| 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/folio_batch.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 "mm_init.h"
|
| 60 | #include "page_alloc.h"
|
| 61 | #include "shuffle.h"
|
| 62 | #include "page_reporting.h"
|
| 63 |
|
| 64 |
|
| 65 | typedef int __bitwise fpi_t;
|
| 66 |
|
| 67 |
|
| 68 | #define FPI_NONE ((__force fpi_t)0)
|
| 69 |
|
| 70 | |
| 71 | * Skip free page reporting notification for the (possibly merged) page.
|
| 72 | * This does not hinder free page reporting from grabbing the page,
|
| 73 | * reporting it and marking it "reported" - it only skips notifying
|
| 74 | * the free page reporting infrastructure about a newly freed page. For
|
| 75 | * example, used when temporarily pulling a page from a freelist and
|
| 76 | * putting it back unmodified.
|
| 77 |
|
| 78 | #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
|
| 79 |
|
| 80 | |
| 81 | * Place the (possibly merged) page to the tail of the freelist. Will ignore
|
| 82 | * page shuffling (relevant code - e.g., memory onlining - is expected to
|
| 83 | * shuffle the whole zone).
|
| 84 | *
|
| 85 | * Note: No code should rely on this flag for correctness - it's purely
|
| 86 | * to allow for optimizations when handing back either fresh pages
|
| 87 | * (memory onlining) or untouched pages (page isolation, free page
|
| 88 | * reporting).
|
| 89 |
|
| 90 | #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
|
| 91 |
|
| 92 |
|
| 93 | #define FPI_NOLOCK ((__force fpi_t)BIT(2))
|
| 94 |
|
| 95 |
|
| 96 | #define FPI_PREPARED ((__force fpi_t)BIT(3))
|
| 97 |
|
| 98 |
|
| 99 | static DEFINE_MUTEX(pcp_batch_high_lock);
|
| 100 | #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
|
| 101 |
|
| 102 | |
| 103 | * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
|
| 104 | * a migration causing the wrong PCP to be locked and remote memory being
|
| 105 | * potentially allocated, pin the task to the CPU for the lookup+lock.
|
| 106 | * preempt_disable is used on !RT because it is faster than migrate_disable.
|
| 107 | * migrate_disable is used on RT because otherwise RT spinlock usage is
|
| 108 | * interfered with and a high priority task cannot preempt the allocator.
|
| 109 |
|
| 110 | #ifndef CONFIG_PREEMPT_RT
|
| 111 | #define pcpu_task_pin() preempt_disable()
|
| 112 | #define pcpu_task_unpin() preempt_enable()
|
| 113 | #else
|
| 114 | #define pcpu_task_pin() migrate_disable()
|
| 115 | #define pcpu_task_unpin() migrate_enable()
|
| 116 | #endif
|
| 117 |
|
| 118 | |
| 119 | * A helper to lookup and trylock pcp with embedded spinlock.
|
| 120 | * The return value should be used with the unlock helper.
|
| 121 | * NULL return value means the trylock failed.
|
| 122 |
|
| 123 | #ifdef CONFIG_SMP
|
| 124 | #define pcp_spin_trylock(ptr) \
|
| 125 | ({ \
|
| 126 | struct per_cpu_pages *_ret; \
|
| 127 | pcpu_task_pin(); \
|
| 128 | _ret = this_cpu_ptr(ptr); \
|
| 129 | if (!spin_trylock(&_ret->lock)) { \
|
| 130 | pcpu_task_unpin(); \
|
| 131 | _ret = NULL; \
|
| 132 | } \
|
| 133 | _ret; \
|
| 134 | })
|
| 135 |
|
| 136 | #define pcp_spin_unlock(ptr) \
|
| 137 | ({ \
|
| 138 | spin_unlock(&ptr->lock); \
|
| 139 | pcpu_task_unpin(); \
|
| 140 | })
|
| 141 |
|
| 142 | |
| 143 | * On CONFIG_SMP=n the UP implementation of spin_trylock() never fails and thus
|
| 144 | * is not compatible with our locking scheme. However we do not need pcp for
|
| 145 | * scalability in the first place, so just make all the trylocks fail and take
|
| 146 | * the slow path unconditionally.
|
| 147 |
|
| 148 | #else
|
| 149 | #define pcp_spin_trylock(ptr) \
|
| 150 | NULL
|
| 151 |
|
| 152 | #define pcp_spin_unlock(ptr) \
|
| 153 | BUG_ON(1)
|
| 154 | #endif
|
| 155 |
|
| 156 | |
| 157 | * In some cases we do not need to pin the task to the CPU because we are
|
| 158 | * already given a specific cpu's pcp pointer.
|
| 159 |
|
| 160 | #define pcp_spin_lock_nopin(ptr) \
|
| 161 | spin_lock(&(ptr)->lock)
|
| 162 | #define pcp_spin_unlock_nopin(ptr) \
|
| 163 | spin_unlock(&(ptr)->lock)
|
| 164 |
|
| 165 | #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
|
| 166 | DEFINE_PER_CPU(int, numa_node);
|
| 167 | EXPORT_PER_CPU_SYMBOL(numa_node);
|
| 168 | #endif
|
| 169 |
|
| 170 | #ifdef CONFIG_NUMA
|
| 171 | DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
|
| 172 | #endif
|
| 173 |
|
| 174 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES
|
| 175 | |
| 176 | * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
|
| 177 | * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
|
| 178 | * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
|
| 179 | * defined in <linux/topology.h>.
|
| 180 |
|
| 181 | DEFINE_PER_CPU(int, _numa_mem_);
|
| 182 | EXPORT_PER_CPU_SYMBOL(_numa_mem_);
|
| 183 | #endif
|
| 184 |
|
| 185 | static DEFINE_MUTEX(pcpu_drain_mutex);
|
| 186 |
|
| 187 | #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
|
| 188 | volatile unsigned long latent_entropy __latent_entropy;
|
| 189 | EXPORT_SYMBOL(latent_entropy);
|
| 190 | #endif
|
| 191 |
|
| 192 | |
| 193 | * Array of node states.
|
| 194 |
|
| 195 | nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
|
| 196 | [N_POSSIBLE] = NODE_MASK_ALL,
|
| 197 | [N_ONLINE] = { { [0] = 1UL } },
|
| 198 | #ifndef CONFIG_NUMA
|
| 199 | [N_NORMAL_MEMORY] = { { [0] = 1UL } },
|
| 200 | #ifdef CONFIG_HIGHMEM
|
| 201 | [N_HIGH_MEMORY] = { { [0] = 1UL } },
|
| 202 | #endif
|
| 203 | [N_MEMORY] = { { [0] = 1UL } },
|
| 204 | [N_CPU] = { { [0] = 1UL } },
|
| 205 | #endif
|
| 206 | };
|
| 207 | EXPORT_SYMBOL(node_states);
|
| 208 |
|
| 209 | gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
|
| 210 |
|
| 211 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
|
| 212 | unsigned int pageblock_order __read_mostly;
|
| 213 | #endif
|
| 214 |
|
| 215 | static void __free_pages_ok(struct page *page, unsigned int order,
|
| 216 | fpi_t fpi_flags);
|
| 217 | static void reserve_highatomic_pageblock(struct page *page, int order,
|
| 218 | struct zone *zone);
|
| 219 |
|
| 220 | |
| 221 | * results with 256, 32 in the lowmem_reserve sysctl:
|
| 222 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
|
| 223 | * 1G machine -> (16M dma, 784M normal, 224M high)
|
| 224 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
|
| 225 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
|
| 226 | * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
|
| 227 | *
|
| 228 | * TBD: should special case ZONE_DMA32 machines here - in those we normally
|
| 229 | * don't need any ZONE_NORMAL reservation
|
| 230 |
|
| 231 | static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
|
| 232 | #ifdef CONFIG_ZONE_DMA
|
| 233 | [ZONE_DMA] = 256,
|
| 234 | #endif
|
| 235 | #ifdef CONFIG_ZONE_DMA32
|
| 236 | [ZONE_DMA32] = 256,
|
| 237 | #endif
|
| 238 | [ZONE_NORMAL] = 32,
|
| 239 | #ifdef CONFIG_HIGHMEM
|
| 240 | [ZONE_HIGHMEM] = 0,
|
| 241 | #endif
|
| 242 | [ZONE_MOVABLE] = 0,
|
| 243 | };
|
| 244 |
|
| 245 | char * const zone_names[MAX_NR_ZONES] = {
|
| 246 | #ifdef CONFIG_ZONE_DMA
|
| 247 | "DMA",
|
| 248 | #endif
|
| 249 | #ifdef CONFIG_ZONE_DMA32
|
| 250 | "DMA32",
|
| 251 | #endif
|
| 252 | "Normal",
|
| 253 | #ifdef CONFIG_HIGHMEM
|
| 254 | "HighMem",
|
| 255 | #endif
|
| 256 | "Movable",
|
| 257 | #ifdef CONFIG_ZONE_DEVICE
|
| 258 | "Device",
|
| 259 | #endif
|
| 260 | };
|
| 261 |
|
| 262 | const char * const migratetype_names[MIGRATE_TYPES] = {
|
| 263 | "Unmovable",
|
| 264 | "Movable",
|
| 265 | "Reclaimable",
|
| 266 | "HighAtomic",
|
| 267 | #ifdef CONFIG_CMA
|
| 268 | "CMA",
|
| 269 | #endif
|
| 270 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 271 | "Isolate",
|
| 272 | #endif
|
| 273 | };
|
| 274 |
|
| 275 | int min_free_kbytes = 1024;
|
| 276 | int user_min_free_kbytes = -1;
|
| 277 | static int watermark_boost_factor __read_mostly = 15000;
|
| 278 | static int watermark_scale_factor = 10;
|
| 279 | int defrag_mode;
|
| 280 |
|
| 281 |
|
| 282 | int movable_zone;
|
| 283 | EXPORT_SYMBOL(movable_zone);
|
| 284 |
|
| 285 | #if MAX_NUMNODES > 1
|
| 286 | unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
|
| 287 | unsigned int nr_online_nodes __read_mostly = 1;
|
| 288 | EXPORT_SYMBOL(nr_node_ids);
|
| 289 | EXPORT_SYMBOL(nr_online_nodes);
|
| 290 | #endif
|
| 291 |
|
| 292 | |
| 293 | * When page allocations stall for longer than a threshold,
|
| 294 | * ALLOC_STALL_WARN_MSECS, leave a warning in the kernel log. Only one warning
|
| 295 | * will be printed during this duration for the entire system.
|
| 296 |
|
| 297 | #define ALLOC_STALL_WARN_MSECS (10 * 1000UL)
|
| 298 | static unsigned long alloc_stall_warn_jiffies = INITIAL_JIFFIES;
|
| 299 |
|
| 300 | static bool page_contains_unaccepted(struct page *page, unsigned int order);
|
| 301 | static bool cond_accept_memory(struct zone *zone, unsigned int order,
|
| 302 | int alloc_flags);
|
| 303 | static bool __free_unaccepted(struct page *page);
|
| 304 |
|
| 305 | int page_group_by_mobility_disabled __read_mostly;
|
| 306 |
|
| 307 | #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
|
| 308 | |
| 309 | * During boot we initialize deferred pages on-demand, as needed, but once
|
| 310 | * page_alloc_init_late() has finished, the deferred pages are all initialized,
|
| 311 | * and we can permanently disable that path.
|
| 312 |
|
| 313 | DEFINE_STATIC_KEY_TRUE(deferred_pages);
|
| 314 |
|
| 315 | |
| 316 | * deferred_grow_zone() is __init, but it is called from
|
| 317 | * get_page_from_freelist() during early boot until deferred_pages permanently
|
| 318 | * disables this call. This is why we have refdata wrapper to avoid warning,
|
| 319 | * and to ensure that the function body gets unloaded.
|
| 320 |
|
| 321 | static bool __ref
|
| 322 | _deferred_grow_zone(struct zone *zone, unsigned int order)
|
| 323 | {
|
| 324 | return deferred_grow_zone(zone, order);
|
| 325 | }
|
| 326 | #else
|
| 327 | static inline bool _deferred_grow_zone(struct zone *zone, unsigned int order)
|
| 328 | {
|
| 329 | return false;
|
| 330 | }
|
| 331 | #endif
|
| 332 |
|
| 333 |
|
| 334 | static inline unsigned long *get_pageblock_bitmap(const struct page *page,
|
| 335 | unsigned long pfn)
|
| 336 | {
|
| 337 | #ifdef CONFIG_SPARSEMEM
|
| 338 | return section_to_usemap(__pfn_to_section(pfn));
|
| 339 | #else
|
| 340 | return page_zone(page)->pageblock_flags;
|
| 341 | #endif
|
| 342 | }
|
| 343 |
|
| 344 | static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
|
| 345 | {
|
| 346 | #ifdef CONFIG_SPARSEMEM
|
| 347 | pfn &= (PAGES_PER_SECTION-1);
|
| 348 | #else
|
| 349 | pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
|
| 350 | #endif
|
| 351 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
|
| 352 | }
|
| 353 |
|
| 354 | static __always_inline bool is_standalone_pb_bit(enum pageblock_bits pb_bit)
|
| 355 | {
|
| 356 | return pb_bit >= PB_compact_skip && pb_bit < __NR_PAGEBLOCK_BITS;
|
| 357 | }
|
| 358 |
|
| 359 | static __always_inline void
|
| 360 | get_pfnblock_bitmap_bitidx(const struct page *page, unsigned long pfn,
|
| 361 | unsigned long **bitmap_word, unsigned long *bitidx)
|
| 362 | {
|
| 363 | unsigned long *bitmap;
|
| 364 | unsigned long word_bitidx;
|
| 365 |
|
| 366 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 367 | BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 8);
|
| 368 | #else
|
| 369 | BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
|
| 370 | #endif
|
| 371 | BUILD_BUG_ON(__MIGRATE_TYPE_END > PAGEBLOCK_MIGRATETYPE_MASK);
|
| 372 | VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
|
| 373 |
|
| 374 | bitmap = get_pageblock_bitmap(page, pfn);
|
| 375 | *bitidx = pfn_to_bitidx(page, pfn);
|
| 376 | word_bitidx = *bitidx / BITS_PER_LONG;
|
| 377 | *bitidx &= (BITS_PER_LONG - 1);
|
| 378 | *bitmap_word = &bitmap[word_bitidx];
|
| 379 | }
|
| 380 |
|
| 381 |
|
| 382 | |
| 383 | * __get_pfnblock_flags_mask - Return the requested group of flags for
|
| 384 | * a pageblock_nr_pages block of pages
|
| 385 | * @page: The page within the block of interest
|
| 386 | * @pfn: The target page frame number
|
| 387 | * @mask: mask of bits that the caller is interested in
|
| 388 | *
|
| 389 | * Return: pageblock_bits flags
|
| 390 |
|
| 391 | static unsigned long __get_pfnblock_flags_mask(const struct page *page,
|
| 392 | unsigned long pfn,
|
| 393 | unsigned long mask)
|
| 394 | {
|
| 395 | unsigned long *bitmap_word;
|
| 396 | unsigned long bitidx;
|
| 397 | unsigned long word;
|
| 398 |
|
| 399 | get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
|
| 400 | |
| 401 | * This races, without locks, with set_pfnblock_migratetype(). Ensure
|
| 402 | * a consistent read of the memory array, so that results, even though
|
| 403 | * racy, are not corrupted.
|
| 404 |
|
| 405 | word = READ_ONCE(*bitmap_word);
|
| 406 | return (word >> bitidx) & mask;
|
| 407 | }
|
| 408 |
|
| 409 | |
| 410 | * get_pfnblock_bit - Check if a standalone bit of a pageblock is set
|
| 411 | * @page: The page within the block of interest
|
| 412 | * @pfn: The target page frame number
|
| 413 | * @pb_bit: pageblock bit to check
|
| 414 | *
|
| 415 | * Return: true if the bit is set, otherwise false
|
| 416 |
|
| 417 | bool get_pfnblock_bit(const struct page *page, unsigned long pfn,
|
| 418 | enum pageblock_bits pb_bit)
|
| 419 | {
|
| 420 | unsigned long *bitmap_word;
|
| 421 | unsigned long bitidx;
|
| 422 |
|
| 423 | if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
|
| 424 | return false;
|
| 425 |
|
| 426 | get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
|
| 427 |
|
| 428 | return test_bit(bitidx + pb_bit, bitmap_word);
|
| 429 | }
|
| 430 |
|
| 431 | |
| 432 | * get_pfnblock_migratetype - Return the migratetype of a pageblock
|
| 433 | * @page: The page within the block of interest
|
| 434 | * @pfn: The target page frame number
|
| 435 | *
|
| 436 | * Return: The migratetype of the pageblock
|
| 437 | *
|
| 438 | * Use get_pfnblock_migratetype() if caller already has both @page and @pfn
|
| 439 | * to save a call to page_to_pfn().
|
| 440 |
|
| 441 | enum migratetype
|
| 442 | get_pfnblock_migratetype(const struct page *page, unsigned long pfn)
|
| 443 | {
|
| 444 | unsigned long mask = PAGEBLOCK_MIGRATETYPE_MASK | PAGEBLOCK_ISO_MASK;
|
| 445 | unsigned long flags;
|
| 446 |
|
| 447 | flags = __get_pfnblock_flags_mask(page, pfn, mask);
|
| 448 |
|
| 449 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 450 | if (flags & BIT(PB_migrate_isolate))
|
| 451 | return MIGRATE_ISOLATE;
|
| 452 | #endif
|
| 453 | return flags & PAGEBLOCK_MIGRATETYPE_MASK;
|
| 454 | }
|
| 455 |
|
| 456 | |
| 457 | * __set_pfnblock_flags_mask - Set the requested group of flags for
|
| 458 | * a pageblock_nr_pages block of pages
|
| 459 | * @page: The page within the block of interest
|
| 460 | * @pfn: The target page frame number
|
| 461 | * @flags: The flags to set
|
| 462 | * @mask: mask of bits that the caller is interested in
|
| 463 |
|
| 464 | static void __set_pfnblock_flags_mask(struct page *page, unsigned long pfn,
|
| 465 | unsigned long flags, unsigned long mask)
|
| 466 | {
|
| 467 | unsigned long *bitmap_word;
|
| 468 | unsigned long bitidx;
|
| 469 | unsigned long word;
|
| 470 |
|
| 471 | get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
|
| 472 |
|
| 473 | mask <<= bitidx;
|
| 474 | flags <<= bitidx;
|
| 475 |
|
| 476 | word = READ_ONCE(*bitmap_word);
|
| 477 | do {
|
| 478 | } while (!try_cmpxchg(bitmap_word, &word, (word & ~mask) | flags));
|
| 479 | }
|
| 480 |
|
| 481 | |
| 482 | * set_pfnblock_bit - Set a standalone bit of a pageblock
|
| 483 | * @page: The page within the block of interest
|
| 484 | * @pfn: The target page frame number
|
| 485 | * @pb_bit: pageblock bit to set
|
| 486 |
|
| 487 | void set_pfnblock_bit(const struct page *page, unsigned long pfn,
|
| 488 | enum pageblock_bits pb_bit)
|
| 489 | {
|
| 490 | unsigned long *bitmap_word;
|
| 491 | unsigned long bitidx;
|
| 492 |
|
| 493 | if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
|
| 494 | return;
|
| 495 |
|
| 496 | get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
|
| 497 |
|
| 498 | set_bit(bitidx + pb_bit, bitmap_word);
|
| 499 | }
|
| 500 |
|
| 501 | |
| 502 | * clear_pfnblock_bit - Clear a standalone bit of a pageblock
|
| 503 | * @page: The page within the block of interest
|
| 504 | * @pfn: The target page frame number
|
| 505 | * @pb_bit: pageblock bit to clear
|
| 506 |
|
| 507 | void clear_pfnblock_bit(const struct page *page, unsigned long pfn,
|
| 508 | enum pageblock_bits pb_bit)
|
| 509 | {
|
| 510 | unsigned long *bitmap_word;
|
| 511 | unsigned long bitidx;
|
| 512 |
|
| 513 | if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
|
| 514 | return;
|
| 515 |
|
| 516 | get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
|
| 517 |
|
| 518 | clear_bit(bitidx + pb_bit, bitmap_word);
|
| 519 | }
|
| 520 |
|
| 521 | |
| 522 | * set_pageblock_migratetype - Set the migratetype of a pageblock
|
| 523 | * @page: The page within the block of interest
|
| 524 | * @migratetype: migratetype to set
|
| 525 |
|
| 526 | static void set_pageblock_migratetype(struct page *page,
|
| 527 | enum migratetype migratetype)
|
| 528 | {
|
| 529 | if (unlikely(page_group_by_mobility_disabled &&
|
| 530 | migratetype < MIGRATE_PCPTYPES))
|
| 531 | migratetype = MIGRATE_UNMOVABLE;
|
| 532 |
|
| 533 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 534 | if (migratetype == MIGRATE_ISOLATE) {
|
| 535 | VM_WARN_ONCE(1,
|
| 536 | "Use set_pageblock_isolate() for pageblock isolation");
|
| 537 | return;
|
| 538 | }
|
| 539 | VM_WARN_ONCE(get_pageblock_isolate(page),
|
| 540 | "Use clear_pageblock_isolate() to unisolate pageblock");
|
| 541 |
|
| 542 | #endif
|
| 543 | __set_pfnblock_flags_mask(page, page_to_pfn(page),
|
| 544 | (unsigned long)migratetype,
|
| 545 | PAGEBLOCK_MIGRATETYPE_MASK | PAGEBLOCK_ISO_MASK);
|
| 546 | }
|
| 547 |
|
| 548 | void __meminit init_pageblock_migratetype(struct page *page,
|
| 549 | enum migratetype migratetype,
|
| 550 | bool isolate)
|
| 551 | {
|
| 552 | unsigned long flags;
|
| 553 |
|
| 554 | if (unlikely(page_group_by_mobility_disabled &&
|
| 555 | migratetype < MIGRATE_PCPTYPES))
|
| 556 | migratetype = MIGRATE_UNMOVABLE;
|
| 557 |
|
| 558 | flags = migratetype;
|
| 559 |
|
| 560 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 561 | if (migratetype == MIGRATE_ISOLATE) {
|
| 562 | VM_WARN_ONCE(
|
| 563 | 1,
|
| 564 | "Set isolate=true to isolate pageblock with a migratetype");
|
| 565 | return;
|
| 566 | }
|
| 567 | if (isolate)
|
| 568 | flags |= BIT(PB_migrate_isolate);
|
| 569 | #endif
|
| 570 | __set_pfnblock_flags_mask(page, page_to_pfn(page), flags,
|
| 571 | PAGEBLOCK_MIGRATETYPE_MASK | PAGEBLOCK_ISO_MASK);
|
| 572 | }
|
| 573 |
|
| 574 | #ifdef CONFIG_DEBUG_VM
|
| 575 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
|
| 576 | {
|
| 577 | int ret;
|
| 578 | unsigned seq;
|
| 579 | unsigned long pfn = page_to_pfn(page);
|
| 580 | unsigned long sp, start_pfn;
|
| 581 |
|
| 582 | do {
|
| 583 | seq = zone_span_seqbegin(zone);
|
| 584 | start_pfn = zone->zone_start_pfn;
|
| 585 | sp = zone->spanned_pages;
|
| 586 | ret = !zone_spans_pfn(zone, pfn);
|
| 587 | } while (zone_span_seqretry(zone, seq));
|
| 588 |
|
| 589 | if (ret)
|
| 590 | pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
|
| 591 | pfn, zone_to_nid(zone), zone->name,
|
| 592 | start_pfn, start_pfn + sp);
|
| 593 |
|
| 594 | return ret;
|
| 595 | }
|
| 596 |
|
| 597 | |
| 598 | * Temporary debugging check for pages not lying within a given zone.
|
| 599 |
|
| 600 | static bool __maybe_unused bad_range(struct zone *zone, struct page *page)
|
| 601 | {
|
| 602 | if (page_outside_zone_boundaries(zone, page))
|
| 603 | return true;
|
| 604 | if (zone != page_zone(page))
|
| 605 | return true;
|
| 606 |
|
| 607 | return false;
|
| 608 | }
|
| 609 | #else
|
| 610 | static inline bool __maybe_unused bad_range(struct zone *zone, struct page *page)
|
| 611 | {
|
| 612 | return false;
|
| 613 | }
|
| 614 | #endif
|
| 615 |
|
| 616 | static void bad_page(struct page *page, const char *reason)
|
| 617 | {
|
| 618 | static unsigned long resume;
|
| 619 | static unsigned long nr_shown;
|
| 620 | static unsigned long nr_unshown;
|
| 621 |
|
| 622 | |
| 623 | * Allow a burst of 60 reports, then keep quiet for that minute;
|
| 624 | * or allow a steady drip of one report per second.
|
| 625 |
|
| 626 | if (nr_shown == 60) {
|
| 627 | if (time_before(jiffies, resume)) {
|
| 628 | nr_unshown++;
|
| 629 | goto out;
|
| 630 | }
|
| 631 | if (nr_unshown) {
|
| 632 | pr_alert(
|
| 633 | "BUG: Bad page state: %lu messages suppressed\n",
|
| 634 | nr_unshown);
|
| 635 | nr_unshown = 0;
|
| 636 | }
|
| 637 | nr_shown = 0;
|
| 638 | }
|
| 639 | if (nr_shown++ == 0)
|
| 640 | resume = jiffies + 60 * HZ;
|
| 641 |
|
| 642 | pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
|
| 643 | current->comm, page_to_pfn(page));
|
| 644 | dump_page(page, reason);
|
| 645 |
|
| 646 | print_modules();
|
| 647 | dump_stack();
|
| 648 | out:
|
| 649 |
|
| 650 | if (PageBuddy(page))
|
| 651 | __ClearPageBuddy(page);
|
| 652 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
|
| 653 | }
|
| 654 |
|
| 655 | static inline unsigned int order_to_pindex(int migratetype, int order)
|
| 656 | {
|
| 657 | if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
|
| 658 | bool movable = migratetype == MIGRATE_MOVABLE;
|
| 659 |
|
| 660 | if (order > PAGE_ALLOC_COSTLY_ORDER)
|
| 661 | return NR_LOWORDER_PCP_LISTS + movable;
|
| 662 | }
|
| 663 |
|
| 664 | return (MIGRATE_PCPTYPES * order) + migratetype;
|
| 665 | }
|
| 666 |
|
| 667 | static inline int pindex_to_order(unsigned int pindex)
|
| 668 | {
|
| 669 | int order = pindex / MIGRATE_PCPTYPES;
|
| 670 |
|
| 671 | if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
|
| 672 | if (pindex >= NR_LOWORDER_PCP_LISTS)
|
| 673 | order = HPAGE_PMD_ORDER;
|
| 674 | }
|
| 675 |
|
| 676 | return order;
|
| 677 | }
|
| 678 |
|
| 679 | static inline bool pcp_allowed_order(unsigned int order)
|
| 680 | {
|
| 681 | if (order <= PAGE_ALLOC_COSTLY_ORDER)
|
| 682 | return true;
|
| 683 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE
|
| 684 | if (is_pmd_order(order))
|
| 685 | return true;
|
| 686 | #endif
|
| 687 | return false;
|
| 688 | }
|
| 689 |
|
| 690 | |
| 691 | * Higher-order pages are called "compound pages". They are structured thusly:
|
| 692 | *
|
| 693 | * The first PAGE_SIZE page is called the "head page" and have PG_head set.
|
| 694 | *
|
| 695 | * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
|
| 696 | * in bit 0 of page->compound_info. The rest of bits is pointer to head page.
|
| 697 | *
|
| 698 | * The first tail page's ->compound_order holds the order of allocation.
|
| 699 | * This usage means that zero-order pages may not be compound.
|
| 700 |
|
| 701 |
|
| 702 | void prep_compound_page(struct page *page, unsigned int order)
|
| 703 | {
|
| 704 | int i;
|
| 705 | int nr_pages = 1 << order;
|
| 706 |
|
| 707 | __SetPageHead(page);
|
| 708 | for (i = 1; i < nr_pages; i++)
|
| 709 | prep_compound_tail(page + i, page, order);
|
| 710 |
|
| 711 | prep_compound_head(page, order);
|
| 712 | }
|
| 713 |
|
| 714 | static inline void set_buddy_order(struct page *page, unsigned int order)
|
| 715 | {
|
| 716 | set_page_private(page, order);
|
| 717 | __SetPageBuddy(page);
|
| 718 | }
|
| 719 |
|
| 720 | #ifdef CONFIG_COMPACTION
|
| 721 | static inline struct capture_control *task_capc(struct zone *zone)
|
| 722 | {
|
| 723 | struct capture_control *capc = current->capture_control;
|
| 724 |
|
| 725 | return unlikely(capc) &&
|
| 726 | !(current->flags & PF_KTHREAD) &&
|
| 727 | !capc->page &&
|
| 728 | capc->zone == zone ? capc : NULL;
|
| 729 | }
|
| 730 |
|
| 731 | static inline bool
|
| 732 | compaction_capture(struct capture_control *capc, struct page *page,
|
| 733 | int order, int migratetype)
|
| 734 | {
|
| 735 | if (!capc || order != capc->order)
|
| 736 | return false;
|
| 737 |
|
| 738 |
|
| 739 | if (is_migrate_cma(migratetype) ||
|
| 740 | is_migrate_isolate(migratetype))
|
| 741 | return false;
|
| 742 |
|
| 743 | |
| 744 | * Do not let lower order allocations pollute a movable pageblock
|
| 745 | * unless compaction is also requesting movable pages.
|
| 746 | * This might let an unmovable request use a reclaimable pageblock
|
| 747 | * and vice-versa but no more than normal fallback logic which can
|
| 748 | * have trouble finding a high-order free page.
|
| 749 |
|
| 750 | if (order < pageblock_order && migratetype == MIGRATE_MOVABLE &&
|
| 751 | capc->migratetype != MIGRATE_MOVABLE)
|
| 752 | return false;
|
| 753 |
|
| 754 | if (migratetype != capc->migratetype)
|
| 755 | trace_mm_page_alloc_extfrag(page, capc->order, order,
|
| 756 | capc->migratetype, migratetype);
|
| 757 |
|
| 758 | capc->page = page;
|
| 759 | return true;
|
| 760 | }
|
| 761 |
|
| 762 | #else
|
| 763 | static inline struct capture_control *task_capc(struct zone *zone)
|
| 764 | {
|
| 765 | return NULL;
|
| 766 | }
|
| 767 |
|
| 768 | static inline bool
|
| 769 | compaction_capture(struct capture_control *capc, struct page *page,
|
| 770 | int order, int migratetype)
|
| 771 | {
|
| 772 | return false;
|
| 773 | }
|
| 774 | #endif
|
| 775 |
|
| 776 | static inline void account_freepages(struct zone *zone, int nr_pages,
|
| 777 | int migratetype)
|
| 778 | {
|
| 779 | lockdep_assert_held(&zone->lock);
|
| 780 |
|
| 781 | if (is_migrate_isolate(migratetype))
|
| 782 | return;
|
| 783 |
|
| 784 | __mod_zone_page_state(zone, NR_FREE_PAGES, nr_pages);
|
| 785 |
|
| 786 | if (is_migrate_cma(migratetype))
|
| 787 | __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, nr_pages);
|
| 788 | else if (migratetype == MIGRATE_HIGHATOMIC)
|
| 789 | WRITE_ONCE(zone->nr_free_highatomic,
|
| 790 | zone->nr_free_highatomic + nr_pages);
|
| 791 | }
|
| 792 |
|
| 793 |
|
| 794 | static inline void __add_to_free_list(struct page *page, struct zone *zone,
|
| 795 | unsigned int order, int migratetype,
|
| 796 | bool tail)
|
| 797 | {
|
| 798 | struct free_area *area = &zone->free_area[order];
|
| 799 | int nr_pages = 1 << order;
|
| 800 |
|
| 801 | VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
|
| 802 | "page type is %d, passed migratetype is %d (nr=%d)\n",
|
| 803 | get_pageblock_migratetype(page), migratetype, nr_pages);
|
| 804 |
|
| 805 | if (tail)
|
| 806 | list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
|
| 807 | else
|
| 808 | list_add(&page->buddy_list, &area->free_list[migratetype]);
|
| 809 | area->nr_free++;
|
| 810 |
|
| 811 | if (order >= pageblock_order && !is_migrate_isolate(migratetype))
|
| 812 | __mod_zone_page_state(zone, NR_FREE_PAGES_BLOCKS, nr_pages);
|
| 813 | }
|
| 814 |
|
| 815 | |
| 816 | * Used for pages which are on another list. Move the pages to the tail
|
| 817 | * of the list - so the moved pages won't immediately be considered for
|
| 818 | * allocation again (e.g., optimization for memory onlining).
|
| 819 |
|
| 820 | static inline void move_to_free_list(struct page *page, struct zone *zone,
|
| 821 | unsigned int order, int old_mt, int new_mt)
|
| 822 | {
|
| 823 | struct free_area *area = &zone->free_area[order];
|
| 824 | int nr_pages = 1 << order;
|
| 825 |
|
| 826 |
|
| 827 | VM_WARN_ONCE(get_pageblock_migratetype(page) != old_mt,
|
| 828 | "page type is %d, passed migratetype is %d (nr=%d)\n",
|
| 829 | get_pageblock_migratetype(page), old_mt, nr_pages);
|
| 830 |
|
| 831 | list_move_tail(&page->buddy_list, &area->free_list[new_mt]);
|
| 832 |
|
| 833 | account_freepages(zone, -nr_pages, old_mt);
|
| 834 | account_freepages(zone, nr_pages, new_mt);
|
| 835 |
|
| 836 | if (order >= pageblock_order &&
|
| 837 | is_migrate_isolate(old_mt) != is_migrate_isolate(new_mt)) {
|
| 838 | if (!is_migrate_isolate(old_mt))
|
| 839 | nr_pages = -nr_pages;
|
| 840 | __mod_zone_page_state(zone, NR_FREE_PAGES_BLOCKS, nr_pages);
|
| 841 | }
|
| 842 | }
|
| 843 |
|
| 844 | static inline void __del_page_from_free_list(struct page *page, struct zone *zone,
|
| 845 | unsigned int order, int migratetype)
|
| 846 | {
|
| 847 | int nr_pages = 1 << order;
|
| 848 |
|
| 849 | VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
|
| 850 | "page type is %d, passed migratetype is %d (nr=%d)\n",
|
| 851 | get_pageblock_migratetype(page), migratetype, nr_pages);
|
| 852 |
|
| 853 |
|
| 854 | if (page_reported(page))
|
| 855 | __ClearPageReported(page);
|
| 856 |
|
| 857 | list_del(&page->buddy_list);
|
| 858 | __ClearPageBuddy(page);
|
| 859 | set_page_private(page, 0);
|
| 860 | zone->free_area[order].nr_free--;
|
| 861 |
|
| 862 | if (order >= pageblock_order && !is_migrate_isolate(migratetype))
|
| 863 | __mod_zone_page_state(zone, NR_FREE_PAGES_BLOCKS, -nr_pages);
|
| 864 | }
|
| 865 |
|
| 866 | static inline void del_page_from_free_list(struct page *page, struct zone *zone,
|
| 867 | unsigned int order, int migratetype)
|
| 868 | {
|
| 869 | __del_page_from_free_list(page, zone, order, migratetype);
|
| 870 | account_freepages(zone, -(1 << order), migratetype);
|
| 871 | }
|
| 872 |
|
| 873 | static inline struct page *get_page_from_free_area(struct free_area *area,
|
| 874 | int migratetype)
|
| 875 | {
|
| 876 | return list_first_entry_or_null(&area->free_list[migratetype],
|
| 877 | struct page, buddy_list);
|
| 878 | }
|
| 879 |
|
| 880 | |
| 881 | * If this is less than the 2nd largest possible page, check if the buddy
|
| 882 | * of the next-higher order is free. If it is, it's possible
|
| 883 | * that pages are being freed that will coalesce soon. In case,
|
| 884 | * that is happening, add the free page to the tail of the list
|
| 885 | * so it's less likely to be used soon and more likely to be merged
|
| 886 | * as a 2-level higher order page
|
| 887 |
|
| 888 | static inline bool
|
| 889 | buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
|
| 890 | struct page *page, unsigned int order)
|
| 891 | {
|
| 892 | unsigned long higher_page_pfn;
|
| 893 | struct page *higher_page;
|
| 894 |
|
| 895 | if (order >= MAX_PAGE_ORDER - 1)
|
| 896 | return false;
|
| 897 |
|
| 898 | higher_page_pfn = buddy_pfn & pfn;
|
| 899 | higher_page = page + (higher_page_pfn - pfn);
|
| 900 |
|
| 901 | return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
|
| 902 | NULL) != NULL;
|
| 903 | }
|
| 904 |
|
| 905 | static void change_pageblock_range(struct page *pageblock_page,
|
| 906 | int start_order, int migratetype)
|
| 907 | {
|
| 908 | int nr_pageblocks = 1 << (start_order - pageblock_order);
|
| 909 |
|
| 910 | while (nr_pageblocks--) {
|
| 911 | set_pageblock_migratetype(pageblock_page, migratetype);
|
| 912 | pageblock_page += pageblock_nr_pages;
|
| 913 | }
|
| 914 | }
|
| 915 |
|
| 916 | |
| 917 | * Freeing function for a buddy system allocator.
|
| 918 | *
|
| 919 | * The concept of a buddy system is to maintain direct-mapped table
|
| 920 | * (containing bit values) for memory blocks of various "orders".
|
| 921 | * The bottom level table contains the map for the smallest allocatable
|
| 922 | * units of memory (here, pages), and each level above it describes
|
| 923 | * pairs of units from the levels below, hence, "buddies".
|
| 924 | * At a high level, all that happens here is marking the table entry
|
| 925 | * at the bottom level available, and propagating the changes upward
|
| 926 | * as necessary, plus some accounting needed to play nicely with other
|
| 927 | * parts of the VM system.
|
| 928 | * At each level, we keep a list of pages, which are heads of continuous
|
| 929 | * free pages of length of (1 << order) and marked with PageBuddy.
|
| 930 | * Page's order is recorded in page_private(page) field.
|
| 931 | * So when we are allocating or freeing one, we can derive the state of the
|
| 932 | * other. That is, if we allocate a small block, and both were
|
| 933 | * free, the remainder of the region must be split into blocks.
|
| 934 | * If a block is freed, and its buddy is also free, then this
|
| 935 | * triggers coalescing into a block of larger size.
|
| 936 | *
|
| 937 | * -- nyc
|
| 938 |
|
| 939 |
|
| 940 | static inline void __free_one_page(struct page *page,
|
| 941 | unsigned long pfn,
|
| 942 | struct zone *zone, unsigned int order,
|
| 943 | int migratetype, fpi_t fpi_flags)
|
| 944 | {
|
| 945 | struct capture_control *capc = task_capc(zone);
|
| 946 | unsigned long buddy_pfn = 0;
|
| 947 | unsigned long combined_pfn;
|
| 948 | struct page *buddy;
|
| 949 | bool to_tail;
|
| 950 |
|
| 951 | VM_BUG_ON(!zone_is_initialized(zone));
|
| 952 | VM_BUG_ON_PAGE(page->flags.f & PAGE_FLAGS_CHECK_AT_PREP, page);
|
| 953 |
|
| 954 | VM_BUG_ON(migratetype == -1);
|
| 955 | VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
|
| 956 | VM_BUG_ON_PAGE(bad_range(zone, page), page);
|
| 957 |
|
| 958 | account_freepages(zone, 1 << order, migratetype);
|
| 959 |
|
| 960 | while (order < MAX_PAGE_ORDER) {
|
| 961 | int buddy_mt = migratetype;
|
| 962 |
|
| 963 | if (compaction_capture(capc, page, order, migratetype)) {
|
| 964 | account_freepages(zone, -(1 << order), migratetype);
|
| 965 | return;
|
| 966 | }
|
| 967 |
|
| 968 | buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
|
| 969 | if (!buddy)
|
| 970 | goto done_merging;
|
| 971 |
|
| 972 | if (unlikely(order >= pageblock_order)) {
|
| 973 | |
| 974 | * We want to prevent merge between freepages on pageblock
|
| 975 | * without fallbacks and normal pageblock. Without this,
|
| 976 | * pageblock isolation could cause incorrect freepage or CMA
|
| 977 | * accounting or HIGHATOMIC accounting.
|
| 978 |
|
| 979 | buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
|
| 980 |
|
| 981 | if (migratetype != buddy_mt &&
|
| 982 | (!migratetype_is_mergeable(migratetype) ||
|
| 983 | !migratetype_is_mergeable(buddy_mt)))
|
| 984 | goto done_merging;
|
| 985 | }
|
| 986 |
|
| 987 | |
| 988 | * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
|
| 989 | * merge with it and move up one order.
|
| 990 |
|
| 991 | if (page_is_guard(buddy))
|
| 992 | clear_page_guard(zone, buddy, order);
|
| 993 | else
|
| 994 | __del_page_from_free_list(buddy, zone, order, buddy_mt);
|
| 995 |
|
| 996 | if (unlikely(buddy_mt != migratetype)) {
|
| 997 | |
| 998 | * Match buddy type. This ensures that an
|
| 999 | * expand() down the line puts the sub-blocks
|
| 1000 | * on the right freelists.
|
| 1001 |
|
| 1002 | change_pageblock_range(buddy, order, migratetype);
|
| 1003 | }
|
| 1004 |
|
| 1005 | combined_pfn = buddy_pfn & pfn;
|
| 1006 | page = page + (combined_pfn - pfn);
|
| 1007 | pfn = combined_pfn;
|
| 1008 | order++;
|
| 1009 | }
|
| 1010 |
|
| 1011 | done_merging:
|
| 1012 | set_buddy_order(page, order);
|
| 1013 |
|
| 1014 | if (fpi_flags & FPI_TO_TAIL)
|
| 1015 | to_tail = true;
|
| 1016 | else if (is_shuffle_order(order))
|
| 1017 | to_tail = shuffle_pick_tail();
|
| 1018 | else
|
| 1019 | to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
|
| 1020 |
|
| 1021 | __add_to_free_list(page, zone, order, migratetype, to_tail);
|
| 1022 |
|
| 1023 |
|
| 1024 | if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
|
| 1025 | page_reporting_notify_free(order);
|
| 1026 | }
|
| 1027 |
|
| 1028 | |
| 1029 | * A bad page could be due to a number of fields. Instead of multiple branches,
|
| 1030 | * try and check multiple fields with one check. The caller must do a detailed
|
| 1031 | * check if necessary.
|
| 1032 |
|
| 1033 | static inline bool page_expected_state(struct page *page,
|
| 1034 | unsigned long check_flags)
|
| 1035 | {
|
| 1036 | if (unlikely(atomic_read(&page->_mapcount) != -1))
|
| 1037 | return false;
|
| 1038 |
|
| 1039 | if (unlikely((unsigned long)page->mapping |
|
| 1040 | page_ref_count(page) |
|
| 1041 | #ifdef CONFIG_MEMCG
|
| 1042 | page->memcg_data |
|
| 1043 | #endif
|
| 1044 | page_pool_page_is_pp(page) |
|
| 1045 | (page->flags.f & check_flags)))
|
| 1046 | return false;
|
| 1047 |
|
| 1048 | return true;
|
| 1049 | }
|
| 1050 |
|
| 1051 | static const char *page_bad_reason(struct page *page, unsigned long flags)
|
| 1052 | {
|
| 1053 | const char *bad_reason = NULL;
|
| 1054 |
|
| 1055 | if (unlikely(atomic_read(&page->_mapcount) != -1))
|
| 1056 | bad_reason = "nonzero mapcount";
|
| 1057 | if (unlikely(page->mapping != NULL))
|
| 1058 | bad_reason = "non-NULL mapping";
|
| 1059 | if (unlikely(page_ref_count(page) != 0))
|
| 1060 | bad_reason = "nonzero _refcount";
|
| 1061 | if (unlikely(page->flags.f & flags)) {
|
| 1062 | if (flags == PAGE_FLAGS_CHECK_AT_PREP)
|
| 1063 | bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
|
| 1064 | else
|
| 1065 | bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
|
| 1066 | }
|
| 1067 | #ifdef CONFIG_MEMCG
|
| 1068 | if (unlikely(page->memcg_data))
|
| 1069 | bad_reason = "page still charged to cgroup";
|
| 1070 | #endif
|
| 1071 | if (unlikely(page_pool_page_is_pp(page)))
|
| 1072 | bad_reason = "page_pool leak";
|
| 1073 | return bad_reason;
|
| 1074 | }
|
| 1075 |
|
| 1076 | static inline bool free_page_is_bad(struct page *page)
|
| 1077 | {
|
| 1078 | if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
|
| 1079 | return false;
|
| 1080 |
|
| 1081 |
|
| 1082 | bad_page(page, page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
|
| 1083 | return true;
|
| 1084 | }
|
| 1085 |
|
| 1086 | static inline bool is_check_pages_enabled(void)
|
| 1087 | {
|
| 1088 | return static_branch_unlikely(&check_pages_enabled);
|
| 1089 | }
|
| 1090 |
|
| 1091 | static int free_tail_page_prepare(struct page *head_page, struct page *page)
|
| 1092 | {
|
| 1093 | struct folio *folio = (struct folio *)head_page;
|
| 1094 | int ret = 1;
|
| 1095 |
|
| 1096 | |
| 1097 | * We rely page->lru.next never has bit 0 set, unless the page
|
| 1098 | * is PageTail(). Let's make sure that's true even for poisoned ->lru.
|
| 1099 |
|
| 1100 | BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
|
| 1101 |
|
| 1102 | if (!is_check_pages_enabled()) {
|
| 1103 | ret = 0;
|
| 1104 | goto out;
|
| 1105 | }
|
| 1106 | switch (page - head_page) {
|
| 1107 | case 1:
|
| 1108 |
|
| 1109 | if (unlikely(folio_large_mapcount(folio))) {
|
| 1110 | bad_page(page, "nonzero large_mapcount");
|
| 1111 | goto out;
|
| 1112 | }
|
| 1113 | if (IS_ENABLED(CONFIG_PAGE_MAPCOUNT) &&
|
| 1114 | unlikely(atomic_read(&folio->_nr_pages_mapped))) {
|
| 1115 | bad_page(page, "nonzero nr_pages_mapped");
|
| 1116 | goto out;
|
| 1117 | }
|
| 1118 | if (IS_ENABLED(CONFIG_MM_ID)) {
|
| 1119 | if (unlikely(folio->_mm_id_mapcount[0] != -1)) {
|
| 1120 | bad_page(page, "nonzero mm mapcount 0");
|
| 1121 | goto out;
|
| 1122 | }
|
| 1123 | if (unlikely(folio->_mm_id_mapcount[1] != -1)) {
|
| 1124 | bad_page(page, "nonzero mm mapcount 1");
|
| 1125 | goto out;
|
| 1126 | }
|
| 1127 | }
|
| 1128 | if (IS_ENABLED(CONFIG_64BIT)) {
|
| 1129 | if (unlikely(atomic_read(&folio->_entire_mapcount) + 1)) {
|
| 1130 | bad_page(page, "nonzero entire_mapcount");
|
| 1131 | goto out;
|
| 1132 | }
|
| 1133 | if (unlikely(atomic_read(&folio->_pincount))) {
|
| 1134 | bad_page(page, "nonzero pincount");
|
| 1135 | goto out;
|
| 1136 | }
|
| 1137 | }
|
| 1138 | break;
|
| 1139 | case 2:
|
| 1140 |
|
| 1141 | if (unlikely(!list_empty(&folio->_deferred_list))) {
|
| 1142 | bad_page(page, "on deferred list");
|
| 1143 | goto out;
|
| 1144 | }
|
| 1145 | if (!IS_ENABLED(CONFIG_64BIT)) {
|
| 1146 | if (unlikely(atomic_read(&folio->_entire_mapcount) + 1)) {
|
| 1147 | bad_page(page, "nonzero entire_mapcount");
|
| 1148 | goto out;
|
| 1149 | }
|
| 1150 | if (unlikely(atomic_read(&folio->_pincount))) {
|
| 1151 | bad_page(page, "nonzero pincount");
|
| 1152 | goto out;
|
| 1153 | }
|
| 1154 | }
|
| 1155 | break;
|
| 1156 | case 3:
|
| 1157 |
|
| 1158 | if (IS_ENABLED(CONFIG_HUGETLB_PAGE))
|
| 1159 | break;
|
| 1160 | fallthrough;
|
| 1161 | default:
|
| 1162 | if (page->mapping != TAIL_MAPPING) {
|
| 1163 | bad_page(page, "corrupted mapping in tail page");
|
| 1164 | goto out;
|
| 1165 | }
|
| 1166 | break;
|
| 1167 | }
|
| 1168 | if (unlikely(!PageTail(page))) {
|
| 1169 | bad_page(page, "PageTail not set");
|
| 1170 | goto out;
|
| 1171 | }
|
| 1172 | if (unlikely(compound_head(page) != head_page)) {
|
| 1173 | bad_page(page, "compound_head not consistent");
|
| 1174 | goto out;
|
| 1175 | }
|
| 1176 | ret = 0;
|
| 1177 | out:
|
| 1178 | page->mapping = NULL;
|
| 1179 | clear_compound_head(page);
|
| 1180 | return ret;
|
| 1181 | }
|
| 1182 |
|
| 1183 | |
| 1184 | * Skip KASAN memory poisoning when either:
|
| 1185 | *
|
| 1186 | * 1. For generic KASAN: deferred memory initialization has not yet completed.
|
| 1187 | * Tag-based KASAN modes skip pages freed via deferred memory initialization
|
| 1188 | * using page tags instead (see below).
|
| 1189 | * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
|
| 1190 | * that error detection is disabled for accesses via the page address.
|
| 1191 | *
|
| 1192 | * Pages will have match-all tags in the following circumstances:
|
| 1193 | *
|
| 1194 | * 1. Pages are being initialized for the first time, including during deferred
|
| 1195 | * memory init; see the call to page_kasan_tag_reset in __init_single_page.
|
| 1196 | * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
|
| 1197 | * exception of pages unpoisoned by kasan_unpoison_vmalloc.
|
| 1198 | * 3. The allocation was excluded from being checked due to sampling,
|
| 1199 | * see the call to kasan_unpoison_pages.
|
| 1200 | *
|
| 1201 | * Poisoning pages during deferred memory init will greatly lengthen the
|
| 1202 | * process and cause problem in large memory systems as the deferred pages
|
| 1203 | * initialization is done with interrupt disabled.
|
| 1204 | *
|
| 1205 | * Assuming that there will be no reference to those newly initialized
|
| 1206 | * pages before they are ever allocated, this should have no effect on
|
| 1207 | * KASAN memory tracking as the poison will be properly inserted at page
|
| 1208 | * allocation time. The only corner case is when pages are allocated by
|
| 1209 | * on-demand allocation and then freed again before the deferred pages
|
| 1210 | * initialization is done, but this is not likely to happen.
|
| 1211 |
|
| 1212 | static inline bool should_skip_kasan_poison(struct page *page)
|
| 1213 | {
|
| 1214 | if (IS_ENABLED(CONFIG_KASAN_GENERIC))
|
| 1215 | return deferred_pages_enabled();
|
| 1216 |
|
| 1217 | return page_kasan_tag(page) == KASAN_TAG_KERNEL;
|
| 1218 | }
|
| 1219 |
|
| 1220 | static void clear_highpages_kasan_tagged(struct page *page, int numpages)
|
| 1221 | {
|
| 1222 |
|
| 1223 | kasan_disable_current();
|
| 1224 | if (!IS_ENABLED(CONFIG_HIGHMEM)) {
|
| 1225 | clear_pages(kasan_reset_tag(page_address(page)), numpages);
|
| 1226 | } else {
|
| 1227 | int i;
|
| 1228 |
|
| 1229 | for (i = 0; i < numpages; i++)
|
| 1230 | clear_highpage_kasan_tagged(page + i);
|
| 1231 | }
|
| 1232 | kasan_enable_current();
|
| 1233 | }
|
| 1234 |
|
| 1235 | #ifdef CONFIG_MEM_ALLOC_PROFILING
|
| 1236 |
|
| 1237 |
|
| 1238 | void __clear_page_tag_ref(struct page *page)
|
| 1239 | {
|
| 1240 | union pgtag_ref_handle handle;
|
| 1241 | union codetag_ref ref;
|
| 1242 |
|
| 1243 | if (get_page_tag_ref(page, &ref, &handle)) {
|
| 1244 | set_codetag_empty(&ref);
|
| 1245 | update_page_tag_ref(handle, &ref);
|
| 1246 | put_page_tag_ref(handle);
|
| 1247 | }
|
| 1248 | }
|
| 1249 |
|
| 1250 |
|
| 1251 | static noinline
|
| 1252 | void __pgalloc_tag_add(struct page *page, struct task_struct *task,
|
| 1253 | unsigned int nr, unsigned int alloc_flags)
|
| 1254 | {
|
| 1255 | union pgtag_ref_handle handle;
|
| 1256 | union codetag_ref ref;
|
| 1257 |
|
| 1258 | if (likely(get_page_tag_ref(page, &ref, &handle))) {
|
| 1259 | alloc_tag_add(&ref, task->alloc_tag, PAGE_SIZE * nr);
|
| 1260 | update_page_tag_ref(handle, &ref);
|
| 1261 | put_page_tag_ref(handle);
|
| 1262 | } else {
|
| 1263 | |
| 1264 | * page_ext is not available yet, record the pfn so we can
|
| 1265 | * clear the tag ref later when page_ext is initialized.
|
| 1266 |
|
| 1267 | alloc_tag_add_early_pfn(page_to_pfn(page), alloc_flags);
|
| 1268 | if (task->alloc_tag)
|
| 1269 | alloc_tag_set_inaccurate(task->alloc_tag);
|
| 1270 | }
|
| 1271 | }
|
| 1272 |
|
| 1273 | static inline void pgalloc_tag_add(struct page *page, struct task_struct *task,
|
| 1274 | unsigned int nr, unsigned int alloc_flags)
|
| 1275 | {
|
| 1276 | if (mem_alloc_profiling_enabled())
|
| 1277 | __pgalloc_tag_add(page, task, nr, alloc_flags);
|
| 1278 | }
|
| 1279 |
|
| 1280 |
|
| 1281 | static noinline
|
| 1282 | void __pgalloc_tag_sub(struct page *page, unsigned int nr)
|
| 1283 | {
|
| 1284 | union pgtag_ref_handle handle;
|
| 1285 | union codetag_ref ref;
|
| 1286 |
|
| 1287 | if (get_page_tag_ref(page, &ref, &handle)) {
|
| 1288 | alloc_tag_sub(&ref, PAGE_SIZE * nr);
|
| 1289 | update_page_tag_ref(handle, &ref);
|
| 1290 | put_page_tag_ref(handle);
|
| 1291 | }
|
| 1292 | }
|
| 1293 |
|
| 1294 | static inline void pgalloc_tag_sub(struct page *page, unsigned int nr)
|
| 1295 | {
|
| 1296 | if (mem_alloc_profiling_enabled())
|
| 1297 | __pgalloc_tag_sub(page, nr);
|
| 1298 | }
|
| 1299 |
|
| 1300 |
|
| 1301 | static inline void pgalloc_tag_sub_pages(struct alloc_tag *tag, unsigned int nr)
|
| 1302 | {
|
| 1303 | if (tag)
|
| 1304 | this_cpu_sub(tag->counters->bytes, PAGE_SIZE * nr);
|
| 1305 | }
|
| 1306 |
|
| 1307 | #else
|
| 1308 |
|
| 1309 | static inline void pgalloc_tag_add(struct page *page, struct task_struct *task,
|
| 1310 | unsigned int nr, unsigned int alloc_flags) {}
|
| 1311 | static inline void pgalloc_tag_sub(struct page *page, unsigned int nr) {}
|
| 1312 | static inline void pgalloc_tag_sub_pages(struct alloc_tag *tag, unsigned int nr) {}
|
| 1313 |
|
| 1314 | #endif
|
| 1315 |
|
| 1316 | static __always_inline bool __free_pages_prepare(struct page *page,
|
| 1317 | unsigned int order, fpi_t fpi_flags)
|
| 1318 | {
|
| 1319 | int bad = 0;
|
| 1320 | bool skip_kasan_poison = should_skip_kasan_poison(page);
|
| 1321 | bool init = want_init_on_free();
|
| 1322 | bool compound = PageCompound(page);
|
| 1323 | struct folio *folio = page_folio(page);
|
| 1324 |
|
| 1325 | if (fpi_flags & FPI_PREPARED)
|
| 1326 | return true;
|
| 1327 |
|
| 1328 | VM_BUG_ON_PAGE(PageTail(page), page);
|
| 1329 |
|
| 1330 | trace_mm_page_free(page, order);
|
| 1331 | kmsan_free_page(page, order);
|
| 1332 |
|
| 1333 | if (memcg_kmem_online() && PageMemcgKmem(page))
|
| 1334 | __memcg_kmem_uncharge_page(page, order);
|
| 1335 |
|
| 1336 | |
| 1337 | * In rare cases, when truncation or holepunching raced with
|
| 1338 | * munlock after VM_LOCKED was cleared, Mlocked may still be
|
| 1339 | * found set here. This does not indicate a problem, unless
|
| 1340 | * "unevictable_pgs_cleared" appears worryingly large.
|
| 1341 |
|
| 1342 | if (unlikely(folio_test_mlocked(folio))) {
|
| 1343 | long nr_pages = folio_nr_pages(folio);
|
| 1344 |
|
| 1345 | __folio_clear_mlocked(folio);
|
| 1346 | zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
|
| 1347 | count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
|
| 1348 | }
|
| 1349 |
|
| 1350 | if (unlikely(PageHWPoison(page)) && !order) {
|
| 1351 |
|
| 1352 | reset_page_owner(page, order);
|
| 1353 | page_table_check_free(page, order);
|
| 1354 | pgalloc_tag_sub(page, 1 << order);
|
| 1355 |
|
| 1356 | |
| 1357 | * The page is isolated and accounted for.
|
| 1358 | * Mark the codetag as empty to avoid accounting error
|
| 1359 | * when the page is freed by unpoison_memory().
|
| 1360 |
|
| 1361 | clear_page_tag_ref(page);
|
| 1362 | return false;
|
| 1363 | }
|
| 1364 |
|
| 1365 | VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
|
| 1366 |
|
| 1367 | |
| 1368 | * Check tail pages before head page information is cleared to
|
| 1369 | * avoid checking PageCompound for order-0 pages.
|
| 1370 |
|
| 1371 | if (unlikely(order)) {
|
| 1372 | int i;
|
| 1373 |
|
| 1374 | if (compound) {
|
| 1375 | page[1].flags.f &= ~PAGE_FLAGS_SECOND;
|
| 1376 | #ifdef NR_PAGES_IN_LARGE_FOLIO
|
| 1377 | folio->_nr_pages = 0;
|
| 1378 | #endif
|
| 1379 | }
|
| 1380 | for (i = 1; i < (1 << order); i++) {
|
| 1381 | struct page *tail_page = page + i;
|
| 1382 |
|
| 1383 | if (compound)
|
| 1384 | bad += free_tail_page_prepare(page, tail_page);
|
| 1385 | if (is_check_pages_enabled()) {
|
| 1386 | if (free_page_is_bad(tail_page)) {
|
| 1387 | bad++;
|
| 1388 | continue;
|
| 1389 | }
|
| 1390 |
|
| 1391 | if (tail_page->private) {
|
| 1392 | bad_page(tail_page, "nonzero private");
|
| 1393 | bad++;
|
| 1394 | continue;
|
| 1395 | }
|
| 1396 | }
|
| 1397 | tail_page->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
|
| 1398 | }
|
| 1399 | }
|
| 1400 | if (folio_test_anon(folio)) {
|
| 1401 | mod_mthp_stat(order, MTHP_STAT_NR_ANON, -1);
|
| 1402 | folio->mapping = NULL;
|
| 1403 | }
|
| 1404 | if (unlikely(page_has_type(page)))
|
| 1405 |
|
| 1406 | page->page_type = UINT_MAX;
|
| 1407 |
|
| 1408 | if (is_check_pages_enabled()) {
|
| 1409 | if (free_page_is_bad(page))
|
| 1410 | bad++;
|
| 1411 | if (bad)
|
| 1412 | return false;
|
| 1413 | }
|
| 1414 |
|
| 1415 | page_cpupid_reset_last(page);
|
| 1416 | page->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
|
| 1417 | page->private = 0;
|
| 1418 | reset_page_owner(page, order);
|
| 1419 | page_table_check_free(page, order);
|
| 1420 | pgalloc_tag_sub(page, 1 << order);
|
| 1421 |
|
| 1422 | if (!PageHighMem(page) && !(fpi_flags & FPI_NOLOCK)) {
|
| 1423 | debug_check_no_locks_freed(page_address(page),
|
| 1424 | PAGE_SIZE << order);
|
| 1425 | debug_check_no_obj_freed(page_address(page),
|
| 1426 | PAGE_SIZE << order);
|
| 1427 | }
|
| 1428 |
|
| 1429 | kernel_poison_pages(page, 1 << order);
|
| 1430 |
|
| 1431 | |
| 1432 | * As memory initialization might be integrated into KASAN,
|
| 1433 | * KASAN poisoning and memory initialization code must be
|
| 1434 | * kept together to avoid discrepancies in behavior.
|
| 1435 | *
|
| 1436 | * With hardware tag-based KASAN, memory tags must be set before the
|
| 1437 | * page becomes unavailable via debug_pagealloc or arch_free_page.
|
| 1438 |
|
| 1439 | if (!skip_kasan_poison) {
|
| 1440 | kasan_poison_pages(page, order, init);
|
| 1441 |
|
| 1442 |
|
| 1443 | if (kasan_has_integrated_init())
|
| 1444 | init = false;
|
| 1445 | }
|
| 1446 | if (init)
|
| 1447 | clear_highpages_kasan_tagged(page, 1 << order);
|
| 1448 |
|
| 1449 | |
| 1450 | * arch_free_page() can make the page's contents inaccessible. s390
|
| 1451 | * does this. So nothing which can access the page's contents should
|
| 1452 | * happen after this.
|
| 1453 |
|
| 1454 | arch_free_page(page, order);
|
| 1455 |
|
| 1456 | debug_pagealloc_unmap_pages(page, 1 << order);
|
| 1457 |
|
| 1458 | return true;
|
| 1459 | }
|
| 1460 |
|
| 1461 | bool free_pages_prepare(struct page *page, unsigned int order)
|
| 1462 | {
|
| 1463 | return __free_pages_prepare(page, order, FPI_NONE);
|
| 1464 | }
|
| 1465 |
|
| 1466 | |
| 1467 | * Frees a number of pages from the PCP lists
|
| 1468 | * Assumes all pages on list are in same zone.
|
| 1469 | * count is the number of pages to free.
|
| 1470 |
|
| 1471 | static void free_pcppages_bulk(struct zone *zone, int count,
|
| 1472 | struct per_cpu_pages *pcp,
|
| 1473 | int pindex)
|
| 1474 | {
|
| 1475 | unsigned int order;
|
| 1476 | struct page *page;
|
| 1477 |
|
| 1478 | |
| 1479 | * Ensure proper count is passed which otherwise would stuck in the
|
| 1480 | * below while (list_empty(list)) loop.
|
| 1481 |
|
| 1482 | count = min(pcp->count, count);
|
| 1483 |
|
| 1484 |
|
| 1485 | pindex = pindex - 1;
|
| 1486 |
|
| 1487 | guard(spinlock_irqsave)(&zone->lock);
|
| 1488 |
|
| 1489 | while (count > 0) {
|
| 1490 | struct list_head *list;
|
| 1491 | int nr_pages;
|
| 1492 |
|
| 1493 |
|
| 1494 | do {
|
| 1495 | if (++pindex > NR_PCP_LISTS - 1)
|
| 1496 | pindex = 0;
|
| 1497 | list = &pcp->lists[pindex];
|
| 1498 | } while (list_empty(list));
|
| 1499 |
|
| 1500 | order = pindex_to_order(pindex);
|
| 1501 | nr_pages = 1 << order;
|
| 1502 | do {
|
| 1503 | unsigned long pfn;
|
| 1504 | int mt;
|
| 1505 |
|
| 1506 | page = list_last_entry(list, struct page, pcp_list);
|
| 1507 | pfn = page_to_pfn(page);
|
| 1508 | mt = get_pfnblock_migratetype(page, pfn);
|
| 1509 |
|
| 1510 |
|
| 1511 | list_del(&page->pcp_list);
|
| 1512 | count -= nr_pages;
|
| 1513 | pcp->count -= nr_pages;
|
| 1514 |
|
| 1515 | __free_one_page(page, pfn, zone, order, mt, FPI_NONE);
|
| 1516 | trace_mm_page_pcpu_drain(page, order, mt);
|
| 1517 | } while (count > 0 && !list_empty(list));
|
| 1518 | }
|
| 1519 | }
|
| 1520 |
|
| 1521 |
|
| 1522 | static void split_large_buddy(struct zone *zone, struct page *page,
|
| 1523 | unsigned long pfn, int order, fpi_t fpi)
|
| 1524 | {
|
| 1525 | unsigned long end = pfn + (1 << order);
|
| 1526 |
|
| 1527 | VM_WARN_ON_ONCE(!IS_ALIGNED(pfn, 1 << order));
|
| 1528 |
|
| 1529 | VM_WARN_ON_ONCE(PageBuddy(page));
|
| 1530 |
|
| 1531 | if (order > pageblock_order)
|
| 1532 | order = pageblock_order;
|
| 1533 |
|
| 1534 | do {
|
| 1535 | int mt = get_pfnblock_migratetype(page, pfn);
|
| 1536 |
|
| 1537 | __free_one_page(page, pfn, zone, order, mt, fpi);
|
| 1538 | pfn += 1 << order;
|
| 1539 | if (pfn == end)
|
| 1540 | break;
|
| 1541 | page = pfn_to_page(pfn);
|
| 1542 | } while (1);
|
| 1543 | }
|
| 1544 |
|
| 1545 | static void add_page_to_zone_llist(struct zone *zone, struct page *page,
|
| 1546 | unsigned int order)
|
| 1547 | {
|
| 1548 |
|
| 1549 | page->private = order;
|
| 1550 |
|
| 1551 | llist_add(&page->pcp_llist, &zone->trylock_free_pages);
|
| 1552 | }
|
| 1553 |
|
| 1554 | static void free_one_page(struct zone *zone, struct page *page,
|
| 1555 | unsigned long pfn, unsigned int order,
|
| 1556 | fpi_t fpi_flags)
|
| 1557 | {
|
| 1558 | struct llist_head *llhead;
|
| 1559 | unsigned long flags;
|
| 1560 |
|
| 1561 | if (unlikely(fpi_flags & FPI_NOLOCK)) {
|
| 1562 | if (!can_spin_trylock() || !spin_trylock_irqsave(&zone->lock, flags)) {
|
| 1563 | add_page_to_zone_llist(zone, page, order);
|
| 1564 | return;
|
| 1565 | }
|
| 1566 | } else {
|
| 1567 | spin_lock_irqsave(&zone->lock, flags);
|
| 1568 | }
|
| 1569 |
|
| 1570 |
|
| 1571 | llhead = &zone->trylock_free_pages;
|
| 1572 | if (unlikely(!llist_empty(llhead) && !(fpi_flags & FPI_NOLOCK))) {
|
| 1573 | struct llist_node *llnode;
|
| 1574 | struct page *p, *tmp;
|
| 1575 |
|
| 1576 | llnode = llist_del_all(llhead);
|
| 1577 | llist_for_each_entry_safe(p, tmp, llnode, pcp_llist) {
|
| 1578 | unsigned int p_order = p->private;
|
| 1579 |
|
| 1580 | split_large_buddy(zone, p, page_to_pfn(p), p_order, fpi_flags);
|
| 1581 | __count_vm_events(PGFREE, 1 << p_order);
|
| 1582 | }
|
| 1583 | }
|
| 1584 | split_large_buddy(zone, page, pfn, order, fpi_flags);
|
| 1585 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 1586 |
|
| 1587 | __count_vm_events(PGFREE, 1 << order);
|
| 1588 | }
|
| 1589 |
|
| 1590 | static void __free_pages_ok(struct page *page, unsigned int order,
|
| 1591 | fpi_t fpi_flags)
|
| 1592 | {
|
| 1593 | unsigned long pfn = page_to_pfn(page);
|
| 1594 | struct zone *zone = page_zone(page);
|
| 1595 |
|
| 1596 | if (__free_pages_prepare(page, order, fpi_flags))
|
| 1597 | free_one_page(zone, page, pfn, order, fpi_flags);
|
| 1598 | }
|
| 1599 |
|
| 1600 | void __meminit __free_pages_core(struct page *page, unsigned int order,
|
| 1601 | enum meminit_context context)
|
| 1602 | {
|
| 1603 | unsigned int nr_pages = 1 << order;
|
| 1604 | struct page *p = page;
|
| 1605 | unsigned int loop;
|
| 1606 |
|
| 1607 | |
| 1608 | * When initializing the memmap, __init_single_page() sets the refcount
|
| 1609 | * of all pages to 1 ("allocated"/"not free"). We have to set the
|
| 1610 | * refcount of all involved pages to 0.
|
| 1611 | *
|
| 1612 | * Note that hotplugged memory pages are initialized to PageOffline().
|
| 1613 | * Pages freed from memblock might be marked as reserved.
|
| 1614 |
|
| 1615 | if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG) &&
|
| 1616 | unlikely(context == MEMINIT_HOTPLUG)) {
|
| 1617 | for (loop = 0; loop < nr_pages; loop++, p++) {
|
| 1618 | VM_WARN_ON_ONCE(PageReserved(p));
|
| 1619 | __ClearPageOffline(p);
|
| 1620 | set_page_count(p, 0);
|
| 1621 | }
|
| 1622 |
|
| 1623 | adjust_managed_page_count(page, nr_pages);
|
| 1624 | } else {
|
| 1625 | for (loop = 0; loop < nr_pages; loop++, p++) {
|
| 1626 | __ClearPageReserved(p);
|
| 1627 | set_page_count(p, 0);
|
| 1628 | }
|
| 1629 |
|
| 1630 |
|
| 1631 | atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
|
| 1632 | }
|
| 1633 |
|
| 1634 | if (page_contains_unaccepted(page, order)) {
|
| 1635 | if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
|
| 1636 | return;
|
| 1637 |
|
| 1638 | accept_memory(page_to_phys(page), PAGE_SIZE << order);
|
| 1639 | }
|
| 1640 |
|
| 1641 | |
| 1642 | * Bypass PCP and place fresh pages right to the tail, primarily
|
| 1643 | * relevant for memory onlining.
|
| 1644 |
|
| 1645 | __free_pages_ok(page, order, FPI_TO_TAIL);
|
| 1646 | }
|
| 1647 |
|
| 1648 | |
| 1649 | * Check that the whole (or subset of) a pageblock given by the interval of
|
| 1650 | * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
|
| 1651 | * with the migration of free compaction scanner.
|
| 1652 | *
|
| 1653 | * Return struct page pointer of start_pfn, or NULL if checks were not passed.
|
| 1654 | *
|
| 1655 | * It's possible on some configurations to have a setup like node0 node1 node0
|
| 1656 | * i.e. it's possible that all pages within a zones range of pages do not
|
| 1657 | * belong to a single zone. We assume that a border between node0 and node1
|
| 1658 | * can occur within a single pageblock, but not a node0 node1 node0
|
| 1659 | * interleaving within a single pageblock. It is therefore sufficient to check
|
| 1660 | * the first and last page of a pageblock and avoid checking each individual
|
| 1661 | * page in a pageblock.
|
| 1662 | *
|
| 1663 | * Note: the function may return non-NULL struct page even for a page block
|
| 1664 | * which contains a memory hole (i.e. there is no physical memory for a subset
|
| 1665 | * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
|
| 1666 | * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
|
| 1667 | * even though the start pfn is online and valid. This should be safe most of
|
| 1668 | * the time because struct pages are still initialized via init_unavailable_range()
|
| 1669 | * and pfn walkers shouldn't touch any physical memory range for which they do
|
| 1670 | * not recognize any specific metadata in struct pages.
|
| 1671 |
|
| 1672 | struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
|
| 1673 | unsigned long end_pfn, struct zone *zone)
|
| 1674 | {
|
| 1675 | struct page *start_page;
|
| 1676 | struct page *end_page;
|
| 1677 |
|
| 1678 |
|
| 1679 | end_pfn--;
|
| 1680 |
|
| 1681 | if (!pfn_valid(end_pfn))
|
| 1682 | return NULL;
|
| 1683 |
|
| 1684 | start_page = pfn_to_online_page(start_pfn);
|
| 1685 | if (!start_page)
|
| 1686 | return NULL;
|
| 1687 |
|
| 1688 | if (page_zone(start_page) != zone)
|
| 1689 | return NULL;
|
| 1690 |
|
| 1691 | end_page = pfn_to_page(end_pfn);
|
| 1692 |
|
| 1693 |
|
| 1694 | if (page_zone_id(start_page) != page_zone_id(end_page))
|
| 1695 | return NULL;
|
| 1696 |
|
| 1697 | return start_page;
|
| 1698 | }
|
| 1699 |
|
| 1700 | |
| 1701 | * The order of subdivision here is critical for the IO subsystem.
|
| 1702 | * Please do not alter this order without good reasons and regression
|
| 1703 | * testing. Specifically, as large blocks of memory are subdivided,
|
| 1704 | * the order in which smaller blocks are delivered depends on the order
|
| 1705 | * they're subdivided in this function. This is the primary factor
|
| 1706 | * influencing the order in which pages are delivered to the IO
|
| 1707 | * subsystem according to empirical testing, and this is also justified
|
| 1708 | * by considering the behavior of a buddy system containing a single
|
| 1709 | * large block of memory acted on by a series of small allocations.
|
| 1710 | * This behavior is a critical factor in sglist merging's success.
|
| 1711 | *
|
| 1712 | * -- nyc
|
| 1713 |
|
| 1714 | static inline unsigned int expand(struct zone *zone, struct page *page, int low,
|
| 1715 | int high, int migratetype)
|
| 1716 | {
|
| 1717 | unsigned int size = 1 << high;
|
| 1718 | unsigned int nr_added = 0;
|
| 1719 |
|
| 1720 | while (high > low) {
|
| 1721 | high--;
|
| 1722 | size >>= 1;
|
| 1723 | VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
|
| 1724 |
|
| 1725 | |
| 1726 | * Mark as guard pages (or page), that will allow to
|
| 1727 | * merge back to allocator when buddy will be freed.
|
| 1728 | * Corresponding page table entries will not be touched,
|
| 1729 | * pages will stay not present in virtual address space
|
| 1730 |
|
| 1731 | if (set_page_guard(zone, &page[size], high))
|
| 1732 | continue;
|
| 1733 |
|
| 1734 | __add_to_free_list(&page[size], zone, high, migratetype, false);
|
| 1735 | set_buddy_order(&page[size], high);
|
| 1736 | nr_added += size;
|
| 1737 | }
|
| 1738 |
|
| 1739 | return nr_added;
|
| 1740 | }
|
| 1741 |
|
| 1742 | static __always_inline void page_del_and_expand(struct zone *zone,
|
| 1743 | struct page *page, int low,
|
| 1744 | int high, int migratetype)
|
| 1745 | {
|
| 1746 | int nr_pages = 1 << high;
|
| 1747 |
|
| 1748 | __del_page_from_free_list(page, zone, high, migratetype);
|
| 1749 | nr_pages -= expand(zone, page, low, high, migratetype);
|
| 1750 | account_freepages(zone, -nr_pages, migratetype);
|
| 1751 | }
|
| 1752 |
|
| 1753 | static void check_new_page_bad(struct page *page)
|
| 1754 | {
|
| 1755 | if (unlikely(PageHWPoison(page))) {
|
| 1756 |
|
| 1757 | if (PageBuddy(page))
|
| 1758 | __ClearPageBuddy(page);
|
| 1759 | return;
|
| 1760 | }
|
| 1761 |
|
| 1762 | bad_page(page,
|
| 1763 | page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
|
| 1764 | }
|
| 1765 |
|
| 1766 | |
| 1767 | * This page is about to be returned from the page allocator
|
| 1768 |
|
| 1769 | static bool check_new_page(struct page *page)
|
| 1770 | {
|
| 1771 | if (likely(page_expected_state(page,
|
| 1772 | PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
|
| 1773 | return false;
|
| 1774 |
|
| 1775 | check_new_page_bad(page);
|
| 1776 | return true;
|
| 1777 | }
|
| 1778 |
|
| 1779 | static inline bool check_new_pages(struct page *page, unsigned int order)
|
| 1780 | {
|
| 1781 | if (is_check_pages_enabled()) {
|
| 1782 | for (int i = 0; i < (1 << order); i++) {
|
| 1783 | struct page *p = page + i;
|
| 1784 |
|
| 1785 | if (check_new_page(p))
|
| 1786 | return true;
|
| 1787 | }
|
| 1788 | }
|
| 1789 |
|
| 1790 | return false;
|
| 1791 | }
|
| 1792 |
|
| 1793 | static inline bool should_skip_kasan_unpoison(gfp_t flags)
|
| 1794 | {
|
| 1795 |
|
| 1796 | if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
|
| 1797 | IS_ENABLED(CONFIG_KASAN_SW_TAGS))
|
| 1798 | return false;
|
| 1799 |
|
| 1800 |
|
| 1801 | if (!kasan_hw_tags_enabled())
|
| 1802 | return true;
|
| 1803 |
|
| 1804 | |
| 1805 | * With hardware tag-based KASAN enabled, skip if this has been
|
| 1806 | * requested via __GFP_SKIP_KASAN.
|
| 1807 |
|
| 1808 | return flags & __GFP_SKIP_KASAN;
|
| 1809 | }
|
| 1810 |
|
| 1811 | static inline bool should_skip_init(gfp_t flags)
|
| 1812 | {
|
| 1813 |
|
| 1814 | if (!kasan_hw_tags_enabled())
|
| 1815 | return false;
|
| 1816 |
|
| 1817 |
|
| 1818 | return (flags & __GFP_SKIP_ZERO);
|
| 1819 | }
|
| 1820 |
|
| 1821 | inline void post_alloc_hook(struct page *page, unsigned int order,
|
| 1822 | gfp_t gfp_flags, unsigned int alloc_flags)
|
| 1823 | {
|
| 1824 | const bool zero_tags = gfp_flags & __GFP_ZEROTAGS;
|
| 1825 | bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
|
| 1826 | !should_skip_init(gfp_flags);
|
| 1827 | int i;
|
| 1828 |
|
| 1829 | set_page_private(page, 0);
|
| 1830 |
|
| 1831 | arch_alloc_page(page, order);
|
| 1832 | debug_pagealloc_map_pages(page, 1 << order);
|
| 1833 |
|
| 1834 | |
| 1835 | * Page unpoisoning must happen before memory initialization.
|
| 1836 | * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
|
| 1837 | * allocations and the page unpoisoning code will complain.
|
| 1838 |
|
| 1839 | kernel_unpoison_pages(page, 1 << order);
|
| 1840 |
|
| 1841 | |
| 1842 | * As memory initialization might be integrated into KASAN,
|
| 1843 | * KASAN unpoisoning and memory initialization code must be
|
| 1844 | * kept together to avoid discrepancies in behavior.
|
| 1845 |
|
| 1846 |
|
| 1847 | |
| 1848 | * Clearing tags can efficiently clear the memory for us as well, if
|
| 1849 | * required.
|
| 1850 |
|
| 1851 | if (zero_tags)
|
| 1852 | init = tag_clear_highpages(page, 1 << order, init);
|
| 1853 |
|
| 1854 | if (!should_skip_kasan_unpoison(gfp_flags) &&
|
| 1855 | kasan_unpoison_pages(page, order, init)) {
|
| 1856 |
|
| 1857 | if (kasan_has_integrated_init())
|
| 1858 | init = false;
|
| 1859 | } else {
|
| 1860 | |
| 1861 | * If memory tags have not been set by KASAN, reset the page
|
| 1862 | * tags to ensure page_address() dereferencing does not fault.
|
| 1863 |
|
| 1864 | for (i = 0; i != 1 << order; ++i)
|
| 1865 | page_kasan_tag_reset(page + i);
|
| 1866 | }
|
| 1867 |
|
| 1868 | if (init)
|
| 1869 | clear_highpages_kasan_tagged(page, 1 << order);
|
| 1870 |
|
| 1871 | set_page_owner(page, order, gfp_flags);
|
| 1872 | page_table_check_alloc(page, order);
|
| 1873 | pgalloc_tag_add(page, current, 1 << order, alloc_flags);
|
| 1874 | }
|
| 1875 |
|
| 1876 | static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
|
| 1877 | unsigned int alloc_flags)
|
| 1878 | {
|
| 1879 | post_alloc_hook(page, order, gfp_flags, alloc_flags);
|
| 1880 |
|
| 1881 | if (order && (gfp_flags & __GFP_COMP))
|
| 1882 | prep_compound_page(page, order);
|
| 1883 |
|
| 1884 | |
| 1885 | * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
|
| 1886 | * allocate the page. The expectation is that the caller is taking
|
| 1887 | * steps that will free more memory. The caller should avoid the page
|
| 1888 | * being used for !PFMEMALLOC purposes.
|
| 1889 |
|
| 1890 | if (alloc_flags & ALLOC_NO_WATERMARKS)
|
| 1891 | set_page_pfmemalloc(page);
|
| 1892 | else
|
| 1893 | clear_page_pfmemalloc(page);
|
| 1894 | }
|
| 1895 |
|
| 1896 | |
| 1897 | * Go through the free lists for the given migratetype and remove
|
| 1898 | * the smallest available page from the freelists
|
| 1899 |
|
| 1900 | static __always_inline
|
| 1901 | struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
|
| 1902 | int migratetype)
|
| 1903 | {
|
| 1904 | unsigned int current_order;
|
| 1905 | struct free_area *area;
|
| 1906 | struct page *page;
|
| 1907 |
|
| 1908 |
|
| 1909 | for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
|
| 1910 | area = &(zone->free_area[current_order]);
|
| 1911 | page = get_page_from_free_area(area, migratetype);
|
| 1912 | if (!page)
|
| 1913 | continue;
|
| 1914 |
|
| 1915 | page_del_and_expand(zone, page, order, current_order,
|
| 1916 | migratetype);
|
| 1917 | trace_mm_page_alloc_zone_locked(page, order, migratetype,
|
| 1918 | pcp_allowed_order(order) &&
|
| 1919 | migratetype < MIGRATE_PCPTYPES);
|
| 1920 | return page;
|
| 1921 | }
|
| 1922 |
|
| 1923 | return NULL;
|
| 1924 | }
|
| 1925 |
|
| 1926 |
|
| 1927 | |
| 1928 | * This array describes the order lists are fallen back to when
|
| 1929 | * the free lists for the desirable migrate type are depleted
|
| 1930 | *
|
| 1931 | * The other migratetypes do not have fallbacks.
|
| 1932 |
|
| 1933 | static int fallbacks[MIGRATE_PCPTYPES][MIGRATE_PCPTYPES - 1] = {
|
| 1934 | [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
|
| 1935 | [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
|
| 1936 | [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
|
| 1937 | };
|
| 1938 |
|
| 1939 | #ifdef CONFIG_CMA
|
| 1940 | static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
|
| 1941 | unsigned int order)
|
| 1942 | {
|
| 1943 | return __rmqueue_smallest(zone, order, MIGRATE_CMA);
|
| 1944 | }
|
| 1945 | #else
|
| 1946 | static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
|
| 1947 | unsigned int order) { return NULL; }
|
| 1948 | #endif
|
| 1949 |
|
| 1950 | |
| 1951 | * Move all free pages of a block to new type's freelist. Caller needs to
|
| 1952 | * change the block type.
|
| 1953 |
|
| 1954 | static int __move_freepages_block(struct zone *zone, unsigned long start_pfn,
|
| 1955 | int old_mt, int new_mt)
|
| 1956 | {
|
| 1957 | struct page *page;
|
| 1958 | unsigned long pfn, end_pfn;
|
| 1959 | unsigned int order;
|
| 1960 | int pages_moved = 0;
|
| 1961 |
|
| 1962 | VM_WARN_ON(start_pfn & (pageblock_nr_pages - 1));
|
| 1963 | end_pfn = pageblock_end_pfn(start_pfn);
|
| 1964 |
|
| 1965 | for (pfn = start_pfn; pfn < end_pfn;) {
|
| 1966 | page = pfn_to_page(pfn);
|
| 1967 | if (!PageBuddy(page)) {
|
| 1968 | pfn++;
|
| 1969 | continue;
|
| 1970 | }
|
| 1971 |
|
| 1972 |
|
| 1973 | VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
|
| 1974 | VM_BUG_ON_PAGE(page_zone(page) != zone, page);
|
| 1975 |
|
| 1976 | order = buddy_order(page);
|
| 1977 |
|
| 1978 | move_to_free_list(page, zone, order, old_mt, new_mt);
|
| 1979 |
|
| 1980 | pfn += 1 << order;
|
| 1981 | pages_moved += 1 << order;
|
| 1982 | }
|
| 1983 |
|
| 1984 | return pages_moved;
|
| 1985 | }
|
| 1986 |
|
| 1987 | static bool prep_move_freepages_block(struct zone *zone, struct page *page,
|
| 1988 | unsigned long *start_pfn,
|
| 1989 | int *num_free, int *num_movable)
|
| 1990 | {
|
| 1991 | unsigned long pfn, start, end;
|
| 1992 |
|
| 1993 | pfn = page_to_pfn(page);
|
| 1994 | start = pageblock_start_pfn(pfn);
|
| 1995 | end = pageblock_end_pfn(pfn);
|
| 1996 |
|
| 1997 | |
| 1998 | * The caller only has the lock for @zone, don't touch ranges
|
| 1999 | * that straddle into other zones. While we could move part of
|
| 2000 | * the range that's inside the zone, this call is usually
|
| 2001 | * accompanied by other operations such as migratetype updates
|
| 2002 | * which also should be locked.
|
| 2003 |
|
| 2004 | if (!zone_spans_pfn(zone, start))
|
| 2005 | return false;
|
| 2006 | if (!zone_spans_pfn(zone, end - 1))
|
| 2007 | return false;
|
| 2008 |
|
| 2009 | *start_pfn = start;
|
| 2010 |
|
| 2011 | if (num_free) {
|
| 2012 | *num_free = 0;
|
| 2013 | *num_movable = 0;
|
| 2014 | for (pfn = start; pfn < end;) {
|
| 2015 | page = pfn_to_page(pfn);
|
| 2016 | if (PageBuddy(page)) {
|
| 2017 | int nr = 1 << buddy_order(page);
|
| 2018 |
|
| 2019 | *num_free += nr;
|
| 2020 | pfn += nr;
|
| 2021 | continue;
|
| 2022 | }
|
| 2023 | |
| 2024 | * We assume that pages that could be isolated for
|
| 2025 | * migration are movable. But we don't actually try
|
| 2026 | * isolating, as that would be expensive.
|
| 2027 |
|
| 2028 | if (PageLRU(page) || page_has_movable_ops(page))
|
| 2029 | (*num_movable)++;
|
| 2030 | pfn++;
|
| 2031 | }
|
| 2032 | }
|
| 2033 |
|
| 2034 | return true;
|
| 2035 | }
|
| 2036 |
|
| 2037 | static int move_freepages_block(struct zone *zone, struct page *page,
|
| 2038 | int old_mt, int new_mt)
|
| 2039 | {
|
| 2040 | unsigned long start_pfn;
|
| 2041 | int res;
|
| 2042 |
|
| 2043 | if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
|
| 2044 | return -1;
|
| 2045 |
|
| 2046 | res = __move_freepages_block(zone, start_pfn, old_mt, new_mt);
|
| 2047 | set_pageblock_migratetype(pfn_to_page(start_pfn), new_mt);
|
| 2048 |
|
| 2049 | return res;
|
| 2050 |
|
| 2051 | }
|
| 2052 |
|
| 2053 | #ifdef CONFIG_MEMORY_ISOLATION
|
| 2054 |
|
| 2055 | static unsigned long find_large_buddy(unsigned long start_pfn)
|
| 2056 | {
|
| 2057 | |
| 2058 | * If start_pfn is not an order-0 PageBuddy, next PageBuddy containing
|
| 2059 | * start_pfn has minimal order of __ffs(start_pfn) + 1. Start checking
|
| 2060 | * the order with __ffs(start_pfn). If start_pfn is order-0 PageBuddy,
|
| 2061 | * the starting order does not matter.
|
| 2062 |
|
| 2063 | int order = start_pfn ? __ffs(start_pfn) : MAX_PAGE_ORDER;
|
| 2064 | struct page *page;
|
| 2065 | unsigned long pfn = start_pfn;
|
| 2066 |
|
| 2067 | while (!PageBuddy(page = pfn_to_page(pfn))) {
|
| 2068 |
|
| 2069 | if (++order > MAX_PAGE_ORDER)
|
| 2070 | return start_pfn;
|
| 2071 | pfn &= ~0UL << order;
|
| 2072 | }
|
| 2073 |
|
| 2074 | |
| 2075 | * Found a preceding buddy, but does it straddle?
|
| 2076 |
|
| 2077 | if (pfn + (1 << buddy_order(page)) > start_pfn)
|
| 2078 | return pfn;
|
| 2079 |
|
| 2080 |
|
| 2081 | return start_pfn;
|
| 2082 | }
|
| 2083 |
|
| 2084 | static inline void toggle_pageblock_isolate(struct page *page, bool isolate)
|
| 2085 | {
|
| 2086 | if (isolate)
|
| 2087 | set_pageblock_isolate(page);
|
| 2088 | else
|
| 2089 | clear_pageblock_isolate(page);
|
| 2090 | }
|
| 2091 |
|
| 2092 | |
| 2093 | * __move_freepages_block_isolate - move free pages in block for page isolation
|
| 2094 | * @zone: the zone
|
| 2095 | * @page: the pageblock page
|
| 2096 | * @isolate: to isolate the given pageblock or unisolate it
|
| 2097 | *
|
| 2098 | * This is similar to move_freepages_block(), but handles the special
|
| 2099 | * case encountered in page isolation, where the block of interest
|
| 2100 | * might be part of a larger buddy spanning multiple pageblocks.
|
| 2101 | *
|
| 2102 | * Unlike the regular page allocator path, which moves pages while
|
| 2103 | * stealing buddies off the freelist, page isolation is interested in
|
| 2104 | * arbitrary pfn ranges that may have overlapping buddies on both ends.
|
| 2105 | *
|
| 2106 | * This function handles that. Straddling buddies are split into
|
| 2107 | * individual pageblocks. Only the block of interest is moved.
|
| 2108 | *
|
| 2109 | * Returns %true if pages could be moved, %false otherwise.
|
| 2110 |
|
| 2111 | static bool __move_freepages_block_isolate(struct zone *zone,
|
| 2112 | struct page *page, bool isolate)
|
| 2113 | {
|
| 2114 | unsigned long start_pfn, buddy_pfn;
|
| 2115 | int from_mt;
|
| 2116 | int to_mt;
|
| 2117 | struct page *buddy;
|
| 2118 |
|
| 2119 | if (isolate == get_pageblock_isolate(page)) {
|
| 2120 | VM_WARN_ONCE(1, "%s a pageblock that is already in that state",
|
| 2121 | isolate ? "Isolate" : "Unisolate");
|
| 2122 | return false;
|
| 2123 | }
|
| 2124 |
|
| 2125 | if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
|
| 2126 | return false;
|
| 2127 |
|
| 2128 |
|
| 2129 | if (pageblock_order == MAX_PAGE_ORDER)
|
| 2130 | goto move;
|
| 2131 |
|
| 2132 | buddy_pfn = find_large_buddy(start_pfn);
|
| 2133 | buddy = pfn_to_page(buddy_pfn);
|
| 2134 |
|
| 2135 | if (PageBuddy(buddy) && buddy_order(buddy) > pageblock_order) {
|
| 2136 | int order = buddy_order(buddy);
|
| 2137 |
|
| 2138 | del_page_from_free_list(buddy, zone, order,
|
| 2139 | get_pfnblock_migratetype(buddy, buddy_pfn));
|
| 2140 | toggle_pageblock_isolate(page, isolate);
|
| 2141 | split_large_buddy(zone, buddy, buddy_pfn, order, FPI_NONE);
|
| 2142 | return true;
|
| 2143 | }
|
| 2144 |
|
| 2145 | move:
|
| 2146 |
|
| 2147 | if (isolate) {
|
| 2148 | from_mt = __get_pfnblock_flags_mask(page, page_to_pfn(page),
|
| 2149 | PAGEBLOCK_MIGRATETYPE_MASK);
|
| 2150 | to_mt = MIGRATE_ISOLATE;
|
| 2151 | } else {
|
| 2152 | from_mt = MIGRATE_ISOLATE;
|
| 2153 | to_mt = __get_pfnblock_flags_mask(page, page_to_pfn(page),
|
| 2154 | PAGEBLOCK_MIGRATETYPE_MASK);
|
| 2155 | }
|
| 2156 |
|
| 2157 | __move_freepages_block(zone, start_pfn, from_mt, to_mt);
|
| 2158 | toggle_pageblock_isolate(pfn_to_page(start_pfn), isolate);
|
| 2159 |
|
| 2160 | return true;
|
| 2161 | }
|
| 2162 |
|
| 2163 | bool pageblock_isolate_and_move_free_pages(struct zone *zone, struct page *page)
|
| 2164 | {
|
| 2165 | return __move_freepages_block_isolate(zone, page, true);
|
| 2166 | }
|
| 2167 |
|
| 2168 | bool pageblock_unisolate_and_move_free_pages(struct zone *zone, struct page *page)
|
| 2169 | {
|
| 2170 | return __move_freepages_block_isolate(zone, page, false);
|
| 2171 | }
|
| 2172 |
|
| 2173 | #endif
|
| 2174 |
|
| 2175 | static inline bool boost_watermark(struct zone *zone)
|
| 2176 | {
|
| 2177 | unsigned long max_boost;
|
| 2178 |
|
| 2179 | if (!watermark_boost_factor)
|
| 2180 | return false;
|
| 2181 | |
| 2182 | * Don't bother in zones that are unlikely to produce results.
|
| 2183 | * On small machines, including kdump capture kernels running
|
| 2184 | * in a small area, boosting the watermark can cause an out of
|
| 2185 | * memory situation immediately.
|
| 2186 |
|
| 2187 | if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
|
| 2188 | return false;
|
| 2189 |
|
| 2190 | max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
|
| 2191 | watermark_boost_factor, 10000);
|
| 2192 |
|
| 2193 | |
| 2194 | * high watermark may be uninitialised if fragmentation occurs
|
| 2195 | * very early in boot so do not boost. We do not fall
|
| 2196 | * through and boost by pageblock_nr_pages as failing
|
| 2197 | * allocations that early means that reclaim is not going
|
| 2198 | * to help and it may even be impossible to reclaim the
|
| 2199 | * boosted watermark resulting in a hang.
|
| 2200 |
|
| 2201 | if (!max_boost)
|
| 2202 | return false;
|
| 2203 |
|
| 2204 | max_boost = max(pageblock_nr_pages, max_boost);
|
| 2205 |
|
| 2206 | zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
|
| 2207 | max_boost);
|
| 2208 |
|
| 2209 | return true;
|
| 2210 | }
|
| 2211 |
|
| 2212 | |
| 2213 | * When we are falling back to another migratetype during allocation, should we
|
| 2214 | * try to claim an entire block to satisfy further allocations, instead of
|
| 2215 | * polluting multiple pageblocks?
|
| 2216 |
|
| 2217 | static bool should_try_claim_block(unsigned int order, int start_mt)
|
| 2218 | {
|
| 2219 | |
| 2220 | * Leaving this order check is intended, although there is
|
| 2221 | * relaxed order check in next check. The reason is that
|
| 2222 | * we can actually claim the whole pageblock if this condition met,
|
| 2223 | * but, below check doesn't guarantee it and that is just heuristic
|
| 2224 | * so could be changed anytime.
|
| 2225 |
|
| 2226 | if (order >= pageblock_order)
|
| 2227 | return true;
|
| 2228 |
|
| 2229 | |
| 2230 | * Above a certain threshold, always try to claim, as it's likely there
|
| 2231 | * will be more free pages in the pageblock.
|
| 2232 |
|
| 2233 | if (order >= pageblock_order / 2)
|
| 2234 | return true;
|
| 2235 |
|
| 2236 | |
| 2237 | * Unmovable/reclaimable allocations would cause permanent
|
| 2238 | * fragmentations if they fell back to allocating from a movable block
|
| 2239 | * (polluting it), so we try to claim the whole block regardless of the
|
| 2240 | * allocation size. Later movable allocations can always steal from this
|
| 2241 | * block, which is less problematic.
|
| 2242 |
|
| 2243 | if (start_mt == MIGRATE_RECLAIMABLE || start_mt == MIGRATE_UNMOVABLE)
|
| 2244 | return true;
|
| 2245 |
|
| 2246 | if (page_group_by_mobility_disabled)
|
| 2247 | return true;
|
| 2248 |
|
| 2249 | |
| 2250 | * Movable pages won't cause permanent fragmentation, so when you alloc
|
| 2251 | * small pages, we just need to temporarily steal unmovable or
|
| 2252 | * reclaimable pages that are closest to the request size. After a
|
| 2253 | * while, memory compaction may occur to form large contiguous pages,
|
| 2254 | * and the next movable allocation may not need to steal.
|
| 2255 |
|
| 2256 | return false;
|
| 2257 | }
|
| 2258 |
|
| 2259 | |
| 2260 | * Check whether there is a suitable fallback freepage with requested order.
|
| 2261 | * If claimable is true, this function returns fallback_mt only if
|
| 2262 | * we would do this whole-block claiming. This would help to reduce
|
| 2263 | * fragmentation due to mixed migratetype pages in one pageblock.
|
| 2264 |
|
| 2265 | enum fallback_result
|
| 2266 | find_suitable_fallback(struct free_area *area, unsigned int order,
|
| 2267 | int migratetype, bool claimable, int *mt_out)
|
| 2268 | {
|
| 2269 | int i;
|
| 2270 |
|
| 2271 | if (claimable && !should_try_claim_block(order, migratetype))
|
| 2272 | return FALLBACK_NOCLAIM;
|
| 2273 |
|
| 2274 | if (area->nr_free == 0)
|
| 2275 | return FALLBACK_EMPTY;
|
| 2276 |
|
| 2277 | for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
|
| 2278 | int fallback_mt = fallbacks[migratetype][i];
|
| 2279 |
|
| 2280 | if (!free_area_empty(area, fallback_mt)) {
|
| 2281 | if (mt_out)
|
| 2282 | *mt_out = fallback_mt;
|
| 2283 | return FALLBACK_FOUND;
|
| 2284 | }
|
| 2285 | }
|
| 2286 |
|
| 2287 | return FALLBACK_EMPTY;
|
| 2288 | }
|
| 2289 |
|
| 2290 | |
| 2291 | * This function implements actual block claiming behaviour. If order is large
|
| 2292 | * enough, we can claim the whole pageblock for the requested migratetype. If
|
| 2293 | * not, we check the pageblock for constituent pages; if at least half of the
|
| 2294 | * pages are free or compatible, we can still claim the whole block, so pages
|
| 2295 | * freed in the future will be put on the correct free list.
|
| 2296 |
|
| 2297 | static struct page *
|
| 2298 | try_to_claim_block(struct zone *zone, struct page *page,
|
| 2299 | int current_order, int order, int start_type,
|
| 2300 | int block_type, unsigned int alloc_flags)
|
| 2301 | {
|
| 2302 | int free_pages, movable_pages, alike_pages;
|
| 2303 | unsigned long start_pfn;
|
| 2304 |
|
| 2305 |
|
| 2306 | if (current_order >= pageblock_order) {
|
| 2307 | unsigned int nr_added;
|
| 2308 |
|
| 2309 | del_page_from_free_list(page, zone, current_order, block_type);
|
| 2310 | change_pageblock_range(page, current_order, start_type);
|
| 2311 | nr_added = expand(zone, page, order, current_order, start_type);
|
| 2312 | account_freepages(zone, nr_added, start_type);
|
| 2313 | return page;
|
| 2314 | }
|
| 2315 |
|
| 2316 | |
| 2317 | * Boost watermarks to increase reclaim pressure to reduce the
|
| 2318 | * likelihood of future fallbacks. Wake kswapd now as the node
|
| 2319 | * may be balanced overall and kswapd will not wake naturally.
|
| 2320 |
|
| 2321 | if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
|
| 2322 | set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
|
| 2323 |
|
| 2324 |
|
| 2325 | if (!prep_move_freepages_block(zone, page, &start_pfn, &free_pages,
|
| 2326 | &movable_pages))
|
| 2327 | return NULL;
|
| 2328 |
|
| 2329 | |
| 2330 | * Determine how many pages are compatible with our allocation.
|
| 2331 | * For movable allocation, it's the number of movable pages which
|
| 2332 | * we just obtained. For other types it's a bit more tricky.
|
| 2333 |
|
| 2334 | if (start_type == MIGRATE_MOVABLE) {
|
| 2335 | alike_pages = movable_pages;
|
| 2336 | } else {
|
| 2337 | |
| 2338 | * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
|
| 2339 | * to MOVABLE pageblock, consider all non-movable pages as
|
| 2340 | * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
|
| 2341 | * vice versa, be conservative since we can't distinguish the
|
| 2342 | * exact migratetype of non-movable pages.
|
| 2343 |
|
| 2344 | if (block_type == MIGRATE_MOVABLE)
|
| 2345 | alike_pages = pageblock_nr_pages
|
| 2346 | - (free_pages + movable_pages);
|
| 2347 | else
|
| 2348 | alike_pages = 0;
|
| 2349 | }
|
| 2350 | |
| 2351 | * If a sufficient number of pages in the block are either free or of
|
| 2352 | * compatible migratability as our allocation, claim the whole block.
|
| 2353 |
|
| 2354 | if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
|
| 2355 | page_group_by_mobility_disabled) {
|
| 2356 | __move_freepages_block(zone, start_pfn, block_type, start_type);
|
| 2357 | set_pageblock_migratetype(pfn_to_page(start_pfn), start_type);
|
| 2358 | return __rmqueue_smallest(zone, order, start_type);
|
| 2359 | }
|
| 2360 |
|
| 2361 | return NULL;
|
| 2362 | }
|
| 2363 |
|
| 2364 | |
| 2365 | * Try to allocate from some fallback migratetype by claiming the entire block,
|
| 2366 | * i.e. converting it to the allocation's start migratetype.
|
| 2367 | *
|
| 2368 | * The use of signed ints for order and current_order is a deliberate
|
| 2369 | * deviation from the rest of this file, to make the for loop
|
| 2370 | * condition simpler.
|
| 2371 |
|
| 2372 | static __always_inline struct page *
|
| 2373 | __rmqueue_claim(struct zone *zone, int order, int start_migratetype,
|
| 2374 | unsigned int alloc_flags)
|
| 2375 | {
|
| 2376 | struct free_area *area;
|
| 2377 | int current_order;
|
| 2378 | int min_order = order;
|
| 2379 | struct page *page;
|
| 2380 | int fallback_mt;
|
| 2381 |
|
| 2382 | |
| 2383 | * Do not steal pages from freelists belonging to other pageblocks
|
| 2384 | * i.e. orders < pageblock_order. If there are no local zones free,
|
| 2385 | * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
|
| 2386 |
|
| 2387 | if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
|
| 2388 | min_order = pageblock_order;
|
| 2389 |
|
| 2390 | |
| 2391 | * Find the largest available free page in the other list. This roughly
|
| 2392 | * approximates finding the pageblock with the most free pages, which
|
| 2393 | * would be too costly to do exactly.
|
| 2394 |
|
| 2395 | for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
|
| 2396 | --current_order) {
|
| 2397 | enum fallback_result result;
|
| 2398 |
|
| 2399 | area = &(zone->free_area[current_order]);
|
| 2400 | result = find_suitable_fallback(area, current_order,
|
| 2401 | start_migratetype, true, &fallback_mt);
|
| 2402 |
|
| 2403 | if (result == FALLBACK_EMPTY)
|
| 2404 | continue;
|
| 2405 |
|
| 2406 | if (result == FALLBACK_NOCLAIM)
|
| 2407 | break;
|
| 2408 |
|
| 2409 | page = get_page_from_free_area(area, fallback_mt);
|
| 2410 | page = try_to_claim_block(zone, page, current_order, order,
|
| 2411 | start_migratetype, fallback_mt,
|
| 2412 | alloc_flags);
|
| 2413 | if (page) {
|
| 2414 | trace_mm_page_alloc_extfrag(page, order, current_order,
|
| 2415 | start_migratetype, fallback_mt);
|
| 2416 | return page;
|
| 2417 | }
|
| 2418 | }
|
| 2419 |
|
| 2420 | return NULL;
|
| 2421 | }
|
| 2422 |
|
| 2423 | |
| 2424 | * Try to steal a single page from some fallback migratetype. Leave the rest of
|
| 2425 | * the block as its current migratetype, potentially causing fragmentation.
|
| 2426 |
|
| 2427 | static __always_inline struct page *
|
| 2428 | __rmqueue_steal(struct zone *zone, int order, int start_migratetype)
|
| 2429 | {
|
| 2430 | struct free_area *area;
|
| 2431 | int current_order;
|
| 2432 | struct page *page;
|
| 2433 | int fallback_mt;
|
| 2434 |
|
| 2435 | for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
|
| 2436 | enum fallback_result result;
|
| 2437 |
|
| 2438 | area = &(zone->free_area[current_order]);
|
| 2439 | result = find_suitable_fallback(area, current_order, start_migratetype,
|
| 2440 | false, &fallback_mt);
|
| 2441 | if (result == FALLBACK_EMPTY)
|
| 2442 | continue;
|
| 2443 |
|
| 2444 | page = get_page_from_free_area(area, fallback_mt);
|
| 2445 | page_del_and_expand(zone, page, order, current_order, fallback_mt);
|
| 2446 | trace_mm_page_alloc_extfrag(page, order, current_order,
|
| 2447 | start_migratetype, fallback_mt);
|
| 2448 | return page;
|
| 2449 | }
|
| 2450 |
|
| 2451 | return NULL;
|
| 2452 | }
|
| 2453 |
|
| 2454 | enum rmqueue_mode {
|
| 2455 | RMQUEUE_NORMAL,
|
| 2456 | RMQUEUE_CMA,
|
| 2457 | RMQUEUE_CLAIM,
|
| 2458 | RMQUEUE_STEAL,
|
| 2459 | };
|
| 2460 |
|
| 2461 | |
| 2462 | * Do the hard work of removing an element from the buddy allocator.
|
| 2463 | * Call me with the zone->lock already held.
|
| 2464 |
|
| 2465 | static __always_inline struct page *
|
| 2466 | __rmqueue(struct zone *zone, unsigned int order, int migratetype,
|
| 2467 | unsigned int alloc_flags, enum rmqueue_mode *mode)
|
| 2468 | {
|
| 2469 | struct page *page;
|
| 2470 |
|
| 2471 | if (IS_ENABLED(CONFIG_CMA)) {
|
| 2472 | |
| 2473 | * Balance movable allocations between regular and CMA areas by
|
| 2474 | * allocating from CMA when over half of the zone's free memory
|
| 2475 | * is in the CMA area.
|
| 2476 |
|
| 2477 | if (alloc_flags & ALLOC_CMA &&
|
| 2478 | zone_page_state(zone, NR_FREE_CMA_PAGES) >
|
| 2479 | zone_page_state(zone, NR_FREE_PAGES) / 2) {
|
| 2480 | page = __rmqueue_cma_fallback(zone, order);
|
| 2481 | if (page)
|
| 2482 | return page;
|
| 2483 | }
|
| 2484 | }
|
| 2485 |
|
| 2486 | |
| 2487 | * First try the freelists of the requested migratetype, then try
|
| 2488 | * fallbacks modes with increasing levels of fragmentation risk.
|
| 2489 | *
|
| 2490 | * The fallback logic is expensive and rmqueue_bulk() calls in
|
| 2491 | * a loop with the zone->lock held, meaning the freelists are
|
| 2492 | * not subject to any outside changes. Remember in *mode where
|
| 2493 | * we found pay dirt, to save us the search on the next call.
|
| 2494 |
|
| 2495 | switch (*mode) {
|
| 2496 | case RMQUEUE_NORMAL:
|
| 2497 | page = __rmqueue_smallest(zone, order, migratetype);
|
| 2498 | if (page)
|
| 2499 | return page;
|
| 2500 | fallthrough;
|
| 2501 | case RMQUEUE_CMA:
|
| 2502 | if (alloc_flags & ALLOC_CMA) {
|
| 2503 | page = __rmqueue_cma_fallback(zone, order);
|
| 2504 | if (page) {
|
| 2505 | *mode = RMQUEUE_CMA;
|
| 2506 | return page;
|
| 2507 | }
|
| 2508 | }
|
| 2509 | fallthrough;
|
| 2510 | case RMQUEUE_CLAIM:
|
| 2511 | page = __rmqueue_claim(zone, order, migratetype, alloc_flags);
|
| 2512 | if (page) {
|
| 2513 |
|
| 2514 | *mode = RMQUEUE_NORMAL;
|
| 2515 | return page;
|
| 2516 | }
|
| 2517 | fallthrough;
|
| 2518 | case RMQUEUE_STEAL:
|
| 2519 | if (!(alloc_flags & ALLOC_NOFRAGMENT)) {
|
| 2520 | page = __rmqueue_steal(zone, order, migratetype);
|
| 2521 | if (page) {
|
| 2522 | *mode = RMQUEUE_STEAL;
|
| 2523 | return page;
|
| 2524 | }
|
| 2525 | }
|
| 2526 | }
|
| 2527 | return NULL;
|
| 2528 | }
|
| 2529 |
|
| 2530 | |
| 2531 | * Obtain a specified number of elements from the buddy allocator, all under
|
| 2532 | * a single hold of the lock, for efficiency. Add them to the supplied list.
|
| 2533 | * Returns the number of new pages which were placed at *list.
|
| 2534 |
|
| 2535 | static int rmqueue_bulk(struct zone *zone, unsigned int order,
|
| 2536 | unsigned long count, struct list_head *list,
|
| 2537 | int migratetype, unsigned int alloc_flags)
|
| 2538 | {
|
| 2539 | enum rmqueue_mode rmqm = RMQUEUE_NORMAL;
|
| 2540 | unsigned long flags;
|
| 2541 | int i;
|
| 2542 |
|
| 2543 | if (unlikely(alloc_flags & ALLOC_NOLOCK)) {
|
| 2544 | if (!spin_trylock_irqsave(&zone->lock, flags))
|
| 2545 | return 0;
|
| 2546 | } else {
|
| 2547 | spin_lock_irqsave(&zone->lock, flags);
|
| 2548 | }
|
| 2549 | for (i = 0; i < count; ++i) {
|
| 2550 | struct page *page = __rmqueue(zone, order, migratetype,
|
| 2551 | alloc_flags, &rmqm);
|
| 2552 | if (unlikely(page == NULL))
|
| 2553 | break;
|
| 2554 |
|
| 2555 | |
| 2556 | * Split buddy pages returned by expand() are received here in
|
| 2557 | * physical page order. The page is added to the tail of
|
| 2558 | * caller's list. From the callers perspective, the linked list
|
| 2559 | * is ordered by page number under some conditions. This is
|
| 2560 | * useful for IO devices that can forward direction from the
|
| 2561 | * head, thus also in the physical page order. This is useful
|
| 2562 | * for IO devices that can merge IO requests if the physical
|
| 2563 | * pages are ordered properly.
|
| 2564 |
|
| 2565 | list_add_tail(&page->pcp_list, list);
|
| 2566 | }
|
| 2567 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 2568 |
|
| 2569 | return i;
|
| 2570 | }
|
| 2571 |
|
| 2572 | |
| 2573 | * Called from the vmstat counter updater to decay the PCP high.
|
| 2574 | * Return whether there are addition works to do.
|
| 2575 |
|
| 2576 | bool decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
|
| 2577 | {
|
| 2578 | int high_min, to_drain, to_drain_batched, batch;
|
| 2579 | bool todo = false;
|
| 2580 |
|
| 2581 | high_min = READ_ONCE(pcp->high_min);
|
| 2582 | batch = READ_ONCE(pcp->batch);
|
| 2583 | |
| 2584 | * Decrease pcp->high periodically to try to free possible
|
| 2585 | * idle PCP pages. And, avoid to free too many pages to
|
| 2586 | * control latency. This caps pcp->high decrement too.
|
| 2587 |
|
| 2588 | if (pcp->high > high_min) {
|
| 2589 | pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
|
| 2590 | pcp->high - (pcp->high >> 3), high_min);
|
| 2591 | if (pcp->high > high_min)
|
| 2592 | todo = true;
|
| 2593 | }
|
| 2594 |
|
| 2595 | to_drain = pcp->count - pcp->high;
|
| 2596 | while (to_drain > 0) {
|
| 2597 | to_drain_batched = min(to_drain, batch);
|
| 2598 | pcp_spin_lock_nopin(pcp);
|
| 2599 | free_pcppages_bulk(zone, to_drain_batched, pcp, 0);
|
| 2600 | pcp_spin_unlock_nopin(pcp);
|
| 2601 | todo = true;
|
| 2602 |
|
| 2603 | to_drain -= to_drain_batched;
|
| 2604 | }
|
| 2605 |
|
| 2606 | return todo;
|
| 2607 | }
|
| 2608 |
|
| 2609 | #ifdef CONFIG_NUMA
|
| 2610 | |
| 2611 | * Called from the vmstat counter updater to drain pagesets of this
|
| 2612 | * currently executing processor on remote nodes after they have
|
| 2613 | * expired.
|
| 2614 |
|
| 2615 | void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
|
| 2616 | {
|
| 2617 | int to_drain, batch;
|
| 2618 |
|
| 2619 | batch = READ_ONCE(pcp->batch);
|
| 2620 | to_drain = min(pcp->count, batch);
|
| 2621 | if (to_drain > 0) {
|
| 2622 | pcp_spin_lock_nopin(pcp);
|
| 2623 | free_pcppages_bulk(zone, to_drain, pcp, 0);
|
| 2624 | pcp_spin_unlock_nopin(pcp);
|
| 2625 | }
|
| 2626 | }
|
| 2627 | #endif
|
| 2628 |
|
| 2629 | |
| 2630 | * Drain pcplists of the indicated processor and zone.
|
| 2631 |
|
| 2632 | static void drain_pages_zone(unsigned int cpu, struct zone *zone)
|
| 2633 | {
|
| 2634 | struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
|
| 2635 | int count;
|
| 2636 |
|
| 2637 | do {
|
| 2638 | pcp_spin_lock_nopin(pcp);
|
| 2639 | count = pcp->count;
|
| 2640 | if (count) {
|
| 2641 | int to_drain = min(count,
|
| 2642 | pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
|
| 2643 |
|
| 2644 | free_pcppages_bulk(zone, to_drain, pcp, 0);
|
| 2645 | count -= to_drain;
|
| 2646 | }
|
| 2647 | pcp_spin_unlock_nopin(pcp);
|
| 2648 | } while (count);
|
| 2649 | }
|
| 2650 |
|
| 2651 | |
| 2652 | * Drain pcplists of all zones on the indicated processor.
|
| 2653 |
|
| 2654 | static void drain_pages(unsigned int cpu)
|
| 2655 | {
|
| 2656 | struct zone *zone;
|
| 2657 |
|
| 2658 | for_each_populated_zone(zone) {
|
| 2659 | drain_pages_zone(cpu, zone);
|
| 2660 | }
|
| 2661 | }
|
| 2662 |
|
| 2663 | |
| 2664 | * Spill all of this CPU's per-cpu pages back into the buddy allocator.
|
| 2665 |
|
| 2666 | void drain_local_pages(struct zone *zone)
|
| 2667 | {
|
| 2668 | int cpu = smp_processor_id();
|
| 2669 |
|
| 2670 | if (zone)
|
| 2671 | drain_pages_zone(cpu, zone);
|
| 2672 | else
|
| 2673 | drain_pages(cpu);
|
| 2674 | }
|
| 2675 |
|
| 2676 | |
| 2677 | * The implementation of drain_all_pages(), exposing an extra parameter to
|
| 2678 | * drain on all cpus.
|
| 2679 | *
|
| 2680 | * drain_all_pages() is optimized to only execute on cpus where pcplists are
|
| 2681 | * not empty. The check for non-emptiness can however race with a free to
|
| 2682 | * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
|
| 2683 | * that need the guarantee that every CPU has drained can disable the
|
| 2684 | * optimizing racy check.
|
| 2685 |
|
| 2686 | static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
|
| 2687 | {
|
| 2688 | int cpu;
|
| 2689 |
|
| 2690 | |
| 2691 | * Allocate in the BSS so we won't require allocation in
|
| 2692 | * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
|
| 2693 |
|
| 2694 | static cpumask_t cpus_with_pcps;
|
| 2695 |
|
| 2696 | |
| 2697 | * Do not drain if one is already in progress unless it's specific to
|
| 2698 | * a zone. Such callers are primarily CMA and memory hotplug and need
|
| 2699 | * the drain to be complete when the call returns.
|
| 2700 |
|
| 2701 | if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
|
| 2702 | if (!zone)
|
| 2703 | return;
|
| 2704 | mutex_lock(&pcpu_drain_mutex);
|
| 2705 | }
|
| 2706 |
|
| 2707 | |
| 2708 | * We don't care about racing with CPU hotplug event
|
| 2709 | * as offline notification will cause the notified
|
| 2710 | * cpu to drain that CPU pcps and on_each_cpu_mask
|
| 2711 | * disables preemption as part of its processing
|
| 2712 |
|
| 2713 | for_each_online_cpu(cpu) {
|
| 2714 | struct per_cpu_pages *pcp;
|
| 2715 | struct zone *z;
|
| 2716 | bool has_pcps = false;
|
| 2717 |
|
| 2718 | if (force_all_cpus) {
|
| 2719 | |
| 2720 | * The pcp.count check is racy, some callers need a
|
| 2721 | * guarantee that no cpu is missed.
|
| 2722 |
|
| 2723 | has_pcps = true;
|
| 2724 | } else if (zone) {
|
| 2725 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
|
| 2726 | if (pcp->count)
|
| 2727 | has_pcps = true;
|
| 2728 | } else {
|
| 2729 | for_each_populated_zone(z) {
|
| 2730 | pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
|
| 2731 | if (pcp->count) {
|
| 2732 | has_pcps = true;
|
| 2733 | break;
|
| 2734 | }
|
| 2735 | }
|
| 2736 | }
|
| 2737 |
|
| 2738 | if (has_pcps)
|
| 2739 | cpumask_set_cpu(cpu, &cpus_with_pcps);
|
| 2740 | else
|
| 2741 | cpumask_clear_cpu(cpu, &cpus_with_pcps);
|
| 2742 | }
|
| 2743 |
|
| 2744 | for_each_cpu(cpu, &cpus_with_pcps) {
|
| 2745 | if (zone)
|
| 2746 | drain_pages_zone(cpu, zone);
|
| 2747 | else
|
| 2748 | drain_pages(cpu);
|
| 2749 | }
|
| 2750 |
|
| 2751 | mutex_unlock(&pcpu_drain_mutex);
|
| 2752 | }
|
| 2753 |
|
| 2754 | |
| 2755 | * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
|
| 2756 | *
|
| 2757 | * When zone parameter is non-NULL, spill just the single zone's pages.
|
| 2758 |
|
| 2759 | void drain_all_pages(struct zone *zone)
|
| 2760 | {
|
| 2761 | __drain_all_pages(zone, false);
|
| 2762 | }
|
| 2763 |
|
| 2764 | static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
|
| 2765 | {
|
| 2766 | int min_nr_free, max_nr_free;
|
| 2767 |
|
| 2768 |
|
| 2769 | if (unlikely(free_high))
|
| 2770 | return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
|
| 2771 |
|
| 2772 |
|
| 2773 | if (unlikely(high < batch))
|
| 2774 | return 1;
|
| 2775 |
|
| 2776 |
|
| 2777 | min_nr_free = batch;
|
| 2778 | max_nr_free = high - batch;
|
| 2779 |
|
| 2780 | |
| 2781 | * Increase the batch number to the number of the consecutive
|
| 2782 | * freed pages to reduce zone lock contention.
|
| 2783 |
|
| 2784 | batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
|
| 2785 |
|
| 2786 | return batch;
|
| 2787 | }
|
| 2788 |
|
| 2789 | static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
|
| 2790 | int batch, bool free_high)
|
| 2791 | {
|
| 2792 | int high, high_min, high_max;
|
| 2793 |
|
| 2794 | high_min = READ_ONCE(pcp->high_min);
|
| 2795 | high_max = READ_ONCE(pcp->high_max);
|
| 2796 | high = pcp->high = clamp(pcp->high, high_min, high_max);
|
| 2797 |
|
| 2798 | if (unlikely(!high))
|
| 2799 | return 0;
|
| 2800 |
|
| 2801 | if (unlikely(free_high)) {
|
| 2802 | pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
|
| 2803 | high_min);
|
| 2804 | return 0;
|
| 2805 | }
|
| 2806 |
|
| 2807 | |
| 2808 | * If reclaim is active, limit the number of pages that can be
|
| 2809 | * stored on pcp lists
|
| 2810 |
|
| 2811 | if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
|
| 2812 | int free_count = max_t(int, pcp->free_count, batch);
|
| 2813 |
|
| 2814 | pcp->high = max(high - free_count, high_min);
|
| 2815 | return min(batch << 2, pcp->high);
|
| 2816 | }
|
| 2817 |
|
| 2818 | if (high_min == high_max)
|
| 2819 | return high;
|
| 2820 |
|
| 2821 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
|
| 2822 | int free_count = max_t(int, pcp->free_count, batch);
|
| 2823 |
|
| 2824 | pcp->high = max(high - free_count, high_min);
|
| 2825 | high = max(pcp->count, high_min);
|
| 2826 | } else if (pcp->count >= high) {
|
| 2827 | int need_high = pcp->free_count + batch;
|
| 2828 |
|
| 2829 |
|
| 2830 | if (pcp->high < need_high)
|
| 2831 | pcp->high = clamp(need_high, high_min, high_max);
|
| 2832 | }
|
| 2833 |
|
| 2834 | return high;
|
| 2835 | }
|
| 2836 |
|
| 2837 | |
| 2838 | * Tune pcp alloc factor and adjust count & free_count. Free pages to bring the
|
| 2839 | * pcp's watermarks below high.
|
| 2840 | *
|
| 2841 | * May return a freed pcp, if during page freeing the pcp spinlock cannot be
|
| 2842 | * reacquired. Return true if pcp is locked, false otherwise.
|
| 2843 |
|
| 2844 | static bool free_frozen_page_commit(struct zone *zone,
|
| 2845 | struct per_cpu_pages *pcp, struct page *page, int migratetype,
|
| 2846 | unsigned int order, fpi_t fpi_flags)
|
| 2847 | {
|
| 2848 | int high, batch;
|
| 2849 | int to_free, to_free_batched;
|
| 2850 | int pindex;
|
| 2851 | int cpu = smp_processor_id();
|
| 2852 | int ret = true;
|
| 2853 | bool free_high = false;
|
| 2854 |
|
| 2855 | |
| 2856 | * On freeing, reduce the number of pages that are batch allocated.
|
| 2857 | * See nr_pcp_alloc() where alloc_factor is increased for subsequent
|
| 2858 | * allocations.
|
| 2859 |
|
| 2860 | pcp->alloc_factor >>= 1;
|
| 2861 | __count_vm_events(PGFREE, 1 << order);
|
| 2862 | pindex = order_to_pindex(migratetype, order);
|
| 2863 | list_add(&page->pcp_list, &pcp->lists[pindex]);
|
| 2864 | pcp->count += 1 << order;
|
| 2865 |
|
| 2866 | batch = READ_ONCE(pcp->batch);
|
| 2867 | |
| 2868 | * As high-order pages other than THP's stored on PCP can contribute
|
| 2869 | * to fragmentation, limit the number stored when PCP is heavily
|
| 2870 | * freeing without allocation. The remainder after bulk freeing
|
| 2871 | * stops will be drained from vmstat refresh context.
|
| 2872 |
|
| 2873 | if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
|
| 2874 | free_high = (pcp->free_count >= (batch + pcp->high_min / 2) &&
|
| 2875 | (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
|
| 2876 | (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
|
| 2877 | pcp->count >= batch));
|
| 2878 | pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
|
| 2879 | } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
|
| 2880 | pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
|
| 2881 | }
|
| 2882 | if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
|
| 2883 | pcp->free_count += (1 << order);
|
| 2884 |
|
| 2885 | if (unlikely(fpi_flags & FPI_NOLOCK)) {
|
| 2886 | |
| 2887 | * Do not attempt to take a zone lock. Let pcp->count get
|
| 2888 | * over high mark temporarily.
|
| 2889 |
|
| 2890 | return true;
|
| 2891 | }
|
| 2892 |
|
| 2893 | high = nr_pcp_high(pcp, zone, batch, free_high);
|
| 2894 | if (pcp->count < high)
|
| 2895 | return true;
|
| 2896 |
|
| 2897 | to_free = nr_pcp_free(pcp, batch, high, free_high);
|
| 2898 | while (to_free > 0 && pcp->count > 0) {
|
| 2899 | to_free_batched = min(to_free, batch);
|
| 2900 | free_pcppages_bulk(zone, to_free_batched, pcp, pindex);
|
| 2901 | to_free -= to_free_batched;
|
| 2902 |
|
| 2903 | if (to_free == 0 || pcp->count == 0)
|
| 2904 | break;
|
| 2905 |
|
| 2906 | pcp_spin_unlock(pcp);
|
| 2907 |
|
| 2908 | pcp = pcp_spin_trylock(zone->per_cpu_pageset);
|
| 2909 | if (!pcp) {
|
| 2910 | ret = false;
|
| 2911 | break;
|
| 2912 | }
|
| 2913 |
|
| 2914 | |
| 2915 | * Check if this thread has been migrated to a different CPU.
|
| 2916 | * If that is the case, give up and indicate that the pcp is
|
| 2917 | * returned in an unlocked state.
|
| 2918 |
|
| 2919 | if (smp_processor_id() != cpu) {
|
| 2920 | pcp_spin_unlock(pcp);
|
| 2921 | ret = false;
|
| 2922 | break;
|
| 2923 | }
|
| 2924 | }
|
| 2925 |
|
| 2926 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
|
| 2927 | zone_watermark_ok(zone, 0, high_wmark_pages(zone),
|
| 2928 | ZONE_MOVABLE, 0)) {
|
| 2929 | struct pglist_data *pgdat = zone->zone_pgdat;
|
| 2930 | clear_bit(ZONE_BELOW_HIGH, &zone->flags);
|
| 2931 |
|
| 2932 | |
| 2933 | * Assume that memory pressure on this node is gone and may be
|
| 2934 | * in a reclaimable state. If a memory fallback node exists,
|
| 2935 | * direct reclaim may not have been triggered, causing a
|
| 2936 | * 'hopeless node' to stay in that state for a while. Let
|
| 2937 | * kswapd work again by resetting kswapd_failures.
|
| 2938 |
|
| 2939 | if (kswapd_test_hopeless(pgdat) &&
|
| 2940 | next_memory_node(pgdat->node_id) < MAX_NUMNODES)
|
| 2941 | kswapd_clear_hopeless(pgdat, KSWAPD_CLEAR_HOPELESS_PCP);
|
| 2942 | }
|
| 2943 | return ret;
|
| 2944 | }
|
| 2945 |
|
| 2946 | |
| 2947 | * Free a pcp page
|
| 2948 |
|
| 2949 | static void __free_frozen_pages(struct page *page, unsigned int order,
|
| 2950 | fpi_t fpi_flags)
|
| 2951 | {
|
| 2952 | struct per_cpu_pages *pcp;
|
| 2953 | struct zone *zone;
|
| 2954 | unsigned long pfn = page_to_pfn(page);
|
| 2955 | int migratetype;
|
| 2956 |
|
| 2957 | if (!pcp_allowed_order(order)) {
|
| 2958 | __free_pages_ok(page, order, fpi_flags);
|
| 2959 | return;
|
| 2960 | }
|
| 2961 |
|
| 2962 | if (!__free_pages_prepare(page, order, fpi_flags))
|
| 2963 | return;
|
| 2964 |
|
| 2965 | |
| 2966 | * We only track unmovable, reclaimable and movable on pcp lists.
|
| 2967 | * Place ISOLATE pages on the isolated list because they are being
|
| 2968 | * offlined but treat HIGHATOMIC and CMA as movable pages so we can
|
| 2969 | * get those areas back if necessary. Otherwise, we may have to free
|
| 2970 | * excessively into the page allocator
|
| 2971 |
|
| 2972 | zone = page_zone(page);
|
| 2973 | migratetype = get_pfnblock_migratetype(page, pfn);
|
| 2974 | if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
|
| 2975 | if (unlikely(is_migrate_isolate(migratetype))) {
|
| 2976 | free_one_page(zone, page, pfn, order, fpi_flags);
|
| 2977 | return;
|
| 2978 | }
|
| 2979 | migratetype = MIGRATE_MOVABLE;
|
| 2980 | }
|
| 2981 |
|
| 2982 | if (unlikely((fpi_flags & FPI_NOLOCK) && !can_spin_trylock())) {
|
| 2983 | add_page_to_zone_llist(zone, page, order);
|
| 2984 | return;
|
| 2985 | }
|
| 2986 | pcp = pcp_spin_trylock(zone->per_cpu_pageset);
|
| 2987 | if (pcp) {
|
| 2988 | if (!free_frozen_page_commit(zone, pcp, page, migratetype,
|
| 2989 | order, fpi_flags))
|
| 2990 | return;
|
| 2991 | pcp_spin_unlock(pcp);
|
| 2992 | } else {
|
| 2993 | free_one_page(zone, page, pfn, order, fpi_flags);
|
| 2994 | }
|
| 2995 | }
|
| 2996 |
|
| 2997 | void free_frozen_pages(struct page *page, unsigned int order)
|
| 2998 | {
|
| 2999 | __free_frozen_pages(page, order, FPI_NONE);
|
| 3000 | }
|
| 3001 |
|
| 3002 | void free_frozen_pages_nolock(struct page *page, unsigned int order)
|
| 3003 | {
|
| 3004 | __free_frozen_pages(page, order, FPI_NOLOCK);
|
| 3005 | }
|
| 3006 |
|
| 3007 | |
| 3008 | * Free a batch of folios
|
| 3009 |
|
| 3010 | void free_unref_folios(struct folio_batch *folios)
|
| 3011 | {
|
| 3012 | struct per_cpu_pages *pcp = NULL;
|
| 3013 | struct zone *locked_zone = NULL;
|
| 3014 | int i, j;
|
| 3015 |
|
| 3016 |
|
| 3017 | for (i = 0, j = 0; i < folios->nr; i++) {
|
| 3018 | struct folio *folio = folios->folios[i];
|
| 3019 | unsigned long pfn = folio_pfn(folio);
|
| 3020 | unsigned int order = folio_order(folio);
|
| 3021 |
|
| 3022 | if (!__free_pages_prepare(&folio->page, order, FPI_NONE))
|
| 3023 | continue;
|
| 3024 | |
| 3025 | * Free orders not handled on the PCP directly to the
|
| 3026 | * allocator.
|
| 3027 |
|
| 3028 | if (!pcp_allowed_order(order)) {
|
| 3029 | free_one_page(folio_zone(folio), &folio->page,
|
| 3030 | pfn, order, FPI_NONE);
|
| 3031 | continue;
|
| 3032 | }
|
| 3033 | folio->private = (void *)(unsigned long)order;
|
| 3034 | if (j != i)
|
| 3035 | folios->folios[j] = folio;
|
| 3036 | j++;
|
| 3037 | }
|
| 3038 | folios->nr = j;
|
| 3039 |
|
| 3040 | for (i = 0; i < folios->nr; i++) {
|
| 3041 | struct folio *folio = folios->folios[i];
|
| 3042 | struct zone *zone = folio_zone(folio);
|
| 3043 | unsigned long pfn = folio_pfn(folio);
|
| 3044 | unsigned int order = (unsigned long)folio->private;
|
| 3045 | int migratetype;
|
| 3046 |
|
| 3047 | folio->private = NULL;
|
| 3048 | migratetype = get_pfnblock_migratetype(&folio->page, pfn);
|
| 3049 |
|
| 3050 |
|
| 3051 | if (zone != locked_zone ||
|
| 3052 | is_migrate_isolate(migratetype)) {
|
| 3053 | if (pcp) {
|
| 3054 | pcp_spin_unlock(pcp);
|
| 3055 | locked_zone = NULL;
|
| 3056 | pcp = NULL;
|
| 3057 | }
|
| 3058 |
|
| 3059 | |
| 3060 | * Free isolated pages directly to the
|
| 3061 | * allocator, see comment in free_frozen_pages.
|
| 3062 |
|
| 3063 | if (is_migrate_isolate(migratetype)) {
|
| 3064 | free_one_page(zone, &folio->page, pfn,
|
| 3065 | order, FPI_NONE);
|
| 3066 | continue;
|
| 3067 | }
|
| 3068 |
|
| 3069 | |
| 3070 | * trylock is necessary as folios may be getting freed
|
| 3071 | * from IRQ or SoftIRQ context after an IO completion.
|
| 3072 |
|
| 3073 | pcp = pcp_spin_trylock(zone->per_cpu_pageset);
|
| 3074 | if (unlikely(!pcp)) {
|
| 3075 | free_one_page(zone, &folio->page, pfn,
|
| 3076 | order, FPI_NONE);
|
| 3077 | continue;
|
| 3078 | }
|
| 3079 | locked_zone = zone;
|
| 3080 | }
|
| 3081 |
|
| 3082 | |
| 3083 | * Non-isolated types over MIGRATE_PCPTYPES get added
|
| 3084 | * to the MIGRATE_MOVABLE pcp list.
|
| 3085 |
|
| 3086 | if (unlikely(migratetype >= MIGRATE_PCPTYPES))
|
| 3087 | migratetype = MIGRATE_MOVABLE;
|
| 3088 |
|
| 3089 | trace_mm_page_free_batched(&folio->page);
|
| 3090 | if (!free_frozen_page_commit(zone, pcp, &folio->page,
|
| 3091 | migratetype, order, FPI_NONE)) {
|
| 3092 | pcp = NULL;
|
| 3093 | locked_zone = NULL;
|
| 3094 | }
|
| 3095 | }
|
| 3096 |
|
| 3097 | if (pcp)
|
| 3098 | pcp_spin_unlock(pcp);
|
| 3099 | folio_batch_reinit(folios);
|
| 3100 | }
|
| 3101 |
|
| 3102 | static void __split_page(struct page *page, unsigned int order)
|
| 3103 | {
|
| 3104 | VM_WARN_ON_PAGE(PageCompound(page), page);
|
| 3105 |
|
| 3106 | split_page_owner(page, order, 0);
|
| 3107 | pgalloc_tag_split(page_folio(page), order, 0);
|
| 3108 | split_page_memcg(page, order);
|
| 3109 | }
|
| 3110 |
|
| 3111 | |
| 3112 | * split_page takes a non-compound higher-order page, and splits it into
|
| 3113 | * n (1<<order) sub-pages: page[0..n]
|
| 3114 | * Each sub-page must be freed individually.
|
| 3115 | *
|
| 3116 | * Note: this is probably too low level an operation for use in drivers.
|
| 3117 | * Please consult with lkml before using this in your driver.
|
| 3118 |
|
| 3119 | void split_page(struct page *page, unsigned int order)
|
| 3120 | {
|
| 3121 | int i;
|
| 3122 |
|
| 3123 | VM_WARN_ON_PAGE(!page_count(page), page);
|
| 3124 |
|
| 3125 | for (i = 1; i < (1 << order); i++)
|
| 3126 | set_page_refcounted(page + i);
|
| 3127 |
|
| 3128 | __split_page(page, order);
|
| 3129 | }
|
| 3130 | EXPORT_SYMBOL_GPL(split_page);
|
| 3131 |
|
| 3132 | int __isolate_free_page(struct page *page, unsigned int order)
|
| 3133 | {
|
| 3134 | struct zone *zone = page_zone(page);
|
| 3135 | int mt = get_pageblock_migratetype(page);
|
| 3136 |
|
| 3137 | if (!is_migrate_isolate(mt)) {
|
| 3138 | unsigned long watermark;
|
| 3139 | |
| 3140 | * Obey watermarks as if the page was being allocated. We can
|
| 3141 | * emulate a high-order watermark check with a raised order-0
|
| 3142 | * watermark, because we already know our high-order page
|
| 3143 | * exists.
|
| 3144 |
|
| 3145 | watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
|
| 3146 | if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
|
| 3147 | return 0;
|
| 3148 | }
|
| 3149 |
|
| 3150 | del_page_from_free_list(page, zone, order, mt);
|
| 3151 |
|
| 3152 | |
| 3153 | * Set the pageblock if the isolated page is at least half of a
|
| 3154 | * pageblock
|
| 3155 |
|
| 3156 | if (order >= pageblock_order - 1) {
|
| 3157 | struct page *endpage = page + (1 << order) - 1;
|
| 3158 | for (; page < endpage; page += pageblock_nr_pages) {
|
| 3159 | int mt = get_pageblock_migratetype(page);
|
| 3160 | |
| 3161 | * Only change normal pageblocks (i.e., they can merge
|
| 3162 | * with others)
|
| 3163 |
|
| 3164 | if (migratetype_is_mergeable(mt))
|
| 3165 | move_freepages_block(zone, page, mt,
|
| 3166 | MIGRATE_MOVABLE);
|
| 3167 | }
|
| 3168 | }
|
| 3169 |
|
| 3170 | return 1UL << order;
|
| 3171 | }
|
| 3172 |
|
| 3173 | |
| 3174 | * __putback_isolated_page - Return a now-isolated page back where we got it
|
| 3175 | * @page: Page that was isolated
|
| 3176 | * @order: Order of the isolated page
|
| 3177 | * @mt: The page's pageblock's migratetype
|
| 3178 | *
|
| 3179 | * This function is meant to return a page pulled from the free lists via
|
| 3180 | * __isolate_free_page back to the free lists they were pulled from.
|
| 3181 |
|
| 3182 | void __putback_isolated_page(struct page *page, unsigned int order, int mt)
|
| 3183 | {
|
| 3184 | struct zone *zone = page_zone(page);
|
| 3185 |
|
| 3186 |
|
| 3187 | lockdep_assert_held(&zone->lock);
|
| 3188 |
|
| 3189 |
|
| 3190 | __free_one_page(page, page_to_pfn(page), zone, order, mt,
|
| 3191 | FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
|
| 3192 | }
|
| 3193 |
|
| 3194 | |
| 3195 | * Update NUMA hit/miss statistics
|
| 3196 |
|
| 3197 | static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
|
| 3198 | long nr_account)
|
| 3199 | {
|
| 3200 | #ifdef CONFIG_NUMA
|
| 3201 | enum numa_stat_item local_stat = NUMA_LOCAL;
|
| 3202 |
|
| 3203 |
|
| 3204 | if (!static_branch_likely(&vm_numa_stat_key))
|
| 3205 | return;
|
| 3206 |
|
| 3207 | if (zone_to_nid(z) != numa_node_id())
|
| 3208 | local_stat = NUMA_OTHER;
|
| 3209 |
|
| 3210 | if (zone_to_nid(z) == zone_to_nid(preferred_zone))
|
| 3211 | __count_numa_events(z, NUMA_HIT, nr_account);
|
| 3212 | else {
|
| 3213 | __count_numa_events(z, NUMA_MISS, nr_account);
|
| 3214 | __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
|
| 3215 | }
|
| 3216 | __count_numa_events(z, local_stat, nr_account);
|
| 3217 | #endif
|
| 3218 | }
|
| 3219 |
|
| 3220 | static __always_inline
|
| 3221 | struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
|
| 3222 | unsigned int order, unsigned int alloc_flags,
|
| 3223 | int migratetype)
|
| 3224 | {
|
| 3225 | struct page *page;
|
| 3226 | unsigned long flags;
|
| 3227 |
|
| 3228 | do {
|
| 3229 | page = NULL;
|
| 3230 | if (unlikely(alloc_flags & ALLOC_NOLOCK)) {
|
| 3231 | if (!spin_trylock_irqsave(&zone->lock, flags))
|
| 3232 | return NULL;
|
| 3233 | } else {
|
| 3234 | spin_lock_irqsave(&zone->lock, flags);
|
| 3235 | }
|
| 3236 | if (alloc_flags & ALLOC_HIGHATOMIC)
|
| 3237 | page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
|
| 3238 | if (!page) {
|
| 3239 | enum rmqueue_mode rmqm = RMQUEUE_NORMAL;
|
| 3240 |
|
| 3241 | page = __rmqueue(zone, order, migratetype, alloc_flags, &rmqm);
|
| 3242 |
|
| 3243 | |
| 3244 | * If the allocation fails, allow OOM handling and
|
| 3245 | * order-0 (atomic) allocs access to HIGHATOMIC
|
| 3246 | * reserves as failing now is worse than failing a
|
| 3247 | * high-order atomic allocation in the future.
|
| 3248 |
|
| 3249 | if (!page && (alloc_flags & (ALLOC_OOM|ALLOC_NON_BLOCK)))
|
| 3250 | page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
|
| 3251 |
|
| 3252 | if (!page) {
|
| 3253 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 3254 | return NULL;
|
| 3255 | }
|
| 3256 | }
|
| 3257 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 3258 | } while (check_new_pages(page, order));
|
| 3259 |
|
| 3260 | |
| 3261 | * Slowpath (precarious) high-atomic allocations may reserve
|
| 3262 | * a pageblock for future use.
|
| 3263 |
|
| 3264 | if (unlikely((alloc_flags & ALLOC_HIGHATOMIC) &&
|
| 3265 | ((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN)))
|
| 3266 | reserve_highatomic_pageblock(page, order, zone);
|
| 3267 |
|
| 3268 | __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
|
| 3269 | zone_statistics(preferred_zone, zone, 1);
|
| 3270 |
|
| 3271 | return page;
|
| 3272 | }
|
| 3273 |
|
| 3274 | static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
|
| 3275 | {
|
| 3276 | int high, base_batch, batch, max_nr_alloc;
|
| 3277 | int high_max, high_min;
|
| 3278 |
|
| 3279 | base_batch = READ_ONCE(pcp->batch);
|
| 3280 | high_min = READ_ONCE(pcp->high_min);
|
| 3281 | high_max = READ_ONCE(pcp->high_max);
|
| 3282 | high = pcp->high = clamp(pcp->high, high_min, high_max);
|
| 3283 |
|
| 3284 |
|
| 3285 | if (unlikely(high < base_batch))
|
| 3286 | return 1;
|
| 3287 |
|
| 3288 | if (order)
|
| 3289 | batch = base_batch;
|
| 3290 | else
|
| 3291 | batch = (base_batch << pcp->alloc_factor);
|
| 3292 |
|
| 3293 | |
| 3294 | * If we had larger pcp->high, we could avoid to allocate from
|
| 3295 | * zone.
|
| 3296 |
|
| 3297 | if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
|
| 3298 | high = pcp->high = min(high + batch, high_max);
|
| 3299 |
|
| 3300 | if (!order) {
|
| 3301 | max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
|
| 3302 | |
| 3303 | * Double the number of pages allocated each time there is
|
| 3304 | * subsequent allocation of order-0 pages without any freeing.
|
| 3305 |
|
| 3306 | if (batch <= max_nr_alloc &&
|
| 3307 | pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
|
| 3308 | pcp->alloc_factor++;
|
| 3309 | batch = min(batch, max_nr_alloc);
|
| 3310 | }
|
| 3311 |
|
| 3312 | |
| 3313 | * Scale batch relative to order if batch implies free pages
|
| 3314 | * can be stored on the PCP. Batch can be 1 for small zones or
|
| 3315 | * for boot pagesets which should never store free pages as
|
| 3316 | * the pages may belong to arbitrary zones.
|
| 3317 |
|
| 3318 | if (batch > 1)
|
| 3319 | batch = max(batch >> order, 2);
|
| 3320 |
|
| 3321 | return batch;
|
| 3322 | }
|
| 3323 |
|
| 3324 |
|
| 3325 | static inline
|
| 3326 | struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
|
| 3327 | int migratetype,
|
| 3328 | unsigned int alloc_flags,
|
| 3329 | struct per_cpu_pages *pcp,
|
| 3330 | struct list_head *list)
|
| 3331 | {
|
| 3332 | struct page *page;
|
| 3333 |
|
| 3334 | do {
|
| 3335 | if (list_empty(list)) {
|
| 3336 | int batch = nr_pcp_alloc(pcp, zone, order);
|
| 3337 | int alloced;
|
| 3338 |
|
| 3339 | |
| 3340 | * Don't refill the list for a higher order atomic
|
| 3341 | * allocation under memory pressure, as this would
|
| 3342 | * not build up any HIGHATOMIC reserves, which
|
| 3343 | * might be needed soon.
|
| 3344 | *
|
| 3345 | * Instead, direct it towards the reserves by
|
| 3346 | * returning NULL, which will make the caller fall
|
| 3347 | * back to rmqueue_buddy. This will try to use the
|
| 3348 | * reserves first and grow them if needed.
|
| 3349 |
|
| 3350 | if (alloc_flags & ALLOC_HIGHATOMIC)
|
| 3351 | return NULL;
|
| 3352 |
|
| 3353 | alloced = rmqueue_bulk(zone, order,
|
| 3354 | batch, list,
|
| 3355 | migratetype, alloc_flags);
|
| 3356 |
|
| 3357 | pcp->count += alloced << order;
|
| 3358 | if (unlikely(list_empty(list)))
|
| 3359 | return NULL;
|
| 3360 | }
|
| 3361 |
|
| 3362 | page = list_first_entry(list, struct page, pcp_list);
|
| 3363 | list_del(&page->pcp_list);
|
| 3364 | pcp->count -= 1 << order;
|
| 3365 | } while (check_new_pages(page, order));
|
| 3366 |
|
| 3367 | return page;
|
| 3368 | }
|
| 3369 |
|
| 3370 |
|
| 3371 | static struct page *rmqueue_pcplist(struct zone *preferred_zone,
|
| 3372 | struct zone *zone, unsigned int order,
|
| 3373 | int migratetype, unsigned int alloc_flags)
|
| 3374 | {
|
| 3375 | struct per_cpu_pages *pcp;
|
| 3376 | struct list_head *list;
|
| 3377 | struct page *page;
|
| 3378 |
|
| 3379 |
|
| 3380 | pcp = pcp_spin_trylock(zone->per_cpu_pageset);
|
| 3381 | if (!pcp)
|
| 3382 | return NULL;
|
| 3383 |
|
| 3384 | |
| 3385 | * On allocation, reduce the number of pages that are batch freed.
|
| 3386 | * See nr_pcp_free() where free_factor is increased for subsequent
|
| 3387 | * frees.
|
| 3388 |
|
| 3389 | pcp->free_count >>= 1;
|
| 3390 | list = &pcp->lists[order_to_pindex(migratetype, order)];
|
| 3391 | page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
|
| 3392 | pcp_spin_unlock(pcp);
|
| 3393 | if (page) {
|
| 3394 | __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
|
| 3395 | zone_statistics(preferred_zone, zone, 1);
|
| 3396 | }
|
| 3397 | return page;
|
| 3398 | }
|
| 3399 |
|
| 3400 | |
| 3401 | * Allocate a page from the given zone.
|
| 3402 | * Use pcplists for THP or "cheap" high-order allocations.
|
| 3403 |
|
| 3404 |
|
| 3405 | |
| 3406 | * Do not instrument rmqueue() with KMSAN. This function may call
|
| 3407 | * __msan_poison_alloca() through a call to set_pfnblock_migratetype().
|
| 3408 | * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
|
| 3409 | * may call rmqueue() again, which will result in a deadlock.
|
| 3410 |
|
| 3411 | __no_sanitize_memory
|
| 3412 | static inline
|
| 3413 | struct page *rmqueue(struct zone *preferred_zone,
|
| 3414 | struct zone *zone, unsigned int order,
|
| 3415 | gfp_t gfp_flags, unsigned int alloc_flags,
|
| 3416 | int migratetype)
|
| 3417 | {
|
| 3418 | struct page *page;
|
| 3419 |
|
| 3420 | if (likely(pcp_allowed_order(order))) {
|
| 3421 | page = rmqueue_pcplist(preferred_zone, zone, order,
|
| 3422 | migratetype, alloc_flags);
|
| 3423 | if (likely(page))
|
| 3424 | goto out;
|
| 3425 | }
|
| 3426 |
|
| 3427 | page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
|
| 3428 | migratetype);
|
| 3429 |
|
| 3430 | out:
|
| 3431 |
|
| 3432 | if ((alloc_flags & ALLOC_KSWAPD) &&
|
| 3433 | unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
|
| 3434 | clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
|
| 3435 | wakeup_kswapd(zone, 0, 0, zone_idx(zone));
|
| 3436 | }
|
| 3437 |
|
| 3438 | VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
|
| 3439 | return page;
|
| 3440 | }
|
| 3441 |
|
| 3442 | |
| 3443 | * Reserve the pageblock(s) surrounding an allocation request for
|
| 3444 | * exclusive use of high-order atomic allocations if there are no
|
| 3445 | * empty page blocks that contain a page with a suitable order
|
| 3446 |
|
| 3447 | static void reserve_highatomic_pageblock(struct page *page, int order,
|
| 3448 | struct zone *zone)
|
| 3449 | {
|
| 3450 | int mt;
|
| 3451 | unsigned long max_managed;
|
| 3452 |
|
| 3453 | |
| 3454 | * The number reserved as: minimum is 1 pageblock, maximum is
|
| 3455 | * roughly 1% of a zone. But if 1% of a zone falls below a
|
| 3456 | * pageblock size, then don't reserve any pageblocks.
|
| 3457 | * Check is race-prone but harmless.
|
| 3458 |
|
| 3459 | if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
|
| 3460 | return;
|
| 3461 | max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
|
| 3462 | if (zone->nr_reserved_highatomic >= max_managed)
|
| 3463 | return;
|
| 3464 |
|
| 3465 | guard(spinlock_irqsave)(&zone->lock);
|
| 3466 |
|
| 3467 |
|
| 3468 | if (zone->nr_reserved_highatomic >= max_managed)
|
| 3469 | return;
|
| 3470 |
|
| 3471 |
|
| 3472 | mt = get_pageblock_migratetype(page);
|
| 3473 |
|
| 3474 | if (!migratetype_is_mergeable(mt))
|
| 3475 | return;
|
| 3476 |
|
| 3477 | if (order < pageblock_order) {
|
| 3478 | if (move_freepages_block(zone, page, mt, MIGRATE_HIGHATOMIC) == -1)
|
| 3479 | return;
|
| 3480 | zone->nr_reserved_highatomic += pageblock_nr_pages;
|
| 3481 | } else {
|
| 3482 | change_pageblock_range(page, order, MIGRATE_HIGHATOMIC);
|
| 3483 | zone->nr_reserved_highatomic += 1 << order;
|
| 3484 | }
|
| 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 | struct zoneref *z;
|
| 3501 | struct zone *zone;
|
| 3502 | struct page *page;
|
| 3503 | int order;
|
| 3504 | int ret;
|
| 3505 |
|
| 3506 | for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
|
| 3507 | ac->nodemask) {
|
| 3508 | |
| 3509 | * Preserve at least one pageblock unless memory pressure
|
| 3510 | * is really high.
|
| 3511 |
|
| 3512 | if (!force && zone->nr_reserved_highatomic <=
|
| 3513 | pageblock_nr_pages)
|
| 3514 | continue;
|
| 3515 |
|
| 3516 | guard(spinlock_irqsave)(&zone->lock);
|
| 3517 | for (order = 0; order < NR_PAGE_ORDERS; order++) {
|
| 3518 | struct free_area *area = &(zone->free_area[order]);
|
| 3519 | unsigned long size;
|
| 3520 |
|
| 3521 | page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
|
| 3522 | if (!page)
|
| 3523 | continue;
|
| 3524 |
|
| 3525 | size = max(pageblock_nr_pages, 1UL << order);
|
| 3526 | |
| 3527 | * It should never happen but changes to
|
| 3528 | * locking could inadvertently allow a per-cpu
|
| 3529 | * drain to add pages to MIGRATE_HIGHATOMIC
|
| 3530 | * while unreserving so be safe and watch for
|
| 3531 | * underflows.
|
| 3532 |
|
| 3533 | if (WARN_ON_ONCE(size > zone->nr_reserved_highatomic))
|
| 3534 | size = zone->nr_reserved_highatomic;
|
| 3535 | zone->nr_reserved_highatomic -= size;
|
| 3536 |
|
| 3537 | |
| 3538 | * Convert to ac->migratetype and avoid the normal
|
| 3539 | * pageblock stealing heuristics. Minimally, the caller
|
| 3540 | * is doing the work and needs the pages. More
|
| 3541 | * importantly, if the block was always converted to
|
| 3542 | * MIGRATE_UNMOVABLE or another type then the number
|
| 3543 | * of pageblocks that cannot be completely freed
|
| 3544 | * may increase.
|
| 3545 |
|
| 3546 | if (order < pageblock_order)
|
| 3547 | ret = move_freepages_block(zone, page,
|
| 3548 | MIGRATE_HIGHATOMIC,
|
| 3549 | ac->migratetype);
|
| 3550 | else {
|
| 3551 | move_to_free_list(page, zone, order,
|
| 3552 | MIGRATE_HIGHATOMIC,
|
| 3553 | ac->migratetype);
|
| 3554 | change_pageblock_range(page, order,
|
| 3555 | ac->migratetype);
|
| 3556 | ret = 1;
|
| 3557 | }
|
| 3558 | |
| 3559 | * Reserving the block(s) already succeeded,
|
| 3560 | * so this should not fail on zone boundaries.
|
| 3561 |
|
| 3562 | WARN_ON_ONCE(ret == -1);
|
| 3563 | if (ret > 0)
|
| 3564 | return ret;
|
| 3565 | }
|
| 3566 | }
|
| 3567 |
|
| 3568 | return false;
|
| 3569 | }
|
| 3570 |
|
| 3571 | static inline long __zone_watermark_unusable_free(struct zone *z,
|
| 3572 | unsigned int order, unsigned int alloc_flags)
|
| 3573 | {
|
| 3574 | long unusable_free = (1 << order) - 1;
|
| 3575 |
|
| 3576 | |
| 3577 | * If the caller does not have rights to reserves below the min
|
| 3578 | * watermark then subtract the free pages reserved for highatomic.
|
| 3579 |
|
| 3580 | if (likely(!(alloc_flags & ALLOC_RESERVES)))
|
| 3581 | unusable_free += READ_ONCE(z->nr_free_highatomic);
|
| 3582 |
|
| 3583 | #ifdef CONFIG_CMA
|
| 3584 |
|
| 3585 | if (!(alloc_flags & ALLOC_CMA))
|
| 3586 | unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
|
| 3587 | #endif
|
| 3588 |
|
| 3589 | return unusable_free;
|
| 3590 | }
|
| 3591 |
|
| 3592 | |
| 3593 | * Return true if free base pages are above 'mark'. For high-order checks it
|
| 3594 | * will return true of the order-0 watermark is reached and there is at least
|
| 3595 | * one free page of a suitable size. Checking now avoids taking the zone lock
|
| 3596 | * to check in the allocation paths if no pages are free.
|
| 3597 |
|
| 3598 | bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
|
| 3599 | int highest_zoneidx, unsigned int alloc_flags,
|
| 3600 | long free_pages)
|
| 3601 | {
|
| 3602 | long min = mark;
|
| 3603 | int o;
|
| 3604 |
|
| 3605 |
|
| 3606 | free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
|
| 3607 |
|
| 3608 | if (unlikely(alloc_flags & ALLOC_RESERVES)) {
|
| 3609 | |
| 3610 | * __GFP_HIGH allows access to 50% of the min reserve as well
|
| 3611 | * as OOM.
|
| 3612 |
|
| 3613 | if (alloc_flags & ALLOC_MIN_RESERVE) {
|
| 3614 | min -= min / 2;
|
| 3615 |
|
| 3616 | |
| 3617 | * Non-blocking allocations (e.g. GFP_ATOMIC) can
|
| 3618 | * access more reserves than just __GFP_HIGH. Other
|
| 3619 | * non-blocking allocations requests such as GFP_NOWAIT
|
| 3620 | * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
|
| 3621 | * access to the min reserve.
|
| 3622 |
|
| 3623 | if (alloc_flags & ALLOC_NON_BLOCK)
|
| 3624 | min -= min / 4;
|
| 3625 | }
|
| 3626 |
|
| 3627 | |
| 3628 | * OOM victims can try even harder than the normal reserve
|
| 3629 | * users on the grounds that it's definitely going to be in
|
| 3630 | * the exit path shortly and free memory. Any allocation it
|
| 3631 | * makes during the free path will be small and short-lived.
|
| 3632 |
|
| 3633 | if (alloc_flags & ALLOC_OOM)
|
| 3634 | min -= min / 2;
|
| 3635 | }
|
| 3636 |
|
| 3637 | |
| 3638 | * Check watermarks for an order-0 allocation request. If these
|
| 3639 | * are not met, then a high-order request also cannot go ahead
|
| 3640 | * even if a suitable page happened to be free.
|
| 3641 |
|
| 3642 | if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
|
| 3643 | return false;
|
| 3644 |
|
| 3645 |
|
| 3646 | if (!order)
|
| 3647 | return true;
|
| 3648 |
|
| 3649 |
|
| 3650 | for (o = order; o < NR_PAGE_ORDERS; o++) {
|
| 3651 | struct free_area *area = &z->free_area[o];
|
| 3652 | int mt;
|
| 3653 |
|
| 3654 | if (!area->nr_free)
|
| 3655 | continue;
|
| 3656 |
|
| 3657 | for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
|
| 3658 | if (!free_area_empty(area, mt))
|
| 3659 | return true;
|
| 3660 | }
|
| 3661 |
|
| 3662 | #ifdef CONFIG_CMA
|
| 3663 | if ((alloc_flags & ALLOC_CMA) &&
|
| 3664 | !free_area_empty(area, MIGRATE_CMA)) {
|
| 3665 | return true;
|
| 3666 | }
|
| 3667 | #endif
|
| 3668 | if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
|
| 3669 | !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
|
| 3670 | return true;
|
| 3671 | }
|
| 3672 | }
|
| 3673 | return false;
|
| 3674 | }
|
| 3675 |
|
| 3676 | bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
|
| 3677 | int highest_zoneidx, unsigned int alloc_flags)
|
| 3678 | {
|
| 3679 | return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
|
| 3680 | zone_page_state(z, NR_FREE_PAGES));
|
| 3681 | }
|
| 3682 |
|
| 3683 | static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
|
| 3684 | unsigned long mark, int highest_zoneidx,
|
| 3685 | unsigned int alloc_flags, gfp_t gfp_mask)
|
| 3686 | {
|
| 3687 | long free_pages;
|
| 3688 |
|
| 3689 | free_pages = zone_page_state(z, NR_FREE_PAGES);
|
| 3690 |
|
| 3691 | |
| 3692 | * Fast check for order-0 only. If this fails then the reserves
|
| 3693 | * need to be calculated.
|
| 3694 |
|
| 3695 | if (!order) {
|
| 3696 | long usable_free;
|
| 3697 | long reserved;
|
| 3698 |
|
| 3699 | usable_free = free_pages;
|
| 3700 | reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
|
| 3701 |
|
| 3702 |
|
| 3703 | usable_free -= min(usable_free, reserved);
|
| 3704 | if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
|
| 3705 | return true;
|
| 3706 | }
|
| 3707 |
|
| 3708 | if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
|
| 3709 | free_pages))
|
| 3710 | return true;
|
| 3711 |
|
| 3712 | |
| 3713 | * Ignore watermark boosting for __GFP_HIGH order-0 allocations
|
| 3714 | * when checking the min watermark. The min watermark is the
|
| 3715 | * point where boosting is ignored so that kswapd is woken up
|
| 3716 | * when below the low watermark.
|
| 3717 |
|
| 3718 | if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
|
| 3719 | && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
|
| 3720 | mark = z->_watermark[WMARK_MIN];
|
| 3721 | return __zone_watermark_ok(z, order, mark, highest_zoneidx,
|
| 3722 | alloc_flags, free_pages);
|
| 3723 | }
|
| 3724 |
|
| 3725 | return false;
|
| 3726 | }
|
| 3727 |
|
| 3728 | #ifdef CONFIG_NUMA
|
| 3729 | int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
|
| 3730 |
|
| 3731 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
|
| 3732 | {
|
| 3733 | return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
|
| 3734 | node_reclaim_distance;
|
| 3735 | }
|
| 3736 | #else
|
| 3737 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
|
| 3738 | {
|
| 3739 | return true;
|
| 3740 | }
|
| 3741 | #endif
|
| 3742 |
|
| 3743 | |
| 3744 | * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
|
| 3745 | * fragmentation is subtle. If the preferred zone was HIGHMEM then
|
| 3746 | * premature use of a lower zone may cause lowmem pressure problems that
|
| 3747 | * are worse than fragmentation. If the next zone is ZONE_DMA then it is
|
| 3748 | * probably too small. It only makes sense to spread allocations to avoid
|
| 3749 | * fragmentation between the Normal and DMA32 zones.
|
| 3750 |
|
| 3751 | static inline unsigned int
|
| 3752 | alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
|
| 3753 | {
|
| 3754 | unsigned int alloc_flags = 0;
|
| 3755 |
|
| 3756 | if (gfp_mask & __GFP_KSWAPD_RECLAIM)
|
| 3757 | alloc_flags |= ALLOC_KSWAPD;
|
| 3758 |
|
| 3759 | if (defrag_mode) {
|
| 3760 | alloc_flags |= ALLOC_NOFRAGMENT;
|
| 3761 | return alloc_flags;
|
| 3762 | }
|
| 3763 |
|
| 3764 | #ifdef CONFIG_ZONE_DMA32
|
| 3765 | if (!zone)
|
| 3766 | return alloc_flags;
|
| 3767 |
|
| 3768 | if (zone_idx(zone) != ZONE_NORMAL)
|
| 3769 | return alloc_flags;
|
| 3770 |
|
| 3771 | |
| 3772 | * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
|
| 3773 | * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
|
| 3774 | * on UMA that if Normal is populated then so is DMA32.
|
| 3775 |
|
| 3776 | BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
|
| 3777 | if (nr_online_nodes > 1 && !populated_zone(--zone))
|
| 3778 | return alloc_flags;
|
| 3779 |
|
| 3780 | alloc_flags |= ALLOC_NOFRAGMENT;
|
| 3781 | #endif
|
| 3782 | return alloc_flags;
|
| 3783 | }
|
| 3784 |
|
| 3785 |
|
| 3786 | static inline unsigned int alloc_flags_cma(gfp_t gfp_mask)
|
| 3787 | {
|
| 3788 | #ifdef CONFIG_CMA
|
| 3789 | if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
|
| 3790 | return ALLOC_CMA;
|
| 3791 | #endif
|
| 3792 | return ALLOC_DEFAULT;
|
| 3793 | }
|
| 3794 |
|
| 3795 | |
| 3796 | * get_page_from_freelist goes through the zonelist trying to allocate
|
| 3797 | * a page.
|
| 3798 |
|
| 3799 | static struct page *
|
| 3800 | get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
|
| 3801 | const struct alloc_context *ac)
|
| 3802 | {
|
| 3803 | struct zoneref *z;
|
| 3804 | struct zone *zone;
|
| 3805 | struct pglist_data *last_pgdat = NULL;
|
| 3806 | bool last_pgdat_dirty_ok = false;
|
| 3807 | bool no_fallback;
|
| 3808 | bool skip_kswapd_nodes = nr_online_nodes > 1;
|
| 3809 | bool skipped_kswapd_nodes = false;
|
| 3810 |
|
| 3811 | retry:
|
| 3812 | |
| 3813 | * Scan zonelist, looking for a zone with enough free.
|
| 3814 | * See also cpuset_current_node_allowed() comment in kernel/cgroup/cpuset.c.
|
| 3815 |
|
| 3816 | no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
|
| 3817 | z = ac->preferred_zoneref;
|
| 3818 | for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
|
| 3819 | ac->nodemask) {
|
| 3820 | struct page *page;
|
| 3821 | unsigned long mark;
|
| 3822 |
|
| 3823 | if (cpusets_enabled() &&
|
| 3824 | (alloc_flags & ALLOC_CPUSET) &&
|
| 3825 | !__cpuset_zone_allowed(zone, gfp_mask))
|
| 3826 | continue;
|
| 3827 | |
| 3828 | * When allocating a page cache page for writing, we
|
| 3829 | * want to get it from a node that is within its dirty
|
| 3830 | * limit, such that no single node holds more than its
|
| 3831 | * proportional share of globally allowed dirty pages.
|
| 3832 | * The dirty limits take into account the node's
|
| 3833 | * lowmem reserves and high watermark so that kswapd
|
| 3834 | * should be able to balance it without having to
|
| 3835 | * write pages from its LRU list.
|
| 3836 | *
|
| 3837 | * XXX: For now, allow allocations to potentially
|
| 3838 | * exceed the per-node dirty limit in the slowpath
|
| 3839 | * (spread_dirty_pages unset) before going into reclaim,
|
| 3840 | * which is important when on a NUMA setup the allowed
|
| 3841 | * nodes are together not big enough to reach the
|
| 3842 | * global limit. The proper fix for these situations
|
| 3843 | * will require awareness of nodes in the
|
| 3844 | * dirty-throttling and the flusher threads.
|
| 3845 |
|
| 3846 | if (ac->spread_dirty_pages) {
|
| 3847 | if (last_pgdat != zone->zone_pgdat) {
|
| 3848 | last_pgdat = zone->zone_pgdat;
|
| 3849 | last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
|
| 3850 | }
|
| 3851 |
|
| 3852 | if (!last_pgdat_dirty_ok)
|
| 3853 | continue;
|
| 3854 | }
|
| 3855 |
|
| 3856 | if (no_fallback && !defrag_mode && nr_online_nodes > 1 &&
|
| 3857 | zone != zonelist_zone(ac->preferred_zoneref)) {
|
| 3858 | int local_nid;
|
| 3859 |
|
| 3860 | |
| 3861 | * If moving to a remote node, retry but allow
|
| 3862 | * fragmenting fallbacks. Locality is more important
|
| 3863 | * than fragmentation avoidance.
|
| 3864 |
|
| 3865 | local_nid = zonelist_node_idx(ac->preferred_zoneref);
|
| 3866 | if (zone_to_nid(zone) != local_nid) {
|
| 3867 | alloc_flags &= ~ALLOC_NOFRAGMENT;
|
| 3868 | goto retry;
|
| 3869 | }
|
| 3870 | }
|
| 3871 |
|
| 3872 | |
| 3873 | * If kswapd is already active on a node, keep looking
|
| 3874 | * for other nodes that might be idle. This can happen
|
| 3875 | * if another process has NUMA bindings and is causing
|
| 3876 | * kswapd wakeups on only some nodes. Avoid accidental
|
| 3877 | * "node_reclaim_mode"-like behavior in this case.
|
| 3878 |
|
| 3879 | if (skip_kswapd_nodes &&
|
| 3880 | !waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
|
| 3881 | skipped_kswapd_nodes = true;
|
| 3882 | continue;
|
| 3883 | }
|
| 3884 |
|
| 3885 | cond_accept_memory(zone, order, alloc_flags);
|
| 3886 |
|
| 3887 | |
| 3888 | * Detect whether the number of free pages is below high
|
| 3889 | * watermark. If so, we will decrease pcp->high and free
|
| 3890 | * PCP pages in free path to reduce the possibility of
|
| 3891 | * premature page reclaiming. Detection is done here to
|
| 3892 | * avoid to do that in hotter free path.
|
| 3893 |
|
| 3894 | if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
|
| 3895 | goto check_alloc_wmark;
|
| 3896 |
|
| 3897 | mark = high_wmark_pages(zone);
|
| 3898 | if (zone_watermark_fast(zone, order, mark,
|
| 3899 | ac->highest_zoneidx, alloc_flags,
|
| 3900 | gfp_mask))
|
| 3901 | goto try_this_zone;
|
| 3902 | else
|
| 3903 | set_bit(ZONE_BELOW_HIGH, &zone->flags);
|
| 3904 |
|
| 3905 | check_alloc_wmark:
|
| 3906 | mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
|
| 3907 | if (!zone_watermark_fast(zone, order, mark,
|
| 3908 | ac->highest_zoneidx, alloc_flags,
|
| 3909 | gfp_mask)) {
|
| 3910 | if (cond_accept_memory(zone, order, alloc_flags))
|
| 3911 | goto try_this_zone;
|
| 3912 |
|
| 3913 | |
| 3914 | * Watermark failed for this zone, but see if we can
|
| 3915 | * grow this zone if it contains deferred pages.
|
| 3916 |
|
| 3917 | if (deferred_pages_enabled()) {
|
| 3918 | if (_deferred_grow_zone(zone, order))
|
| 3919 | goto try_this_zone;
|
| 3920 | }
|
| 3921 |
|
| 3922 | BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
|
| 3923 | if (alloc_flags & ALLOC_NO_WATERMARKS)
|
| 3924 | goto try_this_zone;
|
| 3925 |
|
| 3926 | if (!node_reclaim_enabled() ||
|
| 3927 | !zone_allows_reclaim(zonelist_zone(ac->preferred_zoneref), zone))
|
| 3928 | continue;
|
| 3929 |
|
| 3930 | if (!node_reclaim(zone->zone_pgdat, gfp_mask, order))
|
| 3931 | continue;
|
| 3932 |
|
| 3933 |
|
| 3934 | if (!zone_watermark_ok(zone, order, mark,
|
| 3935 | ac->highest_zoneidx, alloc_flags))
|
| 3936 | continue;
|
| 3937 | }
|
| 3938 |
|
| 3939 | try_this_zone:
|
| 3940 | page = rmqueue(zonelist_zone(ac->preferred_zoneref), zone, order,
|
| 3941 | gfp_mask, alloc_flags, ac->migratetype);
|
| 3942 | if (page) {
|
| 3943 | prep_new_page(page, order, gfp_mask, alloc_flags);
|
| 3944 |
|
| 3945 | return page;
|
| 3946 | } else {
|
| 3947 | if (cond_accept_memory(zone, order, alloc_flags))
|
| 3948 | goto try_this_zone;
|
| 3949 |
|
| 3950 |
|
| 3951 | if (deferred_pages_enabled()) {
|
| 3952 | if (_deferred_grow_zone(zone, order))
|
| 3953 | goto try_this_zone;
|
| 3954 | }
|
| 3955 | }
|
| 3956 | }
|
| 3957 |
|
| 3958 | |
| 3959 | * If we skipped over nodes with active kswapds and found no
|
| 3960 | * idle nodes, retry and place anywhere the watermarks permit.
|
| 3961 |
|
| 3962 | if (skip_kswapd_nodes && skipped_kswapd_nodes) {
|
| 3963 | skip_kswapd_nodes = false;
|
| 3964 | goto retry;
|
| 3965 | }
|
| 3966 |
|
| 3967 | |
| 3968 | * It's possible on a UMA machine to get through all zones that are
|
| 3969 | * fragmented. If avoiding fragmentation, reset and try again.
|
| 3970 |
|
| 3971 | if (no_fallback && !defrag_mode) {
|
| 3972 | alloc_flags &= ~ALLOC_NOFRAGMENT;
|
| 3973 | goto retry;
|
| 3974 | }
|
| 3975 |
|
| 3976 | return NULL;
|
| 3977 | }
|
| 3978 |
|
| 3979 | static void warn_alloc_show_mem(gfp_t gfp_mask, const nodemask_t *nodemask)
|
| 3980 | {
|
| 3981 | unsigned int filter = SHOW_MEM_FILTER_NODES;
|
| 3982 |
|
| 3983 | |
| 3984 | * This documents exceptions given to allocations in certain
|
| 3985 | * contexts that are allowed to allocate outside current's set
|
| 3986 | * of allowed nodes.
|
| 3987 |
|
| 3988 | if (!(gfp_mask & __GFP_NOMEMALLOC))
|
| 3989 | if (tsk_is_oom_victim(current) ||
|
| 3990 | (current->flags & (PF_MEMALLOC | PF_EXITING)))
|
| 3991 | filter &= ~SHOW_MEM_FILTER_NODES;
|
| 3992 | if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
|
| 3993 | filter &= ~SHOW_MEM_FILTER_NODES;
|
| 3994 |
|
| 3995 | __show_mem(filter, nodemask, gfp_zone(gfp_mask));
|
| 3996 | mem_cgroup_show_protected_memory(NULL);
|
| 3997 | }
|
| 3998 |
|
| 3999 | void warn_alloc(gfp_t gfp_mask, const nodemask_t *nodemask, const char *fmt, ...)
|
| 4000 | {
|
| 4001 | struct va_format vaf;
|
| 4002 | va_list args;
|
| 4003 | static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
|
| 4004 |
|
| 4005 | if ((gfp_mask & __GFP_NOWARN) ||
|
| 4006 | !__ratelimit(&nopage_rs) ||
|
| 4007 | ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
|
| 4008 | return;
|
| 4009 |
|
| 4010 | va_start(args, fmt);
|
| 4011 | vaf.fmt = fmt;
|
| 4012 | vaf.va = &args;
|
| 4013 | pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
|
| 4014 | current->comm, &vaf, gfp_mask, &gfp_mask,
|
| 4015 | nodemask_pr_args(nodemask));
|
| 4016 | va_end(args);
|
| 4017 |
|
| 4018 | cpuset_print_current_mems_allowed();
|
| 4019 | pr_cont("\n");
|
| 4020 | dump_stack();
|
| 4021 | warn_alloc_show_mem(gfp_mask, nodemask);
|
| 4022 | }
|
| 4023 |
|
| 4024 | static inline struct page *
|
| 4025 | __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
|
| 4026 | unsigned int alloc_flags,
|
| 4027 | const struct alloc_context *ac)
|
| 4028 | {
|
| 4029 | struct page *page;
|
| 4030 |
|
| 4031 | page = get_page_from_freelist(gfp_mask, order,
|
| 4032 | alloc_flags|ALLOC_CPUSET, ac);
|
| 4033 | |
| 4034 | * fallback to ignore cpuset restriction if our nodes
|
| 4035 | * are depleted
|
| 4036 |
|
| 4037 | if (!page)
|
| 4038 | page = get_page_from_freelist(gfp_mask, order,
|
| 4039 | alloc_flags, ac);
|
| 4040 | return page;
|
| 4041 | }
|
| 4042 |
|
| 4043 | static inline struct page *
|
| 4044 | __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
|
| 4045 | const struct alloc_context *ac, unsigned long *did_some_progress)
|
| 4046 | {
|
| 4047 | struct oom_control oc = {
|
| 4048 | .zonelist = ac->zonelist,
|
| 4049 | .nodemask = ac->nodemask,
|
| 4050 | .memcg = NULL,
|
| 4051 | .gfp_mask = gfp_mask,
|
| 4052 | .order = order,
|
| 4053 | };
|
| 4054 | struct page *page;
|
| 4055 |
|
| 4056 | *did_some_progress = 0;
|
| 4057 |
|
| 4058 | |
| 4059 | * Acquire the oom lock. If that fails, somebody else is
|
| 4060 | * making progress for us.
|
| 4061 |
|
| 4062 | if (!mutex_trylock(&oom_lock)) {
|
| 4063 | *did_some_progress = 1;
|
| 4064 | schedule_timeout_uninterruptible(1);
|
| 4065 | return NULL;
|
| 4066 | }
|
| 4067 |
|
| 4068 | |
| 4069 | * Go through the zonelist yet one more time, keep very high watermark
|
| 4070 | * here, this is only to catch a parallel oom killing, we must fail if
|
| 4071 | * we're still under heavy pressure. But make sure that this reclaim
|
| 4072 | * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
|
| 4073 | * allocation which will never fail due to oom_lock already held.
|
| 4074 |
|
| 4075 | page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
|
| 4076 | ~__GFP_DIRECT_RECLAIM, order,
|
| 4077 | ac->alloc_flags|ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
|
| 4078 | if (page)
|
| 4079 | goto out;
|
| 4080 |
|
| 4081 |
|
| 4082 | if (current->flags & PF_DUMPCORE)
|
| 4083 | goto out;
|
| 4084 |
|
| 4085 | if (order > PAGE_ALLOC_COSTLY_ORDER)
|
| 4086 | goto out;
|
| 4087 | |
| 4088 | * We have already exhausted all our reclaim opportunities without any
|
| 4089 | * success so it is time to admit defeat. We will skip the OOM killer
|
| 4090 | * because it is very likely that the caller has a more reasonable
|
| 4091 | * fallback than shooting a random task.
|
| 4092 | *
|
| 4093 | * The OOM killer may not free memory on a specific node.
|
| 4094 |
|
| 4095 | if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
|
| 4096 | goto out;
|
| 4097 |
|
| 4098 | if (ac->highest_zoneidx < ZONE_NORMAL)
|
| 4099 | goto out;
|
| 4100 | if (pm_suspended_storage())
|
| 4101 | goto out;
|
| 4102 | |
| 4103 | * XXX: GFP_NOFS allocations should rather fail than rely on
|
| 4104 | * other request to make a forward progress.
|
| 4105 | * We are in an unfortunate situation where out_of_memory cannot
|
| 4106 | * do much for this context but let's try it to at least get
|
| 4107 | * access to memory reserved if the current task is killed (see
|
| 4108 | * out_of_memory). Once filesystems are ready to handle allocation
|
| 4109 | * failures more gracefully we should just bail out here.
|
| 4110 |
|
| 4111 |
|
| 4112 |
|
| 4113 | if (out_of_memory(&oc) ||
|
| 4114 | WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
|
| 4115 | *did_some_progress = 1;
|
| 4116 |
|
| 4117 | |
| 4118 | * Help non-failing allocations by giving them access to memory
|
| 4119 | * reserves
|
| 4120 |
|
| 4121 | if (gfp_mask & __GFP_NOFAIL)
|
| 4122 | page = __alloc_pages_cpuset_fallback(gfp_mask, order,
|
| 4123 | ac->alloc_flags|ALLOC_NO_WATERMARKS, ac);
|
| 4124 | }
|
| 4125 | out:
|
| 4126 | mutex_unlock(&oom_lock);
|
| 4127 | return page;
|
| 4128 | }
|
| 4129 |
|
| 4130 | |
| 4131 | * Maximum number of compaction retries with a progress before OOM
|
| 4132 | * killer is consider as the only way to move forward.
|
| 4133 |
|
| 4134 | #define MAX_COMPACT_RETRIES 16
|
| 4135 |
|
| 4136 | #ifdef CONFIG_COMPACTION
|
| 4137 |
|
| 4138 | static struct page *
|
| 4139 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
|
| 4140 | unsigned int alloc_flags, const struct alloc_context *ac,
|
| 4141 | enum compact_priority prio, enum compact_result *compact_result)
|
| 4142 | {
|
| 4143 | struct page *page = NULL;
|
| 4144 | unsigned long pflags;
|
| 4145 | unsigned int noreclaim_flag;
|
| 4146 | struct capture_control capc = {
|
| 4147 | .zone = NULL,
|
| 4148 | .migratetype = ac->migratetype,
|
| 4149 | .order = order,
|
| 4150 | .page = NULL,
|
| 4151 | };
|
| 4152 | int compact_order = order;
|
| 4153 |
|
| 4154 | |
| 4155 | * If fallbacks are not permitted (defrag_mode), we either
|
| 4156 | * need to reclaim space in a block of matching type, or clear
|
| 4157 | * out an entire block to allow __rmqueue_claim() to convert.
|
| 4158 | *
|
| 4159 | * Reclaim by itself is primarily freeing space in movable
|
| 4160 | * blocks, since that's where the LRU pages live. So this
|
| 4161 | * works for movable requests, but not for others.
|
| 4162 | *
|
| 4163 | * For those, promote the order to help make blocks, instead
|
| 4164 | * of spinning in reclaim alone unproductively.
|
| 4165 |
|
| 4166 | if ((alloc_flags & ALLOC_NOFRAGMENT) && ac->migratetype != MIGRATE_MOVABLE)
|
| 4167 | compact_order = max(order, pageblock_order);
|
| 4168 |
|
| 4169 | if (!compact_order)
|
| 4170 | return NULL;
|
| 4171 |
|
| 4172 | psi_memstall_enter(&pflags);
|
| 4173 | delayacct_compact_start();
|
| 4174 | fs_reclaim_acquire(gfp_mask);
|
| 4175 | noreclaim_flag = memalloc_noreclaim_save();
|
| 4176 |
|
| 4177 | |
| 4178 | * Make sure the structs are really initialized before we expose the
|
| 4179 | * capture control, in case we are interrupted and the interrupt handler
|
| 4180 | * frees a page.
|
| 4181 |
|
| 4182 | barrier();
|
| 4183 | WRITE_ONCE(current->capture_control, &capc);
|
| 4184 |
|
| 4185 | *compact_result = try_to_compact_pages(gfp_mask, compact_order,
|
| 4186 | alloc_flags, ac, prio, &capc);
|
| 4187 |
|
| 4188 | |
| 4189 | * Make sure we hide capture control first before we read the captured
|
| 4190 | * page pointer, otherwise an interrupt could free and capture a page
|
| 4191 | * and we would leak it.
|
| 4192 |
|
| 4193 | WRITE_ONCE(current->capture_control, NULL);
|
| 4194 | page = READ_ONCE(capc.page);
|
| 4195 |
|
| 4196 | |
| 4197 | * Technically, it is also possible that compaction is skipped but
|
| 4198 | * the page is still captured out of luck(IRQ came and freed the page).
|
| 4199 | * Returning COMPACT_SUCCESS in such cases helps in properly accounting
|
| 4200 | * the COMPACT[STALL|FAIL] when compaction is skipped.
|
| 4201 |
|
| 4202 | if (page)
|
| 4203 | *compact_result = COMPACT_SUCCESS;
|
| 4204 |
|
| 4205 | memalloc_noreclaim_restore(noreclaim_flag);
|
| 4206 | fs_reclaim_release(gfp_mask);
|
| 4207 | psi_memstall_leave(&pflags);
|
| 4208 | delayacct_compact_end();
|
| 4209 |
|
| 4210 | if (*compact_result == COMPACT_SKIPPED ||
|
| 4211 | *compact_result == COMPACT_DEFERRED)
|
| 4212 | return NULL;
|
| 4213 | |
| 4214 | * At least in one zone compaction wasn't deferred or skipped, so let's
|
| 4215 | * count a compaction stall
|
| 4216 |
|
| 4217 | count_vm_event(COMPACTSTALL);
|
| 4218 |
|
| 4219 |
|
| 4220 | if (page)
|
| 4221 | prep_new_page(page, order, gfp_mask, alloc_flags);
|
| 4222 |
|
| 4223 |
|
| 4224 | if (!page)
|
| 4225 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
|
| 4226 |
|
| 4227 | if (page) {
|
| 4228 | struct zone *zone = page_zone(page);
|
| 4229 |
|
| 4230 | zone->compact_blockskip_flush = false;
|
| 4231 | compaction_defer_reset(zone, compact_order, true);
|
| 4232 | count_vm_event(COMPACTSUCCESS);
|
| 4233 | return page;
|
| 4234 | }
|
| 4235 |
|
| 4236 | |
| 4237 | * It's bad if compaction run occurs and fails. The most likely reason
|
| 4238 | * is that pages exist, but not enough to satisfy watermarks.
|
| 4239 |
|
| 4240 | count_vm_event(COMPACTFAIL);
|
| 4241 |
|
| 4242 | cond_resched();
|
| 4243 |
|
| 4244 | return NULL;
|
| 4245 | }
|
| 4246 |
|
| 4247 | static inline bool
|
| 4248 | should_compact_retry(gfp_t gfp_mask, struct alloc_context *ac, int order,
|
| 4249 | int alloc_flags,
|
| 4250 | enum compact_result compact_result,
|
| 4251 | enum compact_priority *compact_priority,
|
| 4252 | int *compaction_retries)
|
| 4253 | {
|
| 4254 | int max_retries = MAX_COMPACT_RETRIES;
|
| 4255 | int min_priority;
|
| 4256 | bool ret = false;
|
| 4257 | int retries = *compaction_retries;
|
| 4258 | enum compact_priority priority = *compact_priority;
|
| 4259 |
|
| 4260 | if (!order)
|
| 4261 | return false;
|
| 4262 |
|
| 4263 | if (fatal_signal_pending(current))
|
| 4264 | return false;
|
| 4265 |
|
| 4266 | |
| 4267 | * Compaction was skipped due to a lack of free order-0
|
| 4268 | * migration targets. Continue if reclaim can help.
|
| 4269 |
|
| 4270 | if (compact_result == COMPACT_SKIPPED) {
|
| 4271 | ret = compaction_zonelist_suitable(ac, order, alloc_flags,
|
| 4272 | gfp_mask);
|
| 4273 | goto out;
|
| 4274 | }
|
| 4275 |
|
| 4276 | |
| 4277 | * Compaction managed to coalesce some page blocks, but the
|
| 4278 | * allocation failed presumably due to a race. Retry some.
|
| 4279 |
|
| 4280 | if (compact_result == COMPACT_SUCCESS) {
|
| 4281 | |
| 4282 | * !costly requests are much more important than
|
| 4283 | * __GFP_RETRY_MAYFAIL costly ones because they are de
|
| 4284 | * facto nofail and invoke OOM killer to move on while
|
| 4285 | * costly can fail and users are ready to cope with
|
| 4286 | * that. 1/4 retries is rather arbitrary but we would
|
| 4287 | * need much more detailed feedback from compaction to
|
| 4288 | * make a better decision.
|
| 4289 |
|
| 4290 | if (order > PAGE_ALLOC_COSTLY_ORDER)
|
| 4291 | max_retries /= 4;
|
| 4292 |
|
| 4293 | if (++(*compaction_retries) <= max_retries) {
|
| 4294 | ret = true;
|
| 4295 | goto out;
|
| 4296 | }
|
| 4297 | }
|
| 4298 |
|
| 4299 | |
| 4300 | * Compaction failed. Retry with increasing priority.
|
| 4301 |
|
| 4302 | min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
|
| 4303 | MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
|
| 4304 |
|
| 4305 | if (*compact_priority > min_priority) {
|
| 4306 | (*compact_priority)--;
|
| 4307 | *compaction_retries = 0;
|
| 4308 | ret = true;
|
| 4309 | }
|
| 4310 | out:
|
| 4311 | trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
|
| 4312 | return ret;
|
| 4313 | }
|
| 4314 | #else
|
| 4315 | static inline struct page *
|
| 4316 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
|
| 4317 | unsigned int alloc_flags, const struct alloc_context *ac,
|
| 4318 | enum compact_priority prio, enum compact_result *compact_result)
|
| 4319 | {
|
| 4320 | *compact_result = COMPACT_SKIPPED;
|
| 4321 | return NULL;
|
| 4322 | }
|
| 4323 |
|
| 4324 | static inline bool
|
| 4325 | should_compact_retry(gfp_t gfp_mask, struct alloc_context *ac, int order,
|
| 4326 | int alloc_flags,
|
| 4327 | enum compact_result compact_result,
|
| 4328 | enum compact_priority *compact_priority,
|
| 4329 | int *compaction_retries)
|
| 4330 | {
|
| 4331 | struct zone *zone;
|
| 4332 | struct zoneref *z;
|
| 4333 |
|
| 4334 | if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
|
| 4335 | return false;
|
| 4336 |
|
| 4337 | |
| 4338 | * There are setups with compaction disabled which would prefer to loop
|
| 4339 | * inside the allocator rather than hit the oom killer prematurely.
|
| 4340 | * Let's give them a good hope and keep retrying while the order-0
|
| 4341 | * watermarks are OK.
|
| 4342 |
|
| 4343 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
|
| 4344 | ac->highest_zoneidx, ac->nodemask) {
|
| 4345 | if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
|
| 4346 | ac->highest_zoneidx, alloc_flags))
|
| 4347 | return true;
|
| 4348 | }
|
| 4349 | return false;
|
| 4350 | }
|
| 4351 | #endif
|
| 4352 |
|
| 4353 | #ifdef CONFIG_LOCKDEP
|
| 4354 | static struct lockdep_map __fs_reclaim_map =
|
| 4355 | STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
|
| 4356 |
|
| 4357 | static bool __need_reclaim(gfp_t gfp_mask)
|
| 4358 | {
|
| 4359 |
|
| 4360 | if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
|
| 4361 | return false;
|
| 4362 |
|
| 4363 |
|
| 4364 | if (current->flags & PF_MEMALLOC)
|
| 4365 | return false;
|
| 4366 |
|
| 4367 | if (gfp_mask & __GFP_NOLOCKDEP)
|
| 4368 | return false;
|
| 4369 |
|
| 4370 | return true;
|
| 4371 | }
|
| 4372 |
|
| 4373 | void __fs_reclaim_acquire(unsigned long ip)
|
| 4374 | {
|
| 4375 | lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
|
| 4376 | }
|
| 4377 |
|
| 4378 | void __fs_reclaim_release(unsigned long ip)
|
| 4379 | {
|
| 4380 | lock_release(&__fs_reclaim_map, ip);
|
| 4381 | }
|
| 4382 |
|
| 4383 | void fs_reclaim_acquire(gfp_t gfp_mask)
|
| 4384 | {
|
| 4385 | gfp_mask = current_gfp_context(gfp_mask);
|
| 4386 |
|
| 4387 | if (__need_reclaim(gfp_mask)) {
|
| 4388 | if (gfp_mask & __GFP_FS)
|
| 4389 | __fs_reclaim_acquire(_RET_IP_);
|
| 4390 |
|
| 4391 | #ifdef CONFIG_MMU_NOTIFIER
|
| 4392 | lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
|
| 4393 | lock_map_release(&__mmu_notifier_invalidate_range_start_map);
|
| 4394 | #endif
|
| 4395 |
|
| 4396 | }
|
| 4397 | }
|
| 4398 | EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
|
| 4399 |
|
| 4400 | void fs_reclaim_release(gfp_t gfp_mask)
|
| 4401 | {
|
| 4402 | gfp_mask = current_gfp_context(gfp_mask);
|
| 4403 |
|
| 4404 | if (__need_reclaim(gfp_mask)) {
|
| 4405 | if (gfp_mask & __GFP_FS)
|
| 4406 | __fs_reclaim_release(_RET_IP_);
|
| 4407 | }
|
| 4408 | }
|
| 4409 | EXPORT_SYMBOL_GPL(fs_reclaim_release);
|
| 4410 | #endif
|
| 4411 |
|
| 4412 | |
| 4413 | * Zonelists may change due to hotplug during allocation. Detect when zonelists
|
| 4414 | * have been rebuilt so allocation retries. Reader side does not lock and
|
| 4415 | * retries the allocation if zonelist changes. Writer side is protected by the
|
| 4416 | * embedded spin_lock.
|
| 4417 |
|
| 4418 | static DEFINE_SEQLOCK(zonelist_update_seq);
|
| 4419 |
|
| 4420 | static unsigned int zonelist_iter_begin(void)
|
| 4421 | {
|
| 4422 | if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
|
| 4423 | return read_seqbegin(&zonelist_update_seq);
|
| 4424 |
|
| 4425 | return 0;
|
| 4426 | }
|
| 4427 |
|
| 4428 | static unsigned int check_retry_zonelist(unsigned int seq)
|
| 4429 | {
|
| 4430 | if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
|
| 4431 | return read_seqretry(&zonelist_update_seq, seq);
|
| 4432 |
|
| 4433 | return seq;
|
| 4434 | }
|
| 4435 |
|
| 4436 |
|
| 4437 | static unsigned long
|
| 4438 | __perform_reclaim(gfp_t gfp_mask, unsigned int order,
|
| 4439 | const struct alloc_context *ac)
|
| 4440 | {
|
| 4441 | unsigned int noreclaim_flag;
|
| 4442 | unsigned long progress;
|
| 4443 |
|
| 4444 | cond_resched();
|
| 4445 |
|
| 4446 |
|
| 4447 | cpuset_memory_pressure_bump();
|
| 4448 | fs_reclaim_acquire(gfp_mask);
|
| 4449 | noreclaim_flag = memalloc_noreclaim_save();
|
| 4450 |
|
| 4451 | progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
|
| 4452 | ac->nodemask);
|
| 4453 |
|
| 4454 | memalloc_noreclaim_restore(noreclaim_flag);
|
| 4455 | fs_reclaim_release(gfp_mask);
|
| 4456 |
|
| 4457 | cond_resched();
|
| 4458 |
|
| 4459 | return progress;
|
| 4460 | }
|
| 4461 |
|
| 4462 |
|
| 4463 | static inline struct page *
|
| 4464 | __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
|
| 4465 | unsigned int alloc_flags, const struct alloc_context *ac,
|
| 4466 | unsigned long *did_some_progress)
|
| 4467 | {
|
| 4468 | struct page *page = NULL;
|
| 4469 | unsigned long pflags;
|
| 4470 | bool drained = false;
|
| 4471 | int reclaim_order = order;
|
| 4472 |
|
| 4473 |
|
| 4474 | if ((alloc_flags & ALLOC_NOFRAGMENT) && ac->migratetype != MIGRATE_MOVABLE)
|
| 4475 | reclaim_order = max(order, pageblock_order);
|
| 4476 |
|
| 4477 | psi_memstall_enter(&pflags);
|
| 4478 | *did_some_progress = __perform_reclaim(gfp_mask, reclaim_order, ac);
|
| 4479 | if (unlikely(!(*did_some_progress)))
|
| 4480 | goto out;
|
| 4481 |
|
| 4482 | retry:
|
| 4483 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
|
| 4484 |
|
| 4485 | |
| 4486 | * If an allocation failed after direct reclaim, it could be because
|
| 4487 | * pages are pinned on the per-cpu lists or in high alloc reserves.
|
| 4488 | * Shrink them and try again
|
| 4489 |
|
| 4490 | if (!page && !drained) {
|
| 4491 | unreserve_highatomic_pageblock(ac, false);
|
| 4492 | drain_all_pages(NULL);
|
| 4493 | drained = true;
|
| 4494 | goto retry;
|
| 4495 | }
|
| 4496 | out:
|
| 4497 | psi_memstall_leave(&pflags);
|
| 4498 |
|
| 4499 | return page;
|
| 4500 | }
|
| 4501 |
|
| 4502 | static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
|
| 4503 | const struct alloc_context *ac)
|
| 4504 | {
|
| 4505 | struct zoneref *z;
|
| 4506 | struct zone *zone;
|
| 4507 | pg_data_t *last_pgdat = NULL;
|
| 4508 | enum zone_type highest_zoneidx = ac->highest_zoneidx;
|
| 4509 | unsigned int reclaim_order;
|
| 4510 |
|
| 4511 | if (defrag_mode)
|
| 4512 | reclaim_order = max(order, pageblock_order);
|
| 4513 | else
|
| 4514 | reclaim_order = order;
|
| 4515 |
|
| 4516 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
|
| 4517 | ac->nodemask) {
|
| 4518 | if (!managed_zone(zone))
|
| 4519 | continue;
|
| 4520 | if (last_pgdat == zone->zone_pgdat)
|
| 4521 | continue;
|
| 4522 | wakeup_kswapd(zone, gfp_mask, reclaim_order, highest_zoneidx);
|
| 4523 | last_pgdat = zone->zone_pgdat;
|
| 4524 | }
|
| 4525 | }
|
| 4526 |
|
| 4527 | static inline unsigned int
|
| 4528 | alloc_flags_nonblocking(gfp_t gfp_mask, unsigned int order)
|
| 4529 | {
|
| 4530 | unsigned int alloc_flags = 0;
|
| 4531 |
|
| 4532 | if (gfp_mask & __GFP_DIRECT_RECLAIM)
|
| 4533 | return 0;
|
| 4534 |
|
| 4535 | |
| 4536 | * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
|
| 4537 | * if it can't schedule.
|
| 4538 |
|
| 4539 | if (gfp_mask & __GFP_NOMEMALLOC)
|
| 4540 | return 0;
|
| 4541 |
|
| 4542 | alloc_flags |= ALLOC_NON_BLOCK;
|
| 4543 |
|
| 4544 | if (order > 0 && (gfp_mask & __GFP_HIGH))
|
| 4545 | alloc_flags |= ALLOC_HIGHATOMIC;
|
| 4546 |
|
| 4547 | return alloc_flags;
|
| 4548 | }
|
| 4549 |
|
| 4550 | static inline unsigned int
|
| 4551 | alloc_flags_slowpath(gfp_t gfp_mask, unsigned int order)
|
| 4552 | {
|
| 4553 | unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
|
| 4554 |
|
| 4555 | |
| 4556 | * The caller may dip into page reserves a bit more if the caller
|
| 4557 | * cannot run direct reclaim, or if the caller has realtime scheduling
|
| 4558 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
|
| 4559 | * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
|
| 4560 |
|
| 4561 | if (gfp_mask & __GFP_HIGH)
|
| 4562 | alloc_flags |= ALLOC_MIN_RESERVE;
|
| 4563 | if (gfp_mask & __GFP_KSWAPD_RECLAIM)
|
| 4564 | alloc_flags |= ALLOC_KSWAPD;
|
| 4565 |
|
| 4566 | alloc_flags |= alloc_flags_nonblocking(gfp_mask, order);
|
| 4567 |
|
| 4568 | if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
|
| 4569 | |
| 4570 | * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
|
| 4571 | * GFP_ATOMIC) rather than fail, see the comment for
|
| 4572 | * cpuset_current_node_allowed().
|
| 4573 |
|
| 4574 | if (alloc_flags & ALLOC_MIN_RESERVE)
|
| 4575 | alloc_flags &= ~ALLOC_CPUSET;
|
| 4576 | } else if (unlikely(rt_or_dl_task(current)) && in_task())
|
| 4577 | alloc_flags |= ALLOC_MIN_RESERVE;
|
| 4578 |
|
| 4579 | alloc_flags |= alloc_flags_cma(gfp_mask);
|
| 4580 |
|
| 4581 | if (defrag_mode)
|
| 4582 | alloc_flags |= ALLOC_NOFRAGMENT;
|
| 4583 |
|
| 4584 | return alloc_flags;
|
| 4585 | }
|
| 4586 |
|
| 4587 | static bool oom_reserves_allowed(struct task_struct *tsk)
|
| 4588 | {
|
| 4589 | if (!tsk_is_oom_victim(tsk))
|
| 4590 | return false;
|
| 4591 |
|
| 4592 | |
| 4593 | * !MMU doesn't have oom reaper so give access to memory reserves
|
| 4594 | * only to the thread with TIF_MEMDIE set
|
| 4595 |
|
| 4596 | if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
|
| 4597 | return false;
|
| 4598 |
|
| 4599 | return true;
|
| 4600 | }
|
| 4601 |
|
| 4602 | |
| 4603 | * Distinguish requests which really need access to full memory
|
| 4604 | * reserves from oom victims which can live with a portion of it
|
| 4605 |
|
| 4606 | static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
|
| 4607 | {
|
| 4608 | if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
|
| 4609 | return 0;
|
| 4610 | if (gfp_mask & __GFP_MEMALLOC)
|
| 4611 | return ALLOC_NO_WATERMARKS;
|
| 4612 | if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
|
| 4613 | return ALLOC_NO_WATERMARKS;
|
| 4614 | if (!in_interrupt()) {
|
| 4615 | if (current->flags & PF_MEMALLOC)
|
| 4616 | return ALLOC_NO_WATERMARKS;
|
| 4617 | else if (oom_reserves_allowed(current))
|
| 4618 | return ALLOC_OOM;
|
| 4619 | }
|
| 4620 |
|
| 4621 | return 0;
|
| 4622 | }
|
| 4623 |
|
| 4624 | bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
|
| 4625 | {
|
| 4626 | return !!__gfp_pfmemalloc_flags(gfp_mask);
|
| 4627 | }
|
| 4628 |
|
| 4629 | |
| 4630 | * Checks whether it makes sense to retry the reclaim to make a forward progress
|
| 4631 | * for the given allocation request.
|
| 4632 | *
|
| 4633 | * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
|
| 4634 | * without success, or when we couldn't even meet the watermark if we
|
| 4635 | * reclaimed all remaining pages on the LRU lists.
|
| 4636 | *
|
| 4637 | * Returns true if a retry is viable or false to enter the oom path.
|
| 4638 |
|
| 4639 | static inline bool
|
| 4640 | should_reclaim_retry(gfp_t gfp_mask, unsigned order,
|
| 4641 | struct alloc_context *ac, int alloc_flags,
|
| 4642 | bool did_some_progress, int *no_progress_loops)
|
| 4643 | {
|
| 4644 | struct zone *zone;
|
| 4645 | struct zoneref *z;
|
| 4646 | bool ret = false;
|
| 4647 |
|
| 4648 | |
| 4649 | * Costly allocations might have made a progress but this doesn't mean
|
| 4650 | * their order will become available due to high fragmentation so
|
| 4651 | * always increment the no progress counter for them
|
| 4652 |
|
| 4653 | if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
|
| 4654 | *no_progress_loops = 0;
|
| 4655 | else
|
| 4656 | (*no_progress_loops)++;
|
| 4657 |
|
| 4658 | if (*no_progress_loops > MAX_RECLAIM_RETRIES)
|
| 4659 | goto out;
|
| 4660 |
|
| 4661 |
|
| 4662 | |
| 4663 | * Keep reclaiming pages while there is a chance this will lead
|
| 4664 | * somewhere. If none of the target zones can satisfy our allocation
|
| 4665 | * request even if all reclaimable pages are considered then we are
|
| 4666 | * screwed and have to go OOM.
|
| 4667 |
|
| 4668 | for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
|
| 4669 | ac->highest_zoneidx, ac->nodemask) {
|
| 4670 | unsigned long available;
|
| 4671 | unsigned long reclaimable;
|
| 4672 | unsigned long min_wmark = min_wmark_pages(zone);
|
| 4673 | bool wmark;
|
| 4674 |
|
| 4675 | if (cpusets_enabled() &&
|
| 4676 | (alloc_flags & ALLOC_CPUSET) &&
|
| 4677 | !__cpuset_zone_allowed(zone, gfp_mask))
|
| 4678 | continue;
|
| 4679 |
|
| 4680 | available = reclaimable = zone_reclaimable_pages(zone);
|
| 4681 | available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
|
| 4682 |
|
| 4683 | |
| 4684 | * Would the allocation succeed if we reclaimed all
|
| 4685 | * reclaimable pages?
|
| 4686 |
|
| 4687 | wmark = __zone_watermark_ok(zone, order, min_wmark,
|
| 4688 | ac->highest_zoneidx, alloc_flags, available);
|
| 4689 | trace_reclaim_retry_zone(z, order, reclaimable,
|
| 4690 | available, min_wmark, *no_progress_loops, wmark);
|
| 4691 | if (wmark) {
|
| 4692 | ret = true;
|
| 4693 | break;
|
| 4694 | }
|
| 4695 | }
|
| 4696 |
|
| 4697 | |
| 4698 | * Memory allocation/reclaim might be called from a WQ context and the
|
| 4699 | * current implementation of the WQ concurrency control doesn't
|
| 4700 | * recognize that a particular WQ is congested if the worker thread is
|
| 4701 | * looping without ever sleeping. Therefore we have to do a short sleep
|
| 4702 | * here rather than calling cond_resched().
|
| 4703 |
|
| 4704 | if (current->flags & PF_WQ_WORKER)
|
| 4705 | schedule_timeout_uninterruptible(1);
|
| 4706 | else
|
| 4707 | cond_resched();
|
| 4708 | out:
|
| 4709 |
|
| 4710 | if (!ret)
|
| 4711 | return unreserve_highatomic_pageblock(ac, true);
|
| 4712 |
|
| 4713 | return ret;
|
| 4714 | }
|
| 4715 |
|
| 4716 | static inline bool
|
| 4717 | check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
|
| 4718 | {
|
| 4719 | |
| 4720 | * It's possible that cpuset's mems_allowed and the nodemask from
|
| 4721 | * mempolicy don't intersect. This should be normally dealt with by
|
| 4722 | * policy_nodemask(), but it's possible to race with cpuset update in
|
| 4723 | * such a way the check therein was true, and then it became false
|
| 4724 | * before we got our cpuset_mems_cookie here.
|
| 4725 | * This assumes that for all allocations, ac->nodemask can come only
|
| 4726 | * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
|
| 4727 | * when it does not intersect with the cpuset restrictions) or the
|
| 4728 | * caller can deal with a violated nodemask.
|
| 4729 |
|
| 4730 | if (cpusets_enabled() && ac->nodemask &&
|
| 4731 | !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
|
| 4732 | ac->nodemask = NULL;
|
| 4733 | return true;
|
| 4734 | }
|
| 4735 |
|
| 4736 | |
| 4737 | * When updating a task's mems_allowed or mempolicy nodemask, it is
|
| 4738 | * possible to race with parallel threads in such a way that our
|
| 4739 | * allocation can fail while the mask is being updated. If we are about
|
| 4740 | * to fail, check if the cpuset changed during allocation and if so,
|
| 4741 | * retry.
|
| 4742 |
|
| 4743 | if (read_mems_allowed_retry(cpuset_mems_cookie))
|
| 4744 | return true;
|
| 4745 |
|
| 4746 | return false;
|
| 4747 | }
|
| 4748 |
|
| 4749 | static void check_alloc_stall_warn(gfp_t gfp_mask, const nodemask_t *nodemask,
|
| 4750 | unsigned int order, unsigned long alloc_start_time)
|
| 4751 | {
|
| 4752 | static DEFINE_SPINLOCK(alloc_stall_lock);
|
| 4753 | unsigned long stall_msecs = jiffies_to_msecs(jiffies - alloc_start_time);
|
| 4754 |
|
| 4755 | if (likely(stall_msecs < ALLOC_STALL_WARN_MSECS))
|
| 4756 | return;
|
| 4757 | if (time_is_after_jiffies(READ_ONCE(alloc_stall_warn_jiffies)))
|
| 4758 | return;
|
| 4759 | if (gfp_mask & __GFP_NOWARN)
|
| 4760 | return;
|
| 4761 |
|
| 4762 | if (!spin_trylock(&alloc_stall_lock))
|
| 4763 | return;
|
| 4764 |
|
| 4765 |
|
| 4766 | if (time_is_after_jiffies(alloc_stall_warn_jiffies)) {
|
| 4767 | spin_unlock(&alloc_stall_lock);
|
| 4768 | return;
|
| 4769 | }
|
| 4770 |
|
| 4771 | WRITE_ONCE(alloc_stall_warn_jiffies, jiffies + msecs_to_jiffies(ALLOC_STALL_WARN_MSECS));
|
| 4772 | spin_unlock(&alloc_stall_lock);
|
| 4773 |
|
| 4774 | pr_warn("%s: page allocation stall for %lu secs: order:%d, mode:%#x(%pGg) nodemask=%*pbl",
|
| 4775 | current->comm, stall_msecs / MSEC_PER_SEC, order, gfp_mask, &gfp_mask,
|
| 4776 | nodemask_pr_args(nodemask));
|
| 4777 | cpuset_print_current_mems_allowed();
|
| 4778 | pr_cont("\n");
|
| 4779 | dump_stack();
|
| 4780 | warn_alloc_show_mem(gfp_mask, nodemask);
|
| 4781 | }
|
| 4782 |
|
| 4783 | static inline struct page *
|
| 4784 | __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
|
| 4785 | struct alloc_context *ac)
|
| 4786 | {
|
| 4787 | bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
|
| 4788 | bool can_compact = can_direct_reclaim && gfp_compaction_allowed(gfp_mask);
|
| 4789 | bool nofail = gfp_mask & __GFP_NOFAIL;
|
| 4790 | const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
|
| 4791 | struct page *page = NULL;
|
| 4792 | unsigned int alloc_flags;
|
| 4793 | unsigned long did_some_progress;
|
| 4794 | enum compact_priority compact_priority;
|
| 4795 | enum compact_result compact_result;
|
| 4796 | int compaction_retries;
|
| 4797 | int no_progress_loops;
|
| 4798 | unsigned int cpuset_mems_cookie;
|
| 4799 | unsigned int zonelist_iter_cookie;
|
| 4800 | int reserve_flags;
|
| 4801 | bool compact_first = false;
|
| 4802 | bool can_retry_reserves = true;
|
| 4803 | unsigned long alloc_start_time = jiffies;
|
| 4804 |
|
| 4805 | if (unlikely(nofail)) {
|
| 4806 | |
| 4807 | * Also we don't support __GFP_NOFAIL without __GFP_DIRECT_RECLAIM,
|
| 4808 | * otherwise, we may result in lockup.
|
| 4809 |
|
| 4810 | WARN_ON_ONCE(!can_direct_reclaim);
|
| 4811 | |
| 4812 | * PF_MEMALLOC request from this context is rather bizarre
|
| 4813 | * because we cannot reclaim anything and only can loop waiting
|
| 4814 | * for somebody to do a work for us.
|
| 4815 |
|
| 4816 | WARN_ON_ONCE(current->flags & PF_MEMALLOC);
|
| 4817 | }
|
| 4818 |
|
| 4819 | restart:
|
| 4820 | compaction_retries = 0;
|
| 4821 | no_progress_loops = 0;
|
| 4822 | compact_result = COMPACT_SKIPPED;
|
| 4823 | compact_priority = DEF_COMPACT_PRIORITY;
|
| 4824 | cpuset_mems_cookie = read_mems_allowed_begin();
|
| 4825 | zonelist_iter_cookie = zonelist_iter_begin();
|
| 4826 |
|
| 4827 | |
| 4828 | * For costly allocations, try direct compaction first, as it's likely
|
| 4829 | * that we have enough base pages and don't need to reclaim. For non-
|
| 4830 | * movable high-order allocations, do that as well, as compaction will
|
| 4831 | * try prevent permanent fragmentation by migrating from blocks of the
|
| 4832 | * same migratetype.
|
| 4833 |
|
| 4834 | if (can_compact && (costly_order || (order > 0 &&
|
| 4835 | ac->migratetype != MIGRATE_MOVABLE))) {
|
| 4836 | compact_first = true;
|
| 4837 | compact_priority = INIT_COMPACT_PRIORITY;
|
| 4838 | }
|
| 4839 |
|
| 4840 | |
| 4841 | * The fast path uses conservative alloc_flags to succeed only until
|
| 4842 | * kswapd needs to be woken up, and to avoid the cost of setting up
|
| 4843 | * alloc_flags precisely. So we do that now.
|
| 4844 | *
|
| 4845 | * Can't just or alloc_flags if it contains WMARK bits, but those flags
|
| 4846 | * shouldn't be set in ac->alloc_flags.
|
| 4847 |
|
| 4848 | VM_WARN_ON(ac->alloc_flags & ALLOC_WMARK_MASK);
|
| 4849 | alloc_flags = ac->alloc_flags | alloc_flags_slowpath(gfp_mask, order);
|
| 4850 |
|
| 4851 | |
| 4852 | * We need to recalculate the starting point for the zonelist iterator
|
| 4853 | * because we might have used different nodemask in the fast path, or
|
| 4854 | * there was a cpuset modification and we are retrying - otherwise we
|
| 4855 | * could end up iterating over non-eligible zones endlessly.
|
| 4856 |
|
| 4857 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
|
| 4858 | ac->highest_zoneidx, ac->nodemask);
|
| 4859 | if (!zonelist_zone(ac->preferred_zoneref))
|
| 4860 | goto nopage;
|
| 4861 |
|
| 4862 | |
| 4863 | * Check for insane configurations where the cpuset doesn't contain
|
| 4864 | * any suitable zone to satisfy the request - e.g. non-movable
|
| 4865 | * GFP_HIGHUSER allocations from MOVABLE nodes only.
|
| 4866 |
|
| 4867 | if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
|
| 4868 | struct zoneref *z = first_zones_zonelist(ac->zonelist,
|
| 4869 | ac->highest_zoneidx,
|
| 4870 | &cpuset_current_mems_allowed);
|
| 4871 | if (!zonelist_zone(z))
|
| 4872 | goto nopage;
|
| 4873 | }
|
| 4874 |
|
| 4875 | retry:
|
| 4876 |
|
| 4877 | if (alloc_flags & ALLOC_KSWAPD)
|
| 4878 | wake_all_kswapds(order, gfp_mask, ac);
|
| 4879 |
|
| 4880 | |
| 4881 | * The adjusted alloc_flags might result in immediate success, so try
|
| 4882 | * that first
|
| 4883 |
|
| 4884 | page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
|
| 4885 | if (page)
|
| 4886 | goto got_pg;
|
| 4887 |
|
| 4888 | reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
|
| 4889 | if (reserve_flags)
|
| 4890 | alloc_flags = alloc_flags_cma(gfp_mask) | reserve_flags |
|
| 4891 | ac->alloc_flags | (alloc_flags & ALLOC_KSWAPD);
|
| 4892 |
|
| 4893 | |
| 4894 | * Reset the nodemask and zonelist iterators if memory policies can be
|
| 4895 | * ignored. These allocations are high priority and system rather than
|
| 4896 | * user oriented.
|
| 4897 |
|
| 4898 | if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
|
| 4899 | ac->nodemask = NULL;
|
| 4900 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
|
| 4901 | ac->highest_zoneidx, ac->nodemask);
|
| 4902 |
|
| 4903 | |
| 4904 | * The first time we adjust anything due to being allowed to
|
| 4905 | * ignore memory policies or watermarks, retry immediately. This
|
| 4906 | * allows us to keep the first allocation attempt optimistic so
|
| 4907 | * it can succeed in a zone that is still above watermarks.
|
| 4908 |
|
| 4909 | if (can_retry_reserves) {
|
| 4910 | can_retry_reserves = false;
|
| 4911 | goto retry;
|
| 4912 | }
|
| 4913 | }
|
| 4914 |
|
| 4915 |
|
| 4916 | if (!can_direct_reclaim) {
|
| 4917 | |
| 4918 | * Reclaim/compaction cannot run, so defrag_mode's strategy
|
| 4919 | * of enforcing ALLOC_NOFRAGMENT cannot be fulfilled. Allow
|
| 4920 | * fallbacks rather than failing the allocation outright.
|
| 4921 |
|
| 4922 | if (defrag_mode && (alloc_flags & ALLOC_NOFRAGMENT) &&
|
| 4923 | (gfp_mask & __GFP_KSWAPD_RECLAIM)) {
|
| 4924 | alloc_flags &= ~ALLOC_NOFRAGMENT;
|
| 4925 | goto retry;
|
| 4926 | }
|
| 4927 | goto nopage;
|
| 4928 | }
|
| 4929 |
|
| 4930 |
|
| 4931 | if (current->flags & PF_MEMALLOC)
|
| 4932 | goto nopage;
|
| 4933 |
|
| 4934 |
|
| 4935 | check_alloc_stall_warn(gfp_mask, ac->nodemask, order, alloc_start_time);
|
| 4936 |
|
| 4937 |
|
| 4938 | if (!compact_first) {
|
| 4939 | page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags,
|
| 4940 | ac, &did_some_progress);
|
| 4941 | if (page)
|
| 4942 | goto got_pg;
|
| 4943 | }
|
| 4944 |
|
| 4945 |
|
| 4946 | page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
|
| 4947 | compact_priority, &compact_result);
|
| 4948 | if (page)
|
| 4949 | goto got_pg;
|
| 4950 |
|
| 4951 | if (compact_first) {
|
| 4952 | |
| 4953 | * THP page faults may attempt local node only first, but are
|
| 4954 | * then allowed to only compact, not reclaim, see
|
| 4955 | * alloc_pages_mpol().
|
| 4956 | *
|
| 4957 | * Compaction has failed above and we don't want such THP
|
| 4958 | * allocations to put reclaim pressure on a single node in a
|
| 4959 | * situation where other nodes might have plenty of available
|
| 4960 | * memory.
|
| 4961 |
|
| 4962 | if (gfp_has_flags(gfp_mask, __GFP_NORETRY | __GFP_THISNODE))
|
| 4963 | goto nopage;
|
| 4964 |
|
| 4965 | |
| 4966 | * For the initial compaction attempt we have lowered its
|
| 4967 | * priority. Restore it for further retries, if those are
|
| 4968 | * allowed. With __GFP_NORETRY there will be a single round of
|
| 4969 | * reclaim and compaction with the lowered priority.
|
| 4970 |
|
| 4971 | if (!(gfp_mask & __GFP_NORETRY))
|
| 4972 | compact_priority = DEF_COMPACT_PRIORITY;
|
| 4973 |
|
| 4974 | compact_first = false;
|
| 4975 | goto retry;
|
| 4976 | }
|
| 4977 |
|
| 4978 |
|
| 4979 | if (gfp_mask & __GFP_NORETRY)
|
| 4980 | goto nopage;
|
| 4981 |
|
| 4982 | |
| 4983 | * Do not retry costly high order allocations unless they are
|
| 4984 | * __GFP_RETRY_MAYFAIL and we can compact
|
| 4985 |
|
| 4986 | if (costly_order && (!can_compact ||
|
| 4987 | !(gfp_mask & __GFP_RETRY_MAYFAIL)))
|
| 4988 | goto nopage;
|
| 4989 |
|
| 4990 | |
| 4991 | * Deal with possible cpuset update races or zonelist updates to avoid
|
| 4992 | * infinite retries. No "goto retry;" can be placed above this check
|
| 4993 | * unless it can execute just once.
|
| 4994 |
|
| 4995 | if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
|
| 4996 | check_retry_zonelist(zonelist_iter_cookie))
|
| 4997 | goto restart;
|
| 4998 |
|
| 4999 | if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
|
| 5000 | did_some_progress > 0, &no_progress_loops))
|
| 5001 | goto retry;
|
| 5002 |
|
| 5003 | |
| 5004 | * It doesn't make any sense to retry for the compaction if the order-0
|
| 5005 | * reclaim is not able to make any progress because the current
|
| 5006 | * implementation of the compaction depends on the sufficient amount
|
| 5007 | * of free memory (see __compaction_suitable)
|
| 5008 |
|
| 5009 | if (did_some_progress > 0 && can_compact &&
|
| 5010 | should_compact_retry(gfp_mask, ac, order, alloc_flags,
|
| 5011 | compact_result, &compact_priority,
|
| 5012 | &compaction_retries))
|
| 5013 | goto retry;
|
| 5014 |
|
| 5015 |
|
| 5016 | if (defrag_mode && (alloc_flags & ALLOC_NOFRAGMENT)) {
|
| 5017 | alloc_flags &= ~ALLOC_NOFRAGMENT;
|
| 5018 | goto retry;
|
| 5019 | }
|
| 5020 |
|
| 5021 | |
| 5022 | * Deal with possible cpuset update races or zonelist updates to avoid
|
| 5023 | * a unnecessary OOM kill.
|
| 5024 |
|
| 5025 | if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
|
| 5026 | check_retry_zonelist(zonelist_iter_cookie))
|
| 5027 | goto restart;
|
| 5028 |
|
| 5029 |
|
| 5030 | page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
|
| 5031 | if (page)
|
| 5032 | goto got_pg;
|
| 5033 |
|
| 5034 |
|
| 5035 | if (tsk_is_oom_victim(current) &&
|
| 5036 | (alloc_flags & ALLOC_OOM ||
|
| 5037 | (gfp_mask & __GFP_NOMEMALLOC)))
|
| 5038 | goto nopage;
|
| 5039 |
|
| 5040 |
|
| 5041 | if (did_some_progress) {
|
| 5042 | no_progress_loops = 0;
|
| 5043 | goto retry;
|
| 5044 | }
|
| 5045 |
|
| 5046 | nopage:
|
| 5047 | |
| 5048 | * Deal with possible cpuset update races or zonelist updates to avoid
|
| 5049 | * a unnecessary OOM kill.
|
| 5050 |
|
| 5051 | if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
|
| 5052 | check_retry_zonelist(zonelist_iter_cookie))
|
| 5053 | goto restart;
|
| 5054 |
|
| 5055 | |
| 5056 | * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
|
| 5057 | * we always retry
|
| 5058 |
|
| 5059 | if (unlikely(nofail)) {
|
| 5060 | unsigned int alloc_flags = ac->alloc_flags | ALLOC_MIN_RESERVE;
|
| 5061 |
|
| 5062 | |
| 5063 | * Lacking direct_reclaim we can't do anything to reclaim memory,
|
| 5064 | * we disregard these unreasonable nofail requests and still
|
| 5065 | * return NULL
|
| 5066 |
|
| 5067 | if (!can_direct_reclaim)
|
| 5068 | goto fail;
|
| 5069 |
|
| 5070 | |
| 5071 | * Help non-failing allocations by giving some access to memory
|
| 5072 | * reserves normally used for high priority non-blocking
|
| 5073 | * allocations but do not use ALLOC_NO_WATERMARKS because this
|
| 5074 | * could deplete whole memory reserves which would just make
|
| 5075 | * the situation worse.
|
| 5076 |
|
| 5077 | page = __alloc_pages_cpuset_fallback(gfp_mask, order, alloc_flags, ac);
|
| 5078 | if (page)
|
| 5079 | goto got_pg;
|
| 5080 |
|
| 5081 | cond_resched();
|
| 5082 | goto retry;
|
| 5083 | }
|
| 5084 | fail:
|
| 5085 | warn_alloc(gfp_mask, ac->nodemask,
|
| 5086 | "page allocation failure: order:%u", order);
|
| 5087 | got_pg:
|
| 5088 | return page;
|
| 5089 | }
|
| 5090 |
|
| 5091 | static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
|
| 5092 | int preferred_nid, nodemask_t *nodemask,
|
| 5093 | struct alloc_context *ac, gfp_t *alloc_gfp,
|
| 5094 | unsigned int *alloc_flags)
|
| 5095 | {
|
| 5096 | ac->highest_zoneidx = gfp_zone(gfp_mask);
|
| 5097 | ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
|
| 5098 | ac->nodemask = nodemask;
|
| 5099 | ac->migratetype = gfp_migratetype(gfp_mask);
|
| 5100 |
|
| 5101 | if (cpusets_enabled()) {
|
| 5102 | *alloc_gfp |= __GFP_HARDWALL;
|
| 5103 | |
| 5104 | * When we are in the interrupt context, it is irrelevant
|
| 5105 | * to the current task context. It means that any node ok.
|
| 5106 |
|
| 5107 | if (in_task() && !ac->nodemask)
|
| 5108 | ac->nodemask = &cpuset_current_mems_allowed;
|
| 5109 | else
|
| 5110 | *alloc_flags |= ALLOC_CPUSET;
|
| 5111 | }
|
| 5112 |
|
| 5113 | might_alloc(gfp_mask);
|
| 5114 |
|
| 5115 | |
| 5116 | * Don't invoke should_fail logic, since it may call
|
| 5117 | * get_random_u32() and printk() which need to spin_lock.
|
| 5118 |
|
| 5119 | if (!(*alloc_flags & ALLOC_NOLOCK) &&
|
| 5120 | should_fail_alloc_page(gfp_mask, order))
|
| 5121 | return false;
|
| 5122 |
|
| 5123 | *alloc_flags |= alloc_flags_cma(gfp_mask);
|
| 5124 |
|
| 5125 |
|
| 5126 | ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
|
| 5127 |
|
| 5128 | |
| 5129 | * The preferred zone is used for statistics but crucially it is
|
| 5130 | * also used as the starting point for the zonelist iterator. It
|
| 5131 | * may get reset for allocations that ignore memory policies.
|
| 5132 |
|
| 5133 | ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
|
| 5134 | ac->highest_zoneidx, ac->nodemask);
|
| 5135 |
|
| 5136 | return true;
|
| 5137 | }
|
| 5138 |
|
| 5139 | |
| 5140 | * __alloc_pages_bulk - Allocate a number of order-0 pages to an array
|
| 5141 | * @gfp: GFP flags for the allocation
|
| 5142 | * @preferred_nid: The preferred NUMA node ID to allocate from
|
| 5143 | * @nodemask: Set of nodes to allocate from, may be NULL
|
| 5144 | * @nr_pages: The number of pages desired in the array
|
| 5145 | * @page_array: Array to store the pages
|
| 5146 | *
|
| 5147 | * This is a batched version of the page allocator that attempts to allocate
|
| 5148 | * @nr_pages quickly. Pages are added to @page_array.
|
| 5149 | *
|
| 5150 | * Note that only the elements in @page_array that were cleared to %NULL on
|
| 5151 | * entry are populated with newly allocated pages. @nr_pages is the maximum
|
| 5152 | * number of pages that will be stored in the array.
|
| 5153 | *
|
| 5154 | * Returns the number of pages in @page_array, including ones already
|
| 5155 | * allocated on entry. This can be less than the number requested in @nr_pages,
|
| 5156 | * but all empty slots are filled from the beginning. I.e., if all slots in
|
| 5157 | * @page_array were set to %NULL on entry, the slots from 0 to the return value
|
| 5158 | * - 1 will be filled.
|
| 5159 |
|
| 5160 | unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
|
| 5161 | nodemask_t *nodemask, int nr_pages,
|
| 5162 | struct page **page_array)
|
| 5163 | {
|
| 5164 | struct page *page;
|
| 5165 | struct zone *zone;
|
| 5166 | struct zoneref *z;
|
| 5167 | struct per_cpu_pages *pcp;
|
| 5168 | struct list_head *pcp_list;
|
| 5169 | struct alloc_context ac;
|
| 5170 | unsigned int alloc_flags = ALLOC_WMARK_LOW;
|
| 5171 | int nr_populated = 0, nr_account = 0;
|
| 5172 |
|
| 5173 | |
| 5174 | * Skip populated array elements to determine if any pages need
|
| 5175 | * to be allocated before disabling IRQs.
|
| 5176 |
|
| 5177 | while (nr_populated < nr_pages && page_array[nr_populated])
|
| 5178 | nr_populated++;
|
| 5179 |
|
| 5180 |
|
| 5181 | if (unlikely(nr_pages <= 0))
|
| 5182 | goto out;
|
| 5183 |
|
| 5184 |
|
| 5185 | if (unlikely(nr_pages - nr_populated == 0))
|
| 5186 | goto out;
|
| 5187 |
|
| 5188 |
|
| 5189 | if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
|
| 5190 | goto failed;
|
| 5191 |
|
| 5192 |
|
| 5193 | if (nr_pages - nr_populated == 1)
|
| 5194 | goto failed;
|
| 5195 |
|
| 5196 | #ifdef CONFIG_PAGE_OWNER
|
| 5197 | |
| 5198 | * PAGE_OWNER may recurse into the allocator to allocate space to
|
| 5199 | * save the stack with pagesets.lock held. Releasing/reacquiring
|
| 5200 | * removes much of the performance benefit of bulk allocation so
|
| 5201 | * force the caller to allocate one page at a time as it'll have
|
| 5202 | * similar performance to added complexity to the bulk allocator.
|
| 5203 |
|
| 5204 | if (static_branch_unlikely(&page_owner_inited))
|
| 5205 | goto failed;
|
| 5206 | #endif
|
| 5207 |
|
| 5208 |
|
| 5209 | gfp &= gfp_allowed_mask;
|
| 5210 | if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &gfp, &alloc_flags))
|
| 5211 | goto out;
|
| 5212 |
|
| 5213 |
|
| 5214 | z = ac.preferred_zoneref;
|
| 5215 | for_next_zone_zonelist_nodemask(zone, z, ac.highest_zoneidx, ac.nodemask) {
|
| 5216 | unsigned long mark;
|
| 5217 |
|
| 5218 | if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
|
| 5219 | !__cpuset_zone_allowed(zone, gfp)) {
|
| 5220 | continue;
|
| 5221 | }
|
| 5222 |
|
| 5223 | if (nr_online_nodes > 1 && zone != zonelist_zone(ac.preferred_zoneref) &&
|
| 5224 | zone_to_nid(zone) != zonelist_node_idx(ac.preferred_zoneref)) {
|
| 5225 | goto failed;
|
| 5226 | }
|
| 5227 |
|
| 5228 | cond_accept_memory(zone, 0, alloc_flags);
|
| 5229 | retry_this_zone:
|
| 5230 | mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages - nr_populated;
|
| 5231 | if (zone_watermark_fast(zone, 0, mark,
|
| 5232 | zonelist_zone_idx(ac.preferred_zoneref),
|
| 5233 | alloc_flags, gfp)) {
|
| 5234 | break;
|
| 5235 | }
|
| 5236 |
|
| 5237 | if (cond_accept_memory(zone, 0, alloc_flags))
|
| 5238 | goto retry_this_zone;
|
| 5239 |
|
| 5240 |
|
| 5241 | if (deferred_pages_enabled()) {
|
| 5242 | if (_deferred_grow_zone(zone, 0))
|
| 5243 | goto retry_this_zone;
|
| 5244 | }
|
| 5245 | }
|
| 5246 |
|
| 5247 | |
| 5248 | * If there are no allowed local zones that meets the watermarks then
|
| 5249 | * try to allocate a single page and reclaim if necessary.
|
| 5250 |
|
| 5251 | if (unlikely(!zone))
|
| 5252 | goto failed;
|
| 5253 |
|
| 5254 |
|
| 5255 | pcp = pcp_spin_trylock(zone->per_cpu_pageset);
|
| 5256 | if (!pcp)
|
| 5257 | goto failed;
|
| 5258 |
|
| 5259 |
|
| 5260 | pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
|
| 5261 | while (nr_populated < nr_pages) {
|
| 5262 |
|
| 5263 |
|
| 5264 | if (page_array[nr_populated]) {
|
| 5265 | nr_populated++;
|
| 5266 | continue;
|
| 5267 | }
|
| 5268 |
|
| 5269 | page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
|
| 5270 | pcp, pcp_list);
|
| 5271 | if (unlikely(!page)) {
|
| 5272 |
|
| 5273 | if (!nr_account) {
|
| 5274 | pcp_spin_unlock(pcp);
|
| 5275 | goto failed;
|
| 5276 | }
|
| 5277 | break;
|
| 5278 | }
|
| 5279 | nr_account++;
|
| 5280 |
|
| 5281 | prep_new_page(page, 0, gfp, ALLOC_DEFAULT);
|
| 5282 | set_page_refcounted(page);
|
| 5283 | page_array[nr_populated++] = page;
|
| 5284 | }
|
| 5285 |
|
| 5286 | pcp_spin_unlock(pcp);
|
| 5287 |
|
| 5288 | __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
|
| 5289 | zone_statistics(zonelist_zone(ac.preferred_zoneref), zone, nr_account);
|
| 5290 |
|
| 5291 | out:
|
| 5292 | return nr_populated;
|
| 5293 |
|
| 5294 | failed:
|
| 5295 | page = __alloc_pages_noprof(gfp, 0, preferred_nid, nodemask, ALLOC_DEFAULT);
|
| 5296 | if (page)
|
| 5297 | page_array[nr_populated++] = page;
|
| 5298 | goto out;
|
| 5299 | }
|
| 5300 | EXPORT_SYMBOL_GPL(alloc_pages_bulk_noprof);
|
| 5301 |
|
| 5302 | |
| 5303 | * free_pages_bulk - Free an array of order-0 pages
|
| 5304 | * @page_array: Array of pages to free
|
| 5305 | * @nr_pages: The number of pages in the array
|
| 5306 | *
|
| 5307 | * Free the order-0 pages. Adjacent entries whose PFNs form a contiguous
|
| 5308 | * run are released with a single __free_contig_range() call.
|
| 5309 | *
|
| 5310 | * This assumes page_array is sorted in ascending PFN order. Without that,
|
| 5311 | * the function still frees all pages, but contiguous runs may not be
|
| 5312 | * detected and the freeing pattern can degrade to freeing one page at a
|
| 5313 | * time.
|
| 5314 | *
|
| 5315 | * Context: Sleepable process context only; calls cond_resched()
|
| 5316 |
|
| 5317 | void free_pages_bulk(struct page **page_array, unsigned long nr_pages)
|
| 5318 | {
|
| 5319 | while (nr_pages) {
|
| 5320 | unsigned long nr_contig = num_pages_contiguous(page_array, nr_pages);
|
| 5321 |
|
| 5322 | __free_contig_range(page_to_pfn(*page_array), nr_contig);
|
| 5323 |
|
| 5324 | nr_pages -= nr_contig;
|
| 5325 | page_array += nr_contig;
|
| 5326 | cond_resched();
|
| 5327 | }
|
| 5328 | }
|
| 5329 |
|
| 5330 | static inline bool alloc_order_allowed(gfp_t gfp, unsigned int order,
|
| 5331 | unsigned int alloc_flags)
|
| 5332 | {
|
| 5333 | if (alloc_flags & ALLOC_NOLOCK)
|
| 5334 | return pcp_allowed_order(order);
|
| 5335 |
|
| 5336 | |
| 5337 | * There are several places where we assume that the order value is sane
|
| 5338 | * so bail out early if the request is out of bound.
|
| 5339 |
|
| 5340 | return !(WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp));
|
| 5341 | }
|
| 5342 |
|
| 5343 | static inline bool alloc_nolock_allowed(void)
|
| 5344 | {
|
| 5345 | if (!can_spin_trylock())
|
| 5346 | return false;
|
| 5347 |
|
| 5348 |
|
| 5349 | if (deferred_pages_enabled())
|
| 5350 | return false;
|
| 5351 |
|
| 5352 | return true;
|
| 5353 | }
|
| 5354 |
|
| 5355 | |
| 5356 | * GFP flags to set for ALLOC_NOLOCK i.e. alloc_pages_nolock().
|
| 5357 | *
|
| 5358 | * Do not specify __GFP_DIRECT_RECLAIM, since direct claim is not allowed.
|
| 5359 | * Do not specify __GFP_KSWAPD_RECLAIM either, since wake up of kswapd
|
| 5360 | * is not safe in arbitrary context.
|
| 5361 | *
|
| 5362 | * These two are the conditions for gfpflags_allow_spinning() being true.
|
| 5363 | *
|
| 5364 | * Specify __GFP_NOWARN since failing alloc_pages_nolock() is not a reason
|
| 5365 | * to warn. Also warn would trigger printk() which is unsafe from
|
| 5366 | * various contexts. We cannot use printk_deferred_enter() to mitigate,
|
| 5367 | * since the running context is unknown.
|
| 5368 | *
|
| 5369 | * Specify __GFP_ZERO to make sure that call to kmsan_alloc_page() below
|
| 5370 | * is safe in any context. Also zeroing the page is mandatory for
|
| 5371 | * BPF use cases.
|
| 5372 | *
|
| 5373 | * Though __GFP_NOMEMALLOC is not checked in the code path below,
|
| 5374 | * specify it here to highlight that alloc_pages_nolock()
|
| 5375 | * doesn't want to deplete reserves.
|
| 5376 |
|
| 5377 | static const gfp_t gfp_nolock = __GFP_NOWARN | __GFP_ZERO | __GFP_NOMEMALLOC |
|
| 5378 | __GFP_COMP;
|
| 5379 |
|
| 5380 | |
| 5381 | * This is the 'heart' of the zoned buddy allocator.
|
| 5382 |
|
| 5383 | struct page *__alloc_frozen_pages_noprof(gfp_t gfp, unsigned int order,
|
| 5384 | int preferred_nid, nodemask_t *nodemask, unsigned int alloc_flags)
|
| 5385 | {
|
| 5386 | struct page *page;
|
| 5387 | gfp_t alloc_gfp;
|
| 5388 | struct alloc_context ac = {
|
| 5389 | .alloc_flags = alloc_flags,
|
| 5390 | };
|
| 5391 | unsigned int fastpath_alloc_flags = alloc_flags;
|
| 5392 |
|
| 5393 |
|
| 5394 | if (WARN_ON(alloc_flags & ~(ALLOC_NOLOCK | ALLOC_NO_CODETAG)))
|
| 5395 | return NULL;
|
| 5396 |
|
| 5397 | if (!alloc_order_allowed(gfp, order, alloc_flags))
|
| 5398 | return NULL;
|
| 5399 |
|
| 5400 | if (alloc_flags & ALLOC_NOLOCK) {
|
| 5401 |
|
| 5402 | VM_WARN_ON_ONCE(gfp & ~(__GFP_ACCOUNT | gfp_nolock));
|
| 5403 | if (!alloc_nolock_allowed())
|
| 5404 | return NULL;
|
| 5405 | gfp |= gfp_nolock;
|
| 5406 | fastpath_alloc_flags |= ALLOC_WMARK_MIN;
|
| 5407 | } else {
|
| 5408 | fastpath_alloc_flags |= ALLOC_WMARK_LOW;
|
| 5409 | }
|
| 5410 |
|
| 5411 | gfp &= gfp_allowed_mask;
|
| 5412 | |
| 5413 | * Apply scoped allocation constraints. This is mainly about GFP_NOFS
|
| 5414 | * resp. GFP_NOIO which has to be inherited for all allocation requests
|
| 5415 | * from a particular context which has been marked by
|
| 5416 | * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
|
| 5417 | * movable zones are not used during allocation.
|
| 5418 |
|
| 5419 | gfp = current_gfp_context(gfp);
|
| 5420 | alloc_gfp = gfp;
|
| 5421 | if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
|
| 5422 | &alloc_gfp, &fastpath_alloc_flags))
|
| 5423 | return NULL;
|
| 5424 |
|
| 5425 | if (!(alloc_flags & ALLOC_NOLOCK)) {
|
| 5426 | |
| 5427 | * Forbid the first pass from falling back to types that
|
| 5428 | * fragment memory until all local zones are considered.
|
| 5429 |
|
| 5430 | fastpath_alloc_flags |= alloc_flags_nofragment(
|
| 5431 | zonelist_zone(ac.preferred_zoneref), gfp);
|
| 5432 | }
|
| 5433 | fastpath_alloc_flags |= alloc_flags_nonblocking(gfp, order) & ALLOC_HIGHATOMIC;
|
| 5434 |
|
| 5435 |
|
| 5436 | page = get_page_from_freelist(alloc_gfp, order, fastpath_alloc_flags, &ac);
|
| 5437 | if (likely(page) || (alloc_flags & ALLOC_NOLOCK))
|
| 5438 | goto out;
|
| 5439 |
|
| 5440 | alloc_gfp = gfp;
|
| 5441 | ac.spread_dirty_pages = false;
|
| 5442 |
|
| 5443 | |
| 5444 | * Restore the original nodemask if it was potentially replaced with
|
| 5445 | * &cpuset_current_mems_allowed to optimize the fast-path attempt.
|
| 5446 |
|
| 5447 | ac.nodemask = nodemask;
|
| 5448 |
|
| 5449 | page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
|
| 5450 |
|
| 5451 | out:
|
| 5452 | if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
|
| 5453 | unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
|
| 5454 | __free_frozen_pages(page, order,
|
| 5455 | alloc_flags & ALLOC_NOLOCK ? FPI_NOLOCK : 0);
|
| 5456 | page = NULL;
|
| 5457 | }
|
| 5458 |
|
| 5459 | trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
|
| 5460 | kmsan_alloc_page(page, order, alloc_gfp);
|
| 5461 |
|
| 5462 | return page;
|
| 5463 | }
|
| 5464 | EXPORT_SYMBOL(__alloc_frozen_pages_noprof);
|
| 5465 |
|
| 5466 | struct page *__alloc_pages_noprof(gfp_t gfp, unsigned int order,
|
| 5467 | int preferred_nid, nodemask_t *nodemask, unsigned int alloc_flags)
|
| 5468 | {
|
| 5469 | struct page *page;
|
| 5470 |
|
| 5471 | page = __alloc_frozen_pages_noprof(gfp, order, preferred_nid, nodemask,
|
| 5472 | alloc_flags);
|
| 5473 | if (page)
|
| 5474 | set_page_refcounted(page);
|
| 5475 | return page;
|
| 5476 | }
|
| 5477 |
|
| 5478 | struct page *alloc_pages_node_noprof(int nid, gfp_t gfp_mask, unsigned int order)
|
| 5479 | {
|
| 5480 | if (nid == NUMA_NO_NODE)
|
| 5481 | nid = numa_mem_id();
|
| 5482 |
|
| 5483 | warn_if_node_offline(nid, gfp_mask);
|
| 5484 |
|
| 5485 | return __alloc_pages_noprof(gfp_mask, order, nid, NULL, ALLOC_DEFAULT);
|
| 5486 | }
|
| 5487 | EXPORT_SYMBOL(alloc_pages_node_noprof);
|
| 5488 |
|
| 5489 | struct folio *__folio_alloc_noprof(gfp_t gfp, unsigned int order, int preferred_nid,
|
| 5490 | nodemask_t *nodemask)
|
| 5491 | {
|
| 5492 | struct page *page = __alloc_pages_noprof(gfp | __GFP_COMP, order,
|
| 5493 | preferred_nid, nodemask, ALLOC_DEFAULT);
|
| 5494 | return page_rmappable_folio(page);
|
| 5495 | }
|
| 5496 | EXPORT_SYMBOL(__folio_alloc_noprof);
|
| 5497 |
|
| 5498 | |
| 5499 | * Common helper functions. Never use with __GFP_HIGHMEM because the returned
|
| 5500 | * address cannot represent highmem pages. Use alloc_pages and then kmap if
|
| 5501 | * you need to access high mem.
|
| 5502 |
|
| 5503 | unsigned long get_free_pages_noprof(gfp_t gfp_mask, unsigned int order)
|
| 5504 | {
|
| 5505 | struct page *page;
|
| 5506 |
|
| 5507 | page = alloc_pages_noprof(gfp_mask & ~__GFP_HIGHMEM, order);
|
| 5508 | if (!page)
|
| 5509 | return 0;
|
| 5510 | return (unsigned long) page_address(page);
|
| 5511 | }
|
| 5512 | EXPORT_SYMBOL(get_free_pages_noprof);
|
| 5513 |
|
| 5514 | unsigned long get_zeroed_page_noprof(gfp_t gfp_mask)
|
| 5515 | {
|
| 5516 | return get_free_pages_noprof(gfp_mask | __GFP_ZERO, 0);
|
| 5517 | }
|
| 5518 | EXPORT_SYMBOL(get_zeroed_page_noprof);
|
| 5519 |
|
| 5520 | static void ___free_pages(struct page *page, unsigned int order,
|
| 5521 | fpi_t fpi_flags)
|
| 5522 | {
|
| 5523 |
|
| 5524 | int head = PageHead(page);
|
| 5525 |
|
| 5526 | struct alloc_tag *tag = pgalloc_tag_get(page);
|
| 5527 |
|
| 5528 | if (put_page_testzero(page))
|
| 5529 | __free_frozen_pages(page, order, fpi_flags);
|
| 5530 | else if (!head) {
|
| 5531 | pgalloc_tag_sub_pages(tag, (1 << order) - 1);
|
| 5532 | while (order-- > 0) {
|
| 5533 | |
| 5534 | * The "tail" pages of this non-compound high-order
|
| 5535 | * page will have no code tags, so to avoid warnings
|
| 5536 | * mark them as empty.
|
| 5537 |
|
| 5538 | clear_page_tag_ref(page + (1 << order));
|
| 5539 | __free_frozen_pages(page + (1 << order), order,
|
| 5540 | fpi_flags);
|
| 5541 | }
|
| 5542 | }
|
| 5543 | }
|
| 5544 |
|
| 5545 | |
| 5546 | * __free_pages - Free pages allocated with alloc_pages().
|
| 5547 | * @page: The page pointer returned from alloc_pages().
|
| 5548 | * @order: The order of the allocation.
|
| 5549 | *
|
| 5550 | * This function can free multi-page allocations that are not compound
|
| 5551 | * pages. It does not check that the @order passed in matches that of
|
| 5552 | * the allocation, so it is easy to leak memory. Freeing more memory
|
| 5553 | * than was allocated will probably emit a warning.
|
| 5554 | *
|
| 5555 | * If the last reference to this page is speculative, it will be released
|
| 5556 | * by put_page() which only frees the first page of a non-compound
|
| 5557 | * allocation. To prevent the remaining pages from being leaked, we free
|
| 5558 | * the subsequent pages here. If you want to use the page's reference
|
| 5559 | * count to decide when to free the allocation, you should allocate a
|
| 5560 | * compound page, and use put_page() instead of __free_pages().
|
| 5561 | *
|
| 5562 | * Context: May be called in interrupt context or while holding a normal
|
| 5563 | * spinlock, but not in NMI context or while holding a raw spinlock.
|
| 5564 |
|
| 5565 | void __free_pages(struct page *page, unsigned int order)
|
| 5566 | {
|
| 5567 | ___free_pages(page, order, FPI_NONE);
|
| 5568 | }
|
| 5569 | EXPORT_SYMBOL(__free_pages);
|
| 5570 |
|
| 5571 | |
| 5572 | * Can be called while holding raw_spin_lock or from IRQ and NMI for any
|
| 5573 | * page type (not only those that came from alloc_pages_nolock)
|
| 5574 |
|
| 5575 | void free_pages_nolock(struct page *page, unsigned int order)
|
| 5576 | {
|
| 5577 | ___free_pages(page, order, FPI_NOLOCK);
|
| 5578 | }
|
| 5579 |
|
| 5580 | |
| 5581 | * free_pages - Free pages allocated with __get_free_pages().
|
| 5582 | * @addr: The virtual address tied to a page returned from __get_free_pages().
|
| 5583 | * @order: The order of the allocation.
|
| 5584 | *
|
| 5585 | * This function behaves the same as __free_pages(). Use this function
|
| 5586 | * to free pages when you only have a valid virtual address. If you have
|
| 5587 | * the page, call __free_pages() instead.
|
| 5588 |
|
| 5589 | void free_pages(unsigned long addr, unsigned int order)
|
| 5590 | {
|
| 5591 | if (addr != 0) {
|
| 5592 | VM_BUG_ON(!virt_addr_valid((void *)addr));
|
| 5593 | __free_pages(virt_to_page((void *)addr), order);
|
| 5594 | }
|
| 5595 | }
|
| 5596 |
|
| 5597 | EXPORT_SYMBOL(free_pages);
|
| 5598 |
|
| 5599 | static void *make_alloc_exact(unsigned long addr, unsigned int order,
|
| 5600 | size_t size)
|
| 5601 | {
|
| 5602 | if (addr) {
|
| 5603 | unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
|
| 5604 | struct page *page = virt_to_page((void *)addr);
|
| 5605 | struct page *last = page + nr;
|
| 5606 |
|
| 5607 | __split_page(page, order);
|
| 5608 | while (page < --last)
|
| 5609 | set_page_refcounted(last);
|
| 5610 |
|
| 5611 | last = page + (1UL << order);
|
| 5612 | for (page += nr; page < last; page++)
|
| 5613 | __free_pages_ok(page, 0, FPI_TO_TAIL);
|
| 5614 | }
|
| 5615 | return (void *)addr;
|
| 5616 | }
|
| 5617 |
|
| 5618 | |
| 5619 | * alloc_pages_exact - allocate an exact number physically-contiguous pages.
|
| 5620 | * @size: the number of bytes to allocate
|
| 5621 | * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
|
| 5622 | *
|
| 5623 | * This function is similar to alloc_pages(), except that it allocates the
|
| 5624 | * minimum number of pages to satisfy the request. alloc_pages() can only
|
| 5625 | * allocate memory in power-of-two pages.
|
| 5626 | *
|
| 5627 | * This function is also limited by MAX_PAGE_ORDER.
|
| 5628 | *
|
| 5629 | * Memory allocated by this function must be released by free_pages_exact().
|
| 5630 | *
|
| 5631 | * Return: pointer to the allocated area or %NULL in case of error.
|
| 5632 |
|
| 5633 | void *alloc_pages_exact_noprof(size_t size, gfp_t gfp_mask)
|
| 5634 | {
|
| 5635 | unsigned int order = get_order(size);
|
| 5636 | unsigned long addr;
|
| 5637 |
|
| 5638 | if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
|
| 5639 | gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
|
| 5640 |
|
| 5641 | addr = get_free_pages_noprof(gfp_mask, order);
|
| 5642 | return make_alloc_exact(addr, order, size);
|
| 5643 | }
|
| 5644 | EXPORT_SYMBOL(alloc_pages_exact_noprof);
|
| 5645 |
|
| 5646 | |
| 5647 | * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
|
| 5648 | * pages on a node.
|
| 5649 | * @nid: the preferred node ID where memory should be allocated
|
| 5650 | * @size: the number of bytes to allocate
|
| 5651 | * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
|
| 5652 | *
|
| 5653 | * Like alloc_pages_exact(), but try to allocate on node nid first before falling
|
| 5654 | * back.
|
| 5655 | *
|
| 5656 | * Return: pointer to the allocated area or %NULL in case of error.
|
| 5657 |
|
| 5658 | void * __meminit alloc_pages_exact_nid_noprof(int nid, size_t size, gfp_t gfp_mask)
|
| 5659 | {
|
| 5660 | unsigned int order = get_order(size);
|
| 5661 | struct page *p;
|
| 5662 |
|
| 5663 | if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
|
| 5664 | gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
|
| 5665 |
|
| 5666 | p = alloc_pages_node_noprof(nid, gfp_mask, order);
|
| 5667 | if (!p)
|
| 5668 | return NULL;
|
| 5669 | return make_alloc_exact((unsigned long)page_address(p), order, size);
|
| 5670 | }
|
| 5671 |
|
| 5672 | |
| 5673 | * free_pages_exact - release memory allocated via alloc_pages_exact()
|
| 5674 | * @virt: the value returned by alloc_pages_exact.
|
| 5675 | * @size: size of allocation, same value as passed to alloc_pages_exact().
|
| 5676 | *
|
| 5677 | * Release the memory allocated by a previous call to alloc_pages_exact.
|
| 5678 |
|
| 5679 | void free_pages_exact(void *virt, size_t size)
|
| 5680 | {
|
| 5681 | unsigned long addr = (unsigned long)virt;
|
| 5682 | unsigned long end = addr + PAGE_ALIGN(size);
|
| 5683 |
|
| 5684 | while (addr < end) {
|
| 5685 | free_page(addr);
|
| 5686 | addr += PAGE_SIZE;
|
| 5687 | }
|
| 5688 | }
|
| 5689 | EXPORT_SYMBOL(free_pages_exact);
|
| 5690 |
|
| 5691 | |
| 5692 | * nr_free_zone_pages - count number of pages beyond high watermark
|
| 5693 | * @offset: The zone index of the highest zone
|
| 5694 | *
|
| 5695 | * nr_free_zone_pages() counts the number of pages which are beyond the
|
| 5696 | * high watermark within all zones at or below a given zone index. For each
|
| 5697 | * zone, the number of pages is calculated as:
|
| 5698 | *
|
| 5699 | * nr_free_zone_pages = managed_pages - high_pages
|
| 5700 | *
|
| 5701 | * Return: number of pages beyond high watermark.
|
| 5702 |
|
| 5703 | static unsigned long nr_free_zone_pages(int offset)
|
| 5704 | {
|
| 5705 | struct zoneref *z;
|
| 5706 | struct zone *zone;
|
| 5707 |
|
| 5708 |
|
| 5709 | unsigned long sum = 0;
|
| 5710 |
|
| 5711 | struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
|
| 5712 |
|
| 5713 | for_each_zone_zonelist(zone, z, zonelist, offset) {
|
| 5714 | unsigned long size = zone_managed_pages(zone);
|
| 5715 | unsigned long high = high_wmark_pages(zone);
|
| 5716 | if (size > high)
|
| 5717 | sum += size - high;
|
| 5718 | }
|
| 5719 |
|
| 5720 | return sum;
|
| 5721 | }
|
| 5722 |
|
| 5723 | |
| 5724 | * nr_free_buffer_pages - count number of pages beyond high watermark
|
| 5725 | *
|
| 5726 | * nr_free_buffer_pages() counts the number of pages which are beyond the high
|
| 5727 | * watermark within ZONE_DMA and ZONE_NORMAL.
|
| 5728 | *
|
| 5729 | * Return: number of pages beyond high watermark within ZONE_DMA and
|
| 5730 | * ZONE_NORMAL.
|
| 5731 |
|
| 5732 | unsigned long nr_free_buffer_pages(void)
|
| 5733 | {
|
| 5734 | return nr_free_zone_pages(gfp_zone(GFP_USER));
|
| 5735 | }
|
| 5736 | EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
|
| 5737 |
|
| 5738 | static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
|
| 5739 | {
|
| 5740 | zoneref->zone = zone;
|
| 5741 | zoneref->zone_idx = zone_idx(zone);
|
| 5742 | }
|
| 5743 |
|
| 5744 | |
| 5745 | * Builds allocation fallback zone lists.
|
| 5746 | *
|
| 5747 | * Add all populated zones of a node to the zonelist.
|
| 5748 |
|
| 5749 | static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
|
| 5750 | {
|
| 5751 | struct zone *zone;
|
| 5752 | enum zone_type zone_type = MAX_NR_ZONES;
|
| 5753 | int nr_zones = 0;
|
| 5754 |
|
| 5755 | do {
|
| 5756 | zone_type--;
|
| 5757 | zone = pgdat->node_zones + zone_type;
|
| 5758 | if (populated_zone(zone)) {
|
| 5759 | zoneref_set_zone(zone, &zonerefs[nr_zones++]);
|
| 5760 | check_highest_zone(zone_type);
|
| 5761 | }
|
| 5762 | } while (zone_type);
|
| 5763 |
|
| 5764 | return nr_zones;
|
| 5765 | }
|
| 5766 |
|
| 5767 | #ifdef CONFIG_NUMA
|
| 5768 |
|
| 5769 | static int __parse_numa_zonelist_order(char *s)
|
| 5770 | {
|
| 5771 | |
| 5772 | * We used to support different zonelists modes but they turned
|
| 5773 | * out to be just not useful. Let's keep the warning in place
|
| 5774 | * if somebody still use the cmd line parameter so that we do
|
| 5775 | * not fail it silently
|
| 5776 |
|
| 5777 | if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
|
| 5778 | pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
|
| 5779 | return -EINVAL;
|
| 5780 | }
|
| 5781 | return 0;
|
| 5782 | }
|
| 5783 |
|
| 5784 | static char numa_zonelist_order[] = "Node";
|
| 5785 | #define NUMA_ZONELIST_ORDER_LEN 16
|
| 5786 | |
| 5787 | * sysctl handler for numa_zonelist_order
|
| 5788 |
|
| 5789 | static int numa_zonelist_order_handler(const struct ctl_table *table, int write,
|
| 5790 | void *buffer, size_t *length, loff_t *ppos)
|
| 5791 | {
|
| 5792 | if (write)
|
| 5793 | return __parse_numa_zonelist_order(buffer);
|
| 5794 | return proc_dostring(table, write, buffer, length, ppos);
|
| 5795 | }
|
| 5796 |
|
| 5797 | static int node_load[MAX_NUMNODES];
|
| 5798 |
|
| 5799 | |
| 5800 | * find_next_best_node - find the next node that should appear in a given node's fallback list
|
| 5801 | * @node: node whose fallback list we're appending
|
| 5802 | * @used_node_mask: nodemask_t of already used nodes
|
| 5803 | *
|
| 5804 | * We use a number of factors to determine which is the next node that should
|
| 5805 | * appear on a given node's fallback list. The node should not have appeared
|
| 5806 | * already in @node's fallback list, and it should be the next closest node
|
| 5807 | * according to the distance array (which contains arbitrary distance values
|
| 5808 | * from each node to each node in the system), and should also prefer nodes
|
| 5809 | * with no CPUs, since presumably they'll have very little allocation pressure
|
| 5810 | * on them otherwise.
|
| 5811 | *
|
| 5812 | * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
|
| 5813 |
|
| 5814 | int find_next_best_node(int node, nodemask_t *used_node_mask)
|
| 5815 | {
|
| 5816 | int n, val;
|
| 5817 | int min_val = INT_MAX;
|
| 5818 | int best_node = NUMA_NO_NODE;
|
| 5819 |
|
| 5820 | |
| 5821 | * Use the local node if we haven't already, but for memoryless local
|
| 5822 | * node, we should skip it and fall back to other nodes.
|
| 5823 |
|
| 5824 | if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
|
| 5825 | node_set(node, *used_node_mask);
|
| 5826 | return node;
|
| 5827 | }
|
| 5828 |
|
| 5829 | for_each_node_state(n, N_MEMORY) {
|
| 5830 |
|
| 5831 |
|
| 5832 | if (node_isset(n, *used_node_mask))
|
| 5833 | continue;
|
| 5834 |
|
| 5835 |
|
| 5836 | val = node_distance(node, n);
|
| 5837 |
|
| 5838 |
|
| 5839 | val += (n < node);
|
| 5840 |
|
| 5841 |
|
| 5842 | if (!cpumask_empty(cpumask_of_node(n)))
|
| 5843 | val += PENALTY_FOR_NODE_WITH_CPUS;
|
| 5844 |
|
| 5845 |
|
| 5846 | val *= MAX_NUMNODES;
|
| 5847 | val += node_load[n];
|
| 5848 |
|
| 5849 | if (val < min_val) {
|
| 5850 | min_val = val;
|
| 5851 | best_node = n;
|
| 5852 | }
|
| 5853 | }
|
| 5854 |
|
| 5855 | if (best_node >= 0)
|
| 5856 | node_set(best_node, *used_node_mask);
|
| 5857 |
|
| 5858 | return best_node;
|
| 5859 | }
|
| 5860 |
|
| 5861 |
|
| 5862 | |
| 5863 | * Build zonelists ordered by node and zones within node.
|
| 5864 | * This results in maximum locality--normal zone overflows into local
|
| 5865 | * DMA zone, if any--but risks exhausting DMA zone.
|
| 5866 |
|
| 5867 | static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
|
| 5868 | unsigned nr_nodes)
|
| 5869 | {
|
| 5870 | struct zoneref *zonerefs;
|
| 5871 | int i;
|
| 5872 |
|
| 5873 | zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
|
| 5874 |
|
| 5875 | for (i = 0; i < nr_nodes; i++) {
|
| 5876 | int nr_zones;
|
| 5877 |
|
| 5878 | pg_data_t *node = NODE_DATA(node_order[i]);
|
| 5879 |
|
| 5880 | nr_zones = build_zonerefs_node(node, zonerefs);
|
| 5881 | zonerefs += nr_zones;
|
| 5882 | }
|
| 5883 | zonerefs->zone = NULL;
|
| 5884 | zonerefs->zone_idx = 0;
|
| 5885 | }
|
| 5886 |
|
| 5887 | |
| 5888 | * Build __GFP_THISNODE zonelists
|
| 5889 |
|
| 5890 | static void build_thisnode_zonelists(pg_data_t *pgdat)
|
| 5891 | {
|
| 5892 | struct zoneref *zonerefs;
|
| 5893 | int nr_zones;
|
| 5894 |
|
| 5895 | zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
|
| 5896 | nr_zones = build_zonerefs_node(pgdat, zonerefs);
|
| 5897 | zonerefs += nr_zones;
|
| 5898 | zonerefs->zone = NULL;
|
| 5899 | zonerefs->zone_idx = 0;
|
| 5900 | }
|
| 5901 |
|
| 5902 | static void build_zonelists(pg_data_t *pgdat)
|
| 5903 | {
|
| 5904 | static int node_order[MAX_NUMNODES];
|
| 5905 | int node, nr_nodes = 0;
|
| 5906 | nodemask_t used_mask = NODE_MASK_NONE;
|
| 5907 | int local_node, prev_node;
|
| 5908 |
|
| 5909 |
|
| 5910 | local_node = pgdat->node_id;
|
| 5911 | prev_node = local_node;
|
| 5912 |
|
| 5913 | memset(node_order, 0, sizeof(node_order));
|
| 5914 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
|
| 5915 | |
| 5916 | * We don't want to pressure a particular node.
|
| 5917 | * So adding penalty to the first node in same
|
| 5918 | * distance group to make it round-robin.
|
| 5919 |
|
| 5920 | if (node_distance(local_node, node) !=
|
| 5921 | node_distance(local_node, prev_node))
|
| 5922 | node_load[node] += 1;
|
| 5923 |
|
| 5924 | node_order[nr_nodes++] = node;
|
| 5925 | prev_node = node;
|
| 5926 | }
|
| 5927 |
|
| 5928 | build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
|
| 5929 | build_thisnode_zonelists(pgdat);
|
| 5930 | pr_info("Fallback order for Node %d: ", local_node);
|
| 5931 | for (node = 0; node < nr_nodes; node++)
|
| 5932 | pr_cont("%d ", node_order[node]);
|
| 5933 | pr_cont("\n");
|
| 5934 | }
|
| 5935 |
|
| 5936 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES
|
| 5937 | |
| 5938 | * Return node id of node used for "local" allocations.
|
| 5939 | * I.e., first node id of first zone in arg node's generic zonelist.
|
| 5940 | * Used for initializing percpu 'numa_mem', which is used primarily
|
| 5941 | * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
|
| 5942 |
|
| 5943 | int local_memory_node(int node)
|
| 5944 | {
|
| 5945 | struct zoneref *z;
|
| 5946 |
|
| 5947 | z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
|
| 5948 | gfp_zone(GFP_KERNEL),
|
| 5949 | NULL);
|
| 5950 | return zonelist_node_idx(z);
|
| 5951 | }
|
| 5952 | #endif
|
| 5953 |
|
| 5954 | static void setup_min_unmapped_ratio(void);
|
| 5955 | static void setup_min_slab_ratio(void);
|
| 5956 | #else
|
| 5957 |
|
| 5958 | static void build_zonelists(pg_data_t *pgdat)
|
| 5959 | {
|
| 5960 | struct zoneref *zonerefs;
|
| 5961 | int nr_zones;
|
| 5962 |
|
| 5963 | zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
|
| 5964 | nr_zones = build_zonerefs_node(pgdat, zonerefs);
|
| 5965 | zonerefs += nr_zones;
|
| 5966 |
|
| 5967 | zonerefs->zone = NULL;
|
| 5968 | zonerefs->zone_idx = 0;
|
| 5969 | }
|
| 5970 |
|
| 5971 | #endif
|
| 5972 |
|
| 5973 | |
| 5974 | * Boot pageset table. One per cpu which is going to be used for all
|
| 5975 | * zones and all nodes. The parameters will be set in such a way
|
| 5976 | * that an item put on a list will immediately be handed over to
|
| 5977 | * the buddy list. This is safe since pageset manipulation is done
|
| 5978 | * with interrupts disabled.
|
| 5979 | *
|
| 5980 | * The boot_pagesets must be kept even after bootup is complete for
|
| 5981 | * unused processors and/or zones. They do play a role for bootstrapping
|
| 5982 | * hotplugged processors.
|
| 5983 | *
|
| 5984 | * zoneinfo_show() and maybe other functions do
|
| 5985 | * not check if the processor is online before following the pageset pointer.
|
| 5986 | * Other parts of the kernel may not check if the zone is available.
|
| 5987 |
|
| 5988 | static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
|
| 5989 |
|
| 5990 | #define BOOT_PAGESET_HIGH 0
|
| 5991 | #define BOOT_PAGESET_BATCH 1
|
| 5992 | static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
|
| 5993 | static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
|
| 5994 |
|
| 5995 | static void __build_all_zonelists(void *data)
|
| 5996 | {
|
| 5997 | int nid;
|
| 5998 | int __maybe_unused cpu;
|
| 5999 | pg_data_t *self = data;
|
| 6000 | unsigned long flags;
|
| 6001 |
|
| 6002 | |
| 6003 | * The zonelist_update_seq must be acquired with irqsave because the
|
| 6004 | * reader can be invoked from IRQ with GFP_ATOMIC.
|
| 6005 |
|
| 6006 | write_seqlock_irqsave(&zonelist_update_seq, flags);
|
| 6007 | |
| 6008 | * Also disable synchronous printk() to prevent any printk() from
|
| 6009 | * trying to hold port->lock, for
|
| 6010 | * tty_insert_flip_string_and_push_buffer() on other CPU might be
|
| 6011 | * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
|
| 6012 |
|
| 6013 | printk_deferred_enter();
|
| 6014 |
|
| 6015 | #ifdef CONFIG_NUMA
|
| 6016 | memset(node_load, 0, sizeof(node_load));
|
| 6017 | #endif
|
| 6018 |
|
| 6019 | |
| 6020 | * This node is hotadded and no memory is yet present. So just
|
| 6021 | * building zonelists is fine - no need to touch other nodes.
|
| 6022 |
|
| 6023 | if (self && !node_online(self->node_id)) {
|
| 6024 | build_zonelists(self);
|
| 6025 | } else {
|
| 6026 | |
| 6027 | * All possible nodes have pgdat preallocated
|
| 6028 | * in free_area_init
|
| 6029 |
|
| 6030 | for_each_node(nid) {
|
| 6031 | pg_data_t *pgdat = NODE_DATA(nid);
|
| 6032 |
|
| 6033 | build_zonelists(pgdat);
|
| 6034 | }
|
| 6035 |
|
| 6036 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES
|
| 6037 | |
| 6038 | * We now know the "local memory node" for each node--
|
| 6039 | * i.e., the node of the first zone in the generic zonelist.
|
| 6040 | * Set up numa_mem percpu variable for on-line cpus. During
|
| 6041 | * boot, only the boot cpu should be on-line; we'll init the
|
| 6042 | * secondary cpus' numa_mem as they come on-line. During
|
| 6043 | * node/memory hotplug, we'll fixup all on-line cpus.
|
| 6044 |
|
| 6045 | for_each_online_cpu(cpu)
|
| 6046 | set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
|
| 6047 | #endif
|
| 6048 | }
|
| 6049 |
|
| 6050 | printk_deferred_exit();
|
| 6051 | write_sequnlock_irqrestore(&zonelist_update_seq, flags);
|
| 6052 | }
|
| 6053 |
|
| 6054 | static noinline void __init
|
| 6055 | build_all_zonelists_init(void)
|
| 6056 | {
|
| 6057 | int cpu;
|
| 6058 |
|
| 6059 | __build_all_zonelists(NULL);
|
| 6060 |
|
| 6061 | |
| 6062 | * Initialize the boot_pagesets that are going to be used
|
| 6063 | * for bootstrapping processors. The real pagesets for
|
| 6064 | * each zone will be allocated later when the per cpu
|
| 6065 | * allocator is available.
|
| 6066 | *
|
| 6067 | * boot_pagesets are used also for bootstrapping offline
|
| 6068 | * cpus if the system is already booted because the pagesets
|
| 6069 | * are needed to initialize allocators on a specific cpu too.
|
| 6070 | * F.e. the percpu allocator needs the page allocator which
|
| 6071 | * needs the percpu allocator in order to allocate its pagesets
|
| 6072 | * (a chicken-egg dilemma).
|
| 6073 |
|
| 6074 | for_each_possible_cpu(cpu)
|
| 6075 | per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
|
| 6076 |
|
| 6077 | mminit_verify_zonelist();
|
| 6078 | cpuset_init_current_mems_allowed();
|
| 6079 | }
|
| 6080 |
|
| 6081 | |
| 6082 | * unless system_state == SYSTEM_BOOTING.
|
| 6083 | *
|
| 6084 | * __ref due to call of __init annotated helper build_all_zonelists_init
|
| 6085 | * [protected by SYSTEM_BOOTING].
|
| 6086 |
|
| 6087 | void __ref build_all_zonelists(pg_data_t *pgdat)
|
| 6088 | {
|
| 6089 | unsigned long vm_total_pages;
|
| 6090 |
|
| 6091 | if (system_state == SYSTEM_BOOTING) {
|
| 6092 | build_all_zonelists_init();
|
| 6093 | } else {
|
| 6094 | __build_all_zonelists(pgdat);
|
| 6095 |
|
| 6096 | }
|
| 6097 |
|
| 6098 | vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
|
| 6099 | |
| 6100 | * Disable grouping by mobility if the number of pages in the
|
| 6101 | * system is too low to allow the mechanism to work. It would be
|
| 6102 | * more accurate, but expensive to check per-zone. This check is
|
| 6103 | * made on memory-hotadd so a system can start with mobility
|
| 6104 | * disabled and enable it later
|
| 6105 |
|
| 6106 | if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
|
| 6107 | page_group_by_mobility_disabled = 1;
|
| 6108 | else
|
| 6109 | page_group_by_mobility_disabled = 0;
|
| 6110 |
|
| 6111 | pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
|
| 6112 | nr_online_nodes,
|
| 6113 | str_off_on(page_group_by_mobility_disabled),
|
| 6114 | vm_total_pages);
|
| 6115 | #ifdef CONFIG_NUMA
|
| 6116 | pr_info("Policy zone: %s\n", zone_names[policy_zone]);
|
| 6117 | #endif
|
| 6118 | }
|
| 6119 |
|
| 6120 | static int zone_batchsize(struct zone *zone)
|
| 6121 | {
|
| 6122 | #ifdef CONFIG_MMU
|
| 6123 | int batch;
|
| 6124 |
|
| 6125 | |
| 6126 | * The number of pages to batch allocate is either ~0.025%
|
| 6127 | * of the zone or 256KB, whichever is smaller. The batch
|
| 6128 | * size is striking a balance between allocation latency
|
| 6129 | * and zone lock contention.
|
| 6130 |
|
| 6131 | batch = min(zone_managed_pages(zone) >> 12, SZ_256K / PAGE_SIZE);
|
| 6132 | if (batch <= 1)
|
| 6133 | return 1;
|
| 6134 |
|
| 6135 | |
| 6136 | * Clamp the batch to a 2^n - 1 value. Having a power
|
| 6137 | * of 2 value was found to be more likely to have
|
| 6138 | * suboptimal cache aliasing properties in some cases.
|
| 6139 | *
|
| 6140 | * For example if 2 tasks are alternately allocating
|
| 6141 | * batches of pages, one task can end up with a lot
|
| 6142 | * of pages of one half of the possible page colors
|
| 6143 | * and the other with pages of the other colors.
|
| 6144 |
|
| 6145 | batch = rounddown_pow_of_two(batch + batch/2) - 1;
|
| 6146 |
|
| 6147 | return batch;
|
| 6148 |
|
| 6149 | #else
|
| 6150 | |
| 6151 | * conditions.
|
| 6152 | *
|
| 6153 | * The problem is that NOMMU needs to be able to allocate large chunks
|
| 6154 | * of contiguous memory as there's no hardware page translation to
|
| 6155 | * assemble apparent contiguous memory from discontiguous pages.
|
| 6156 | *
|
| 6157 | * Queueing large contiguous runs of pages for batching, however,
|
| 6158 | * causes the pages to actually be freed in smaller chunks. As there
|
| 6159 | * can be a significant delay between the individual batches being
|
| 6160 | * recycled, this leads to the once large chunks of space being
|
| 6161 | * fragmented and becoming unavailable for high-order allocations.
|
| 6162 |
|
| 6163 | return 1;
|
| 6164 | #endif
|
| 6165 | }
|
| 6166 |
|
| 6167 | static int percpu_pagelist_high_fraction;
|
| 6168 | static int zone_highsize(struct zone *zone, int batch, int cpu_online,
|
| 6169 | int high_fraction)
|
| 6170 | {
|
| 6171 | #ifdef CONFIG_MMU
|
| 6172 | int high;
|
| 6173 | int nr_split_cpus;
|
| 6174 | unsigned long total_pages;
|
| 6175 |
|
| 6176 | if (!high_fraction) {
|
| 6177 | |
| 6178 | * By default, the high value of the pcp is based on the zone
|
| 6179 | * low watermark so that if they are full then background
|
| 6180 | * reclaim will not be started prematurely.
|
| 6181 |
|
| 6182 | total_pages = low_wmark_pages(zone);
|
| 6183 | } else {
|
| 6184 | |
| 6185 | * If percpu_pagelist_high_fraction is configured, the high
|
| 6186 | * value is based on a fraction of the managed pages in the
|
| 6187 | * zone.
|
| 6188 |
|
| 6189 | total_pages = zone_managed_pages(zone) / high_fraction;
|
| 6190 | }
|
| 6191 |
|
| 6192 | |
| 6193 | * Split the high value across all online CPUs local to the zone. Note
|
| 6194 | * that early in boot that CPUs may not be online yet and that during
|
| 6195 | * CPU hotplug that the cpumask is not yet updated when a CPU is being
|
| 6196 | * onlined. For memory nodes that have no CPUs, split the high value
|
| 6197 | * across all online CPUs to mitigate the risk that reclaim is triggered
|
| 6198 | * prematurely due to pages stored on pcp lists.
|
| 6199 |
|
| 6200 | nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
|
| 6201 | if (!nr_split_cpus)
|
| 6202 | nr_split_cpus = num_online_cpus();
|
| 6203 | high = total_pages / nr_split_cpus;
|
| 6204 |
|
| 6205 | |
| 6206 | * Ensure high is at least batch*4. The multiple is based on the
|
| 6207 | * historical relationship between high and batch.
|
| 6208 |
|
| 6209 | high = max(high, batch << 2);
|
| 6210 |
|
| 6211 | return high;
|
| 6212 | #else
|
| 6213 | return 0;
|
| 6214 | #endif
|
| 6215 | }
|
| 6216 |
|
| 6217 | |
| 6218 | * pcp->high and pcp->batch values are related and generally batch is lower
|
| 6219 | * than high. They are also related to pcp->count such that count is lower
|
| 6220 | * than high, and as soon as it reaches high, the pcplist is flushed.
|
| 6221 | *
|
| 6222 | * However, guaranteeing these relations at all times would require e.g. write
|
| 6223 | * barriers here but also careful usage of read barriers at the read side, and
|
| 6224 | * thus be prone to error and bad for performance. Thus the update only prevents
|
| 6225 | * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
|
| 6226 | * should ensure they can cope with those fields changing asynchronously, and
|
| 6227 | * fully trust only the pcp->count field on the local CPU with interrupts
|
| 6228 | * disabled.
|
| 6229 | *
|
| 6230 | * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
|
| 6231 | * outside of boot time (or some other assurance that no concurrent updaters
|
| 6232 | * exist).
|
| 6233 |
|
| 6234 | static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
|
| 6235 | unsigned long high_max, unsigned long batch)
|
| 6236 | {
|
| 6237 | WRITE_ONCE(pcp->batch, batch);
|
| 6238 | WRITE_ONCE(pcp->high_min, high_min);
|
| 6239 | WRITE_ONCE(pcp->high_max, high_max);
|
| 6240 | }
|
| 6241 |
|
| 6242 | static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
|
| 6243 | {
|
| 6244 | int pindex;
|
| 6245 |
|
| 6246 | memset(pcp, 0, sizeof(*pcp));
|
| 6247 | memset(pzstats, 0, sizeof(*pzstats));
|
| 6248 |
|
| 6249 | spin_lock_init(&pcp->lock);
|
| 6250 | for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
|
| 6251 | INIT_LIST_HEAD(&pcp->lists[pindex]);
|
| 6252 |
|
| 6253 | |
| 6254 | * Set batch and high values safe for a boot pageset. A true percpu
|
| 6255 | * pageset's initialization will update them subsequently. Here we don't
|
| 6256 | * need to be as careful as pageset_update() as nobody can access the
|
| 6257 | * pageset yet.
|
| 6258 |
|
| 6259 | pcp->high_min = BOOT_PAGESET_HIGH;
|
| 6260 | pcp->high_max = BOOT_PAGESET_HIGH;
|
| 6261 | pcp->batch = BOOT_PAGESET_BATCH;
|
| 6262 | }
|
| 6263 |
|
| 6264 | static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
|
| 6265 | unsigned long high_max, unsigned long batch)
|
| 6266 | {
|
| 6267 | struct per_cpu_pages *pcp;
|
| 6268 | int cpu;
|
| 6269 |
|
| 6270 | for_each_possible_cpu(cpu) {
|
| 6271 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
|
| 6272 | pageset_update(pcp, high_min, high_max, batch);
|
| 6273 | }
|
| 6274 | }
|
| 6275 |
|
| 6276 | |
| 6277 | * Calculate and set new high and batch values for all per-cpu pagesets of a
|
| 6278 | * zone based on the zone's size.
|
| 6279 |
|
| 6280 | static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
|
| 6281 | {
|
| 6282 | int new_high_min, new_high_max, new_batch;
|
| 6283 |
|
| 6284 | new_batch = zone_batchsize(zone);
|
| 6285 | if (percpu_pagelist_high_fraction) {
|
| 6286 | new_high_min = zone_highsize(zone, new_batch, cpu_online,
|
| 6287 | percpu_pagelist_high_fraction);
|
| 6288 | |
| 6289 | * PCP high is tuned manually, disable auto-tuning via
|
| 6290 | * setting high_min and high_max to the manual value.
|
| 6291 |
|
| 6292 | new_high_max = new_high_min;
|
| 6293 | } else {
|
| 6294 | new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
|
| 6295 | new_high_max = zone_highsize(zone, new_batch, cpu_online,
|
| 6296 | MIN_PERCPU_PAGELIST_HIGH_FRACTION);
|
| 6297 | }
|
| 6298 |
|
| 6299 | if (zone->pageset_high_min == new_high_min &&
|
| 6300 | zone->pageset_high_max == new_high_max &&
|
| 6301 | zone->pageset_batch == new_batch)
|
| 6302 | return;
|
| 6303 |
|
| 6304 | zone->pageset_high_min = new_high_min;
|
| 6305 | zone->pageset_high_max = new_high_max;
|
| 6306 | zone->pageset_batch = new_batch;
|
| 6307 |
|
| 6308 | __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
|
| 6309 | new_batch);
|
| 6310 | }
|
| 6311 |
|
| 6312 | void __meminit setup_zone_pageset(struct zone *zone)
|
| 6313 | {
|
| 6314 | int cpu;
|
| 6315 |
|
| 6316 |
|
| 6317 | if (sizeof(struct per_cpu_zonestat) > 0)
|
| 6318 | zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
|
| 6319 |
|
| 6320 | zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
|
| 6321 | for_each_possible_cpu(cpu) {
|
| 6322 | struct per_cpu_pages *pcp;
|
| 6323 | struct per_cpu_zonestat *pzstats;
|
| 6324 |
|
| 6325 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
|
| 6326 | pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
|
| 6327 | per_cpu_pages_init(pcp, pzstats);
|
| 6328 | }
|
| 6329 |
|
| 6330 | zone_set_pageset_high_and_batch(zone, 0);
|
| 6331 | }
|
| 6332 |
|
| 6333 | |
| 6334 | * The zone indicated has a new number of managed_pages; batch sizes and percpu
|
| 6335 | * page high values need to be recalculated.
|
| 6336 |
|
| 6337 | static void zone_pcp_update(struct zone *zone, int cpu_online)
|
| 6338 | {
|
| 6339 | mutex_lock(&pcp_batch_high_lock);
|
| 6340 | zone_set_pageset_high_and_batch(zone, cpu_online);
|
| 6341 | mutex_unlock(&pcp_batch_high_lock);
|
| 6342 | }
|
| 6343 |
|
| 6344 | static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu)
|
| 6345 | {
|
| 6346 | struct per_cpu_pages *pcp;
|
| 6347 | struct cpu_cacheinfo *cci;
|
| 6348 |
|
| 6349 | pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
|
| 6350 | cci = get_cpu_cacheinfo(cpu);
|
| 6351 | |
| 6352 | * If data cache slice of CPU is large enough, "pcp->batch"
|
| 6353 | * pages can be preserved in PCP before draining PCP for
|
| 6354 | * consecutive high-order pages freeing without allocation.
|
| 6355 | * This can reduce zone lock contention without hurting
|
| 6356 | * cache-hot pages sharing.
|
| 6357 |
|
| 6358 | pcp_spin_lock_nopin(pcp);
|
| 6359 | if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
|
| 6360 | pcp->flags |= PCPF_FREE_HIGH_BATCH;
|
| 6361 | else
|
| 6362 | pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
|
| 6363 | pcp_spin_unlock_nopin(pcp);
|
| 6364 | }
|
| 6365 |
|
| 6366 | void setup_pcp_cacheinfo(unsigned int cpu)
|
| 6367 | {
|
| 6368 | struct zone *zone;
|
| 6369 |
|
| 6370 | for_each_populated_zone(zone)
|
| 6371 | zone_pcp_update_cacheinfo(zone, cpu);
|
| 6372 | }
|
| 6373 |
|
| 6374 | |
| 6375 | * Allocate per cpu pagesets and initialize them.
|
| 6376 | * Before this call only boot pagesets were available.
|
| 6377 |
|
| 6378 | void __init setup_per_cpu_pageset(void)
|
| 6379 | {
|
| 6380 | struct pglist_data *pgdat;
|
| 6381 | struct zone *zone;
|
| 6382 | int __maybe_unused cpu;
|
| 6383 |
|
| 6384 | for_each_populated_zone(zone)
|
| 6385 | setup_zone_pageset(zone);
|
| 6386 |
|
| 6387 | #ifdef CONFIG_NUMA
|
| 6388 | |
| 6389 | * Unpopulated zones continue using the boot pagesets.
|
| 6390 | * The numa stats for these pagesets need to be reset.
|
| 6391 | * Otherwise, they will end up skewing the stats of
|
| 6392 | * the nodes these zones are associated with.
|
| 6393 |
|
| 6394 | for_each_possible_cpu(cpu) {
|
| 6395 | struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
|
| 6396 | memset(pzstats->vm_numa_event, 0,
|
| 6397 | sizeof(pzstats->vm_numa_event));
|
| 6398 | }
|
| 6399 | #endif
|
| 6400 |
|
| 6401 | for_each_online_pgdat(pgdat)
|
| 6402 | pgdat->per_cpu_nodestats =
|
| 6403 | alloc_percpu(struct per_cpu_nodestat);
|
| 6404 | }
|
| 6405 |
|
| 6406 | __meminit void zone_pcp_init(struct zone *zone)
|
| 6407 | {
|
| 6408 | |
| 6409 | * per cpu subsystem is not up at this point. The following code
|
| 6410 | * relies on the ability of the linker to provide the
|
| 6411 | * offset of a (static) per cpu variable into the per cpu area.
|
| 6412 |
|
| 6413 | zone->per_cpu_pageset = &boot_pageset;
|
| 6414 | zone->per_cpu_zonestats = &boot_zonestats;
|
| 6415 | zone->pageset_high_min = BOOT_PAGESET_HIGH;
|
| 6416 | zone->pageset_high_max = BOOT_PAGESET_HIGH;
|
| 6417 | zone->pageset_batch = BOOT_PAGESET_BATCH;
|
| 6418 |
|
| 6419 | if (populated_zone(zone))
|
| 6420 | pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
|
| 6421 | zone->present_pages, zone_batchsize(zone));
|
| 6422 | }
|
| 6423 |
|
| 6424 | static void setup_per_zone_lowmem_reserve(void);
|
| 6425 |
|
| 6426 | void adjust_managed_page_count(struct page *page, long count)
|
| 6427 | {
|
| 6428 | atomic_long_add(count, &page_zone(page)->managed_pages);
|
| 6429 | totalram_pages_add(count);
|
| 6430 | setup_per_zone_lowmem_reserve();
|
| 6431 | }
|
| 6432 | EXPORT_SYMBOL(adjust_managed_page_count);
|
| 6433 |
|
| 6434 | |
| 6435 | * free_reserved_pages - free reserved pages
|
| 6436 | * @page: First page to free.
|
| 6437 | * @order: The page order to free.
|
| 6438 | *
|
| 6439 | * Free pages allocated through memblock during boot, letting the buddy
|
| 6440 | * manage them from now on.
|
| 6441 | *
|
| 6442 | * @page must be naturally aligned to the order and the order must not
|
| 6443 | * exceed MAX_PAGE_ORDER. All pages must be reserved.
|
| 6444 |
|
| 6445 | void free_reserved_pages(struct page *page, unsigned int order)
|
| 6446 | {
|
| 6447 | const unsigned long nr_pages = 1UL << order;
|
| 6448 | int i;
|
| 6449 |
|
| 6450 | VM_WARN_ON_ONCE(!IS_ALIGNED(page_to_pfn(page), nr_pages));
|
| 6451 | VM_WARN_ON_ONCE(order > MAX_PAGE_ORDER);
|
| 6452 |
|
| 6453 | for (i = 0; i < nr_pages; i++) {
|
| 6454 | clear_page_tag_ref(page + i);
|
| 6455 | set_page_count(page + i, 0);
|
| 6456 | ClearPageReserved(page + i);
|
| 6457 | }
|
| 6458 | adjust_managed_page_count(page, nr_pages);
|
| 6459 | __free_frozen_pages(page, order, FPI_NONE);
|
| 6460 | }
|
| 6461 | EXPORT_SYMBOL(free_reserved_pages);
|
| 6462 |
|
| 6463 | static int page_alloc_cpu_dead(unsigned int cpu)
|
| 6464 | {
|
| 6465 | struct zone *zone;
|
| 6466 |
|
| 6467 | lru_add_drain_cpu(cpu);
|
| 6468 | mlock_drain_remote(cpu);
|
| 6469 | drain_pages(cpu);
|
| 6470 |
|
| 6471 | |
| 6472 | * Spill the event counters of the dead processor
|
| 6473 | * into the current processors event counters.
|
| 6474 | * This artificially elevates the count of the current
|
| 6475 | * processor.
|
| 6476 |
|
| 6477 | vm_events_fold_cpu(cpu);
|
| 6478 |
|
| 6479 | |
| 6480 | * Zero the differential counters of the dead processor
|
| 6481 | * so that the vm statistics are consistent.
|
| 6482 | *
|
| 6483 | * This is only okay since the processor is dead and cannot
|
| 6484 | * race with what we are doing.
|
| 6485 |
|
| 6486 | cpu_vm_stats_fold(cpu);
|
| 6487 |
|
| 6488 | for_each_populated_zone(zone)
|
| 6489 | zone_pcp_update(zone, 0);
|
| 6490 |
|
| 6491 | return 0;
|
| 6492 | }
|
| 6493 |
|
| 6494 | static int page_alloc_cpu_online(unsigned int cpu)
|
| 6495 | {
|
| 6496 | struct zone *zone;
|
| 6497 |
|
| 6498 | for_each_populated_zone(zone)
|
| 6499 | zone_pcp_update(zone, 1);
|
| 6500 | return 0;
|
| 6501 | }
|
| 6502 |
|
| 6503 | void __init page_alloc_init_cpuhp(void)
|
| 6504 | {
|
| 6505 | int ret;
|
| 6506 |
|
| 6507 | ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
|
| 6508 | "mm/page_alloc:pcp",
|
| 6509 | page_alloc_cpu_online,
|
| 6510 | page_alloc_cpu_dead);
|
| 6511 | WARN_ON(ret < 0);
|
| 6512 | }
|
| 6513 |
|
| 6514 | |
| 6515 | * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
|
| 6516 | * or min_free_kbytes changes.
|
| 6517 |
|
| 6518 | static void calculate_totalreserve_pages(void)
|
| 6519 | {
|
| 6520 | struct pglist_data *pgdat;
|
| 6521 | unsigned long reserve_pages = 0;
|
| 6522 | enum zone_type i, j;
|
| 6523 |
|
| 6524 | for_each_online_pgdat(pgdat) {
|
| 6525 |
|
| 6526 | pgdat->totalreserve_pages = 0;
|
| 6527 |
|
| 6528 | for (i = 0; i < MAX_NR_ZONES; i++) {
|
| 6529 | struct zone *zone = pgdat->node_zones + i;
|
| 6530 | long max = 0;
|
| 6531 | unsigned long managed_pages = zone_managed_pages(zone);
|
| 6532 |
|
| 6533 | |
| 6534 | * lowmem_reserve[j] is monotonically non-decreasing
|
| 6535 | * in j for a given zone (see
|
| 6536 | * setup_per_zone_lowmem_reserve()). The maximum
|
| 6537 | * valid reserve lives at the highest index with a
|
| 6538 | * non-zero value, so scan backwards and stop at the
|
| 6539 | * first hit.
|
| 6540 |
|
| 6541 | for (j = MAX_NR_ZONES - 1; j > i; j--) {
|
| 6542 | if (!zone->lowmem_reserve[j])
|
| 6543 | continue;
|
| 6544 |
|
| 6545 | max = zone->lowmem_reserve[j];
|
| 6546 | break;
|
| 6547 | }
|
| 6548 |
|
| 6549 | max += high_wmark_pages(zone);
|
| 6550 |
|
| 6551 | max = min_t(unsigned long, max, managed_pages);
|
| 6552 |
|
| 6553 | pgdat->totalreserve_pages += max;
|
| 6554 |
|
| 6555 | reserve_pages += max;
|
| 6556 | }
|
| 6557 | }
|
| 6558 | totalreserve_pages = reserve_pages;
|
| 6559 | trace_mm_calculate_totalreserve_pages(totalreserve_pages);
|
| 6560 | }
|
| 6561 |
|
| 6562 | |
| 6563 | * setup_per_zone_lowmem_reserve - called whenever
|
| 6564 | * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
|
| 6565 | * has a correct pages reserved value, so an adequate number of
|
| 6566 | * pages are left in the zone after a successful __alloc_pages().
|
| 6567 |
|
| 6568 | static void setup_per_zone_lowmem_reserve(void)
|
| 6569 | {
|
| 6570 | struct pglist_data *pgdat;
|
| 6571 | enum zone_type i, j;
|
| 6572 | |
| 6573 | * For a given zone node_zones[i], lowmem_reserve[j] (j > i)
|
| 6574 | * represents how many pages in zone i must effectively be kept
|
| 6575 | * in reserve when deciding whether an allocation class that is
|
| 6576 | * allowed to allocate from zones up to j may fall back into
|
| 6577 | * zone i.
|
| 6578 | *
|
| 6579 | * As j increases, the allocation class can use a strictly larger
|
| 6580 | * set of fallback zones and therefore must not be allowed to
|
| 6581 | * deplete low zones more aggressively than a less flexible one.
|
| 6582 | * As a result, lowmem_reserve[j] is required to be monotonically
|
| 6583 | * non-decreasing in j for each zone i. Callers such as
|
| 6584 | * calculate_totalreserve_pages() rely on this monotonicity when
|
| 6585 | * selecting the maximum reserve entry.
|
| 6586 |
|
| 6587 | for_each_online_pgdat(pgdat) {
|
| 6588 | for (i = 0; i < MAX_NR_ZONES - 1; i++) {
|
| 6589 | struct zone *zone = &pgdat->node_zones[i];
|
| 6590 | int ratio = sysctl_lowmem_reserve_ratio[i];
|
| 6591 | bool clear = !ratio || !zone_managed_pages(zone);
|
| 6592 | unsigned long managed_pages = 0;
|
| 6593 |
|
| 6594 | for (j = i + 1; j < MAX_NR_ZONES; j++) {
|
| 6595 | struct zone *upper_zone = &pgdat->node_zones[j];
|
| 6596 |
|
| 6597 | managed_pages += zone_managed_pages(upper_zone);
|
| 6598 |
|
| 6599 | if (clear)
|
| 6600 | zone->lowmem_reserve[j] = 0;
|
| 6601 | else
|
| 6602 | zone->lowmem_reserve[j] = managed_pages / ratio;
|
| 6603 | trace_mm_setup_per_zone_lowmem_reserve(zone, upper_zone,
|
| 6604 | zone->lowmem_reserve[j]);
|
| 6605 | }
|
| 6606 | }
|
| 6607 | }
|
| 6608 |
|
| 6609 |
|
| 6610 | calculate_totalreserve_pages();
|
| 6611 | }
|
| 6612 |
|
| 6613 | static void __setup_per_zone_wmarks(void)
|
| 6614 | {
|
| 6615 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
|
| 6616 | unsigned long lowmem_pages = 0;
|
| 6617 | struct zone *zone;
|
| 6618 | unsigned long flags;
|
| 6619 |
|
| 6620 |
|
| 6621 | for_each_zone(zone) {
|
| 6622 | if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
|
| 6623 | lowmem_pages += zone_managed_pages(zone);
|
| 6624 | }
|
| 6625 |
|
| 6626 | for_each_zone(zone) {
|
| 6627 | u64 tmp;
|
| 6628 |
|
| 6629 | spin_lock_irqsave(&zone->lock, flags);
|
| 6630 | tmp = (u64)pages_min * zone_managed_pages(zone);
|
| 6631 | tmp = div64_ul(tmp, lowmem_pages);
|
| 6632 | if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
|
| 6633 | |
| 6634 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't
|
| 6635 | * need highmem and movable zones pages, so cap pages_min
|
| 6636 | * to a small value here.
|
| 6637 | *
|
| 6638 | * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
|
| 6639 | * deltas control async page reclaim, and so should
|
| 6640 | * not be capped for highmem and movable zones.
|
| 6641 |
|
| 6642 | unsigned long min_pages;
|
| 6643 |
|
| 6644 | min_pages = zone_managed_pages(zone) / 1024;
|
| 6645 | min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
|
| 6646 | zone->_watermark[WMARK_MIN] = min_pages;
|
| 6647 | } else {
|
| 6648 | |
| 6649 | * If it's a lowmem zone, reserve a number of pages
|
| 6650 | * proportionate to the zone's size.
|
| 6651 |
|
| 6652 | zone->_watermark[WMARK_MIN] = tmp;
|
| 6653 | }
|
| 6654 |
|
| 6655 | |
| 6656 | * Set the kswapd watermarks distance according to the
|
| 6657 | * scale factor in proportion to available memory, but
|
| 6658 | * ensure a minimum size on small systems.
|
| 6659 |
|
| 6660 | tmp = max_t(u64, tmp >> 2,
|
| 6661 | mult_frac(zone_managed_pages(zone),
|
| 6662 | watermark_scale_factor, 10000));
|
| 6663 |
|
| 6664 | zone->watermark_boost = 0;
|
| 6665 | zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
|
| 6666 | zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
|
| 6667 | zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
|
| 6668 | trace_mm_setup_per_zone_wmarks(zone);
|
| 6669 |
|
| 6670 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 6671 | }
|
| 6672 |
|
| 6673 |
|
| 6674 | calculate_totalreserve_pages();
|
| 6675 | }
|
| 6676 |
|
| 6677 | |
| 6678 | * setup_per_zone_wmarks - called when min_free_kbytes changes
|
| 6679 | * or when memory is hot-{added|removed}
|
| 6680 | *
|
| 6681 | * Ensures that the watermark[min,low,high] values for each zone are set
|
| 6682 | * correctly with respect to min_free_kbytes.
|
| 6683 |
|
| 6684 | void setup_per_zone_wmarks(void)
|
| 6685 | {
|
| 6686 | struct zone *zone;
|
| 6687 | static DEFINE_SPINLOCK(lock);
|
| 6688 |
|
| 6689 | spin_lock(&lock);
|
| 6690 | __setup_per_zone_wmarks();
|
| 6691 | spin_unlock(&lock);
|
| 6692 |
|
| 6693 | |
| 6694 | * The watermark size have changed so update the pcpu batch
|
| 6695 | * and high limits or the limits may be inappropriate.
|
| 6696 |
|
| 6697 | for_each_zone(zone)
|
| 6698 | zone_pcp_update(zone, 0);
|
| 6699 | }
|
| 6700 |
|
| 6701 | |
| 6702 | * Initialise min_free_kbytes.
|
| 6703 | *
|
| 6704 | * For small machines we want it small (128k min). For large machines
|
| 6705 | * we want it large (256MB max). But it is not linear, because network
|
| 6706 | * bandwidth does not increase linearly with machine size. We use
|
| 6707 | *
|
| 6708 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
|
| 6709 | * min_free_kbytes = sqrt(lowmem_kbytes * 16)
|
| 6710 | *
|
| 6711 | * which yields
|
| 6712 | *
|
| 6713 | * 16MB: 512k
|
| 6714 | * 32MB: 724k
|
| 6715 | * 64MB: 1024k
|
| 6716 | * 128MB: 1448k
|
| 6717 | * 256MB: 2048k
|
| 6718 | * 512MB: 2896k
|
| 6719 | * 1024MB: 4096k
|
| 6720 | * 2048MB: 5792k
|
| 6721 | * 4096MB: 8192k
|
| 6722 | * 8192MB: 11584k
|
| 6723 | * 16384MB: 16384k
|
| 6724 |
|
| 6725 | void calculate_min_free_kbytes(void)
|
| 6726 | {
|
| 6727 | unsigned long lowmem_kbytes;
|
| 6728 | int new_min_free_kbytes;
|
| 6729 |
|
| 6730 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
|
| 6731 | new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
|
| 6732 |
|
| 6733 | if (new_min_free_kbytes > user_min_free_kbytes)
|
| 6734 | min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
|
| 6735 | else
|
| 6736 | pr_warn_ratelimited("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
|
| 6737 | new_min_free_kbytes, user_min_free_kbytes);
|
| 6738 |
|
| 6739 | }
|
| 6740 |
|
| 6741 | int __meminit init_per_zone_wmark_min(void)
|
| 6742 | {
|
| 6743 | calculate_min_free_kbytes();
|
| 6744 | setup_per_zone_wmarks();
|
| 6745 | refresh_zone_stat_thresholds();
|
| 6746 | setup_per_zone_lowmem_reserve();
|
| 6747 |
|
| 6748 | #ifdef CONFIG_NUMA
|
| 6749 | setup_min_unmapped_ratio();
|
| 6750 | setup_min_slab_ratio();
|
| 6751 | #endif
|
| 6752 |
|
| 6753 | khugepaged_min_free_kbytes_update();
|
| 6754 |
|
| 6755 | return 0;
|
| 6756 | }
|
| 6757 | postcore_initcall(init_per_zone_wmark_min)
|
| 6758 |
|
| 6759 | |
| 6760 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
|
| 6761 | * that we can call two helper functions whenever min_free_kbytes
|
| 6762 | * changes.
|
| 6763 |
|
| 6764 | static int min_free_kbytes_sysctl_handler(const struct ctl_table *table, int write,
|
| 6765 | void *buffer, size_t *length, loff_t *ppos)
|
| 6766 | {
|
| 6767 | int rc;
|
| 6768 |
|
| 6769 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6770 | if (rc)
|
| 6771 | return rc;
|
| 6772 |
|
| 6773 | if (write) {
|
| 6774 | user_min_free_kbytes = min_free_kbytes;
|
| 6775 | setup_per_zone_wmarks();
|
| 6776 | }
|
| 6777 | return 0;
|
| 6778 | }
|
| 6779 |
|
| 6780 | static int watermark_scale_factor_sysctl_handler(const struct ctl_table *table, int write,
|
| 6781 | void *buffer, size_t *length, loff_t *ppos)
|
| 6782 | {
|
| 6783 | int rc;
|
| 6784 |
|
| 6785 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6786 | if (rc)
|
| 6787 | return rc;
|
| 6788 |
|
| 6789 | if (write)
|
| 6790 | setup_per_zone_wmarks();
|
| 6791 |
|
| 6792 | return 0;
|
| 6793 | }
|
| 6794 |
|
| 6795 | #ifdef CONFIG_NUMA
|
| 6796 | static void setup_min_unmapped_ratio(void)
|
| 6797 | {
|
| 6798 | pg_data_t *pgdat;
|
| 6799 | struct zone *zone;
|
| 6800 |
|
| 6801 | for_each_online_pgdat(pgdat)
|
| 6802 | pgdat->min_unmapped_pages = 0;
|
| 6803 |
|
| 6804 | for_each_zone(zone)
|
| 6805 | zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
|
| 6806 | sysctl_min_unmapped_ratio) / 100;
|
| 6807 | }
|
| 6808 |
|
| 6809 |
|
| 6810 | static int sysctl_min_unmapped_ratio_sysctl_handler(const struct ctl_table *table, int write,
|
| 6811 | void *buffer, size_t *length, loff_t *ppos)
|
| 6812 | {
|
| 6813 | int rc;
|
| 6814 |
|
| 6815 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6816 | if (rc)
|
| 6817 | return rc;
|
| 6818 |
|
| 6819 | if (write)
|
| 6820 | setup_min_unmapped_ratio();
|
| 6821 |
|
| 6822 | return 0;
|
| 6823 | }
|
| 6824 |
|
| 6825 | static void setup_min_slab_ratio(void)
|
| 6826 | {
|
| 6827 | pg_data_t *pgdat;
|
| 6828 | struct zone *zone;
|
| 6829 |
|
| 6830 | for_each_online_pgdat(pgdat)
|
| 6831 | pgdat->min_slab_pages = 0;
|
| 6832 |
|
| 6833 | for_each_zone(zone)
|
| 6834 | zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
|
| 6835 | sysctl_min_slab_ratio) / 100;
|
| 6836 | }
|
| 6837 |
|
| 6838 | static int sysctl_min_slab_ratio_sysctl_handler(const struct ctl_table *table, int write,
|
| 6839 | void *buffer, size_t *length, loff_t *ppos)
|
| 6840 | {
|
| 6841 | int rc;
|
| 6842 |
|
| 6843 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6844 | if (rc)
|
| 6845 | return rc;
|
| 6846 |
|
| 6847 | if (write)
|
| 6848 | setup_min_slab_ratio();
|
| 6849 |
|
| 6850 | return 0;
|
| 6851 | }
|
| 6852 | #endif
|
| 6853 |
|
| 6854 | |
| 6855 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
|
| 6856 | * proc_dointvec_minmax() so that we can call
|
| 6857 | * setup_per_zone_lowmem_reserve() when the sysctl is written.
|
| 6858 | *
|
| 6859 | * The reserve ratio obviously has absolutely no relation with the
|
| 6860 | * minimum watermarks. The lowmem reserve ratio can only make sense
|
| 6861 | * if in function of the boot time zone sizes.
|
| 6862 |
|
| 6863 | static int lowmem_reserve_ratio_sysctl_handler(const struct ctl_table *table,
|
| 6864 | int write, void *buffer, size_t *length, loff_t *ppos)
|
| 6865 | {
|
| 6866 | struct ctl_table tmp = *table;
|
| 6867 | int ratio[ARRAY_SIZE(sysctl_lowmem_reserve_ratio)];
|
| 6868 | int rc;
|
| 6869 |
|
| 6870 | if (!write)
|
| 6871 | return proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6872 |
|
| 6873 | |
| 6874 | * proc_dointvec_max() works incrementally. Use a buffer and only set
|
| 6875 | * the values if all of them parse cleanly.
|
| 6876 |
|
| 6877 | memcpy(ratio, sysctl_lowmem_reserve_ratio, sizeof(ratio));
|
| 6878 | tmp.data = ratio;
|
| 6879 |
|
| 6880 | rc = proc_dointvec_minmax(&tmp, write, buffer, length, ppos);
|
| 6881 | if (rc)
|
| 6882 | return rc;
|
| 6883 |
|
| 6884 | memcpy(sysctl_lowmem_reserve_ratio, ratio, sizeof(ratio));
|
| 6885 | setup_per_zone_lowmem_reserve();
|
| 6886 |
|
| 6887 | return 0;
|
| 6888 | }
|
| 6889 |
|
| 6890 | |
| 6891 | * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
|
| 6892 | * cpu. It is the fraction of total pages in each zone that a hot per cpu
|
| 6893 | * pagelist can have before it gets flushed back to buddy allocator.
|
| 6894 |
|
| 6895 | static int percpu_pagelist_high_fraction_sysctl_handler(const struct ctl_table *table,
|
| 6896 | int write, void *buffer, size_t *length, loff_t *ppos)
|
| 6897 | {
|
| 6898 | struct zone *zone;
|
| 6899 | int old_percpu_pagelist_high_fraction;
|
| 6900 | int ret;
|
| 6901 |
|
| 6902 | |
| 6903 | * Avoid using pcp_batch_high_lock for reads as the value is read
|
| 6904 | * atomically and a race with offlining is harmless.
|
| 6905 |
|
| 6906 |
|
| 6907 | if (!write)
|
| 6908 | return proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6909 |
|
| 6910 | mutex_lock(&pcp_batch_high_lock);
|
| 6911 | old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
|
| 6912 |
|
| 6913 | ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
|
| 6914 | if (ret < 0)
|
| 6915 | goto out;
|
| 6916 |
|
| 6917 |
|
| 6918 | if (percpu_pagelist_high_fraction &&
|
| 6919 | percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
|
| 6920 | percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
|
| 6921 | ret = -EINVAL;
|
| 6922 | goto out;
|
| 6923 | }
|
| 6924 |
|
| 6925 |
|
| 6926 | if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
|
| 6927 | goto out;
|
| 6928 |
|
| 6929 | for_each_populated_zone(zone)
|
| 6930 | zone_set_pageset_high_and_batch(zone, 0);
|
| 6931 | out:
|
| 6932 | mutex_unlock(&pcp_batch_high_lock);
|
| 6933 | return ret;
|
| 6934 | }
|
| 6935 |
|
| 6936 | static const struct ctl_table page_alloc_sysctl_table[] = {
|
| 6937 | {
|
| 6938 | .procname = "min_free_kbytes",
|
| 6939 | .data = &min_free_kbytes,
|
| 6940 | .maxlen = sizeof(min_free_kbytes),
|
| 6941 | .mode = 0644,
|
| 6942 | .proc_handler = min_free_kbytes_sysctl_handler,
|
| 6943 | .extra1 = SYSCTL_ZERO,
|
| 6944 | },
|
| 6945 | {
|
| 6946 | .procname = "watermark_boost_factor",
|
| 6947 | .data = &watermark_boost_factor,
|
| 6948 | .maxlen = sizeof(watermark_boost_factor),
|
| 6949 | .mode = 0644,
|
| 6950 | .proc_handler = proc_dointvec_minmax,
|
| 6951 | .extra1 = SYSCTL_ZERO,
|
| 6952 | },
|
| 6953 | {
|
| 6954 | .procname = "watermark_scale_factor",
|
| 6955 | .data = &watermark_scale_factor,
|
| 6956 | .maxlen = sizeof(watermark_scale_factor),
|
| 6957 | .mode = 0644,
|
| 6958 | .proc_handler = watermark_scale_factor_sysctl_handler,
|
| 6959 | .extra1 = SYSCTL_ONE,
|
| 6960 | .extra2 = SYSCTL_THREE_THOUSAND,
|
| 6961 | },
|
| 6962 | {
|
| 6963 | .procname = "defrag_mode",
|
| 6964 | .data = &defrag_mode,
|
| 6965 | .maxlen = sizeof(defrag_mode),
|
| 6966 | .mode = 0644,
|
| 6967 | .proc_handler = proc_dointvec_minmax,
|
| 6968 | .extra1 = SYSCTL_ZERO,
|
| 6969 | .extra2 = SYSCTL_ONE,
|
| 6970 | },
|
| 6971 | {
|
| 6972 | .procname = "percpu_pagelist_high_fraction",
|
| 6973 | .data = &percpu_pagelist_high_fraction,
|
| 6974 | .maxlen = sizeof(percpu_pagelist_high_fraction),
|
| 6975 | .mode = 0644,
|
| 6976 | .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
|
| 6977 | .extra1 = SYSCTL_ZERO,
|
| 6978 | },
|
| 6979 | {
|
| 6980 | .procname = "lowmem_reserve_ratio",
|
| 6981 | .data = &sysctl_lowmem_reserve_ratio,
|
| 6982 | .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
|
| 6983 | .mode = 0644,
|
| 6984 | .proc_handler = lowmem_reserve_ratio_sysctl_handler,
|
| 6985 | .extra1 = SYSCTL_ZERO,
|
| 6986 | },
|
| 6987 | #ifdef CONFIG_NUMA
|
| 6988 | {
|
| 6989 | .procname = "numa_zonelist_order",
|
| 6990 | .data = &numa_zonelist_order,
|
| 6991 | .maxlen = NUMA_ZONELIST_ORDER_LEN,
|
| 6992 | .mode = 0644,
|
| 6993 | .proc_handler = numa_zonelist_order_handler,
|
| 6994 | },
|
| 6995 | {
|
| 6996 | .procname = "min_unmapped_ratio",
|
| 6997 | .data = &sysctl_min_unmapped_ratio,
|
| 6998 | .maxlen = sizeof(sysctl_min_unmapped_ratio),
|
| 6999 | .mode = 0644,
|
| 7000 | .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
|
| 7001 | .extra1 = SYSCTL_ZERO,
|
| 7002 | .extra2 = SYSCTL_ONE_HUNDRED,
|
| 7003 | },
|
| 7004 | {
|
| 7005 | .procname = "min_slab_ratio",
|
| 7006 | .data = &sysctl_min_slab_ratio,
|
| 7007 | .maxlen = sizeof(sysctl_min_slab_ratio),
|
| 7008 | .mode = 0644,
|
| 7009 | .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
|
| 7010 | .extra1 = SYSCTL_ZERO,
|
| 7011 | .extra2 = SYSCTL_ONE_HUNDRED,
|
| 7012 | },
|
| 7013 | #endif
|
| 7014 | };
|
| 7015 |
|
| 7016 | void __init page_alloc_sysctl_init(void)
|
| 7017 | {
|
| 7018 | register_sysctl_init("vm", page_alloc_sysctl_table);
|
| 7019 | }
|
| 7020 |
|
| 7021 | static void free_prepared_contig_range(struct page *page,
|
| 7022 | unsigned long nr_pages)
|
| 7023 | {
|
| 7024 | unsigned long pfn = page_to_pfn(page);
|
| 7025 |
|
| 7026 | while (nr_pages) {
|
| 7027 | unsigned int order;
|
| 7028 |
|
| 7029 |
|
| 7030 | order = pfn ? __ffs(pfn) : MAX_PAGE_ORDER;
|
| 7031 |
|
| 7032 | order = min_t(unsigned int, order, ilog2(nr_pages));
|
| 7033 | order = min_t(unsigned int, order, MAX_PAGE_ORDER);
|
| 7034 |
|
| 7035 | |
| 7036 | * Free the chunk as a single block. Our caller has already
|
| 7037 | * called free_pages_prepare() for each order-0 page.
|
| 7038 |
|
| 7039 | __free_frozen_pages(page, order, FPI_PREPARED);
|
| 7040 |
|
| 7041 | pfn += 1UL << order;
|
| 7042 | page += 1UL << order;
|
| 7043 | nr_pages -= 1UL << order;
|
| 7044 | }
|
| 7045 | }
|
| 7046 |
|
| 7047 | static void __free_contig_range_common(unsigned long pfn, unsigned long nr_pages,
|
| 7048 | bool is_frozen)
|
| 7049 | {
|
| 7050 | struct page *page, *start = NULL;
|
| 7051 | unsigned long nr_start = 0;
|
| 7052 | unsigned long start_sec;
|
| 7053 | unsigned long i;
|
| 7054 |
|
| 7055 | for (i = 0; i < nr_pages; i++) {
|
| 7056 | bool can_free = true;
|
| 7057 |
|
| 7058 | |
| 7059 | * Contiguous PFNs might not have contiguous "struct pages"
|
| 7060 | * in some kernel configs: page++ across a section boundary
|
| 7061 | * is undefined. Use pfn_to_page() for each PFN.
|
| 7062 |
|
| 7063 | page = pfn_to_page(pfn + i);
|
| 7064 |
|
| 7065 | VM_WARN_ON_ONCE(PageHead(page));
|
| 7066 | VM_WARN_ON_ONCE(PageTail(page));
|
| 7067 |
|
| 7068 | if (!is_frozen)
|
| 7069 | can_free = put_page_testzero(page);
|
| 7070 |
|
| 7071 | if (can_free)
|
| 7072 | can_free = free_pages_prepare(page, 0);
|
| 7073 |
|
| 7074 | if (!can_free) {
|
| 7075 | if (start) {
|
| 7076 | free_prepared_contig_range(start, i - nr_start);
|
| 7077 | start = NULL;
|
| 7078 | }
|
| 7079 | continue;
|
| 7080 | }
|
| 7081 |
|
| 7082 | if (start && memdesc_section(&page->flags) != start_sec) {
|
| 7083 | free_prepared_contig_range(start, i - nr_start);
|
| 7084 | start = page;
|
| 7085 | nr_start = i;
|
| 7086 | start_sec = memdesc_section(&page->flags);
|
| 7087 | } else if (!start) {
|
| 7088 | start = page;
|
| 7089 | nr_start = i;
|
| 7090 | start_sec = memdesc_section(&page->flags);
|
| 7091 | }
|
| 7092 | }
|
| 7093 |
|
| 7094 | if (start)
|
| 7095 | free_prepared_contig_range(start, nr_pages - nr_start);
|
| 7096 | }
|
| 7097 |
|
| 7098 | |
| 7099 | * __free_contig_range - Free contiguous range of order-0 pages.
|
| 7100 | * @pfn: Page frame number of the first page in the range.
|
| 7101 | * @nr_pages: Number of pages to free.
|
| 7102 | *
|
| 7103 | * For each order-0 struct page in the physically contiguous range, put a
|
| 7104 | * reference. Free any page who's reference count falls to zero. The
|
| 7105 | * implementation is functionally equivalent to, but significantly faster than
|
| 7106 | * calling __free_page() for each struct page in a loop.
|
| 7107 | *
|
| 7108 | * Memory allocated with alloc_pages(order>=1) then subsequently split to
|
| 7109 | * order-0 with split_page() is an example of appropriate contiguous pages that
|
| 7110 | * can be freed with this API.
|
| 7111 | *
|
| 7112 | * Context: May be called in interrupt context or while holding a normal
|
| 7113 | * spinlock, but not in NMI context or while holding a raw spinlock.
|
| 7114 |
|
| 7115 | void __free_contig_range(unsigned long pfn, unsigned long nr_pages)
|
| 7116 | {
|
| 7117 | __free_contig_range_common(pfn, nr_pages, false);
|
| 7118 | }
|
| 7119 |
|
| 7120 | #ifdef CONFIG_CONTIG_ALLOC
|
| 7121 |
|
| 7122 | static void alloc_contig_dump_pages(struct list_head *page_list)
|
| 7123 | {
|
| 7124 | DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
|
| 7125 |
|
| 7126 | if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
|
| 7127 | struct page *page;
|
| 7128 |
|
| 7129 | dump_stack();
|
| 7130 | list_for_each_entry(page, page_list, lru)
|
| 7131 | dump_page(page, "migration failure");
|
| 7132 | }
|
| 7133 | }
|
| 7134 |
|
| 7135 |
|
| 7136 | static int __alloc_contig_migrate_range(struct compact_control *cc,
|
| 7137 | unsigned long start, unsigned long end)
|
| 7138 | {
|
| 7139 |
|
| 7140 | unsigned int nr_reclaimed;
|
| 7141 | unsigned long pfn = start;
|
| 7142 | unsigned int tries = 0;
|
| 7143 | int ret = 0;
|
| 7144 | struct migration_target_control mtc = {
|
| 7145 | .nid = zone_to_nid(cc->zone),
|
| 7146 | .gfp_mask = cc->gfp_mask,
|
| 7147 | .reason = MR_CONTIG_RANGE,
|
| 7148 | };
|
| 7149 |
|
| 7150 | lru_cache_disable();
|
| 7151 |
|
| 7152 | while (pfn < end || !list_empty(&cc->migratepages)) {
|
| 7153 | if (fatal_signal_pending(current)) {
|
| 7154 | ret = -EINTR;
|
| 7155 | break;
|
| 7156 | }
|
| 7157 |
|
| 7158 | if (list_empty(&cc->migratepages)) {
|
| 7159 | cc->nr_migratepages = 0;
|
| 7160 | ret = isolate_migratepages_range(cc, pfn, end);
|
| 7161 | if (ret && ret != -EAGAIN)
|
| 7162 | break;
|
| 7163 | pfn = cc->migrate_pfn;
|
| 7164 | tries = 0;
|
| 7165 | } else if (++tries == 5) {
|
| 7166 | ret = -EBUSY;
|
| 7167 | break;
|
| 7168 | }
|
| 7169 |
|
| 7170 | nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
|
| 7171 | &cc->migratepages);
|
| 7172 | cc->nr_migratepages -= nr_reclaimed;
|
| 7173 |
|
| 7174 | ret = migrate_pages(&cc->migratepages, alloc_migration_target,
|
| 7175 | NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
|
| 7176 |
|
| 7177 | |
| 7178 | * On -ENOMEM, migrate_pages() bails out right away. It is pointless
|
| 7179 | * to retry again over this error, so do the same here.
|
| 7180 |
|
| 7181 | if (ret == -ENOMEM)
|
| 7182 | break;
|
| 7183 | }
|
| 7184 |
|
| 7185 | lru_cache_enable();
|
| 7186 | if (ret < 0) {
|
| 7187 | if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
|
| 7188 | alloc_contig_dump_pages(&cc->migratepages);
|
| 7189 | putback_movable_pages(&cc->migratepages);
|
| 7190 | }
|
| 7191 |
|
| 7192 | return (ret < 0) ? ret : 0;
|
| 7193 | }
|
| 7194 |
|
| 7195 | static void split_free_frozen_pages(struct list_head *list, gfp_t gfp_mask)
|
| 7196 | {
|
| 7197 | int order;
|
| 7198 |
|
| 7199 | for (order = 0; order < NR_PAGE_ORDERS; order++) {
|
| 7200 | struct page *page, *next;
|
| 7201 | int nr_pages = 1 << order;
|
| 7202 |
|
| 7203 | list_for_each_entry_safe(page, next, &list[order], lru) {
|
| 7204 | int i;
|
| 7205 |
|
| 7206 | post_alloc_hook(page, order, gfp_mask, ALLOC_DEFAULT);
|
| 7207 | if (!order)
|
| 7208 | continue;
|
| 7209 |
|
| 7210 | __split_page(page, order);
|
| 7211 |
|
| 7212 |
|
| 7213 | list_del(&page->lru);
|
| 7214 | for (i = 0; i < nr_pages; i++)
|
| 7215 | list_add_tail(&page[i].lru, &list[0]);
|
| 7216 | }
|
| 7217 | }
|
| 7218 | }
|
| 7219 |
|
| 7220 | static int __alloc_contig_verify_gfp_mask(gfp_t gfp_mask, gfp_t *gfp_cc_mask)
|
| 7221 | {
|
| 7222 | const gfp_t reclaim_mask = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
|
| 7223 | const gfp_t action_mask = __GFP_COMP | __GFP_RETRY_MAYFAIL | __GFP_NOWARN |
|
| 7224 | __GFP_ZERO | __GFP_ZEROTAGS | __GFP_SKIP_ZERO |
|
| 7225 | __GFP_SKIP_KASAN;
|
| 7226 | const gfp_t cc_action_mask = __GFP_RETRY_MAYFAIL | __GFP_NOWARN;
|
| 7227 |
|
| 7228 | |
| 7229 | * We are given the range to allocate; node, mobility and placement
|
| 7230 | * hints are irrelevant at this point. We'll simply ignore them.
|
| 7231 |
|
| 7232 | gfp_mask &= ~(GFP_ZONEMASK | __GFP_RECLAIMABLE | __GFP_WRITE |
|
| 7233 | __GFP_HARDWALL | __GFP_THISNODE | __GFP_MOVABLE);
|
| 7234 |
|
| 7235 | |
| 7236 | * We only support most reclaim flags (but not NOFAIL/NORETRY), and
|
| 7237 | * selected action flags.
|
| 7238 |
|
| 7239 | if (gfp_mask & ~(reclaim_mask | action_mask))
|
| 7240 | return -EINVAL;
|
| 7241 |
|
| 7242 | |
| 7243 | * Flags to control page compaction/migration/reclaim, to free up our
|
| 7244 | * page range. Migratable pages are movable, __GFP_MOVABLE is implied
|
| 7245 | * for them.
|
| 7246 | *
|
| 7247 | * Traditionally we always had __GFP_RETRY_MAYFAIL set, keep doing that
|
| 7248 | * to not degrade callers.
|
| 7249 |
|
| 7250 | *gfp_cc_mask = (gfp_mask & (reclaim_mask | cc_action_mask)) |
|
| 7251 | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL;
|
| 7252 | return 0;
|
| 7253 | }
|
| 7254 |
|
| 7255 | static void __free_contig_frozen_range(unsigned long pfn, unsigned long nr_pages)
|
| 7256 | {
|
| 7257 | __free_contig_range_common(pfn, nr_pages, true);
|
| 7258 | }
|
| 7259 |
|
| 7260 | |
| 7261 | * alloc_contig_frozen_range() -- tries to allocate given range of frozen pages
|
| 7262 | * @start: start PFN to allocate
|
| 7263 | * @end: one-past-the-last PFN to allocate
|
| 7264 | * @alloc_flags: allocation information
|
| 7265 | * @gfp_mask: GFP mask. Node/zone/placement hints are ignored; only some
|
| 7266 | * action and reclaim modifiers are supported. Reclaim modifiers
|
| 7267 | * control allocation behavior during compaction/migration/reclaim.
|
| 7268 | *
|
| 7269 | * The PFN range does not have to be pageblock aligned. The PFN range must
|
| 7270 | * belong to a single zone.
|
| 7271 | *
|
| 7272 | * The first thing this routine does is attempt to MIGRATE_ISOLATE all
|
| 7273 | * pageblocks in the range. Once isolated, the pageblocks should not
|
| 7274 | * be modified by others.
|
| 7275 | *
|
| 7276 | * All frozen pages which PFN is in [start, end) are allocated for the
|
| 7277 | * caller, and they could be freed with free_contig_frozen_range(),
|
| 7278 | * free_frozen_pages() also could be used to free compound frozen pages
|
| 7279 | * directly.
|
| 7280 | *
|
| 7281 | * Return: zero on success or negative error code.
|
| 7282 |
|
| 7283 | int alloc_contig_frozen_range_noprof(unsigned long start, unsigned long end,
|
| 7284 | acr_flags_t alloc_flags, gfp_t gfp_mask)
|
| 7285 | {
|
| 7286 | const unsigned int order = ilog2(end - start);
|
| 7287 | unsigned long outer_start, outer_end;
|
| 7288 | int ret = 0;
|
| 7289 |
|
| 7290 | struct compact_control cc = {
|
| 7291 | .nr_migratepages = 0,
|
| 7292 | .order = -1,
|
| 7293 | .zone = page_zone(pfn_to_page(start)),
|
| 7294 | .mode = MIGRATE_SYNC,
|
| 7295 | .ignore_skip_hint = true,
|
| 7296 | .no_set_skip_hint = true,
|
| 7297 | .alloc_contig = true,
|
| 7298 | };
|
| 7299 | INIT_LIST_HEAD(&cc.migratepages);
|
| 7300 | enum pb_isolate_mode mode = (alloc_flags & ACR_FLAGS_CMA) ?
|
| 7301 | PB_ISOLATE_MODE_CMA_ALLOC :
|
| 7302 | PB_ISOLATE_MODE_OTHER;
|
| 7303 |
|
| 7304 | |
| 7305 | * In contrast to the buddy, we allow for orders here that exceed
|
| 7306 | * MAX_PAGE_ORDER, so we must manually make sure that we are not
|
| 7307 | * exceeding the maximum folio order.
|
| 7308 |
|
| 7309 | if (WARN_ON_ONCE((gfp_mask & __GFP_COMP) && order > MAX_FOLIO_ORDER))
|
| 7310 | return -EINVAL;
|
| 7311 |
|
| 7312 | gfp_mask = current_gfp_context(gfp_mask);
|
| 7313 | if (__alloc_contig_verify_gfp_mask(gfp_mask, (gfp_t *)&cc.gfp_mask))
|
| 7314 | return -EINVAL;
|
| 7315 |
|
| 7316 | |
| 7317 | * What we do here is we mark all pageblocks in range as
|
| 7318 | * MIGRATE_ISOLATE. Because pageblock and max order pages may
|
| 7319 | * have different sizes, and due to the way page allocator
|
| 7320 | * work, start_isolate_page_range() has special handlings for this.
|
| 7321 | *
|
| 7322 | * Once the pageblocks are marked as MIGRATE_ISOLATE, we
|
| 7323 | * migrate the pages from an unaligned range (ie. pages that
|
| 7324 | * we are interested in). This will put all the pages in
|
| 7325 | * range back to page allocator as MIGRATE_ISOLATE.
|
| 7326 | *
|
| 7327 | * When this is done, we take the pages in range from page
|
| 7328 | * allocator removing them from the buddy system. This way
|
| 7329 | * page allocator will never consider using them.
|
| 7330 | *
|
| 7331 | * This lets us mark the pageblocks back as
|
| 7332 | * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
|
| 7333 | * aligned range but not in the unaligned, original range are
|
| 7334 | * put back to page allocator so that buddy can use them.
|
| 7335 |
|
| 7336 |
|
| 7337 | ret = start_isolate_page_range(start, end, mode);
|
| 7338 | if (ret)
|
| 7339 | goto done;
|
| 7340 |
|
| 7341 | drain_all_pages(cc.zone);
|
| 7342 |
|
| 7343 | |
| 7344 | * In case of -EBUSY, we'd like to know which page causes problem.
|
| 7345 | * So, just fall through. test_pages_isolated() has a tracepoint
|
| 7346 | * which will report the busy page.
|
| 7347 | *
|
| 7348 | * It is possible that busy pages could become available before
|
| 7349 | * the call to test_pages_isolated, and the range will actually be
|
| 7350 | * allocated. So, if we fall through be sure to clear ret so that
|
| 7351 | * -EBUSY is not accidentally used or returned to caller.
|
| 7352 |
|
| 7353 | ret = __alloc_contig_migrate_range(&cc, start, end);
|
| 7354 | if (ret && ret != -EBUSY)
|
| 7355 | goto done;
|
| 7356 |
|
| 7357 | |
| 7358 | * When in-use hugetlb pages are migrated, they may simply be released
|
| 7359 | * back into the free hugepage pool instead of being returned to the
|
| 7360 | * buddy system. After the migration of in-use huge pages is completed,
|
| 7361 | * we will invoke replace_free_hugepage_folios() to ensure that these
|
| 7362 | * hugepages are properly released to the buddy system.
|
| 7363 |
|
| 7364 | ret = replace_free_hugepage_folios(start, end);
|
| 7365 | if (ret)
|
| 7366 | goto done;
|
| 7367 |
|
| 7368 | |
| 7369 | * Pages from [start, end) are within a pageblock_nr_pages
|
| 7370 | * aligned blocks that are marked as MIGRATE_ISOLATE. What's
|
| 7371 | * more, all pages in [start, end) are free in page allocator.
|
| 7372 | * What we are going to do is to allocate all pages from
|
| 7373 | * [start, end) (that is remove them from page allocator).
|
| 7374 | *
|
| 7375 | * The only problem is that pages at the beginning and at the
|
| 7376 | * end of interesting range may be not aligned with pages that
|
| 7377 | * page allocator holds, ie. they can be part of higher order
|
| 7378 | * pages. Because of this, we reserve the bigger range and
|
| 7379 | * once this is done free the pages we are not interested in.
|
| 7380 | *
|
| 7381 | * We don't have to hold zone->lock here because the pages are
|
| 7382 | * isolated thus they won't get removed from buddy.
|
| 7383 |
|
| 7384 | outer_start = find_large_buddy(start);
|
| 7385 |
|
| 7386 |
|
| 7387 | if (test_pages_isolated(outer_start, end, mode)) {
|
| 7388 | ret = -EBUSY;
|
| 7389 | goto done;
|
| 7390 | }
|
| 7391 |
|
| 7392 |
|
| 7393 | outer_end = isolate_freepages_range(&cc, outer_start, end);
|
| 7394 | if (!outer_end) {
|
| 7395 | ret = -EBUSY;
|
| 7396 | goto done;
|
| 7397 | }
|
| 7398 |
|
| 7399 | if (!(gfp_mask & __GFP_COMP)) {
|
| 7400 | split_free_frozen_pages(cc.freepages, gfp_mask);
|
| 7401 |
|
| 7402 |
|
| 7403 | if (start != outer_start)
|
| 7404 | __free_contig_frozen_range(outer_start, start - outer_start);
|
| 7405 | if (end != outer_end)
|
| 7406 | __free_contig_frozen_range(end, outer_end - end);
|
| 7407 | } else if (start == outer_start && end == outer_end && is_power_of_2(end - start)) {
|
| 7408 | struct page *head = pfn_to_page(start);
|
| 7409 |
|
| 7410 | check_new_pages(head, order);
|
| 7411 | prep_new_page(head, order, gfp_mask, ALLOC_DEFAULT);
|
| 7412 | } else {
|
| 7413 | ret = -EINVAL;
|
| 7414 | WARN(true, "PFN range: requested [%lu, %lu), allocated [%lu, %lu)\n",
|
| 7415 | start, end, outer_start, outer_end);
|
| 7416 | }
|
| 7417 | done:
|
| 7418 | undo_isolate_page_range(start, end);
|
| 7419 | return ret;
|
| 7420 | }
|
| 7421 | EXPORT_SYMBOL(alloc_contig_frozen_range_noprof);
|
| 7422 |
|
| 7423 | |
| 7424 | * alloc_contig_range() -- tries to allocate given range of pages
|
| 7425 | * @start: start PFN to allocate
|
| 7426 | * @end: one-past-the-last PFN to allocate
|
| 7427 | * @alloc_flags: allocation information
|
| 7428 | * @gfp_mask: GFP mask.
|
| 7429 | *
|
| 7430 | * This routine is a wrapper around alloc_contig_frozen_range(), it can't
|
| 7431 | * be used to allocate compound pages, the refcount of each allocated page
|
| 7432 | * will be set to one.
|
| 7433 | *
|
| 7434 | * All pages which PFN is in [start, end) are allocated for the caller,
|
| 7435 | * and should be freed with free_contig_range() or by manually calling
|
| 7436 | * __free_page() on each allocated page.
|
| 7437 | *
|
| 7438 | * Return: zero on success or negative error code.
|
| 7439 |
|
| 7440 | int alloc_contig_range_noprof(unsigned long start, unsigned long end,
|
| 7441 | acr_flags_t alloc_flags, gfp_t gfp_mask)
|
| 7442 | {
|
| 7443 | int ret;
|
| 7444 |
|
| 7445 | if (WARN_ON(gfp_mask & __GFP_COMP))
|
| 7446 | return -EINVAL;
|
| 7447 |
|
| 7448 | ret = alloc_contig_frozen_range_noprof(start, end, alloc_flags, gfp_mask);
|
| 7449 | if (!ret)
|
| 7450 | set_pages_refcounted(pfn_to_page(start), end - start);
|
| 7451 |
|
| 7452 | return ret;
|
| 7453 | }
|
| 7454 | EXPORT_SYMBOL(alloc_contig_range_noprof);
|
| 7455 |
|
| 7456 | static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
|
| 7457 | unsigned long nr_pages, bool skip_hugetlb,
|
| 7458 | bool *skipped_hugetlb)
|
| 7459 | {
|
| 7460 | unsigned long end_pfn = start_pfn + nr_pages;
|
| 7461 | struct page *page;
|
| 7462 |
|
| 7463 | while (start_pfn < end_pfn) {
|
| 7464 | unsigned long step = 1;
|
| 7465 |
|
| 7466 | page = pfn_to_online_page(start_pfn);
|
| 7467 | if (!page)
|
| 7468 | return false;
|
| 7469 |
|
| 7470 | if (page_zone(page) != z)
|
| 7471 | return false;
|
| 7472 |
|
| 7473 | if (page_is_unmovable(z, page, PB_ISOLATE_MODE_OTHER, &step))
|
| 7474 | return false;
|
| 7475 |
|
| 7476 | |
| 7477 | * Only consider ranges containing hugepages if those pages are
|
| 7478 | * smaller than the requested contiguous region. e.g.:
|
| 7479 | * Move 2MB pages to free up a 1GB range.
|
| 7480 | * Don't move 1GB pages to free up a 2MB range.
|
| 7481 | *
|
| 7482 | * This makes contiguous allocation more reliable if multiple
|
| 7483 | * hugepage sizes are used without causing needless movement.
|
| 7484 |
|
| 7485 | if (PageHuge(page)) {
|
| 7486 | unsigned int order;
|
| 7487 |
|
| 7488 | if (skip_hugetlb) {
|
| 7489 | *skipped_hugetlb = true;
|
| 7490 | return false;
|
| 7491 | }
|
| 7492 |
|
| 7493 | page = compound_head(page);
|
| 7494 | order = compound_order(page);
|
| 7495 | if ((order >= MAX_FOLIO_ORDER) ||
|
| 7496 | (nr_pages <= (1 << order)))
|
| 7497 | return false;
|
| 7498 | }
|
| 7499 |
|
| 7500 | start_pfn += step;
|
| 7501 | }
|
| 7502 | return true;
|
| 7503 | }
|
| 7504 |
|
| 7505 | static bool zone_spans_last_pfn(const struct zone *zone,
|
| 7506 | unsigned long start_pfn, unsigned long nr_pages)
|
| 7507 | {
|
| 7508 | unsigned long last_pfn = start_pfn + nr_pages - 1;
|
| 7509 |
|
| 7510 | return zone_spans_pfn(zone, last_pfn);
|
| 7511 | }
|
| 7512 |
|
| 7513 | |
| 7514 | * alloc_contig_frozen_pages() -- tries to find and allocate contiguous range of frozen pages
|
| 7515 | * @nr_pages: Number of contiguous pages to allocate
|
| 7516 | * @gfp_mask: GFP mask. Node/zone/placement hints limit the search; only some
|
| 7517 | * action and reclaim modifiers are supported. Reclaim modifiers
|
| 7518 | * control allocation behavior during compaction/migration/reclaim.
|
| 7519 | * @nid: Target node
|
| 7520 | * @nodemask: Mask for other possible nodes
|
| 7521 | *
|
| 7522 | * This routine is a wrapper around alloc_contig_frozen_range(). It scans over
|
| 7523 | * zones on an applicable zonelist to find a contiguous pfn range which can then
|
| 7524 | * be tried for allocation with alloc_contig_frozen_range(). This routine is
|
| 7525 | * intended for allocation requests which can not be fulfilled with the buddy
|
| 7526 | * allocator.
|
| 7527 | *
|
| 7528 | * The allocated memory is always aligned to a page boundary. If nr_pages is a
|
| 7529 | * power of two, then allocated range is also guaranteed to be aligned to same
|
| 7530 | * nr_pages (e.g. 1GB request would be aligned to 1GB).
|
| 7531 | *
|
| 7532 | * Allocated frozen pages need be freed with free_contig_frozen_range(),
|
| 7533 | * or by manually calling free_frozen_pages() on each allocated frozen
|
| 7534 | * non-compound page, for compound frozen pages could be freed with
|
| 7535 | * free_frozen_pages() directly.
|
| 7536 | *
|
| 7537 | * Return: pointer to contiguous frozen pages on success, or NULL if not successful.
|
| 7538 |
|
| 7539 | struct page *alloc_contig_frozen_pages_noprof(unsigned long nr_pages,
|
| 7540 | gfp_t gfp_mask, int nid, nodemask_t *nodemask)
|
| 7541 | {
|
| 7542 | unsigned long ret, pfn, flags;
|
| 7543 | struct zonelist *zonelist;
|
| 7544 | struct zone *zone;
|
| 7545 | struct zoneref *z;
|
| 7546 | bool skip_hugetlb = true;
|
| 7547 | bool skipped_hugetlb = false;
|
| 7548 |
|
| 7549 | retry:
|
| 7550 | zonelist = node_zonelist(nid, gfp_mask);
|
| 7551 | for_each_zone_zonelist_nodemask(zone, z, zonelist,
|
| 7552 | gfp_zone(gfp_mask), nodemask) {
|
| 7553 | spin_lock_irqsave(&zone->lock, flags);
|
| 7554 |
|
| 7555 | pfn = ALIGN(zone->zone_start_pfn, nr_pages);
|
| 7556 | while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
|
| 7557 | if (pfn_range_valid_contig(zone, pfn, nr_pages,
|
| 7558 | skip_hugetlb,
|
| 7559 | &skipped_hugetlb)) {
|
| 7560 | |
| 7561 | * We release the zone lock here because
|
| 7562 | * alloc_contig_frozen_range() will also lock
|
| 7563 | * the zone at some point. If there's an
|
| 7564 | * allocation spinning on this lock, it may
|
| 7565 | * win the race and cause allocation to fail.
|
| 7566 |
|
| 7567 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7568 | ret = alloc_contig_frozen_range_noprof(pfn,
|
| 7569 | pfn + nr_pages,
|
| 7570 | ACR_FLAGS_NONE,
|
| 7571 | gfp_mask);
|
| 7572 | if (!ret)
|
| 7573 | return pfn_to_page(pfn);
|
| 7574 | spin_lock_irqsave(&zone->lock, flags);
|
| 7575 | }
|
| 7576 | pfn += nr_pages;
|
| 7577 | }
|
| 7578 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7579 | }
|
| 7580 | |
| 7581 | * If we failed, retry the search, but treat regions with HugeTLB pages
|
| 7582 | * as valid targets. This retains fast-allocations on first pass
|
| 7583 | * without trying to migrate HugeTLB pages (which may fail). On the
|
| 7584 | * second pass, we will try moving HugeTLB pages when those pages are
|
| 7585 | * smaller than the requested contiguous region size.
|
| 7586 |
|
| 7587 | if (skip_hugetlb && skipped_hugetlb) {
|
| 7588 | skip_hugetlb = false;
|
| 7589 | goto retry;
|
| 7590 | }
|
| 7591 | return NULL;
|
| 7592 | }
|
| 7593 | EXPORT_SYMBOL(alloc_contig_frozen_pages_noprof);
|
| 7594 |
|
| 7595 | |
| 7596 | * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
|
| 7597 | * @nr_pages: Number of contiguous pages to allocate
|
| 7598 | * @gfp_mask: GFP mask.
|
| 7599 | * @nid: Target node
|
| 7600 | * @nodemask: Mask for other possible nodes
|
| 7601 | *
|
| 7602 | * This routine is a wrapper around alloc_contig_frozen_pages(), it can't
|
| 7603 | * be used to allocate compound pages, the refcount of each allocated page
|
| 7604 | * will be set to one.
|
| 7605 | *
|
| 7606 | * Allocated pages can be freed with free_contig_range() or by manually
|
| 7607 | * calling __free_page() on each allocated page.
|
| 7608 | *
|
| 7609 | * Return: pointer to contiguous pages on success, or NULL if not successful.
|
| 7610 |
|
| 7611 | struct page *alloc_contig_pages_noprof(unsigned long nr_pages, gfp_t gfp_mask,
|
| 7612 | int nid, nodemask_t *nodemask)
|
| 7613 | {
|
| 7614 | struct page *page;
|
| 7615 |
|
| 7616 | if (WARN_ON(gfp_mask & __GFP_COMP))
|
| 7617 | return NULL;
|
| 7618 |
|
| 7619 | page = alloc_contig_frozen_pages_noprof(nr_pages, gfp_mask, nid,
|
| 7620 | nodemask);
|
| 7621 | if (page)
|
| 7622 | set_pages_refcounted(page, nr_pages);
|
| 7623 |
|
| 7624 | return page;
|
| 7625 | }
|
| 7626 | EXPORT_SYMBOL(alloc_contig_pages_noprof);
|
| 7627 |
|
| 7628 | |
| 7629 | * free_contig_frozen_range() -- free the contiguous range of frozen pages
|
| 7630 | * @pfn: start PFN to free
|
| 7631 | * @nr_pages: Number of contiguous frozen pages to free
|
| 7632 | *
|
| 7633 | * This can be used to free the allocated compound/non-compound frozen pages.
|
| 7634 |
|
| 7635 | void free_contig_frozen_range(unsigned long pfn, unsigned long nr_pages)
|
| 7636 | {
|
| 7637 | struct page *first_page = pfn_to_page(pfn);
|
| 7638 | const unsigned int order = ilog2(nr_pages);
|
| 7639 |
|
| 7640 | if (WARN_ON_ONCE(first_page != compound_head(first_page)))
|
| 7641 | return;
|
| 7642 |
|
| 7643 | if (PageHead(first_page)) {
|
| 7644 | WARN_ON_ONCE(order != compound_order(first_page));
|
| 7645 | free_frozen_pages(first_page, order);
|
| 7646 | return;
|
| 7647 | }
|
| 7648 |
|
| 7649 | __free_contig_frozen_range(pfn, nr_pages);
|
| 7650 | }
|
| 7651 | EXPORT_SYMBOL(free_contig_frozen_range);
|
| 7652 |
|
| 7653 | |
| 7654 | * free_contig_range() -- free the contiguous range of pages
|
| 7655 | * @pfn: start PFN to free
|
| 7656 | * @nr_pages: Number of contiguous pages to free
|
| 7657 | *
|
| 7658 | * This can be only used to free the allocated non-compound pages.
|
| 7659 |
|
| 7660 | void free_contig_range(unsigned long pfn, unsigned long nr_pages)
|
| 7661 | {
|
| 7662 | if (WARN_ON_ONCE(PageHead(pfn_to_page(pfn))))
|
| 7663 | return;
|
| 7664 |
|
| 7665 | __free_contig_range(pfn, nr_pages);
|
| 7666 | }
|
| 7667 | EXPORT_SYMBOL(free_contig_range);
|
| 7668 | #endif
|
| 7669 |
|
| 7670 | |
| 7671 | * Effectively disable pcplists for the zone by setting the high limit to 0
|
| 7672 | * and draining all cpus. A concurrent page freeing on another CPU that's about
|
| 7673 | * to put the page on pcplist will either finish before the drain and the page
|
| 7674 | * will be drained, or observe the new high limit and skip the pcplist.
|
| 7675 | *
|
| 7676 | * Must be paired with a call to zone_pcp_enable().
|
| 7677 |
|
| 7678 | void zone_pcp_disable(struct zone *zone)
|
| 7679 | {
|
| 7680 | mutex_lock(&pcp_batch_high_lock);
|
| 7681 | __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
|
| 7682 | __drain_all_pages(zone, true);
|
| 7683 | }
|
| 7684 |
|
| 7685 | void zone_pcp_enable(struct zone *zone)
|
| 7686 | {
|
| 7687 | __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
|
| 7688 | zone->pageset_high_max, zone->pageset_batch);
|
| 7689 | mutex_unlock(&pcp_batch_high_lock);
|
| 7690 | }
|
| 7691 |
|
| 7692 | void zone_pcp_reset(struct zone *zone)
|
| 7693 | {
|
| 7694 | int cpu;
|
| 7695 | struct per_cpu_zonestat *pzstats;
|
| 7696 |
|
| 7697 | if (zone->per_cpu_pageset != &boot_pageset) {
|
| 7698 | for_each_online_cpu(cpu) {
|
| 7699 | pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
|
| 7700 | drain_zonestat(zone, pzstats);
|
| 7701 | }
|
| 7702 | free_percpu(zone->per_cpu_pageset);
|
| 7703 | zone->per_cpu_pageset = &boot_pageset;
|
| 7704 | if (zone->per_cpu_zonestats != &boot_zonestats) {
|
| 7705 | free_percpu(zone->per_cpu_zonestats);
|
| 7706 | zone->per_cpu_zonestats = &boot_zonestats;
|
| 7707 | }
|
| 7708 | }
|
| 7709 | }
|
| 7710 |
|
| 7711 | #ifdef CONFIG_MEMORY_HOTREMOVE
|
| 7712 | |
| 7713 | * All pages in the range must be in a single zone, must not contain holes,
|
| 7714 | * must span full sections, and must be isolated before calling this function.
|
| 7715 | *
|
| 7716 | * Returns the number of managed (non-PageOffline()) pages in the range: the
|
| 7717 | * number of pages for which memory offlining code must adjust managed page
|
| 7718 | * counters using adjust_managed_page_count().
|
| 7719 |
|
| 7720 | unsigned long __offline_isolated_pages(unsigned long start_pfn,
|
| 7721 | unsigned long end_pfn)
|
| 7722 | {
|
| 7723 | unsigned long already_offline = 0;
|
| 7724 | unsigned long pfn = start_pfn;
|
| 7725 | struct page *page;
|
| 7726 | struct zone *zone;
|
| 7727 | unsigned int order;
|
| 7728 |
|
| 7729 | offline_mem_sections(pfn, end_pfn);
|
| 7730 | zone = page_zone(pfn_to_page(pfn));
|
| 7731 | guard(spinlock_irqsave)(&zone->lock);
|
| 7732 | while (pfn < end_pfn) {
|
| 7733 | page = pfn_to_page(pfn);
|
| 7734 | |
| 7735 | * The HWPoisoned page may be not in buddy system, and
|
| 7736 | * page_count() is not 0.
|
| 7737 |
|
| 7738 | if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
|
| 7739 | pfn++;
|
| 7740 | continue;
|
| 7741 | }
|
| 7742 | |
| 7743 | * At this point all remaining PageOffline() pages have a
|
| 7744 | * reference count of 0 and can simply be skipped.
|
| 7745 |
|
| 7746 | if (PageOffline(page)) {
|
| 7747 | BUG_ON(page_count(page));
|
| 7748 | BUG_ON(PageBuddy(page));
|
| 7749 | already_offline++;
|
| 7750 | pfn++;
|
| 7751 | continue;
|
| 7752 | }
|
| 7753 |
|
| 7754 | BUG_ON(page_count(page));
|
| 7755 | BUG_ON(!PageBuddy(page));
|
| 7756 | VM_WARN_ON(get_pageblock_migratetype(page) != MIGRATE_ISOLATE);
|
| 7757 | order = buddy_order(page);
|
| 7758 | del_page_from_free_list(page, zone, order, MIGRATE_ISOLATE);
|
| 7759 | pfn += (1 << order);
|
| 7760 | }
|
| 7761 |
|
| 7762 | return end_pfn - start_pfn - already_offline;
|
| 7763 | }
|
| 7764 | #endif
|
| 7765 |
|
| 7766 | |
| 7767 | * This function returns a stable result only if called under zone lock.
|
| 7768 |
|
| 7769 | bool is_free_buddy_page(const struct page *page)
|
| 7770 | {
|
| 7771 | unsigned long pfn = page_to_pfn(page);
|
| 7772 | unsigned int order;
|
| 7773 |
|
| 7774 | for (order = 0; order < NR_PAGE_ORDERS; order++) {
|
| 7775 | const struct page *head = page - (pfn & ((1 << order) - 1));
|
| 7776 |
|
| 7777 | if (PageBuddy(head) &&
|
| 7778 | buddy_order_unsafe(head) >= order)
|
| 7779 | break;
|
| 7780 | }
|
| 7781 |
|
| 7782 | return order <= MAX_PAGE_ORDER;
|
| 7783 | }
|
| 7784 | EXPORT_SYMBOL(is_free_buddy_page);
|
| 7785 |
|
| 7786 | #ifdef CONFIG_MEMORY_FAILURE
|
| 7787 | static inline void add_to_free_list(struct page *page, struct zone *zone,
|
| 7788 | unsigned int order, int migratetype,
|
| 7789 | bool tail)
|
| 7790 | {
|
| 7791 | __add_to_free_list(page, zone, order, migratetype, tail);
|
| 7792 | account_freepages(zone, 1 << order, migratetype);
|
| 7793 | }
|
| 7794 |
|
| 7795 | |
| 7796 | * Break down a higher-order page in sub-pages, and keep our target out of
|
| 7797 | * buddy allocator.
|
| 7798 |
|
| 7799 | static void break_down_buddy_pages(struct zone *zone, struct page *page,
|
| 7800 | struct page *target, int low, int high,
|
| 7801 | int migratetype)
|
| 7802 | {
|
| 7803 | unsigned long size = 1 << high;
|
| 7804 | struct page *current_buddy;
|
| 7805 |
|
| 7806 | while (high > low) {
|
| 7807 | high--;
|
| 7808 | size >>= 1;
|
| 7809 |
|
| 7810 | if (target >= &page[size]) {
|
| 7811 | current_buddy = page;
|
| 7812 | page = page + size;
|
| 7813 | } else {
|
| 7814 | current_buddy = page + size;
|
| 7815 | }
|
| 7816 |
|
| 7817 | if (set_page_guard(zone, current_buddy, high))
|
| 7818 | continue;
|
| 7819 |
|
| 7820 | add_to_free_list(current_buddy, zone, high, migratetype, false);
|
| 7821 | set_buddy_order(current_buddy, high);
|
| 7822 | }
|
| 7823 | }
|
| 7824 |
|
| 7825 | |
| 7826 | * Take a page that will be marked as poisoned off the buddy allocator.
|
| 7827 |
|
| 7828 | bool take_page_off_buddy(struct page *page)
|
| 7829 | {
|
| 7830 | struct zone *zone = page_zone(page);
|
| 7831 | unsigned long pfn = page_to_pfn(page);
|
| 7832 | unsigned int order;
|
| 7833 |
|
| 7834 | guard(spinlock_irqsave)(&zone->lock);
|
| 7835 | for (order = 0; order < NR_PAGE_ORDERS; order++) {
|
| 7836 | struct page *page_head = page - (pfn & ((1 << order) - 1));
|
| 7837 | int page_order = buddy_order(page_head);
|
| 7838 |
|
| 7839 | if (PageBuddy(page_head) && page_order >= order) {
|
| 7840 | unsigned long pfn_head = page_to_pfn(page_head);
|
| 7841 | int migratetype = get_pfnblock_migratetype(page_head,
|
| 7842 | pfn_head);
|
| 7843 |
|
| 7844 | del_page_from_free_list(page_head, zone, page_order,
|
| 7845 | migratetype);
|
| 7846 | break_down_buddy_pages(zone, page_head, page, 0,
|
| 7847 | page_order, migratetype);
|
| 7848 | SetPageHWPoisonTakenOff(page);
|
| 7849 | return true;
|
| 7850 | }
|
| 7851 | if (page_count(page_head) > 0)
|
| 7852 | break;
|
| 7853 | }
|
| 7854 | return false;
|
| 7855 | }
|
| 7856 |
|
| 7857 | |
| 7858 | * Cancel takeoff done by take_page_off_buddy().
|
| 7859 |
|
| 7860 | bool put_page_back_buddy(struct page *page)
|
| 7861 | {
|
| 7862 | struct zone *zone = page_zone(page);
|
| 7863 |
|
| 7864 | guard(spinlock_irqsave)(&zone->lock);
|
| 7865 | if (put_page_testzero(page)) {
|
| 7866 | unsigned long pfn = page_to_pfn(page);
|
| 7867 | int migratetype = get_pfnblock_migratetype(page, pfn);
|
| 7868 |
|
| 7869 | ClearPageHWPoisonTakenOff(page);
|
| 7870 | __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
|
| 7871 | if (TestClearPageHWPoison(page))
|
| 7872 | return true;
|
| 7873 | }
|
| 7874 |
|
| 7875 | return false;
|
| 7876 | }
|
| 7877 | #endif
|
| 7878 |
|
| 7879 | bool has_managed_zone(enum zone_type zone)
|
| 7880 | {
|
| 7881 | struct pglist_data *pgdat;
|
| 7882 |
|
| 7883 | for_each_online_pgdat(pgdat) {
|
| 7884 | if (managed_zone(&pgdat->node_zones[zone]))
|
| 7885 | return true;
|
| 7886 | }
|
| 7887 | return false;
|
| 7888 | }
|
| 7889 |
|
| 7890 | #ifdef CONFIG_UNACCEPTED_MEMORY
|
| 7891 |
|
| 7892 | static bool lazy_accept = true;
|
| 7893 |
|
| 7894 | static int __init accept_memory_parse(char *p)
|
| 7895 | {
|
| 7896 | if (!strcmp(p, "lazy")) {
|
| 7897 | lazy_accept = true;
|
| 7898 | return 0;
|
| 7899 | } else if (!strcmp(p, "eager")) {
|
| 7900 | lazy_accept = false;
|
| 7901 | return 0;
|
| 7902 | } else {
|
| 7903 | return -EINVAL;
|
| 7904 | }
|
| 7905 | }
|
| 7906 | early_param("accept_memory", accept_memory_parse);
|
| 7907 |
|
| 7908 | static bool page_contains_unaccepted(struct page *page, unsigned int order)
|
| 7909 | {
|
| 7910 | phys_addr_t start = page_to_phys(page);
|
| 7911 |
|
| 7912 | return range_contains_unaccepted_memory(start, PAGE_SIZE << order);
|
| 7913 | }
|
| 7914 |
|
| 7915 | static void __accept_page(struct zone *zone, unsigned long *flags,
|
| 7916 | struct page *page)
|
| 7917 | {
|
| 7918 | list_del(&page->lru);
|
| 7919 | account_freepages(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
|
| 7920 | __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
|
| 7921 | __ClearPageUnaccepted(page);
|
| 7922 | spin_unlock_irqrestore(&zone->lock, *flags);
|
| 7923 |
|
| 7924 | accept_memory(page_to_phys(page), PAGE_SIZE << MAX_PAGE_ORDER);
|
| 7925 |
|
| 7926 | __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
|
| 7927 | }
|
| 7928 |
|
| 7929 | void accept_page(struct page *page)
|
| 7930 | {
|
| 7931 | struct zone *zone = page_zone(page);
|
| 7932 | unsigned long flags;
|
| 7933 |
|
| 7934 | spin_lock_irqsave(&zone->lock, flags);
|
| 7935 | if (!PageUnaccepted(page)) {
|
| 7936 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7937 | return;
|
| 7938 | }
|
| 7939 |
|
| 7940 |
|
| 7941 | __accept_page(zone, &flags, page);
|
| 7942 | }
|
| 7943 |
|
| 7944 | static bool try_to_accept_memory_one(struct zone *zone)
|
| 7945 | {
|
| 7946 | unsigned long flags;
|
| 7947 | struct page *page;
|
| 7948 |
|
| 7949 | spin_lock_irqsave(&zone->lock, flags);
|
| 7950 | page = list_first_entry_or_null(&zone->unaccepted_pages,
|
| 7951 | struct page, lru);
|
| 7952 | if (!page) {
|
| 7953 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 7954 | return false;
|
| 7955 | }
|
| 7956 |
|
| 7957 |
|
| 7958 | __accept_page(zone, &flags, page);
|
| 7959 |
|
| 7960 | return true;
|
| 7961 | }
|
| 7962 |
|
| 7963 | static bool cond_accept_memory(struct zone *zone, unsigned int order,
|
| 7964 | int alloc_flags)
|
| 7965 | {
|
| 7966 | long to_accept, wmark;
|
| 7967 | bool ret = false;
|
| 7968 |
|
| 7969 | if (list_empty(&zone->unaccepted_pages))
|
| 7970 | return false;
|
| 7971 |
|
| 7972 |
|
| 7973 | if (alloc_flags & ALLOC_NOLOCK)
|
| 7974 | return false;
|
| 7975 |
|
| 7976 | wmark = promo_wmark_pages(zone);
|
| 7977 |
|
| 7978 | |
| 7979 | * Watermarks have not been initialized yet.
|
| 7980 | *
|
| 7981 | * Accepting one MAX_ORDER page to ensure progress.
|
| 7982 |
|
| 7983 | if (!wmark)
|
| 7984 | return try_to_accept_memory_one(zone);
|
| 7985 |
|
| 7986 |
|
| 7987 | to_accept = wmark -
|
| 7988 | (zone_page_state(zone, NR_FREE_PAGES) -
|
| 7989 | __zone_watermark_unusable_free(zone, order, 0) -
|
| 7990 | zone_page_state(zone, NR_UNACCEPTED));
|
| 7991 |
|
| 7992 | while (to_accept > 0) {
|
| 7993 | if (!try_to_accept_memory_one(zone))
|
| 7994 | break;
|
| 7995 | ret = true;
|
| 7996 | to_accept -= MAX_ORDER_NR_PAGES;
|
| 7997 | }
|
| 7998 |
|
| 7999 | return ret;
|
| 8000 | }
|
| 8001 |
|
| 8002 | static bool __free_unaccepted(struct page *page)
|
| 8003 | {
|
| 8004 | struct zone *zone = page_zone(page);
|
| 8005 | unsigned long flags;
|
| 8006 |
|
| 8007 | if (!lazy_accept)
|
| 8008 | return false;
|
| 8009 |
|
| 8010 | spin_lock_irqsave(&zone->lock, flags);
|
| 8011 | list_add_tail(&page->lru, &zone->unaccepted_pages);
|
| 8012 | account_freepages(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
|
| 8013 | __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
|
| 8014 | __SetPageUnaccepted(page);
|
| 8015 | spin_unlock_irqrestore(&zone->lock, flags);
|
| 8016 |
|
| 8017 | return true;
|
| 8018 | }
|
| 8019 |
|
| 8020 | #else
|
| 8021 |
|
| 8022 | static bool page_contains_unaccepted(struct page *page, unsigned int order)
|
| 8023 | {
|
| 8024 | return false;
|
| 8025 | }
|
| 8026 |
|
| 8027 | static bool cond_accept_memory(struct zone *zone, unsigned int order,
|
| 8028 | int alloc_flags)
|
| 8029 | {
|
| 8030 | return false;
|
| 8031 | }
|
| 8032 |
|
| 8033 | static bool __free_unaccepted(struct page *page)
|
| 8034 | {
|
| 8035 | BUILD_BUG();
|
| 8036 | return false;
|
| 8037 | }
|
| 8038 |
|
| 8039 | #endif
|
| 8040 |
|
| 8041 | struct page *alloc_frozen_pages_nolock_noprof(gfp_t gfp_flags, int nid, unsigned int order)
|
| 8042 | {
|
| 8043 | if (nid == NUMA_NO_NODE)
|
| 8044 | nid = numa_node_id();
|
| 8045 |
|
| 8046 | return __alloc_frozen_pages_noprof(gfp_flags, order, nid, NULL, ALLOC_NOLOCK);
|
| 8047 | }
|
| 8048 | |
| 8049 | * alloc_pages_nolock - opportunistic reentrant allocation from any context
|
| 8050 | * @gfp_flags: GFP flags. Only __GFP_ACCOUNT, plus some flags that get set
|
| 8051 | * internally regardless (see %gfp_nolock) are allowed.
|
| 8052 | * @nid: node to allocate from
|
| 8053 | * @order: allocation order size
|
| 8054 | *
|
| 8055 | * Allocates pages of a given order from the given node. This is safe to
|
| 8056 | * call from any context where RCU is watching (from atomic, NMI, and also
|
| 8057 | * reentrant allocator -> tracepoint -> alloc_pages_nolock_noprof).
|
| 8058 | * Allocation is best effort and to be expected to fail easily so nobody should
|
| 8059 | * rely on the success. Failures are not reported via warn_alloc().
|
| 8060 | * See always fail conditions below.
|
| 8061 | *
|
| 8062 | * Return: allocated page or NULL on failure. NULL does not mean EBUSY or EAGAIN.
|
| 8063 | * It means ENOMEM. There is no reason to call it again and expect !NULL.
|
| 8064 |
|
| 8065 | struct page *alloc_pages_nolock_noprof(gfp_t gfp_flags, int nid, unsigned int order)
|
| 8066 | {
|
| 8067 | struct page *page;
|
| 8068 |
|
| 8069 | page = alloc_frozen_pages_nolock_noprof(gfp_flags, nid, order);
|
| 8070 | if (page)
|
| 8071 | set_page_refcounted(page);
|
| 8072 | return page;
|
| 8073 | }
|
| 8074 | EXPORT_SYMBOL_GPL(alloc_pages_nolock_noprof);
|