_of_memory). Once filesystems are ready to handle allocation
|
| 4135 | * failures more gracefully we should just bail out here.
|
| 4136 |
|
| 4137 |
|
| 4138 |
|
| 4139 | if (out_of_memory(&oc) ||
|
| 4140 | WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
|
| 4141 | *did_some_progress = 1;
|
| 4142 |
|
| 4143 | |
| 4144 | * Help non-failing allocations by giving them access to memory
|
| 4145 | * reserves
|
| 4146 |
|
| 4147 | if (gfp_mask & __GFP_NOFAIL)
|
| 4148 | page = __alloc_pages_cpuset_fallback(gfp_mask, order,
|
| 4149 | ALLOC_NO_WATERMARKS, ac);
|
| 4150 | }
|
| 4151 | out:
|
| 4152 | mutex_unlock(&oom_lock);
|
| 4153 | return page;
|
| 4154 | }
|
| 4155 |
|
| 4156 | |
| 4157 | * Maximum number of compaction retries with a progress before OOM
|
| 4158 | * killer is consider as the only way to move forward.
|
| 4159 |
|
| 4160 | #define MAX_COMPACT_RETRIES 16
|
| 4161 |
|
| 4162 | #ifdef CONFIG_COMPACTION
|
| 4163 |
|
| 4164 | static struct page *
|
| 4165 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
|
| 4166 | unsigned int alloc_flags, const struct alloc_context *ac,
|
| 4167 | enum compact_priority prio, enum compact_result *compact_result)
|
| 4168 | {
|
| 4169 | struct page *page = NULL;
|
| 4170 | unsigned long pflags;
|
| 4171 | unsigned int noreclaim_flag;
|
| 4172 |
|
| 4173 | if (!order)
|
| 4174 | return NULL;
|
| 4175 |
|
| 4176 | psi_memstall_enter(&pflags);
|
| 4177 | delayacct_compact_start();
|
| 4178 | noreclaim_flag = memalloc_noreclaim_save();
|
| 4179 |
|
| 4180 | *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
|
| 4181 | prio, &page);
|
| 4182 |
|
| 4183 | memalloc_noreclaim_restore(noreclaim_flag);
|
| 4184 | psi_memstall_leave(&pflags);
|
| 4185 | delayacct_compact_end();
|
| 4186 |
|
| 4187 | if (*compact_result == COMPACT_SKIPPED)
|
| 4188 | return NULL;
|
| 4189 | |
| 4190 | * At least in one zone compaction wasn't deferred or skipped, so let's
|
| 4191 | * count a compaction stall
|
| 4192 |
|
| 4193 | count_vm_event(COMPACTSTALL);
|
| 4194 |
|
| 4195 |
|
| 4196 | if (page)
|
| 4197 | prep_new_page(page, order, gfp_mask, alloc_flags);
|
| 4198 |
|
| 4199 |
|
| 4200 | if (!page)
|
| 4201 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
|
| 4202 |
|
| 4203 | if (page) {
|
| 4204 | struct zone *zone = page_zone(page);
|
| 4205 |
|
| 4206 | zone->compact_blockskip_flush = false;
|
| 4207 | compaction_defer_reset(zone, order, true);
|
| 4208 | count_vm_event(COMPACTSUCCESS);
|
| 4209 | return page;
|
| 4210 | }
|
| 4211 |
|
| 4212 | |
| 4213 | * It's bad if compaction run occurs and fails. The most likely reason
|
| 4214 | * is that pages exist, but not enough to satisfy watermarks.
|
| 4215 |
|
| 4216 | count_vm_event(COMPACTFAIL);
|
| 4217 |
|
| 4218 | cond_resched();
|
| 4219 |
|
| 4220 | return NULL;
|
| 4221 | }
|
| 4222 |
|
| 4223 | static inline bool
|
| 4224 | should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
|
| 4225 | enum compact_result compact_result,
|
| 4226 | enum compact_priority *compact_priority,
|
| 4227 | int *compaction_retries)
|
| 4228 | {
|
| 4229 | int max_retries = MAX_COMPACT_RETRIES;
|
| 4230 | int min_priority;
|
| 4231 | bool ret = false;
|
| 4232 | int retries = *compaction_retries;
|
| 4233 | enum compact_priority priority = *compact_priority;
|
| 4234 |
|
| 4235 | if (!order)
|
| 4236 | return false;
|
| 4237 |
|
| 4238 | if (fatal_signal_pending(current))
|
| 4239 | return false;
|
| 4240 |
|
| 4241 | |
| 4242 | * Compaction was skipped due to a lack of free order-0
|
| 4243 | * migration targets. Continue if reclaim can help.
|
| 4244 |
|
| 4245 | if (compact_result == COMPACT_SKIPPED) {
|
| 4246 | ret = compaction_zonelist_suitable(ac, order, alloc_flags);
|
| 4247 | goto out;
|
| 4248 | }
|
| 4249 |
|
| 4250 | |
| 4251 | * Compaction managed to coalesce some page blocks, but the
|
| 4252 | * allocation failed presumably due to a race. Retry some.
|
| 4253 |
|
| 4254 | if (compact_result == COMPACT_SUCCESS) {
|
| 4255 | |
| 4256 | * !costly requests are much more important than
|
| 4257 | * __GFP_RETRY_MAYFAIL costly ones because they are de
|
| 4258 | * facto nofail and invoke OOM killer to move on while
|
| 4259 | * costly can fail and users are ready to cope with
|
| 4260 | * that. 1/4 retries is rather arbitrary but we would
|
| 4261 | * need much more detailed feedback from compaction to
|
| 4262 | * make a better decision.
|
| 4263 |
|
| 4264 | if (order > PAGE_ALLOC_COSTLY_ORDER)
|
| 4265 | max_retries /= 4;
|
| 4266 |
|
| 4267 | if (++(*compaction_retries) <= max_retries) {
|
| 4268 | ret = true;
|
| 4269 | goto out;
|
| 4270 | }
|
| 4271 | }
|
| 4272 |
|
| 4273 | |
| 4274 | * Compaction failed. Retry with increasing priority.
|
| 4275 |
|
| 4276 | min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
|
| 4277 | MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
|
| 4278 |
|
| 4279 | if (*compact_priority > min_priority) {
|
| 4280 | (*compact_priority)--;
|
| 4281 | *compaction_retries = 0;
|
| 4282 | ret = true;
|
| 4283 | }
|
| 4284 | out:
|
| 4285 | trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
|
| 4286 | return ret;
|
| 4287 | }
|
| 4288 | #else
|
| 4289 | static inline struct page *
|
| 4290 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
|
| 4291 | unsigned int alloc_flags, const struct alloc_context *ac,
|
| 4292 | enum compact_priority prio, enum compact_result *compact_result)
|
| 4293 | {
|
| 4294 | *compact_result = COMPACT_SKIPPED;
|
| 4295 | return NULL;
|
| 4296 | }
|
| 4297 |
|
| 4298 | static inline bool
|
| 4299 | should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
|
| 4300 | enum compact_result compact_result,
|
| 4301 | enum compact_priority *compact_priority,
|
| 4302 | int *compaction_retries)
|
| 4303 | {
|
| 4304 | struct zone *zone;
|
| 4305 | struct zoneref *z;
|
| 4306 |
|
| 4307 | if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
|
| 4308 | return false;
|
| 4309 |
|
| 4310 | |
| 4311 | * There are setups with compaction disabled which would prefer to loop
|
| 4312 | * inside the allocator rather than hit the oom killer prematurely.
|
| 4313 | * Let's give them a good hope and keep retrying while the order-0
|
| 4314 | * watermarks are OK.
|
| 4315 |
|
| 4316 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
|
| 4317 | ac->highest_zoneidx, ac->nodemask) {
|
| 4318 | if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
|
| 4319 | ac->highest_zoneidx, alloc_flags))
|
| 4320 | return true;
|
| 4321 | }
|
| 4322 | return false;
|
| 4323 | }
|
| 4324 | #endif
|
| 4325 |
|
| 4326 | #ifdef CONFIG_LOCKDEP
|
| 4327 | static struct lockdep_map __fs_reclaim_map =
|
| 4328 | STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
|
| 4329 |
|
| 4330 | static bool __need_reclaim(gfp_t gfp_mask)
|
| 4331 | {
|
| 4332 |
|
| 4333 | if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
|
| 4334 | return false;
|
| 4335 |
|
| 4336 |
|
| 4337 | if (current->flags & PF_MEMALLOC)
|
| 4338 | return false;
|
| 4339 |
|
| 4340 | if (gfp_mask & __GFP_NOLOCKDEP)
|
| 4341 | return false;
|
| 4342 |
|
| 4343 | return true;
|
| 4344 | }
|
| 4345 |
|
| 4346 | void __fs_reclaim_acquire(unsigned long ip)
|
| 4347 | {
|
| 4348 | lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
|
| 4349 | }
|
| 4350 |
|
| 4351 | void __fs_reclaim_release(unsigned long ip)
|
| 4352 | {
|
| 4353 | lock_release(&__fs_reclaim_map, ip);
|
| 4354 | }
|
| 4355 |
|
| 4356 | void fs_reclaim_acquire(gfp_t gfp_mask)
|
| 4357 | {
|
| 4358 | gfp_mask = current_gfp_context(gfp_mask);
|
| 4359 |
|
| 4360 | if (__need_reclaim(gfp_mask)) {
|
| 4361 | if (gfp_mask & __GFP_FS)
|
| 4362 | __fs_reclaim_acquire(_RET_IP_);
|
| 4363 |
|
| 4364 | #ifdef CONFIG_MMU_NOTIFIER
|
| 4365 | lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
|
| 4366 | lock_map_release(&__mmu_notifier_invalidate_range_start_map);
|
| 4367 | #endif
|
| 4368 |
|
| 4369 | }
|
| 4370 | }
|
| 4371 | EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
|
| 4372 |
|
| 4373 | void fs_reclaim_release(gfp_t gfp_mask)
|
| 4374 | {
|
| 4375 | gfp_mask = current_gfp_context(gfp_mask);
|
| 4376 |
|
| 4377 | if (__need_reclaim(gfp_mask)) {
|
| 4378 | if (gfp_mask & __GFP_FS)
|
| 4379 | __fs_reclaim_release(_RET_IP_);
|
| 4380 | }
|
| 4381 | }
|
| 4382 | EXPORT_SYMBOL_GPL(fs_reclaim_release);
|
| 4383 | #endif
|
| 4384 |
|
| 4385 | |
| 4386 | * Zonelists may change due to hotplug during allocation. Detect when zonelists
|
| 4387 | * have been rebuilt so allocation retries. Reader side does not lock and
|
| 4388 | * retries the allocation if zonelist changes. Writer side is protected by the
|
| 4389 | * embedded spin_lock.
|
| 4390 |
|
| 4391 | static DEFINE_SEQLOCK(zonelist_update_seq);
|
| 4392 |
|
| 4393 | static unsigned int zonelist_iter_begin(void)
|
| 4394 | {
|
| 4395 | if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
|
| 4396 | return read_seqbegin(&zonelist_update_seq);
|
| 4397 |
|
| 4398 | return 0;
|
| 4399 | }
|
| 4400 |
|
| 4401 | static unsigned int check_retry_zonelist(unsigned int seq)
|
| 4402 | {
|
| 4403 | if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
|
| 4404 | return read_seqretry(&zonelist_update_seq, seq);
|
| 4405 |
|
| 4406 | return seq;
|
| 4407 | }
|
| 4408 |
|
| 4409 |
|
| 4410 | static unsigned long
|
| 4411 | __perform_reclaim(gfp_t gfp_mask, unsigned int order,
|
| 4412 | const struct alloc_context *ac)
|
| 4413 | {
|
| 4414 | unsigned int noreclaim_flag;
|
| 4415 | unsigned long progress;
|
| 4416 |
|
| 4417 | cond_resched();
|
| 4418 |
|
| 4419 |
|
| 4420 | cpuset_memory_pressure_bump();
|
| 4421 | fs_reclaim_acquire(gfp_mask);
|
| 4422 | noreclaim_flag = memalloc_noreclaim_save();
|
| 4423 |
|
| 4424 | progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
|
| 4425 | ac->nodemask);
|
| 4426 |
|
| 4427 | memalloc_noreclaim_restore(noreclaim_flag);
|
| 4428 | fs_reclaim_release(gfp_mask);
|
| 4429 |
|
| 4430 | cond_resched();
|
| 4431 |
|
| 4432 | return progress;
|
| 4433 | }
|
| 4434 |
|
| 4435 |
|
| 4436 | static inline struct page *
|
| 4437 | __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
|
| 4438 | unsigned int alloc_flags, const struct alloc_context *ac,
|
| 4439 | unsigned long *did_some_progress)
|
| 4440 | {
|
| 4441 | struct page *page = NULL;
|
| 4442 | unsigned long pflags;
|
| 4443 | bool drained = false;
|
| 4444 |
|
| 4445 | psi_memstall_enter(&pflags);
|
| 4446 | *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
|
| 4447 | if (unlikely(!(*did_some_progress)))
|
| 4448 | goto out;
|
| 4449 |
|
| 4450 | retry:
|
| 4451 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
|
| 4452 |
|
| 4453 | |
| 4454 | * If an allocation failed after direct reclaim, it could be because
|
| 4455 | * pages are pinned on the per-cpu lists or in high alloc reserves.
|
| 4456 | * Shrink them and try again
|
| 4457 |
|
| 4458 | if (!page && !drained) {
|
| 4459 | unreserve_highatomic_pageblock(ac, false);
|
| 4460 | drain_all_pages(NULL);
|
| 4461 | drained = true;
|
| 4462 | goto retry;
|
| 4463 | }
|
| 4464 | out:
|
| 4465 | psi_memstall_leave(&pflags);
|
| 4466 |
|
| 4467 | return page;
|
| 4468 | }
|
| 4469 |
|
| 4470 | static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
|
| 4471 | const struct alloc_context *ac)
|
| 4472 | {
|
| 4473 | struct zoneref *z;
|
| 4474 | struct zone *zone;
|
| 4475 | pg_data_t *last_pgdat = NULL;
|
| 4476 | enum zone_type highest_zoneidx = ac->highest_zoneidx;
|
| 4477 | unsigned int reclaim_order;
|
| 4478 |
|
| 4479 | if (defrag_mode)
|
| 4480 | reclaim_order = max(order, pageblock_order);
|
| 4481 | else
|
| 4482 | reclaim_order = order;
|
| 4483 |
|
| 4484 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
|
| 4485 | ac->nodemask) {
|
| 4486 | if (!managed_zone(zone))
|
| 4487 | continue;
|
| 4488 | if (last_pgdat == zone->zone_pgdat)
|
| 4489 | continue;
|
| 4490 | wakeup_kswapd(zone, gfp_mask, reclaim_order, highest_zoneidx);
|
| 4491 | last_pgdat = zone->zone_pgdat;
|
| 4492 | }
|
| 4493 | }
|
| 4494 |
|
| 4495 | static inline unsigned int
|
| 4496 | gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
|
| 4497 | {
|
| 4498 | unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
|
| 4499 |
|
| 4500 | |
| 4501 | * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
|
| 4502 | * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
|
| 4503 | * to save two branches.
|
| 4504 |
|
| 4505 | BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
|
| 4506 | BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
|
| 4507 |
|
| 4508 | |
| 4509 | * The caller may dip into page reserves a bit more if the caller
|
| 4510 | * cannot run direct reclaim, or if the caller has realtime scheduling
|
| 4511 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
|
| 4512 | * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
|
| 4513 |
|
| 4514 | alloc_flags |= (__force int)
|
| 4515 | (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
|
| 4516 |
|
| 4517 | if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
|
| 4518 | |
| 4519 | * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
|
| 4520 | * if it can't schedule.
|
| 4521 |
|
| 4522 | if (!(gfp_mask & __GFP_NOMEMALLOC)) {
|
| 4523 | alloc_flags |= ALLOC_NON_BLOCK;
|
| 4524 |
|
| 4525 | if (order > 0 && (alloc_flags & ALLOC_MIN_RESERVE))
|
| 4526 | alloc_flags |= ALLOC_HIGHATOMIC;
|
| 4527 | }
|
| 4528 |
|
| 4529 | |
| 4530 | * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
|
| 4531 | * GFP_ATOMIC) rather than fail, see the comment for
|
| 4532 | * cpuset_current_node_allowed().
|
| 4533 |
|
| 4534 | if (alloc_flags & ALLOC_MIN_RESERVE)
|
| 4535 | alloc_flags &= ~ALLOC_CPUSET;
|
| 4536 | } else if (unlikely(rt_or_dl_task(current)) && in_task())
|
| 4537 | alloc_flags |= ALLOC_MIN_RESERVE;
|
| 4538 |
|
| 4539 | alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
|
| 4540 |
|
| 4541 | if (defrag_mode)
|
| 4542 | alloc_flags |= ALLOC_NOFRAGMENT;
|
| 4543 |
|
| 4544 | return alloc_flags;
|
| 4545 | }
|
| 4546 |
|
| 4547 | static bool oom_reserves_allowed(struct task_struct *tsk)
|
| 4548 | {
|
| 4549 | if (!tsk_is_oom_victim(tsk))
|
| 4550 | return false;
|
| 4551 |
|
| 4552 | |
| 4553 | * !MMU doesn't have oom reaper so give access to memory reserves
|
| 4554 | * only to the thread with TIF_MEMDIE set
|
| 4555 |
|
| 4556 | if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
|
| 4557 | return false;
|
| 4558 |
|
| 4559 | return true;
|
| 4560 | }
|
| 4561 |
|
| 4562 | |
| 4563 | * Distinguish requests which really need access to full memory
|
| 4564 | * reserves from oom victims which can live with a portion of it
|
| 4565 |
|
| 4566 | static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
|
| 4567 | {
|
| 4568 | if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
|
| 4569 | return 0;
|
| 4570 | if (gfp_mask & __GFP_MEMALLOC)
|
| 4571 | return ALLOC_NO_WATERMARKS;
|
| 4572 | if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
|
| 4573 | return ALLOC_NO_WATERMARKS;
|
| 4574 | if (!in_interrupt()) {
|
| 4575 | if (current->flags & PF_MEMALLOC)
|
| 4576 | return ALLOC_NO_WATERMARKS;
|
| 4577 | else if (oom_reserves_allowed(current))
|
| 4578 | return ALLOC_OOM;
|
| 4579 | }
|
| 4580 |
|
| 4581 | return 0;
|
| 4582 | }
|
| 4583 |
|
| 4584 | bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
|
| 4585 | {
|
| 4586 | return !!__gfp_pfmemalloc_flags(gfp_mask);
|
| 4587 | }
|
| 4588 |
|
| 4589 | |
| 4590 | * Checks whether it makes sense to retry the reclaim to make a forward progress
|
| 4591 | * for the given allocation request.
|
| 4592 | *
|
| 4593 | * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
|
| 4594 | * without success, or when we couldn't even meet the watermark if we
|
| 4595 | * reclaimed all remaining pages on the LRU lists.
|
| 4596 | *
|
| 4597 | * Returns true if a retry is viable or false to enter the oom path.
|
| 4598 |
|
| 4599 | static inline bool
|
| 4600 | should_reclaim_retry(gfp_t gfp_mask, unsigned order,
|
| 4601 | struct alloc_context *ac, int alloc_flags,
|
| 4602 | bool did_some_progress, int *no_progress_loops)
|
| 4603 | {
|
| 4604 | struct zone *zone;
|
| 4605 | struct zoneref *z;
|
| 4606 | bool ret = false;
|
| 4607 |
|
| 4608 | |
| 4609 | * Costly allocations might have made a progress but this doesn't mean
|
| 4610 | * their order will become available due to high fragmentation so
|
| 4611 | * always increment the no progress counter for them
|
| 4612 |
|
| 4613 | if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
|
| 4614 | *no_progress_loops = 0;
|
| 4615 | else
|
| 4616 | (*no_progress_loops)++;
|
| 4617 |
|
| 4618 | if (*no_progress_loops > MAX_RECLAIM_RETRIES)
|
| 4619 | goto out;
|
| 4620 |
|
| 4621 |
|
| 4622 | |
| 4623 | * Keep reclaiming pages while there is a chance this will lead
|
| 4624 | * somewhere. If none of the target zones can satisfy our allocation
|
| 4625 | * request even if all reclaimable pages are considered then we are
|
| 4626 | * screwed and have to go OOM.
|
| 4627 |
|
| 4628 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
|
| 4629 | ac->highest_zoneidx, ac->nodemask) {
|
| 4630 | unsigned long available;
|
| 4631 | unsigned long reclaimable;
|
| 4632 | unsigned long min_wmark = min_wmark_pages(zone);
|
| 4633 | bool wmark;
|
| 4634 |
|
| 4635 | if (cpusets_enabled() &&
|
| 4636 | (alloc_flags & ALLOC_CPUSET) &&
|
| 4637 | !__cpuset_zone_allowed(zone, gfp_mask))
|
| 4638 | continue;
|
| 4639 |
|
| 4640 | available = reclaimable = zone_reclaimable_pages(zone);
|
| 4641 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
|
| 4642 |
|
| 4643 | |
| 4644 | * Would the allocation succeed if we reclaimed all
|
| 4645 | * reclaimable pages?
|
| 4646 |
|
| 4647 | wmark = __zone_watermark_ok(zone, order, min_wmark,
|
| 4648 | ac->highest_zoneidx, alloc_flags, available);
|
| 4649 | trace_reclaim_retry_zone(z, order, reclaimable,
|
| 4650 | available, min_wmark, *no_progress_loops, wmark);
|
| 4651 | if (wmark) {
|
| 4652 | ret = true;
|
| 4653 | break;
|
| 4654 | }
|
| 4655 | }
|
| 4656 |
|
| 4657 | |
| 4658 | * Memory allocation/reclaim might be called from a WQ context and the
|
| 4659 | * current implementation of the WQ concurrency control doesn't
|
| 4660 | * recognize that a particular WQ is congested if the worker thread is
|
| 4661 | * looping without ever sleeping. Therefore we have to do a short sleep
|
| 4662 | * here rather than calling cond_resched().
|
| 4663 |
|
| 4664 | if (current->flags & PF_WQ_WORKER)
|
| 4665 | schedule_timeout_uninterruptible(1);
|
| 4666 | else
|
| 4667 | cond_resched();
|
| 4668 | out:
|
| 4669 |
|
| 4670 | if (!ret)
|
| 4671 | return unreserve_highatomic_pageblock(ac, true);
|
| 4672 |
|
| 4673 | return ret;
|
| 4674 | }
|
| 4675 |
|
| 4676 | static inline bool
|
| 4677 | check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
|
| 4678 | {
|
| 4679 | |
| 4680 | * It's possible that cpuset's mems_allowed and the nodemask from
|
| 4681 | * mempolicy don't intersect. This should be normally dealt with by
|
| 4682 | * policy_nodemask(), but it's possible to race with cpuset update in
|
| 4683 | * such a way the check therein was true, and then it became false
|
| 4684 | * before we got our cpuset_mems_cookie here.
|
| 4685 | * This assumes that for all allocations, ac->nodemask can come only
|
| 4686 | * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
|
| 4687 | * when it does not intersect with the cpuset restrictions) or the
|
| 4688 | * caller can deal with a violated nodemask.
|
| 4689 |
|
| 4690 | if (cpusets_enabled() && ac->nodemask &&
|
| 4691 | !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
|
| 4692 | ac->nodemask = NULL;
|
| 4693 | return true;
|
| 4694 | }
|
| 4695 |
|
| 4696 | |
| 4697 | * When updating a task's mems_allowed or mempolicy nodemask, it is
|
| 4698 | * possible to race with parallel threads in such a way that our
|
| 4699 | * allocation can fail while the mask is being updated. If we are about
|
| 4700 | * to fail, check if the cpuset changed during allocation and if so,
|
| 4701 | * retry.
|
| 4702 |
|
| 4703 | if (read_mems_allowed_retry(cpuset_mems_cookie))
|
| 4704 | return true;
|
| 4705 |
|
| 4706 | return false;
|
| 4707 | }
|
| 4708 |
|
| 4709 | static inline struct page *
|
| 4710 | __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
|
| 4711 | struct alloc_context *ac)
|
| 4712 | {
|
| 4713 | bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
|
| 4714 | bool can_compact = can_direct_reclaim && gfp_compaction_allowed(gfp_mask);
|
| 4715 | bool nofail = gfp_mask & __GFP_NOFAIL;
|
| 4716 | const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
|
| 4717 | struct page *page = NULL;
|
| 4718 | unsigned int alloc_flags;
|
| 4719 | unsigned long did_some_progress;
|
| 4720 | enum compact_priority compact_priority;
|
| 4721 | enum compact_result compact_result;
|
| 4722 | int compaction_retries;
|
| 4723 | int no_progress_loops;
|
| 4724 | unsigned int cpuset_mems_cookie;
|
| 4725 | unsigned int zonelist_iter_cookie;
|
| 4726 | int reserve_flags;
|
| 4727 | bool compact_first = false;
|
| 4728 | bool can_retry_reserves = true;
|
| 4729 |
|
| 4730 | if (unlikely(nofail)) {
|
| 4731 | |
| 4732 | * Also we don't support __GFP_NOFAIL without __GFP_DIRECT_RECLAIM,
|
| 4733 | * otherwise, we may result in lockup.
|
| 4734 |
|
| 4735 | WARN_ON_ONCE(!can_direct_reclaim);
|
| 4736 | |
| 4737 | * PF_MEMALLOC request from this context is rather bizarre
|
| 4738 | * because we cannot reclaim anything and only can loop waiting
|
| 4739 | * for somebody to do a work for us.
|
| 4740 |
|
| 4741 | WARN_ON_ONCE(current->flags & PF_MEMALLOC);
|
| 4742 | }
|
| 4743 |
|
| 4744 | restart:
|
| 4745 | compaction_retries = 0;
|
| 4746 | no_progress_loops = 0;
|
| 4747 | compact_result = COMPACT_SKIPPED;
|
| 4748 | compact_priority = DEF_COMPACT_PRIORITY;
|
| 4749 | cpuset_mems_cookie = read_mems_allowed_begin();
|
| 4750 | zonelist_iter_cookie = zonelist_iter_begin();
|
| 4751 |
|
| 4752 | |
| 4753 | * For costly allocations, try direct compaction first, as it's likely
|
| 4754 | * that we have enough base pages and don't need to reclaim. For non-
|
| 4755 | * movable high-order allocations, do that as well, as compaction will
|
| 4756 | * try prevent permanent fragmentation by migrating from blocks of the
|
| 4757 | * same migratetype.
|
| 4758 |
|
| 4759 | if (can_compact && (costly_order || (order > 0 &&
|
| 4760 | ac->migratetype != MIGRATE_MOVABLE))) {
|
| 4761 | compact_first = true;
|
| 4762 | compact_priority = INIT_COMPACT_PRIORITY;
|
| 4763 | }
|
| 4764 |
|
| 4765 | |
| 4766 | * The fast path uses conservative alloc_flags to succeed only until
|
| 4767 | * kswapd needs to be woken up, and to avoid the cost of setting up
|
| 4768 | * alloc_flags precisely. So we do that now.
|
| 4769 |
|
| 4770 | alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
|
| 4771 |
|
| 4772 | |
| 4773 | * We need to recalculate the starting point for the zonelist iterator
|
| 4774 | * because we might have used different nodemask in the fast path, or
|
| 4775 | * there was a cpuset modification and we are retrying - otherwise we
|
| 4776 | * could end up iterating over non-eligible zones endlessly.
|
| 4777 |
|
| 4778 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
|
| 4779 | ac->highest_zoneidx, ac->nodemask);
|
| 4780 | if (!zonelist_zone(ac->preferred_zoneref))
|
| 4781 | goto nopage;
|
| 4782 |
|
| 4783 | |
| 4784 | * Check for insane configurations where the cpuset doesn't contain
|
| 4785 | * any suitable zone to satisfy the request - e.g. non-movable
|
| 4786 | * GFP_HIGHUSER allocations from MOVABLE nodes only.
|
| 4787 |
|
| 4788 | if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
|
| 4789 | struct zoneref *z = first_zones_zonelist(ac->zonelist,
|
| 4790 | ac->highest_zoneidx,
|
| 4791 | &cpuset_current_mems_allowed);
|
| 4792 | if (!zonelist_zone(z))
|
| 4793 | goto nopage;
|
| 4794 | }
|
| 4795 |
|
| 4796 | retry:
|
| 4797 |
|
| 4798 | if (alloc_flags & ALLOC_KSWAPD)
|
| 4799 | wake_all_kswapds(order, gfp_mask, ac);
|
| 4800 |
|
| 4801 | |
| 4802 | * The adjusted alloc_flags might result in immediate success, so try
|
| 4803 | * that first
|
| 4804 |
|
| 4805 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
|
| 4806 | if (page)
|
| 4807 | goto got_pg;
|
| 4808 |
|
| 4809 | reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
|
| 4810 | if (reserve_flags)
|
| 4811 | alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
|
| 4812 | (alloc_flags & ALLOC_KSWAPD);
|
| 4813 |
|
| 4814 | |
| 4815 | * Reset the nodemask and zonelist iterators if memory policies can be
|
| 4816 | * ignored. These allocations are high priority and system rather than
|
| 4817 | * user oriented.
|
| 4818 |
|
| 4819 | if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
|
| 4820 | ac->nodemask = NULL;
|
| 4821 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
|
| 4822 | ac->highest_zoneidx, ac->nodemask);
|
| 4823 |
|
| 4824 | |
| 4825 | * The first time we adjust anything due to being allowed to
|
| 4826 | * ignore memory policies or watermarks, retry immediately. This
|
| 4827 | * allows us to keep the first allocation attempt optimistic so
|
| 4828 | * it can succeed in a zone that is still above watermarks.
|
| 4829 |
|
| 4830 | if (can_retry_reserves) {
|
| 4831 | can_retry_reserves = false;
|
| 4832 | goto retry;
|
| 4833 | }
|
| 4834 | }
|
| 4835 |
|
| 4836 |
|
| 4837 | if (!can_direct_reclaim)
|
| 4838 | goto nopage;
|
| 4839 |
|
| 4840 |
|
| 4841 | if (current->flags & PF_MEMALLOC)
|
| 4842 | goto nopage;
|
| 4843 |
|
| 4844 |
|
| 4845 | if (!compact_first) {
|
| 4846 | page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags,
|
| 4847 | ac, &did_some_progress);
|
| 4848 | if (page)
|
| 4849 | goto got_pg;
|
| 4850 | }
|
| 4851 |
|
| 4852 |
|
| 4853 | page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
|
| 4854 | compact_priority, &compact_result);
|
| 4855 | if (page)
|
| 4856 | goto got_pg;
|
| 4857 |
|
| 4858 | if (compact_first) {
|
| 4859 | |
| 4860 | * THP page faults may attempt local node only first, but are
|
| 4861 | * then allowed to only compact, not reclaim, see
|
| 4862 | * alloc_pages_mpol().
|
| 4863 | *
|
| 4864 | * Compaction has failed above and we don't want such THP
|
| 4865 | * allocations to put reclaim pressure on a single node in a
|
| 4866 | * situation where other nodes might have plenty of available
|
| 4867 | * memory.
|
| 4868 |
|
| 4869 | if (gfp_has_flags(gfp_mask, __GFP_NORETRY | __GFP_THISNODE))
|
| 4870 | goto nopage;
|
| 4871 |
|
| 4872 | |
| 4873 | * For the initial compaction attempt we have lowered its
|
| 4874 | * priority. Restore it for further retries, if those are
|
| 4875 | * allowed. With __GFP_NORETRY there will be a single round of
|
| 4876 | * reclaim and compaction with the lowered priority.
|
| 4877 |
|
| 4878 | if (!(gfp_mask & __GFP_NORETRY))
|
| 4879 | compact_priority = DEF_COMPACT_PRIORITY;
|
| 4880 |
|
| 4881 | compact_first = false;
|
| 4882 | goto retry;
|
| 4883 | }
|
| 4884 |
|
| 4885 |
|
| 4886 | if (gfp_mask & __GFP_NORETRY)
|
| 4887 | goto nopage;
|
| 4888 |
|
| 4889 | |
| 4890 | * Do not retry costly high order allocations unless they are
|
| 4891 | * __GFP_RETRY_MAYFAIL and we can compact
|
| 4892 |
|
| 4893 | if (costly_order && (!can_compact ||
|
| 4894 | !(gfp_mask & __GFP_RETRY_MAYFAIL)))
|
| 4895 | goto nopage;
|
| 4896 |
|
| 4897 | |
| 4898 | * Deal with possible cpuset update races or zonelist updates to avoid
|
| 4899 | * infinite retries. No "goto retry;" can be placed above this check
|
| 4900 | * unless it can execute just once.
|
| 4901 |
|
| 4902 | if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
|
| 4903 | check_retry_zonelist(zonelist_iter_cookie))
|
| 4904 | goto restart;
|
| 4905 |
|
| 4906 | if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
|
| 4907 | did_some_progress > 0, &no_progress_loops))
|
| 4908 | goto retry;
|
| 4909 |
|
| 4910 | |
| 4911 | * It doesn't make any sense to retry for the compaction if the order-0
|
| 4912 | * reclaim is not able to make any progress because the current
|
| 4913 | * implementation of the compaction depends on the sufficient amount
|
| 4914 | * of free memory (see __compaction_suitable)
|
| 4915 |
|
| 4916 | if (did_some_progress > 0 && can_compact &&
|
| 4917 | should_compact_retry(ac, order, alloc_flags,
|
| 4918 | compact_result, &compact_priority,
|
| 4919 | &compaction_retries))
|
| 4920 | goto retry;
|
| 4921 |
|
| 4922 |
|
| 4923 | if (defrag_mode && (alloc_flags & ALLOC_NOFRAGMENT)) {
|
| 4924 | alloc_flags &= ~ALLOC_NOFRAGMENT;
|
| 4925 | goto retry;
|
| 4926 | }
|
| 4927 |
|
| 4928 | |
| 4929 | * Deal with possible cpuset update races or zonelist updates to avoid
|
| 4930 | * a unnecessary OOM kill.
|
| 4931 |
|
| 4932 | if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
|
| 4933 | check_retry_zonelist(zonelist_iter_cookie))
|
| 4934 | goto restart;
|
| 4935 |
|
| 4936 |
|
| 4937 | page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
|
| 4938 | if (page)
|
| 4939 | goto got_pg;
|
| 4940 |
|
| 4941 |
|
| 4942 | if (tsk_is_oom_victim(current) &&
|
| 4943 | (alloc_flags & ALLOC_OOM ||
|
| 4944 | (gfp_mask & __GFP_NOMEMALLOC)))
|
| 4945 | goto nopage;
|
| 4946 |
|
| 4947 |
|
| 4948 | if (did_some_progress) {
|
| 4949 | no_progress_loops = 0;
|
| 4950 | goto retry;
|
| 4951 | }
|
| 4952 |
|
| 4953 | nopage:
|
| 4954 | |
| 4955 | * Deal with possible cpuset update races or zonelist updates to avoid
|
| 4956 | * a unnecessary OOM kill.
|
| 4957 |
|
| 4958 | if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
|
| 4959 | check_retry_zonelist(zonelist_iter_cookie))
|
| 4960 | goto restart;
|
| 4961 |
|
| 4962 | |
| 4963 | * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
|
| 4964 | * we always retry
|
| 4965 |
|
| 4966 | if (unlikely(nofail)) {
|
| 4967 | |
| 4968 | * Lacking direct_reclaim we can't do anything to reclaim memory,
|
| 4969 | * we disregard these unreasonable nofail requests and still
|
| 4970 | * return NULL
|
| 4971 |
|
| 4972 | if (!can_direct_reclaim)
|
| 4973 | goto fail;
|
| 4974 |
|
| 4975 | |
| 4976 | * Help non-failing allocations by giving some access to memory
|
| 4977 | * reserves normally used for high priority non-blocking
|
| 4978 | * allocations but do not use ALLOC_NO_WATERMARKS because this
|
| 4979 | * could deplete whole memory reserves which would just make
|
| 4980 | * the situation worse.
|
| 4981 |
|
| 4982 | page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
|
| 4983 | if (page)
|
| 4984 | goto got_pg;
|
| 4985 |
|
| 4986 | cond_resched();
|
| 4987 | goto retry;
|
| 4988 | }
|
| 4989 | fail:
|
| 4990 | warn_alloc(gfp_mask, ac->nodemask,
|
| 4991 | "page allocation failure: order:%u", order);
|
| 4992 | got_pg:
|
| 4993 | return page;
|
| 4994 | }
|
| 4995 |
|
| 4996 | static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
|
| 4997 | int preferred_nid, nodemask_t *nodemask,
|
| 4998 | struct alloc_context *ac, gfp_t *alloc_gfp,
|
| 4999 | unsigned int *alloc_flags)
|
| 5000 | {
|
| 5001 | ac->highest_zoneidx = gfp_zone(gfp_mask);
|
| 5002 | ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
|
| 5003 | ac->nodemask = nodemask;
|
| 5004 | ac->migratetype = gfp_migratetype(gfp_mask);
|
| 5005 |
|
| 5006 | if (cpusets_enabled()) {
|
| 5007 | *alloc_gfp |= __GFP_HARDWALL;
|
| 5008 | |
| 5009 | * When we are in the interrupt context, it is irrelevant
|
| 5010 | * to the current task context. It means that any node ok.
|
| 5011 |
|
| 5012 | if (in_task() && !ac->nodemask)
|
| 5013 | ac->nodemask = &cpuset_current_mems_allowed;
|
| 5014 | else
|
| 5015 | *alloc_flags |= ALLOC_CPUSET;
|
| 5016 | }
|
| 5017 |
|
| 5018 | might_alloc(gfp_mask);
|
| 5019 |
|
| 5020 | |
| 5021 | * Don't invoke should_fail logic, since it may call
|
| 5022 | * get_random_u32() and printk() which need to spin_lock.
|
| 5023 |
|
| 5024 | if (!(*alloc_flags & ALLOC_TRYLOCK) &&
|
| 5025 | should_fail_alloc_page(gfp_mask, order))
|
| 5026 | return false;
|
| 5027 |
|
| 5028 | *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
|
| 5029 |
|
| 5030 |
|
| 5031 | ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
|
| 5032 |
|
| 5033 | |
| 5034 | * The preferred zone is used for statistics but crucially it is
|
| 5035 | * also used as the starting point for the zonelist iterator. It
|
| 5036 | * may get reset for allocations that ignore memory policies.
|
| 5037 |
|
| 5038 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
|
| 5039 | ac->highest_zoneidx, ac->nodemask);
|
| 5040 |
|
| 5041 | return true;
|
| 5042 | }
|
| 5043 |
|
| 5044 | |
| 5045 | * __alloc_pages_bulk - Allocate a number of order-0 pages to an array
|
| 5046 | * @gfp: GFP flags for the allocation
|
| 5047 | * @preferred_nid: The preferred NUMA node ID to allocate from
|
| 5048 | * @nodemask: Set of nodes to allocate from, may be NULL
|
| 5049 | * @nr_pages: The number of pages desired in the array
|
| 5050 | * @page_array: Array to store the pages
|
| 5051 | *
|
| 5052 | * This is a batched version of the page allocator that attempts to allocate
|
| 5053 | * @nr_pages quickly. Pages are added to @page_array.
|
| 5054 | *
|
| 5055 | * Note that only the elements in @page_array that were cleared to %NULL on
|
| 5056 | * entry are populated with newly allocated pages. @nr_pages is the maximum
|
| 5057 | * number of pages that will be stored in the array.
|
| 5058 | *
|
| 5059 | * Returns the number of pages in @page_array, including ones already
|
| 5060 | * allocated on entry. This can be less than the number requested in @nr_pages,
|
| 5061 | * but all empty slots are filled from the beginning. I.e., if all slots in
|
| 5062 | * @page_array were set to %NULL on entry, the slots from 0 to the return value
|
| 5063 | * - 1 will be filled.
|
| 5064 |
|
| 5065 | unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
|
| 5066 | nodemask_t *nodemask, int nr_pages,
|
| 5067 | struct page **page_array)
|
| 5068 | {
|
| 5069 | struct page *page;
|
| 5070 | unsigned long UP_flags;
|
| 5071 | struct zone *zone;
|
| 5072 | struct zoneref *z;
|
| 5073 | struct per_cpu_pages *pcp;
|
| 5074 | struct list_head *pcp_list;
|
| 5075 | struct alloc_context ac;
|
| 5076 | gfp_t alloc_gfp;
|
| 5077 | unsigned int alloc_flags = ALLOC_WMARK_LOW;
|
| 5078 | int nr_populated = 0, nr_account = 0;
|
| 5079 |
|
| 5080 | |
| 5081 | * Skip populated array elements to determine if any pages need
|
| 5082 | * to be allocated before disabling IRQs.
|
| 5083 |
|
| 5084 | while (nr_populated < nr_pages && page_array[nr_populated])
|
| 5085 | nr_populated++;
|
| 5086 |
|
| 5087 |
|
| 5088 | if (unlikely(nr_pages <= 0))
|
| 5089 | goto out;
|
| 5090 |
|
| 5091 |
|
| 5092 | if (unlikely(nr_pages - nr_populated == 0))
|
| 5093 | goto out;
|
| 5094 |
|
| 5095 |
|
| 5096 | if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
|
| 5097 | goto failed;
|
| 5098 |
|
| 5099 |
|
| 5100 | if (nr_pages - nr_populated == 1)
|
| 5101 | goto failed;
|
| 5102 |
|
| 5103 | #ifdef CONFIG_PAGE_OWNER
|
| 5104 | |
| 5105 | * PAGE_OWNER may recurse into the allocator to allocate space to
|
| 5106 | * save the stack with pagesets.lock held. Releasing/reacquiring
|
| 5107 | * removes much of the performance benefit of bulk allocation so
|
| 5108 | * force the caller to allocate one page at a time as it'll have
|
| 5109 | * similar performance to added complexity to the bulk allocator.
|
| 5110 |
|
| 5111 | if (static_branch_unlikely(&page_owner_inited))
|
| 5112 | goto failed;
|
| 5113 | #endif
|
| 5114 |
|
| 5115 |
|
| 5116 | gfp &= gfp_allowed_mask;
|
| 5117 | alloc_gfp = gfp;
|
| 5118 | if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
|
| 5119 | goto out;
|
| 5120 | gfp = alloc_gfp;
|
| 5121 |
|
| 5122 |
|
| 5123 | z = ac.preferred_zoneref;
|
| 5124 | for_next_zone_zonelist_nodemask(zone, z, ac.highest_zoneidx, ac.nodemask) {
|
| 5125 | unsigned long mark;
|
| 5126 |
|
| 5127 | if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
|
| 5128 | !__cpuset_zone_allowed(zone, gfp)) {
|
| 5129 | continue;
|
| 5130 | }
|
| 5131 |
|
| 5132 | if (nr_online_nodes > 1 && zone != zonelist_zone(ac.preferred_zoneref) &&
|
| 5133 | zone_to_nid(zone) != zonelist_node_idx(ac.preferred_zoneref)) {
|
| 5134 | goto failed;
|
| 5135 | }
|
| 5136 |
|
| 5137 | cond_accept_memory(zone, 0, alloc_flags);
|
| 5138 | retry_this_zone:
|
| 5139 | mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
|
| 5140 | if (zone_watermark_fast(zone, 0, mark,
|
| 5141 | zonelist_zone_idx(ac.preferred_zoneref),
|
| 5142 | alloc_flags, gfp)) {
|
| 5143 | break;
|
| 5144 | }
|
| 5145 |
|
| 5146 | if (cond_accept_memory(zone, 0, alloc_flags))
|
| 5147 | goto retry_this_zone;
|
| 5148 |
|
| 5149 |
|
| 5150 | if (deferred_pages_enabled()) {
|
| 5151 | if (_deferred_grow_zone(zone, 0))
|
| 5152 | goto retry_this_zone;
|
| 5153 | }
|
| 5154 | }
|
| 5155 |
|
| 5156 | |
| 5157 | * If there are no allowed local zones that meets the watermarks then
|
| 5158 | * try to allocate a single page and reclaim if necessary.
|
| 5159 |
|
| 5160 | if (unlikely(!zone))
|
| 5161 | goto failed;
|
| 5162 |
|
| 5163 |
|
| 5164 | pcp = pcp_spin_trylock(zone->per_cpu_pageset, UP_flags);
|
| 5165 | if (!pcp)
|
| 5166 | goto failed;
|
| 5167 |
|
| 5168 |
|
| 5169 | pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
|
| 5170 | while (nr_populated < nr_pages) {
|
| 5171 |
|
| 5172 |
|
| 5173 | if (page_array[nr_populated]) {
|
| 5174 | nr_populated++;
|
| 5175 | continue;
|
| 5176 | }
|
| 5177 |
|
| 5178 | page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
|
| 5179 | pcp, pcp_list);
|
| 5180 | if (unlikely(!page)) {
|
| 5181 |
|
| 5182 | if (!nr_account) {
|
| 5183 | pcp_spin_unlock(pcp, UP_flags);
|
| 5184 | goto failed;
|
| 5185 | }
|
| 5186 | break;
|
| 5187 | }
|
| 5188 | nr_account++;
|
| 5189 |
|
| 5190 | prep_new_page(page, 0, gfp, 0);
|
| 5191 | set_page_refcounted(page);
|
| 5192 | page_array[nr_populated++] = page;
|
| 5193 | }
|
| 5194 |
|
| 5195 | pcp_spin_unlock(pcp, UP_flags);
|
| 5196 |
|
| 5197 | __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
|
| 5198 | zone_statistics(zonelist_zone(ac.preferred_zoneref), zone, nr_account);
|
| 5199 |
|
| 5200 | out:
|
| 5201 | return nr_populated;
|
| 5202 |
|
| 5203 | failed:
|
| 5204 | page = __alloc_pages_noprof(gfp, 0, preferred_nid, nodemask);
|
| 5205 | if (page)
|
| 5206 | page_array[nr_populated++] = page;
|
| 5207 | goto out;
|
| 5208 | }
|
| 5209 | EXPORT_SYMBOL_GPL(alloc_pages_bulk_noprof);
|
| 5210 |
|
| 5211 | |
| 5212 | * This is the 'heart' of the zoned buddy allocator.
|
| 5213 |
|
| 5214 | struct page *__alloc_frozen_pages_noprof(gfp_t gfp, unsigned int order,
|
| 5215 | int preferred_nid, nodemask_t *nodemask)
|
| 5216 | {
|
| 5217 | struct page *page;
|
| 5218 | unsigned int alloc_flags = ALLOC_WMARK_LOW;
|
| 5219 | gfp_t alloc_gfp;
|
| 5220 | struct alloc_context ac = { };
|
| 5221 |
|
| 5222 | |
| 5223 | * There are several places where we assume that the order value is sane
|
| 5224 | * so bail out early if the request is out of bound.
|
| 5225 |
|
| 5226 | if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
|
| 5227 | return NULL;
|
| 5228 |
|
| 5229 | gfp &= gfp_allowed_mask;
|
| 5230 | |
| 5231 | * Apply scoped allocation constraints. This is mainly about GFP_NOFS
|
| 5232 | * resp. GFP_NOIO which has to be inherited for all allocation requests
|
| 5233 | * from a particular context which has been marked by
|
| 5234 | * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
|
| 5235 | * movable zones are not used during allocation.
|
| 5236 |
|
| 5237 | gfp = current_gfp_context(gfp);
|
| 5238 | alloc_gfp = gfp;
|
| 5239 | if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
|
| 5240 | &alloc_gfp, &alloc_flags))
|
| 5241 | return NULL;
|
| 5242 |
|
| 5243 | |
| 5244 | * Forbid the first pass from falling back to types that fragment
|
| 5245 | * memory until all local zones are considered.
|
| 5246 |
|
| 5247 | alloc_flags |= alloc_flags_nofragment(zonelist_zone(ac.preferred_zoneref), gfp);
|
| 5248 |
|
| 5249 |
|
| 5250 | page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
|
| 5251 | if (likely(page))
|
| 5252 | goto out;
|
| 5253 |
|
| 5254 | alloc_gfp = gfp;
|
| 5255 | ac.spread_dirty_pages = false;
|
| 5256 |
|
| 5257 | |
| 5258 | * Restore the original nodemask if it was potentially replaced with
|
| 5259 | * &cpuset_current_mems_allowed to optimize the fast-path attempt.
|
| 5260 |
|
| 5261 | ac.nodemask = nodemask;
|
| 5262 |
|
| 5263 | page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
|
| 5264 |
|
| 5265 | out:
|
| 5266 | if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
|
| 5267 | unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
|
| 5268 | free_frozen_pages(page, order);
|
| 5269 | page = NULL;
|
| 5270 | }
|
| 5271 |
|
| 5272 | trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
|
| 5273 | kmsan_alloc_page(page, order, alloc_gfp);
|
| 5274 |
|
| 5275 | return page;
|
| 5276 | }
|
| 5277 | EXPORT_SYMBOL(__alloc_frozen_pages_noprof);
|
| 5278 |
|
| 5279 | struct page *__alloc_pages_noprof(gfp_t gfp, unsigned int order,
|
| 5280 | int preferred_nid, nodemask_t *nodemask)
|
| 5281 | {
|
| 5282 | struct page *page;
|
| 5283 |
|
| 5284 | page = __alloc_frozen_pages_noprof(gfp, order, preferred_nid, nodemask);
|
| 5285 | if (page)
|
| 5286 | set_page_refcounted(page);
|
| 5287 | return page;
|
| 5288 | }
|
| 5289 | EXPORT_SYMBOL(__alloc_pages_noprof);
|
| 5290 |
|
| 5291 | struct folio *__folio_alloc_noprof(gfp_t gfp, unsigned int order, int preferred_nid,
|
| 5292 | nodemask_t *nodemask)
|
| 5293 | {
|
| 5294 | struct page *page = __alloc_pages_noprof(gfp | __GFP_COMP, order,
|
| 5295 | preferred_nid, nodemask);
|
| 5296 | return page_rmappable_folio(page);
|
| 5297 | }
|
| 5298 | EXPORT_SYMBOL(__folio_alloc_noprof);
|
| 5299 |
|
| 5300 | |
| 5301 | * Common helper functions. Never use with __GFP_HIGHMEM because the returned
|
| 5302 | * address cannot represent highmem pages. Use alloc_pages and then kmap if
|
| 5303 | * you need to access high mem.
|
| 5304 |
|
| 5305 | unsigned long get_free_pages_noprof(gfp_t gfp_mask, unsigned int order)
|
| 5306 | {
|
| 5307 | struct page *page;
|
| 5308 |
|
| 5309 | page = alloc_pages_noprof(gfp_mask & ~__GFP_HIGHMEM, order);
|
| 5310 | if (!page)
|
| 5311 | return 0;
|
| 5312 | return (unsigned long) page_address(page);
|
| 5313 | }
|
| 5314 | EXPORT_SYMBOL(get_free_pages_noprof);
|
| 5315 |
|
| 5316 | unsigned long get_zeroed_page_noprof(gfp_t gfp_mask)
|
| 5317 | {
|
| 5318 | return get_free_pages_noprof(gfp_mask | __GFP_ZERO, 0);
|
| 5319 | }
|
| 5320 | EXPORT_SYMBOL(get_zeroed_page_noprof);
|
| 5321 |
|
| 5322 | static void ___free_pages(struct page *page, unsigned int order,
|
| 5323 | fpi_t fpi_flags)
|
| 5324 | {
|
| 5325 |
|
| 5326 | int head = PageHead(page);
|
| 5327 |
|
| 5328 | struct alloc_tag *tag = pgalloc_tag_get(page);
|
| 5329 |
|
| 5330 | if (put_page_testzero(page))
|
| 5331 | __free_frozen_pages(page, order, fpi_flags);
|
| 5332 | else if (!head) {
|
| 5333 | pgalloc_tag_sub_pages(tag, (1 << order) - 1);
|
| 5334 | while (order-- > 0) {
|
| 5335 | |
| 5336 | * The "tail" pages of this non-compound high-order
|
| 5337 | * page will have no code tags, so to avoid warnings
|
| 5338 | * mark them as empty.
|
| 5339 |
|
| 5340 | clear_page_tag_ref(page + (1 << order));
|
| 5341 | __free_frozen_pages(page + (1 << order), order,
|
| 5342 | fpi_flags);
|
| 5343 | }
|
| 5344 | }
|
| 5345 | }
|
| 5346 |
|
| 5347 | |
| 5348 | * __free_pages - Free pages allocated with alloc_pages().
|
| 5349 | * @page: The page pointer returned from alloc_pages().
|
| 5350 | * @order: The order of the allocation.
|
| 5351 | *
|
| 5352 | * This function can free multi-page allocations that are not compound
|
| 5353 | * pages. It does not check that the @order passed in matches that of
|
| 5354 | * the allocation, so it is easy to leak memory. Freeing more memory
|
| 5355 | * than was allocated will probably emit a warning.
|
| 5356 | *
|
| 5357 | * If the last reference to this page is speculative, it will be released
|
| 5358 | * by put_page() which only frees the first page of a non-compound
|
| 5359 | * allocation. To prevent the remaining pages from being leaked, we free
|
| 5360 | * the subsequent pages here. If you want to use the page's reference
|
| 5361 | * count to decide when to free the allocation, you should allocate a
|
| 5362 | * compound page, and use put_page() instead of __free_pages().
|
| 5363 | *
|
| 5364 | * Context: May be called in interrupt context or while holding a normal
|
| 5365 | * spinlock, but not in NMI context or while holding a raw spinlock.
|
| 5366 |
|
| 5367 | void __free_pages(struct page *page, unsigned int order)
|
| 5368 | {
|
| 5369 | ___free_pages(page, order, FPI_NONE);
|
| 5370 | }
|
| 5371 | EXPORT_SYMBOL(__free_pages);
|
| 5372 |
|
| 5373 | |
| 5374 | * Can be called while holding raw_spin_lock or from IRQ and NMI for any
|
| 5375 | * page type (not only those that came from alloc_pages_nolock)
|
| 5376 |
|
| 5377 | void free_pages_nolock(struct page *page, unsigned int order)
|
| 5378 | {
|
| 5379 | ___free_pages(page, order, FPI_TRYLOCK);
|
| 5380 | }
|
| 5381 |
|
| 5382 | |
| 5383 | * free_pages - Free pages allocated with __get_free_pages().
|
| 5384 | * @addr: The virtual address tied to a page returned from __get_free_pages().
|
| 5385 | * @order: The order of the allocation.
|
| 5386 | *
|
| 5387 | * This function behaves the same as __free_pages(). Use this function
|
| 5388 | * to free pages when you only have a valid virtual address. If you have
|
| 5389 | * the page, call __free_pages() instead.
|
| 5390 |
|
| 5391 | void free_pages(unsigned long addr, unsigned int order)
|
| 5392 | {
|
| 5393 | if (addr != 0) {
|
| 5394 | VM_BUG_ON(!virt_addr_valid((void *)addr));
|
| 5395 | __free_pages(virt_to_page((void *)addr), order);
|
| 5396 | }
|
| 5397 | }
|
| 5398 |
|
| 5399 | EXPORT_SYMBOL(free_pages);
|
| 5400 |
|
| 5401 | static void *make_alloc_exact(unsigned long addr, unsigned int order,
|
| 5402 | size_t size)
|
| 5403 | {
|
| 5404 | if (addr) {
|
| 5405 | unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
|
| 5406 | struct page *page = virt_to_page((void *)addr);
|
| 5407 | struct page *last = page + nr;
|
| 5408 |
|
| 5409 | __split_page(page, order);
|
| 5410 | while (page < --last)
|
| 5411 | set_page_refcounted(last);
|
| 5412 |
|
| 5413 | last = page + (1UL << order);
|
| 5414 | for (page += nr; page < last; page++)
|
| 5415 | __free_pages_ok(page, 0, FPI_TO_TAIL);
|
| 5416 | }
|
| 5417 | return (void *)addr;
|
| 5418 | }
|
| 5419 |
|
| 5420 | |
| 5421 | * alloc_pages_exact - allocate an exact number physically-contiguous pages.
|
| 5422 | * @size: the number of bytes to allocate
|
| 5423 | * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
|
| 5424 | *
|
| 5425 | * This function is similar to alloc_pages(), except that it allocates the
|
| 5426 | * minimum number of pages to satisfy the request. alloc_pages() can only
|
| 5427 | * allocate memory in power-of-two pages.
|
| 5428 | *
|
| 5429 | * This function is also limited by MAX_PAGE_ORDER.
|
| 5430 | *
|
| 5431 | * Memory allocated by this function must be released by free_pages_exact().
|
| 5432 | *
|
| 5433 | * Return: pointer to the allocated area or %NULL in case of error.
|
| 5434 |
|
| 5435 | void *alloc_pages_exact_noprof(size_t size, gfp_t gfp_mask)
|
| 5436 | {
|
| 5437 | unsigned int order = get_order(size);
|
| 5438 | unsigned long addr;
|
| 5439 |
|
| 5440 | if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
|
| 5441 | gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
|
| 5442 |
|
| 5443 | addr = get_free_pages_noprof(gfp_mask, order);
|
| 5444 | return make_alloc_exact(addr, order, size);
|
| 5445 | }
|
| 5446 | EXPORT_SYMBOL(alloc_pages_exact_noprof);
|
| 5447 |
|
| 5448 | |
| 5449 | * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
|
| 5450 | * pages on a node.
|
| 5451 | * @nid: the preferred node ID where memory should be allocated
|
| 5452 | * @size: the number of bytes to allocate
|
| 5453 | * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
|
| 5454 | *
|
| 5455 | * Like alloc_pages_exact(), but try to allocate on node nid first before falling
|
| 5456 | * back.
|
| 5457 | *
|
| 5458 | * Return: pointer to the allocated area or %NULL in case of error.
|
| 5459 |
|
| 5460 | void * __meminit alloc_pages_exact_nid_noprof(int nid, size_t size, gfp_t gfp_mask)
|
| 5461 | {
|
| 5462 | unsigned int order = get_order(size);
|
| 5463 | struct page *p;
|
| 5464 |
|
| 5465 | if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
|
| 5466 | gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
|
| 5467 |
|
| 5468 | p = alloc_pages_node_noprof(nid, gfp_mask, order);
|
| 5469 | if (!p)
|
| 5470 | return NULL;
|
| 5471 | return make_alloc_exact((unsigned long)page_address(p), order, size);
|
| 5472 | }
|
| 5473 |
|
| 5474 | |
| 5475 | * free_pages_exact - release memory allocated via alloc_pages_exact()
|
| 5476 | * @virt: the value returned by alloc_pages_exact.
|
| 5477 | * @size: size of allocation, same value as passed to alloc_pages_exact().
|
| 5478 | *
|
| 5479 | * Release the memory allocated by a previous call to alloc_pages_exact.
|
| 5480 |
|
| 5481 | void free_pages_exact(void *virt, size_t size)
|
| 5482 | {
|
| 5483 | unsigned long addr = (unsigned long)virt;
|
| 5484 | unsigned long end = addr + PAGE_ALIGN(size);
|
| 5485 |
|
| 5486 | while (addr < end) {
|
| 5487 | free_page(addr);
|
| 5488 | addr += PAGE_SIZE;
|
| 5489 | }
|
| 5490 | }
|
| 5491 | EXPORT_SYMBOL(free_pages_exact);
|
| 5492 |
|
| 5493 | |
| 5494 | * nr_free_zone_pages - count number of pages beyond high watermark
|
| 5495 | * @offset: The zone index of the highest zone
|
| 5496 | *
|
| 5497 | * nr_free_zone_pages() counts the number of pages which are beyond the
|
| 5498 | * high watermark within all zones at or below a given zone index. For each
|
| 5499 | * zone, the number of pages is calculated as:
|
| 5500 | *
|
| 5501 | * nr_free_zone_pages = managed_pages - high_pages
|
| 5502 | *
|
| 5503 | * Return: number of pages beyond high watermark.
|
| 5504 |
|
| 5505 | static unsigned long nr_free_zone_pages(int offset)
|
| 5506 | {
|
| 5507 | struct zoneref *z;
|
| 5508 | struct zone *zone;
|
| 5509 |
|
| 5510 |
|
| 5511 | unsigned long sum = 0;
|
| 5512 |
|
| 5513 | struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
|
| 5514 |
|
| 5515 | for_each_zone_zonelist(zone, z, zonelist, offset) {
|
| 5516 | unsigned long size = zone_managed_pages(zone);
|
| 5517 | unsigned long high = high_wmark_pages(zone);
|
| 5518 | if (size > high)
|
| 5519 | sum += size - high;
|
| 5520 | }
|
| 5521 |
|
| 5522 | return sum;
|
| 5523 | }
|
| 5524 |
|
| 5525 | |
| 5526 | * nr_free_buffer_pages - count number of pages beyond high watermark
|
| 5527 | *
|
| 5528 | * nr_free_buffer_pages() counts the number of pages which are beyond the high
|
| 5529 | * watermark within ZONE_DMA and ZONE_NORMAL.
|
| 5530 | *
|
| 5531 | * Return: number of pages beyond high watermark within ZONE_DMA and
|
| 5532 | * ZONE_NORMAL.
|
| 5533 |
|
| 5534 | unsigned long nr_free_buffer_pages(void)
|
| 5535 | {
|
| 5536 | return nr_free_zone_pages(gfp_zone(GFP_USER));
|
| 5537 | }
|
| 5538 | EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
|
| 5539 |
|
| 5540 | static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
|
| 5541 | {
|
| 5542 | zoneref->zone = zone;
|
| 5543 | zoneref->zone_idx = zone_idx(zone);
|
| 5544 | }
|
| 5545 |
|
| 5546 | |
| 5547 | * Builds allocation fallback zone lists.
|
| 5548 | *
|
| 5549 | * Add all populated zones of a node to the zonelist.
|
| 5550 |
|
| 5551 | static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
|
| 5552 | {
|
| 5553 | struct zone *zone;
|
| 5554 | enum zone_type zone_type = MAX_NR_ZONES;
|
| 5555 | int nr_zones = 0;
|
| 5556 |
|
| 5557 | do {
|
| 5558 | zone_type--;
|
| 5559 | zone = pgdat->node_zones + zone_type;
|
| 5560 | if (populated_zone(zone)) {
|
| 5561 | zoneref_set_zone(zone, &zonerefs[nr_zones++]);
|
| 5562 | check_highest_zone(zone_type);
|
| 5563 | }
|
| 5564 | } while (zone_type);
|
| 5565 |
|
| 5566 | return nr_zones;
|
| 5567 | }
|
| 5568 |
|
| 5569 | #ifdef CONFIG_NUMA
|
| 5570 |
|
| 5571 | static int __parse_numa_zonelist_order(char *s)
|
| 5572 | {
|
| 5573 | |
| 5574 | * We used to support different zonelists modes but they turned
|
| 5575 | * out to be just not useful. Let's keep the warning in place
|
| 5576 | * if somebody still use the cmd line parameter so that we do
|
| 5577 | * not fail it silently
|
| 5578 |
|
| 5579 | if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
|
| 5580 | pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
|
| 5581 | return -EINVAL;
|
| 5582 | }
|
| 5583 | return 0;
|
| 5584 | }
|
| 5585 |
|
| 5586 | static char numa_zonelist_order[] = "Node";
|
| 5587 | #define NUMA_ZONELIST_ORDER_LEN 16
|
| 5588 | |
| 5589 | * sysctl handler for numa_zonelist_order
|
| 5590 |
|
| 5591 | static int numa_zonelist_order_handler(const struct ctl_table *table, int write,
|
| 5592 | void *buffer, size_t *length, loff_t *ppos)
|
| 5593 | {
|
| 5594 | if (write)
|
| 5595 | return __parse_numa_zonelist_order(buffer);
|
| 5596 | return proc_dostring(table, write, buffer, length, ppos);
|
| 5597 | }
|
| 5598 |
|
| 5599 | static int node_load[MAX_NUMNODES];
|
| 5600 |
|
| 5601 | |
| 5602 | * find_next_best_node - find the next node that should appear in a given node's fallback list
|
| 5603 | * @node: node whose fallback list we're appending
|
| 5604 | * @used_node_mask: nodemask_t of already used nodes
|
| 5605 | *
|
| 5606 | * We use a number of factors to determine which is the next node that should
|
| 5607 | * appear on a given node's fallback list. The node should not have appeared
|
| 5608 | * already in @node's fallback list, and it should be the next closest node
|
| 5609 | * according to the distance array (which contains arbitrary distance values
|
| 5610 | * from each node to each node in the system), and should also prefer nodes
|
| 5611 | * with no CPUs, since presumably they'll have very little allocation pressure
|
| 5612 | * on them otherwise.
|
| 5613 | *
|
| 5614 | * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
|
| 5615 |
|
| 5616 | int find_next_best_node(int node, nodemask_t *used_node_mask)
|
| 5617 | {
|
| 5618 | int n, val;
|
| 5619 | int min_val = INT_MAX;
|
| 5620 | int best_node = NUMA_NO_NODE;
|
| 5621 |
|
| 5622 | |
| 5623 | * Use the local node if we haven't already, but for memoryless local
|
| 5624 | * node, we should skip it and fall back to other nodes.
|
| 5625 |
|
| 5626 | if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
|
| 5627 | node_set(node, *used_node_mask);
|
| 5628 | return node;
|
| 5629 | }
|
| 5630 |
|
| 5631 | for_each_node_state(n, N_MEMORY) {
|
| 5632 |
|
| 5633 |
|
| 5634 | if (node_isset(n, *used_node_mask))
|
| 5635 | continue;
|
| 5636 |
|
| 5637 |
|
| 5638 | val = node_distance(node, n);
|
| 5639 |
|
| 5640 |
|
| 5641 | val += (n < node);
|
| 5642 |
|
| 5643 |
|
| 5644 | if (!cpumask_empty(cpumask_of_node(n)))
|
| 5645 | val += PENALTY_FOR_NODE_WITH_CPUS;
|
| 5646 |
|
| 5647 |
|
| 5648 | val *= MAX_NUMNODES;
|
| 5649 | val += node_load[n];
|
| 5650 |
|
| 5651 | if (val < min_val) {
|
| 5652 | min_val = val;
|
| 5653 | best_node = n;
|
| 5654 | }
|
| 5655 | }
|
| 5656 |
|
| 5657 | if (best_node >= 0)
|
| 5658 | node_set(best_node, *used_node_mask);
|
| 5659 |
|
| 5660 | return best_node;
|
| 5661 | }
|
| 5662 |
|
| 5663 |
|
| 5664 | |
| 5665 | * Build zonelists ordered by node and zones within node.
|
| 5666 | * This results in maximum locality--normal zone overflows into local
|
| 5667 | * DMA zone, if any--but risks exhausting DMA zone.
|
| 5668 |
|
| 5669 | static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
|
| 5670 | unsigned nr_nodes)
|
| 5671 | {
|
| 5672 | struct zoneref *zonerefs;
|
| 5673 | int i;
|
| 5674 |
|
| 5675 | zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
|
| 5676 |
|
| 5677 | for (i = 0; i < nr_nodes; i++) {
|
| 5678 | int nr_zones;
|
| 5679 |
|
| 5680 | pg_data_t *node = NODE_DATA(node_order[i]);
|
| 5681 |
|
| 5682 | nr_zones = build_zonerefs_node(node, zonerefs);
|
| 5683 | zonerefs += nr_zones;
|
| 5684 | }
|
| 5685 | zonerefs->zone = NULL;
|
| 5686 | zonerefs->zone_idx = 0;
|
| 5687 | }
|
| 5688 |
|
| 5689 | |
| 5690 | * Build __GFP_THISNODE zonelists
|
| 5691 |
|
| 5692 | static void build_thisnode_zonelists(pg_data_t *pgdat)
|
| 5693 | {
|
| 5694 | struct zoneref *zonerefs;
|
| 5695 | int nr_zones;
|
| 5696 |
|
| 5697 | zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
|
| 5698 | nr_zones = build_zonerefs_node(pgdat, zonerefs);
|
| 5699 | zonerefs += nr_zones;
|
| 5700 | zonerefs->zone = NULL;
|
| 5701 | zonerefs->zone_idx = 0;
|
| 5702 | }
|
| 5703 |
|
| 5704 | static void build_zonelists(pg_data_t *pgdat)
|
| 5705 | {
|
| 5706 | static int node_order[MAX_NUMNODES];
|
| 5707 | int node, nr_nodes = 0;
|
| 5708 | nodemask_t used_mask = NODE_MASK_NONE;
|
| 5709 | int local_node, prev_node;
|
| 5710 |
|
| 5711 |
|
| 5712 | local_node = pgdat->node_id;
|
| 5713 | prev_node = local_node;
|
| 5714 |
|
| 5715 | memset(node_order, 0, sizeof(node_order));
|
| 5716 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
|
| 5717 | |
| 5718 | * We don't want to pressure a particular node.
|
| 5719 | * So adding penalty to the first node in same
|
| 5720 | * distance group to make it round-robin.
|
| 5721 |
|
| 5722 | if (node_distance(local_node, node) !=
|
| 5723 | node_distance(local_node, prev_node))
|
| 5724 | node_load[node] += 1;
|
| 5725 |
|
| 5726 | node_order[nr_nodes++] = node;
|
| 5727 | prev_node = node;
|
| 5728 | }
|
| 5729 |
|
| 5730 | build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
|
| 5731 | build_thisnode_zonelists(pgdat);
|
| 5732 | pr_info("Fallback order for Node %d: ", local_node);
|
| 5733 | for (node = 0; node < nr_nodes; node++)
|
| 5734 | pr_cont("%d ", node_order[node]);
|
| 5735 | pr_cont("\n");
|
| 5736 | }
|
| 5737 |
|
| 5738 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES
|
| 5739 | |
| 5740 | * Return node id of node used for "local" allocations.
|
| 5741 | * I.e., first node id of first zone in arg node's generic zonelist.
|
| 5742 | * Used for initializing percpu 'numa_mem', which is used primarily
|
| 5743 | * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
|
| 5744 |
|
| 5745 | int local_memory_node(int node)
|
| 5746 | {
|
| 5747 | struct zoneref *z;
|
| 5748 |
|
| 5749 | z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
|
| 5750 | gfp_zone(GFP_KERNEL),
|
| 5751 | NULL);
|
| 5752 | return zonelist_node_idx(z);
|
| 5753 | }
|
| 5754 | #endif
|
| 5755 |
|
| 5756 | static void setup_min_unmapped_ratio(void);
|
| 5757 | static void setup_min_slab_ratio(void);
|
| 5758 | #else
|
| 5759 |
|
| 5760 | static void build_zonelists(pg_data_t *pgdat)
|
| 5761 | {
|
| 5762 | struct zoneref *zonerefs;
|
| 5763 | int nr_zones;
|
| 5764 |
|
| 5765 | zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
|
| 5766 | nr_zones = build_zonerefs_node(pgdat, zonerefs);
|
| 5767 | zonerefs += nr_zones;
|
| 5768 |
|
| 5769 | zonerefs->zone = NULL;
|
| 5770 | zonerefs->zone_idx = 0;
|
| 5771 | }
|
| 5772 |
|
| 5773 | #endif
|
| 5774 |
|
| 5775 | |
| 5776 | * Boot pageset table. One per cpu which is going to be used for all
|
| 5777 | * zones and all nodes. The parameters will be set in such a way
|
| 5778 | * that an item put on a list will immediately be handed over to
|
| 5779 | * the buddy list. This is safe since pageset manipulation is done
|
| 5780 | * with interrupts disabled.
|
| 5781 | *
|
| 5782 | * The boot_pagesets must be kept even after bootup is complete for
|
| 5783 | * unused processors and/or zones. They do play a role for bootstrapping
|
| 5784 | * hotplugged processors.
|
| 5785 | *
|
| 5786 | * zoneinfo_show() and maybe other functions do
|
| 5787 | * not check if the processor is online before following the pageset pointer.
|
| 5788 | * Other parts of the kernel may not check if the zone is available.
|
| 5789 |
|
| 5790 | static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
|
| 5791 |
|
| 5792 | #define BOOT_PAGESET_HIGH 0
|
| 5793 | #define BOOT_PAGESET_BATCH 1
|
| 5794 | static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
|
| 5795 | static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
|
| 5796 |
|
| 5797 | static void __build_all_zonelists(void *data)
|
| 5798 | {
|
| 5799 | int nid;
|
| 5800 | int __maybe_unused cpu;
|
| 5801 | pg_data_t *self = data;
|
| 5802 | unsigned long flags;
|
| 5803 |
|
| 5804 | |
| 5805 | * The zonelist_update_seq must be acquired with irqsave because the
|
| 5806 | * reader can be invoked from IRQ with GFP_ATOMIC.
|
| 5807 |
|
| 5808 | write_seqlock_irqsave(&zonelist_update_seq, flags);
|
| 5809 | |
| 5810 | * Also disable synchronous printk() to prevent any printk() from
|
| 5811 | * trying to hold port->lock, for
|
| 5812 | * tty_insert_flip_string_and_push_buffer() on other CPU might be
|
| 5813 | * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
|
| 5814 |
|
| 5815 | printk_deferred_enter();
|
| 5816 |
|
| 5817 | #ifdef CONFIG_NUMA
|
| 5818 | memset(node_load, 0, sizeof(node_load));
|
| 5819 | #endif
|
| 5820 |
|
| 5821 | |
| 5822 | * This node is hotadded and no memory is yet present. So just
|
| 5823 | * building zonelists is fine - no need to touch other nodes.
|
| 5824 |
|
| 5825 | if (self && !node_online(self->node_id)) {
|
| 5826 | build_zonelists(self);
|
| 5827 | } else {
|
| 5828 | |
| 5829 | * All possible nodes have pgdat preallocated
|
| 5830 | * in free_area_init
|
| 5831 |
|
| 5832 | for_each_node(nid) {
|
| 5833 | pg_data_t *pgdat = NODE_DATA(nid);
|
| 5834 |
|
| 5835 | build_zonelists(pgdat);
|
| 5836 | }
|
| 5837 |
|
| 5838 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES
|
| 5839 | |
| 5840 | * We now know the "local memory node" for each node--
|
| 5841 | * i.e., the node of the first zone in the generic zonelist.
|
| 5842 | * Set up numa_mem percpu variable for on-line cpus. During
|
| 5843 | * boot, only the boot cpu should be on-line; we'll init the
|
| 5844 | * secondary cpus' numa_mem as they come on-line. During
|
| 5845 | * node/memory hotplug, we'll fixup all on-line cpus.
|
| 5846 |
|
| 5847 | for_each_online_cpu(cpu)
|
| 5848 | set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
|
| 5849 | #endif
|
| 5850 | }
|
| 5851 |
|
| 5852 | printk_deferred_exit();
|
| 5853 | write_sequnlock_irqrestore(&zonelist_update_seq, flags);
|
| 5854 | }
|
| 5855 |
|
| 5856 | static noinline void __init
|
| 5857 | build_all_zonelists_init(void)
|
| 5858 | {
|
| 5859 | int cpu;
|
| 5860 |
|
| 5861 | __build_all_zonelists(NULL);
|
| 5862 |
|
| 5863 | |
| 5864 | * Initialize the boot_pagesets that are going to be used
|
| 5865 | * for bootstrapping processors. The real pagesets for
|
| 5866 | * each zone will be allocated later when the per cpu
|
| 5867 | * allocator is available.
|
| 5868 | *
|
| 5869 | * boot_pagesets are used also for bootstrapping offline
|
| 5870 | * cpus if the system is already booted because the pagesets
|
| 5871 | * are needed to initialize allocators on a specific cpu too.
|
| 5872 | * F.e. the percpu allocator needs the page allocator which
|
| 5873 | * needs the percpu allocator in order to allocate its pagesets
|
| 5874 | * (a chicken-egg dilemma).
|
| 5875 |
|
| 5876 | for_each_possible_cpu(cpu)
|
| 5877 | per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
|
| 5878 |
|
| 5879 | mminit_verify_zonelist();
|
| 5880 | cpuset_init_current_mems_allowed();
|
| 5881 | }
|
| 5882 |
|
| 5883 | |
| 5884 | * unless system_state == SYSTEM_BOOTING.
|
| 5885 | *
|
| 5886 | * __ref due to call of __init annotated helper build_all_zonelists_init
|
| 5887 | * [protected by SYSTEM_BOOTING].
|
| 5888 |
|
| 5889 | void __ref build_all_zonelists(pg_data_t *pgdat)
|
| 5890 | {
|
| 5891 | unsigned long vm_total_pages;
|
| 5892 |
|
| 5893 | if (system_state == SYSTEM_BOOTING) {
|
| 5894 | build_all_zonelists_init();
|
| 5895 | } else {
|
| 5896 | __build_all_zonelists(pgdat);
|
| 5897 |
|
| 5898 | }
|
| 5899 |
|
| 5900 | vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
|
| 5901 | |
| 5902 | * Disable grouping by mobility if the number of pages in the
|
| 5903 | * system is too low to allow the mechanism to work. It would be
|
| 5904 | * more accurate, but expensive to check per-zone. This check is
|
| 5905 | * made on memory-hotadd so a system can start with mobility
|
| 5906 | * disabled and enable it later
|
| 5907 |
|
| 5908 | if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
|
| 5909 | page_group_by_mobility_disabled = 1;
|
| 5910 | else
|
| 5911 | page_group_by_mobility_disabled = 0;
|
| 5912 |
|
| 5913 | pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
|
| 5914 | nr_online_nodes,
|
| 5915 | str_off_on(page_group_by_mobility_disabled),
|
| 5916 | vm_total_pages);
|
| 5917 | #ifdef CONFIG_NUMA
|
| 5918 | pr_info("Policy zone: %s\n", zone_names[policy_zone]);
|
| 5919 | #endif
|
| 5920 | }
|
| 5921 |
|
| 5922 | static int zone_batchsize(struct zone *zone)
|
| 5923 | {
|
| 5924 | #ifdef CONFIG_MMU
|
| 5925 | int batch;
|
| 5926 |
|
| 5927 | |
| 5928 | * The number of pages to batch allocate is either ~0.025%
|
| 5929 | * of the zone or 256KB, whichever is smaller. The batch
|
| 5930 | * size is striking a balance between allocation latency
|
| 5931 | * and zone lock contention.
|
| 5932 |
|
| 5933 | batch = min(zone_managed_pages(zone) >> 12, SZ_256K / PAGE_SIZE);
|
| 5934 | if (batch <= 1)
|
| 5935 | return 1;
|
| 5936 |
|
| 5937 | |
| 5938 | * Clamp the batch to a 2^n - 1 value. Having a power
|
| 5939 | * of 2 value was found to be more likely to have
|
| 5940 | * suboptimal cache aliasing properties in some cases.
|
| 5941 | *
|
| 5942 | * For example if 2 tasks are alternately allocating
|
| 5943 | * batches of pages, one task can end up with a lot
|
| 5944 | * of pages of one half of the possible page colors
|
| 5945 | * and the other with pages of the other colors.
|
| 5946 |
|
| 5947 | batch = rounddown_pow_of_two(batch + batch/2) - 1;
|
| 5948 |
|
| 5949 | return batch;
|
| 5950 |
|
| 5951 | #else
|
| 5952 | |
| 5953 | * conditions.
|
| 5954 | *
|
| 5955 | * The problem is that NOMMU needs to be able to allocate large chunks
|
| 5956 | * of contiguous memory as there's no hardware page translation to
|
| 5957 | * assemble apparent contiguous memory from discontiguous pages.
|
| 5958 | *
|
| 5959 | * Queueing large contiguous runs of pages for batching, however,
|
| 5960 | * causes the pages to actually be freed in smaller chunks. As there
|
| 5961 | * can be a significant delay between the individual batches being
|
| 5962 | * recycled, this leads to the once large chunks of space being
|
| 5963 | * fragmented and becoming unavailable for high-order allocations.
|
| 5964 |
|
| 5965 | return 1;
|
| 5966 | #endif
|
| 5967 | }
|
| 5968 |
|
| 5969 | static int percpu_pagelist_high_fraction;
|
| 5970 | static int zone_highsize(struct zone *zone, int batch, int cpu_online,
|
| 5971 | int high_fraction)
|
| 5972 | {
|
| 5973 | #ifdef CONFIG_MMU
|
| 5974 | int high;
|
| 5975 | int nr_split_cpus;
|
| 5976 | unsigned long total_pages;
|
| 5977 |
|
| 5978 | if (!high_fraction) {
|
| 5979 | |
| 5980 | * By default, the high value of the pcp is based on the zone
|
| 5981 | * low watermark so that if they are full then background
|
| 5982 | * reclaim will not be started prematurely.
|
| 5983 |
|
| 5984 | total_pages = low_wmark_pages(zone);
|
| 5985 | } else {
|
| 5986 | |
| 5987 | * If percpu_pagelist_high_fraction is configured, the high
|
| 5988 | * value is based on a fraction of the managed pages in the
|
| 5989 | * zone.
|
| 5990 |
|
| 5991 | total_pages = zone_managed_pages(zone) / high_fraction;
|
| 5992 | }
|
| 5993 |
|
| 5994 | |
| 5995 | * Split the high value across all online CPUs local to the zone. Note
|
| 5996 | * that early in boot that CPUs may not be online yet and that during
|
| 5997 | * CPU hotplug that the cpumask is not yet updated when a CPU is being
|
| 5998 | * onlined. For memory nodes that have no CPUs, split the high value
|
| 5999 | * across all online CPUs to mitigate the risk that reclaim is triggered
|
| 6000 | * prematurely due to pages stored on pcp lists.
|
| 6001 |
|
| 6002 | nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
|
| 6003 | if (!nr_split_cpus)
|
| 6004 | nr_split_cpus = num_online_cpus();
|
| 6005 | high = total_pages / nr_split_cpus;
|
| 6006 |
|
| 6007 | |
| 6008 | * Ensure high is at least batch*4. The multiple is based on the
|
| 6009 | * historical relationship between high and batch.
|
| 6010 |
|
| 6011 | high = max(high, batch << 2);
|
| 6012 |
|
| 6013 | return high;
|
| 6014 | #else
|
| 6015 | return 0;
|
| 6016 | #endif
|
| 6017 | }
|
| 6018 |
|
| 6019 | |
| 6020 | * pcp->high and pcp->batch values are related and generally batch is lower
|
| 6021 | * than high. They are also related to pcp->count such that count is lower
|
| 6022 | * than high, and as soon as it reaches high, the pcplist is flushed.
|
| 6023 | *
|
| 6024 | * However, guaranteeing these relations at all times would require e.g. write
|
| 6025 | * barriers here but also careful usage of read barriers at the read side, and
|
| 6026 | * thus be prone to error and bad for performance. Thus the update only prevents
|
| 6027 | * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
|
| 6028 | * should ensure they can cope with those fields changing asynchronously, and
|
| 6029 | * fully trust only the pcp->count field on the local CPU with interrupts
|
| 6030 | * disabled.
|
| 6031 | *
|
| 6032 | * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
|
| 6033 | * outside of boot time (or some other assurance that no concurrent updaters
|
| 6034 | * exist).
|
| 6035 |
|
| 6036 | static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
|
| 6037 | unsigned long high_max, unsigned long batch)
|
| 6038 | {
|
| 6039 | WRITE_ONCE(pcp->batch, batch);
|
| 6040 | WRITE_ONCE(pcp->high_min, high_min);
|
| 6041 | WRITE_ONCE(pcp->high_max, high_max);
|
| 6042 | }
|
| 6043 |
|
| 6044 | static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
|
| 6045 | {
|
| 6046 | int pindex;
|
| 6047 |
|
| 6048 | memset(pcp, 0, sizeof(*pcp));
|
| 6049 | memset(pzstats, 0, sizeof(*pzstats));
|
| 6050 |
|
| 6051 | spin_lock_init(&pcp->lock);
|
| 6052 | for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
|
| 6053 | INIT_LIST_HEAD(&pcp->lists[pindex]);
|
| 6054 |
|
| 6055 | |
| 6056 | * Set batch and high values safe for a boot pageset. A true percpu
|
| 6057 | * pageset's initialization will update them subsequently. Here we don't
|
| 6058 | * need to be as careful as pageset_update() as nobody can access the
|
| 6059 | * pageset yet.
|
| 6060 |
|
| 6061 | pcp->high_min = BOOT_PAGESET_HIGH;
|
| 6062 | pcp->high_max = BOOT_PAGESET_HIGH;
|
| 6063 | pcp->batch = BOOT_PAGESET_BATCH;
|
| 6064 | }
|
| 6065 |
|
| 6066 | static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
|
| 6067 | unsigned long high_max, unsigned long batch)
|
| 6068 | {
|
| 6069 | struct per_cpu_pages *pcp;
|
| 6070 | int cpu;
|
| 6071 |
|
| 6072 | for_each_possible_cpu(cpu) {
|
| 6073 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
|
| 6074 | pageset_update(pcp, high_min, high_max, batch);
|
| 6075 | }
|
| 6076 | }
|
| 6077 |
|
| 6078 | |
| 6079 | * Calculate and set new high and batch values for all per-cpu pagesets of a
|
| 6080 | * zone based on the zone's size.
|
| 6081 |
|
| 6082 | static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
|
| 6083 | {
|
| 6084 | int new_high_min, new_high_max, new_batch;
|
| 6085 |
|
| 6086 | new_batch = zone_batchsize(zone);
|
| 6087 | if (percpu_pagelist_high_fraction) {
|
| 6088 | new_high_min = zone_highsize(zone, new_batch, cpu_online,
|
| 6089 | percpu_pagelist_high_fraction);
|
| 6090 | |
| 6091 | * PCP high is tuned manually, disable auto-tuning via
|
| 6092 | * setting high_min and high_max to the manual value.
|
| 6093 |
|
| 6094 | new_high_max = new_high_min;
|
| 6095 | } else {
|
| 6096 | new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
|
| 6097 | new_high_max = zone_highsize(zone, new_batch, cpu_online,
|
| 6098 | MIN_PERCPU_PAGELIST_HIGH_FRACTION);
|
| 6099 | }
|
| 6100 |
|
| 6101 | if (zone->pageset_high_min == new_high_min &&
|
| 6102 | zone->pageset_high_max == new_high_max &&
|
| 6103 | zone->pageset_batch == new_batch)
|
| 6104 | return;
|
| 6105 |
|
| 6106 | zone->pageset_high_min = new_high_min;
|
| 6107 | zone->pageset_high_max = new_high_max;
|
| 6108 | zone->pageset_batch = new_batch;
|
| 6109 |
|
| 6110 | __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
|
| 6111 | new_batch);
|
| 6112 | }
|
| 6113 |
|
| 6114 | void __meminit setup_zone_pageset(struct zone *zone)
|
| 6115 | {
|
| 6116 | int cpu;
|
| 6117 |
|
| 6118 |
|
| 6119 | if (sizeof(struct per_cpu_zonestat) > 0)
|
| 6120 | zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
|
| 6121 |
|
| 6122 | zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
|
| 6123 | for_each_possible_cpu(cpu) {
|
| 6124 | struct per_cpu_pages *pcp;
|
| 6125 | struct per_cpu_zonestat *pzstats;
|
| 6126 |
|
| 6127 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
|
| 6128 | pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
|
| 6129 | per_cpu_pages_init(pcp, pzstats);
|
| 6130 | }
|
| 6131 |
|
| 6132 | zone_set_pageset_high_and_batch(zone, 0);
|
| 6133 | }
|
| 6134 |
|
| 6135 | |
| 6136 | * The zone indicated has a new number of managed_pages; batch sizes and percpu
|
| 6137 | * page high values need to be recalculated.
|
| 6138 |
|
| 6139 | static void zone_pcp_update(struct zone *zone, int cpu_online)
|
| 6140 | {
|
| 6141 | mutex_lock(&pcp_batch_high_lock);
|
| 6142 | zone_set_pageset_high_and_batch(zone, cpu_online);
|
| 6143 | mutex_unlock(&pcp_batch_high_lock);
|
| 6144 | }
|
| 6145 |
|
| 6146 | static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu)
|
| 6147 | {
|
| 6148 | struct per_cpu_pages *pcp;
|
| 6149 | struct cpu_cacheinfo *cci;
|
| 6150 | unsigned long UP_flags;
|
| 6151 |
|
| 6152 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
|
| 6153 | cci = get_cpu_cacheinfo(cpu);
|
| 6154 | |
| 6155 | * If data cache slice of CPU is large enough, "pcp->batch"
|
| 6156 | * pages can be preserved in PCP before draining PCP for
|
| 6157 | * consecutive high-order pages freeing without allocation.
|
| 6158 | * This can reduce zone lock contention without hurting
|
| 6159 | * cache-hot pages sharing.
|
| 6160 |
|
| 6161 | pcp_spin_lock_maybe_irqsave(pcp, UP_flags);
|
| 6162 | if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
|
| 6163 | pcp->flags |= PCPF_FREE_HIGH_BATCH;
|
| 6164 | else
|
| 6165 | pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
|
| 6166 | pcp_spin_unlock_maybe_irqrestore(pcp, UP_flags);
|
| 6167 | }
|
| 6168 |
|
| 6169 | void setup_pcp_cacheinfo(unsigned int cpu)
|
| 6170 | {
|
| 6171 | struct zone *zone;
|
| 6172 |
|
| 6173 | for_each_populated_zone(zone)
|
| 6174 | zone_pcp_update_cacheinfo(zone, cpu);
|
| 6175 | }
|
| 6176 |
|
| 6177 | |
| 6178 | * Allocate per cpu pagesets and initialize them.
|
| 6179 | * Before this call only boot pagesets were available.
|
| 6180 |
|
| 6181 | void __init setup_per_cpu_pageset(void)
|
| 6182 | {
|
| 6183 | struct pglist_data *pgdat;
|
| 6184 | struct zone *zone;
|
| 6185 | int __maybe_unused cpu;
|
| 6186 |
|
| 6187 | for_each_populated_zone(zone)
|
| 6188 | setup_zone_pageset(zone);
|
| 6189 |
|
| 6190 | #ifdef CONFIG_NUMA
|
| 6191 | |
| 6192 | * Unpopulated zones continue using the boot pagesets.
|
| 6193 | * The numa stats for these pagesets need to be reset.
|
| 6194 | * Otherwise, they will end up skewing the stats of
|
| 6195 | * the nodes these zones are associated with.
|
| 6196 |
|
| 6197 | for_each_possible_cpu(cpu) {
|
| 6198 | struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
|
| 6199 | memset(pzstats->vm_numa_event, 0,
|
| 6200 | sizeof(pzstats->vm_numa_event));
|
| 6201 | }
|
| 6202 | #endif
|
| 6203 |
|
| 6204 | for_each_online_pgdat(pgdat)
|
| 6205 | pgdat->per_cpu_nodestats =
|
| 6206 | alloc_percpu(struct per_cpu_nodestat);
|
| 6207 | }
|
| 6208 |
|
| 6209 | __meminit void zone_pcp_init(struct zone *zone)
|
| 6210 | {
|
| 6211 | |
| 6212 | * per cpu subsystem is not up at this point. The following code
|
| 6213 | * relies on the ability of the linker to provide the
|
| 6214 | * offset of a (static) per cpu variable into the per cpu area.
|
| 6215 |
|
| 6216 | zone->per_cpu_pageset = &boot_pageset;
|
| 6217 | zone->per_cpu_zonestats = &boot_zonestats;
|
| 6218 | zone->pageset_high_min = BOOT_PAGESET_HIGH;
|
| 6219 | zone->pageset_high_max = BOOT_PAGESET_HIGH;
|
| 6220 | zone->pageset_batch = BOOT_PAGESET_BATCH;
|
| 6221 |
|
| 6222 | if (populated_zone(zone))
|
| 6223 | pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
|
| 6224 | zone->present_pages, zone_batchsize(zone));
|
| 6225 | }
|
| 6226 |
|
| 6227 | static void setup_per_zone_lowmem_reserve(void);
|
| 6228 |
|
| 6229 | void adjust_managed_page_count(struct page *page, long count)
|
| 6230 | {
|
| 6231 | atomic_long_add(count, &page_zone(page)->managed_pages);
|
| 6232 | totalram_pages_add(count);
|
| 6233 | setup_per_zone_lowmem_reserve();
|
| 6234 | }
|
| 6235 | EXPORT_SYMBOL(adjust_managed_page_count);
|
| 6236 |
|
| 6237 | unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
|
| 6238 | {
|
| 6239 | void *pos;
|
| 6240 | unsigned long pages = 0;
|
| 6241 |
|
| 6242 | start = (void *)PAGE_ALIGN((unsigned long)start);
|
| 6243 | end = (void *)((unsigned long)end & PAGE_MASK);
|
| 6244 | for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
|
| 6245 | struct page *page = virt_to_page(pos);
|
| 6246 | void *direct_map_addr;
|
| 6247 |
|
| 6248 | |
| 6249 | * 'direct_map_addr' might be different from 'pos'
|
| 6250 | * because some architectures' virt_to_page()
|
| 6251 | * work with aliases. Getting the direct map
|
| 6252 | * address ensures that we get a _writeable_
|
| 6253 | * alias for the memset().
|
| 6254 |
|
| 6255 | direct_map_addr = page_address(page);
|
| 6256 | |
| 6257 | * Perform a kasan-unchecked memset() since this memory
|
| 6258 | * has not been initialized.
|
| 6259 |
|
| 6260 | direct_map_addr = kasan_reset_tag(direct_map_addr);
|
| 6261 | if ((unsigned int)poison <= 0xFF)
|
| 6262 | memset(direct_map_addr, poison, PAGE_SIZE);
|
| 6263 |
|
| 6264 | free_reserved_page(page);
|
| 6265 | }
|
| 6266 |
|
| 6267 | if (pages && s)
|
| 6268 | pr_info("Freeing %s memory: %ldK\n", s, K(pages));
|
| 6269 |
|
| 6270 | return pages;
|
| 6271 | }
|
| 6272 |
|
| 6273 | void free_reserved_page(struct page *page)
|
| 6274 | {
|
| 6275 | clear_page_tag_ref(page);
|
| 6276 | ClearPageReserved(page);
|
| 6277 | init_page_count(page);
|
| 6278 | __free_page(page);
|
| 6279 | adjust_managed_page_count(page, 1);
|
| 6280 | }
|
| 6281 | EXPORT_SYMBOL(free_reserved_page);
|
| 6282 |
|
| 6283 | static int page_alloc_cpu_dead(unsigned int cpu)
|
| 6284 | {
|
| 6285 | struct zone *zone;
|
| 6286 |
|
| 6287 | lru_add_drain_cpu(cpu);
|
| 6288 | mlock_drain_remote(cpu);
|
| 6289 | drain_pages(cpu);
|
| 6290 |
|
| 6291 | |
| 6292 | * Spill the event counters of the dead processor
|
| 6293 | * into the current processors event counters.
|
| 6294 | * This artificially elevates the count of the current
|
| 6295 | * processor.
|
| 6296 |
|
| 6297 | vm_events_fold_cpu(cpu);
|
| 6298 |
|
| 6299 | |
| 6300 | * Zero the differential counters of the dead processor
|
| 6301 | * so that the vm statistics are consistent.
|
| 6302 | *
|
| 6303 | * This is only okay since the processor is dead and cannot
|
| 6304 | * race with what we are doing.
|
| 6305 |
|
| 6306 | cpu_vm_stats_fold(cpu);
|
| 6307 |
|
| 6308 | for_each_populated_zone(zone)
|
| 6309 | zone_pcp_update(zone, 0);
|
| 6310 |
|
| 6311 | return 0;
|
| 6312 | }
|
| 6313 |
|
| 6314 | static int page_alloc_cpu_online(unsigned int cpu)
|
| 6315 | {
|
| 6316 | struct zone *zone;
|
| 6317 |
|
| 6318 | for_each_populated_zone(zone)
|
| 6319 | zone_pcp_update(zone, 1);
|
| 6320 | return 0;
|
| 6321 | }
|
| 6322 |
|
| 6323 | void __init page_alloc_init_cpuhp(void)
|
| 6324 | {
|
| 6325 | int ret;
|
| 6326 |
|
| 6327 | ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
|
| 6328 | "mm/page_alloc:pcp",
|
| 6329 | page_alloc_cpu_online,
|
| 6330 | page_alloc_cpu_dead);
|
| 6331 | WARN_ON(ret < 0);
|
| 6332 | }
|
| 6333 |
|
| 6334 | |
| 6335 | * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
|
| 6336 | * or min_free_kbytes changes.
|
| 6337 |
|
| 6338 | static void calculate_totalreserve_pages(void)
|
| 6339 | {
|
| 6340 | struct pglist_data *pgdat;
|
| 6341 | unsigned long reserve_pages = 0;
|
| 6342 | enum zone_type i, j;
|
| 6343 |
|
| 6344 | for_each_online_pgdat(pgdat) {
|
| 6345 |
|
| 6346 | pgdat->totalreserve_pages = 0;
|
| 6347 |
|
| 6348 | for (i = 0; i < MAX_NR_ZONES; i++) {
|
| 6349 | struct zone *zone = pgdat->node_zones + i;
|
| 6350 | long max = 0;
|
| 6351 | unsigned long managed_pages = zone_managed_pages(zone);
|
| 6352 |
|
| 6353 | |
| 6354 | * lowmem_reserve[j] is monotonically non-decreasing
|
| 6355 | * in j for a given zone (see
|
| 6356 | * setup_per_zone_lowmem_reserve()). The maximum
|
| 6357 | * valid reserve lives at the highest index with a
|
| 6358 | * non-zero value, so scan backwards and stop at the
|
| 6359 | * first hit.
|
| 6360 |
|
| 6361 | for (j = MAX_NR_ZONES - 1; j > i; j--) {
|
| 6362 | if (!zone->lowmem_reserve[j])
|
| 6363 | continue;
|
| 6364 |
|
| 6365 | max = zone->lowmem_reserve[j];
|
| 6366 | break;
|
| 6367 | }
|
| 6368 |
|
| 6369 | max += high_wmark_pages(zone);
|
| 6370 |
|
| 6371 | max = min_t(unsigned long, max, managed_pages);
|
| 6372 |
|
| 6373 | pgdat->totalreserve_pages += max;
|
| 6374 |
|
| 6375 | reserve_pages += max;
|
| 6376 | }
|
| 6377 | }
|
| 6378 | totalreserve_pages = reserve_pages;
|
| 6379 | trace_mm_calculate_totalreserve_pages(totalreserve_pages);
|
| 6380 | }
|
| 6381 |
|
| 6382 | |
| 6383 | * setup_per_zone_lowmem_reserve - called whenever
|
| 6384 | * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
|
| 6385 | * has a correct pages reserved value, so an adequate number of
|
| 6386 | * pages are left in the zone after a successful __alloc_pages().
|
| 6387 |
|
| 6388 | static void setup_per_zone_lowmem_reserve(void)
|
| 6389 | {
|
| 6390 | struct pglist_data *pgdat;
|
| 6391 | enum zone_type i, j;
|
| 6392 | |
| 6393 | * For a given zone node_zones[i], lowmem_reserve[j] (j > i)
|
| 6394 | * represents how many pages in zone i must effectively be kept
|
| 6395 | * in reserve when deciding whether an allocation class that is
|
| 6396 | * allowed to allocate from zones up to j may fall back into
|
| 6397 | * zone i.
|
| 6398 | *
|
| 6399 | * As j increases, the allocation class can use a strictly larger
|
| 6400 | * set of fallback zones and therefore must not be allowed to
|
| 6401 | * deplete low zones more aggressively than a less flexible one.
|
| 6402 | * As a result, lowmem_reserve[j] is required to be monotonically
|
| 6403 | * non-decreasing in j for each zone i. Callers such as
|
| 6404 | * calculate_totalreserve_pages() rely on this monotonicity when
|
| 6405 | * selecting the maximum reserve entry.
|
| 6406 |
|
| 6407 | for_each_online_pgdat(pgdat) {
|
| 6408 | for (i = 0; i < MAX_NR_ZONES - 1; i++) {
|
| 6409 | struct zone *zone = &pgdat->node_zones[i];
|
| 6410 | int ratio = sysctl_lowmem_reserve_ratio[i];
|
| 6411 | bool clear = !ratio || !zone_managed_pages(zone);
|
| 6412 | unsigned long managed_pages = 0;
|
| 6413 |
|
| 6414 | for (j = i + 1; j < MAX_NR_ZONES; j++) {
|
| 6415 | struct zone *upper_zone = &pgdat->node_zones[j];
|
| 6416 |
|
| 6417 | managed_pages += zone_managed_pages(upper_zone);
|
| 6418 |
|
| 6419 | if (clear)
|
| 6420 | zone->lowmem_reserve[j] = 0;
|
| 6421 | else
|
| 6422 | zone->lowmem_reserve[j] = managed_pages / ratio;
|
| 6423 | trace_mm_setup_per_zone_lowmem_reserve(zone, upper_zone,
|
| 6424 | zone->lowmem_reserve[j]);
|
| 6425 | }
|
| 6426 | }
|
| 6427 | }
|
| 6428 |
|
| 6429 |
|
| 6430 | calculate_totalreserve_pages();
|
| 6431 | }
|
| 6432 |
|
| 6433 | static void __setup_per_zone_wmarks(void)
|
| 6434 | {
|
| 6435 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
|
| 6436 | unsigned long lowmem_pages = 0;
|
| 6437 | struct zone *zone;
|
| 6438 | unsigned long flags;
|
| 6439 |
|
| 6440 |
|
| 6441 | for_each_zone(zone) {
|
| 6442 | if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
|
| 6443 | lowmem_pages += zone_managed_pages(zone);
|
| 6444 | }
|
| 6445 |
|
| 6446 | for_each_zone(zone) {
|
| 6447 | u64 tmp;
|
| 6448 |
|
| 6449 | spin_lock_irqsave(&zone->lock, flags);
|
| 6450 | tmp = (u64)pages_min * zone_managed_pages(zone);
|
| 6451 | tmp = div64_ul(tmp, lowmem_pages);
|
| 6452 | if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
|
| 6453 | |
| 6454 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't
|
| 6455 | * need highmem and movable zones pages, so cap pages_min
|
| 6456 | * to a small value here.
|
| 6457 | *
|
| 6458 | * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
|
| 6459 | * deltas control async page reclaim, and so should
|
| 6460 | * not be capped for highmem and movable zones.
|
| 6461 |
|
| 6462 | unsigned long min_pages;
|
| 6463 |
|
| 6464 | min_pages = zone_managed_pages(zone) / 1024;
|
| 6465 | min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
|
| 6466 | zone->_watermark[WMARK_MIN] = min_pages;
|
| 6467 | } else {
|
| 6468 | |
| 6469 | * If it's a lowmem zone, reserve a number of pages
|
| 6470 | * proportionate to the zone's size.
|
| 6471 |
|
| 6472 | zone->_watermark[WMARK_MIN] = tmp;
|
| 6473 | }
|
| 6474 |
|
| 6475 | |
| 6476 | * Set the kswapd watermarks distance according to the
|
| 6477 | * scale factor in proportion to available memory, but
|
| 6478 | * ensure a minimum size on small systems.
|
| 6479 |
|
| 6480 | tmp = max_t(u64, tmp >> 2,
|
| 6481 | mult_frac(zone_managed_pages(zone),
|
| 6482 | watermark_scale_factor, 10000));
|
| 6483 |
|
| 6484 | zone->watermark_boost = 0;
|
| 6485 | zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
|
| 6486 | zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
|
| 6487 | zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
|
| 6488 | trace_mm_setup_per_zone_wmarks(zone);
|
| 6489 |
|
| 6490 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 6491 | }
|
| 6492 |
|
| 6493 |
|
| 6494 | calculate_totalreserve_pages();
|
| 6495 | }
|
| 6496 |
|
| 6497 | |
| 6498 | * setup_per_zone_wmarks - called when min_free_kbytes changes
|
| 6499 | * or when memory is hot-{added|removed}
|
| 6500 | *
|
| 6501 | * Ensures that the watermark[min,low,high] values for each zone are set
|
| 6502 | * correctly with respect to min_free_kbytes.
|
| 6503 |
|
| 6504 | void setup_per_zone_wmarks(void)
|
| 6505 | {
|
| 6506 | struct zone *zone;
|
| 6507 | static DEFINE_SPINLOCK(lock);
|
| 6508 |
|
| 6509 | spin_lock(&lock);
|
| 6510 | __setup_per_zone_wmarks();
|
| 6511 | spin_unlock(&lock);
|
| 6512 |
|
| 6513 | |
| 6514 | * The watermark size have changed so update the pcpu batch
|
| 6515 | * and high limits or the limits may be inappropriate.
|
| 6516 |
|
| 6517 | for_each_zone(zone)
|
| 6518 | zone_pcp_update(zone, 0);
|
| 6519 | }
|
| 6520 |
|
| 6521 | |
| 6522 | * Initialise min_free_kbytes.
|
| 6523 | *
|
| 6524 | * For small machines we want it small (128k min). For large machines
|
| 6525 | * we want it large (256MB max). But it is not linear, because network
|
| 6526 | * bandwidth does not increase linearly with machine size. We use
|
| 6527 | *
|
| 6528 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
|
| 6529 | * min_free_kbytes = sqrt(lowmem_kbytes * 16)
|
| 6530 | *
|
| 6531 | * which yields
|
| 6532 | *
|
| 6533 | * 16MB: 512k
|
| 6534 | * 32MB: 724k
|
| 6535 | * 64MB: 1024k
|
| 6536 | * 128MB: 1448k
|
| 6537 | * 256MB: 2048k
|
| 6538 | * 512MB: 2896k
|
| 6539 | * 1024MB: 4096k
|
| 6540 | * 2048MB: 5792k
|
| 6541 | * 4096MB: 8192k
|
| 6542 | * 8192MB: 11584k
|
| 6543 | * 16384MB: 16384k
|
| 6544 |
|
| 6545 | void calculate_min_free_kbytes(void)
|
| 6546 | {
|
| 6547 | unsigned long lowmem_kbytes;
|
| 6548 | int new_min_free_kbytes;
|
| 6549 |
|
| 6550 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
|
| 6551 | new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
|
| 6552 |
|
| 6553 | if (new_min_free_kbytes > user_min_free_kbytes)
|
| 6554 | min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
|
| 6555 | else
|
| 6556 | pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
|
| 6557 | new_min_free_kbytes, user_min_free_kbytes);
|
| 6558 |
|
| 6559 | }
|
| 6560 |
|
| 6561 | int __meminit init_per_zone_wmark_min(void)
|
| 6562 | {
|
| 6563 | calculate_min_free_kbytes();
|
| 6564 | setup_per_zone_wmarks();
|
| 6565 | refresh_zone_stat_thresholds();
|
| 6566 | setup_per_zone_lowmem_reserve();
|
| 6567 |
|
| 6568 | #ifdef CONFIG_NUMA
|
| 6569 | setup_min_unmapped_ratio();
|
| 6570 | setup_min_slab_ratio();
|
| 6571 | #endif
|
| 6572 |
|
| 6573 | khugepaged_min_free_kbytes_update();
|
| 6574 |
|
| 6575 | return 0;
|
| 6576 | }
|
| 6577 | postcore_initcall(init_per_zone_wmark_min)
|
| 6578 |
|
| 6579 | |
| 6580 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
|
| 6581 | * that we can call two helper functions whenever min_free_kbytes
|
| 6582 | * changes.
|
| 6583 |
|
| 6584 | static int min_free_kbytes_sysctl_handler(const struct ctl_table *table, int write,
|
| 6585 | void *buffer, size_t *length, loff_t *ppos)
|
| 6586 | {
|
| 6587 | int rc;
|
| 6588 |
|
| 6589 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6590 | if (rc)
|
| 6591 | return rc;
|
| 6592 |
|
| 6593 | if (write) {
|
| 6594 | user_min_free_kbytes = min_free_kbytes;
|
| 6595 | setup_per_zone_wmarks();
|
| 6596 | }
|
| 6597 | return 0;
|
| 6598 | }
|
| 6599 |
|
| 6600 | static int watermark_scale_factor_sysctl_handler(const struct ctl_table *table, int write,
|
| 6601 | void *buffer, size_t *length, loff_t *ppos)
|
| 6602 | {
|
| 6603 | int rc;
|
| 6604 |
|
| 6605 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6606 | if (rc)
|
| 6607 | return rc;
|
| 6608 |
|
| 6609 | if (write)
|
| 6610 | setup_per_zone_wmarks();
|
| 6611 |
|
| 6612 | return 0;
|
| 6613 | }
|
| 6614 |
|
| 6615 | #ifdef CONFIG_NUMA
|
| 6616 | static void setup_min_unmapped_ratio(void)
|
| 6617 | {
|
| 6618 | pg_data_t *pgdat;
|
| 6619 | struct zone *zone;
|
| 6620 |
|
| 6621 | for_each_online_pgdat(pgdat)
|
| 6622 | pgdat->min_unmapped_pages = 0;
|
| 6623 |
|
| 6624 | for_each_zone(zone)
|
| 6625 | zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
|
| 6626 | sysctl_min_unmapped_ratio) / 100;
|
| 6627 | }
|
| 6628 |
|
| 6629 |
|
| 6630 | static int sysctl_min_unmapped_ratio_sysctl_handler(const struct ctl_table *table, int write,
|
| 6631 | void *buffer, size_t *length, loff_t *ppos)
|
| 6632 | {
|
| 6633 | int rc;
|
| 6634 |
|
| 6635 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6636 | if (rc)
|
| 6637 | return rc;
|
| 6638 |
|
| 6639 | setup_min_unmapped_ratio();
|
| 6640 |
|
| 6641 | return 0;
|
| 6642 | }
|
| 6643 |
|
| 6644 | static void setup_min_slab_ratio(void)
|
| 6645 | {
|
| 6646 | pg_data_t *pgdat;
|
| 6647 | struct zone *zone;
|
| 6648 |
|
| 6649 | for_each_online_pgdat(pgdat)
|
| 6650 | pgdat->min_slab_pages = 0;
|
| 6651 |
|
| 6652 | for_each_zone(zone)
|
| 6653 | zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
|
| 6654 | sysctl_min_slab_ratio) / 100;
|
| 6655 | }
|
| 6656 |
|
| 6657 | static int sysctl_min_slab_ratio_sysctl_handler(const struct ctl_table *table, int write,
|
| 6658 | void *buffer, size_t *length, loff_t *ppos)
|
| 6659 | {
|
| 6660 | int rc;
|
| 6661 |
|
| 6662 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6663 | if (rc)
|
| 6664 | return rc;
|
| 6665 |
|
| 6666 | setup_min_slab_ratio();
|
| 6667 |
|
| 6668 | return 0;
|
| 6669 | }
|
| 6670 | #endif
|
| 6671 |
|
| 6672 | |
| 6673 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
|
| 6674 | * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
|
| 6675 | * whenever sysctl_lowmem_reserve_ratio changes.
|
| 6676 | *
|
| 6677 | * The reserve ratio obviously has absolutely no relation with the
|
| 6678 | * minimum watermarks. The lowmem reserve ratio can only make sense
|
| 6679 | * if in function of the boot time zone sizes.
|
| 6680 |
|
| 6681 | static int lowmem_reserve_ratio_sysctl_handler(const struct ctl_table *table,
|
| 6682 | int write, void *buffer, size_t *length, loff_t *ppos)
|
| 6683 | {
|
| 6684 | int i;
|
| 6685 |
|
| 6686 | proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6687 |
|
| 6688 | for (i = 0; i < MAX_NR_ZONES; i++) {
|
| 6689 | if (sysctl_lowmem_reserve_ratio[i] < 1)
|
| 6690 | sysctl_lowmem_reserve_ratio[i] = 0;
|
| 6691 | }
|
| 6692 |
|
| 6693 | setup_per_zone_lowmem_reserve();
|
| 6694 | return 0;
|
| 6695 | }
|
| 6696 |
|
| 6697 | |
| 6698 | * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
|
| 6699 | * cpu. It is the fraction of total pages in each zone that a hot per cpu
|
| 6700 | * pagelist can have before it gets flushed back to buddy allocator.
|
| 6701 |
|
| 6702 | static int percpu_pagelist_high_fraction_sysctl_handler(const struct ctl_table *table,
|
| 6703 | int write, void *buffer, size_t *length, loff_t *ppos)
|
| 6704 | {
|
| 6705 | struct zone *zone;
|
| 6706 | int old_percpu_pagelist_high_fraction;
|
| 6707 | int ret;
|
| 6708 |
|
| 6709 | |
| 6710 | * Avoid using pcp_batch_high_lock for reads as the value is read
|
| 6711 | * atomically and a race with offlining is harmless.
|
| 6712 |
|
| 6713 |
|
| 6714 | if (!write)
|
| 6715 | return proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6716 |
|
| 6717 | mutex_lock(&pcp_batch_high_lock);
|
| 6718 | old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
|
| 6719 |
|
| 6720 | ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6721 | if (ret < 0)
|
| 6722 | goto out;
|
| 6723 |
|
| 6724 |
|
| 6725 | if (percpu_pagelist_high_fraction &&
|
| 6726 | percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
|
| 6727 | percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
|
| 6728 | ret = -EINVAL;
|
| 6729 | goto out;
|
| 6730 | }
|
| 6731 |
|
| 6732 |
|
| 6733 | if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
|
| 6734 | goto out;
|
| 6735 |
|
| 6736 | for_each_populated_zone(zone)
|
| 6737 | zone_set_pageset_high_and_batch(zone, 0);
|
| 6738 | out:
|
| 6739 | mutex_unlock(&pcp_batch_high_lock);
|
| 6740 | return ret;
|
| 6741 | }
|
| 6742 |
|
| 6743 | static const struct ctl_table page_alloc_sysctl_table[] = {
|
| 6744 | {
|
| 6745 | .procname = "min_free_kbytes",
|
| 6746 | .data = &min_free_kbytes,
|
| 6747 | .maxlen = sizeof(min_free_kbytes),
|
| 6748 | .mode = 0644,
|
| 6749 | .proc_handler = min_free_kbytes_sysctl_handler,
|
| 6750 | .extra1 = SYSCTL_ZERO,
|
| 6751 | },
|
| 6752 | {
|
| 6753 | .procname = "watermark_boost_factor",
|
| 6754 | .data = &watermark_boost_factor,
|
| 6755 | .maxlen = sizeof(watermark_boost_factor),
|
| 6756 | .mode = 0644,
|
| 6757 | .proc_handler = proc_dointvec_minmax,
|
| 6758 | .extra1 = SYSCTL_ZERO,
|
| 6759 | },
|
| 6760 | {
|
| 6761 | .procname = "watermark_scale_factor",
|
| 6762 | .data = &watermark_scale_factor,
|
| 6763 | .maxlen = sizeof(watermark_scale_factor),
|
| 6764 | .mode = 0644,
|
| 6765 | .proc_handler = watermark_scale_factor_sysctl_handler,
|
| 6766 | .extra1 = SYSCTL_ONE,
|
| 6767 | .extra2 = SYSCTL_THREE_THOUSAND,
|
| 6768 | },
|
| 6769 | {
|
| 6770 | .procname = "defrag_mode",
|
| 6771 | .data = &defrag_mode,
|
| 6772 | .maxlen = sizeof(defrag_mode),
|
| 6773 | .mode = 0644,
|
| 6774 | .proc_handler = proc_dointvec_minmax,
|
| 6775 | .extra1 = SYSCTL_ZERO,
|
| 6776 | .extra2 = SYSCTL_ONE,
|
| 6777 | },
|
| 6778 | {
|
| 6779 | .procname = "percpu_pagelist_high_fraction",
|
| 6780 | .data = &percpu_pagelist_high_fraction,
|
| 6781 | .maxlen = sizeof(percpu_pagelist_high_fraction),
|
| 6782 | .mode = 0644,
|
| 6783 | .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
|
| 6784 | .extra1 = SYSCTL_ZERO,
|
| 6785 | },
|
| 6786 | {
|
| 6787 | .procname = "lowmem_reserve_ratio",
|
| 6788 | .data = &sysctl_lowmem_reserve_ratio,
|
| 6789 | .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
|
| 6790 | .mode = 0644,
|
| 6791 | .proc_handler = lowmem_reserve_ratio_sysctl_handler,
|
| 6792 | },
|
| 6793 | #ifdef CONFIG_NUMA
|
| 6794 | {
|
| 6795 | .procname = "numa_zonelist_order",
|
| 6796 | .data = &numa_zonelist_order,
|
| 6797 | .maxlen = NUMA_ZONELIST_ORDER_LEN,
|
| 6798 | .mode = 0644,
|
| 6799 | .proc_handler = numa_zonelist_order_handler,
|
| 6800 | },
|
| 6801 | {
|
| 6802 | .procname = "min_unmapped_ratio",
|
| 6803 | .data = &sysctl_min_unmapped_ratio,
|
| 6804 | .maxlen = sizeof(sysctl_min_unmapped_ratio),
|
| 6805 | .mode = 0644,
|
| 6806 | .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
|
| 6807 | .extra1 = SYSCTL_ZERO,
|
| 6808 | .extra2 = SYSCTL_ONE_HUNDRED,
|
| 6809 | },
|
| 6810 | {
|
| 6811 | .procname = "min_slab_ratio",
|
| 6812 | .data = &sysctl_min_slab_ratio,
|
| 6813 | .maxlen = sizeof(sysctl_min_slab_ratio),
|
| 6814 | .mode = 0644,
|
| 6815 | .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
|
| 6816 | .extra1 = SYSCTL_ZERO,
|
| 6817 | .extra2 = SYSCTL_ONE_HUNDRED,
|
| 6818 | },
|
| 6819 | #endif
|
| 6820 | };
|
| 6821 |
|
| 6822 | void __init page_alloc_sysctl_init(void)
|
| 6823 | {
|
| 6824 | register_sysctl_init("vm", page_alloc_sysctl_table);
|
| 6825 | }
|
| 6826 |
|
| 6827 | #ifdef CONFIG_CONTIG_ALLOC
|
| 6828 |
|
| 6829 | static void alloc_contig_dump_pages(struct list_head *page_list)
|
| 6830 | {
|
| 6831 | DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
|
| 6832 |
|
| 6833 | if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
|
| 6834 | struct page *page;
|
| 6835 |
|
| 6836 | dump_stack();
|
| 6837 | list_for_each_entry(page, page_list, lru)
|
| 6838 | dump_page(page, "migration failure");
|
| 6839 | }
|
| 6840 | }
|
| 6841 |
|
| 6842 |
|
| 6843 | static int __alloc_contig_migrate_range(struct compact_control *cc,
|
| 6844 | unsigned long start, unsigned long end)
|
| 6845 | {
|
| 6846 |
|
| 6847 | unsigned int nr_reclaimed;
|
| 6848 | unsigned long pfn = start;
|
| 6849 | unsigned int tries = 0;
|
| 6850 | int ret = 0;
|
| 6851 | struct migration_target_control mtc = {
|
| 6852 | .nid = zone_to_nid(cc->zone),
|
| 6853 | .gfp_mask = cc->gfp_mask,
|
| 6854 | .reason = MR_CONTIG_RANGE,
|
| 6855 | };
|
| 6856 |
|
| 6857 | lru_cache_disable();
|
| 6858 |
|
| 6859 | while (pfn < end || !list_empty(&cc->migratepages)) {
|
| 6860 | if (fatal_signal_pending(current)) {
|
| 6861 | ret = -EINTR;
|
| 6862 | break;
|
| 6863 | }
|
| 6864 |
|
| 6865 | if (list_empty(&cc->migratepages)) {
|
| 6866 | cc->nr_migratepages = 0;
|
| 6867 | ret = isolate_migratepages_range(cc, pfn, end);
|
| 6868 | if (ret && ret != -EAGAIN)
|
| 6869 | break;
|
| 6870 | pfn = cc->migrate_pfn;
|
| 6871 | tries = 0;
|
| 6872 | } else if (++tries == 5) {
|
| 6873 | ret = -EBUSY;
|
| 6874 | break;
|
| 6875 | }
|
| 6876 |
|
| 6877 | nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
|
| 6878 | &cc->migratepages);
|
| 6879 | cc->nr_migratepages -= nr_reclaimed;
|
| 6880 |
|
| 6881 | ret = migrate_pages(&cc->migratepages, alloc_migration_target,
|
| 6882 | NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
|
| 6883 |
|
| 6884 | |
| 6885 | * On -ENOMEM, migrate_pages() bails out right away. It is pointless
|
| 6886 | * to retry again over this error, so do the same here.
|
| 6887 |
|
| 6888 | if (ret == -ENOMEM)
|
| 6889 | break;
|
| 6890 | }
|
| 6891 |
|
| 6892 | lru_cache_enable();
|
| 6893 | if (ret < 0) {
|
| 6894 | if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
|
| 6895 | alloc_contig_dump_pages(&cc->migratepages);
|
| 6896 | putback_movable_pages(&cc->migratepages);
|
| 6897 | }
|
| 6898 |
|
| 6899 | return (ret < 0) ? ret : 0;
|
| 6900 | }
|
| 6901 |
|
| 6902 | static void split_free_frozen_pages(struct list_head *list, gfp_t gfp_mask)
|
| 6903 | {
|
| 6904 | int order;
|
| 6905 |
|
| 6906 | for (order = 0; order < NR_PAGE_ORDERS; order++) {
|
| 6907 | struct page *page, *next;
|
| 6908 | int nr_pages = 1 << order;
|
| 6909 |
|
| 6910 | list_for_each_entry_safe(page, next, &list[order], lru) {
|
| 6911 | int i;
|
| 6912 |
|
| 6913 | post_alloc_hook(page, order, gfp_mask);
|
| 6914 | if (!order)
|
| 6915 | continue;
|
| 6916 |
|
| 6917 | __split_page(page, order);
|
| 6918 |
|
| 6919 |
|
| 6920 | list_del(&page->lru);
|
| 6921 | for (i = 0; i < nr_pages; i++)
|
| 6922 | list_add_tail(&page[i].lru, &list[0]);
|
| 6923 | }
|
| 6924 | }
|
| 6925 | }
|
| 6926 |
|
| 6927 | static int __alloc_contig_verify_gfp_mask(gfp_t gfp_mask, gfp_t *gfp_cc_mask)
|
| 6928 | {
|
| 6929 | const gfp_t reclaim_mask = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
|
| 6930 | const gfp_t action_mask = __GFP_COMP | __GFP_RETRY_MAYFAIL | __GFP_NOWARN |
|
| 6931 | __GFP_ZERO | __GFP_ZEROTAGS | __GFP_SKIP_ZERO |
|
| 6932 | __GFP_SKIP_KASAN;
|
| 6933 | const gfp_t cc_action_mask = __GFP_RETRY_MAYFAIL | __GFP_NOWARN;
|
| 6934 |
|
| 6935 | |
| 6936 | * We are given the range to allocate; node, mobility and placement
|
| 6937 | * hints are irrelevant at this point. We'll simply ignore them.
|
| 6938 |
|
| 6939 | gfp_mask &= ~(GFP_ZONEMASK | __GFP_RECLAIMABLE | __GFP_WRITE |
|
| 6940 | __GFP_HARDWALL | __GFP_THISNODE | __GFP_MOVABLE);
|
| 6941 |
|
| 6942 | |
| 6943 | * We only support most reclaim flags (but not NOFAIL/NORETRY), and
|
| 6944 | * selected action flags.
|
| 6945 |
|
| 6946 | if (gfp_mask & ~(reclaim_mask | action_mask))
|
| 6947 | return -EINVAL;
|
| 6948 |
|
| 6949 | |
| 6950 | * Flags to control page compaction/migration/reclaim, to free up our
|
| 6951 | * page range. Migratable pages are movable, __GFP_MOVABLE is implied
|
| 6952 | * for them.
|
| 6953 | *
|
| 6954 | * Traditionally we always had __GFP_RETRY_MAYFAIL set, keep doing that
|
| 6955 | * to not degrade callers.
|
| 6956 |
|
| 6957 | *gfp_cc_mask = (gfp_mask & (reclaim_mask | cc_action_mask)) |
|
| 6958 | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL;
|
| 6959 | return 0;
|
| 6960 | }
|
| 6961 |
|
| 6962 | static void __free_contig_frozen_range(unsigned long pfn, unsigned long nr_pages)
|
| 6963 | {
|
| 6964 | for (; nr_pages--; pfn++)
|
| 6965 | free_frozen_pages(pfn_to_page(pfn), 0);
|
| 6966 | }
|
| 6967 |
|
| 6968 | |
| 6969 | * alloc_contig_frozen_range() -- tries to allocate given range of frozen pages
|
| 6970 | * @start: start PFN to allocate
|
| 6971 | * @end: one-past-the-last PFN to allocate
|
| 6972 | * @alloc_flags: allocation information
|
| 6973 | * @gfp_mask: GFP mask. Node/zone/placement hints are ignored; only some
|
| 6974 | * action and reclaim modifiers are supported. Reclaim modifiers
|
| 6975 | * control allocation behavior during compaction/migration/reclaim.
|
| 6976 | *
|
| 6977 | * The PFN range does not have to be pageblock aligned. The PFN range must
|
| 6978 | * belong to a single zone.
|
| 6979 | *
|
| 6980 | * The first thing this routine does is attempt to MIGRATE_ISOLATE all
|
| 6981 | * pageblocks in the range. Once isolated, the pageblocks should not
|
| 6982 | * be modified by others.
|
| 6983 | *
|
| 6984 | * All frozen pages which PFN is in [start, end) are allocated for the
|
| 6985 | * caller, and they could be freed with free_contig_frozen_range(),
|
| 6986 | * free_frozen_pages() also could be used to free compound frozen pages
|
| 6987 | * directly.
|
| 6988 | *
|
| 6989 | * Return: zero on success or negative error code.
|
| 6990 |
|
| 6991 | int alloc_contig_frozen_range_noprof(unsigned long start, unsigned long end,
|
| 6992 | acr_flags_t alloc_flags, gfp_t gfp_mask)
|
| 6993 | {
|
| 6994 | const unsigned int order = ilog2(end - start);
|
| 6995 | unsigned long outer_start, outer_end;
|
| 6996 | int ret = 0;
|
| 6997 |
|
| 6998 | struct compact_control cc = {
|
| 6999 | .nr_migratepages = 0,
|
| 7000 | .order = -1,
|
| 7001 | .zone = page_zone(pfn_to_page(start)),
|
| 7002 | .mode = MIGRATE_SYNC,
|
| 7003 | .ignore_skip_hint = true,
|
| 7004 | .no_set_skip_hint = true,
|
| 7005 | .alloc_contig = true,
|
| 7006 | };
|
| 7007 | INIT_LIST_HEAD(&cc.migratepages);
|
| 7008 | enum pb_isolate_mode mode = (alloc_flags & ACR_FLAGS_CMA) ?
|
| 7009 | PB_ISOLATE_MODE_CMA_ALLOC :
|
| 7010 | PB_ISOLATE_MODE_OTHER;
|
| 7011 |
|
| 7012 | |
| 7013 | * In contrast to the buddy, we allow for orders here that exceed
|
| 7014 | * MAX_PAGE_ORDER, so we must manually make sure that we are not
|
| 7015 | * exceeding the maximum folio order.
|
| 7016 |
|
| 7017 | if (WARN_ON_ONCE((gfp_mask & __GFP_COMP) && order > MAX_FOLIO_ORDER))
|
| 7018 | return -EINVAL;
|
| 7019 |
|
| 7020 | gfp_mask = current_gfp_context(gfp_mask);
|
| 7021 | if (__alloc_contig_verify_gfp_mask(gfp_mask, (gfp_t *)&cc.gfp_mask))
|
| 7022 | return -EINVAL;
|
| 7023 |
|
| 7024 | |
| 7025 | * What we do here is we mark all pageblocks in range as
|
| 7026 | * MIGRATE_ISOLATE. Because pageblock and max order pages may
|
| 7027 | * have different sizes, and due to the way page allocator
|
| 7028 | * work, start_isolate_page_range() has special handlings for this.
|
| 7029 | *
|
| 7030 | * Once the pageblocks are marked as MIGRATE_ISOLATE, we
|
| 7031 | * migrate the pages from an unaligned range (ie. pages that
|
| 7032 | * we are interested in). This will put all the pages in
|
| 7033 | * range back to page allocator as MIGRATE_ISOLATE.
|
| 7034 | *
|
| 7035 | * When this is done, we take the pages in range from page
|
| 7036 | * allocator removing them from the buddy system. This way
|
| 7037 | * page allocator will never consider using them.
|
| 7038 | *
|
| 7039 | * This lets us mark the pageblocks back as
|
| 7040 | * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
|
| 7041 | * aligned range but not in the unaligned, original range are
|
| 7042 | * put back to page allocator so that buddy can use them.
|
| 7043 |
|
| 7044 |
|
| 7045 | ret = start_isolate_page_range(start, end, mode);
|
| 7046 | if (ret)
|
| 7047 | goto done;
|
| 7048 |
|
| 7049 | drain_all_pages(cc.zone);
|
| 7050 |
|
| 7051 | |
| 7052 | * In case of -EBUSY, we'd like to know which page causes problem.
|
| 7053 | * So, just fall through. test_pages_isolated() has a tracepoint
|
| 7054 | * which will report the busy page.
|
| 7055 | *
|
| 7056 | * It is possible that busy pages could become available before
|
| 7057 | * the call to test_pages_isolated, and the range will actually be
|
| 7058 | * allocated. So, if we fall through be sure to clear ret so that
|
| 7059 | * -EBUSY is not accidentally used or returned to caller.
|
| 7060 |
|
| 7061 | ret = __alloc_contig_migrate_range(&cc, start, end);
|
| 7062 | if (ret && ret != -EBUSY)
|
| 7063 | goto done;
|
| 7064 |
|
| 7065 | |
| 7066 | * When in-use hugetlb pages are migrated, they may simply be released
|
| 7067 | * back into the free hugepage pool instead of being returned to the
|
| 7068 | * buddy system. After the migration of in-use huge pages is completed,
|
| 7069 | * we will invoke replace_free_hugepage_folios() to ensure that these
|
| 7070 | * hugepages are properly released to the buddy system.
|
| 7071 |
|
| 7072 | ret = replace_free_hugepage_folios(start, end);
|
| 7073 | if (ret)
|
| 7074 | goto done;
|
| 7075 |
|
| 7076 | |
| 7077 | * Pages from [start, end) are within a pageblock_nr_pages
|
| 7078 | * aligned blocks that are marked as MIGRATE_ISOLATE. What's
|
| 7079 | * more, all pages in [start, end) are free in page allocator.
|
| 7080 | * What we are going to do is to allocate all pages from
|
| 7081 | * [start, end) (that is remove them from page allocator).
|
| 7082 | *
|
| 7083 | * The only problem is that pages at the beginning and at the
|
| 7084 | * end of interesting range may be not aligned with pages that
|
| 7085 | * page allocator holds, ie. they can be part of higher order
|
| 7086 | * pages. Because of this, we reserve the bigger range and
|
| 7087 | * once this is done free the pages we are not interested in.
|
| 7088 | *
|
| 7089 | * We don't have to hold zone->lock here because the pages are
|
| 7090 | * isolated thus they won't get removed from buddy.
|
| 7091 |
|
| 7092 | outer_start = find_large_buddy(start);
|
| 7093 |
|
| 7094 |
|
| 7095 | if (test_pages_isolated(outer_start, end, mode)) {
|
| 7096 | ret = -EBUSY;
|
| 7097 | goto done;
|
| 7098 | }
|
| 7099 |
|
| 7100 |
|
| 7101 | outer_end = isolate_freepages_range(&cc, outer_start, end);
|
| 7102 | if (!outer_end) {
|
| 7103 | ret = -EBUSY;
|
| 7104 | goto done;
|
| 7105 | }
|
| 7106 |
|
| 7107 | if (!(gfp_mask & __GFP_COMP)) {
|
| 7108 | split_free_frozen_pages(cc.freepages, gfp_mask);
|
| 7109 |
|
| 7110 |
|
| 7111 | if (start != outer_start)
|
| 7112 | __free_contig_frozen_range(outer_start, start - outer_start);
|
| 7113 | if (end != outer_end)
|
| 7114 | __free_contig_frozen_range(end, outer_end - end);
|
| 7115 | } else if (start == outer_start && end == outer_end && is_power_of_2(end - start)) {
|
| 7116 | struct page *head = pfn_to_page(start);
|
| 7117 |
|
| 7118 | check_new_pages(head, order);
|
| 7119 | prep_new_page(head, order, gfp_mask, 0);
|
| 7120 | } else {
|
| 7121 | ret = -EINVAL;
|
| 7122 | WARN(true, "PFN range: requested [%lu, %lu), allocated [%lu, %lu)\n",
|
| 7123 | start, end, outer_start, outer_end);
|
| 7124 | }
|
| 7125 | done:
|
| 7126 | undo_isolate_page_range(start, end);
|
| 7127 | return ret;
|
| 7128 | }
|
| 7129 | EXPORT_SYMBOL(alloc_contig_frozen_range_noprof);
|
| 7130 |
|
| 7131 | |
| 7132 | * alloc_contig_range() -- tries to allocate given range of pages
|
| 7133 | * @start: start PFN to allocate
|
| 7134 | * @end: one-past-the-last PFN to allocate
|
| 7135 | * @alloc_flags: allocation information
|
| 7136 | * @gfp_mask: GFP mask.
|
| 7137 | *
|
| 7138 | * This routine is a wrapper around alloc_contig_frozen_range(), it can't
|
| 7139 | * be used to allocate compound pages, the refcount of each allocated page
|
| 7140 | * will be set to one.
|
| 7141 | *
|
| 7142 | * All pages which PFN is in [start, end) are allocated for the caller,
|
| 7143 | * and should be freed with free_contig_range() or by manually calling
|
| 7144 | * __free_page() on each allocated page.
|
| 7145 | *
|
| 7146 | * Return: zero on success or negative error code.
|
| 7147 |
|
| 7148 | int alloc_contig_range_noprof(unsigned long start, unsigned long end,
|
| 7149 | acr_flags_t alloc_flags, gfp_t gfp_mask)
|
| 7150 | {
|
| 7151 | int ret;
|
| 7152 |
|
| 7153 | if (WARN_ON(gfp_mask & __GFP_COMP))
|
| 7154 | return -EINVAL;
|
| 7155 |
|
| 7156 | ret = alloc_contig_frozen_range_noprof(start, end, alloc_flags, gfp_mask);
|
| 7157 | if (!ret)
|
| 7158 | set_pages_refcounted(pfn_to_page(start), end - start);
|
| 7159 |
|
| 7160 | return ret;
|
| 7161 | }
|
| 7162 | EXPORT_SYMBOL(alloc_contig_range_noprof);
|
| 7163 |
|
| 7164 | static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
|
| 7165 | unsigned long nr_pages, bool skip_hugetlb,
|
| 7166 | bool *skipped_hugetlb)
|
| 7167 | {
|
| 7168 | unsigned long end_pfn = start_pfn + nr_pages;
|
| 7169 | struct page *page;
|
| 7170 |
|
| 7171 | while (start_pfn < end_pfn) {
|
| 7172 | unsigned long step = 1;
|
| 7173 |
|
| 7174 | page = pfn_to_online_page(start_pfn);
|
| 7175 | if (!page)
|
| 7176 | return false;
|
| 7177 |
|
| 7178 | if (page_zone(page) != z)
|
| 7179 | return false;
|
| 7180 |
|
| 7181 | if (page_is_unmovable(z, page, PB_ISOLATE_MODE_OTHER, &step))
|
| 7182 | return false;
|
| 7183 |
|
| 7184 | |
| 7185 | * Only consider ranges containing hugepages if those pages are
|
| 7186 | * smaller than the requested contiguous region. e.g.:
|
| 7187 | * Move 2MB pages to free up a 1GB range.
|
| 7188 | * Don't move 1GB pages to free up a 2MB range.
|
| 7189 | *
|
| 7190 | * This makes contiguous allocation more reliable if multiple
|
| 7191 | * hugepage sizes are used without causing needless movement.
|
| 7192 |
|
| 7193 | if (PageHuge(page)) {
|
| 7194 | unsigned int order;
|
| 7195 |
|
| 7196 | if (skip_hugetlb) {
|
| 7197 | *skipped_hugetlb = true;
|
| 7198 | return false;
|
| 7199 | }
|
| 7200 |
|
| 7201 | page = compound_head(page);
|
| 7202 | order = compound_order(page);
|
| 7203 | if ((order >= MAX_FOLIO_ORDER) ||
|
| 7204 | (nr_pages <= (1 << order)))
|
| 7205 | return false;
|
| 7206 | }
|
| 7207 |
|
| 7208 | start_pfn += step;
|
| 7209 | }
|
| 7210 | return true;
|
| 7211 | }
|
| 7212 |
|
| 7213 | static bool zone_spans_last_pfn(const struct zone *zone,
|
| 7214 | unsigned long start_pfn, unsigned long nr_pages)
|
| 7215 | {
|
| 7216 | unsigned long last_pfn = start_pfn + nr_pages - 1;
|
| 7217 |
|
| 7218 | return zone_spans_pfn(zone, last_pfn);
|
| 7219 | }
|
| 7220 |
|
| 7221 | |
| 7222 | * alloc_contig_frozen_pages() -- tries to find and allocate contiguous range of frozen pages
|
| 7223 | * @nr_pages: Number of contiguous pages to allocate
|
| 7224 | * @gfp_mask: GFP mask. Node/zone/placement hints limit the search; only some
|
| 7225 | * action and reclaim modifiers are supported. Reclaim modifiers
|
| 7226 | * control allocation behavior during compaction/migration/reclaim.
|
| 7227 | * @nid: Target node
|
| 7228 | * @nodemask: Mask for other possible nodes
|
| 7229 | *
|
| 7230 | * This routine is a wrapper around alloc_contig_frozen_range(). It scans over
|
| 7231 | * zones on an applicable zonelist to find a contiguous pfn range which can then
|
| 7232 | * be tried for allocation with alloc_contig_frozen_range(). This routine is
|
| 7233 | * intended for allocation requests which can not be fulfilled with the buddy
|
| 7234 | * allocator.
|
| 7235 | *
|
| 7236 | * The allocated memory is always aligned to a page boundary. If nr_pages is a
|
| 7237 | * power of two, then allocated range is also guaranteed to be aligned to same
|
| 7238 | * nr_pages (e.g. 1GB request would be aligned to 1GB).
|
| 7239 | *
|
| 7240 | * Allocated frozen pages need be freed with free_contig_frozen_range(),
|
| 7241 | * or by manually calling free_frozen_pages() on each allocated frozen
|
| 7242 | * non-compound page, for compound frozen pages could be freed with
|
| 7243 | * free_frozen_pages() directly.
|
| 7244 | *
|
| 7245 | * Return: pointer to contiguous frozen pages on success, or NULL if not successful.
|
| 7246 |
|
| 7247 | struct page *alloc_contig_frozen_pages_noprof(unsigned long nr_pages,
|
| 7248 | gfp_t gfp_mask, int nid, nodemask_t *nodemask)
|
| 7249 | {
|
| 7250 | unsigned long ret, pfn, flags;
|
| 7251 | struct zonelist *zonelist;
|
| 7252 | struct zone *zone;
|
| 7253 | struct zoneref *z;
|
| 7254 | bool skip_hugetlb = true;
|
| 7255 | bool skipped_hugetlb = false;
|
| 7256 |
|
| 7257 | retry:
|
| 7258 | zonelist = node_zonelist(nid, gfp_mask);
|
| 7259 | for_each_zone_zonelist_nodemask(zone, z, zonelist,
|
| 7260 | gfp_zone(gfp_mask), nodemask) {
|
| 7261 | spin_lock_irqsave(&zone->lock, flags);
|
| 7262 |
|
| 7263 | pfn = ALIGN(zone->zone_start_pfn, nr_pages);
|
| 7264 | while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
|
| 7265 | if (pfn_range_valid_contig(zone, pfn, nr_pages,
|
| 7266 | skip_hugetlb,
|
| 7267 | &skipped_hugetlb)) {
|
| 7268 | |
| 7269 | * We release the zone lock here because
|
| 7270 | * alloc_contig_frozen_range() will also lock
|
| 7271 | * the zone at some point. If there's an
|
| 7272 | * allocation spinning on this lock, it may
|
| 7273 | * win the race and cause allocation to fail.
|
| 7274 |
|
| 7275 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7276 | ret = alloc_contig_frozen_range_noprof(pfn,
|
| 7277 | pfn + nr_pages,
|
| 7278 | ACR_FLAGS_NONE,
|
| 7279 | gfp_mask);
|
| 7280 | if (!ret)
|
| 7281 | return pfn_to_page(pfn);
|
| 7282 | spin_lock_irqsave(&zone->lock, flags);
|
| 7283 | }
|
| 7284 | pfn += nr_pages;
|
| 7285 | }
|
| 7286 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7287 | }
|
| 7288 | |
| 7289 | * If we failed, retry the search, but treat regions with HugeTLB pages
|
| 7290 | * as valid targets. This retains fast-allocations on first pass
|
| 7291 | * without trying to migrate HugeTLB pages (which may fail). On the
|
| 7292 | * second pass, we will try moving HugeTLB pages when those pages are
|
| 7293 | * smaller than the requested contiguous region size.
|
| 7294 |
|
| 7295 | if (skip_hugetlb && skipped_hugetlb) {
|
| 7296 | skip_hugetlb = false;
|
| 7297 | goto retry;
|
| 7298 | }
|
| 7299 | return NULL;
|
| 7300 | }
|
| 7301 | EXPORT_SYMBOL(alloc_contig_frozen_pages_noprof);
|
| 7302 |
|
| 7303 | |
| 7304 | * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
|
| 7305 | * @nr_pages: Number of contiguous pages to allocate
|
| 7306 | * @gfp_mask: GFP mask.
|
| 7307 | * @nid: Target node
|
| 7308 | * @nodemask: Mask for other possible nodes
|
| 7309 | *
|
| 7310 | * This routine is a wrapper around alloc_contig_frozen_pages(), it can't
|
| 7311 | * be used to allocate compound pages, the refcount of each allocated page
|
| 7312 | * will be set to one.
|
| 7313 | *
|
| 7314 | * Allocated pages can be freed with free_contig_range() or by manually
|
| 7315 | * calling __free_page() on each allocated page.
|
| 7316 | *
|
| 7317 | * Return: pointer to contiguous pages on success, or NULL if not successful.
|
| 7318 |
|
| 7319 | struct page *alloc_contig_pages_noprof(unsigned long nr_pages, gfp_t gfp_mask,
|
| 7320 | int nid, nodemask_t *nodemask)
|
| 7321 | {
|
| 7322 | struct page *page;
|
| 7323 |
|
| 7324 | if (WARN_ON(gfp_mask & __GFP_COMP))
|
| 7325 | return NULL;
|
| 7326 |
|
| 7327 | page = alloc_contig_frozen_pages_noprof(nr_pages, gfp_mask, nid,
|
| 7328 | nodemask);
|
| 7329 | if (page)
|
| 7330 | set_pages_refcounted(page, nr_pages);
|
| 7331 |
|
| 7332 | return page;
|
| 7333 | }
|
| 7334 | EXPORT_SYMBOL(alloc_contig_pages_noprof);
|
| 7335 |
|
| 7336 | |
| 7337 | * free_contig_frozen_range() -- free the contiguous range of frozen pages
|
| 7338 | * @pfn: start PFN to free
|
| 7339 | * @nr_pages: Number of contiguous frozen pages to free
|
| 7340 | *
|
| 7341 | * This can be used to free the allocated compound/non-compound frozen pages.
|
| 7342 |
|
| 7343 | void free_contig_frozen_range(unsigned long pfn, unsigned long nr_pages)
|
| 7344 | {
|
| 7345 | struct page *first_page = pfn_to_page(pfn);
|
| 7346 | const unsigned int order = ilog2(nr_pages);
|
| 7347 |
|
| 7348 | if (WARN_ON_ONCE(first_page != compound_head(first_page)))
|
| 7349 | return;
|
| 7350 |
|
| 7351 | if (PageHead(first_page)) {
|
| 7352 | WARN_ON_ONCE(order != compound_order(first_page));
|
| 7353 | free_frozen_pages(first_page, order);
|
| 7354 | return;
|
| 7355 | }
|
| 7356 |
|
| 7357 | __free_contig_frozen_range(pfn, nr_pages);
|
| 7358 | }
|
| 7359 | EXPORT_SYMBOL(free_contig_frozen_range);
|
| 7360 |
|
| 7361 | |
| 7362 | * free_contig_range() -- free the contiguous range of pages
|
| 7363 | * @pfn: start PFN to free
|
| 7364 | * @nr_pages: Number of contiguous pages to free
|
| 7365 | *
|
| 7366 | * This can be only used to free the allocated non-compound pages.
|
| 7367 |
|
| 7368 | void free_contig_range(unsigned long pfn, unsigned long nr_pages)
|
| 7369 | {
|
| 7370 | if (WARN_ON_ONCE(PageHead(pfn_to_page(pfn))))
|
| 7371 | return;
|
| 7372 |
|
| 7373 | for (; nr_pages--; pfn++)
|
| 7374 | __free_page(pfn_to_page(pfn));
|
| 7375 | }
|
| 7376 | EXPORT_SYMBOL(free_contig_range);
|
| 7377 | #endif
|
| 7378 |
|
| 7379 | |
| 7380 | * Effectively disable pcplists for the zone by setting the high limit to 0
|
| 7381 | * and draining all cpus. A concurrent page freeing on another CPU that's about
|
| 7382 | * to put the page on pcplist will either finish before the drain and the page
|
| 7383 | * will be drained, or observe the new high limit and skip the pcplist.
|
| 7384 | *
|
| 7385 | * Must be paired with a call to zone_pcp_enable().
|
| 7386 |
|
| 7387 | void zone_pcp_disable(struct zone *zone)
|
| 7388 | {
|
| 7389 | mutex_lock(&pcp_batch_high_lock);
|
| 7390 | __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
|
| 7391 | __drain_all_pages(zone, true);
|
| 7392 | }
|
| 7393 |
|
| 7394 | void zone_pcp_enable(struct zone *zone)
|
| 7395 | {
|
| 7396 | __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
|
| 7397 | zone->pageset_high_max, zone->pageset_batch);
|
| 7398 | mutex_unlock(&pcp_batch_high_lock);
|
| 7399 | }
|
| 7400 |
|
| 7401 | void zone_pcp_reset(struct zone *zone)
|
| 7402 | {
|
| 7403 | int cpu;
|
| 7404 | struct per_cpu_zonestat *pzstats;
|
| 7405 |
|
| 7406 | if (zone->per_cpu_pageset != &boot_pageset) {
|
| 7407 | for_each_online_cpu(cpu) {
|
| 7408 | pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
|
| 7409 | drain_zonestat(zone, pzstats);
|
| 7410 | }
|
| 7411 | free_percpu(zone->per_cpu_pageset);
|
| 7412 | zone->per_cpu_pageset = &boot_pageset;
|
| 7413 | if (zone->per_cpu_zonestats != &boot_zonestats) {
|
| 7414 | free_percpu(zone->per_cpu_zonestats);
|
| 7415 | zone->per_cpu_zonestats = &boot_zonestats;
|
| 7416 | }
|
| 7417 | }
|
| 7418 | }
|
| 7419 |
|
| 7420 | #ifdef CONFIG_MEMORY_HOTREMOVE
|
| 7421 | |
| 7422 | * All pages in the range must be in a single zone, must not contain holes,
|
| 7423 | * must span full sections, and must be isolated before calling this function.
|
| 7424 | *
|
| 7425 | * Returns the number of managed (non-PageOffline()) pages in the range: the
|
| 7426 | * number of pages for which memory offlining code must adjust managed page
|
| 7427 | * counters using adjust_managed_page_count().
|
| 7428 |
|
| 7429 | unsigned long __offline_isolated_pages(unsigned long start_pfn,
|
| 7430 | unsigned long end_pfn)
|
| 7431 | {
|
| 7432 | unsigned long already_offline = 0, flags;
|
| 7433 | unsigned long pfn = start_pfn;
|
| 7434 | struct page *page;
|
| 7435 | struct zone *zone;
|
| 7436 | unsigned int order;
|
| 7437 |
|
| 7438 | offline_mem_sections(pfn, end_pfn);
|
| 7439 | zone = page_zone(pfn_to_page(pfn));
|
| 7440 | spin_lock_irqsave(&zone->lock, flags);
|
| 7441 | while (pfn < end_pfn) {
|
| 7442 | page = pfn_to_page(pfn);
|
| 7443 | |
| 7444 | * The HWPoisoned page may be not in buddy system, and
|
| 7445 | * page_count() is not 0.
|
| 7446 |
|
| 7447 | if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
|
| 7448 | pfn++;
|
| 7449 | continue;
|
| 7450 | }
|
| 7451 | |
| 7452 | * At this point all remaining PageOffline() pages have a
|
| 7453 | * reference count of 0 and can simply be skipped.
|
| 7454 |
|
| 7455 | if (PageOffline(page)) {
|
| 7456 | BUG_ON(page_count(page));
|
| 7457 | BUG_ON(PageBuddy(page));
|
| 7458 | already_offline++;
|
| 7459 | pfn++;
|
| 7460 | continue;
|
| 7461 | }
|
| 7462 |
|
| 7463 | BUG_ON(page_count(page));
|
| 7464 | BUG_ON(!PageBuddy(page));
|
| 7465 | VM_WARN_ON(get_pageblock_migratetype(page) != MIGRATE_ISOLATE);
|
| 7466 | order = buddy_order(page);
|
| 7467 | del_page_from_free_list(page, zone, order, MIGRATE_ISOLATE);
|
| 7468 | pfn += (1 << order);
|
| 7469 | }
|
| 7470 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7471 |
|
| 7472 | return end_pfn - start_pfn - already_offline;
|
| 7473 | }
|
| 7474 | #endif
|
| 7475 |
|
| 7476 | |
| 7477 | * This function returns a stable result only if called under zone lock.
|
| 7478 |
|
| 7479 | bool is_free_buddy_page(const struct page *page)
|
| 7480 | {
|
| 7481 | unsigned long pfn = page_to_pfn(page);
|
| 7482 | unsigned int order;
|
| 7483 |
|
| 7484 | for (order = 0; order < NR_PAGE_ORDERS; order++) {
|
| 7485 | const struct page *head = page - (pfn & ((1 << order) - 1));
|
| 7486 |
|
| 7487 | if (PageBuddy(head) &&
|
| 7488 | buddy_order_unsafe(head) >= order)
|
| 7489 | break;
|
| 7490 | }
|
| 7491 |
|
| 7492 | return order <= MAX_PAGE_ORDER;
|
| 7493 | }
|
| 7494 | EXPORT_SYMBOL(is_free_buddy_page);
|
| 7495 |
|
| 7496 | #ifdef CONFIG_MEMORY_FAILURE
|
| 7497 | static inline void add_to_free_list(struct page *page, struct zone *zone,
|
| 7498 | unsigned int order, int migratetype,
|
| 7499 | bool tail)
|
| 7500 | {
|
| 7501 | __add_to_free_list(page, zone, order, migratetype, tail);
|
| 7502 | account_freepages(zone, 1 << order, migratetype);
|
| 7503 | }
|
| 7504 |
|
| 7505 | |
| 7506 | * Break down a higher-order page in sub-pages, and keep our target out of
|
| 7507 | * buddy allocator.
|
| 7508 |
|
| 7509 | static void break_down_buddy_pages(struct zone *zone, struct page *page,
|
| 7510 | struct page *target, int low, int high,
|
| 7511 | int migratetype)
|
| 7512 | {
|
| 7513 | unsigned long size = 1 << high;
|
| 7514 | struct page *current_buddy;
|
| 7515 |
|
| 7516 | while (high > low) {
|
| 7517 | high--;
|
| 7518 | size >>= 1;
|
| 7519 |
|
| 7520 | if (target >= &page[size]) {
|
| 7521 | current_buddy = page;
|
| 7522 | page = page + size;
|
| 7523 | } else {
|
| 7524 | current_buddy = page + size;
|
| 7525 | }
|
| 7526 |
|
| 7527 | if (set_page_guard(zone, current_buddy, high))
|
| 7528 | continue;
|
| 7529 |
|
| 7530 | add_to_free_list(current_buddy, zone, high, migratetype, false);
|
| 7531 | set_buddy_order(current_buddy, high);
|
| 7532 | }
|
| 7533 | }
|
| 7534 |
|
| 7535 | |
| 7536 | * Take a page that will be marked as poisoned off the buddy allocator.
|
| 7537 |
|
| 7538 | bool take_page_off_buddy(struct page *page)
|
| 7539 | {
|
| 7540 | struct zone *zone = page_zone(page);
|
| 7541 | unsigned long pfn = page_to_pfn(page);
|
| 7542 | unsigned long flags;
|
| 7543 | unsigned int order;
|
| 7544 | bool ret = false;
|
| 7545 |
|
| 7546 | spin_lock_irqsave(&zone->lock, flags);
|
| 7547 | for (order = 0; order < NR_PAGE_ORDERS; order++) {
|
| 7548 | struct page *page_head = page - (pfn & ((1 << order) - 1));
|
| 7549 | int page_order = buddy_order(page_head);
|
| 7550 |
|
| 7551 | if (PageBuddy(page_head) && page_order >= order) {
|
| 7552 | unsigned long pfn_head = page_to_pfn(page_head);
|
| 7553 | int migratetype = get_pfnblock_migratetype(page_head,
|
| 7554 | pfn_head);
|
| 7555 |
|
| 7556 | del_page_from_free_list(page_head, zone, page_order,
|
| 7557 | migratetype);
|
| 7558 | break_down_buddy_pages(zone, page_head, page, 0,
|
| 7559 | page_order, migratetype);
|
| 7560 | SetPageHWPoisonTakenOff(page);
|
| 7561 | ret = true;
|
| 7562 | break;
|
| 7563 | }
|
| 7564 | if (page_count(page_head) > 0)
|
| 7565 | break;
|
| 7566 | }
|
| 7567 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7568 | return ret;
|
| 7569 | }
|
| 7570 |
|
| 7571 | |
| 7572 | * Cancel takeoff done by take_page_off_buddy().
|
| 7573 |
|
| 7574 | bool put_page_back_buddy(struct page *page)
|
| 7575 | {
|
| 7576 | struct zone *zone = page_zone(page);
|
| 7577 | unsigned long flags;
|
| 7578 | bool ret = false;
|
| 7579 |
|
| 7580 | spin_lock_irqsave(&zone->lock, flags);
|
| 7581 | if (put_page_testzero(page)) {
|
| 7582 | unsigned long pfn = page_to_pfn(page);
|
| 7583 | int migratetype = get_pfnblock_migratetype(page, pfn);
|
| 7584 |
|
| 7585 | ClearPageHWPoisonTakenOff(page);
|
| 7586 | __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
|
| 7587 | if (TestClearPageHWPoison(page)) {
|
| 7588 | ret = true;
|
| 7589 | }
|
| 7590 | }
|
| 7591 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7592 |
|
| 7593 | return ret;
|
| 7594 | }
|
| 7595 | #endif
|
| 7596 |
|
| 7597 | bool has_managed_zone(enum zone_type zone)
|
| 7598 | {
|
| 7599 | struct pglist_data *pgdat;
|
| 7600 |
|
| 7601 | for_each_online_pgdat(pgdat) {
|
| 7602 | if (managed_zone(&pgdat->node_zones[zone]))
|
| 7603 | return true;
|
| 7604 | }
|
| 7605 | return false;
|
| 7606 | }
|
| 7607 |
|
| 7608 | #ifdef CONFIG_UNACCEPTED_MEMORY
|
| 7609 |
|
| 7610 | static bool lazy_accept = true;
|
| 7611 |
|
| 7612 | static int __init accept_memory_parse(char *p)
|
| 7613 | {
|
| 7614 | if (!strcmp(p, "lazy")) {
|
| 7615 | lazy_accept = true;
|
| 7616 | return 0;
|
| 7617 | } else if (!strcmp(p, "eager")) {
|
| 7618 | lazy_accept = false;
|
| 7619 | return 0;
|
| 7620 | } else {
|
| 7621 | return -EINVAL;
|
| 7622 | }
|
| 7623 | }
|
| 7624 | early_param("accept_memory", accept_memory_parse);
|
| 7625 |
|
| 7626 | static bool page_contains_unaccepted(struct page *page, unsigned int order)
|
| 7627 | {
|
| 7628 | phys_addr_t start = page_to_phys(page);
|
| 7629 |
|
| 7630 | return range_contains_unaccepted_memory(start, PAGE_SIZE << order);
|
| 7631 | }
|
| 7632 |
|
| 7633 | static void __accept_page(struct zone *zone, unsigned long *flags,
|
| 7634 | struct page *page)
|
| 7635 | {
|
| 7636 | list_del(&page->lru);
|
| 7637 | account_freepages(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
|
| 7638 | __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
|
| 7639 | __ClearPageUnaccepted(page);
|
| 7640 | spin_unlock_irqrestore(&zone->lock, *flags);
|
| 7641 |
|
| 7642 | accept_memory(page_to_phys(page), PAGE_SIZE << MAX_PAGE_ORDER);
|
| 7643 |
|
| 7644 | __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
|
| 7645 | }
|
| 7646 |
|
| 7647 | void accept_page(struct page *page)
|
| 7648 | {
|
| 7649 | struct zone *zone = page_zone(page);
|
| 7650 | unsigned long flags;
|
| 7651 |
|
| 7652 | spin_lock_irqsave(&zone->lock, flags);
|
| 7653 | if (!PageUnaccepted(page)) {
|
| 7654 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7655 | return;
|
| 7656 | }
|
| 7657 |
|
| 7658 |
|
| 7659 | __accept_page(zone, &flags, page);
|
| 7660 | }
|
| 7661 |
|
| 7662 | static bool try_to_accept_memory_one(struct zone *zone)
|
| 7663 | {
|
| 7664 | unsigned long flags;
|
| 7665 | struct page *page;
|
| 7666 |
|
| 7667 | spin_lock_irqsave(&zone->lock, flags);
|
| 7668 | page = list_first_entry_or_null(&zone->unaccepted_pages,
|
| 7669 | struct page, lru);
|
| 7670 | if (!page) {
|
| 7671 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7672 | return false;
|
| 7673 | }
|
| 7674 |
|
| 7675 |
|
| 7676 | __accept_page(zone, &flags, page);
|
| 7677 |
|
| 7678 | return true;
|
| 7679 | }
|
| 7680 |
|
| 7681 | static bool cond_accept_memory(struct zone *zone, unsigned int order,
|
| 7682 | int alloc_flags)
|
| 7683 | {
|
| 7684 | long to_accept, wmark;
|
| 7685 | bool ret = false;
|
| 7686 |
|
| 7687 | if (list_empty(&zone->unaccepted_pages))
|
| 7688 | return false;
|
| 7689 |
|
| 7690 |
|
| 7691 | if (alloc_flags & ALLOC_TRYLOCK)
|
| 7692 | return false;
|
| 7693 |
|
| 7694 | wmark = promo_wmark_pages(zone);
|
| 7695 |
|
| 7696 | |
| 7697 | * Watermarks have not been initialized yet.
|
| 7698 | *
|
| 7699 | * Accepting one MAX_ORDER page to ensure progress.
|
| 7700 |
|
| 7701 | if (!wmark)
|
| 7702 | return try_to_accept_memory_one(zone);
|
| 7703 |
|
| 7704 |
|
| 7705 | to_accept = wmark -
|
| 7706 | (zone_page_state(zone, NR_FREE_PAGES) -
|
| 7707 | __zone_watermark_unusable_free(zone, order, 0) -
|
| 7708 | zone_page_state(zone, NR_UNACCEPTED));
|
| 7709 |
|
| 7710 | while (to_accept > 0) {
|
| 7711 | if (!try_to_accept_memory_one(zone))
|
| 7712 | break;
|
| 7713 | ret = true;
|
| 7714 | to_accept -= MAX_ORDER_NR_PAGES;
|
| 7715 | }
|
| 7716 |
|
| 7717 | return ret;
|
| 7718 | }
|
| 7719 |
|
| 7720 | static bool __free_unaccepted(struct page *page)
|
| 7721 | {
|
| 7722 | struct zone *zone = page_zone(page);
|
| 7723 | unsigned long flags;
|
| 7724 |
|
| 7725 | if (!lazy_accept)
|
| 7726 | return false;
|
| 7727 |
|
| 7728 | spin_lock_irqsave(&zone->lock, flags);
|
| 7729 | list_add_tail(&page->lru, &zone->unaccepted_pages);
|
| 7730 | account_freepages(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
|
| 7731 | __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
|
| 7732 | __SetPageUnaccepted(page);
|
| 7733 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7734 |
|
| 7735 | return true;
|
| 7736 | }
|
| 7737 |
|
| 7738 | #else
|
| 7739 |
|
| 7740 | static bool page_contains_unaccepted(struct page *page, unsigned int order)
|
| 7741 | {
|
| 7742 | return false;
|
| 7743 | }
|
| 7744 |
|
| 7745 | static bool cond_accept_memory(struct zone *zone, unsigned int order,
|
| 7746 | int alloc_flags)
|
| 7747 | {
|
| 7748 | return false;
|
| 7749 | }
|
| 7750 |
|
| 7751 | static bool __free_unaccepted(struct page *page)
|
| 7752 | {
|
| 7753 | BUILD_BUG();
|
| 7754 | return false;
|
| 7755 | }
|
| 7756 |
|
| 7757 | #endif
|
| 7758 |
|
| 7759 | struct page *alloc_frozen_pages_nolock_noprof(gfp_t gfp_flags, int nid, unsigned int order)
|
| 7760 | {
|
| 7761 | |
| 7762 | * Do not specify __GFP_DIRECT_RECLAIM, since direct claim is not allowed.
|
| 7763 | * Do not specify __GFP_KSWAPD_RECLAIM either, since wake up of kswapd
|
| 7764 | * is not safe in arbitrary context.
|
| 7765 | *
|
| 7766 | * These two are the conditions for gfpflags_allow_spinning() being true.
|
| 7767 | *
|
| 7768 | * Specify __GFP_NOWARN since failing alloc_pages_nolock() is not a reason
|
| 7769 | * to warn. Also warn would trigger printk() which is unsafe from
|
| 7770 | * various contexts. We cannot use printk_deferred_enter() to mitigate,
|
| 7771 | * since the running context is unknown.
|
| 7772 | *
|
| 7773 | * Specify __GFP_ZERO to make sure that call to kmsan_alloc_page() below
|
| 7774 | * is safe in any context. Also zeroing the page is mandatory for
|
| 7775 | * BPF use cases.
|
| 7776 | *
|
| 7777 | * Though __GFP_NOMEMALLOC is not checked in the code path below,
|
| 7778 | * specify it here to highlight that alloc_pages_nolock()
|
| 7779 | * doesn't want to deplete reserves.
|
| 7780 |
|
| 7781 | gfp_t alloc_gfp = __GFP_NOWARN | __GFP_ZERO | __GFP_NOMEMALLOC | __GFP_COMP
|
| 7782 | | gfp_flags;
|
| 7783 | unsigned int alloc_flags = ALLOC_TRYLOCK;
|
| 7784 | struct alloc_context ac = { };
|
| 7785 | struct page *page;
|
| 7786 |
|
| 7787 | VM_WARN_ON_ONCE(gfp_flags & ~__GFP_ACCOUNT);
|
| 7788 | |
| 7789 | * In PREEMPT_RT spin_trylock() will call raw_spin_lock() which is
|
| 7790 | * unsafe in NMI. If spin_trylock() is called from hard IRQ the current
|
| 7791 | * task may be waiting for one rt_spin_lock, but rt_spin_trylock() will
|
| 7792 | * mark the task as the owner of another rt_spin_lock which will
|
| 7793 | * confuse PI logic, so return immediately if called from hard IRQ or
|
| 7794 | * NMI.
|
| 7795 | *
|
| 7796 | * Note, irqs_disabled() case is ok. This function can be called
|
| 7797 | * from raw_spin_lock_irqsave region.
|
| 7798 |
|
| 7799 | if (IS_ENABLED(CONFIG_PREEMPT_RT) && (in_nmi() || in_hardirq()))
|
| 7800 | return NULL;
|
| 7801 | if (!pcp_allowed_order(order))
|
| 7802 | return NULL;
|
| 7803 |
|
| 7804 |
|
| 7805 | if (deferred_pages_enabled())
|
| 7806 | return NULL;
|
| 7807 |
|
| 7808 | if (nid == NUMA_NO_NODE)
|
| 7809 | nid = numa_node_id();
|
| 7810 |
|
| 7811 | prepare_alloc_pages(alloc_gfp, order, nid, NULL, &ac,
|
| 7812 | &alloc_gfp, &alloc_flags);
|
| 7813 |
|
| 7814 | |
| 7815 | * Best effort allocation from percpu free list.
|
| 7816 | * If it's empty attempt to spin_trylock zone->lock.
|
| 7817 |
|
| 7818 | page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
|
| 7819 |
|
| 7820 |
|
| 7821 |
|
| 7822 | if (memcg_kmem_online() && page && (gfp_flags & __GFP_ACCOUNT) &&
|
| 7823 | unlikely(__memcg_kmem_charge_page(page, alloc_gfp, order) != 0)) {
|
| 7824 | __free_frozen_pages(page, order, FPI_TRYLOCK);
|
| 7825 | page = NULL;
|
| 7826 | }
|
| 7827 | trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
|
| 7828 | kmsan_alloc_page(page, order, alloc_gfp);
|
| 7829 | return page;
|
| 7830 | }
|
| 7831 | |
| 7832 | * alloc_pages_nolock - opportunistic reentrant allocation from any context
|
| 7833 | * @gfp_flags: GFP flags. Only __GFP_ACCOUNT allowed.
|
| 7834 | * @nid: node to allocate from
|
| 7835 | * @order: allocation order size
|
| 7836 | *
|
| 7837 | * Allocates pages of a given order from the given node. This is safe to
|
| 7838 | * call from any context (from atomic, NMI, and also reentrant
|
| 7839 | * allocator -> tracepoint -> alloc_pages_nolock_noprof).
|
| 7840 | * Allocation is best effort and to be expected to fail easily so nobody should
|
| 7841 | * rely on the success. Failures are not reported via warn_alloc().
|
| 7842 | * See always fail conditions below.
|
| 7843 | *
|
| 7844 | * Return: allocated page or NULL on failure. NULL does not mean EBUSY or EAGAIN.
|
| 7845 | * It means ENOMEM. There is no reason to call it again and expect !NULL.
|
| 7846 |
|
| 7847 | struct page *alloc_pages_nolock_noprof(gfp_t gfp_flags, int nid, unsigned int order)
|
| 7848 | {
|
| 7849 | struct page *page;
|
| 7850 |
|
| 7851 | page = alloc_frozen_pages_nolock_noprof(gfp_flags, nid, order);
|
| 7852 | if (page)
|
| 7853 | set_page_refcounted(page);
|
| 7854 | return page;
|
| 7855 | }
|
| 7856 | EXPORT_SYMBOL_GPL(alloc_pages_nolock_noprof);
|