| 1 | /* |
| 2 | * linux/mm/page_alloc.c |
| 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/swap.h> |
| 20 | #include <linux/interrupt.h> |
| 21 | #include <linux/pagemap.h> |
| 22 | #include <linux/jiffies.h> |
| 23 | #include <linux/bootmem.h> |
| 24 | #include <linux/memblock.h> |
| 25 | #include <linux/compiler.h> |
| 26 | #include <linux/kernel.h> |
| 27 | #include <linux/kmemcheck.h> |
| 28 | #include <linux/module.h> |
| 29 | #include <linux/suspend.h> |
| 30 | #include <linux/pagevec.h> |
| 31 | #include <linux/blkdev.h> |
| 32 | #include <linux/slab.h> |
| 33 | #include <linux/ratelimit.h> |
| 34 | #include <linux/oom.h> |
| 35 | #include <linux/notifier.h> |
| 36 | #include <linux/topology.h> |
| 37 | #include <linux/sysctl.h> |
| 38 | #include <linux/cpu.h> |
| 39 | #include <linux/cpuset.h> |
| 40 | #include <linux/memory_hotplug.h> |
| 41 | #include <linux/nodemask.h> |
| 42 | #include <linux/vmalloc.h> |
| 43 | #include <linux/vmstat.h> |
| 44 | #include <linux/mempolicy.h> |
| 45 | #include <linux/stop_machine.h> |
| 46 | #include <linux/sort.h> |
| 47 | #include <linux/pfn.h> |
| 48 | #include <linux/backing-dev.h> |
| 49 | #include <linux/fault-inject.h> |
| 50 | #include <linux/page-isolation.h> |
| 51 | #include <linux/page_cgroup.h> |
| 52 | #include <linux/debugobjects.h> |
| 53 | #include <linux/kmemleak.h> |
| 54 | #include <linux/compaction.h> |
| 55 | #include <trace/events/kmem.h> |
| 56 | #include <linux/ftrace_event.h> |
| 57 | #include <linux/memcontrol.h> |
| 58 | #include <linux/prefetch.h> |
| 59 | #include <linux/mm_inline.h> |
| 60 | #include <linux/migrate.h> |
| 61 | #include <linux/page-debug-flags.h> |
| 62 | #include <linux/hugetlb.h> |
| 63 | #include <linux/sched/rt.h> |
| 64 | |
| 65 | #include <asm/sections.h> |
| 66 | #include <asm/tlbflush.h> |
| 67 | #include <asm/div64.h> |
| 68 | #include "internal.h" |
| 69 | |
| 70 | /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */ |
| 71 | static DEFINE_MUTEX(pcp_batch_high_lock); |
| 72 | |
| 73 | #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID |
| 74 | DEFINE_PER_CPU(int, numa_node); |
| 75 | EXPORT_PER_CPU_SYMBOL(numa_node); |
| 76 | #endif |
| 77 | |
| 78 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES |
| 79 | /* |
| 80 | * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. |
| 81 | * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. |
| 82 | * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() |
| 83 | * defined in <linux/topology.h>. |
| 84 | */ |
| 85 | DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ |
| 86 | EXPORT_PER_CPU_SYMBOL(_numa_mem_); |
| 87 | #endif |
| 88 | |
| 89 | /* |
| 90 | * Array of node states. |
| 91 | */ |
| 92 | nodemask_t node_states[NR_NODE_STATES] __read_mostly = { |
| 93 | [N_POSSIBLE] = NODE_MASK_ALL, |
| 94 | [N_ONLINE] = { { [0] = 1UL } }, |
| 95 | #ifndef CONFIG_NUMA |
| 96 | [N_NORMAL_MEMORY] = { { [0] = 1UL } }, |
| 97 | #ifdef CONFIG_HIGHMEM |
| 98 | [N_HIGH_MEMORY] = { { [0] = 1UL } }, |
| 99 | #endif |
| 100 | #ifdef CONFIG_MOVABLE_NODE |
| 101 | [N_MEMORY] = { { [0] = 1UL } }, |
| 102 | #endif |
| 103 | [N_CPU] = { { [0] = 1UL } }, |
| 104 | #endif /* NUMA */ |
| 105 | }; |
| 106 | EXPORT_SYMBOL(node_states); |
| 107 | |
| 108 | /* Protect totalram_pages and zone->managed_pages */ |
| 109 | static DEFINE_SPINLOCK(managed_page_count_lock); |
| 110 | |
| 111 | unsigned long totalram_pages __read_mostly; |
| 112 | unsigned long totalreserve_pages __read_mostly; |
| 113 | /* |
| 114 | * When calculating the number of globally allowed dirty pages, there |
| 115 | * is a certain number of per-zone reserves that should not be |
| 116 | * considered dirtyable memory. This is the sum of those reserves |
| 117 | * over all existing zones that contribute dirtyable memory. |
| 118 | */ |
| 119 | unsigned long dirty_balance_reserve __read_mostly; |
| 120 | |
| 121 | int percpu_pagelist_fraction; |
| 122 | gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; |
| 123 | |
| 124 | #ifdef CONFIG_PM_SLEEP |
| 125 | /* |
| 126 | * The following functions are used by the suspend/hibernate code to temporarily |
| 127 | * change gfp_allowed_mask in order to avoid using I/O during memory allocations |
| 128 | * while devices are suspended. To avoid races with the suspend/hibernate code, |
| 129 | * they should always be called with pm_mutex held (gfp_allowed_mask also should |
| 130 | * only be modified with pm_mutex held, unless the suspend/hibernate code is |
| 131 | * guaranteed not to run in parallel with that modification). |
| 132 | */ |
| 133 | |
| 134 | static gfp_t saved_gfp_mask; |
| 135 | |
| 136 | void pm_restore_gfp_mask(void) |
| 137 | { |
| 138 | WARN_ON(!mutex_is_locked(&pm_mutex)); |
| 139 | if (saved_gfp_mask) { |
| 140 | gfp_allowed_mask = saved_gfp_mask; |
| 141 | saved_gfp_mask = 0; |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | void pm_restrict_gfp_mask(void) |
| 146 | { |
| 147 | WARN_ON(!mutex_is_locked(&pm_mutex)); |
| 148 | WARN_ON(saved_gfp_mask); |
| 149 | saved_gfp_mask = gfp_allowed_mask; |
| 150 | gfp_allowed_mask &= ~GFP_IOFS; |
| 151 | } |
| 152 | |
| 153 | bool pm_suspended_storage(void) |
| 154 | { |
| 155 | if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS) |
| 156 | return false; |
| 157 | return true; |
| 158 | } |
| 159 | #endif /* CONFIG_PM_SLEEP */ |
| 160 | |
| 161 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE |
| 162 | int pageblock_order __read_mostly; |
| 163 | #endif |
| 164 | |
| 165 | static void __free_pages_ok(struct page *page, unsigned int order); |
| 166 | |
| 167 | /* |
| 168 | * results with 256, 32 in the lowmem_reserve sysctl: |
| 169 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) |
| 170 | * 1G machine -> (16M dma, 784M normal, 224M high) |
| 171 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA |
| 172 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL |
| 173 | * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA |
| 174 | * |
| 175 | * TBD: should special case ZONE_DMA32 machines here - in those we normally |
| 176 | * don't need any ZONE_NORMAL reservation |
| 177 | */ |
| 178 | int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { |
| 179 | #ifdef CONFIG_ZONE_DMA |
| 180 | 256, |
| 181 | #endif |
| 182 | #ifdef CONFIG_ZONE_DMA32 |
| 183 | 256, |
| 184 | #endif |
| 185 | #ifdef CONFIG_HIGHMEM |
| 186 | 32, |
| 187 | #endif |
| 188 | 32, |
| 189 | }; |
| 190 | |
| 191 | EXPORT_SYMBOL(totalram_pages); |
| 192 | |
| 193 | static char * const zone_names[MAX_NR_ZONES] = { |
| 194 | #ifdef CONFIG_ZONE_DMA |
| 195 | "DMA", |
| 196 | #endif |
| 197 | #ifdef CONFIG_ZONE_DMA32 |
| 198 | "DMA32", |
| 199 | #endif |
| 200 | "Normal", |
| 201 | #ifdef CONFIG_HIGHMEM |
| 202 | "HighMem", |
| 203 | #endif |
| 204 | "Movable", |
| 205 | }; |
| 206 | |
| 207 | int min_free_kbytes = 1024; |
| 208 | int user_min_free_kbytes = -1; |
| 209 | |
| 210 | static unsigned long __meminitdata nr_kernel_pages; |
| 211 | static unsigned long __meminitdata nr_all_pages; |
| 212 | static unsigned long __meminitdata dma_reserve; |
| 213 | |
| 214 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP |
| 215 | static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; |
| 216 | static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; |
| 217 | static unsigned long __initdata required_kernelcore; |
| 218 | static unsigned long __initdata required_movablecore; |
| 219 | static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; |
| 220 | |
| 221 | /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ |
| 222 | int movable_zone; |
| 223 | EXPORT_SYMBOL(movable_zone); |
| 224 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ |
| 225 | |
| 226 | #if MAX_NUMNODES > 1 |
| 227 | int nr_node_ids __read_mostly = MAX_NUMNODES; |
| 228 | int nr_online_nodes __read_mostly = 1; |
| 229 | EXPORT_SYMBOL(nr_node_ids); |
| 230 | EXPORT_SYMBOL(nr_online_nodes); |
| 231 | #endif |
| 232 | |
| 233 | int page_group_by_mobility_disabled __read_mostly; |
| 234 | |
| 235 | void set_pageblock_migratetype(struct page *page, int migratetype) |
| 236 | { |
| 237 | if (unlikely(page_group_by_mobility_disabled && |
| 238 | migratetype < MIGRATE_PCPTYPES)) |
| 239 | migratetype = MIGRATE_UNMOVABLE; |
| 240 | |
| 241 | set_pageblock_flags_group(page, (unsigned long)migratetype, |
| 242 | PB_migrate, PB_migrate_end); |
| 243 | } |
| 244 | |
| 245 | bool oom_killer_disabled __read_mostly; |
| 246 | |
| 247 | #ifdef CONFIG_DEBUG_VM |
| 248 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page) |
| 249 | { |
| 250 | int ret = 0; |
| 251 | unsigned seq; |
| 252 | unsigned long pfn = page_to_pfn(page); |
| 253 | unsigned long sp, start_pfn; |
| 254 | |
| 255 | do { |
| 256 | seq = zone_span_seqbegin(zone); |
| 257 | start_pfn = zone->zone_start_pfn; |
| 258 | sp = zone->spanned_pages; |
| 259 | if (!zone_spans_pfn(zone, pfn)) |
| 260 | ret = 1; |
| 261 | } while (zone_span_seqretry(zone, seq)); |
| 262 | |
| 263 | if (ret) |
| 264 | pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n", |
| 265 | pfn, zone_to_nid(zone), zone->name, |
| 266 | start_pfn, start_pfn + sp); |
| 267 | |
| 268 | return ret; |
| 269 | } |
| 270 | |
| 271 | static int page_is_consistent(struct zone *zone, struct page *page) |
| 272 | { |
| 273 | if (!pfn_valid_within(page_to_pfn(page))) |
| 274 | return 0; |
| 275 | if (zone != page_zone(page)) |
| 276 | return 0; |
| 277 | |
| 278 | return 1; |
| 279 | } |
| 280 | /* |
| 281 | * Temporary debugging check for pages not lying within a given zone. |
| 282 | */ |
| 283 | static int bad_range(struct zone *zone, struct page *page) |
| 284 | { |
| 285 | if (page_outside_zone_boundaries(zone, page)) |
| 286 | return 1; |
| 287 | if (!page_is_consistent(zone, page)) |
| 288 | return 1; |
| 289 | |
| 290 | return 0; |
| 291 | } |
| 292 | #else |
| 293 | static inline int bad_range(struct zone *zone, struct page *page) |
| 294 | { |
| 295 | return 0; |
| 296 | } |
| 297 | #endif |
| 298 | |
| 299 | static void bad_page(struct page *page, const char *reason, |
| 300 | unsigned long bad_flags) |
| 301 | { |
| 302 | static unsigned long resume; |
| 303 | static unsigned long nr_shown; |
| 304 | static unsigned long nr_unshown; |
| 305 | |
| 306 | /* Don't complain about poisoned pages */ |
| 307 | if (PageHWPoison(page)) { |
| 308 | page_mapcount_reset(page); /* remove PageBuddy */ |
| 309 | return; |
| 310 | } |
| 311 | |
| 312 | /* |
| 313 | * Allow a burst of 60 reports, then keep quiet for that minute; |
| 314 | * or allow a steady drip of one report per second. |
| 315 | */ |
| 316 | if (nr_shown == 60) { |
| 317 | if (time_before(jiffies, resume)) { |
| 318 | nr_unshown++; |
| 319 | goto out; |
| 320 | } |
| 321 | if (nr_unshown) { |
| 322 | printk(KERN_ALERT |
| 323 | "BUG: Bad page state: %lu messages suppressed\n", |
| 324 | nr_unshown); |
| 325 | nr_unshown = 0; |
| 326 | } |
| 327 | nr_shown = 0; |
| 328 | } |
| 329 | if (nr_shown++ == 0) |
| 330 | resume = jiffies + 60 * HZ; |
| 331 | |
| 332 | printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", |
| 333 | current->comm, page_to_pfn(page)); |
| 334 | dump_page_badflags(page, reason, bad_flags); |
| 335 | |
| 336 | print_modules(); |
| 337 | dump_stack(); |
| 338 | out: |
| 339 | /* Leave bad fields for debug, except PageBuddy could make trouble */ |
| 340 | page_mapcount_reset(page); /* remove PageBuddy */ |
| 341 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
| 342 | } |
| 343 | |
| 344 | /* |
| 345 | * Higher-order pages are called "compound pages". They are structured thusly: |
| 346 | * |
| 347 | * The first PAGE_SIZE page is called the "head page". |
| 348 | * |
| 349 | * The remaining PAGE_SIZE pages are called "tail pages". |
| 350 | * |
| 351 | * All pages have PG_compound set. All tail pages have their ->first_page |
| 352 | * pointing at the head page. |
| 353 | * |
| 354 | * The first tail page's ->lru.next holds the address of the compound page's |
| 355 | * put_page() function. Its ->lru.prev holds the order of allocation. |
| 356 | * This usage means that zero-order pages may not be compound. |
| 357 | */ |
| 358 | |
| 359 | static void free_compound_page(struct page *page) |
| 360 | { |
| 361 | __free_pages_ok(page, compound_order(page)); |
| 362 | } |
| 363 | |
| 364 | void prep_compound_page(struct page *page, unsigned long order) |
| 365 | { |
| 366 | int i; |
| 367 | int nr_pages = 1 << order; |
| 368 | |
| 369 | set_compound_page_dtor(page, free_compound_page); |
| 370 | set_compound_order(page, order); |
| 371 | __SetPageHead(page); |
| 372 | for (i = 1; i < nr_pages; i++) { |
| 373 | struct page *p = page + i; |
| 374 | set_page_count(p, 0); |
| 375 | p->first_page = page; |
| 376 | /* Make sure p->first_page is always valid for PageTail() */ |
| 377 | smp_wmb(); |
| 378 | __SetPageTail(p); |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | /* update __split_huge_page_refcount if you change this function */ |
| 383 | static int destroy_compound_page(struct page *page, unsigned long order) |
| 384 | { |
| 385 | int i; |
| 386 | int nr_pages = 1 << order; |
| 387 | int bad = 0; |
| 388 | |
| 389 | if (unlikely(compound_order(page) != order)) { |
| 390 | bad_page(page, "wrong compound order", 0); |
| 391 | bad++; |
| 392 | } |
| 393 | |
| 394 | __ClearPageHead(page); |
| 395 | |
| 396 | for (i = 1; i < nr_pages; i++) { |
| 397 | struct page *p = page + i; |
| 398 | |
| 399 | if (unlikely(!PageTail(p))) { |
| 400 | bad_page(page, "PageTail not set", 0); |
| 401 | bad++; |
| 402 | } else if (unlikely(p->first_page != page)) { |
| 403 | bad_page(page, "first_page not consistent", 0); |
| 404 | bad++; |
| 405 | } |
| 406 | __ClearPageTail(p); |
| 407 | } |
| 408 | |
| 409 | return bad; |
| 410 | } |
| 411 | |
| 412 | static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) |
| 413 | { |
| 414 | int i; |
| 415 | |
| 416 | /* |
| 417 | * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO |
| 418 | * and __GFP_HIGHMEM from hard or soft interrupt context. |
| 419 | */ |
| 420 | VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); |
| 421 | for (i = 0; i < (1 << order); i++) |
| 422 | clear_highpage(page + i); |
| 423 | } |
| 424 | |
| 425 | #ifdef CONFIG_DEBUG_PAGEALLOC |
| 426 | unsigned int _debug_guardpage_minorder; |
| 427 | |
| 428 | static int __init debug_guardpage_minorder_setup(char *buf) |
| 429 | { |
| 430 | unsigned long res; |
| 431 | |
| 432 | if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) { |
| 433 | printk(KERN_ERR "Bad debug_guardpage_minorder value\n"); |
| 434 | return 0; |
| 435 | } |
| 436 | _debug_guardpage_minorder = res; |
| 437 | printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res); |
| 438 | return 0; |
| 439 | } |
| 440 | __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup); |
| 441 | |
| 442 | static inline void set_page_guard_flag(struct page *page) |
| 443 | { |
| 444 | __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); |
| 445 | } |
| 446 | |
| 447 | static inline void clear_page_guard_flag(struct page *page) |
| 448 | { |
| 449 | __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); |
| 450 | } |
| 451 | #else |
| 452 | static inline void set_page_guard_flag(struct page *page) { } |
| 453 | static inline void clear_page_guard_flag(struct page *page) { } |
| 454 | #endif |
| 455 | |
| 456 | static inline void set_page_order(struct page *page, int order) |
| 457 | { |
| 458 | set_page_private(page, order); |
| 459 | __SetPageBuddy(page); |
| 460 | } |
| 461 | |
| 462 | static inline void rmv_page_order(struct page *page) |
| 463 | { |
| 464 | __ClearPageBuddy(page); |
| 465 | set_page_private(page, 0); |
| 466 | } |
| 467 | |
| 468 | /* |
| 469 | * Locate the struct page for both the matching buddy in our |
| 470 | * pair (buddy1) and the combined O(n+1) page they form (page). |
| 471 | * |
| 472 | * 1) Any buddy B1 will have an order O twin B2 which satisfies |
| 473 | * the following equation: |
| 474 | * B2 = B1 ^ (1 << O) |
| 475 | * For example, if the starting buddy (buddy2) is #8 its order |
| 476 | * 1 buddy is #10: |
| 477 | * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 |
| 478 | * |
| 479 | * 2) Any buddy B will have an order O+1 parent P which |
| 480 | * satisfies the following equation: |
| 481 | * P = B & ~(1 << O) |
| 482 | * |
| 483 | * Assumption: *_mem_map is contiguous at least up to MAX_ORDER |
| 484 | */ |
| 485 | static inline unsigned long |
| 486 | __find_buddy_index(unsigned long page_idx, unsigned int order) |
| 487 | { |
| 488 | return page_idx ^ (1 << order); |
| 489 | } |
| 490 | |
| 491 | /* |
| 492 | * This function checks whether a page is free && is the buddy |
| 493 | * we can do coalesce a page and its buddy if |
| 494 | * (a) the buddy is not in a hole && |
| 495 | * (b) the buddy is in the buddy system && |
| 496 | * (c) a page and its buddy have the same order && |
| 497 | * (d) a page and its buddy are in the same zone. |
| 498 | * |
| 499 | * For recording whether a page is in the buddy system, we set ->_mapcount |
| 500 | * PAGE_BUDDY_MAPCOUNT_VALUE. |
| 501 | * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is |
| 502 | * serialized by zone->lock. |
| 503 | * |
| 504 | * For recording page's order, we use page_private(page). |
| 505 | */ |
| 506 | static inline int page_is_buddy(struct page *page, struct page *buddy, |
| 507 | int order) |
| 508 | { |
| 509 | if (!pfn_valid_within(page_to_pfn(buddy))) |
| 510 | return 0; |
| 511 | |
| 512 | if (page_is_guard(buddy) && page_order(buddy) == order) { |
| 513 | VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); |
| 514 | |
| 515 | if (page_zone_id(page) != page_zone_id(buddy)) |
| 516 | return 0; |
| 517 | |
| 518 | return 1; |
| 519 | } |
| 520 | |
| 521 | if (PageBuddy(buddy) && page_order(buddy) == order) { |
| 522 | VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); |
| 523 | |
| 524 | /* |
| 525 | * zone check is done late to avoid uselessly |
| 526 | * calculating zone/node ids for pages that could |
| 527 | * never merge. |
| 528 | */ |
| 529 | if (page_zone_id(page) != page_zone_id(buddy)) |
| 530 | return 0; |
| 531 | |
| 532 | return 1; |
| 533 | } |
| 534 | return 0; |
| 535 | } |
| 536 | |
| 537 | /* |
| 538 | * Freeing function for a buddy system allocator. |
| 539 | * |
| 540 | * The concept of a buddy system is to maintain direct-mapped table |
| 541 | * (containing bit values) for memory blocks of various "orders". |
| 542 | * The bottom level table contains the map for the smallest allocatable |
| 543 | * units of memory (here, pages), and each level above it describes |
| 544 | * pairs of units from the levels below, hence, "buddies". |
| 545 | * At a high level, all that happens here is marking the table entry |
| 546 | * at the bottom level available, and propagating the changes upward |
| 547 | * as necessary, plus some accounting needed to play nicely with other |
| 548 | * parts of the VM system. |
| 549 | * At each level, we keep a list of pages, which are heads of continuous |
| 550 | * free pages of length of (1 << order) and marked with _mapcount |
| 551 | * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page) |
| 552 | * field. |
| 553 | * So when we are allocating or freeing one, we can derive the state of the |
| 554 | * other. That is, if we allocate a small block, and both were |
| 555 | * free, the remainder of the region must be split into blocks. |
| 556 | * If a block is freed, and its buddy is also free, then this |
| 557 | * triggers coalescing into a block of larger size. |
| 558 | * |
| 559 | * -- nyc |
| 560 | */ |
| 561 | |
| 562 | static inline void __free_one_page(struct page *page, |
| 563 | unsigned long pfn, |
| 564 | struct zone *zone, unsigned int order, |
| 565 | int migratetype) |
| 566 | { |
| 567 | unsigned long page_idx; |
| 568 | unsigned long combined_idx; |
| 569 | unsigned long uninitialized_var(buddy_idx); |
| 570 | struct page *buddy; |
| 571 | |
| 572 | VM_BUG_ON(!zone_is_initialized(zone)); |
| 573 | |
| 574 | if (unlikely(PageCompound(page))) |
| 575 | if (unlikely(destroy_compound_page(page, order))) |
| 576 | return; |
| 577 | |
| 578 | VM_BUG_ON(migratetype == -1); |
| 579 | |
| 580 | page_idx = pfn & ((1 << MAX_ORDER) - 1); |
| 581 | |
| 582 | VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page); |
| 583 | VM_BUG_ON_PAGE(bad_range(zone, page), page); |
| 584 | |
| 585 | while (order < MAX_ORDER-1) { |
| 586 | buddy_idx = __find_buddy_index(page_idx, order); |
| 587 | buddy = page + (buddy_idx - page_idx); |
| 588 | if (!page_is_buddy(page, buddy, order)) |
| 589 | break; |
| 590 | /* |
| 591 | * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, |
| 592 | * merge with it and move up one order. |
| 593 | */ |
| 594 | if (page_is_guard(buddy)) { |
| 595 | clear_page_guard_flag(buddy); |
| 596 | set_page_private(page, 0); |
| 597 | __mod_zone_freepage_state(zone, 1 << order, |
| 598 | migratetype); |
| 599 | } else { |
| 600 | list_del(&buddy->lru); |
| 601 | zone->free_area[order].nr_free--; |
| 602 | rmv_page_order(buddy); |
| 603 | } |
| 604 | combined_idx = buddy_idx & page_idx; |
| 605 | page = page + (combined_idx - page_idx); |
| 606 | page_idx = combined_idx; |
| 607 | order++; |
| 608 | } |
| 609 | set_page_order(page, order); |
| 610 | |
| 611 | /* |
| 612 | * If this is not the largest possible page, check if the buddy |
| 613 | * of the next-highest order is free. If it is, it's possible |
| 614 | * that pages are being freed that will coalesce soon. In case, |
| 615 | * that is happening, add the free page to the tail of the list |
| 616 | * so it's less likely to be used soon and more likely to be merged |
| 617 | * as a higher order page |
| 618 | */ |
| 619 | if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { |
| 620 | struct page *higher_page, *higher_buddy; |
| 621 | combined_idx = buddy_idx & page_idx; |
| 622 | higher_page = page + (combined_idx - page_idx); |
| 623 | buddy_idx = __find_buddy_index(combined_idx, order + 1); |
| 624 | higher_buddy = higher_page + (buddy_idx - combined_idx); |
| 625 | if (page_is_buddy(higher_page, higher_buddy, order + 1)) { |
| 626 | list_add_tail(&page->lru, |
| 627 | &zone->free_area[order].free_list[migratetype]); |
| 628 | goto out; |
| 629 | } |
| 630 | } |
| 631 | |
| 632 | list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); |
| 633 | out: |
| 634 | zone->free_area[order].nr_free++; |
| 635 | } |
| 636 | |
| 637 | static inline int free_pages_check(struct page *page) |
| 638 | { |
| 639 | const char *bad_reason = NULL; |
| 640 | unsigned long bad_flags = 0; |
| 641 | |
| 642 | if (unlikely(page_mapcount(page))) |
| 643 | bad_reason = "nonzero mapcount"; |
| 644 | if (unlikely(page->mapping != NULL)) |
| 645 | bad_reason = "non-NULL mapping"; |
| 646 | if (unlikely(atomic_read(&page->_count) != 0)) |
| 647 | bad_reason = "nonzero _count"; |
| 648 | if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) { |
| 649 | bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set"; |
| 650 | bad_flags = PAGE_FLAGS_CHECK_AT_FREE; |
| 651 | } |
| 652 | if (unlikely(mem_cgroup_bad_page_check(page))) |
| 653 | bad_reason = "cgroup check failed"; |
| 654 | if (unlikely(bad_reason)) { |
| 655 | bad_page(page, bad_reason, bad_flags); |
| 656 | return 1; |
| 657 | } |
| 658 | page_cpupid_reset_last(page); |
| 659 | if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
| 660 | page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; |
| 661 | return 0; |
| 662 | } |
| 663 | |
| 664 | /* |
| 665 | * Frees a number of pages from the PCP lists |
| 666 | * Assumes all pages on list are in same zone, and of same order. |
| 667 | * count is the number of pages to free. |
| 668 | * |
| 669 | * If the zone was previously in an "all pages pinned" state then look to |
| 670 | * see if this freeing clears that state. |
| 671 | * |
| 672 | * And clear the zone's pages_scanned counter, to hold off the "all pages are |
| 673 | * pinned" detection logic. |
| 674 | */ |
| 675 | static void free_pcppages_bulk(struct zone *zone, int count, |
| 676 | struct per_cpu_pages *pcp) |
| 677 | { |
| 678 | int migratetype = 0; |
| 679 | int batch_free = 0; |
| 680 | int to_free = count; |
| 681 | |
| 682 | spin_lock(&zone->lock); |
| 683 | zone->pages_scanned = 0; |
| 684 | |
| 685 | while (to_free) { |
| 686 | struct page *page; |
| 687 | struct list_head *list; |
| 688 | |
| 689 | /* |
| 690 | * Remove pages from lists in a round-robin fashion. A |
| 691 | * batch_free count is maintained that is incremented when an |
| 692 | * empty list is encountered. This is so more pages are freed |
| 693 | * off fuller lists instead of spinning excessively around empty |
| 694 | * lists |
| 695 | */ |
| 696 | do { |
| 697 | batch_free++; |
| 698 | if (++migratetype == MIGRATE_PCPTYPES) |
| 699 | migratetype = 0; |
| 700 | list = &pcp->lists[migratetype]; |
| 701 | } while (list_empty(list)); |
| 702 | |
| 703 | /* This is the only non-empty list. Free them all. */ |
| 704 | if (batch_free == MIGRATE_PCPTYPES) |
| 705 | batch_free = to_free; |
| 706 | |
| 707 | do { |
| 708 | int mt; /* migratetype of the to-be-freed page */ |
| 709 | |
| 710 | page = list_entry(list->prev, struct page, lru); |
| 711 | /* must delete as __free_one_page list manipulates */ |
| 712 | list_del(&page->lru); |
| 713 | mt = get_freepage_migratetype(page); |
| 714 | /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ |
| 715 | __free_one_page(page, page_to_pfn(page), zone, 0, mt); |
| 716 | trace_mm_page_pcpu_drain(page, 0, mt); |
| 717 | if (likely(!is_migrate_isolate_page(page))) { |
| 718 | __mod_zone_page_state(zone, NR_FREE_PAGES, 1); |
| 719 | if (is_migrate_cma(mt)) |
| 720 | __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1); |
| 721 | } |
| 722 | } while (--to_free && --batch_free && !list_empty(list)); |
| 723 | } |
| 724 | spin_unlock(&zone->lock); |
| 725 | } |
| 726 | |
| 727 | static void free_one_page(struct zone *zone, |
| 728 | struct page *page, unsigned long pfn, |
| 729 | int order, |
| 730 | int migratetype) |
| 731 | { |
| 732 | spin_lock(&zone->lock); |
| 733 | zone->pages_scanned = 0; |
| 734 | |
| 735 | __free_one_page(page, pfn, zone, order, migratetype); |
| 736 | if (unlikely(!is_migrate_isolate(migratetype))) |
| 737 | __mod_zone_freepage_state(zone, 1 << order, migratetype); |
| 738 | spin_unlock(&zone->lock); |
| 739 | } |
| 740 | |
| 741 | static bool free_pages_prepare(struct page *page, unsigned int order) |
| 742 | { |
| 743 | int i; |
| 744 | int bad = 0; |
| 745 | |
| 746 | trace_mm_page_free(page, order); |
| 747 | kmemcheck_free_shadow(page, order); |
| 748 | |
| 749 | if (PageAnon(page)) |
| 750 | page->mapping = NULL; |
| 751 | for (i = 0; i < (1 << order); i++) |
| 752 | bad += free_pages_check(page + i); |
| 753 | if (bad) |
| 754 | return false; |
| 755 | |
| 756 | if (!PageHighMem(page)) { |
| 757 | debug_check_no_locks_freed(page_address(page), |
| 758 | PAGE_SIZE << order); |
| 759 | debug_check_no_obj_freed(page_address(page), |
| 760 | PAGE_SIZE << order); |
| 761 | } |
| 762 | arch_free_page(page, order); |
| 763 | kernel_map_pages(page, 1 << order, 0); |
| 764 | |
| 765 | return true; |
| 766 | } |
| 767 | |
| 768 | static void __free_pages_ok(struct page *page, unsigned int order) |
| 769 | { |
| 770 | unsigned long flags; |
| 771 | int migratetype; |
| 772 | unsigned long pfn = page_to_pfn(page); |
| 773 | |
| 774 | if (!free_pages_prepare(page, order)) |
| 775 | return; |
| 776 | |
| 777 | local_irq_save(flags); |
| 778 | __count_vm_events(PGFREE, 1 << order); |
| 779 | migratetype = get_pfnblock_migratetype(page, pfn); |
| 780 | set_freepage_migratetype(page, migratetype); |
| 781 | free_one_page(page_zone(page), page, pfn, order, migratetype); |
| 782 | local_irq_restore(flags); |
| 783 | } |
| 784 | |
| 785 | void __init __free_pages_bootmem(struct page *page, unsigned int order) |
| 786 | { |
| 787 | unsigned int nr_pages = 1 << order; |
| 788 | struct page *p = page; |
| 789 | unsigned int loop; |
| 790 | |
| 791 | prefetchw(p); |
| 792 | for (loop = 0; loop < (nr_pages - 1); loop++, p++) { |
| 793 | prefetchw(p + 1); |
| 794 | __ClearPageReserved(p); |
| 795 | set_page_count(p, 0); |
| 796 | } |
| 797 | __ClearPageReserved(p); |
| 798 | set_page_count(p, 0); |
| 799 | |
| 800 | page_zone(page)->managed_pages += nr_pages; |
| 801 | set_page_refcounted(page); |
| 802 | __free_pages(page, order); |
| 803 | } |
| 804 | |
| 805 | #ifdef CONFIG_CMA |
| 806 | /* Free whole pageblock and set its migration type to MIGRATE_CMA. */ |
| 807 | void __init init_cma_reserved_pageblock(struct page *page) |
| 808 | { |
| 809 | unsigned i = pageblock_nr_pages; |
| 810 | struct page *p = page; |
| 811 | |
| 812 | do { |
| 813 | __ClearPageReserved(p); |
| 814 | set_page_count(p, 0); |
| 815 | } while (++p, --i); |
| 816 | |
| 817 | set_page_refcounted(page); |
| 818 | set_pageblock_migratetype(page, MIGRATE_CMA); |
| 819 | __free_pages(page, pageblock_order); |
| 820 | adjust_managed_page_count(page, pageblock_nr_pages); |
| 821 | } |
| 822 | #endif |
| 823 | |
| 824 | /* |
| 825 | * The order of subdivision here is critical for the IO subsystem. |
| 826 | * Please do not alter this order without good reasons and regression |
| 827 | * testing. Specifically, as large blocks of memory are subdivided, |
| 828 | * the order in which smaller blocks are delivered depends on the order |
| 829 | * they're subdivided in this function. This is the primary factor |
| 830 | * influencing the order in which pages are delivered to the IO |
| 831 | * subsystem according to empirical testing, and this is also justified |
| 832 | * by considering the behavior of a buddy system containing a single |
| 833 | * large block of memory acted on by a series of small allocations. |
| 834 | * This behavior is a critical factor in sglist merging's success. |
| 835 | * |
| 836 | * -- nyc |
| 837 | */ |
| 838 | static inline void expand(struct zone *zone, struct page *page, |
| 839 | int low, int high, struct free_area *area, |
| 840 | int migratetype) |
| 841 | { |
| 842 | unsigned long size = 1 << high; |
| 843 | |
| 844 | while (high > low) { |
| 845 | area--; |
| 846 | high--; |
| 847 | size >>= 1; |
| 848 | VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]); |
| 849 | |
| 850 | #ifdef CONFIG_DEBUG_PAGEALLOC |
| 851 | if (high < debug_guardpage_minorder()) { |
| 852 | /* |
| 853 | * Mark as guard pages (or page), that will allow to |
| 854 | * merge back to allocator when buddy will be freed. |
| 855 | * Corresponding page table entries will not be touched, |
| 856 | * pages will stay not present in virtual address space |
| 857 | */ |
| 858 | INIT_LIST_HEAD(&page[size].lru); |
| 859 | set_page_guard_flag(&page[size]); |
| 860 | set_page_private(&page[size], high); |
| 861 | /* Guard pages are not available for any usage */ |
| 862 | __mod_zone_freepage_state(zone, -(1 << high), |
| 863 | migratetype); |
| 864 | continue; |
| 865 | } |
| 866 | #endif |
| 867 | list_add(&page[size].lru, &area->free_list[migratetype]); |
| 868 | area->nr_free++; |
| 869 | set_page_order(&page[size], high); |
| 870 | } |
| 871 | } |
| 872 | |
| 873 | /* |
| 874 | * This page is about to be returned from the page allocator |
| 875 | */ |
| 876 | static inline int check_new_page(struct page *page) |
| 877 | { |
| 878 | const char *bad_reason = NULL; |
| 879 | unsigned long bad_flags = 0; |
| 880 | |
| 881 | if (unlikely(page_mapcount(page))) |
| 882 | bad_reason = "nonzero mapcount"; |
| 883 | if (unlikely(page->mapping != NULL)) |
| 884 | bad_reason = "non-NULL mapping"; |
| 885 | if (unlikely(atomic_read(&page->_count) != 0)) |
| 886 | bad_reason = "nonzero _count"; |
| 887 | if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) { |
| 888 | bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set"; |
| 889 | bad_flags = PAGE_FLAGS_CHECK_AT_PREP; |
| 890 | } |
| 891 | if (unlikely(mem_cgroup_bad_page_check(page))) |
| 892 | bad_reason = "cgroup check failed"; |
| 893 | if (unlikely(bad_reason)) { |
| 894 | bad_page(page, bad_reason, bad_flags); |
| 895 | return 1; |
| 896 | } |
| 897 | return 0; |
| 898 | } |
| 899 | |
| 900 | static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) |
| 901 | { |
| 902 | int i; |
| 903 | |
| 904 | for (i = 0; i < (1 << order); i++) { |
| 905 | struct page *p = page + i; |
| 906 | if (unlikely(check_new_page(p))) |
| 907 | return 1; |
| 908 | } |
| 909 | |
| 910 | set_page_private(page, 0); |
| 911 | set_page_refcounted(page); |
| 912 | |
| 913 | arch_alloc_page(page, order); |
| 914 | kernel_map_pages(page, 1 << order, 1); |
| 915 | |
| 916 | if (gfp_flags & __GFP_ZERO) |
| 917 | prep_zero_page(page, order, gfp_flags); |
| 918 | |
| 919 | if (order && (gfp_flags & __GFP_COMP)) |
| 920 | prep_compound_page(page, order); |
| 921 | |
| 922 | return 0; |
| 923 | } |
| 924 | |
| 925 | /* |
| 926 | * Go through the free lists for the given migratetype and remove |
| 927 | * the smallest available page from the freelists |
| 928 | */ |
| 929 | static inline |
| 930 | struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, |
| 931 | int migratetype) |
| 932 | { |
| 933 | unsigned int current_order; |
| 934 | struct free_area *area; |
| 935 | struct page *page; |
| 936 | |
| 937 | /* Find a page of the appropriate size in the preferred list */ |
| 938 | for (current_order = order; current_order < MAX_ORDER; ++current_order) { |
| 939 | area = &(zone->free_area[current_order]); |
| 940 | if (list_empty(&area->free_list[migratetype])) |
| 941 | continue; |
| 942 | |
| 943 | page = list_entry(area->free_list[migratetype].next, |
| 944 | struct page, lru); |
| 945 | list_del(&page->lru); |
| 946 | rmv_page_order(page); |
| 947 | area->nr_free--; |
| 948 | expand(zone, page, order, current_order, area, migratetype); |
| 949 | set_freepage_migratetype(page, migratetype); |
| 950 | return page; |
| 951 | } |
| 952 | |
| 953 | return NULL; |
| 954 | } |
| 955 | |
| 956 | |
| 957 | /* |
| 958 | * This array describes the order lists are fallen back to when |
| 959 | * the free lists for the desirable migrate type are depleted |
| 960 | */ |
| 961 | static int fallbacks[MIGRATE_TYPES][4] = { |
| 962 | [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, |
| 963 | [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, |
| 964 | #ifdef CONFIG_CMA |
| 965 | [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, |
| 966 | [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */ |
| 967 | #else |
| 968 | [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, |
| 969 | #endif |
| 970 | [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */ |
| 971 | #ifdef CONFIG_MEMORY_ISOLATION |
| 972 | [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */ |
| 973 | #endif |
| 974 | }; |
| 975 | |
| 976 | /* |
| 977 | * Move the free pages in a range to the free lists of the requested type. |
| 978 | * Note that start_page and end_pages are not aligned on a pageblock |
| 979 | * boundary. If alignment is required, use move_freepages_block() |
| 980 | */ |
| 981 | int move_freepages(struct zone *zone, |
| 982 | struct page *start_page, struct page *end_page, |
| 983 | int migratetype) |
| 984 | { |
| 985 | struct page *page; |
| 986 | unsigned long order; |
| 987 | int pages_moved = 0; |
| 988 | |
| 989 | #ifndef CONFIG_HOLES_IN_ZONE |
| 990 | /* |
| 991 | * page_zone is not safe to call in this context when |
| 992 | * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant |
| 993 | * anyway as we check zone boundaries in move_freepages_block(). |
| 994 | * Remove at a later date when no bug reports exist related to |
| 995 | * grouping pages by mobility |
| 996 | */ |
| 997 | BUG_ON(page_zone(start_page) != page_zone(end_page)); |
| 998 | #endif |
| 999 | |
| 1000 | for (page = start_page; page <= end_page;) { |
| 1001 | /* Make sure we are not inadvertently changing nodes */ |
| 1002 | VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page); |
| 1003 | |
| 1004 | if (!pfn_valid_within(page_to_pfn(page))) { |
| 1005 | page++; |
| 1006 | continue; |
| 1007 | } |
| 1008 | |
| 1009 | if (!PageBuddy(page)) { |
| 1010 | page++; |
| 1011 | continue; |
| 1012 | } |
| 1013 | |
| 1014 | order = page_order(page); |
| 1015 | list_move(&page->lru, |
| 1016 | &zone->free_area[order].free_list[migratetype]); |
| 1017 | set_freepage_migratetype(page, migratetype); |
| 1018 | page += 1 << order; |
| 1019 | pages_moved += 1 << order; |
| 1020 | } |
| 1021 | |
| 1022 | return pages_moved; |
| 1023 | } |
| 1024 | |
| 1025 | int move_freepages_block(struct zone *zone, struct page *page, |
| 1026 | int migratetype) |
| 1027 | { |
| 1028 | unsigned long start_pfn, end_pfn; |
| 1029 | struct page *start_page, *end_page; |
| 1030 | |
| 1031 | start_pfn = page_to_pfn(page); |
| 1032 | start_pfn = start_pfn & ~(pageblock_nr_pages-1); |
| 1033 | start_page = pfn_to_page(start_pfn); |
| 1034 | end_page = start_page + pageblock_nr_pages - 1; |
| 1035 | end_pfn = start_pfn + pageblock_nr_pages - 1; |
| 1036 | |
| 1037 | /* Do not cross zone boundaries */ |
| 1038 | if (!zone_spans_pfn(zone, start_pfn)) |
| 1039 | start_page = page; |
| 1040 | if (!zone_spans_pfn(zone, end_pfn)) |
| 1041 | return 0; |
| 1042 | |
| 1043 | return move_freepages(zone, start_page, end_page, migratetype); |
| 1044 | } |
| 1045 | |
| 1046 | static void change_pageblock_range(struct page *pageblock_page, |
| 1047 | int start_order, int migratetype) |
| 1048 | { |
| 1049 | int nr_pageblocks = 1 << (start_order - pageblock_order); |
| 1050 | |
| 1051 | while (nr_pageblocks--) { |
| 1052 | set_pageblock_migratetype(pageblock_page, migratetype); |
| 1053 | pageblock_page += pageblock_nr_pages; |
| 1054 | } |
| 1055 | } |
| 1056 | |
| 1057 | /* |
| 1058 | * If breaking a large block of pages, move all free pages to the preferred |
| 1059 | * allocation list. If falling back for a reclaimable kernel allocation, be |
| 1060 | * more aggressive about taking ownership of free pages. |
| 1061 | * |
| 1062 | * On the other hand, never change migration type of MIGRATE_CMA pageblocks |
| 1063 | * nor move CMA pages to different free lists. We don't want unmovable pages |
| 1064 | * to be allocated from MIGRATE_CMA areas. |
| 1065 | * |
| 1066 | * Returns the new migratetype of the pageblock (or the same old migratetype |
| 1067 | * if it was unchanged). |
| 1068 | */ |
| 1069 | static int try_to_steal_freepages(struct zone *zone, struct page *page, |
| 1070 | int start_type, int fallback_type) |
| 1071 | { |
| 1072 | int current_order = page_order(page); |
| 1073 | |
| 1074 | /* |
| 1075 | * When borrowing from MIGRATE_CMA, we need to release the excess |
| 1076 | * buddy pages to CMA itself. We also ensure the freepage_migratetype |
| 1077 | * is set to CMA so it is returned to the correct freelist in case |
| 1078 | * the page ends up being not actually allocated from the pcp lists. |
| 1079 | */ |
| 1080 | if (is_migrate_cma(fallback_type)) |
| 1081 | return fallback_type; |
| 1082 | |
| 1083 | /* Take ownership for orders >= pageblock_order */ |
| 1084 | if (current_order >= pageblock_order) { |
| 1085 | change_pageblock_range(page, current_order, start_type); |
| 1086 | return start_type; |
| 1087 | } |
| 1088 | |
| 1089 | if (current_order >= pageblock_order / 2 || |
| 1090 | start_type == MIGRATE_RECLAIMABLE || |
| 1091 | page_group_by_mobility_disabled) { |
| 1092 | int pages; |
| 1093 | |
| 1094 | pages = move_freepages_block(zone, page, start_type); |
| 1095 | |
| 1096 | /* Claim the whole block if over half of it is free */ |
| 1097 | if (pages >= (1 << (pageblock_order-1)) || |
| 1098 | page_group_by_mobility_disabled) { |
| 1099 | |
| 1100 | set_pageblock_migratetype(page, start_type); |
| 1101 | return start_type; |
| 1102 | } |
| 1103 | |
| 1104 | } |
| 1105 | |
| 1106 | return fallback_type; |
| 1107 | } |
| 1108 | |
| 1109 | /* Remove an element from the buddy allocator from the fallback list */ |
| 1110 | static inline struct page * |
| 1111 | __rmqueue_fallback(struct zone *zone, int order, int start_migratetype) |
| 1112 | { |
| 1113 | struct free_area *area; |
| 1114 | int current_order; |
| 1115 | struct page *page; |
| 1116 | int migratetype, new_type, i; |
| 1117 | |
| 1118 | /* Find the largest possible block of pages in the other list */ |
| 1119 | for (current_order = MAX_ORDER-1; current_order >= order; |
| 1120 | --current_order) { |
| 1121 | for (i = 0;; i++) { |
| 1122 | migratetype = fallbacks[start_migratetype][i]; |
| 1123 | |
| 1124 | /* MIGRATE_RESERVE handled later if necessary */ |
| 1125 | if (migratetype == MIGRATE_RESERVE) |
| 1126 | break; |
| 1127 | |
| 1128 | area = &(zone->free_area[current_order]); |
| 1129 | if (list_empty(&area->free_list[migratetype])) |
| 1130 | continue; |
| 1131 | |
| 1132 | page = list_entry(area->free_list[migratetype].next, |
| 1133 | struct page, lru); |
| 1134 | area->nr_free--; |
| 1135 | |
| 1136 | new_type = try_to_steal_freepages(zone, page, |
| 1137 | start_migratetype, |
| 1138 | migratetype); |
| 1139 | |
| 1140 | /* Remove the page from the freelists */ |
| 1141 | list_del(&page->lru); |
| 1142 | rmv_page_order(page); |
| 1143 | |
| 1144 | expand(zone, page, order, current_order, area, |
| 1145 | new_type); |
| 1146 | /* The freepage_migratetype may differ from pageblock's |
| 1147 | * migratetype depending on the decisions in |
| 1148 | * try_to_steal_freepages. This is OK as long as it does |
| 1149 | * not differ for MIGRATE_CMA type. |
| 1150 | */ |
| 1151 | set_freepage_migratetype(page, new_type); |
| 1152 | |
| 1153 | trace_mm_page_alloc_extfrag(page, order, current_order, |
| 1154 | start_migratetype, migratetype, new_type); |
| 1155 | |
| 1156 | return page; |
| 1157 | } |
| 1158 | } |
| 1159 | |
| 1160 | return NULL; |
| 1161 | } |
| 1162 | |
| 1163 | /* |
| 1164 | * Do the hard work of removing an element from the buddy allocator. |
| 1165 | * Call me with the zone->lock already held. |
| 1166 | */ |
| 1167 | static struct page *__rmqueue(struct zone *zone, unsigned int order, |
| 1168 | int migratetype) |
| 1169 | { |
| 1170 | struct page *page; |
| 1171 | |
| 1172 | retry_reserve: |
| 1173 | page = __rmqueue_smallest(zone, order, migratetype); |
| 1174 | |
| 1175 | if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { |
| 1176 | page = __rmqueue_fallback(zone, order, migratetype); |
| 1177 | |
| 1178 | /* |
| 1179 | * Use MIGRATE_RESERVE rather than fail an allocation. goto |
| 1180 | * is used because __rmqueue_smallest is an inline function |
| 1181 | * and we want just one call site |
| 1182 | */ |
| 1183 | if (!page) { |
| 1184 | migratetype = MIGRATE_RESERVE; |
| 1185 | goto retry_reserve; |
| 1186 | } |
| 1187 | } |
| 1188 | |
| 1189 | trace_mm_page_alloc_zone_locked(page, order, migratetype); |
| 1190 | return page; |
| 1191 | } |
| 1192 | |
| 1193 | /* |
| 1194 | * Obtain a specified number of elements from the buddy allocator, all under |
| 1195 | * a single hold of the lock, for efficiency. Add them to the supplied list. |
| 1196 | * Returns the number of new pages which were placed at *list. |
| 1197 | */ |
| 1198 | static int rmqueue_bulk(struct zone *zone, unsigned int order, |
| 1199 | unsigned long count, struct list_head *list, |
| 1200 | int migratetype, int cold) |
| 1201 | { |
| 1202 | int i; |
| 1203 | |
| 1204 | spin_lock(&zone->lock); |
| 1205 | for (i = 0; i < count; ++i) { |
| 1206 | struct page *page = __rmqueue(zone, order, migratetype); |
| 1207 | if (unlikely(page == NULL)) |
| 1208 | break; |
| 1209 | |
| 1210 | /* |
| 1211 | * Split buddy pages returned by expand() are received here |
| 1212 | * in physical page order. The page is added to the callers and |
| 1213 | * list and the list head then moves forward. From the callers |
| 1214 | * perspective, the linked list is ordered by page number in |
| 1215 | * some conditions. This is useful for IO devices that can |
| 1216 | * merge IO requests if the physical pages are ordered |
| 1217 | * properly. |
| 1218 | */ |
| 1219 | if (likely(cold == 0)) |
| 1220 | list_add(&page->lru, list); |
| 1221 | else |
| 1222 | list_add_tail(&page->lru, list); |
| 1223 | list = &page->lru; |
| 1224 | if (is_migrate_cma(get_freepage_migratetype(page))) |
| 1225 | __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, |
| 1226 | -(1 << order)); |
| 1227 | } |
| 1228 | __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); |
| 1229 | spin_unlock(&zone->lock); |
| 1230 | return i; |
| 1231 | } |
| 1232 | |
| 1233 | #ifdef CONFIG_NUMA |
| 1234 | /* |
| 1235 | * Called from the vmstat counter updater to drain pagesets of this |
| 1236 | * currently executing processor on remote nodes after they have |
| 1237 | * expired. |
| 1238 | * |
| 1239 | * Note that this function must be called with the thread pinned to |
| 1240 | * a single processor. |
| 1241 | */ |
| 1242 | void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) |
| 1243 | { |
| 1244 | unsigned long flags; |
| 1245 | int to_drain; |
| 1246 | unsigned long batch; |
| 1247 | |
| 1248 | local_irq_save(flags); |
| 1249 | batch = ACCESS_ONCE(pcp->batch); |
| 1250 | if (pcp->count >= batch) |
| 1251 | to_drain = batch; |
| 1252 | else |
| 1253 | to_drain = pcp->count; |
| 1254 | if (to_drain > 0) { |
| 1255 | free_pcppages_bulk(zone, to_drain, pcp); |
| 1256 | pcp->count -= to_drain; |
| 1257 | } |
| 1258 | local_irq_restore(flags); |
| 1259 | } |
| 1260 | #endif |
| 1261 | |
| 1262 | /* |
| 1263 | * Drain pages of the indicated processor. |
| 1264 | * |
| 1265 | * The processor must either be the current processor and the |
| 1266 | * thread pinned to the current processor or a processor that |
| 1267 | * is not online. |
| 1268 | */ |
| 1269 | static void drain_pages(unsigned int cpu) |
| 1270 | { |
| 1271 | unsigned long flags; |
| 1272 | struct zone *zone; |
| 1273 | |
| 1274 | for_each_populated_zone(zone) { |
| 1275 | struct per_cpu_pageset *pset; |
| 1276 | struct per_cpu_pages *pcp; |
| 1277 | |
| 1278 | local_irq_save(flags); |
| 1279 | pset = per_cpu_ptr(zone->pageset, cpu); |
| 1280 | |
| 1281 | pcp = &pset->pcp; |
| 1282 | if (pcp->count) { |
| 1283 | free_pcppages_bulk(zone, pcp->count, pcp); |
| 1284 | pcp->count = 0; |
| 1285 | } |
| 1286 | local_irq_restore(flags); |
| 1287 | } |
| 1288 | } |
| 1289 | |
| 1290 | /* |
| 1291 | * Spill all of this CPU's per-cpu pages back into the buddy allocator. |
| 1292 | */ |
| 1293 | void drain_local_pages(void *arg) |
| 1294 | { |
| 1295 | drain_pages(smp_processor_id()); |
| 1296 | } |
| 1297 | |
| 1298 | /* |
| 1299 | * Spill all the per-cpu pages from all CPUs back into the buddy allocator. |
| 1300 | * |
| 1301 | * Note that this code is protected against sending an IPI to an offline |
| 1302 | * CPU but does not guarantee sending an IPI to newly hotplugged CPUs: |
| 1303 | * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but |
| 1304 | * nothing keeps CPUs from showing up after we populated the cpumask and |
| 1305 | * before the call to on_each_cpu_mask(). |
| 1306 | */ |
| 1307 | void drain_all_pages(void) |
| 1308 | { |
| 1309 | int cpu; |
| 1310 | struct per_cpu_pageset *pcp; |
| 1311 | struct zone *zone; |
| 1312 | |
| 1313 | /* |
| 1314 | * Allocate in the BSS so we wont require allocation in |
| 1315 | * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y |
| 1316 | */ |
| 1317 | static cpumask_t cpus_with_pcps; |
| 1318 | |
| 1319 | /* |
| 1320 | * We don't care about racing with CPU hotplug event |
| 1321 | * as offline notification will cause the notified |
| 1322 | * cpu to drain that CPU pcps and on_each_cpu_mask |
| 1323 | * disables preemption as part of its processing |
| 1324 | */ |
| 1325 | for_each_online_cpu(cpu) { |
| 1326 | bool has_pcps = false; |
| 1327 | for_each_populated_zone(zone) { |
| 1328 | pcp = per_cpu_ptr(zone->pageset, cpu); |
| 1329 | if (pcp->pcp.count) { |
| 1330 | has_pcps = true; |
| 1331 | break; |
| 1332 | } |
| 1333 | } |
| 1334 | if (has_pcps) |
| 1335 | cpumask_set_cpu(cpu, &cpus_with_pcps); |
| 1336 | else |
| 1337 | cpumask_clear_cpu(cpu, &cpus_with_pcps); |
| 1338 | } |
| 1339 | on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1); |
| 1340 | } |
| 1341 | |
| 1342 | #ifdef CONFIG_HIBERNATION |
| 1343 | |
| 1344 | void mark_free_pages(struct zone *zone) |
| 1345 | { |
| 1346 | unsigned long pfn, max_zone_pfn; |
| 1347 | unsigned long flags; |
| 1348 | int order, t; |
| 1349 | struct list_head *curr; |
| 1350 | |
| 1351 | if (zone_is_empty(zone)) |
| 1352 | return; |
| 1353 | |
| 1354 | spin_lock_irqsave(&zone->lock, flags); |
| 1355 | |
| 1356 | max_zone_pfn = zone_end_pfn(zone); |
| 1357 | for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) |
| 1358 | if (pfn_valid(pfn)) { |
| 1359 | struct page *page = pfn_to_page(pfn); |
| 1360 | |
| 1361 | if (!swsusp_page_is_forbidden(page)) |
| 1362 | swsusp_unset_page_free(page); |
| 1363 | } |
| 1364 | |
| 1365 | for_each_migratetype_order(order, t) { |
| 1366 | list_for_each(curr, &zone->free_area[order].free_list[t]) { |
| 1367 | unsigned long i; |
| 1368 | |
| 1369 | pfn = page_to_pfn(list_entry(curr, struct page, lru)); |
| 1370 | for (i = 0; i < (1UL << order); i++) |
| 1371 | swsusp_set_page_free(pfn_to_page(pfn + i)); |
| 1372 | } |
| 1373 | } |
| 1374 | spin_unlock_irqrestore(&zone->lock, flags); |
| 1375 | } |
| 1376 | #endif /* CONFIG_PM */ |
| 1377 | |
| 1378 | /* |
| 1379 | * Free a 0-order page |
| 1380 | * cold == 1 ? free a cold page : free a hot page |
| 1381 | */ |
| 1382 | void free_hot_cold_page(struct page *page, int cold) |
| 1383 | { |
| 1384 | struct zone *zone = page_zone(page); |
| 1385 | struct per_cpu_pages *pcp; |
| 1386 | unsigned long flags; |
| 1387 | unsigned long pfn = page_to_pfn(page); |
| 1388 | int migratetype; |
| 1389 | |
| 1390 | if (!free_pages_prepare(page, 0)) |
| 1391 | return; |
| 1392 | |
| 1393 | migratetype = get_pfnblock_migratetype(page, pfn); |
| 1394 | set_freepage_migratetype(page, migratetype); |
| 1395 | local_irq_save(flags); |
| 1396 | __count_vm_event(PGFREE); |
| 1397 | |
| 1398 | /* |
| 1399 | * We only track unmovable, reclaimable and movable on pcp lists. |
| 1400 | * Free ISOLATE pages back to the allocator because they are being |
| 1401 | * offlined but treat RESERVE as movable pages so we can get those |
| 1402 | * areas back if necessary. Otherwise, we may have to free |
| 1403 | * excessively into the page allocator |
| 1404 | */ |
| 1405 | if (migratetype >= MIGRATE_PCPTYPES) { |
| 1406 | if (unlikely(is_migrate_isolate(migratetype))) { |
| 1407 | free_one_page(zone, page, pfn, 0, migratetype); |
| 1408 | goto out; |
| 1409 | } |
| 1410 | migratetype = MIGRATE_MOVABLE; |
| 1411 | } |
| 1412 | |
| 1413 | pcp = &this_cpu_ptr(zone->pageset)->pcp; |
| 1414 | if (cold) |
| 1415 | list_add_tail(&page->lru, &pcp->lists[migratetype]); |
| 1416 | else |
| 1417 | list_add(&page->lru, &pcp->lists[migratetype]); |
| 1418 | pcp->count++; |
| 1419 | if (pcp->count >= pcp->high) { |
| 1420 | unsigned long batch = ACCESS_ONCE(pcp->batch); |
| 1421 | free_pcppages_bulk(zone, batch, pcp); |
| 1422 | pcp->count -= batch; |
| 1423 | } |
| 1424 | |
| 1425 | out: |
| 1426 | local_irq_restore(flags); |
| 1427 | } |
| 1428 | |
| 1429 | /* |
| 1430 | * Free a list of 0-order pages |
| 1431 | */ |
| 1432 | void free_hot_cold_page_list(struct list_head *list, int cold) |
| 1433 | { |
| 1434 | struct page *page, *next; |
| 1435 | |
| 1436 | list_for_each_entry_safe(page, next, list, lru) { |
| 1437 | trace_mm_page_free_batched(page, cold); |
| 1438 | free_hot_cold_page(page, cold); |
| 1439 | } |
| 1440 | } |
| 1441 | |
| 1442 | /* |
| 1443 | * split_page takes a non-compound higher-order page, and splits it into |
| 1444 | * n (1<<order) sub-pages: page[0..n] |
| 1445 | * Each sub-page must be freed individually. |
| 1446 | * |
| 1447 | * Note: this is probably too low level an operation for use in drivers. |
| 1448 | * Please consult with lkml before using this in your driver. |
| 1449 | */ |
| 1450 | void split_page(struct page *page, unsigned int order) |
| 1451 | { |
| 1452 | int i; |
| 1453 | |
| 1454 | VM_BUG_ON_PAGE(PageCompound(page), page); |
| 1455 | VM_BUG_ON_PAGE(!page_count(page), page); |
| 1456 | |
| 1457 | #ifdef CONFIG_KMEMCHECK |
| 1458 | /* |
| 1459 | * Split shadow pages too, because free(page[0]) would |
| 1460 | * otherwise free the whole shadow. |
| 1461 | */ |
| 1462 | if (kmemcheck_page_is_tracked(page)) |
| 1463 | split_page(virt_to_page(page[0].shadow), order); |
| 1464 | #endif |
| 1465 | |
| 1466 | for (i = 1; i < (1 << order); i++) |
| 1467 | set_page_refcounted(page + i); |
| 1468 | } |
| 1469 | EXPORT_SYMBOL_GPL(split_page); |
| 1470 | |
| 1471 | static int __isolate_free_page(struct page *page, unsigned int order) |
| 1472 | { |
| 1473 | unsigned long watermark; |
| 1474 | struct zone *zone; |
| 1475 | int mt; |
| 1476 | |
| 1477 | BUG_ON(!PageBuddy(page)); |
| 1478 | |
| 1479 | zone = page_zone(page); |
| 1480 | mt = get_pageblock_migratetype(page); |
| 1481 | |
| 1482 | if (!is_migrate_isolate(mt)) { |
| 1483 | /* Obey watermarks as if the page was being allocated */ |
| 1484 | watermark = low_wmark_pages(zone) + (1 << order); |
| 1485 | if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) |
| 1486 | return 0; |
| 1487 | |
| 1488 | __mod_zone_freepage_state(zone, -(1UL << order), mt); |
| 1489 | } |
| 1490 | |
| 1491 | /* Remove page from free list */ |
| 1492 | list_del(&page->lru); |
| 1493 | zone->free_area[order].nr_free--; |
| 1494 | rmv_page_order(page); |
| 1495 | |
| 1496 | /* Set the pageblock if the isolated page is at least a pageblock */ |
| 1497 | if (order >= pageblock_order - 1) { |
| 1498 | struct page *endpage = page + (1 << order) - 1; |
| 1499 | for (; page < endpage; page += pageblock_nr_pages) { |
| 1500 | int mt = get_pageblock_migratetype(page); |
| 1501 | if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)) |
| 1502 | set_pageblock_migratetype(page, |
| 1503 | MIGRATE_MOVABLE); |
| 1504 | } |
| 1505 | } |
| 1506 | |
| 1507 | return 1UL << order; |
| 1508 | } |
| 1509 | |
| 1510 | /* |
| 1511 | * Similar to split_page except the page is already free. As this is only |
| 1512 | * being used for migration, the migratetype of the block also changes. |
| 1513 | * As this is called with interrupts disabled, the caller is responsible |
| 1514 | * for calling arch_alloc_page() and kernel_map_page() after interrupts |
| 1515 | * are enabled. |
| 1516 | * |
| 1517 | * Note: this is probably too low level an operation for use in drivers. |
| 1518 | * Please consult with lkml before using this in your driver. |
| 1519 | */ |
| 1520 | int split_free_page(struct page *page) |
| 1521 | { |
| 1522 | unsigned int order; |
| 1523 | int nr_pages; |
| 1524 | |
| 1525 | order = page_order(page); |
| 1526 | |
| 1527 | nr_pages = __isolate_free_page(page, order); |
| 1528 | if (!nr_pages) |
| 1529 | return 0; |
| 1530 | |
| 1531 | /* Split into individual pages */ |
| 1532 | set_page_refcounted(page); |
| 1533 | split_page(page, order); |
| 1534 | return nr_pages; |
| 1535 | } |
| 1536 | |
| 1537 | /* |
| 1538 | * Really, prep_compound_page() should be called from __rmqueue_bulk(). But |
| 1539 | * we cheat by calling it from here, in the order > 0 path. Saves a branch |
| 1540 | * or two. |
| 1541 | */ |
| 1542 | static inline |
| 1543 | struct page *buffered_rmqueue(struct zone *preferred_zone, |
| 1544 | struct zone *zone, int order, gfp_t gfp_flags, |
| 1545 | int migratetype) |
| 1546 | { |
| 1547 | unsigned long flags; |
| 1548 | struct page *page; |
| 1549 | int cold = !!(gfp_flags & __GFP_COLD); |
| 1550 | |
| 1551 | again: |
| 1552 | if (likely(order == 0)) { |
| 1553 | struct per_cpu_pages *pcp; |
| 1554 | struct list_head *list; |
| 1555 | |
| 1556 | local_irq_save(flags); |
| 1557 | pcp = &this_cpu_ptr(zone->pageset)->pcp; |
| 1558 | list = &pcp->lists[migratetype]; |
| 1559 | if (list_empty(list)) { |
| 1560 | pcp->count += rmqueue_bulk(zone, 0, |
| 1561 | pcp->batch, list, |
| 1562 | migratetype, cold); |
| 1563 | if (unlikely(list_empty(list))) |
| 1564 | goto failed; |
| 1565 | } |
| 1566 | |
| 1567 | if (cold) |
| 1568 | page = list_entry(list->prev, struct page, lru); |
| 1569 | else |
| 1570 | page = list_entry(list->next, struct page, lru); |
| 1571 | |
| 1572 | list_del(&page->lru); |
| 1573 | pcp->count--; |
| 1574 | } else { |
| 1575 | if (unlikely(gfp_flags & __GFP_NOFAIL)) { |
| 1576 | /* |
| 1577 | * __GFP_NOFAIL is not to be used in new code. |
| 1578 | * |
| 1579 | * All __GFP_NOFAIL callers should be fixed so that they |
| 1580 | * properly detect and handle allocation failures. |
| 1581 | * |
| 1582 | * We most definitely don't want callers attempting to |
| 1583 | * allocate greater than order-1 page units with |
| 1584 | * __GFP_NOFAIL. |
| 1585 | */ |
| 1586 | WARN_ON_ONCE(order > 1); |
| 1587 | } |
| 1588 | spin_lock_irqsave(&zone->lock, flags); |
| 1589 | page = __rmqueue(zone, order, migratetype); |
| 1590 | spin_unlock(&zone->lock); |
| 1591 | if (!page) |
| 1592 | goto failed; |
| 1593 | __mod_zone_freepage_state(zone, -(1 << order), |
| 1594 | get_freepage_migratetype(page)); |
| 1595 | } |
| 1596 | |
| 1597 | __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order)); |
| 1598 | |
| 1599 | __count_zone_vm_events(PGALLOC, zone, 1 << order); |
| 1600 | zone_statistics(preferred_zone, zone, gfp_flags); |
| 1601 | local_irq_restore(flags); |
| 1602 | |
| 1603 | VM_BUG_ON_PAGE(bad_range(zone, page), page); |
| 1604 | if (prep_new_page(page, order, gfp_flags)) |
| 1605 | goto again; |
| 1606 | return page; |
| 1607 | |
| 1608 | failed: |
| 1609 | local_irq_restore(flags); |
| 1610 | return NULL; |
| 1611 | } |
| 1612 | |
| 1613 | #ifdef CONFIG_FAIL_PAGE_ALLOC |
| 1614 | |
| 1615 | static struct { |
| 1616 | struct fault_attr attr; |
| 1617 | |
| 1618 | u32 ignore_gfp_highmem; |
| 1619 | u32 ignore_gfp_wait; |
| 1620 | u32 min_order; |
| 1621 | } fail_page_alloc = { |
| 1622 | .attr = FAULT_ATTR_INITIALIZER, |
| 1623 | .ignore_gfp_wait = 1, |
| 1624 | .ignore_gfp_highmem = 1, |
| 1625 | .min_order = 1, |
| 1626 | }; |
| 1627 | |
| 1628 | static int __init setup_fail_page_alloc(char *str) |
| 1629 | { |
| 1630 | return setup_fault_attr(&fail_page_alloc.attr, str); |
| 1631 | } |
| 1632 | __setup("fail_page_alloc=", setup_fail_page_alloc); |
| 1633 | |
| 1634 | static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) |
| 1635 | { |
| 1636 | if (order < fail_page_alloc.min_order) |
| 1637 | return false; |
| 1638 | if (gfp_mask & __GFP_NOFAIL) |
| 1639 | return false; |
| 1640 | if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) |
| 1641 | return false; |
| 1642 | if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) |
| 1643 | return false; |
| 1644 | |
| 1645 | return should_fail(&fail_page_alloc.attr, 1 << order); |
| 1646 | } |
| 1647 | |
| 1648 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS |
| 1649 | |
| 1650 | static int __init fail_page_alloc_debugfs(void) |
| 1651 | { |
| 1652 | umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; |
| 1653 | struct dentry *dir; |
| 1654 | |
| 1655 | dir = fault_create_debugfs_attr("fail_page_alloc", NULL, |
| 1656 | &fail_page_alloc.attr); |
| 1657 | if (IS_ERR(dir)) |
| 1658 | return PTR_ERR(dir); |
| 1659 | |
| 1660 | if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, |
| 1661 | &fail_page_alloc.ignore_gfp_wait)) |
| 1662 | goto fail; |
| 1663 | if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, |
| 1664 | &fail_page_alloc.ignore_gfp_highmem)) |
| 1665 | goto fail; |
| 1666 | if (!debugfs_create_u32("min-order", mode, dir, |
| 1667 | &fail_page_alloc.min_order)) |
| 1668 | goto fail; |
| 1669 | |
| 1670 | return 0; |
| 1671 | fail: |
| 1672 | debugfs_remove_recursive(dir); |
| 1673 | |
| 1674 | return -ENOMEM; |
| 1675 | } |
| 1676 | |
| 1677 | late_initcall(fail_page_alloc_debugfs); |
| 1678 | |
| 1679 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ |
| 1680 | |
| 1681 | #else /* CONFIG_FAIL_PAGE_ALLOC */ |
| 1682 | |
| 1683 | static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) |
| 1684 | { |
| 1685 | return false; |
| 1686 | } |
| 1687 | |
| 1688 | #endif /* CONFIG_FAIL_PAGE_ALLOC */ |
| 1689 | |
| 1690 | /* |
| 1691 | * Return true if free pages are above 'mark'. This takes into account the order |
| 1692 | * of the allocation. |
| 1693 | */ |
| 1694 | static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark, |
| 1695 | int classzone_idx, int alloc_flags, long free_pages) |
| 1696 | { |
| 1697 | /* free_pages my go negative - that's OK */ |
| 1698 | long min = mark; |
| 1699 | long lowmem_reserve = z->lowmem_reserve[classzone_idx]; |
| 1700 | int o; |
| 1701 | long free_cma = 0; |
| 1702 | |
| 1703 | free_pages -= (1 << order) - 1; |
| 1704 | if (alloc_flags & ALLOC_HIGH) |
| 1705 | min -= min / 2; |
| 1706 | if (alloc_flags & ALLOC_HARDER) |
| 1707 | min -= min / 4; |
| 1708 | #ifdef CONFIG_CMA |
| 1709 | /* If allocation can't use CMA areas don't use free CMA pages */ |
| 1710 | if (!(alloc_flags & ALLOC_CMA)) |
| 1711 | free_cma = zone_page_state(z, NR_FREE_CMA_PAGES); |
| 1712 | #endif |
| 1713 | |
| 1714 | if (free_pages - free_cma <= min + lowmem_reserve) |
| 1715 | return false; |
| 1716 | for (o = 0; o < order; o++) { |
| 1717 | /* At the next order, this order's pages become unavailable */ |
| 1718 | free_pages -= z->free_area[o].nr_free << o; |
| 1719 | |
| 1720 | /* Require fewer higher order pages to be free */ |
| 1721 | min >>= 1; |
| 1722 | |
| 1723 | if (free_pages <= min) |
| 1724 | return false; |
| 1725 | } |
| 1726 | return true; |
| 1727 | } |
| 1728 | |
| 1729 | bool zone_watermark_ok(struct zone *z, int order, unsigned long mark, |
| 1730 | int classzone_idx, int alloc_flags) |
| 1731 | { |
| 1732 | return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, |
| 1733 | zone_page_state(z, NR_FREE_PAGES)); |
| 1734 | } |
| 1735 | |
| 1736 | bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark, |
| 1737 | int classzone_idx, int alloc_flags) |
| 1738 | { |
| 1739 | long free_pages = zone_page_state(z, NR_FREE_PAGES); |
| 1740 | |
| 1741 | if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) |
| 1742 | free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); |
| 1743 | |
| 1744 | return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, |
| 1745 | free_pages); |
| 1746 | } |
| 1747 | |
| 1748 | #ifdef CONFIG_NUMA |
| 1749 | /* |
| 1750 | * zlc_setup - Setup for "zonelist cache". Uses cached zone data to |
| 1751 | * skip over zones that are not allowed by the cpuset, or that have |
| 1752 | * been recently (in last second) found to be nearly full. See further |
| 1753 | * comments in mmzone.h. Reduces cache footprint of zonelist scans |
| 1754 | * that have to skip over a lot of full or unallowed zones. |
| 1755 | * |
| 1756 | * If the zonelist cache is present in the passed zonelist, then |
| 1757 | * returns a pointer to the allowed node mask (either the current |
| 1758 | * tasks mems_allowed, or node_states[N_MEMORY].) |
| 1759 | * |
| 1760 | * If the zonelist cache is not available for this zonelist, does |
| 1761 | * nothing and returns NULL. |
| 1762 | * |
| 1763 | * If the fullzones BITMAP in the zonelist cache is stale (more than |
| 1764 | * a second since last zap'd) then we zap it out (clear its bits.) |
| 1765 | * |
| 1766 | * We hold off even calling zlc_setup, until after we've checked the |
| 1767 | * first zone in the zonelist, on the theory that most allocations will |
| 1768 | * be satisfied from that first zone, so best to examine that zone as |
| 1769 | * quickly as we can. |
| 1770 | */ |
| 1771 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) |
| 1772 | { |
| 1773 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ |
| 1774 | nodemask_t *allowednodes; /* zonelist_cache approximation */ |
| 1775 | |
| 1776 | zlc = zonelist->zlcache_ptr; |
| 1777 | if (!zlc) |
| 1778 | return NULL; |
| 1779 | |
| 1780 | if (time_after(jiffies, zlc->last_full_zap + HZ)) { |
| 1781 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); |
| 1782 | zlc->last_full_zap = jiffies; |
| 1783 | } |
| 1784 | |
| 1785 | allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? |
| 1786 | &cpuset_current_mems_allowed : |
| 1787 | &node_states[N_MEMORY]; |
| 1788 | return allowednodes; |
| 1789 | } |
| 1790 | |
| 1791 | /* |
| 1792 | * Given 'z' scanning a zonelist, run a couple of quick checks to see |
| 1793 | * if it is worth looking at further for free memory: |
| 1794 | * 1) Check that the zone isn't thought to be full (doesn't have its |
| 1795 | * bit set in the zonelist_cache fullzones BITMAP). |
| 1796 | * 2) Check that the zones node (obtained from the zonelist_cache |
| 1797 | * z_to_n[] mapping) is allowed in the passed in allowednodes mask. |
| 1798 | * Return true (non-zero) if zone is worth looking at further, or |
| 1799 | * else return false (zero) if it is not. |
| 1800 | * |
| 1801 | * This check -ignores- the distinction between various watermarks, |
| 1802 | * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is |
| 1803 | * found to be full for any variation of these watermarks, it will |
| 1804 | * be considered full for up to one second by all requests, unless |
| 1805 | * we are so low on memory on all allowed nodes that we are forced |
| 1806 | * into the second scan of the zonelist. |
| 1807 | * |
| 1808 | * In the second scan we ignore this zonelist cache and exactly |
| 1809 | * apply the watermarks to all zones, even it is slower to do so. |
| 1810 | * We are low on memory in the second scan, and should leave no stone |
| 1811 | * unturned looking for a free page. |
| 1812 | */ |
| 1813 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, |
| 1814 | nodemask_t *allowednodes) |
| 1815 | { |
| 1816 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ |
| 1817 | int i; /* index of *z in zonelist zones */ |
| 1818 | int n; /* node that zone *z is on */ |
| 1819 | |
| 1820 | zlc = zonelist->zlcache_ptr; |
| 1821 | if (!zlc) |
| 1822 | return 1; |
| 1823 | |
| 1824 | i = z - zonelist->_zonerefs; |
| 1825 | n = zlc->z_to_n[i]; |
| 1826 | |
| 1827 | /* This zone is worth trying if it is allowed but not full */ |
| 1828 | return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); |
| 1829 | } |
| 1830 | |
| 1831 | /* |
| 1832 | * Given 'z' scanning a zonelist, set the corresponding bit in |
| 1833 | * zlc->fullzones, so that subsequent attempts to allocate a page |
| 1834 | * from that zone don't waste time re-examining it. |
| 1835 | */ |
| 1836 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) |
| 1837 | { |
| 1838 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ |
| 1839 | int i; /* index of *z in zonelist zones */ |
| 1840 | |
| 1841 | zlc = zonelist->zlcache_ptr; |
| 1842 | if (!zlc) |
| 1843 | return; |
| 1844 | |
| 1845 | i = z - zonelist->_zonerefs; |
| 1846 | |
| 1847 | set_bit(i, zlc->fullzones); |
| 1848 | } |
| 1849 | |
| 1850 | /* |
| 1851 | * clear all zones full, called after direct reclaim makes progress so that |
| 1852 | * a zone that was recently full is not skipped over for up to a second |
| 1853 | */ |
| 1854 | static void zlc_clear_zones_full(struct zonelist *zonelist) |
| 1855 | { |
| 1856 | struct zonelist_cache *zlc; /* cached zonelist speedup info */ |
| 1857 | |
| 1858 | zlc = zonelist->zlcache_ptr; |
| 1859 | if (!zlc) |
| 1860 | return; |
| 1861 | |
| 1862 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); |
| 1863 | } |
| 1864 | |
| 1865 | static bool zone_local(struct zone *local_zone, struct zone *zone) |
| 1866 | { |
| 1867 | return local_zone->node == zone->node; |
| 1868 | } |
| 1869 | |
| 1870 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) |
| 1871 | { |
| 1872 | return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) < |
| 1873 | RECLAIM_DISTANCE; |
| 1874 | } |
| 1875 | |
| 1876 | #else /* CONFIG_NUMA */ |
| 1877 | |
| 1878 | static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) |
| 1879 | { |
| 1880 | return NULL; |
| 1881 | } |
| 1882 | |
| 1883 | static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, |
| 1884 | nodemask_t *allowednodes) |
| 1885 | { |
| 1886 | return 1; |
| 1887 | } |
| 1888 | |
| 1889 | static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) |
| 1890 | { |
| 1891 | } |
| 1892 | |
| 1893 | static void zlc_clear_zones_full(struct zonelist *zonelist) |
| 1894 | { |
| 1895 | } |
| 1896 | |
| 1897 | static bool zone_local(struct zone *local_zone, struct zone *zone) |
| 1898 | { |
| 1899 | return true; |
| 1900 | } |
| 1901 | |
| 1902 | static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) |
| 1903 | { |
| 1904 | return true; |
| 1905 | } |
| 1906 | |
| 1907 | #endif /* CONFIG_NUMA */ |
| 1908 | |
| 1909 | /* |
| 1910 | * get_page_from_freelist goes through the zonelist trying to allocate |
| 1911 | * a page. |
| 1912 | */ |
| 1913 | static struct page * |
| 1914 | get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, |
| 1915 | struct zonelist *zonelist, int high_zoneidx, int alloc_flags, |
| 1916 | struct zone *preferred_zone, int migratetype) |
| 1917 | { |
| 1918 | struct zoneref *z; |
| 1919 | struct page *page = NULL; |
| 1920 | int classzone_idx; |
| 1921 | struct zone *zone; |
| 1922 | nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ |
| 1923 | int zlc_active = 0; /* set if using zonelist_cache */ |
| 1924 | int did_zlc_setup = 0; /* just call zlc_setup() one time */ |
| 1925 | bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) && |
| 1926 | (gfp_mask & __GFP_WRITE); |
| 1927 | |
| 1928 | classzone_idx = zone_idx(preferred_zone); |
| 1929 | zonelist_scan: |
| 1930 | /* |
| 1931 | * Scan zonelist, looking for a zone with enough free. |
| 1932 | * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c. |
| 1933 | */ |
| 1934 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
| 1935 | high_zoneidx, nodemask) { |
| 1936 | unsigned long mark; |
| 1937 | |
| 1938 | if (IS_ENABLED(CONFIG_NUMA) && zlc_active && |
| 1939 | !zlc_zone_worth_trying(zonelist, z, allowednodes)) |
| 1940 | continue; |
| 1941 | if (cpusets_enabled() && |
| 1942 | (alloc_flags & ALLOC_CPUSET) && |
| 1943 | !cpuset_zone_allowed_softwall(zone, gfp_mask)) |
| 1944 | continue; |
| 1945 | /* |
| 1946 | * Distribute pages in proportion to the individual |
| 1947 | * zone size to ensure fair page aging. The zone a |
| 1948 | * page was allocated in should have no effect on the |
| 1949 | * time the page has in memory before being reclaimed. |
| 1950 | */ |
| 1951 | if (alloc_flags & ALLOC_FAIR) { |
| 1952 | if (!zone_local(preferred_zone, zone)) |
| 1953 | continue; |
| 1954 | if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0) |
| 1955 | continue; |
| 1956 | } |
| 1957 | /* |
| 1958 | * When allocating a page cache page for writing, we |
| 1959 | * want to get it from a zone that is within its dirty |
| 1960 | * limit, such that no single zone holds more than its |
| 1961 | * proportional share of globally allowed dirty pages. |
| 1962 | * The dirty limits take into account the zone's |
| 1963 | * lowmem reserves and high watermark so that kswapd |
| 1964 | * should be able to balance it without having to |
| 1965 | * write pages from its LRU list. |
| 1966 | * |
| 1967 | * This may look like it could increase pressure on |
| 1968 | * lower zones by failing allocations in higher zones |
| 1969 | * before they are full. But the pages that do spill |
| 1970 | * over are limited as the lower zones are protected |
| 1971 | * by this very same mechanism. It should not become |
| 1972 | * a practical burden to them. |
| 1973 | * |
| 1974 | * XXX: For now, allow allocations to potentially |
| 1975 | * exceed the per-zone dirty limit in the slowpath |
| 1976 | * (ALLOC_WMARK_LOW unset) before going into reclaim, |
| 1977 | * which is important when on a NUMA setup the allowed |
| 1978 | * zones are together not big enough to reach the |
| 1979 | * global limit. The proper fix for these situations |
| 1980 | * will require awareness of zones in the |
| 1981 | * dirty-throttling and the flusher threads. |
| 1982 | */ |
| 1983 | if (consider_zone_dirty && !zone_dirty_ok(zone)) |
| 1984 | continue; |
| 1985 | |
| 1986 | mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; |
| 1987 | if (!zone_watermark_ok(zone, order, mark, |
| 1988 | classzone_idx, alloc_flags)) { |
| 1989 | int ret; |
| 1990 | |
| 1991 | /* Checked here to keep the fast path fast */ |
| 1992 | BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); |
| 1993 | if (alloc_flags & ALLOC_NO_WATERMARKS) |
| 1994 | goto try_this_zone; |
| 1995 | |
| 1996 | if (IS_ENABLED(CONFIG_NUMA) && |
| 1997 | !did_zlc_setup && nr_online_nodes > 1) { |
| 1998 | /* |
| 1999 | * we do zlc_setup if there are multiple nodes |
| 2000 | * and before considering the first zone allowed |
| 2001 | * by the cpuset. |
| 2002 | */ |
| 2003 | allowednodes = zlc_setup(zonelist, alloc_flags); |
| 2004 | zlc_active = 1; |
| 2005 | did_zlc_setup = 1; |
| 2006 | } |
| 2007 | |
| 2008 | if (zone_reclaim_mode == 0 || |
| 2009 | !zone_allows_reclaim(preferred_zone, zone)) |
| 2010 | goto this_zone_full; |
| 2011 | |
| 2012 | /* |
| 2013 | * As we may have just activated ZLC, check if the first |
| 2014 | * eligible zone has failed zone_reclaim recently. |
| 2015 | */ |
| 2016 | if (IS_ENABLED(CONFIG_NUMA) && zlc_active && |
| 2017 | !zlc_zone_worth_trying(zonelist, z, allowednodes)) |
| 2018 | continue; |
| 2019 | |
| 2020 | ret = zone_reclaim(zone, gfp_mask, order); |
| 2021 | switch (ret) { |
| 2022 | case ZONE_RECLAIM_NOSCAN: |
| 2023 | /* did not scan */ |
| 2024 | continue; |
| 2025 | case ZONE_RECLAIM_FULL: |
| 2026 | /* scanned but unreclaimable */ |
| 2027 | continue; |
| 2028 | default: |
| 2029 | /* did we reclaim enough */ |
| 2030 | if (zone_watermark_ok(zone, order, mark, |
| 2031 | classzone_idx, alloc_flags)) |
| 2032 | goto try_this_zone; |
| 2033 | |
| 2034 | /* |
| 2035 | * Failed to reclaim enough to meet watermark. |
| 2036 | * Only mark the zone full if checking the min |
| 2037 | * watermark or if we failed to reclaim just |
| 2038 | * 1<<order pages or else the page allocator |
| 2039 | * fastpath will prematurely mark zones full |
| 2040 | * when the watermark is between the low and |
| 2041 | * min watermarks. |
| 2042 | */ |
| 2043 | if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) || |
| 2044 | ret == ZONE_RECLAIM_SOME) |
| 2045 | goto this_zone_full; |
| 2046 | |
| 2047 | continue; |
| 2048 | } |
| 2049 | } |
| 2050 | |
| 2051 | try_this_zone: |
| 2052 | page = buffered_rmqueue(preferred_zone, zone, order, |
| 2053 | gfp_mask, migratetype); |
| 2054 | if (page) |
| 2055 | break; |
| 2056 | this_zone_full: |
| 2057 | if (IS_ENABLED(CONFIG_NUMA) && zlc_active) |
| 2058 | zlc_mark_zone_full(zonelist, z); |
| 2059 | } |
| 2060 | |
| 2061 | if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) { |
| 2062 | /* Disable zlc cache for second zonelist scan */ |
| 2063 | zlc_active = 0; |
| 2064 | goto zonelist_scan; |
| 2065 | } |
| 2066 | |
| 2067 | if (page) |
| 2068 | /* |
| 2069 | * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was |
| 2070 | * necessary to allocate the page. The expectation is |
| 2071 | * that the caller is taking steps that will free more |
| 2072 | * memory. The caller should avoid the page being used |
| 2073 | * for !PFMEMALLOC purposes. |
| 2074 | */ |
| 2075 | page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS); |
| 2076 | |
| 2077 | return page; |
| 2078 | } |
| 2079 | |
| 2080 | /* |
| 2081 | * Large machines with many possible nodes should not always dump per-node |
| 2082 | * meminfo in irq context. |
| 2083 | */ |
| 2084 | static inline bool should_suppress_show_mem(void) |
| 2085 | { |
| 2086 | bool ret = false; |
| 2087 | |
| 2088 | #if NODES_SHIFT > 8 |
| 2089 | ret = in_interrupt(); |
| 2090 | #endif |
| 2091 | return ret; |
| 2092 | } |
| 2093 | |
| 2094 | static DEFINE_RATELIMIT_STATE(nopage_rs, |
| 2095 | DEFAULT_RATELIMIT_INTERVAL, |
| 2096 | DEFAULT_RATELIMIT_BURST); |
| 2097 | |
| 2098 | void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) |
| 2099 | { |
| 2100 | unsigned int filter = SHOW_MEM_FILTER_NODES; |
| 2101 | |
| 2102 | if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) || |
| 2103 | debug_guardpage_minorder() > 0) |
| 2104 | return; |
| 2105 | |
| 2106 | /* |
| 2107 | * This documents exceptions given to allocations in certain |
| 2108 | * contexts that are allowed to allocate outside current's set |
| 2109 | * of allowed nodes. |
| 2110 | */ |
| 2111 | if (!(gfp_mask & __GFP_NOMEMALLOC)) |
| 2112 | if (test_thread_flag(TIF_MEMDIE) || |
| 2113 | (current->flags & (PF_MEMALLOC | PF_EXITING))) |
| 2114 | filter &= ~SHOW_MEM_FILTER_NODES; |
| 2115 | if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) |
| 2116 | filter &= ~SHOW_MEM_FILTER_NODES; |
| 2117 | |
| 2118 | if (fmt) { |
| 2119 | struct va_format vaf; |
| 2120 | va_list args; |
| 2121 | |
| 2122 | va_start(args, fmt); |
| 2123 | |
| 2124 | vaf.fmt = fmt; |
| 2125 | vaf.va = &args; |
| 2126 | |
| 2127 | pr_warn("%pV", &vaf); |
| 2128 | |
| 2129 | va_end(args); |
| 2130 | } |
| 2131 | |
| 2132 | pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n", |
| 2133 | current->comm, order, gfp_mask); |
| 2134 | |
| 2135 | dump_stack(); |
| 2136 | if (!should_suppress_show_mem()) |
| 2137 | show_mem(filter); |
| 2138 | } |
| 2139 | |
| 2140 | static inline int |
| 2141 | should_alloc_retry(gfp_t gfp_mask, unsigned int order, |
| 2142 | unsigned long did_some_progress, |
| 2143 | unsigned long pages_reclaimed) |
| 2144 | { |
| 2145 | /* Do not loop if specifically requested */ |
| 2146 | if (gfp_mask & __GFP_NORETRY) |
| 2147 | return 0; |
| 2148 | |
| 2149 | /* Always retry if specifically requested */ |
| 2150 | if (gfp_mask & __GFP_NOFAIL) |
| 2151 | return 1; |
| 2152 | |
| 2153 | /* |
| 2154 | * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim |
| 2155 | * making forward progress without invoking OOM. Suspend also disables |
| 2156 | * storage devices so kswapd will not help. Bail if we are suspending. |
| 2157 | */ |
| 2158 | if (!did_some_progress && pm_suspended_storage()) |
| 2159 | return 0; |
| 2160 | |
| 2161 | /* |
| 2162 | * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER |
| 2163 | * means __GFP_NOFAIL, but that may not be true in other |
| 2164 | * implementations. |
| 2165 | */ |
| 2166 | if (order <= PAGE_ALLOC_COSTLY_ORDER) |
| 2167 | return 1; |
| 2168 | |
| 2169 | /* |
| 2170 | * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is |
| 2171 | * specified, then we retry until we no longer reclaim any pages |
| 2172 | * (above), or we've reclaimed an order of pages at least as |
| 2173 | * large as the allocation's order. In both cases, if the |
| 2174 | * allocation still fails, we stop retrying. |
| 2175 | */ |
| 2176 | if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) |
| 2177 | return 1; |
| 2178 | |
| 2179 | return 0; |
| 2180 | } |
| 2181 | |
| 2182 | static inline struct page * |
| 2183 | __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, |
| 2184 | struct zonelist *zonelist, enum zone_type high_zoneidx, |
| 2185 | nodemask_t *nodemask, struct zone *preferred_zone, |
| 2186 | int migratetype) |
| 2187 | { |
| 2188 | struct page *page; |
| 2189 | |
| 2190 | /* Acquire the OOM killer lock for the zones in zonelist */ |
| 2191 | if (!try_set_zonelist_oom(zonelist, gfp_mask)) { |
| 2192 | schedule_timeout_uninterruptible(1); |
| 2193 | return NULL; |
| 2194 | } |
| 2195 | |
| 2196 | /* |
| 2197 | * Go through the zonelist yet one more time, keep very high watermark |
| 2198 | * here, this is only to catch a parallel oom killing, we must fail if |
| 2199 | * we're still under heavy pressure. |
| 2200 | */ |
| 2201 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, |
| 2202 | order, zonelist, high_zoneidx, |
| 2203 | ALLOC_WMARK_HIGH|ALLOC_CPUSET, |
| 2204 | preferred_zone, migratetype); |
| 2205 | if (page) |
| 2206 | goto out; |
| 2207 | |
| 2208 | if (!(gfp_mask & __GFP_NOFAIL)) { |
| 2209 | /* The OOM killer will not help higher order allocs */ |
| 2210 | if (order > PAGE_ALLOC_COSTLY_ORDER) |
| 2211 | goto out; |
| 2212 | /* The OOM killer does not needlessly kill tasks for lowmem */ |
| 2213 | if (high_zoneidx < ZONE_NORMAL) |
| 2214 | goto out; |
| 2215 | /* |
| 2216 | * GFP_THISNODE contains __GFP_NORETRY and we never hit this. |
| 2217 | * Sanity check for bare calls of __GFP_THISNODE, not real OOM. |
| 2218 | * The caller should handle page allocation failure by itself if |
| 2219 | * it specifies __GFP_THISNODE. |
| 2220 | * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. |
| 2221 | */ |
| 2222 | if (gfp_mask & __GFP_THISNODE) |
| 2223 | goto out; |
| 2224 | } |
| 2225 | /* Exhausted what can be done so it's blamo time */ |
| 2226 | out_of_memory(zonelist, gfp_mask, order, nodemask, false); |
| 2227 | |
| 2228 | out: |
| 2229 | clear_zonelist_oom(zonelist, gfp_mask); |
| 2230 | return page; |
| 2231 | } |
| 2232 | |
| 2233 | #ifdef CONFIG_COMPACTION |
| 2234 | /* Try memory compaction for high-order allocations before reclaim */ |
| 2235 | static struct page * |
| 2236 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, |
| 2237 | struct zonelist *zonelist, enum zone_type high_zoneidx, |
| 2238 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, |
| 2239 | int migratetype, enum migrate_mode mode, |
| 2240 | bool *contended_compaction, bool *deferred_compaction, |
| 2241 | unsigned long *did_some_progress) |
| 2242 | { |
| 2243 | if (!order) |
| 2244 | return NULL; |
| 2245 | |
| 2246 | if (compaction_deferred(preferred_zone, order)) { |
| 2247 | *deferred_compaction = true; |
| 2248 | return NULL; |
| 2249 | } |
| 2250 | |
| 2251 | current->flags |= PF_MEMALLOC; |
| 2252 | *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask, |
| 2253 | nodemask, mode, |
| 2254 | contended_compaction); |
| 2255 | current->flags &= ~PF_MEMALLOC; |
| 2256 | |
| 2257 | if (*did_some_progress != COMPACT_SKIPPED) { |
| 2258 | struct page *page; |
| 2259 | |
| 2260 | /* Page migration frees to the PCP lists but we want merging */ |
| 2261 | drain_pages(get_cpu()); |
| 2262 | put_cpu(); |
| 2263 | |
| 2264 | page = get_page_from_freelist(gfp_mask, nodemask, |
| 2265 | order, zonelist, high_zoneidx, |
| 2266 | alloc_flags & ~ALLOC_NO_WATERMARKS, |
| 2267 | preferred_zone, migratetype); |
| 2268 | if (page) { |
| 2269 | preferred_zone->compact_blockskip_flush = false; |
| 2270 | compaction_defer_reset(preferred_zone, order, true); |
| 2271 | count_vm_event(COMPACTSUCCESS); |
| 2272 | return page; |
| 2273 | } |
| 2274 | |
| 2275 | /* |
| 2276 | * It's bad if compaction run occurs and fails. |
| 2277 | * The most likely reason is that pages exist, |
| 2278 | * but not enough to satisfy watermarks. |
| 2279 | */ |
| 2280 | count_vm_event(COMPACTFAIL); |
| 2281 | |
| 2282 | /* |
| 2283 | * As async compaction considers a subset of pageblocks, only |
| 2284 | * defer if the failure was a sync compaction failure. |
| 2285 | */ |
| 2286 | if (mode != MIGRATE_ASYNC) |
| 2287 | defer_compaction(preferred_zone, order); |
| 2288 | |
| 2289 | cond_resched(); |
| 2290 | } |
| 2291 | |
| 2292 | return NULL; |
| 2293 | } |
| 2294 | #else |
| 2295 | static inline struct page * |
| 2296 | __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, |
| 2297 | struct zonelist *zonelist, enum zone_type high_zoneidx, |
| 2298 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, |
| 2299 | int migratetype, enum migrate_mode mode, bool *contended_compaction, |
| 2300 | bool *deferred_compaction, unsigned long *did_some_progress) |
| 2301 | { |
| 2302 | return NULL; |
| 2303 | } |
| 2304 | #endif /* CONFIG_COMPACTION */ |
| 2305 | |
| 2306 | /* Perform direct synchronous page reclaim */ |
| 2307 | static int |
| 2308 | __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist, |
| 2309 | nodemask_t *nodemask) |
| 2310 | { |
| 2311 | struct reclaim_state reclaim_state; |
| 2312 | int progress; |
| 2313 | |
| 2314 | cond_resched(); |
| 2315 | |
| 2316 | /* We now go into synchronous reclaim */ |
| 2317 | cpuset_memory_pressure_bump(); |
| 2318 | current->flags |= PF_MEMALLOC; |
| 2319 | lockdep_set_current_reclaim_state(gfp_mask); |
| 2320 | reclaim_state.reclaimed_slab = 0; |
| 2321 | current->reclaim_state = &reclaim_state; |
| 2322 | |
| 2323 | progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); |
| 2324 | |
| 2325 | current->reclaim_state = NULL; |
| 2326 | lockdep_clear_current_reclaim_state(); |
| 2327 | current->flags &= ~PF_MEMALLOC; |
| 2328 | |
| 2329 | cond_resched(); |
| 2330 | |
| 2331 | return progress; |
| 2332 | } |
| 2333 | |
| 2334 | /* The really slow allocator path where we enter direct reclaim */ |
| 2335 | static inline struct page * |
| 2336 | __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, |
| 2337 | struct zonelist *zonelist, enum zone_type high_zoneidx, |
| 2338 | nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, |
| 2339 | int migratetype, unsigned long *did_some_progress) |
| 2340 | { |
| 2341 | struct page *page = NULL; |
| 2342 | bool drained = false; |
| 2343 | |
| 2344 | *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist, |
| 2345 | nodemask); |
| 2346 | if (unlikely(!(*did_some_progress))) |
| 2347 | return NULL; |
| 2348 | |
| 2349 | /* After successful reclaim, reconsider all zones for allocation */ |
| 2350 | if (IS_ENABLED(CONFIG_NUMA)) |
| 2351 | zlc_clear_zones_full(zonelist); |
| 2352 | |
| 2353 | retry: |
| 2354 | page = get_page_from_freelist(gfp_mask, nodemask, order, |
| 2355 | zonelist, high_zoneidx, |
| 2356 | alloc_flags & ~ALLOC_NO_WATERMARKS, |
| 2357 | preferred_zone, migratetype); |
| 2358 | |
| 2359 | /* |
| 2360 | * If an allocation failed after direct reclaim, it could be because |
| 2361 | * pages are pinned on the per-cpu lists. Drain them and try again |
| 2362 | */ |
| 2363 | if (!page && !drained) { |
| 2364 | drain_all_pages(); |
| 2365 | drained = true; |
| 2366 | goto retry; |
| 2367 | } |
| 2368 | |
| 2369 | return page; |
| 2370 | } |
| 2371 | |
| 2372 | /* |
| 2373 | * This is called in the allocator slow-path if the allocation request is of |
| 2374 | * sufficient urgency to ignore watermarks and take other desperate measures |
| 2375 | */ |
| 2376 | static inline struct page * |
| 2377 | __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, |
| 2378 | struct zonelist *zonelist, enum zone_type high_zoneidx, |
| 2379 | nodemask_t *nodemask, struct zone *preferred_zone, |
| 2380 | int migratetype) |
| 2381 | { |
| 2382 | struct page *page; |
| 2383 | |
| 2384 | do { |
| 2385 | page = get_page_from_freelist(gfp_mask, nodemask, order, |
| 2386 | zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, |
| 2387 | preferred_zone, migratetype); |
| 2388 | |
| 2389 | if (!page && gfp_mask & __GFP_NOFAIL) |
| 2390 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); |
| 2391 | } while (!page && (gfp_mask & __GFP_NOFAIL)); |
| 2392 | |
| 2393 | return page; |
| 2394 | } |
| 2395 | |
| 2396 | static void reset_alloc_batches(struct zonelist *zonelist, |
| 2397 | enum zone_type high_zoneidx, |
| 2398 | struct zone *preferred_zone) |
| 2399 | { |
| 2400 | struct zoneref *z; |
| 2401 | struct zone *zone; |
| 2402 | |
| 2403 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
| 2404 | /* |
| 2405 | * Only reset the batches of zones that were actually |
| 2406 | * considered in the fairness pass, we don't want to |
| 2407 | * trash fairness information for zones that are not |
| 2408 | * actually part of this zonelist's round-robin cycle. |
| 2409 | */ |
| 2410 | if (!zone_local(preferred_zone, zone)) |
| 2411 | continue; |
| 2412 | mod_zone_page_state(zone, NR_ALLOC_BATCH, |
| 2413 | high_wmark_pages(zone) - low_wmark_pages(zone) - |
| 2414 | atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH])); |
| 2415 | } |
| 2416 | } |
| 2417 | |
| 2418 | static void wake_all_kswapds(unsigned int order, |
| 2419 | struct zonelist *zonelist, |
| 2420 | enum zone_type high_zoneidx, |
| 2421 | struct zone *preferred_zone) |
| 2422 | { |
| 2423 | struct zoneref *z; |
| 2424 | struct zone *zone; |
| 2425 | |
| 2426 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) |
| 2427 | wakeup_kswapd(zone, order, zone_idx(preferred_zone)); |
| 2428 | } |
| 2429 | |
| 2430 | static inline int |
| 2431 | gfp_to_alloc_flags(gfp_t gfp_mask) |
| 2432 | { |
| 2433 | int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; |
| 2434 | const gfp_t wait = gfp_mask & __GFP_WAIT; |
| 2435 | |
| 2436 | /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ |
| 2437 | BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); |
| 2438 | |
| 2439 | /* |
| 2440 | * The caller may dip into page reserves a bit more if the caller |
| 2441 | * cannot run direct reclaim, or if the caller has realtime scheduling |
| 2442 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will |
| 2443 | * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). |
| 2444 | */ |
| 2445 | alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); |
| 2446 | |
| 2447 | if (!wait) { |
| 2448 | /* |
| 2449 | * Not worth trying to allocate harder for |
| 2450 | * __GFP_NOMEMALLOC even if it can't schedule. |
| 2451 | */ |
| 2452 | if (!(gfp_mask & __GFP_NOMEMALLOC)) |
| 2453 | alloc_flags |= ALLOC_HARDER; |
| 2454 | /* |
| 2455 | * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. |
| 2456 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. |
| 2457 | */ |
| 2458 | alloc_flags &= ~ALLOC_CPUSET; |
| 2459 | } else if (unlikely(rt_task(current)) && !in_interrupt()) |
| 2460 | alloc_flags |= ALLOC_HARDER; |
| 2461 | |
| 2462 | if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { |
| 2463 | if (gfp_mask & __GFP_MEMALLOC) |
| 2464 | alloc_flags |= ALLOC_NO_WATERMARKS; |
| 2465 | else if (in_serving_softirq() && (current->flags & PF_MEMALLOC)) |
| 2466 | alloc_flags |= ALLOC_NO_WATERMARKS; |
| 2467 | else if (!in_interrupt() && |
| 2468 | ((current->flags & PF_MEMALLOC) || |
| 2469 | unlikely(test_thread_flag(TIF_MEMDIE)))) |
| 2470 | alloc_flags |= ALLOC_NO_WATERMARKS; |
| 2471 | } |
| 2472 | #ifdef CONFIG_CMA |
| 2473 | if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE) |
| 2474 | alloc_flags |= ALLOC_CMA; |
| 2475 | #endif |
| 2476 | return alloc_flags; |
| 2477 | } |
| 2478 | |
| 2479 | bool gfp_pfmemalloc_allowed(gfp_t gfp_mask) |
| 2480 | { |
| 2481 | return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS); |
| 2482 | } |
| 2483 | |
| 2484 | static inline struct page * |
| 2485 | __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, |
| 2486 | struct zonelist *zonelist, enum zone_type high_zoneidx, |
| 2487 | nodemask_t *nodemask, struct zone *preferred_zone, |
| 2488 | int migratetype) |
| 2489 | { |
| 2490 | const gfp_t wait = gfp_mask & __GFP_WAIT; |
| 2491 | struct page *page = NULL; |
| 2492 | int alloc_flags; |
| 2493 | unsigned long pages_reclaimed = 0; |
| 2494 | unsigned long did_some_progress; |
| 2495 | enum migrate_mode migration_mode = MIGRATE_ASYNC; |
| 2496 | bool deferred_compaction = false; |
| 2497 | bool contended_compaction = false; |
| 2498 | |
| 2499 | /* |
| 2500 | * In the slowpath, we sanity check order to avoid ever trying to |
| 2501 | * reclaim >= MAX_ORDER areas which will never succeed. Callers may |
| 2502 | * be using allocators in order of preference for an area that is |
| 2503 | * too large. |
| 2504 | */ |
| 2505 | if (order >= MAX_ORDER) { |
| 2506 | WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); |
| 2507 | return NULL; |
| 2508 | } |
| 2509 | |
| 2510 | /* |
| 2511 | * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and |
| 2512 | * __GFP_NOWARN set) should not cause reclaim since the subsystem |
| 2513 | * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim |
| 2514 | * using a larger set of nodes after it has established that the |
| 2515 | * allowed per node queues are empty and that nodes are |
| 2516 | * over allocated. |
| 2517 | */ |
| 2518 | if (IS_ENABLED(CONFIG_NUMA) && |
| 2519 | (gfp_mask & GFP_THISNODE) == GFP_THISNODE) |
| 2520 | goto nopage; |
| 2521 | |
| 2522 | restart: |
| 2523 | if (!(gfp_mask & __GFP_NO_KSWAPD)) |
| 2524 | wake_all_kswapds(order, zonelist, high_zoneidx, preferred_zone); |
| 2525 | |
| 2526 | /* |
| 2527 | * OK, we're below the kswapd watermark and have kicked background |
| 2528 | * reclaim. Now things get more complex, so set up alloc_flags according |
| 2529 | * to how we want to proceed. |
| 2530 | */ |
| 2531 | alloc_flags = gfp_to_alloc_flags(gfp_mask); |
| 2532 | |
| 2533 | /* |
| 2534 | * Find the true preferred zone if the allocation is unconstrained by |
| 2535 | * cpusets. |
| 2536 | */ |
| 2537 | if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) |
| 2538 | first_zones_zonelist(zonelist, high_zoneidx, NULL, |
| 2539 | &preferred_zone); |
| 2540 | |
| 2541 | rebalance: |
| 2542 | /* This is the last chance, in general, before the goto nopage. */ |
| 2543 | page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, |
| 2544 | high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, |
| 2545 | preferred_zone, migratetype); |
| 2546 | if (page) |
| 2547 | goto got_pg; |
| 2548 | |
| 2549 | /* Allocate without watermarks if the context allows */ |
| 2550 | if (alloc_flags & ALLOC_NO_WATERMARKS) { |
| 2551 | /* |
| 2552 | * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds |
| 2553 | * the allocation is high priority and these type of |
| 2554 | * allocations are system rather than user orientated |
| 2555 | */ |
| 2556 | zonelist = node_zonelist(numa_node_id(), gfp_mask); |
| 2557 | |
| 2558 | page = __alloc_pages_high_priority(gfp_mask, order, |
| 2559 | zonelist, high_zoneidx, nodemask, |
| 2560 | preferred_zone, migratetype); |
| 2561 | if (page) { |
| 2562 | goto got_pg; |
| 2563 | } |
| 2564 | } |
| 2565 | |
| 2566 | /* Atomic allocations - we can't balance anything */ |
| 2567 | if (!wait) { |
| 2568 | /* |
| 2569 | * All existing users of the deprecated __GFP_NOFAIL are |
| 2570 | * blockable, so warn of any new users that actually allow this |
| 2571 | * type of allocation to fail. |
| 2572 | */ |
| 2573 | WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL); |
| 2574 | goto nopage; |
| 2575 | } |
| 2576 | |
| 2577 | /* Avoid recursion of direct reclaim */ |
| 2578 | if (current->flags & PF_MEMALLOC) |
| 2579 | goto nopage; |
| 2580 | |
| 2581 | /* Avoid allocations with no watermarks from looping endlessly */ |
| 2582 | if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) |
| 2583 | goto nopage; |
| 2584 | |
| 2585 | /* |
| 2586 | * Try direct compaction. The first pass is asynchronous. Subsequent |
| 2587 | * attempts after direct reclaim are synchronous |
| 2588 | */ |
| 2589 | page = __alloc_pages_direct_compact(gfp_mask, order, zonelist, |
| 2590 | high_zoneidx, nodemask, alloc_flags, |
| 2591 | preferred_zone, migratetype, |
| 2592 | migration_mode, &contended_compaction, |
| 2593 | &deferred_compaction, |
| 2594 | &did_some_progress); |
| 2595 | if (page) |
| 2596 | goto got_pg; |
| 2597 | |
| 2598 | /* |
| 2599 | * It can become very expensive to allocate transparent hugepages at |
| 2600 | * fault, so use asynchronous memory compaction for THP unless it is |
| 2601 | * khugepaged trying to collapse. |
| 2602 | */ |
| 2603 | if (!(gfp_mask & __GFP_NO_KSWAPD) || (current->flags & PF_KTHREAD)) |
| 2604 | migration_mode = MIGRATE_SYNC_LIGHT; |
| 2605 | |
| 2606 | /* |
| 2607 | * If compaction is deferred for high-order allocations, it is because |
| 2608 | * sync compaction recently failed. In this is the case and the caller |
| 2609 | * requested a movable allocation that does not heavily disrupt the |
| 2610 | * system then fail the allocation instead of entering direct reclaim. |
| 2611 | */ |
| 2612 | if ((deferred_compaction || contended_compaction) && |
| 2613 | (gfp_mask & __GFP_NO_KSWAPD)) |
| 2614 | goto nopage; |
| 2615 | |
| 2616 | /* Try direct reclaim and then allocating */ |
| 2617 | page = __alloc_pages_direct_reclaim(gfp_mask, order, |
| 2618 | zonelist, high_zoneidx, |
| 2619 | nodemask, |
| 2620 | alloc_flags, preferred_zone, |
| 2621 | migratetype, &did_some_progress); |
| 2622 | if (page) |
| 2623 | goto got_pg; |
| 2624 | |
| 2625 | /* |
| 2626 | * If we failed to make any progress reclaiming, then we are |
| 2627 | * running out of options and have to consider going OOM |
| 2628 | */ |
| 2629 | if (!did_some_progress) { |
| 2630 | if (oom_gfp_allowed(gfp_mask)) { |
| 2631 | if (oom_killer_disabled) |
| 2632 | goto nopage; |
| 2633 | /* Coredumps can quickly deplete all memory reserves */ |
| 2634 | if ((current->flags & PF_DUMPCORE) && |
| 2635 | !(gfp_mask & __GFP_NOFAIL)) |
| 2636 | goto nopage; |
| 2637 | page = __alloc_pages_may_oom(gfp_mask, order, |
| 2638 | zonelist, high_zoneidx, |
| 2639 | nodemask, preferred_zone, |
| 2640 | migratetype); |
| 2641 | if (page) |
| 2642 | goto got_pg; |
| 2643 | |
| 2644 | if (!(gfp_mask & __GFP_NOFAIL)) { |
| 2645 | /* |
| 2646 | * The oom killer is not called for high-order |
| 2647 | * allocations that may fail, so if no progress |
| 2648 | * is being made, there are no other options and |
| 2649 | * retrying is unlikely to help. |
| 2650 | */ |
| 2651 | if (order > PAGE_ALLOC_COSTLY_ORDER) |
| 2652 | goto nopage; |
| 2653 | /* |
| 2654 | * The oom killer is not called for lowmem |
| 2655 | * allocations to prevent needlessly killing |
| 2656 | * innocent tasks. |
| 2657 | */ |
| 2658 | if (high_zoneidx < ZONE_NORMAL) |
| 2659 | goto nopage; |
| 2660 | } |
| 2661 | |
| 2662 | goto restart; |
| 2663 | } |
| 2664 | } |
| 2665 | |
| 2666 | /* Check if we should retry the allocation */ |
| 2667 | pages_reclaimed += did_some_progress; |
| 2668 | if (should_alloc_retry(gfp_mask, order, did_some_progress, |
| 2669 | pages_reclaimed)) { |
| 2670 | /* Wait for some write requests to complete then retry */ |
| 2671 | wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); |
| 2672 | goto rebalance; |
| 2673 | } else { |
| 2674 | /* |
| 2675 | * High-order allocations do not necessarily loop after |
| 2676 | * direct reclaim and reclaim/compaction depends on compaction |
| 2677 | * being called after reclaim so call directly if necessary |
| 2678 | */ |
| 2679 | page = __alloc_pages_direct_compact(gfp_mask, order, zonelist, |
| 2680 | high_zoneidx, nodemask, alloc_flags, |
| 2681 | preferred_zone, migratetype, |
| 2682 | migration_mode, &contended_compaction, |
| 2683 | &deferred_compaction, |
| 2684 | &did_some_progress); |
| 2685 | if (page) |
| 2686 | goto got_pg; |
| 2687 | } |
| 2688 | |
| 2689 | nopage: |
| 2690 | warn_alloc_failed(gfp_mask, order, NULL); |
| 2691 | return page; |
| 2692 | got_pg: |
| 2693 | if (kmemcheck_enabled) |
| 2694 | kmemcheck_pagealloc_alloc(page, order, gfp_mask); |
| 2695 | |
| 2696 | return page; |
| 2697 | } |
| 2698 | |
| 2699 | /* |
| 2700 | * This is the 'heart' of the zoned buddy allocator. |
| 2701 | */ |
| 2702 | struct page * |
| 2703 | __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, |
| 2704 | struct zonelist *zonelist, nodemask_t *nodemask) |
| 2705 | { |
| 2706 | enum zone_type high_zoneidx = gfp_zone(gfp_mask); |
| 2707 | struct zone *preferred_zone; |
| 2708 | struct page *page = NULL; |
| 2709 | int migratetype = allocflags_to_migratetype(gfp_mask); |
| 2710 | unsigned int cpuset_mems_cookie; |
| 2711 | int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR; |
| 2712 | |
| 2713 | gfp_mask &= gfp_allowed_mask; |
| 2714 | |
| 2715 | lockdep_trace_alloc(gfp_mask); |
| 2716 | |
| 2717 | might_sleep_if(gfp_mask & __GFP_WAIT); |
| 2718 | |
| 2719 | if (should_fail_alloc_page(gfp_mask, order)) |
| 2720 | return NULL; |
| 2721 | |
| 2722 | /* |
| 2723 | * Check the zones suitable for the gfp_mask contain at least one |
| 2724 | * valid zone. It's possible to have an empty zonelist as a result |
| 2725 | * of GFP_THISNODE and a memoryless node |
| 2726 | */ |
| 2727 | if (unlikely(!zonelist->_zonerefs->zone)) |
| 2728 | return NULL; |
| 2729 | |
| 2730 | retry_cpuset: |
| 2731 | cpuset_mems_cookie = read_mems_allowed_begin(); |
| 2732 | |
| 2733 | /* The preferred zone is used for statistics later */ |
| 2734 | first_zones_zonelist(zonelist, high_zoneidx, |
| 2735 | nodemask ? : &cpuset_current_mems_allowed, |
| 2736 | &preferred_zone); |
| 2737 | if (!preferred_zone) |
| 2738 | goto out; |
| 2739 | |
| 2740 | #ifdef CONFIG_CMA |
| 2741 | if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE) |
| 2742 | alloc_flags |= ALLOC_CMA; |
| 2743 | #endif |
| 2744 | retry: |
| 2745 | /* First allocation attempt */ |
| 2746 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, |
| 2747 | zonelist, high_zoneidx, alloc_flags, |
| 2748 | preferred_zone, migratetype); |
| 2749 | if (unlikely(!page)) { |
| 2750 | /* |
| 2751 | * The first pass makes sure allocations are spread |
| 2752 | * fairly within the local node. However, the local |
| 2753 | * node might have free pages left after the fairness |
| 2754 | * batches are exhausted, and remote zones haven't |
| 2755 | * even been considered yet. Try once more without |
| 2756 | * fairness, and include remote zones now, before |
| 2757 | * entering the slowpath and waking kswapd: prefer |
| 2758 | * spilling to a remote zone over swapping locally. |
| 2759 | */ |
| 2760 | if (alloc_flags & ALLOC_FAIR) { |
| 2761 | reset_alloc_batches(zonelist, high_zoneidx, |
| 2762 | preferred_zone); |
| 2763 | alloc_flags &= ~ALLOC_FAIR; |
| 2764 | goto retry; |
| 2765 | } |
| 2766 | /* |
| 2767 | * Runtime PM, block IO and its error handling path |
| 2768 | * can deadlock because I/O on the device might not |
| 2769 | * complete. |
| 2770 | */ |
| 2771 | gfp_mask = memalloc_noio_flags(gfp_mask); |
| 2772 | page = __alloc_pages_slowpath(gfp_mask, order, |
| 2773 | zonelist, high_zoneidx, nodemask, |
| 2774 | preferred_zone, migratetype); |
| 2775 | } |
| 2776 | |
| 2777 | trace_mm_page_alloc(page, order, gfp_mask, migratetype); |
| 2778 | |
| 2779 | out: |
| 2780 | /* |
| 2781 | * When updating a task's mems_allowed, it is possible to race with |
| 2782 | * parallel threads in such a way that an allocation can fail while |
| 2783 | * the mask is being updated. If a page allocation is about to fail, |
| 2784 | * check if the cpuset changed during allocation and if so, retry. |
| 2785 | */ |
| 2786 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
| 2787 | goto retry_cpuset; |
| 2788 | |
| 2789 | return page; |
| 2790 | } |
| 2791 | EXPORT_SYMBOL(__alloc_pages_nodemask); |
| 2792 | |
| 2793 | /* |
| 2794 | * Common helper functions. |
| 2795 | */ |
| 2796 | unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) |
| 2797 | { |
| 2798 | struct page *page; |
| 2799 | |
| 2800 | /* |
| 2801 | * __get_free_pages() returns a 32-bit address, which cannot represent |
| 2802 | * a highmem page |
| 2803 | */ |
| 2804 | VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); |
| 2805 | |
| 2806 | page = alloc_pages(gfp_mask, order); |
| 2807 | if (!page) |
| 2808 | return 0; |
| 2809 | return (unsigned long) page_address(page); |
| 2810 | } |
| 2811 | EXPORT_SYMBOL(__get_free_pages); |
| 2812 | |
| 2813 | unsigned long get_zeroed_page(gfp_t gfp_mask) |
| 2814 | { |
| 2815 | return __get_free_pages(gfp_mask | __GFP_ZERO, 0); |
| 2816 | } |
| 2817 | EXPORT_SYMBOL(get_zeroed_page); |
| 2818 | |
| 2819 | void __free_pages(struct page *page, unsigned int order) |
| 2820 | { |
| 2821 | if (put_page_testzero(page)) { |
| 2822 | if (order == 0) |
| 2823 | free_hot_cold_page(page, 0); |
| 2824 | else |
| 2825 | __free_pages_ok(page, order); |
| 2826 | } |
| 2827 | } |
| 2828 | |
| 2829 | EXPORT_SYMBOL(__free_pages); |
| 2830 | |
| 2831 | void free_pages(unsigned long addr, unsigned int order) |
| 2832 | { |
| 2833 | if (addr != 0) { |
| 2834 | VM_BUG_ON(!virt_addr_valid((void *)addr)); |
| 2835 | __free_pages(virt_to_page((void *)addr), order); |
| 2836 | } |
| 2837 | } |
| 2838 | |
| 2839 | EXPORT_SYMBOL(free_pages); |
| 2840 | |
| 2841 | /* |
| 2842 | * alloc_kmem_pages charges newly allocated pages to the kmem resource counter |
| 2843 | * of the current memory cgroup. |
| 2844 | * |
| 2845 | * It should be used when the caller would like to use kmalloc, but since the |
| 2846 | * allocation is large, it has to fall back to the page allocator. |
| 2847 | */ |
| 2848 | struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order) |
| 2849 | { |
| 2850 | struct page *page; |
| 2851 | struct mem_cgroup *memcg = NULL; |
| 2852 | |
| 2853 | if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) |
| 2854 | return NULL; |
| 2855 | page = alloc_pages(gfp_mask, order); |
| 2856 | memcg_kmem_commit_charge(page, memcg, order); |
| 2857 | return page; |
| 2858 | } |
| 2859 | |
| 2860 | struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order) |
| 2861 | { |
| 2862 | struct page *page; |
| 2863 | struct mem_cgroup *memcg = NULL; |
| 2864 | |
| 2865 | if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) |
| 2866 | return NULL; |
| 2867 | page = alloc_pages_node(nid, gfp_mask, order); |
| 2868 | memcg_kmem_commit_charge(page, memcg, order); |
| 2869 | return page; |
| 2870 | } |
| 2871 | |
| 2872 | /* |
| 2873 | * __free_kmem_pages and free_kmem_pages will free pages allocated with |
| 2874 | * alloc_kmem_pages. |
| 2875 | */ |
| 2876 | void __free_kmem_pages(struct page *page, unsigned int order) |
| 2877 | { |
| 2878 | memcg_kmem_uncharge_pages(page, order); |
| 2879 | __free_pages(page, order); |
| 2880 | } |
| 2881 | |
| 2882 | void free_kmem_pages(unsigned long addr, unsigned int order) |
| 2883 | { |
| 2884 | if (addr != 0) { |
| 2885 | VM_BUG_ON(!virt_addr_valid((void *)addr)); |
| 2886 | __free_kmem_pages(virt_to_page((void *)addr), order); |
| 2887 | } |
| 2888 | } |
| 2889 | |
| 2890 | static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) |
| 2891 | { |
| 2892 | if (addr) { |
| 2893 | unsigned long alloc_end = addr + (PAGE_SIZE << order); |
| 2894 | unsigned long used = addr + PAGE_ALIGN(size); |
| 2895 | |
| 2896 | split_page(virt_to_page((void *)addr), order); |
| 2897 | while (used < alloc_end) { |
| 2898 | free_page(used); |
| 2899 | used += PAGE_SIZE; |
| 2900 | } |
| 2901 | } |
| 2902 | return (void *)addr; |
| 2903 | } |
| 2904 | |
| 2905 | /** |
| 2906 | * alloc_pages_exact - allocate an exact number physically-contiguous pages. |
| 2907 | * @size: the number of bytes to allocate |
| 2908 | * @gfp_mask: GFP flags for the allocation |
| 2909 | * |
| 2910 | * This function is similar to alloc_pages(), except that it allocates the |
| 2911 | * minimum number of pages to satisfy the request. alloc_pages() can only |
| 2912 | * allocate memory in power-of-two pages. |
| 2913 | * |
| 2914 | * This function is also limited by MAX_ORDER. |
| 2915 | * |
| 2916 | * Memory allocated by this function must be released by free_pages_exact(). |
| 2917 | */ |
| 2918 | void *alloc_pages_exact(size_t size, gfp_t gfp_mask) |
| 2919 | { |
| 2920 | unsigned int order = get_order(size); |
| 2921 | unsigned long addr; |
| 2922 | |
| 2923 | addr = __get_free_pages(gfp_mask, order); |
| 2924 | return make_alloc_exact(addr, order, size); |
| 2925 | } |
| 2926 | EXPORT_SYMBOL(alloc_pages_exact); |
| 2927 | |
| 2928 | /** |
| 2929 | * alloc_pages_exact_nid - allocate an exact number of physically-contiguous |
| 2930 | * pages on a node. |
| 2931 | * @nid: the preferred node ID where memory should be allocated |
| 2932 | * @size: the number of bytes to allocate |
| 2933 | * @gfp_mask: GFP flags for the allocation |
| 2934 | * |
| 2935 | * Like alloc_pages_exact(), but try to allocate on node nid first before falling |
| 2936 | * back. |
| 2937 | * Note this is not alloc_pages_exact_node() which allocates on a specific node, |
| 2938 | * but is not exact. |
| 2939 | */ |
| 2940 | void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) |
| 2941 | { |
| 2942 | unsigned order = get_order(size); |
| 2943 | struct page *p = alloc_pages_node(nid, gfp_mask, order); |
| 2944 | if (!p) |
| 2945 | return NULL; |
| 2946 | return make_alloc_exact((unsigned long)page_address(p), order, size); |
| 2947 | } |
| 2948 | EXPORT_SYMBOL(alloc_pages_exact_nid); |
| 2949 | |
| 2950 | /** |
| 2951 | * free_pages_exact - release memory allocated via alloc_pages_exact() |
| 2952 | * @virt: the value returned by alloc_pages_exact. |
| 2953 | * @size: size of allocation, same value as passed to alloc_pages_exact(). |
| 2954 | * |
| 2955 | * Release the memory allocated by a previous call to alloc_pages_exact. |
| 2956 | */ |
| 2957 | void free_pages_exact(void *virt, size_t size) |
| 2958 | { |
| 2959 | unsigned long addr = (unsigned long)virt; |
| 2960 | unsigned long end = addr + PAGE_ALIGN(size); |
| 2961 | |
| 2962 | while (addr < end) { |
| 2963 | free_page(addr); |
| 2964 | addr += PAGE_SIZE; |
| 2965 | } |
| 2966 | } |
| 2967 | EXPORT_SYMBOL(free_pages_exact); |
| 2968 | |
| 2969 | /** |
| 2970 | * nr_free_zone_pages - count number of pages beyond high watermark |
| 2971 | * @offset: The zone index of the highest zone |
| 2972 | * |
| 2973 | * nr_free_zone_pages() counts the number of counts pages which are beyond the |
| 2974 | * high watermark within all zones at or below a given zone index. For each |
| 2975 | * zone, the number of pages is calculated as: |
| 2976 | * managed_pages - high_pages |
| 2977 | */ |
| 2978 | static unsigned long nr_free_zone_pages(int offset) |
| 2979 | { |
| 2980 | struct zoneref *z; |
| 2981 | struct zone *zone; |
| 2982 | |
| 2983 | /* Just pick one node, since fallback list is circular */ |
| 2984 | unsigned long sum = 0; |
| 2985 | |
| 2986 | struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); |
| 2987 | |
| 2988 | for_each_zone_zonelist(zone, z, zonelist, offset) { |
| 2989 | unsigned long size = zone->managed_pages; |
| 2990 | unsigned long high = high_wmark_pages(zone); |
| 2991 | if (size > high) |
| 2992 | sum += size - high; |
| 2993 | } |
| 2994 | |
| 2995 | return sum; |
| 2996 | } |
| 2997 | |
| 2998 | /** |
| 2999 | * nr_free_buffer_pages - count number of pages beyond high watermark |
| 3000 | * |
| 3001 | * nr_free_buffer_pages() counts the number of pages which are beyond the high |
| 3002 | * watermark within ZONE_DMA and ZONE_NORMAL. |
| 3003 | */ |
| 3004 | unsigned long nr_free_buffer_pages(void) |
| 3005 | { |
| 3006 | return nr_free_zone_pages(gfp_zone(GFP_USER)); |
| 3007 | } |
| 3008 | EXPORT_SYMBOL_GPL(nr_free_buffer_pages); |
| 3009 | |
| 3010 | /** |
| 3011 | * nr_free_pagecache_pages - count number of pages beyond high watermark |
| 3012 | * |
| 3013 | * nr_free_pagecache_pages() counts the number of pages which are beyond the |
| 3014 | * high watermark within all zones. |
| 3015 | */ |
| 3016 | unsigned long nr_free_pagecache_pages(void) |
| 3017 | { |
| 3018 | return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); |
| 3019 | } |
| 3020 | |
| 3021 | static inline void show_node(struct zone *zone) |
| 3022 | { |
| 3023 | if (IS_ENABLED(CONFIG_NUMA)) |
| 3024 | printk("Node %d ", zone_to_nid(zone)); |
| 3025 | } |
| 3026 | |
| 3027 | void si_meminfo(struct sysinfo *val) |
| 3028 | { |
| 3029 | val->totalram = totalram_pages; |
| 3030 | val->sharedram = 0; |
| 3031 | val->freeram = global_page_state(NR_FREE_PAGES); |
| 3032 | val->bufferram = nr_blockdev_pages(); |
| 3033 | val->totalhigh = totalhigh_pages; |
| 3034 | val->freehigh = nr_free_highpages(); |
| 3035 | val->mem_unit = PAGE_SIZE; |
| 3036 | } |
| 3037 | |
| 3038 | EXPORT_SYMBOL(si_meminfo); |
| 3039 | |
| 3040 | #ifdef CONFIG_NUMA |
| 3041 | void si_meminfo_node(struct sysinfo *val, int nid) |
| 3042 | { |
| 3043 | int zone_type; /* needs to be signed */ |
| 3044 | unsigned long managed_pages = 0; |
| 3045 | pg_data_t *pgdat = NODE_DATA(nid); |
| 3046 | |
| 3047 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) |
| 3048 | managed_pages += pgdat->node_zones[zone_type].managed_pages; |
| 3049 | val->totalram = managed_pages; |
| 3050 | val->freeram = node_page_state(nid, NR_FREE_PAGES); |
| 3051 | #ifdef CONFIG_HIGHMEM |
| 3052 | val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages; |
| 3053 | val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], |
| 3054 | NR_FREE_PAGES); |
| 3055 | #else |
| 3056 | val->totalhigh = 0; |
| 3057 | val->freehigh = 0; |
| 3058 | #endif |
| 3059 | val->mem_unit = PAGE_SIZE; |
| 3060 | } |
| 3061 | #endif |
| 3062 | |
| 3063 | /* |
| 3064 | * Determine whether the node should be displayed or not, depending on whether |
| 3065 | * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). |
| 3066 | */ |
| 3067 | bool skip_free_areas_node(unsigned int flags, int nid) |
| 3068 | { |
| 3069 | bool ret = false; |
| 3070 | unsigned int cpuset_mems_cookie; |
| 3071 | |
| 3072 | if (!(flags & SHOW_MEM_FILTER_NODES)) |
| 3073 | goto out; |
| 3074 | |
| 3075 | do { |
| 3076 | cpuset_mems_cookie = read_mems_allowed_begin(); |
| 3077 | ret = !node_isset(nid, cpuset_current_mems_allowed); |
| 3078 | } while (read_mems_allowed_retry(cpuset_mems_cookie)); |
| 3079 | out: |
| 3080 | return ret; |
| 3081 | } |
| 3082 | |
| 3083 | #define K(x) ((x) << (PAGE_SHIFT-10)) |
| 3084 | |
| 3085 | static void show_migration_types(unsigned char type) |
| 3086 | { |
| 3087 | static const char types[MIGRATE_TYPES] = { |
| 3088 | [MIGRATE_UNMOVABLE] = 'U', |
| 3089 | [MIGRATE_RECLAIMABLE] = 'E', |
| 3090 | [MIGRATE_MOVABLE] = 'M', |
| 3091 | [MIGRATE_RESERVE] = 'R', |
| 3092 | #ifdef CONFIG_CMA |
| 3093 | [MIGRATE_CMA] = 'C', |
| 3094 | #endif |
| 3095 | #ifdef CONFIG_MEMORY_ISOLATION |
| 3096 | [MIGRATE_ISOLATE] = 'I', |
| 3097 | #endif |
| 3098 | }; |
| 3099 | char tmp[MIGRATE_TYPES + 1]; |
| 3100 | char *p = tmp; |
| 3101 | int i; |
| 3102 | |
| 3103 | for (i = 0; i < MIGRATE_TYPES; i++) { |
| 3104 | if (type & (1 << i)) |
| 3105 | *p++ = types[i]; |
| 3106 | } |
| 3107 | |
| 3108 | *p = '\0'; |
| 3109 | printk("(%s) ", tmp); |
| 3110 | } |
| 3111 | |
| 3112 | /* |
| 3113 | * Show free area list (used inside shift_scroll-lock stuff) |
| 3114 | * We also calculate the percentage fragmentation. We do this by counting the |
| 3115 | * memory on each free list with the exception of the first item on the list. |
| 3116 | * Suppresses nodes that are not allowed by current's cpuset if |
| 3117 | * SHOW_MEM_FILTER_NODES is passed. |
| 3118 | */ |
| 3119 | void show_free_areas(unsigned int filter) |
| 3120 | { |
| 3121 | int cpu; |
| 3122 | struct zone *zone; |
| 3123 | |
| 3124 | for_each_populated_zone(zone) { |
| 3125 | if (skip_free_areas_node(filter, zone_to_nid(zone))) |
| 3126 | continue; |
| 3127 | show_node(zone); |
| 3128 | printk("%s per-cpu:\n", zone->name); |
| 3129 | |
| 3130 | for_each_online_cpu(cpu) { |
| 3131 | struct per_cpu_pageset *pageset; |
| 3132 | |
| 3133 | pageset = per_cpu_ptr(zone->pageset, cpu); |
| 3134 | |
| 3135 | printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", |
| 3136 | cpu, pageset->pcp.high, |
| 3137 | pageset->pcp.batch, pageset->pcp.count); |
| 3138 | } |
| 3139 | } |
| 3140 | |
| 3141 | printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" |
| 3142 | " active_file:%lu inactive_file:%lu isolated_file:%lu\n" |
| 3143 | " unevictable:%lu" |
| 3144 | " dirty:%lu writeback:%lu unstable:%lu\n" |
| 3145 | " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" |
| 3146 | " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n" |
| 3147 | " free_cma:%lu\n", |
| 3148 | global_page_state(NR_ACTIVE_ANON), |
| 3149 | global_page_state(NR_INACTIVE_ANON), |
| 3150 | global_page_state(NR_ISOLATED_ANON), |
| 3151 | global_page_state(NR_ACTIVE_FILE), |
| 3152 | global_page_state(NR_INACTIVE_FILE), |
| 3153 | global_page_state(NR_ISOLATED_FILE), |
| 3154 | global_page_state(NR_UNEVICTABLE), |
| 3155 | global_page_state(NR_FILE_DIRTY), |
| 3156 | global_page_state(NR_WRITEBACK), |
| 3157 | global_page_state(NR_UNSTABLE_NFS), |
| 3158 | global_page_state(NR_FREE_PAGES), |
| 3159 | global_page_state(NR_SLAB_RECLAIMABLE), |
| 3160 | global_page_state(NR_SLAB_UNRECLAIMABLE), |
| 3161 | global_page_state(NR_FILE_MAPPED), |
| 3162 | global_page_state(NR_SHMEM), |
| 3163 | global_page_state(NR_PAGETABLE), |
| 3164 | global_page_state(NR_BOUNCE), |
| 3165 | global_page_state(NR_FREE_CMA_PAGES)); |
| 3166 | |
| 3167 | for_each_populated_zone(zone) { |
| 3168 | int i; |
| 3169 | |
| 3170 | if (skip_free_areas_node(filter, zone_to_nid(zone))) |
| 3171 | continue; |
| 3172 | show_node(zone); |
| 3173 | printk("%s" |
| 3174 | " free:%lukB" |
| 3175 | " min:%lukB" |
| 3176 | " low:%lukB" |
| 3177 | " high:%lukB" |
| 3178 | " active_anon:%lukB" |
| 3179 | " inactive_anon:%lukB" |
| 3180 | " active_file:%lukB" |
| 3181 | " inactive_file:%lukB" |
| 3182 | " unevictable:%lukB" |
| 3183 | " isolated(anon):%lukB" |
| 3184 | " isolated(file):%lukB" |
| 3185 | " present:%lukB" |
| 3186 | " managed:%lukB" |
| 3187 | " mlocked:%lukB" |
| 3188 | " dirty:%lukB" |
| 3189 | " writeback:%lukB" |
| 3190 | " mapped:%lukB" |
| 3191 | " shmem:%lukB" |
| 3192 | " slab_reclaimable:%lukB" |
| 3193 | " slab_unreclaimable:%lukB" |
| 3194 | " kernel_stack:%lukB" |
| 3195 | " pagetables:%lukB" |
| 3196 | " unstable:%lukB" |
| 3197 | " bounce:%lukB" |
| 3198 | " free_cma:%lukB" |
| 3199 | " writeback_tmp:%lukB" |
| 3200 | " pages_scanned:%lu" |
| 3201 | " all_unreclaimable? %s" |
| 3202 | "\n", |
| 3203 | zone->name, |
| 3204 | K(zone_page_state(zone, NR_FREE_PAGES)), |
| 3205 | K(min_wmark_pages(zone)), |
| 3206 | K(low_wmark_pages(zone)), |
| 3207 | K(high_wmark_pages(zone)), |
| 3208 | K(zone_page_state(zone, NR_ACTIVE_ANON)), |
| 3209 | K(zone_page_state(zone, NR_INACTIVE_ANON)), |
| 3210 | K(zone_page_state(zone, NR_ACTIVE_FILE)), |
| 3211 | K(zone_page_state(zone, NR_INACTIVE_FILE)), |
| 3212 | K(zone_page_state(zone, NR_UNEVICTABLE)), |
| 3213 | K(zone_page_state(zone, NR_ISOLATED_ANON)), |
| 3214 | K(zone_page_state(zone, NR_ISOLATED_FILE)), |
| 3215 | K(zone->present_pages), |
| 3216 | K(zone->managed_pages), |
| 3217 | K(zone_page_state(zone, NR_MLOCK)), |
| 3218 | K(zone_page_state(zone, NR_FILE_DIRTY)), |
| 3219 | K(zone_page_state(zone, NR_WRITEBACK)), |
| 3220 | K(zone_page_state(zone, NR_FILE_MAPPED)), |
| 3221 | K(zone_page_state(zone, NR_SHMEM)), |
| 3222 | K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), |
| 3223 | K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), |
| 3224 | zone_page_state(zone, NR_KERNEL_STACK) * |
| 3225 | THREAD_SIZE / 1024, |
| 3226 | K(zone_page_state(zone, NR_PAGETABLE)), |
| 3227 | K(zone_page_state(zone, NR_UNSTABLE_NFS)), |
| 3228 | K(zone_page_state(zone, NR_BOUNCE)), |
| 3229 | K(zone_page_state(zone, NR_FREE_CMA_PAGES)), |
| 3230 | K(zone_page_state(zone, NR_WRITEBACK_TEMP)), |
| 3231 | zone->pages_scanned, |
| 3232 | (!zone_reclaimable(zone) ? "yes" : "no") |
| 3233 | ); |
| 3234 | printk("lowmem_reserve[]:"); |
| 3235 | for (i = 0; i < MAX_NR_ZONES; i++) |
| 3236 | printk(" %lu", zone->lowmem_reserve[i]); |
| 3237 | printk("\n"); |
| 3238 | } |
| 3239 | |
| 3240 | for_each_populated_zone(zone) { |
| 3241 | unsigned long nr[MAX_ORDER], flags, order, total = 0; |
| 3242 | unsigned char types[MAX_ORDER]; |
| 3243 | |
| 3244 | if (skip_free_areas_node(filter, zone_to_nid(zone))) |
| 3245 | continue; |
| 3246 | show_node(zone); |
| 3247 | printk("%s: ", zone->name); |
| 3248 | |
| 3249 | spin_lock_irqsave(&zone->lock, flags); |
| 3250 | for (order = 0; order < MAX_ORDER; order++) { |
| 3251 | struct free_area *area = &zone->free_area[order]; |
| 3252 | int type; |
| 3253 | |
| 3254 | nr[order] = area->nr_free; |
| 3255 | total += nr[order] << order; |
| 3256 | |
| 3257 | types[order] = 0; |
| 3258 | for (type = 0; type < MIGRATE_TYPES; type++) { |
| 3259 | if (!list_empty(&area->free_list[type])) |
| 3260 | types[order] |= 1 << type; |
| 3261 | } |
| 3262 | } |
| 3263 | spin_unlock_irqrestore(&zone->lock, flags); |
| 3264 | for (order = 0; order < MAX_ORDER; order++) { |
| 3265 | printk("%lu*%lukB ", nr[order], K(1UL) << order); |
| 3266 | if (nr[order]) |
| 3267 | show_migration_types(types[order]); |
| 3268 | } |
| 3269 | printk("= %lukB\n", K(total)); |
| 3270 | } |
| 3271 | |
| 3272 | hugetlb_show_meminfo(); |
| 3273 | |
| 3274 | printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); |
| 3275 | |
| 3276 | show_swap_cache_info(); |
| 3277 | } |
| 3278 | |
| 3279 | static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) |
| 3280 | { |
| 3281 | zoneref->zone = zone; |
| 3282 | zoneref->zone_idx = zone_idx(zone); |
| 3283 | } |
| 3284 | |
| 3285 | /* |
| 3286 | * Builds allocation fallback zone lists. |
| 3287 | * |
| 3288 | * Add all populated zones of a node to the zonelist. |
| 3289 | */ |
| 3290 | static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, |
| 3291 | int nr_zones) |
| 3292 | { |
| 3293 | struct zone *zone; |
| 3294 | enum zone_type zone_type = MAX_NR_ZONES; |
| 3295 | |
| 3296 | do { |
| 3297 | zone_type--; |
| 3298 | zone = pgdat->node_zones + zone_type; |
| 3299 | if (populated_zone(zone)) { |
| 3300 | zoneref_set_zone(zone, |
| 3301 | &zonelist->_zonerefs[nr_zones++]); |
| 3302 | check_highest_zone(zone_type); |
| 3303 | } |
| 3304 | } while (zone_type); |
| 3305 | |
| 3306 | return nr_zones; |
| 3307 | } |
| 3308 | |
| 3309 | |
| 3310 | /* |
| 3311 | * zonelist_order: |
| 3312 | * 0 = automatic detection of better ordering. |
| 3313 | * 1 = order by ([node] distance, -zonetype) |
| 3314 | * 2 = order by (-zonetype, [node] distance) |
| 3315 | * |
| 3316 | * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create |
| 3317 | * the same zonelist. So only NUMA can configure this param. |
| 3318 | */ |
| 3319 | #define ZONELIST_ORDER_DEFAULT 0 |
| 3320 | #define ZONELIST_ORDER_NODE 1 |
| 3321 | #define ZONELIST_ORDER_ZONE 2 |
| 3322 | |
| 3323 | /* zonelist order in the kernel. |
| 3324 | * set_zonelist_order() will set this to NODE or ZONE. |
| 3325 | */ |
| 3326 | static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; |
| 3327 | static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; |
| 3328 | |
| 3329 | |
| 3330 | #ifdef CONFIG_NUMA |
| 3331 | /* The value user specified ....changed by config */ |
| 3332 | static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; |
| 3333 | /* string for sysctl */ |
| 3334 | #define NUMA_ZONELIST_ORDER_LEN 16 |
| 3335 | char numa_zonelist_order[16] = "default"; |
| 3336 | |
| 3337 | /* |
| 3338 | * interface for configure zonelist ordering. |
| 3339 | * command line option "numa_zonelist_order" |
| 3340 | * = "[dD]efault - default, automatic configuration. |
| 3341 | * = "[nN]ode - order by node locality, then by zone within node |
| 3342 | * = "[zZ]one - order by zone, then by locality within zone |
| 3343 | */ |
| 3344 | |
| 3345 | static int __parse_numa_zonelist_order(char *s) |
| 3346 | { |
| 3347 | if (*s == 'd' || *s == 'D') { |
| 3348 | user_zonelist_order = ZONELIST_ORDER_DEFAULT; |
| 3349 | } else if (*s == 'n' || *s == 'N') { |
| 3350 | user_zonelist_order = ZONELIST_ORDER_NODE; |
| 3351 | } else if (*s == 'z' || *s == 'Z') { |
| 3352 | user_zonelist_order = ZONELIST_ORDER_ZONE; |
| 3353 | } else { |
| 3354 | printk(KERN_WARNING |
| 3355 | "Ignoring invalid numa_zonelist_order value: " |
| 3356 | "%s\n", s); |
| 3357 | return -EINVAL; |
| 3358 | } |
| 3359 | return 0; |
| 3360 | } |
| 3361 | |
| 3362 | static __init int setup_numa_zonelist_order(char *s) |
| 3363 | { |
| 3364 | int ret; |
| 3365 | |
| 3366 | if (!s) |
| 3367 | return 0; |
| 3368 | |
| 3369 | ret = __parse_numa_zonelist_order(s); |
| 3370 | if (ret == 0) |
| 3371 | strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); |
| 3372 | |
| 3373 | return ret; |
| 3374 | } |
| 3375 | early_param("numa_zonelist_order", setup_numa_zonelist_order); |
| 3376 | |
| 3377 | /* |
| 3378 | * sysctl handler for numa_zonelist_order |
| 3379 | */ |
| 3380 | int numa_zonelist_order_handler(ctl_table *table, int write, |
| 3381 | void __user *buffer, size_t *length, |
| 3382 | loff_t *ppos) |
| 3383 | { |
| 3384 | char saved_string[NUMA_ZONELIST_ORDER_LEN]; |
| 3385 | int ret; |
| 3386 | static DEFINE_MUTEX(zl_order_mutex); |
| 3387 | |
| 3388 | mutex_lock(&zl_order_mutex); |
| 3389 | if (write) { |
| 3390 | if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) { |
| 3391 | ret = -EINVAL; |
| 3392 | goto out; |
| 3393 | } |
| 3394 | strcpy(saved_string, (char *)table->data); |
| 3395 | } |
| 3396 | ret = proc_dostring(table, write, buffer, length, ppos); |
| 3397 | if (ret) |
| 3398 | goto out; |
| 3399 | if (write) { |
| 3400 | int oldval = user_zonelist_order; |
| 3401 | |
| 3402 | ret = __parse_numa_zonelist_order((char *)table->data); |
| 3403 | if (ret) { |
| 3404 | /* |
| 3405 | * bogus value. restore saved string |
| 3406 | */ |
| 3407 | strncpy((char *)table->data, saved_string, |
| 3408 | NUMA_ZONELIST_ORDER_LEN); |
| 3409 | user_zonelist_order = oldval; |
| 3410 | } else if (oldval != user_zonelist_order) { |
| 3411 | mutex_lock(&zonelists_mutex); |
| 3412 | build_all_zonelists(NULL, NULL); |
| 3413 | mutex_unlock(&zonelists_mutex); |
| 3414 | } |
| 3415 | } |
| 3416 | out: |
| 3417 | mutex_unlock(&zl_order_mutex); |
| 3418 | return ret; |
| 3419 | } |
| 3420 | |
| 3421 | |
| 3422 | #define MAX_NODE_LOAD (nr_online_nodes) |
| 3423 | static int node_load[MAX_NUMNODES]; |
| 3424 | |
| 3425 | /** |
| 3426 | * find_next_best_node - find the next node that should appear in a given node's fallback list |
| 3427 | * @node: node whose fallback list we're appending |
| 3428 | * @used_node_mask: nodemask_t of already used nodes |
| 3429 | * |
| 3430 | * We use a number of factors to determine which is the next node that should |
| 3431 | * appear on a given node's fallback list. The node should not have appeared |
| 3432 | * already in @node's fallback list, and it should be the next closest node |
| 3433 | * according to the distance array (which contains arbitrary distance values |
| 3434 | * from each node to each node in the system), and should also prefer nodes |
| 3435 | * with no CPUs, since presumably they'll have very little allocation pressure |
| 3436 | * on them otherwise. |
| 3437 | * It returns -1 if no node is found. |
| 3438 | */ |
| 3439 | static int find_next_best_node(int node, nodemask_t *used_node_mask) |
| 3440 | { |
| 3441 | int n, val; |
| 3442 | int min_val = INT_MAX; |
| 3443 | int best_node = NUMA_NO_NODE; |
| 3444 | const struct cpumask *tmp = cpumask_of_node(0); |
| 3445 | |
| 3446 | /* Use the local node if we haven't already */ |
| 3447 | if (!node_isset(node, *used_node_mask)) { |
| 3448 | node_set(node, *used_node_mask); |
| 3449 | return node; |
| 3450 | } |
| 3451 | |
| 3452 | for_each_node_state(n, N_MEMORY) { |
| 3453 | |
| 3454 | /* Don't want a node to appear more than once */ |
| 3455 | if (node_isset(n, *used_node_mask)) |
| 3456 | continue; |
| 3457 | |
| 3458 | /* Use the distance array to find the distance */ |
| 3459 | val = node_distance(node, n); |
| 3460 | |
| 3461 | /* Penalize nodes under us ("prefer the next node") */ |
| 3462 | val += (n < node); |
| 3463 | |
| 3464 | /* Give preference to headless and unused nodes */ |
| 3465 | tmp = cpumask_of_node(n); |
| 3466 | if (!cpumask_empty(tmp)) |
| 3467 | val += PENALTY_FOR_NODE_WITH_CPUS; |
| 3468 | |
| 3469 | /* Slight preference for less loaded node */ |
| 3470 | val *= (MAX_NODE_LOAD*MAX_NUMNODES); |
| 3471 | val += node_load[n]; |
| 3472 | |
| 3473 | if (val < min_val) { |
| 3474 | min_val = val; |
| 3475 | best_node = n; |
| 3476 | } |
| 3477 | } |
| 3478 | |
| 3479 | if (best_node >= 0) |
| 3480 | node_set(best_node, *used_node_mask); |
| 3481 | |
| 3482 | return best_node; |
| 3483 | } |
| 3484 | |
| 3485 | |
| 3486 | /* |
| 3487 | * Build zonelists ordered by node and zones within node. |
| 3488 | * This results in maximum locality--normal zone overflows into local |
| 3489 | * DMA zone, if any--but risks exhausting DMA zone. |
| 3490 | */ |
| 3491 | static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) |
| 3492 | { |
| 3493 | int j; |
| 3494 | struct zonelist *zonelist; |
| 3495 | |
| 3496 | zonelist = &pgdat->node_zonelists[0]; |
| 3497 | for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) |
| 3498 | ; |
| 3499 | j = build_zonelists_node(NODE_DATA(node), zonelist, j); |
| 3500 | zonelist->_zonerefs[j].zone = NULL; |
| 3501 | zonelist->_zonerefs[j].zone_idx = 0; |
| 3502 | } |
| 3503 | |
| 3504 | /* |
| 3505 | * Build gfp_thisnode zonelists |
| 3506 | */ |
| 3507 | static void build_thisnode_zonelists(pg_data_t *pgdat) |
| 3508 | { |
| 3509 | int j; |
| 3510 | struct zonelist *zonelist; |
| 3511 | |
| 3512 | zonelist = &pgdat->node_zonelists[1]; |
| 3513 | j = build_zonelists_node(pgdat, zonelist, 0); |
| 3514 | zonelist->_zonerefs[j].zone = NULL; |
| 3515 | zonelist->_zonerefs[j].zone_idx = 0; |
| 3516 | } |
| 3517 | |
| 3518 | /* |
| 3519 | * Build zonelists ordered by zone and nodes within zones. |
| 3520 | * This results in conserving DMA zone[s] until all Normal memory is |
| 3521 | * exhausted, but results in overflowing to remote node while memory |
| 3522 | * may still exist in local DMA zone. |
| 3523 | */ |
| 3524 | static int node_order[MAX_NUMNODES]; |
| 3525 | |
| 3526 | static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) |
| 3527 | { |
| 3528 | int pos, j, node; |
| 3529 | int zone_type; /* needs to be signed */ |
| 3530 | struct zone *z; |
| 3531 | struct zonelist *zonelist; |
| 3532 | |
| 3533 | zonelist = &pgdat->node_zonelists[0]; |
| 3534 | pos = 0; |
| 3535 | for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { |
| 3536 | for (j = 0; j < nr_nodes; j++) { |
| 3537 | node = node_order[j]; |
| 3538 | z = &NODE_DATA(node)->node_zones[zone_type]; |
| 3539 | if (populated_zone(z)) { |
| 3540 | zoneref_set_zone(z, |
| 3541 | &zonelist->_zonerefs[pos++]); |
| 3542 | check_highest_zone(zone_type); |
| 3543 | } |
| 3544 | } |
| 3545 | } |
| 3546 | zonelist->_zonerefs[pos].zone = NULL; |
| 3547 | zonelist->_zonerefs[pos].zone_idx = 0; |
| 3548 | } |
| 3549 | |
| 3550 | static int default_zonelist_order(void) |
| 3551 | { |
| 3552 | int nid, zone_type; |
| 3553 | unsigned long low_kmem_size, total_size; |
| 3554 | struct zone *z; |
| 3555 | int average_size; |
| 3556 | /* |
| 3557 | * ZONE_DMA and ZONE_DMA32 can be very small area in the system. |
| 3558 | * If they are really small and used heavily, the system can fall |
| 3559 | * into OOM very easily. |
| 3560 | * This function detect ZONE_DMA/DMA32 size and configures zone order. |
| 3561 | */ |
| 3562 | /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ |
| 3563 | low_kmem_size = 0; |
| 3564 | total_size = 0; |
| 3565 | for_each_online_node(nid) { |
| 3566 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { |
| 3567 | z = &NODE_DATA(nid)->node_zones[zone_type]; |
| 3568 | if (populated_zone(z)) { |
| 3569 | if (zone_type < ZONE_NORMAL) |
| 3570 | low_kmem_size += z->managed_pages; |
| 3571 | total_size += z->managed_pages; |
| 3572 | } else if (zone_type == ZONE_NORMAL) { |
| 3573 | /* |
| 3574 | * If any node has only lowmem, then node order |
| 3575 | * is preferred to allow kernel allocations |
| 3576 | * locally; otherwise, they can easily infringe |
| 3577 | * on other nodes when there is an abundance of |
| 3578 | * lowmem available to allocate from. |
| 3579 | */ |
| 3580 | return ZONELIST_ORDER_NODE; |
| 3581 | } |
| 3582 | } |
| 3583 | } |
| 3584 | if (!low_kmem_size || /* there are no DMA area. */ |
| 3585 | low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ |
| 3586 | return ZONELIST_ORDER_NODE; |
| 3587 | /* |
| 3588 | * look into each node's config. |
| 3589 | * If there is a node whose DMA/DMA32 memory is very big area on |
| 3590 | * local memory, NODE_ORDER may be suitable. |
| 3591 | */ |
| 3592 | average_size = total_size / |
| 3593 | (nodes_weight(node_states[N_MEMORY]) + 1); |
| 3594 | for_each_online_node(nid) { |
| 3595 | low_kmem_size = 0; |
| 3596 | total_size = 0; |
| 3597 | for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { |
| 3598 | z = &NODE_DATA(nid)->node_zones[zone_type]; |
| 3599 | if (populated_zone(z)) { |
| 3600 | if (zone_type < ZONE_NORMAL) |
| 3601 | low_kmem_size += z->present_pages; |
| 3602 | total_size += z->present_pages; |
| 3603 | } |
| 3604 | } |
| 3605 | if (low_kmem_size && |
| 3606 | total_size > average_size && /* ignore small node */ |
| 3607 | low_kmem_size > total_size * 70/100) |
| 3608 | return ZONELIST_ORDER_NODE; |
| 3609 | } |
| 3610 | return ZONELIST_ORDER_ZONE; |
| 3611 | } |
| 3612 | |
| 3613 | static void set_zonelist_order(void) |
| 3614 | { |
| 3615 | if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) |
| 3616 | current_zonelist_order = default_zonelist_order(); |
| 3617 | else |
| 3618 | current_zonelist_order = user_zonelist_order; |
| 3619 | } |
| 3620 | |
| 3621 | static void build_zonelists(pg_data_t *pgdat) |
| 3622 | { |
| 3623 | int j, node, load; |
| 3624 | enum zone_type i; |
| 3625 | nodemask_t used_mask; |
| 3626 | int local_node, prev_node; |
| 3627 | struct zonelist *zonelist; |
| 3628 | int order = current_zonelist_order; |
| 3629 | |
| 3630 | /* initialize zonelists */ |
| 3631 | for (i = 0; i < MAX_ZONELISTS; i++) { |
| 3632 | zonelist = pgdat->node_zonelists + i; |
| 3633 | zonelist->_zonerefs[0].zone = NULL; |
| 3634 | zonelist->_zonerefs[0].zone_idx = 0; |
| 3635 | } |
| 3636 | |
| 3637 | /* NUMA-aware ordering of nodes */ |
| 3638 | local_node = pgdat->node_id; |
| 3639 | load = nr_online_nodes; |
| 3640 | prev_node = local_node; |
| 3641 | nodes_clear(used_mask); |
| 3642 | |
| 3643 | memset(node_order, 0, sizeof(node_order)); |
| 3644 | j = 0; |
| 3645 | |
| 3646 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { |
| 3647 | /* |
| 3648 | * We don't want to pressure a particular node. |
| 3649 | * So adding penalty to the first node in same |
| 3650 | * distance group to make it round-robin. |
| 3651 | */ |
| 3652 | if (node_distance(local_node, node) != |
| 3653 | node_distance(local_node, prev_node)) |
| 3654 | node_load[node] = load; |
| 3655 | |
| 3656 | prev_node = node; |
| 3657 | load--; |
| 3658 | if (order == ZONELIST_ORDER_NODE) |
| 3659 | build_zonelists_in_node_order(pgdat, node); |
| 3660 | else |
| 3661 | node_order[j++] = node; /* remember order */ |
| 3662 | } |
| 3663 | |
| 3664 | if (order == ZONELIST_ORDER_ZONE) { |
| 3665 | /* calculate node order -- i.e., DMA last! */ |
| 3666 | build_zonelists_in_zone_order(pgdat, j); |
| 3667 | } |
| 3668 | |
| 3669 | build_thisnode_zonelists(pgdat); |
| 3670 | } |
| 3671 | |
| 3672 | /* Construct the zonelist performance cache - see further mmzone.h */ |
| 3673 | static void build_zonelist_cache(pg_data_t *pgdat) |
| 3674 | { |
| 3675 | struct zonelist *zonelist; |
| 3676 | struct zonelist_cache *zlc; |
| 3677 | struct zoneref *z; |
| 3678 | |
| 3679 | zonelist = &pgdat->node_zonelists[0]; |
| 3680 | zonelist->zlcache_ptr = zlc = &zonelist->zlcache; |
| 3681 | bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); |
| 3682 | for (z = zonelist->_zonerefs; z->zone; z++) |
| 3683 | zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); |
| 3684 | } |
| 3685 | |
| 3686 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES |
| 3687 | /* |
| 3688 | * Return node id of node used for "local" allocations. |
| 3689 | * I.e., first node id of first zone in arg node's generic zonelist. |
| 3690 | * Used for initializing percpu 'numa_mem', which is used primarily |
| 3691 | * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. |
| 3692 | */ |
| 3693 | int local_memory_node(int node) |
| 3694 | { |
| 3695 | struct zone *zone; |
| 3696 | |
| 3697 | (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), |
| 3698 | gfp_zone(GFP_KERNEL), |
| 3699 | NULL, |
| 3700 | &zone); |
| 3701 | return zone->node; |
| 3702 | } |
| 3703 | #endif |
| 3704 | |
| 3705 | #else /* CONFIG_NUMA */ |
| 3706 | |
| 3707 | static void set_zonelist_order(void) |
| 3708 | { |
| 3709 | current_zonelist_order = ZONELIST_ORDER_ZONE; |
| 3710 | } |
| 3711 | |
| 3712 | static void build_zonelists(pg_data_t *pgdat) |
| 3713 | { |
| 3714 | int node, local_node; |
| 3715 | enum zone_type j; |
| 3716 | struct zonelist *zonelist; |
| 3717 | |
| 3718 | local_node = pgdat->node_id; |
| 3719 | |
| 3720 | zonelist = &pgdat->node_zonelists[0]; |
| 3721 | j = build_zonelists_node(pgdat, zonelist, 0); |
| 3722 | |
| 3723 | /* |
| 3724 | * Now we build the zonelist so that it contains the zones |
| 3725 | * of all the other nodes. |
| 3726 | * We don't want to pressure a particular node, so when |
| 3727 | * building the zones for node N, we make sure that the |
| 3728 | * zones coming right after the local ones are those from |
| 3729 | * node N+1 (modulo N) |
| 3730 | */ |
| 3731 | for (node = local_node + 1; node < MAX_NUMNODES; node++) { |
| 3732 | if (!node_online(node)) |
| 3733 | continue; |
| 3734 | j = build_zonelists_node(NODE_DATA(node), zonelist, j); |
| 3735 | } |
| 3736 | for (node = 0; node < local_node; node++) { |
| 3737 | if (!node_online(node)) |
| 3738 | continue; |
| 3739 | j = build_zonelists_node(NODE_DATA(node), zonelist, j); |
| 3740 | } |
| 3741 | |
| 3742 | zonelist->_zonerefs[j].zone = NULL; |
| 3743 | zonelist->_zonerefs[j].zone_idx = 0; |
| 3744 | } |
| 3745 | |
| 3746 | /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ |
| 3747 | static void build_zonelist_cache(pg_data_t *pgdat) |
| 3748 | { |
| 3749 | pgdat->node_zonelists[0].zlcache_ptr = NULL; |
| 3750 | } |
| 3751 | |
| 3752 | #endif /* CONFIG_NUMA */ |
| 3753 | |
| 3754 | /* |
| 3755 | * Boot pageset table. One per cpu which is going to be used for all |
| 3756 | * zones and all nodes. The parameters will be set in such a way |
| 3757 | * that an item put on a list will immediately be handed over to |
| 3758 | * the buddy list. This is safe since pageset manipulation is done |
| 3759 | * with interrupts disabled. |
| 3760 | * |
| 3761 | * The boot_pagesets must be kept even after bootup is complete for |
| 3762 | * unused processors and/or zones. They do play a role for bootstrapping |
| 3763 | * hotplugged processors. |
| 3764 | * |
| 3765 | * zoneinfo_show() and maybe other functions do |
| 3766 | * not check if the processor is online before following the pageset pointer. |
| 3767 | * Other parts of the kernel may not check if the zone is available. |
| 3768 | */ |
| 3769 | static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); |
| 3770 | static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); |
| 3771 | static void setup_zone_pageset(struct zone *zone); |
| 3772 | |
| 3773 | /* |
| 3774 | * Global mutex to protect against size modification of zonelists |
| 3775 | * as well as to serialize pageset setup for the new populated zone. |
| 3776 | */ |
| 3777 | DEFINE_MUTEX(zonelists_mutex); |
| 3778 | |
| 3779 | /* return values int ....just for stop_machine() */ |
| 3780 | static int __build_all_zonelists(void *data) |
| 3781 | { |
| 3782 | int nid; |
| 3783 | int cpu; |
| 3784 | pg_data_t *self = data; |
| 3785 | |
| 3786 | #ifdef CONFIG_NUMA |
| 3787 | memset(node_load, 0, sizeof(node_load)); |
| 3788 | #endif |
| 3789 | |
| 3790 | if (self && !node_online(self->node_id)) { |
| 3791 | build_zonelists(self); |
| 3792 | build_zonelist_cache(self); |
| 3793 | } |
| 3794 | |
| 3795 | for_each_online_node(nid) { |
| 3796 | pg_data_t *pgdat = NODE_DATA(nid); |
| 3797 | |
| 3798 | build_zonelists(pgdat); |
| 3799 | build_zonelist_cache(pgdat); |
| 3800 | } |
| 3801 | |
| 3802 | /* |
| 3803 | * Initialize the boot_pagesets that are going to be used |
| 3804 | * for bootstrapping processors. The real pagesets for |
| 3805 | * each zone will be allocated later when the per cpu |
| 3806 | * allocator is available. |
| 3807 | * |
| 3808 | * boot_pagesets are used also for bootstrapping offline |
| 3809 | * cpus if the system is already booted because the pagesets |
| 3810 | * are needed to initialize allocators on a specific cpu too. |
| 3811 | * F.e. the percpu allocator needs the page allocator which |
| 3812 | * needs the percpu allocator in order to allocate its pagesets |
| 3813 | * (a chicken-egg dilemma). |
| 3814 | */ |
| 3815 | for_each_possible_cpu(cpu) { |
| 3816 | setup_pageset(&per_cpu(boot_pageset, cpu), 0); |
| 3817 | |
| 3818 | #ifdef CONFIG_HAVE_MEMORYLESS_NODES |
| 3819 | /* |
| 3820 | * We now know the "local memory node" for each node-- |
| 3821 | * i.e., the node of the first zone in the generic zonelist. |
| 3822 | * Set up numa_mem percpu variable for on-line cpus. During |
| 3823 | * boot, only the boot cpu should be on-line; we'll init the |
| 3824 | * secondary cpus' numa_mem as they come on-line. During |
| 3825 | * node/memory hotplug, we'll fixup all on-line cpus. |
| 3826 | */ |
| 3827 | if (cpu_online(cpu)) |
| 3828 | set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); |
| 3829 | #endif |
| 3830 | } |
| 3831 | |
| 3832 | return 0; |
| 3833 | } |
| 3834 | |
| 3835 | /* |
| 3836 | * Called with zonelists_mutex held always |
| 3837 | * unless system_state == SYSTEM_BOOTING. |
| 3838 | */ |
| 3839 | void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone) |
| 3840 | { |
| 3841 | set_zonelist_order(); |
| 3842 | |
| 3843 | if (system_state == SYSTEM_BOOTING) { |
| 3844 | __build_all_zonelists(NULL); |
| 3845 | mminit_verify_zonelist(); |
| 3846 | cpuset_init_current_mems_allowed(); |
| 3847 | } else { |
| 3848 | #ifdef CONFIG_MEMORY_HOTPLUG |
| 3849 | if (zone) |
| 3850 | setup_zone_pageset(zone); |
| 3851 | #endif |
| 3852 | /* we have to stop all cpus to guarantee there is no user |
| 3853 | of zonelist */ |
| 3854 | stop_machine(__build_all_zonelists, pgdat, NULL); |
| 3855 | /* cpuset refresh routine should be here */ |
| 3856 | } |
| 3857 | vm_total_pages = nr_free_pagecache_pages(); |
| 3858 | /* |
| 3859 | * Disable grouping by mobility if the number of pages in the |
| 3860 | * system is too low to allow the mechanism to work. It would be |
| 3861 | * more accurate, but expensive to check per-zone. This check is |
| 3862 | * made on memory-hotadd so a system can start with mobility |
| 3863 | * disabled and enable it later |
| 3864 | */ |
| 3865 | if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) |
| 3866 | page_group_by_mobility_disabled = 1; |
| 3867 | else |
| 3868 | page_group_by_mobility_disabled = 0; |
| 3869 | |
| 3870 | printk("Built %i zonelists in %s order, mobility grouping %s. " |
| 3871 | "Total pages: %ld\n", |
| 3872 | nr_online_nodes, |
| 3873 | zonelist_order_name[current_zonelist_order], |
| 3874 | page_group_by_mobility_disabled ? "off" : "on", |
| 3875 | vm_total_pages); |
| 3876 | #ifdef CONFIG_NUMA |
| 3877 | printk("Policy zone: %s\n", zone_names[policy_zone]); |
| 3878 | #endif |
| 3879 | } |
| 3880 | |
| 3881 | /* |
| 3882 | * Helper functions to size the waitqueue hash table. |
| 3883 | * Essentially these want to choose hash table sizes sufficiently |
| 3884 | * large so that collisions trying to wait on pages are rare. |
| 3885 | * But in fact, the number of active page waitqueues on typical |
| 3886 | * systems is ridiculously low, less than 200. So this is even |
| 3887 | * conservative, even though it seems large. |
| 3888 | * |
| 3889 | * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to |
| 3890 | * waitqueues, i.e. the size of the waitq table given the number of pages. |
| 3891 | */ |
| 3892 | #define PAGES_PER_WAITQUEUE 256 |
| 3893 | |
| 3894 | #ifndef CONFIG_MEMORY_HOTPLUG |
| 3895 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) |
| 3896 | { |
| 3897 | unsigned long size = 1; |
| 3898 | |
| 3899 | pages /= PAGES_PER_WAITQUEUE; |
| 3900 | |
| 3901 | while (size < pages) |
| 3902 | size <<= 1; |
| 3903 | |
| 3904 | /* |
| 3905 | * Once we have dozens or even hundreds of threads sleeping |
| 3906 | * on IO we've got bigger problems than wait queue collision. |
| 3907 | * Limit the size of the wait table to a reasonable size. |
| 3908 | */ |
| 3909 | size = min(size, 4096UL); |
| 3910 | |
| 3911 | return max(size, 4UL); |
| 3912 | } |
| 3913 | #else |
| 3914 | /* |
| 3915 | * A zone's size might be changed by hot-add, so it is not possible to determine |
| 3916 | * a suitable size for its wait_table. So we use the maximum size now. |
| 3917 | * |
| 3918 | * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: |
| 3919 | * |
| 3920 | * i386 (preemption config) : 4096 x 16 = 64Kbyte. |
| 3921 | * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. |
| 3922 | * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. |
| 3923 | * |
| 3924 | * The maximum entries are prepared when a zone's memory is (512K + 256) pages |
| 3925 | * or more by the traditional way. (See above). It equals: |
| 3926 | * |
| 3927 | * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. |
| 3928 | * ia64(16K page size) : = ( 8G + 4M)byte. |
| 3929 | * powerpc (64K page size) : = (32G +16M)byte. |
| 3930 | */ |
| 3931 | static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) |
| 3932 | { |
| 3933 | return 4096UL; |
| 3934 | } |
| 3935 | #endif |
| 3936 | |
| 3937 | /* |
| 3938 | * This is an integer logarithm so that shifts can be used later |
| 3939 | * to extract the more random high bits from the multiplicative |
| 3940 | * hash function before the remainder is taken. |
| 3941 | */ |
| 3942 | static inline unsigned long wait_table_bits(unsigned long size) |
| 3943 | { |
| 3944 | return ffz(~size); |
| 3945 | } |
| 3946 | |
| 3947 | /* |
| 3948 | * Check if a pageblock contains reserved pages |
| 3949 | */ |
| 3950 | static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) |
| 3951 | { |
| 3952 | unsigned long pfn; |
| 3953 | |
| 3954 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { |
| 3955 | if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) |
| 3956 | return 1; |
| 3957 | } |
| 3958 | return 0; |
| 3959 | } |
| 3960 | |
| 3961 | /* |
| 3962 | * Mark a number of pageblocks as MIGRATE_RESERVE. The number |
| 3963 | * of blocks reserved is based on min_wmark_pages(zone). The memory within |
| 3964 | * the reserve will tend to store contiguous free pages. Setting min_free_kbytes |
| 3965 | * higher will lead to a bigger reserve which will get freed as contiguous |
| 3966 | * blocks as reclaim kicks in |
| 3967 | */ |
| 3968 | static void setup_zone_migrate_reserve(struct zone *zone) |
| 3969 | { |
| 3970 | unsigned long start_pfn, pfn, end_pfn, block_end_pfn; |
| 3971 | struct page *page; |
| 3972 | unsigned long block_migratetype; |
| 3973 | int reserve; |
| 3974 | int old_reserve; |
| 3975 | |
| 3976 | /* |
| 3977 | * Get the start pfn, end pfn and the number of blocks to reserve |
| 3978 | * We have to be careful to be aligned to pageblock_nr_pages to |
| 3979 | * make sure that we always check pfn_valid for the first page in |
| 3980 | * the block. |
| 3981 | */ |
| 3982 | start_pfn = zone->zone_start_pfn; |
| 3983 | end_pfn = zone_end_pfn(zone); |
| 3984 | start_pfn = roundup(start_pfn, pageblock_nr_pages); |
| 3985 | reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> |
| 3986 | pageblock_order; |
| 3987 | |
| 3988 | /* |
| 3989 | * Reserve blocks are generally in place to help high-order atomic |
| 3990 | * allocations that are short-lived. A min_free_kbytes value that |
| 3991 | * would result in more than 2 reserve blocks for atomic allocations |
| 3992 | * is assumed to be in place to help anti-fragmentation for the |
| 3993 | * future allocation of hugepages at runtime. |
| 3994 | */ |
| 3995 | reserve = min(2, reserve); |
| 3996 | old_reserve = zone->nr_migrate_reserve_block; |
| 3997 | |
| 3998 | /* When memory hot-add, we almost always need to do nothing */ |
| 3999 | if (reserve == old_reserve) |
| 4000 | return; |
| 4001 | zone->nr_migrate_reserve_block = reserve; |
| 4002 | |
| 4003 | for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { |
| 4004 | if (!pfn_valid(pfn)) |
| 4005 | continue; |
| 4006 | page = pfn_to_page(pfn); |
| 4007 | |
| 4008 | /* Watch out for overlapping nodes */ |
| 4009 | if (page_to_nid(page) != zone_to_nid(zone)) |
| 4010 | continue; |
| 4011 | |
| 4012 | block_migratetype = get_pageblock_migratetype(page); |
| 4013 | |
| 4014 | /* Only test what is necessary when the reserves are not met */ |
| 4015 | if (reserve > 0) { |
| 4016 | /* |
| 4017 | * Blocks with reserved pages will never free, skip |
| 4018 | * them. |
| 4019 | */ |
| 4020 | block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); |
| 4021 | if (pageblock_is_reserved(pfn, block_end_pfn)) |
| 4022 | continue; |
| 4023 | |
| 4024 | /* If this block is reserved, account for it */ |
| 4025 | if (block_migratetype == MIGRATE_RESERVE) { |
| 4026 | reserve--; |
| 4027 | continue; |
| 4028 | } |
| 4029 | |
| 4030 | /* Suitable for reserving if this block is movable */ |
| 4031 | if (block_migratetype == MIGRATE_MOVABLE) { |
| 4032 | set_pageblock_migratetype(page, |
| 4033 | MIGRATE_RESERVE); |
| 4034 | move_freepages_block(zone, page, |
| 4035 | MIGRATE_RESERVE); |
| 4036 | reserve--; |
| 4037 | continue; |
| 4038 | } |
| 4039 | } else if (!old_reserve) { |
| 4040 | /* |
| 4041 | * At boot time we don't need to scan the whole zone |
| 4042 | * for turning off MIGRATE_RESERVE. |
| 4043 | */ |
| 4044 | break; |
| 4045 | } |
| 4046 | |
| 4047 | /* |
| 4048 | * If the reserve is met and this is a previous reserved block, |
| 4049 | * take it back |
| 4050 | */ |
| 4051 | if (block_migratetype == MIGRATE_RESERVE) { |
| 4052 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); |
| 4053 | move_freepages_block(zone, page, MIGRATE_MOVABLE); |
| 4054 | } |
| 4055 | } |
| 4056 | } |
| 4057 | |
| 4058 | /* |
| 4059 | * Initially all pages are reserved - free ones are freed |
| 4060 | * up by free_all_bootmem() once the early boot process is |
| 4061 | * done. Non-atomic initialization, single-pass. |
| 4062 | */ |
| 4063 | void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, |
| 4064 | unsigned long start_pfn, enum memmap_context context) |
| 4065 | { |
| 4066 | struct page *page; |
| 4067 | unsigned long end_pfn = start_pfn + size; |
| 4068 | unsigned long pfn; |
| 4069 | struct zone *z; |
| 4070 | |
| 4071 | if (highest_memmap_pfn < end_pfn - 1) |
| 4072 | highest_memmap_pfn = end_pfn - 1; |
| 4073 | |
| 4074 | z = &NODE_DATA(nid)->node_zones[zone]; |
| 4075 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { |
| 4076 | /* |
| 4077 | * There can be holes in boot-time mem_map[]s |
| 4078 | * handed to this function. They do not |
| 4079 | * exist on hotplugged memory. |
| 4080 | */ |
| 4081 | if (context == MEMMAP_EARLY) { |
| 4082 | if (!early_pfn_valid(pfn)) |
| 4083 | continue; |
| 4084 | if (!early_pfn_in_nid(pfn, nid)) |
| 4085 | continue; |
| 4086 | } |
| 4087 | page = pfn_to_page(pfn); |
| 4088 | set_page_links(page, zone, nid, pfn); |
| 4089 | mminit_verify_page_links(page, zone, nid, pfn); |
| 4090 | init_page_count(page); |
| 4091 | page_mapcount_reset(page); |
| 4092 | page_cpupid_reset_last(page); |
| 4093 | SetPageReserved(page); |
| 4094 | /* |
| 4095 | * Mark the block movable so that blocks are reserved for |
| 4096 | * movable at startup. This will force kernel allocations |
| 4097 | * to reserve their blocks rather than leaking throughout |
| 4098 | * the address space during boot when many long-lived |
| 4099 | * kernel allocations are made. Later some blocks near |
| 4100 | * the start are marked MIGRATE_RESERVE by |
| 4101 | * setup_zone_migrate_reserve() |
| 4102 | * |
| 4103 | * bitmap is created for zone's valid pfn range. but memmap |
| 4104 | * can be created for invalid pages (for alignment) |
| 4105 | * check here not to call set_pageblock_migratetype() against |
| 4106 | * pfn out of zone. |
| 4107 | */ |
| 4108 | if ((z->zone_start_pfn <= pfn) |
| 4109 | && (pfn < zone_end_pfn(z)) |
| 4110 | && !(pfn & (pageblock_nr_pages - 1))) |
| 4111 | set_pageblock_migratetype(page, MIGRATE_MOVABLE); |
| 4112 | |
| 4113 | INIT_LIST_HEAD(&page->lru); |
| 4114 | #ifdef WANT_PAGE_VIRTUAL |
| 4115 | /* The shift won't overflow because ZONE_NORMAL is below 4G. */ |
| 4116 | if (!is_highmem_idx(zone)) |
| 4117 | set_page_address(page, __va(pfn << PAGE_SHIFT)); |
| 4118 | #endif |
| 4119 | } |
| 4120 | } |
| 4121 | |
| 4122 | static void __meminit zone_init_free_lists(struct zone *zone) |
| 4123 | { |
| 4124 | int order, t; |
| 4125 | for_each_migratetype_order(order, t) { |
| 4126 | INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); |
| 4127 | zone->free_area[order].nr_free = 0; |
| 4128 | } |
| 4129 | } |
| 4130 | |
| 4131 | #ifndef __HAVE_ARCH_MEMMAP_INIT |
| 4132 | #define memmap_init(size, nid, zone, start_pfn) \ |
| 4133 | memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) |
| 4134 | #endif |
| 4135 | |
| 4136 | static int __meminit zone_batchsize(struct zone *zone) |
| 4137 | { |
| 4138 | #ifdef CONFIG_MMU |
| 4139 | int batch; |
| 4140 | |
| 4141 | /* |
| 4142 | * The per-cpu-pages pools are set to around 1000th of the |
| 4143 | * size of the zone. But no more than 1/2 of a meg. |
| 4144 | * |
| 4145 | * OK, so we don't know how big the cache is. So guess. |
| 4146 | */ |
| 4147 | batch = zone->managed_pages / 1024; |
| 4148 | if (batch * PAGE_SIZE > 512 * 1024) |
| 4149 | batch = (512 * 1024) / PAGE_SIZE; |
| 4150 | batch /= 4; /* We effectively *= 4 below */ |
| 4151 | if (batch < 1) |
| 4152 | batch = 1; |
| 4153 | |
| 4154 | /* |
| 4155 | * Clamp the batch to a 2^n - 1 value. Having a power |
| 4156 | * of 2 value was found to be more likely to have |
| 4157 | * suboptimal cache aliasing properties in some cases. |
| 4158 | * |
| 4159 | * For example if 2 tasks are alternately allocating |
| 4160 | * batches of pages, one task can end up with a lot |
| 4161 | * of pages of one half of the possible page colors |
| 4162 | * and the other with pages of the other colors. |
| 4163 | */ |
| 4164 | batch = rounddown_pow_of_two(batch + batch/2) - 1; |
| 4165 | |
| 4166 | return batch; |
| 4167 | |
| 4168 | #else |
| 4169 | /* The deferral and batching of frees should be suppressed under NOMMU |
| 4170 | * conditions. |
| 4171 | * |
| 4172 | * The problem is that NOMMU needs to be able to allocate large chunks |
| 4173 | * of contiguous memory as there's no hardware page translation to |
| 4174 | * assemble apparent contiguous memory from discontiguous pages. |
| 4175 | * |
| 4176 | * Queueing large contiguous runs of pages for batching, however, |
| 4177 | * causes the pages to actually be freed in smaller chunks. As there |
| 4178 | * can be a significant delay between the individual batches being |
| 4179 | * recycled, this leads to the once large chunks of space being |
| 4180 | * fragmented and becoming unavailable for high-order allocations. |
| 4181 | */ |
| 4182 | return 0; |
| 4183 | #endif |
| 4184 | } |
| 4185 | |
| 4186 | /* |
| 4187 | * pcp->high and pcp->batch values are related and dependent on one another: |
| 4188 | * ->batch must never be higher then ->high. |
| 4189 | * The following function updates them in a safe manner without read side |
| 4190 | * locking. |
| 4191 | * |
| 4192 | * Any new users of pcp->batch and pcp->high should ensure they can cope with |
| 4193 | * those fields changing asynchronously (acording the the above rule). |
| 4194 | * |
| 4195 | * mutex_is_locked(&pcp_batch_high_lock) required when calling this function |
| 4196 | * outside of boot time (or some other assurance that no concurrent updaters |
| 4197 | * exist). |
| 4198 | */ |
| 4199 | static void pageset_update(struct per_cpu_pages *pcp, unsigned long high, |
| 4200 | unsigned long batch) |
| 4201 | { |
| 4202 | /* start with a fail safe value for batch */ |
| 4203 | pcp->batch = 1; |
| 4204 | smp_wmb(); |
| 4205 | |
| 4206 | /* Update high, then batch, in order */ |
| 4207 | pcp->high = high; |
| 4208 | smp_wmb(); |
| 4209 | |
| 4210 | pcp->batch = batch; |
| 4211 | } |
| 4212 | |
| 4213 | /* a companion to pageset_set_high() */ |
| 4214 | static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch) |
| 4215 | { |
| 4216 | pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch)); |
| 4217 | } |
| 4218 | |
| 4219 | static void pageset_init(struct per_cpu_pageset *p) |
| 4220 | { |
| 4221 | struct per_cpu_pages *pcp; |
| 4222 | int migratetype; |
| 4223 | |
| 4224 | memset(p, 0, sizeof(*p)); |
| 4225 | |
| 4226 | pcp = &p->pcp; |
| 4227 | pcp->count = 0; |
| 4228 | for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) |
| 4229 | INIT_LIST_HEAD(&pcp->lists[migratetype]); |
| 4230 | } |
| 4231 | |
| 4232 | static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) |
| 4233 | { |
| 4234 | pageset_init(p); |
| 4235 | pageset_set_batch(p, batch); |
| 4236 | } |
| 4237 | |
| 4238 | /* |
| 4239 | * pageset_set_high() sets the high water mark for hot per_cpu_pagelist |
| 4240 | * to the value high for the pageset p. |
| 4241 | */ |
| 4242 | static void pageset_set_high(struct per_cpu_pageset *p, |
| 4243 | unsigned long high) |
| 4244 | { |
| 4245 | unsigned long batch = max(1UL, high / 4); |
| 4246 | if ((high / 4) > (PAGE_SHIFT * 8)) |
| 4247 | batch = PAGE_SHIFT * 8; |
| 4248 | |
| 4249 | pageset_update(&p->pcp, high, batch); |
| 4250 | } |
| 4251 | |
| 4252 | static void __meminit pageset_set_high_and_batch(struct zone *zone, |
| 4253 | struct per_cpu_pageset *pcp) |
| 4254 | { |
| 4255 | if (percpu_pagelist_fraction) |
| 4256 | pageset_set_high(pcp, |
| 4257 | (zone->managed_pages / |
| 4258 | percpu_pagelist_fraction)); |
| 4259 | else |
| 4260 | pageset_set_batch(pcp, zone_batchsize(zone)); |
| 4261 | } |
| 4262 | |
| 4263 | static void __meminit zone_pageset_init(struct zone *zone, int cpu) |
| 4264 | { |
| 4265 | struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); |
| 4266 | |
| 4267 | pageset_init(pcp); |
| 4268 | pageset_set_high_and_batch(zone, pcp); |
| 4269 | } |
| 4270 | |
| 4271 | static void __meminit setup_zone_pageset(struct zone *zone) |
| 4272 | { |
| 4273 | int cpu; |
| 4274 | zone->pageset = alloc_percpu(struct per_cpu_pageset); |
| 4275 | for_each_possible_cpu(cpu) |
| 4276 | zone_pageset_init(zone, cpu); |
| 4277 | } |
| 4278 | |
| 4279 | /* |
| 4280 | * Allocate per cpu pagesets and initialize them. |
| 4281 | * Before this call only boot pagesets were available. |
| 4282 | */ |
| 4283 | void __init setup_per_cpu_pageset(void) |
| 4284 | { |
| 4285 | struct zone *zone; |
| 4286 | |
| 4287 | for_each_populated_zone(zone) |
| 4288 | setup_zone_pageset(zone); |
| 4289 | } |
| 4290 | |
| 4291 | static noinline __init_refok |
| 4292 | int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) |
| 4293 | { |
| 4294 | int i; |
| 4295 | size_t alloc_size; |
| 4296 | |
| 4297 | /* |
| 4298 | * The per-page waitqueue mechanism uses hashed waitqueues |
| 4299 | * per zone. |
| 4300 | */ |
| 4301 | zone->wait_table_hash_nr_entries = |
| 4302 | wait_table_hash_nr_entries(zone_size_pages); |
| 4303 | zone->wait_table_bits = |
| 4304 | wait_table_bits(zone->wait_table_hash_nr_entries); |
| 4305 | alloc_size = zone->wait_table_hash_nr_entries |
| 4306 | * sizeof(wait_queue_head_t); |
| 4307 | |
| 4308 | if (!slab_is_available()) { |
| 4309 | zone->wait_table = (wait_queue_head_t *) |
| 4310 | memblock_virt_alloc_node_nopanic( |
| 4311 | alloc_size, zone->zone_pgdat->node_id); |
| 4312 | } else { |
| 4313 | /* |
| 4314 | * This case means that a zone whose size was 0 gets new memory |
| 4315 | * via memory hot-add. |
| 4316 | * But it may be the case that a new node was hot-added. In |
| 4317 | * this case vmalloc() will not be able to use this new node's |
| 4318 | * memory - this wait_table must be initialized to use this new |
| 4319 | * node itself as well. |
| 4320 | * To use this new node's memory, further consideration will be |
| 4321 | * necessary. |
| 4322 | */ |
| 4323 | zone->wait_table = vmalloc(alloc_size); |
| 4324 | } |
| 4325 | if (!zone->wait_table) |
| 4326 | return -ENOMEM; |
| 4327 | |
| 4328 | for (i = 0; i < zone->wait_table_hash_nr_entries; ++i) |
| 4329 | init_waitqueue_head(zone->wait_table + i); |
| 4330 | |
| 4331 | return 0; |
| 4332 | } |
| 4333 | |
| 4334 | static __meminit void zone_pcp_init(struct zone *zone) |
| 4335 | { |
| 4336 | /* |
| 4337 | * per cpu subsystem is not up at this point. The following code |
| 4338 | * relies on the ability of the linker to provide the |
| 4339 | * offset of a (static) per cpu variable into the per cpu area. |
| 4340 | */ |
| 4341 | zone->pageset = &boot_pageset; |
| 4342 | |
| 4343 | if (populated_zone(zone)) |
| 4344 | printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", |
| 4345 | zone->name, zone->present_pages, |
| 4346 | zone_batchsize(zone)); |
| 4347 | } |
| 4348 | |
| 4349 | int __meminit init_currently_empty_zone(struct zone *zone, |
| 4350 | unsigned long zone_start_pfn, |
| 4351 | unsigned long size, |
| 4352 | enum memmap_context context) |
| 4353 | { |
| 4354 | struct pglist_data *pgdat = zone->zone_pgdat; |
| 4355 | int ret; |
| 4356 | ret = zone_wait_table_init(zone, size); |
| 4357 | if (ret) |
| 4358 | return ret; |
| 4359 | pgdat->nr_zones = zone_idx(zone) + 1; |
| 4360 | |
| 4361 | zone->zone_start_pfn = zone_start_pfn; |
| 4362 | |
| 4363 | mminit_dprintk(MMINIT_TRACE, "memmap_init", |
| 4364 | "Initialising map node %d zone %lu pfns %lu -> %lu\n", |
| 4365 | pgdat->node_id, |
| 4366 | (unsigned long)zone_idx(zone), |
| 4367 | zone_start_pfn, (zone_start_pfn + size)); |
| 4368 | |
| 4369 | zone_init_free_lists(zone); |
| 4370 | |
| 4371 | return 0; |
| 4372 | } |
| 4373 | |
| 4374 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP |
| 4375 | #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID |
| 4376 | /* |
| 4377 | * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. |
| 4378 | * Architectures may implement their own version but if add_active_range() |
| 4379 | * was used and there are no special requirements, this is a convenient |
| 4380 | * alternative |
| 4381 | */ |
| 4382 | int __meminit __early_pfn_to_nid(unsigned long pfn) |
| 4383 | { |
| 4384 | unsigned long start_pfn, end_pfn; |
| 4385 | int nid; |
| 4386 | /* |
| 4387 | * NOTE: The following SMP-unsafe globals are only used early in boot |
| 4388 | * when the kernel is running single-threaded. |
| 4389 | */ |
| 4390 | static unsigned long __meminitdata last_start_pfn, last_end_pfn; |
| 4391 | static int __meminitdata last_nid; |
| 4392 | |
| 4393 | if (last_start_pfn <= pfn && pfn < last_end_pfn) |
| 4394 | return last_nid; |
| 4395 | |
| 4396 | nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn); |
| 4397 | if (nid != -1) { |
| 4398 | last_start_pfn = start_pfn; |
| 4399 | last_end_pfn = end_pfn; |
| 4400 | last_nid = nid; |
| 4401 | } |
| 4402 | |
| 4403 | return nid; |
| 4404 | } |
| 4405 | #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ |
| 4406 | |
| 4407 | int __meminit early_pfn_to_nid(unsigned long pfn) |
| 4408 | { |
| 4409 | int nid; |
| 4410 | |
| 4411 | nid = __early_pfn_to_nid(pfn); |
| 4412 | if (nid >= 0) |
| 4413 | return nid; |
| 4414 | /* just returns 0 */ |
| 4415 | return 0; |
| 4416 | } |
| 4417 | |
| 4418 | #ifdef CONFIG_NODES_SPAN_OTHER_NODES |
| 4419 | bool __meminit early_pfn_in_nid(unsigned long pfn, int node) |
| 4420 | { |
| 4421 | int nid; |
| 4422 | |
| 4423 | nid = __early_pfn_to_nid(pfn); |
| 4424 | if (nid >= 0 && nid != node) |
| 4425 | return false; |
| 4426 | return true; |
| 4427 | } |
| 4428 | #endif |
| 4429 | |
| 4430 | /** |
| 4431 | * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range |
| 4432 | * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. |
| 4433 | * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid |
| 4434 | * |
| 4435 | * If an architecture guarantees that all ranges registered with |
| 4436 | * add_active_ranges() contain no holes and may be freed, this |
| 4437 | * this function may be used instead of calling memblock_free_early_nid() |
| 4438 | * manually. |
| 4439 | */ |
| 4440 | void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn) |
| 4441 | { |
| 4442 | unsigned long start_pfn, end_pfn; |
| 4443 | int i, this_nid; |
| 4444 | |
| 4445 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) { |
| 4446 | start_pfn = min(start_pfn, max_low_pfn); |
| 4447 | end_pfn = min(end_pfn, max_low_pfn); |
| 4448 | |
| 4449 | if (start_pfn < end_pfn) |
| 4450 | memblock_free_early_nid(PFN_PHYS(start_pfn), |
| 4451 | (end_pfn - start_pfn) << PAGE_SHIFT, |
| 4452 | this_nid); |
| 4453 | } |
| 4454 | } |
| 4455 | |
| 4456 | /** |
| 4457 | * sparse_memory_present_with_active_regions - Call memory_present for each active range |
| 4458 | * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. |
| 4459 | * |
| 4460 | * If an architecture guarantees that all ranges registered with |
| 4461 | * add_active_ranges() contain no holes and may be freed, this |
| 4462 | * function may be used instead of calling memory_present() manually. |
| 4463 | */ |
| 4464 | void __init sparse_memory_present_with_active_regions(int nid) |
| 4465 | { |
| 4466 | unsigned long start_pfn, end_pfn; |
| 4467 | int i, this_nid; |
| 4468 | |
| 4469 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) |
| 4470 | memory_present(this_nid, start_pfn, end_pfn); |
| 4471 | } |
| 4472 | |
| 4473 | /** |
| 4474 | * get_pfn_range_for_nid - Return the start and end page frames for a node |
| 4475 | * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. |
| 4476 | * @start_pfn: Passed by reference. On return, it will have the node start_pfn. |
| 4477 | * @end_pfn: Passed by reference. On return, it will have the node end_pfn. |
| 4478 | * |
| 4479 | * It returns the start and end page frame of a node based on information |
| 4480 | * provided by an arch calling add_active_range(). If called for a node |
| 4481 | * with no available memory, a warning is printed and the start and end |
| 4482 | * PFNs will be 0. |
| 4483 | */ |
| 4484 | void __meminit get_pfn_range_for_nid(unsigned int nid, |
| 4485 | unsigned long *start_pfn, unsigned long *end_pfn) |
| 4486 | { |
| 4487 | unsigned long this_start_pfn, this_end_pfn; |
| 4488 | int i; |
| 4489 | |
| 4490 | *start_pfn = -1UL; |
| 4491 | *end_pfn = 0; |
| 4492 | |
| 4493 | for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) { |
| 4494 | *start_pfn = min(*start_pfn, this_start_pfn); |
| 4495 | *end_pfn = max(*end_pfn, this_end_pfn); |
| 4496 | } |
| 4497 | |
| 4498 | if (*start_pfn == -1UL) |
| 4499 | *start_pfn = 0; |
| 4500 | } |
| 4501 | |
| 4502 | /* |
| 4503 | * This finds a zone that can be used for ZONE_MOVABLE pages. The |
| 4504 | * assumption is made that zones within a node are ordered in monotonic |
| 4505 | * increasing memory addresses so that the "highest" populated zone is used |
| 4506 | */ |
| 4507 | static void __init find_usable_zone_for_movable(void) |
| 4508 | { |
| 4509 | int zone_index; |
| 4510 | for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { |
| 4511 | if (zone_index == ZONE_MOVABLE) |
| 4512 | continue; |
| 4513 | |
| 4514 | if (arch_zone_highest_possible_pfn[zone_index] > |
| 4515 | arch_zone_lowest_possible_pfn[zone_index]) |
| 4516 | break; |
| 4517 | } |
| 4518 | |
| 4519 | VM_BUG_ON(zone_index == -1); |
| 4520 | movable_zone = zone_index; |
| 4521 | } |
| 4522 | |
| 4523 | /* |
| 4524 | * The zone ranges provided by the architecture do not include ZONE_MOVABLE |
| 4525 | * because it is sized independent of architecture. Unlike the other zones, |
| 4526 | * the starting point for ZONE_MOVABLE is not fixed. It may be different |
| 4527 | * in each node depending on the size of each node and how evenly kernelcore |
| 4528 | * is distributed. This helper function adjusts the zone ranges |
| 4529 | * provided by the architecture for a given node by using the end of the |
| 4530 | * highest usable zone for ZONE_MOVABLE. This preserves the assumption that |
| 4531 | * zones within a node are in order of monotonic increases memory addresses |
| 4532 | */ |
| 4533 | static void __meminit adjust_zone_range_for_zone_movable(int nid, |
| 4534 | unsigned long zone_type, |
| 4535 | unsigned long node_start_pfn, |
| 4536 | unsigned long node_end_pfn, |
| 4537 | unsigned long *zone_start_pfn, |
| 4538 | unsigned long *zone_end_pfn) |
| 4539 | { |
| 4540 | /* Only adjust if ZONE_MOVABLE is on this node */ |
| 4541 | if (zone_movable_pfn[nid]) { |
| 4542 | /* Size ZONE_MOVABLE */ |
| 4543 | if (zone_type == ZONE_MOVABLE) { |
| 4544 | *zone_start_pfn = zone_movable_pfn[nid]; |
| 4545 | *zone_end_pfn = min(node_end_pfn, |
| 4546 | arch_zone_highest_possible_pfn[movable_zone]); |
| 4547 | |
| 4548 | /* Adjust for ZONE_MOVABLE starting within this range */ |
| 4549 | } else if (*zone_start_pfn < zone_movable_pfn[nid] && |
| 4550 | *zone_end_pfn > zone_movable_pfn[nid]) { |
| 4551 | *zone_end_pfn = zone_movable_pfn[nid]; |
| 4552 | |
| 4553 | /* Check if this whole range is within ZONE_MOVABLE */ |
| 4554 | } else if (*zone_start_pfn >= zone_movable_pfn[nid]) |
| 4555 | *zone_start_pfn = *zone_end_pfn; |
| 4556 | } |
| 4557 | } |
| 4558 | |
| 4559 | /* |
| 4560 | * Return the number of pages a zone spans in a node, including holes |
| 4561 | * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() |
| 4562 | */ |
| 4563 | static unsigned long __meminit zone_spanned_pages_in_node(int nid, |
| 4564 | unsigned long zone_type, |
| 4565 | unsigned long node_start_pfn, |
| 4566 | unsigned long node_end_pfn, |
| 4567 | unsigned long *ignored) |
| 4568 | { |
| 4569 | unsigned long zone_start_pfn, zone_end_pfn; |
| 4570 | |
| 4571 | /* Get the start and end of the zone */ |
| 4572 | zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; |
| 4573 | zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; |
| 4574 | adjust_zone_range_for_zone_movable(nid, zone_type, |
| 4575 | node_start_pfn, node_end_pfn, |
| 4576 | &zone_start_pfn, &zone_end_pfn); |
| 4577 | |
| 4578 | /* Check that this node has pages within the zone's required range */ |
| 4579 | if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) |
| 4580 | return 0; |
| 4581 | |
| 4582 | /* Move the zone boundaries inside the node if necessary */ |
| 4583 | zone_end_pfn = min(zone_end_pfn, node_end_pfn); |
| 4584 | zone_start_pfn = max(zone_start_pfn, node_start_pfn); |
| 4585 | |
| 4586 | /* Return the spanned pages */ |
| 4587 | return zone_end_pfn - zone_start_pfn; |
| 4588 | } |
| 4589 | |
| 4590 | /* |
| 4591 | * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, |
| 4592 | * then all holes in the requested range will be accounted for. |
| 4593 | */ |
| 4594 | unsigned long __meminit __absent_pages_in_range(int nid, |
| 4595 | unsigned long range_start_pfn, |
| 4596 | unsigned long range_end_pfn) |
| 4597 | { |
| 4598 | unsigned long nr_absent = range_end_pfn - range_start_pfn; |
| 4599 | unsigned long start_pfn, end_pfn; |
| 4600 | int i; |
| 4601 | |
| 4602 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { |
| 4603 | start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn); |
| 4604 | end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn); |
| 4605 | nr_absent -= end_pfn - start_pfn; |
| 4606 | } |
| 4607 | return nr_absent; |
| 4608 | } |
| 4609 | |
| 4610 | /** |
| 4611 | * absent_pages_in_range - Return number of page frames in holes within a range |
| 4612 | * @start_pfn: The start PFN to start searching for holes |
| 4613 | * @end_pfn: The end PFN to stop searching for holes |
| 4614 | * |
| 4615 | * It returns the number of pages frames in memory holes within a range. |
| 4616 | */ |
| 4617 | unsigned long __init absent_pages_in_range(unsigned long start_pfn, |
| 4618 | unsigned long end_pfn) |
| 4619 | { |
| 4620 | return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); |
| 4621 | } |
| 4622 | |
| 4623 | /* Return the number of page frames in holes in a zone on a node */ |
| 4624 | static unsigned long __meminit zone_absent_pages_in_node(int nid, |
| 4625 | unsigned long zone_type, |
| 4626 | unsigned long node_start_pfn, |
| 4627 | unsigned long node_end_pfn, |
| 4628 | unsigned long *ignored) |
| 4629 | { |
| 4630 | unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type]; |
| 4631 | unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type]; |
| 4632 | unsigned long zone_start_pfn, zone_end_pfn; |
| 4633 | |
| 4634 | zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high); |
| 4635 | zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high); |
| 4636 | |
| 4637 | adjust_zone_range_for_zone_movable(nid, zone_type, |
| 4638 | node_start_pfn, node_end_pfn, |
| 4639 | &zone_start_pfn, &zone_end_pfn); |
| 4640 | return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); |
| 4641 | } |
| 4642 | |
| 4643 | #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ |
| 4644 | static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, |
| 4645 | unsigned long zone_type, |
| 4646 | unsigned long node_start_pfn, |
| 4647 | unsigned long node_end_pfn, |
| 4648 | unsigned long *zones_size) |
| 4649 | { |
| 4650 | return zones_size[zone_type]; |
| 4651 | } |
| 4652 | |
| 4653 | static inline unsigned long __meminit zone_absent_pages_in_node(int nid, |
| 4654 | unsigned long zone_type, |
| 4655 | unsigned long node_start_pfn, |
| 4656 | unsigned long node_end_pfn, |
| 4657 | unsigned long *zholes_size) |
| 4658 | { |
| 4659 | if (!zholes_size) |
| 4660 | return 0; |
| 4661 | |
| 4662 | return zholes_size[zone_type]; |
| 4663 | } |
| 4664 | |
| 4665 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ |
| 4666 | |
| 4667 | static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, |
| 4668 | unsigned long node_start_pfn, |
| 4669 | unsigned long node_end_pfn, |
| 4670 | unsigned long *zones_size, |
| 4671 | unsigned long *zholes_size) |
| 4672 | { |
| 4673 | unsigned long realtotalpages, totalpages = 0; |
| 4674 | enum zone_type i; |
| 4675 | |
| 4676 | for (i = 0; i < MAX_NR_ZONES; i++) |
| 4677 | totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, |
| 4678 | node_start_pfn, |
| 4679 | node_end_pfn, |
| 4680 | zones_size); |
| 4681 | pgdat->node_spanned_pages = totalpages; |
| 4682 | |
| 4683 | realtotalpages = totalpages; |
| 4684 | for (i = 0; i < MAX_NR_ZONES; i++) |
| 4685 | realtotalpages -= |
| 4686 | zone_absent_pages_in_node(pgdat->node_id, i, |
| 4687 | node_start_pfn, node_end_pfn, |
| 4688 | zholes_size); |
| 4689 | pgdat->node_present_pages = realtotalpages; |
| 4690 | printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, |
| 4691 | realtotalpages); |
| 4692 | } |
| 4693 | |
| 4694 | #ifndef CONFIG_SPARSEMEM |
| 4695 | /* |
| 4696 | * Calculate the size of the zone->blockflags rounded to an unsigned long |
| 4697 | * Start by making sure zonesize is a multiple of pageblock_order by rounding |
| 4698 | * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally |
| 4699 | * round what is now in bits to nearest long in bits, then return it in |
| 4700 | * bytes. |
| 4701 | */ |
| 4702 | static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize) |
| 4703 | { |
| 4704 | unsigned long usemapsize; |
| 4705 | |
| 4706 | zonesize += zone_start_pfn & (pageblock_nr_pages-1); |
| 4707 | usemapsize = roundup(zonesize, pageblock_nr_pages); |
| 4708 | usemapsize = usemapsize >> pageblock_order; |
| 4709 | usemapsize *= NR_PAGEBLOCK_BITS; |
| 4710 | usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); |
| 4711 | |
| 4712 | return usemapsize / 8; |
| 4713 | } |
| 4714 | |
| 4715 | static void __init setup_usemap(struct pglist_data *pgdat, |
| 4716 | struct zone *zone, |
| 4717 | unsigned long zone_start_pfn, |
| 4718 | unsigned long zonesize) |
| 4719 | { |
| 4720 | unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize); |
| 4721 | zone->pageblock_flags = NULL; |
| 4722 | if (usemapsize) |
| 4723 | zone->pageblock_flags = |
| 4724 | memblock_virt_alloc_node_nopanic(usemapsize, |
| 4725 | pgdat->node_id); |
| 4726 | } |
| 4727 | #else |
| 4728 | static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone, |
| 4729 | unsigned long zone_start_pfn, unsigned long zonesize) {} |
| 4730 | #endif /* CONFIG_SPARSEMEM */ |
| 4731 | |
| 4732 | #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE |
| 4733 | |
| 4734 | /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ |
| 4735 | void __paginginit set_pageblock_order(void) |
| 4736 | { |
| 4737 | unsigned int order; |
| 4738 | |
| 4739 | /* Check that pageblock_nr_pages has not already been setup */ |
| 4740 | if (pageblock_order) |
| 4741 | return; |
| 4742 | |
| 4743 | if (HPAGE_SHIFT > PAGE_SHIFT) |
| 4744 | order = HUGETLB_PAGE_ORDER; |
| 4745 | else |
| 4746 | order = MAX_ORDER - 1; |
| 4747 | |
| 4748 | /* |
| 4749 | * Assume the largest contiguous order of interest is a huge page. |
| 4750 | * This value may be variable depending on boot parameters on IA64 and |
| 4751 | * powerpc. |
| 4752 | */ |
| 4753 | pageblock_order = order; |
| 4754 | } |
| 4755 | #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ |
| 4756 | |
| 4757 | /* |
| 4758 | * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() |
| 4759 | * is unused as pageblock_order is set at compile-time. See |
| 4760 | * include/linux/pageblock-flags.h for the values of pageblock_order based on |
| 4761 | * the kernel config |
| 4762 | */ |
| 4763 | void __paginginit set_pageblock_order(void) |
| 4764 | { |
| 4765 | } |
| 4766 | |
| 4767 | #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ |
| 4768 | |
| 4769 | static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages, |
| 4770 | unsigned long present_pages) |
| 4771 | { |
| 4772 | unsigned long pages = spanned_pages; |
| 4773 | |
| 4774 | /* |
| 4775 | * Provide a more accurate estimation if there are holes within |
| 4776 | * the zone and SPARSEMEM is in use. If there are holes within the |
| 4777 | * zone, each populated memory region may cost us one or two extra |
| 4778 | * memmap pages due to alignment because memmap pages for each |
| 4779 | * populated regions may not naturally algined on page boundary. |
| 4780 | * So the (present_pages >> 4) heuristic is a tradeoff for that. |
| 4781 | */ |
| 4782 | if (spanned_pages > present_pages + (present_pages >> 4) && |
| 4783 | IS_ENABLED(CONFIG_SPARSEMEM)) |
| 4784 | pages = present_pages; |
| 4785 | |
| 4786 | return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT; |
| 4787 | } |
| 4788 | |
| 4789 | /* |
| 4790 | * Set up the zone data structures: |
| 4791 | * - mark all pages reserved |
| 4792 | * - mark all memory queues empty |
| 4793 | * - clear the memory bitmaps |
| 4794 | * |
| 4795 | * NOTE: pgdat should get zeroed by caller. |
| 4796 | */ |
| 4797 | static void __paginginit free_area_init_core(struct pglist_data *pgdat, |
| 4798 | unsigned long node_start_pfn, unsigned long node_end_pfn, |
| 4799 | unsigned long *zones_size, unsigned long *zholes_size) |
| 4800 | { |
| 4801 | enum zone_type j; |
| 4802 | int nid = pgdat->node_id; |
| 4803 | unsigned long zone_start_pfn = pgdat->node_start_pfn; |
| 4804 | int ret; |
| 4805 | |
| 4806 | pgdat_resize_init(pgdat); |
| 4807 | #ifdef CONFIG_NUMA_BALANCING |
| 4808 | spin_lock_init(&pgdat->numabalancing_migrate_lock); |
| 4809 | pgdat->numabalancing_migrate_nr_pages = 0; |
| 4810 | pgdat->numabalancing_migrate_next_window = jiffies; |
| 4811 | #endif |
| 4812 | init_waitqueue_head(&pgdat->kswapd_wait); |
| 4813 | init_waitqueue_head(&pgdat->pfmemalloc_wait); |
| 4814 | pgdat_page_cgroup_init(pgdat); |
| 4815 | |
| 4816 | for (j = 0; j < MAX_NR_ZONES; j++) { |
| 4817 | struct zone *zone = pgdat->node_zones + j; |
| 4818 | unsigned long size, realsize, freesize, memmap_pages; |
| 4819 | |
| 4820 | size = zone_spanned_pages_in_node(nid, j, node_start_pfn, |
| 4821 | node_end_pfn, zones_size); |
| 4822 | realsize = freesize = size - zone_absent_pages_in_node(nid, j, |
| 4823 | node_start_pfn, |
| 4824 | node_end_pfn, |
| 4825 | zholes_size); |
| 4826 | |
| 4827 | /* |
| 4828 | * Adjust freesize so that it accounts for how much memory |
| 4829 | * is used by this zone for memmap. This affects the watermark |
| 4830 | * and per-cpu initialisations |
| 4831 | */ |
| 4832 | memmap_pages = calc_memmap_size(size, realsize); |
| 4833 | if (freesize >= memmap_pages) { |
| 4834 | freesize -= memmap_pages; |
| 4835 | if (memmap_pages) |
| 4836 | printk(KERN_DEBUG |
| 4837 | " %s zone: %lu pages used for memmap\n", |
| 4838 | zone_names[j], memmap_pages); |
| 4839 | } else |
| 4840 | printk(KERN_WARNING |
| 4841 | " %s zone: %lu pages exceeds freesize %lu\n", |
| 4842 | zone_names[j], memmap_pages, freesize); |
| 4843 | |
| 4844 | /* Account for reserved pages */ |
| 4845 | if (j == 0 && freesize > dma_reserve) { |
| 4846 | freesize -= dma_reserve; |
| 4847 | printk(KERN_DEBUG " %s zone: %lu pages reserved\n", |
| 4848 | zone_names[0], dma_reserve); |
| 4849 | } |
| 4850 | |
| 4851 | if (!is_highmem_idx(j)) |
| 4852 | nr_kernel_pages += freesize; |
| 4853 | /* Charge for highmem memmap if there are enough kernel pages */ |
| 4854 | else if (nr_kernel_pages > memmap_pages * 2) |
| 4855 | nr_kernel_pages -= memmap_pages; |
| 4856 | nr_all_pages += freesize; |
| 4857 | |
| 4858 | zone->spanned_pages = size; |
| 4859 | zone->present_pages = realsize; |
| 4860 | /* |
| 4861 | * Set an approximate value for lowmem here, it will be adjusted |
| 4862 | * when the bootmem allocator frees pages into the buddy system. |
| 4863 | * And all highmem pages will be managed by the buddy system. |
| 4864 | */ |
| 4865 | zone->managed_pages = is_highmem_idx(j) ? realsize : freesize; |
| 4866 | #ifdef CONFIG_NUMA |
| 4867 | zone->node = nid; |
| 4868 | zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio) |
| 4869 | / 100; |
| 4870 | zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100; |
| 4871 | #endif |
| 4872 | zone->name = zone_names[j]; |
| 4873 | spin_lock_init(&zone->lock); |
| 4874 | spin_lock_init(&zone->lru_lock); |
| 4875 | zone_seqlock_init(zone); |
| 4876 | zone->zone_pgdat = pgdat; |
| 4877 | zone_pcp_init(zone); |
| 4878 | |
| 4879 | /* For bootup, initialized properly in watermark setup */ |
| 4880 | mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages); |
| 4881 | |
| 4882 | lruvec_init(&zone->lruvec); |
| 4883 | if (!size) |
| 4884 | continue; |
| 4885 | |
| 4886 | set_pageblock_order(); |
| 4887 | setup_usemap(pgdat, zone, zone_start_pfn, size); |
| 4888 | ret = init_currently_empty_zone(zone, zone_start_pfn, |
| 4889 | size, MEMMAP_EARLY); |
| 4890 | BUG_ON(ret); |
| 4891 | memmap_init(size, nid, j, zone_start_pfn); |
| 4892 | zone_start_pfn += size; |
| 4893 | } |
| 4894 | } |
| 4895 | |
| 4896 | static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) |
| 4897 | { |
| 4898 | /* Skip empty nodes */ |
| 4899 | if (!pgdat->node_spanned_pages) |
| 4900 | return; |
| 4901 | |
| 4902 | #ifdef CONFIG_FLAT_NODE_MEM_MAP |
| 4903 | /* ia64 gets its own node_mem_map, before this, without bootmem */ |
| 4904 | if (!pgdat->node_mem_map) { |
| 4905 | unsigned long size, start, end; |
| 4906 | struct page *map; |
| 4907 | |
| 4908 | /* |
| 4909 | * The zone's endpoints aren't required to be MAX_ORDER |
| 4910 | * aligned but the node_mem_map endpoints must be in order |
| 4911 | * for the buddy allocator to function correctly. |
| 4912 | */ |
| 4913 | start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); |
| 4914 | end = pgdat_end_pfn(pgdat); |
| 4915 | end = ALIGN(end, MAX_ORDER_NR_PAGES); |
| 4916 | size = (end - start) * sizeof(struct page); |
| 4917 | map = alloc_remap(pgdat->node_id, size); |
| 4918 | if (!map) |
| 4919 | map = memblock_virt_alloc_node_nopanic(size, |
| 4920 | pgdat->node_id); |
| 4921 | pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); |
| 4922 | } |
| 4923 | #ifndef CONFIG_NEED_MULTIPLE_NODES |
| 4924 | /* |
| 4925 | * With no DISCONTIG, the global mem_map is just set as node 0's |
| 4926 | */ |
| 4927 | if (pgdat == NODE_DATA(0)) { |
| 4928 | mem_map = NODE_DATA(0)->node_mem_map; |
| 4929 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP |
| 4930 | if (page_to_pfn(mem_map) != pgdat->node_start_pfn) |
| 4931 | mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); |
| 4932 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ |
| 4933 | } |
| 4934 | #endif |
| 4935 | #endif /* CONFIG_FLAT_NODE_MEM_MAP */ |
| 4936 | } |
| 4937 | |
| 4938 | void __paginginit free_area_init_node(int nid, unsigned long *zones_size, |
| 4939 | unsigned long node_start_pfn, unsigned long *zholes_size) |
| 4940 | { |
| 4941 | pg_data_t *pgdat = NODE_DATA(nid); |
| 4942 | unsigned long start_pfn = 0; |
| 4943 | unsigned long end_pfn = 0; |
| 4944 | |
| 4945 | /* pg_data_t should be reset to zero when it's allocated */ |
| 4946 | WARN_ON(pgdat->nr_zones || pgdat->classzone_idx); |
| 4947 | |
| 4948 | pgdat->node_id = nid; |
| 4949 | pgdat->node_start_pfn = node_start_pfn; |
| 4950 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP |
| 4951 | get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); |
| 4952 | #endif |
| 4953 | calculate_node_totalpages(pgdat, start_pfn, end_pfn, |
| 4954 | zones_size, zholes_size); |
| 4955 | |
| 4956 | alloc_node_mem_map(pgdat); |
| 4957 | #ifdef CONFIG_FLAT_NODE_MEM_MAP |
| 4958 | printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", |
| 4959 | nid, (unsigned long)pgdat, |
| 4960 | (unsigned long)pgdat->node_mem_map); |
| 4961 | #endif |
| 4962 | |
| 4963 | free_area_init_core(pgdat, start_pfn, end_pfn, |
| 4964 | zones_size, zholes_size); |
| 4965 | } |
| 4966 | |
| 4967 | #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP |
| 4968 | |
| 4969 | #if MAX_NUMNODES > 1 |
| 4970 | /* |
| 4971 | * Figure out the number of possible node ids. |
| 4972 | */ |
| 4973 | void __init setup_nr_node_ids(void) |
| 4974 | { |
| 4975 | unsigned int node; |
| 4976 | unsigned int highest = 0; |
| 4977 | |
| 4978 | for_each_node_mask(node, node_possible_map) |
| 4979 | highest = node; |
| 4980 | nr_node_ids = highest + 1; |
| 4981 | } |
| 4982 | #endif |
| 4983 | |
| 4984 | /** |
| 4985 | * node_map_pfn_alignment - determine the maximum internode alignment |
| 4986 | * |
| 4987 | * This function should be called after node map is populated and sorted. |
| 4988 | * It calculates the maximum power of two alignment which can distinguish |
| 4989 | * all the nodes. |
| 4990 | * |
| 4991 | * For example, if all nodes are 1GiB and aligned to 1GiB, the return value |
| 4992 | * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the |
| 4993 | * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is |
| 4994 | * shifted, 1GiB is enough and this function will indicate so. |
| 4995 | * |
| 4996 | * This is used to test whether pfn -> nid mapping of the chosen memory |
| 4997 | * model has fine enough granularity to avoid incorrect mapping for the |
| 4998 | * populated node map. |
| 4999 | * |
| 5000 | * Returns the determined alignment in pfn's. 0 if there is no alignment |
| 5001 | * requirement (single node). |
| 5002 | */ |
| 5003 | unsigned long __init node_map_pfn_alignment(void) |
| 5004 | { |
| 5005 | unsigned long accl_mask = 0, last_end = 0; |
| 5006 | unsigned long start, end, mask; |
| 5007 | int last_nid = -1; |
| 5008 | int i, nid; |
| 5009 | |
| 5010 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) { |
| 5011 | if (!start || last_nid < 0 || last_nid == nid) { |
| 5012 | last_nid = nid; |
| 5013 | last_end = end; |
| 5014 | continue; |
| 5015 | } |
| 5016 | |
| 5017 | /* |
| 5018 | * Start with a mask granular enough to pin-point to the |
| 5019 | * start pfn and tick off bits one-by-one until it becomes |
| 5020 | * too coarse to separate the current node from the last. |
| 5021 | */ |
| 5022 | mask = ~((1 << __ffs(start)) - 1); |
| 5023 | while (mask && last_end <= (start & (mask << 1))) |
| 5024 | mask <<= 1; |
| 5025 | |
| 5026 | /* accumulate all internode masks */ |
| 5027 | accl_mask |= mask; |
| 5028 | } |
| 5029 | |
| 5030 | /* convert mask to number of pages */ |
| 5031 | return ~accl_mask + 1; |
| 5032 | } |
| 5033 | |
| 5034 | /* Find the lowest pfn for a node */ |
| 5035 | static unsigned long __init find_min_pfn_for_node(int nid) |
| 5036 | { |
| 5037 | unsigned long min_pfn = ULONG_MAX; |
| 5038 | unsigned long start_pfn; |
| 5039 | int i; |
| 5040 | |
| 5041 | for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL) |
| 5042 | min_pfn = min(min_pfn, start_pfn); |
| 5043 | |
| 5044 | if (min_pfn == ULONG_MAX) { |
| 5045 | printk(KERN_WARNING |
| 5046 | "Could not find start_pfn for node %d\n", nid); |
| 5047 | return 0; |
| 5048 | } |
| 5049 | |
| 5050 | return min_pfn; |
| 5051 | } |
| 5052 | |
| 5053 | /** |
| 5054 | * find_min_pfn_with_active_regions - Find the minimum PFN registered |
| 5055 | * |
| 5056 | * It returns the minimum PFN based on information provided via |
| 5057 | * add_active_range(). |
| 5058 | */ |
| 5059 | unsigned long __init find_min_pfn_with_active_regions(void) |
| 5060 | { |
| 5061 | return find_min_pfn_for_node(MAX_NUMNODES); |
| 5062 | } |
| 5063 | |
| 5064 | /* |
| 5065 | * early_calculate_totalpages() |
| 5066 | * Sum pages in active regions for movable zone. |
| 5067 | * Populate N_MEMORY for calculating usable_nodes. |
| 5068 | */ |
| 5069 | static unsigned long __init early_calculate_totalpages(void) |
| 5070 | { |
| 5071 | unsigned long totalpages = 0; |
| 5072 | unsigned long start_pfn, end_pfn; |
| 5073 | int i, nid; |
| 5074 | |
| 5075 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { |
| 5076 | unsigned long pages = end_pfn - start_pfn; |
| 5077 | |
| 5078 | totalpages += pages; |
| 5079 | if (pages) |
| 5080 | node_set_state(nid, N_MEMORY); |
| 5081 | } |
| 5082 | return totalpages; |
| 5083 | } |
| 5084 | |
| 5085 | /* |
| 5086 | * Find the PFN the Movable zone begins in each node. Kernel memory |
| 5087 | * is spread evenly between nodes as long as the nodes have enough |
| 5088 | * memory. When they don't, some nodes will have more kernelcore than |
| 5089 | * others |
| 5090 | */ |
| 5091 | static void __init find_zone_movable_pfns_for_nodes(void) |
| 5092 | { |
| 5093 | int i, nid; |
| 5094 | unsigned long usable_startpfn; |
| 5095 | unsigned long kernelcore_node, kernelcore_remaining; |
| 5096 | /* save the state before borrow the nodemask */ |
| 5097 | nodemask_t saved_node_state = node_states[N_MEMORY]; |
| 5098 | unsigned long totalpages = early_calculate_totalpages(); |
| 5099 | int usable_nodes = nodes_weight(node_states[N_MEMORY]); |
| 5100 | struct memblock_region *r; |
| 5101 | |
| 5102 | /* Need to find movable_zone earlier when movable_node is specified. */ |
| 5103 | find_usable_zone_for_movable(); |
| 5104 | |
| 5105 | /* |
| 5106 | * If movable_node is specified, ignore kernelcore and movablecore |
| 5107 | * options. |
| 5108 | */ |
| 5109 | if (movable_node_is_enabled()) { |
| 5110 | for_each_memblock(memory, r) { |
| 5111 | if (!memblock_is_hotpluggable(r)) |
| 5112 | continue; |
| 5113 | |
| 5114 | nid = r->nid; |
| 5115 | |
| 5116 | usable_startpfn = PFN_DOWN(r->base); |
| 5117 | zone_movable_pfn[nid] = zone_movable_pfn[nid] ? |
| 5118 | min(usable_startpfn, zone_movable_pfn[nid]) : |
| 5119 | usable_startpfn; |
| 5120 | } |
| 5121 | |
| 5122 | goto out2; |
| 5123 | } |
| 5124 | |
| 5125 | /* |
| 5126 | * If movablecore=nn[KMG] was specified, calculate what size of |
| 5127 | * kernelcore that corresponds so that memory usable for |
| 5128 | * any allocation type is evenly spread. If both kernelcore |
| 5129 | * and movablecore are specified, then the value of kernelcore |
| 5130 | * will be used for required_kernelcore if it's greater than |
| 5131 | * what movablecore would have allowed. |
| 5132 | */ |
| 5133 | if (required_movablecore) { |
| 5134 | unsigned long corepages; |
| 5135 | |
| 5136 | /* |
| 5137 | * Round-up so that ZONE_MOVABLE is at least as large as what |
| 5138 | * was requested by the user |
| 5139 | */ |
| 5140 | required_movablecore = |
| 5141 | roundup(required_movablecore, MAX_ORDER_NR_PAGES); |
| 5142 | corepages = totalpages - required_movablecore; |
| 5143 | |
| 5144 | required_kernelcore = max(required_kernelcore, corepages); |
| 5145 | } |
| 5146 | |
| 5147 | /* If kernelcore was not specified, there is no ZONE_MOVABLE */ |
| 5148 | if (!required_kernelcore) |
| 5149 | goto out; |
| 5150 | |
| 5151 | /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ |
| 5152 | usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; |
| 5153 | |
| 5154 | restart: |
| 5155 | /* Spread kernelcore memory as evenly as possible throughout nodes */ |
| 5156 | kernelcore_node = required_kernelcore / usable_nodes; |
| 5157 | for_each_node_state(nid, N_MEMORY) { |
| 5158 | unsigned long start_pfn, end_pfn; |
| 5159 | |
| 5160 | /* |
| 5161 | * Recalculate kernelcore_node if the division per node |
| 5162 | * now exceeds what is necessary to satisfy the requested |
| 5163 | * amount of memory for the kernel |
| 5164 | */ |
| 5165 | if (required_kernelcore < kernelcore_node) |
| 5166 | kernelcore_node = required_kernelcore / usable_nodes; |
| 5167 | |
| 5168 | /* |
| 5169 | * As the map is walked, we track how much memory is usable |
| 5170 | * by the kernel using kernelcore_remaining. When it is |
| 5171 | * 0, the rest of the node is usable by ZONE_MOVABLE |
| 5172 | */ |
| 5173 | kernelcore_remaining = kernelcore_node; |
| 5174 | |
| 5175 | /* Go through each range of PFNs within this node */ |
| 5176 | for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { |
| 5177 | unsigned long size_pages; |
| 5178 | |
| 5179 | start_pfn = max(start_pfn, zone_movable_pfn[nid]); |
| 5180 | if (start_pfn >= end_pfn) |
| 5181 | continue; |
| 5182 | |
| 5183 | /* Account for what is only usable for kernelcore */ |
| 5184 | if (start_pfn < usable_startpfn) { |
| 5185 | unsigned long kernel_pages; |
| 5186 | kernel_pages = min(end_pfn, usable_startpfn) |
| 5187 | - start_pfn; |
| 5188 | |
| 5189 | kernelcore_remaining -= min(kernel_pages, |
| 5190 | kernelcore_remaining); |
| 5191 | required_kernelcore -= min(kernel_pages, |
| 5192 | required_kernelcore); |
| 5193 | |
| 5194 | /* Continue if range is now fully accounted */ |
| 5195 | if (end_pfn <= usable_startpfn) { |
| 5196 | |
| 5197 | /* |
| 5198 | * Push zone_movable_pfn to the end so |
| 5199 | * that if we have to rebalance |
| 5200 | * kernelcore across nodes, we will |
| 5201 | * not double account here |
| 5202 | */ |
| 5203 | zone_movable_pfn[nid] = end_pfn; |
| 5204 | continue; |
| 5205 | } |
| 5206 | start_pfn = usable_startpfn; |
| 5207 | } |
| 5208 | |
| 5209 | /* |
| 5210 | * The usable PFN range for ZONE_MOVABLE is from |
| 5211 | * start_pfn->end_pfn. Calculate size_pages as the |
| 5212 | * number of pages used as kernelcore |
| 5213 | */ |
| 5214 | size_pages = end_pfn - start_pfn; |
| 5215 | if (size_pages > kernelcore_remaining) |
| 5216 | size_pages = kernelcore_remaining; |
| 5217 | zone_movable_pfn[nid] = start_pfn + size_pages; |
| 5218 | |
| 5219 | /* |
| 5220 | * Some kernelcore has been met, update counts and |
| 5221 | * break if the kernelcore for this node has been |
| 5222 | * satisfied |
| 5223 | */ |
| 5224 | required_kernelcore -= min(required_kernelcore, |
| 5225 | size_pages); |
| 5226 | kernelcore_remaining -= size_pages; |
| 5227 | if (!kernelcore_remaining) |
| 5228 | break; |
| 5229 | } |
| 5230 | } |
| 5231 | |
| 5232 | /* |
| 5233 | * If there is still required_kernelcore, we do another pass with one |
| 5234 | * less node in the count. This will push zone_movable_pfn[nid] further |
| 5235 | * along on the nodes that still have memory until kernelcore is |
| 5236 | * satisfied |
| 5237 | */ |
| 5238 | usable_nodes--; |
| 5239 | if (usable_nodes && required_kernelcore > usable_nodes) |
| 5240 | goto restart; |
| 5241 | |
| 5242 | out2: |
| 5243 | /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ |
| 5244 | for (nid = 0; nid < MAX_NUMNODES; nid++) |
| 5245 | zone_movable_pfn[nid] = |
| 5246 | roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); |
| 5247 | |
| 5248 | out: |
| 5249 | /* restore the node_state */ |
| 5250 | node_states[N_MEMORY] = saved_node_state; |
| 5251 | } |
| 5252 | |
| 5253 | /* Any regular or high memory on that node ? */ |
| 5254 | static void check_for_memory(pg_data_t *pgdat, int nid) |
| 5255 | { |
| 5256 | enum zone_type zone_type; |
| 5257 | |
| 5258 | if (N_MEMORY == N_NORMAL_MEMORY) |
| 5259 | return; |
| 5260 | |
| 5261 | for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) { |
| 5262 | struct zone *zone = &pgdat->node_zones[zone_type]; |
| 5263 | if (populated_zone(zone)) { |
| 5264 | node_set_state(nid, N_HIGH_MEMORY); |
| 5265 | if (N_NORMAL_MEMORY != N_HIGH_MEMORY && |
| 5266 | zone_type <= ZONE_NORMAL) |
| 5267 | node_set_state(nid, N_NORMAL_MEMORY); |
| 5268 | break; |
| 5269 | } |
| 5270 | } |
| 5271 | } |
| 5272 | |
| 5273 | /** |
| 5274 | * free_area_init_nodes - Initialise all pg_data_t and zone data |
| 5275 | * @max_zone_pfn: an array of max PFNs for each zone |
| 5276 | * |
| 5277 | * This will call free_area_init_node() for each active node in the system. |
| 5278 | * Using the page ranges provided by add_active_range(), the size of each |
| 5279 | * zone in each node and their holes is calculated. If the maximum PFN |
| 5280 | * between two adjacent zones match, it is assumed that the zone is empty. |
| 5281 | * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed |
| 5282 | * that arch_max_dma32_pfn has no pages. It is also assumed that a zone |
| 5283 | * starts where the previous one ended. For example, ZONE_DMA32 starts |
| 5284 | * at arch_max_dma_pfn. |
| 5285 | */ |
| 5286 | void __init free_area_init_nodes(unsigned long *max_zone_pfn) |
| 5287 | { |
| 5288 | unsigned long start_pfn, end_pfn; |
| 5289 | int i, nid; |
| 5290 | |
| 5291 | /* Record where the zone boundaries are */ |
| 5292 | memset(arch_zone_lowest_possible_pfn, 0, |
| 5293 | sizeof(arch_zone_lowest_possible_pfn)); |
| 5294 | memset(arch_zone_highest_possible_pfn, 0, |
| 5295 | sizeof(arch_zone_highest_possible_pfn)); |
| 5296 | arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); |
| 5297 | arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; |
| 5298 | for (i = 1; i < MAX_NR_ZONES; i++) { |
| 5299 | if (i == ZONE_MOVABLE) |
| 5300 | continue; |
| 5301 | arch_zone_lowest_possible_pfn[i] = |
| 5302 | arch_zone_highest_possible_pfn[i-1]; |
| 5303 | arch_zone_highest_possible_pfn[i] = |
| 5304 | max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); |
| 5305 | } |
| 5306 | arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; |
| 5307 | arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; |
| 5308 | |
| 5309 | /* Find the PFNs that ZONE_MOVABLE begins at in each node */ |
| 5310 | memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); |
| 5311 | find_zone_movable_pfns_for_nodes(); |
| 5312 | |
| 5313 | /* Print out the zone ranges */ |
| 5314 | printk("Zone ranges:\n"); |
| 5315 | for (i = 0; i < MAX_NR_ZONES; i++) { |
| 5316 | if (i == ZONE_MOVABLE) |
| 5317 | continue; |
| 5318 | printk(KERN_CONT " %-8s ", zone_names[i]); |
| 5319 | if (arch_zone_lowest_possible_pfn[i] == |
| 5320 | arch_zone_highest_possible_pfn[i]) |
| 5321 | printk(KERN_CONT "empty\n"); |
| 5322 | else |
| 5323 | printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n", |
| 5324 | arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT, |
| 5325 | (arch_zone_highest_possible_pfn[i] |
| 5326 | << PAGE_SHIFT) - 1); |
| 5327 | } |
| 5328 | |
| 5329 | /* Print out the PFNs ZONE_MOVABLE begins at in each node */ |
| 5330 | printk("Movable zone start for each node\n"); |
| 5331 | for (i = 0; i < MAX_NUMNODES; i++) { |
| 5332 | if (zone_movable_pfn[i]) |
| 5333 | printk(" Node %d: %#010lx\n", i, |
| 5334 | zone_movable_pfn[i] << PAGE_SHIFT); |
| 5335 | } |
| 5336 | |
| 5337 | /* Print out the early node map */ |
| 5338 | printk("Early memory node ranges\n"); |
| 5339 | for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) |
| 5340 | printk(" node %3d: [mem %#010lx-%#010lx]\n", nid, |
| 5341 | start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1); |
| 5342 | |
| 5343 | /* Initialise every node */ |
| 5344 | mminit_verify_pageflags_layout(); |
| 5345 | setup_nr_node_ids(); |
| 5346 | for_each_online_node(nid) { |
| 5347 | pg_data_t *pgdat = NODE_DATA(nid); |
| 5348 | free_area_init_node(nid, NULL, |
| 5349 | find_min_pfn_for_node(nid), NULL); |
| 5350 | |
| 5351 | /* Any memory on that node */ |
| 5352 | if (pgdat->node_present_pages) |
| 5353 | node_set_state(nid, N_MEMORY); |
| 5354 | check_for_memory(pgdat, nid); |
| 5355 | } |
| 5356 | } |
| 5357 | |
| 5358 | static int __init cmdline_parse_core(char *p, unsigned long *core) |
| 5359 | { |
| 5360 | unsigned long long coremem; |
| 5361 | if (!p) |
| 5362 | return -EINVAL; |
| 5363 | |
| 5364 | coremem = memparse(p, &p); |
| 5365 | *core = coremem >> PAGE_SHIFT; |
| 5366 | |
| 5367 | /* Paranoid check that UL is enough for the coremem value */ |
| 5368 | WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); |
| 5369 | |
| 5370 | return 0; |
| 5371 | } |
| 5372 | |
| 5373 | /* |
| 5374 | * kernelcore=size sets the amount of memory for use for allocations that |
| 5375 | * cannot be reclaimed or migrated. |
| 5376 | */ |
| 5377 | static int __init cmdline_parse_kernelcore(char *p) |
| 5378 | { |
| 5379 | return cmdline_parse_core(p, &required_kernelcore); |
| 5380 | } |
| 5381 | |
| 5382 | /* |
| 5383 | * movablecore=size sets the amount of memory for use for allocations that |
| 5384 | * can be reclaimed or migrated. |
| 5385 | */ |
| 5386 | static int __init cmdline_parse_movablecore(char *p) |
| 5387 | { |
| 5388 | return cmdline_parse_core(p, &required_movablecore); |
| 5389 | } |
| 5390 | |
| 5391 | early_param("kernelcore", cmdline_parse_kernelcore); |
| 5392 | early_param("movablecore", cmdline_parse_movablecore); |
| 5393 | |
| 5394 | #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ |
| 5395 | |
| 5396 | void adjust_managed_page_count(struct page *page, long count) |
| 5397 | { |
| 5398 | spin_lock(&managed_page_count_lock); |
| 5399 | page_zone(page)->managed_pages += count; |
| 5400 | totalram_pages += count; |
| 5401 | #ifdef CONFIG_HIGHMEM |
| 5402 | if (PageHighMem(page)) |
| 5403 | totalhigh_pages += count; |
| 5404 | #endif |
| 5405 | spin_unlock(&managed_page_count_lock); |
| 5406 | } |
| 5407 | EXPORT_SYMBOL(adjust_managed_page_count); |
| 5408 | |
| 5409 | unsigned long free_reserved_area(void *start, void *end, int poison, char *s) |
| 5410 | { |
| 5411 | void *pos; |
| 5412 | unsigned long pages = 0; |
| 5413 | |
| 5414 | start = (void *)PAGE_ALIGN((unsigned long)start); |
| 5415 | end = (void *)((unsigned long)end & PAGE_MASK); |
| 5416 | for (pos = start; pos < end; pos += PAGE_SIZE, pages++) { |
| 5417 | if ((unsigned int)poison <= 0xFF) |
| 5418 | memset(pos, poison, PAGE_SIZE); |
| 5419 | free_reserved_page(virt_to_page(pos)); |
| 5420 | } |
| 5421 | |
| 5422 | if (pages && s) |
| 5423 | pr_info("Freeing %s memory: %ldK (%p - %p)\n", |
| 5424 | s, pages << (PAGE_SHIFT - 10), start, end); |
| 5425 | |
| 5426 | return pages; |
| 5427 | } |
| 5428 | EXPORT_SYMBOL(free_reserved_area); |
| 5429 | |
| 5430 | #ifdef CONFIG_HIGHMEM |
| 5431 | void free_highmem_page(struct page *page) |
| 5432 | { |
| 5433 | __free_reserved_page(page); |
| 5434 | totalram_pages++; |
| 5435 | page_zone(page)->managed_pages++; |
| 5436 | totalhigh_pages++; |
| 5437 | } |
| 5438 | #endif |
| 5439 | |
| 5440 | |
| 5441 | void __init mem_init_print_info(const char *str) |
| 5442 | { |
| 5443 | unsigned long physpages, codesize, datasize, rosize, bss_size; |
| 5444 | unsigned long init_code_size, init_data_size; |
| 5445 | |
| 5446 | physpages = get_num_physpages(); |
| 5447 | codesize = _etext - _stext; |
| 5448 | datasize = _edata - _sdata; |
| 5449 | rosize = __end_rodata - __start_rodata; |
| 5450 | bss_size = __bss_stop - __bss_start; |
| 5451 | init_data_size = __init_end - __init_begin; |
| 5452 | init_code_size = _einittext - _sinittext; |
| 5453 | |
| 5454 | /* |
| 5455 | * Detect special cases and adjust section sizes accordingly: |
| 5456 | * 1) .init.* may be embedded into .data sections |
| 5457 | * 2) .init.text.* may be out of [__init_begin, __init_end], |
| 5458 | * please refer to arch/tile/kernel/vmlinux.lds.S. |
| 5459 | * 3) .rodata.* may be embedded into .text or .data sections. |
| 5460 | */ |
| 5461 | #define adj_init_size(start, end, size, pos, adj) \ |
| 5462 | do { \ |
| 5463 | if (start <= pos && pos < end && size > adj) \ |
| 5464 | size -= adj; \ |
| 5465 | } while (0) |
| 5466 | |
| 5467 | adj_init_size(__init_begin, __init_end, init_data_size, |
| 5468 | _sinittext, init_code_size); |
| 5469 | adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size); |
| 5470 | adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size); |
| 5471 | adj_init_size(_stext, _etext, codesize, __start_rodata, rosize); |
| 5472 | adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize); |
| 5473 | |
| 5474 | #undef adj_init_size |
| 5475 | |
| 5476 | printk("Memory: %luK/%luK available " |
| 5477 | "(%luK kernel code, %luK rwdata, %luK rodata, " |
| 5478 | "%luK init, %luK bss, %luK reserved" |
| 5479 | #ifdef CONFIG_HIGHMEM |
| 5480 | ", %luK highmem" |
| 5481 | #endif |
| 5482 | "%s%s)\n", |
| 5483 | nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10), |
| 5484 | codesize >> 10, datasize >> 10, rosize >> 10, |
| 5485 | (init_data_size + init_code_size) >> 10, bss_size >> 10, |
| 5486 | (physpages - totalram_pages) << (PAGE_SHIFT-10), |
| 5487 | #ifdef CONFIG_HIGHMEM |
| 5488 | totalhigh_pages << (PAGE_SHIFT-10), |
| 5489 | #endif |
| 5490 | str ? ", " : "", str ? str : ""); |
| 5491 | } |
| 5492 | |
| 5493 | /** |
| 5494 | * set_dma_reserve - set the specified number of pages reserved in the first zone |
| 5495 | * @new_dma_reserve: The number of pages to mark reserved |
| 5496 | * |
| 5497 | * The per-cpu batchsize and zone watermarks are determined by present_pages. |
| 5498 | * In the DMA zone, a significant percentage may be consumed by kernel image |
| 5499 | * and other unfreeable allocations which can skew the watermarks badly. This |
| 5500 | * function may optionally be used to account for unfreeable pages in the |
| 5501 | * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and |
| 5502 | * smaller per-cpu batchsize. |
| 5503 | */ |
| 5504 | void __init set_dma_reserve(unsigned long new_dma_reserve) |
| 5505 | { |
| 5506 | dma_reserve = new_dma_reserve; |
| 5507 | } |
| 5508 | |
| 5509 | void __init free_area_init(unsigned long *zones_size) |
| 5510 | { |
| 5511 | free_area_init_node(0, zones_size, |
| 5512 | __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); |
| 5513 | } |
| 5514 | |
| 5515 | static int page_alloc_cpu_notify(struct notifier_block *self, |
| 5516 | unsigned long action, void *hcpu) |
| 5517 | { |
| 5518 | int cpu = (unsigned long)hcpu; |
| 5519 | |
| 5520 | if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { |
| 5521 | lru_add_drain_cpu(cpu); |
| 5522 | drain_pages(cpu); |
| 5523 | |
| 5524 | /* |
| 5525 | * Spill the event counters of the dead processor |
| 5526 | * into the current processors event counters. |
| 5527 | * This artificially elevates the count of the current |
| 5528 | * processor. |
| 5529 | */ |
| 5530 | vm_events_fold_cpu(cpu); |
| 5531 | |
| 5532 | /* |
| 5533 | * Zero the differential counters of the dead processor |
| 5534 | * so that the vm statistics are consistent. |
| 5535 | * |
| 5536 | * This is only okay since the processor is dead and cannot |
| 5537 | * race with what we are doing. |
| 5538 | */ |
| 5539 | cpu_vm_stats_fold(cpu); |
| 5540 | } |
| 5541 | return NOTIFY_OK; |
| 5542 | } |
| 5543 | |
| 5544 | void __init page_alloc_init(void) |
| 5545 | { |
| 5546 | hotcpu_notifier(page_alloc_cpu_notify, 0); |
| 5547 | } |
| 5548 | |
| 5549 | /* |
| 5550 | * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio |
| 5551 | * or min_free_kbytes changes. |
| 5552 | */ |
| 5553 | static void calculate_totalreserve_pages(void) |
| 5554 | { |
| 5555 | struct pglist_data *pgdat; |
| 5556 | unsigned long reserve_pages = 0; |
| 5557 | enum zone_type i, j; |
| 5558 | |
| 5559 | for_each_online_pgdat(pgdat) { |
| 5560 | for (i = 0; i < MAX_NR_ZONES; i++) { |
| 5561 | struct zone *zone = pgdat->node_zones + i; |
| 5562 | unsigned long max = 0; |
| 5563 | |
| 5564 | /* Find valid and maximum lowmem_reserve in the zone */ |
| 5565 | for (j = i; j < MAX_NR_ZONES; j++) { |
| 5566 | if (zone->lowmem_reserve[j] > max) |
| 5567 | max = zone->lowmem_reserve[j]; |
| 5568 | } |
| 5569 | |
| 5570 | /* we treat the high watermark as reserved pages. */ |
| 5571 | max += high_wmark_pages(zone); |
| 5572 | |
| 5573 | if (max > zone->managed_pages) |
| 5574 | max = zone->managed_pages; |
| 5575 | reserve_pages += max; |
| 5576 | /* |
| 5577 | * Lowmem reserves are not available to |
| 5578 | * GFP_HIGHUSER page cache allocations and |
| 5579 | * kswapd tries to balance zones to their high |
| 5580 | * watermark. As a result, neither should be |
| 5581 | * regarded as dirtyable memory, to prevent a |
| 5582 | * situation where reclaim has to clean pages |
| 5583 | * in order to balance the zones. |
| 5584 | */ |
| 5585 | zone->dirty_balance_reserve = max; |
| 5586 | } |
| 5587 | } |
| 5588 | dirty_balance_reserve = reserve_pages; |
| 5589 | totalreserve_pages = reserve_pages; |
| 5590 | } |
| 5591 | |
| 5592 | /* |
| 5593 | * setup_per_zone_lowmem_reserve - called whenever |
| 5594 | * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone |
| 5595 | * has a correct pages reserved value, so an adequate number of |
| 5596 | * pages are left in the zone after a successful __alloc_pages(). |
| 5597 | */ |
| 5598 | static void setup_per_zone_lowmem_reserve(void) |
| 5599 | { |
| 5600 | struct pglist_data *pgdat; |
| 5601 | enum zone_type j, idx; |
| 5602 | |
| 5603 | for_each_online_pgdat(pgdat) { |
| 5604 | for (j = 0; j < MAX_NR_ZONES; j++) { |
| 5605 | struct zone *zone = pgdat->node_zones + j; |
| 5606 | unsigned long managed_pages = zone->managed_pages; |
| 5607 | |
| 5608 | zone->lowmem_reserve[j] = 0; |
| 5609 | |
| 5610 | idx = j; |
| 5611 | while (idx) { |
| 5612 | struct zone *lower_zone; |
| 5613 | |
| 5614 | idx--; |
| 5615 | |
| 5616 | if (sysctl_lowmem_reserve_ratio[idx] < 1) |
| 5617 | sysctl_lowmem_reserve_ratio[idx] = 1; |
| 5618 | |
| 5619 | lower_zone = pgdat->node_zones + idx; |
| 5620 | lower_zone->lowmem_reserve[j] = managed_pages / |
| 5621 | sysctl_lowmem_reserve_ratio[idx]; |
| 5622 | managed_pages += lower_zone->managed_pages; |
| 5623 | } |
| 5624 | } |
| 5625 | } |
| 5626 | |
| 5627 | /* update totalreserve_pages */ |
| 5628 | calculate_totalreserve_pages(); |
| 5629 | } |
| 5630 | |
| 5631 | static void __setup_per_zone_wmarks(void) |
| 5632 | { |
| 5633 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); |
| 5634 | unsigned long lowmem_pages = 0; |
| 5635 | struct zone *zone; |
| 5636 | unsigned long flags; |
| 5637 | |
| 5638 | /* Calculate total number of !ZONE_HIGHMEM pages */ |
| 5639 | for_each_zone(zone) { |
| 5640 | if (!is_highmem(zone)) |
| 5641 | lowmem_pages += zone->managed_pages; |
| 5642 | } |
| 5643 | |
| 5644 | for_each_zone(zone) { |
| 5645 | u64 tmp; |
| 5646 | |
| 5647 | spin_lock_irqsave(&zone->lock, flags); |
| 5648 | tmp = (u64)pages_min * zone->managed_pages; |
| 5649 | do_div(tmp, lowmem_pages); |
| 5650 | if (is_highmem(zone)) { |
| 5651 | /* |
| 5652 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't |
| 5653 | * need highmem pages, so cap pages_min to a small |
| 5654 | * value here. |
| 5655 | * |
| 5656 | * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) |
| 5657 | * deltas controls asynch page reclaim, and so should |
| 5658 | * not be capped for highmem. |
| 5659 | */ |
| 5660 | unsigned long min_pages; |
| 5661 | |
| 5662 | min_pages = zone->managed_pages / 1024; |
| 5663 | min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL); |
| 5664 | zone->watermark[WMARK_MIN] = min_pages; |
| 5665 | } else { |
| 5666 | /* |
| 5667 | * If it's a lowmem zone, reserve a number of pages |
| 5668 | * proportionate to the zone's size. |
| 5669 | */ |
| 5670 | zone->watermark[WMARK_MIN] = tmp; |
| 5671 | } |
| 5672 | |
| 5673 | zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); |
| 5674 | zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); |
| 5675 | |
| 5676 | __mod_zone_page_state(zone, NR_ALLOC_BATCH, |
| 5677 | high_wmark_pages(zone) - |
| 5678 | low_wmark_pages(zone) - |
| 5679 | zone_page_state(zone, NR_ALLOC_BATCH)); |
| 5680 | |
| 5681 | setup_zone_migrate_reserve(zone); |
| 5682 | spin_unlock_irqrestore(&zone->lock, flags); |
| 5683 | } |
| 5684 | |
| 5685 | /* update totalreserve_pages */ |
| 5686 | calculate_totalreserve_pages(); |
| 5687 | } |
| 5688 | |
| 5689 | /** |
| 5690 | * setup_per_zone_wmarks - called when min_free_kbytes changes |
| 5691 | * or when memory is hot-{added|removed} |
| 5692 | * |
| 5693 | * Ensures that the watermark[min,low,high] values for each zone are set |
| 5694 | * correctly with respect to min_free_kbytes. |
| 5695 | */ |
| 5696 | void setup_per_zone_wmarks(void) |
| 5697 | { |
| 5698 | mutex_lock(&zonelists_mutex); |
| 5699 | __setup_per_zone_wmarks(); |
| 5700 | mutex_unlock(&zonelists_mutex); |
| 5701 | } |
| 5702 | |
| 5703 | /* |
| 5704 | * The inactive anon list should be small enough that the VM never has to |
| 5705 | * do too much work, but large enough that each inactive page has a chance |
| 5706 | * to be referenced again before it is swapped out. |
| 5707 | * |
| 5708 | * The inactive_anon ratio is the target ratio of ACTIVE_ANON to |
| 5709 | * INACTIVE_ANON pages on this zone's LRU, maintained by the |
| 5710 | * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of |
| 5711 | * the anonymous pages are kept on the inactive list. |
| 5712 | * |
| 5713 | * total target max |
| 5714 | * memory ratio inactive anon |
| 5715 | * ------------------------------------- |
| 5716 | * 10MB 1 5MB |
| 5717 | * 100MB 1 50MB |
| 5718 | * 1GB 3 250MB |
| 5719 | * 10GB 10 0.9GB |
| 5720 | * 100GB 31 3GB |
| 5721 | * 1TB 101 10GB |
| 5722 | * 10TB 320 32GB |
| 5723 | */ |
| 5724 | static void __meminit calculate_zone_inactive_ratio(struct zone *zone) |
| 5725 | { |
| 5726 | unsigned int gb, ratio; |
| 5727 | |
| 5728 | /* Zone size in gigabytes */ |
| 5729 | gb = zone->managed_pages >> (30 - PAGE_SHIFT); |
| 5730 | if (gb) |
| 5731 | ratio = int_sqrt(10 * gb); |
| 5732 | else |
| 5733 | ratio = 1; |
| 5734 | |
| 5735 | zone->inactive_ratio = ratio; |
| 5736 | } |
| 5737 | |
| 5738 | static void __meminit setup_per_zone_inactive_ratio(void) |
| 5739 | { |
| 5740 | struct zone *zone; |
| 5741 | |
| 5742 | for_each_zone(zone) |
| 5743 | calculate_zone_inactive_ratio(zone); |
| 5744 | } |
| 5745 | |
| 5746 | /* |
| 5747 | * Initialise min_free_kbytes. |
| 5748 | * |
| 5749 | * For small machines we want it small (128k min). For large machines |
| 5750 | * we want it large (64MB max). But it is not linear, because network |
| 5751 | * bandwidth does not increase linearly with machine size. We use |
| 5752 | * |
| 5753 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: |
| 5754 | * min_free_kbytes = sqrt(lowmem_kbytes * 16) |
| 5755 | * |
| 5756 | * which yields |
| 5757 | * |
| 5758 | * 16MB: 512k |
| 5759 | * 32MB: 724k |
| 5760 | * 64MB: 1024k |
| 5761 | * 128MB: 1448k |
| 5762 | * 256MB: 2048k |
| 5763 | * 512MB: 2896k |
| 5764 | * 1024MB: 4096k |
| 5765 | * 2048MB: 5792k |
| 5766 | * 4096MB: 8192k |
| 5767 | * 8192MB: 11584k |
| 5768 | * 16384MB: 16384k |
| 5769 | */ |
| 5770 | int __meminit init_per_zone_wmark_min(void) |
| 5771 | { |
| 5772 | unsigned long lowmem_kbytes; |
| 5773 | int new_min_free_kbytes; |
| 5774 | |
| 5775 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); |
| 5776 | new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16); |
| 5777 | |
| 5778 | if (new_min_free_kbytes > user_min_free_kbytes) { |
| 5779 | min_free_kbytes = new_min_free_kbytes; |
| 5780 | if (min_free_kbytes < 128) |
| 5781 | min_free_kbytes = 128; |
| 5782 | if (min_free_kbytes > 65536) |
| 5783 | min_free_kbytes = 65536; |
| 5784 | } else { |
| 5785 | pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n", |
| 5786 | new_min_free_kbytes, user_min_free_kbytes); |
| 5787 | } |
| 5788 | setup_per_zone_wmarks(); |
| 5789 | refresh_zone_stat_thresholds(); |
| 5790 | setup_per_zone_lowmem_reserve(); |
| 5791 | setup_per_zone_inactive_ratio(); |
| 5792 | return 0; |
| 5793 | } |
| 5794 | module_init(init_per_zone_wmark_min) |
| 5795 | |
| 5796 | /* |
| 5797 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so |
| 5798 | * that we can call two helper functions whenever min_free_kbytes |
| 5799 | * changes. |
| 5800 | */ |
| 5801 | int min_free_kbytes_sysctl_handler(ctl_table *table, int write, |
| 5802 | void __user *buffer, size_t *length, loff_t *ppos) |
| 5803 | { |
| 5804 | int rc; |
| 5805 | |
| 5806 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); |
| 5807 | if (rc) |
| 5808 | return rc; |
| 5809 | |
| 5810 | if (write) { |
| 5811 | user_min_free_kbytes = min_free_kbytes; |
| 5812 | setup_per_zone_wmarks(); |
| 5813 | } |
| 5814 | return 0; |
| 5815 | } |
| 5816 | |
| 5817 | #ifdef CONFIG_NUMA |
| 5818 | int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, |
| 5819 | void __user *buffer, size_t *length, loff_t *ppos) |
| 5820 | { |
| 5821 | struct zone *zone; |
| 5822 | int rc; |
| 5823 | |
| 5824 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); |
| 5825 | if (rc) |
| 5826 | return rc; |
| 5827 | |
| 5828 | for_each_zone(zone) |
| 5829 | zone->min_unmapped_pages = (zone->managed_pages * |
| 5830 | sysctl_min_unmapped_ratio) / 100; |
| 5831 | return 0; |
| 5832 | } |
| 5833 | |
| 5834 | int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, |
| 5835 | void __user *buffer, size_t *length, loff_t *ppos) |
| 5836 | { |
| 5837 | struct zone *zone; |
| 5838 | int rc; |
| 5839 | |
| 5840 | rc = proc_dointvec_minmax(table, write, buffer, length, ppos); |
| 5841 | if (rc) |
| 5842 | return rc; |
| 5843 | |
| 5844 | for_each_zone(zone) |
| 5845 | zone->min_slab_pages = (zone->managed_pages * |
| 5846 | sysctl_min_slab_ratio) / 100; |
| 5847 | return 0; |
| 5848 | } |
| 5849 | #endif |
| 5850 | |
| 5851 | /* |
| 5852 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around |
| 5853 | * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() |
| 5854 | * whenever sysctl_lowmem_reserve_ratio changes. |
| 5855 | * |
| 5856 | * The reserve ratio obviously has absolutely no relation with the |
| 5857 | * minimum watermarks. The lowmem reserve ratio can only make sense |
| 5858 | * if in function of the boot time zone sizes. |
| 5859 | */ |
| 5860 | int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, |
| 5861 | void __user *buffer, size_t *length, loff_t *ppos) |
| 5862 | { |
| 5863 | proc_dointvec_minmax(table, write, buffer, length, ppos); |
| 5864 | setup_per_zone_lowmem_reserve(); |
| 5865 | return 0; |
| 5866 | } |
| 5867 | |
| 5868 | /* |
| 5869 | * percpu_pagelist_fraction - changes the pcp->high for each zone on each |
| 5870 | * cpu. It is the fraction of total pages in each zone that a hot per cpu |
| 5871 | * pagelist can have before it gets flushed back to buddy allocator. |
| 5872 | */ |
| 5873 | int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, |
| 5874 | void __user *buffer, size_t *length, loff_t *ppos) |
| 5875 | { |
| 5876 | struct zone *zone; |
| 5877 | unsigned int cpu; |
| 5878 | int ret; |
| 5879 | |
| 5880 | ret = proc_dointvec_minmax(table, write, buffer, length, ppos); |
| 5881 | if (!write || (ret < 0)) |
| 5882 | return ret; |
| 5883 | |
| 5884 | mutex_lock(&pcp_batch_high_lock); |
| 5885 | for_each_populated_zone(zone) { |
| 5886 | unsigned long high; |
| 5887 | high = zone->managed_pages / percpu_pagelist_fraction; |
| 5888 | for_each_possible_cpu(cpu) |
| 5889 | pageset_set_high(per_cpu_ptr(zone->pageset, cpu), |
| 5890 | high); |
| 5891 | } |
| 5892 | mutex_unlock(&pcp_batch_high_lock); |
| 5893 | return 0; |
| 5894 | } |
| 5895 | |
| 5896 | int hashdist = HASHDIST_DEFAULT; |
| 5897 | |
| 5898 | #ifdef CONFIG_NUMA |
| 5899 | static int __init set_hashdist(char *str) |
| 5900 | { |
| 5901 | if (!str) |
| 5902 | return 0; |
| 5903 | hashdist = simple_strtoul(str, &str, 0); |
| 5904 | return 1; |
| 5905 | } |
| 5906 | __setup("hashdist=", set_hashdist); |
| 5907 | #endif |
| 5908 | |
| 5909 | /* |
| 5910 | * allocate a large system hash table from bootmem |
| 5911 | * - it is assumed that the hash table must contain an exact power-of-2 |
| 5912 | * quantity of entries |
| 5913 | * - limit is the number of hash buckets, not the total allocation size |
| 5914 | */ |
| 5915 | void *__init alloc_large_system_hash(const char *tablename, |
| 5916 | unsigned long bucketsize, |
| 5917 | unsigned long numentries, |
| 5918 | int scale, |
| 5919 | int flags, |
| 5920 | unsigned int *_hash_shift, |
| 5921 | unsigned int *_hash_mask, |
| 5922 | unsigned long low_limit, |
| 5923 | unsigned long high_limit) |
| 5924 | { |
| 5925 | unsigned long long max = high_limit; |
| 5926 | unsigned long log2qty, size; |
| 5927 | void *table = NULL; |
| 5928 | |
| 5929 | /* allow the kernel cmdline to have a say */ |
| 5930 | if (!numentries) { |
| 5931 | /* round applicable memory size up to nearest megabyte */ |
| 5932 | numentries = nr_kernel_pages; |
| 5933 | |
| 5934 | /* It isn't necessary when PAGE_SIZE >= 1MB */ |
| 5935 | if (PAGE_SHIFT < 20) |
| 5936 | numentries = round_up(numentries, (1<<20)/PAGE_SIZE); |
| 5937 | |
| 5938 | /* limit to 1 bucket per 2^scale bytes of low memory */ |
| 5939 | if (scale > PAGE_SHIFT) |
| 5940 | numentries >>= (scale - PAGE_SHIFT); |
| 5941 | else |
| 5942 | numentries <<= (PAGE_SHIFT - scale); |
| 5943 | |
| 5944 | /* Make sure we've got at least a 0-order allocation.. */ |
| 5945 | if (unlikely(flags & HASH_SMALL)) { |
| 5946 | /* Makes no sense without HASH_EARLY */ |
| 5947 | WARN_ON(!(flags & HASH_EARLY)); |
| 5948 | if (!(numentries >> *_hash_shift)) { |
| 5949 | numentries = 1UL << *_hash_shift; |
| 5950 | BUG_ON(!numentries); |
| 5951 | } |
| 5952 | } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) |
| 5953 | numentries = PAGE_SIZE / bucketsize; |
| 5954 | } |
| 5955 | numentries = roundup_pow_of_two(numentries); |
| 5956 | |
| 5957 | /* limit allocation size to 1/16 total memory by default */ |
| 5958 | if (max == 0) { |
| 5959 | max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; |
| 5960 | do_div(max, bucketsize); |
| 5961 | } |
| 5962 | max = min(max, 0x80000000ULL); |
| 5963 | |
| 5964 | if (numentries < low_limit) |
| 5965 | numentries = low_limit; |
| 5966 | if (numentries > max) |
| 5967 | numentries = max; |
| 5968 | |
| 5969 | log2qty = ilog2(numentries); |
| 5970 | |
| 5971 | do { |
| 5972 | size = bucketsize << log2qty; |
| 5973 | if (flags & HASH_EARLY) |
| 5974 | table = memblock_virt_alloc_nopanic(size, 0); |
| 5975 | else if (hashdist) |
| 5976 | table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); |
| 5977 | else { |
| 5978 | /* |
| 5979 | * If bucketsize is not a power-of-two, we may free |
| 5980 | * some pages at the end of hash table which |
| 5981 | * alloc_pages_exact() automatically does |
| 5982 | */ |
| 5983 | if (get_order(size) < MAX_ORDER) { |
| 5984 | table = alloc_pages_exact(size, GFP_ATOMIC); |
| 5985 | kmemleak_alloc(table, size, 1, GFP_ATOMIC); |
| 5986 | } |
| 5987 | } |
| 5988 | } while (!table && size > PAGE_SIZE && --log2qty); |
| 5989 | |
| 5990 | if (!table) |
| 5991 | panic("Failed to allocate %s hash table\n", tablename); |
| 5992 | |
| 5993 | printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", |
| 5994 | tablename, |
| 5995 | (1UL << log2qty), |
| 5996 | ilog2(size) - PAGE_SHIFT, |
| 5997 | size); |
| 5998 | |
| 5999 | if (_hash_shift) |
| 6000 | *_hash_shift = log2qty; |
| 6001 | if (_hash_mask) |
| 6002 | *_hash_mask = (1 << log2qty) - 1; |
| 6003 | |
| 6004 | return table; |
| 6005 | } |
| 6006 | |
| 6007 | /* Return a pointer to the bitmap storing bits affecting a block of pages */ |
| 6008 | static inline unsigned long *get_pageblock_bitmap(struct zone *zone, |
| 6009 | unsigned long pfn) |
| 6010 | { |
| 6011 | #ifdef CONFIG_SPARSEMEM |
| 6012 | return __pfn_to_section(pfn)->pageblock_flags; |
| 6013 | #else |
| 6014 | return zone->pageblock_flags; |
| 6015 | #endif /* CONFIG_SPARSEMEM */ |
| 6016 | } |
| 6017 | |
| 6018 | static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) |
| 6019 | { |
| 6020 | #ifdef CONFIG_SPARSEMEM |
| 6021 | pfn &= (PAGES_PER_SECTION-1); |
| 6022 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; |
| 6023 | #else |
| 6024 | pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages); |
| 6025 | return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; |
| 6026 | #endif /* CONFIG_SPARSEMEM */ |
| 6027 | } |
| 6028 | |
| 6029 | /** |
| 6030 | * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages |
| 6031 | * @page: The page within the block of interest |
| 6032 | * @start_bitidx: The first bit of interest to retrieve |
| 6033 | * @end_bitidx: The last bit of interest |
| 6034 | * returns pageblock_bits flags |
| 6035 | */ |
| 6036 | unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn, |
| 6037 | unsigned long end_bitidx, |
| 6038 | unsigned long mask) |
| 6039 | { |
| 6040 | struct zone *zone; |
| 6041 | unsigned long *bitmap; |
| 6042 | unsigned long bitidx, word_bitidx; |
| 6043 | unsigned long word; |
| 6044 | |
| 6045 | zone = page_zone(page); |
| 6046 | bitmap = get_pageblock_bitmap(zone, pfn); |
| 6047 | bitidx = pfn_to_bitidx(zone, pfn); |
| 6048 | word_bitidx = bitidx / BITS_PER_LONG; |
| 6049 | bitidx &= (BITS_PER_LONG-1); |
| 6050 | |
| 6051 | word = bitmap[word_bitidx]; |
| 6052 | bitidx += end_bitidx; |
| 6053 | return (word >> (BITS_PER_LONG - bitidx - 1)) & mask; |
| 6054 | } |
| 6055 | |
| 6056 | /** |
| 6057 | * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages |
| 6058 | * @page: The page within the block of interest |
| 6059 | * @start_bitidx: The first bit of interest |
| 6060 | * @end_bitidx: The last bit of interest |
| 6061 | * @flags: The flags to set |
| 6062 | */ |
| 6063 | void set_pfnblock_flags_mask(struct page *page, unsigned long flags, |
| 6064 | unsigned long pfn, |
| 6065 | unsigned long end_bitidx, |
| 6066 | unsigned long mask) |
| 6067 | { |
| 6068 | struct zone *zone; |
| 6069 | unsigned long *bitmap; |
| 6070 | unsigned long bitidx, word_bitidx; |
| 6071 | unsigned long old_word, word; |
| 6072 | |
| 6073 | BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4); |
| 6074 | |
| 6075 | zone = page_zone(page); |
| 6076 | bitmap = get_pageblock_bitmap(zone, pfn); |
| 6077 | bitidx = pfn_to_bitidx(zone, pfn); |
| 6078 | word_bitidx = bitidx / BITS_PER_LONG; |
| 6079 | bitidx &= (BITS_PER_LONG-1); |
| 6080 | |
| 6081 | VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page); |
| 6082 | |
| 6083 | bitidx += end_bitidx; |
| 6084 | mask <<= (BITS_PER_LONG - bitidx - 1); |
| 6085 | flags <<= (BITS_PER_LONG - bitidx - 1); |
| 6086 | |
| 6087 | word = ACCESS_ONCE(bitmap[word_bitidx]); |
| 6088 | for (;;) { |
| 6089 | old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags); |
| 6090 | if (word == old_word) |
| 6091 | break; |
| 6092 | word = old_word; |
| 6093 | } |
| 6094 | } |
| 6095 | |
| 6096 | /* |
| 6097 | * This function checks whether pageblock includes unmovable pages or not. |
| 6098 | * If @count is not zero, it is okay to include less @count unmovable pages |
| 6099 | * |
| 6100 | * PageLRU check without isolation or lru_lock could race so that |
| 6101 | * MIGRATE_MOVABLE block might include unmovable pages. It means you can't |
| 6102 | * expect this function should be exact. |
| 6103 | */ |
| 6104 | bool has_unmovable_pages(struct zone *zone, struct page *page, int count, |
| 6105 | bool skip_hwpoisoned_pages) |
| 6106 | { |
| 6107 | unsigned long pfn, iter, found; |
| 6108 | int mt; |
| 6109 | |
| 6110 | /* |
| 6111 | * For avoiding noise data, lru_add_drain_all() should be called |
| 6112 | * If ZONE_MOVABLE, the zone never contains unmovable pages |
| 6113 | */ |
| 6114 | if (zone_idx(zone) == ZONE_MOVABLE) |
| 6115 | return false; |
| 6116 | mt = get_pageblock_migratetype(page); |
| 6117 | if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt)) |
| 6118 | return false; |
| 6119 | |
| 6120 | pfn = page_to_pfn(page); |
| 6121 | for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { |
| 6122 | unsigned long check = pfn + iter; |
| 6123 | |
| 6124 | if (!pfn_valid_within(check)) |
| 6125 | continue; |
| 6126 | |
| 6127 | page = pfn_to_page(check); |
| 6128 | |
| 6129 | /* |
| 6130 | * Hugepages are not in LRU lists, but they're movable. |
| 6131 | * We need not scan over tail pages bacause we don't |
| 6132 | * handle each tail page individually in migration. |
| 6133 | */ |
| 6134 | if (PageHuge(page)) { |
| 6135 | iter = round_up(iter + 1, 1<<compound_order(page)) - 1; |
| 6136 | continue; |
| 6137 | } |
| 6138 | |
| 6139 | /* |
| 6140 | * We can't use page_count without pin a page |
| 6141 | * because another CPU can free compound page. |
| 6142 | * This check already skips compound tails of THP |
| 6143 | * because their page->_count is zero at all time. |
| 6144 | */ |
| 6145 | if (!atomic_read(&page->_count)) { |
| 6146 | if (PageBuddy(page)) |
| 6147 | iter += (1 << page_order(page)) - 1; |
| 6148 | continue; |
| 6149 | } |
| 6150 | |
| 6151 | /* |
| 6152 | * The HWPoisoned page may be not in buddy system, and |
| 6153 | * page_count() is not 0. |
| 6154 | */ |
| 6155 | if (skip_hwpoisoned_pages && PageHWPoison(page)) |
| 6156 | continue; |
| 6157 | |
| 6158 | if (!PageLRU(page)) |
| 6159 | found++; |
| 6160 | /* |
| 6161 | * If there are RECLAIMABLE pages, we need to check it. |
| 6162 | * But now, memory offline itself doesn't call shrink_slab() |
| 6163 | * and it still to be fixed. |
| 6164 | */ |
| 6165 | /* |
| 6166 | * If the page is not RAM, page_count()should be 0. |
| 6167 | * we don't need more check. This is an _used_ not-movable page. |
| 6168 | * |
| 6169 | * The problematic thing here is PG_reserved pages. PG_reserved |
| 6170 | * is set to both of a memory hole page and a _used_ kernel |
| 6171 | * page at boot. |
| 6172 | */ |
| 6173 | if (found > count) |
| 6174 | return true; |
| 6175 | } |
| 6176 | return false; |
| 6177 | } |
| 6178 | |
| 6179 | bool is_pageblock_removable_nolock(struct page *page) |
| 6180 | { |
| 6181 | struct zone *zone; |
| 6182 | unsigned long pfn; |
| 6183 | |
| 6184 | /* |
| 6185 | * We have to be careful here because we are iterating over memory |
| 6186 | * sections which are not zone aware so we might end up outside of |
| 6187 | * the zone but still within the section. |
| 6188 | * We have to take care about the node as well. If the node is offline |
| 6189 | * its NODE_DATA will be NULL - see page_zone. |
| 6190 | */ |
| 6191 | if (!node_online(page_to_nid(page))) |
| 6192 | return false; |
| 6193 | |
| 6194 | zone = page_zone(page); |
| 6195 | pfn = page_to_pfn(page); |
| 6196 | if (!zone_spans_pfn(zone, pfn)) |
| 6197 | return false; |
| 6198 | |
| 6199 | return !has_unmovable_pages(zone, page, 0, true); |
| 6200 | } |
| 6201 | |
| 6202 | #ifdef CONFIG_CMA |
| 6203 | |
| 6204 | static unsigned long pfn_max_align_down(unsigned long pfn) |
| 6205 | { |
| 6206 | return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES, |
| 6207 | pageblock_nr_pages) - 1); |
| 6208 | } |
| 6209 | |
| 6210 | static unsigned long pfn_max_align_up(unsigned long pfn) |
| 6211 | { |
| 6212 | return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES, |
| 6213 | pageblock_nr_pages)); |
| 6214 | } |
| 6215 | |
| 6216 | /* [start, end) must belong to a single zone. */ |
| 6217 | static int __alloc_contig_migrate_range(struct compact_control *cc, |
| 6218 | unsigned long start, unsigned long end) |
| 6219 | { |
| 6220 | /* This function is based on compact_zone() from compaction.c. */ |
| 6221 | unsigned long nr_reclaimed; |
| 6222 | unsigned long pfn = start; |
| 6223 | unsigned int tries = 0; |
| 6224 | int ret = 0; |
| 6225 | |
| 6226 | migrate_prep(); |
| 6227 | |
| 6228 | while (pfn < end || !list_empty(&cc->migratepages)) { |
| 6229 | if (fatal_signal_pending(current)) { |
| 6230 | ret = -EINTR; |
| 6231 | break; |
| 6232 | } |
| 6233 | |
| 6234 | if (list_empty(&cc->migratepages)) { |
| 6235 | cc->nr_migratepages = 0; |
| 6236 | pfn = isolate_migratepages_range(cc->zone, cc, |
| 6237 | pfn, end, true); |
| 6238 | if (!pfn) { |
| 6239 | ret = -EINTR; |
| 6240 | break; |
| 6241 | } |
| 6242 | tries = 0; |
| 6243 | } else if (++tries == 5) { |
| 6244 | ret = ret < 0 ? ret : -EBUSY; |
| 6245 | break; |
| 6246 | } |
| 6247 | |
| 6248 | nr_reclaimed = reclaim_clean_pages_from_list(cc->zone, |
| 6249 | &cc->migratepages); |
| 6250 | cc->nr_migratepages -= nr_reclaimed; |
| 6251 | |
| 6252 | ret = migrate_pages(&cc->migratepages, alloc_migrate_target, |
| 6253 | NULL, 0, cc->mode, MR_CMA); |
| 6254 | } |
| 6255 | if (ret < 0) { |
| 6256 | putback_movable_pages(&cc->migratepages); |
| 6257 | return ret; |
| 6258 | } |
| 6259 | return 0; |
| 6260 | } |
| 6261 | |
| 6262 | /** |
| 6263 | * alloc_contig_range() -- tries to allocate given range of pages |
| 6264 | * @start: start PFN to allocate |
| 6265 | * @end: one-past-the-last PFN to allocate |
| 6266 | * @migratetype: migratetype of the underlaying pageblocks (either |
| 6267 | * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks |
| 6268 | * in range must have the same migratetype and it must |
| 6269 | * be either of the two. |
| 6270 | * |
| 6271 | * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES |
| 6272 | * aligned, however it's the caller's responsibility to guarantee that |
| 6273 | * we are the only thread that changes migrate type of pageblocks the |
| 6274 | * pages fall in. |
| 6275 | * |
| 6276 | * The PFN range must belong to a single zone. |
| 6277 | * |
| 6278 | * Returns zero on success or negative error code. On success all |
| 6279 | * pages which PFN is in [start, end) are allocated for the caller and |
| 6280 | * need to be freed with free_contig_range(). |
| 6281 | */ |
| 6282 | int alloc_contig_range(unsigned long start, unsigned long end, |
| 6283 | unsigned migratetype) |
| 6284 | { |
| 6285 | unsigned long outer_start, outer_end; |
| 6286 | int ret = 0, order; |
| 6287 | |
| 6288 | struct compact_control cc = { |
| 6289 | .nr_migratepages = 0, |
| 6290 | .order = -1, |
| 6291 | .zone = page_zone(pfn_to_page(start)), |
| 6292 | .mode = MIGRATE_SYNC, |
| 6293 | .ignore_skip_hint = true, |
| 6294 | }; |
| 6295 | INIT_LIST_HEAD(&cc.migratepages); |
| 6296 | |
| 6297 | /* |
| 6298 | * What we do here is we mark all pageblocks in range as |
| 6299 | * MIGRATE_ISOLATE. Because pageblock and max order pages may |
| 6300 | * have different sizes, and due to the way page allocator |
| 6301 | * work, we align the range to biggest of the two pages so |
| 6302 | * that page allocator won't try to merge buddies from |
| 6303 | * different pageblocks and change MIGRATE_ISOLATE to some |
| 6304 | * other migration type. |
| 6305 | * |
| 6306 | * Once the pageblocks are marked as MIGRATE_ISOLATE, we |
| 6307 | * migrate the pages from an unaligned range (ie. pages that |
| 6308 | * we are interested in). This will put all the pages in |
| 6309 | * range back to page allocator as MIGRATE_ISOLATE. |
| 6310 | * |
| 6311 | * When this is done, we take the pages in range from page |
| 6312 | * allocator removing them from the buddy system. This way |
| 6313 | * page allocator will never consider using them. |
| 6314 | * |
| 6315 | * This lets us mark the pageblocks back as |
| 6316 | * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the |
| 6317 | * aligned range but not in the unaligned, original range are |
| 6318 | * put back to page allocator so that buddy can use them. |
| 6319 | */ |
| 6320 | |
| 6321 | ret = start_isolate_page_range(pfn_max_align_down(start), |
| 6322 | pfn_max_align_up(end), migratetype, |
| 6323 | false); |
| 6324 | if (ret) |
| 6325 | return ret; |
| 6326 | |
| 6327 | ret = __alloc_contig_migrate_range(&cc, start, end); |
| 6328 | if (ret) |
| 6329 | goto done; |
| 6330 | |
| 6331 | /* |
| 6332 | * Pages from [start, end) are within a MAX_ORDER_NR_PAGES |
| 6333 | * aligned blocks that are marked as MIGRATE_ISOLATE. What's |
| 6334 | * more, all pages in [start, end) are free in page allocator. |
| 6335 | * What we are going to do is to allocate all pages from |
| 6336 | * [start, end) (that is remove them from page allocator). |
| 6337 | * |
| 6338 | * The only problem is that pages at the beginning and at the |
| 6339 | * end of interesting range may be not aligned with pages that |
| 6340 | * page allocator holds, ie. they can be part of higher order |
| 6341 | * pages. Because of this, we reserve the bigger range and |
| 6342 | * once this is done free the pages we are not interested in. |
| 6343 | * |
| 6344 | * We don't have to hold zone->lock here because the pages are |
| 6345 | * isolated thus they won't get removed from buddy. |
| 6346 | */ |
| 6347 | |
| 6348 | lru_add_drain_all(); |
| 6349 | drain_all_pages(); |
| 6350 | |
| 6351 | order = 0; |
| 6352 | outer_start = start; |
| 6353 | while (!PageBuddy(pfn_to_page(outer_start))) { |
| 6354 | if (++order >= MAX_ORDER) { |
| 6355 | ret = -EBUSY; |
| 6356 | goto done; |
| 6357 | } |
| 6358 | outer_start &= ~0UL << order; |
| 6359 | } |
| 6360 | |
| 6361 | /* Make sure the range is really isolated. */ |
| 6362 | if (test_pages_isolated(outer_start, end, false)) { |
| 6363 | pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n", |
| 6364 | outer_start, end); |
| 6365 | ret = -EBUSY; |
| 6366 | goto done; |
| 6367 | } |
| 6368 | |
| 6369 | |
| 6370 | /* Grab isolated pages from freelists. */ |
| 6371 | outer_end = isolate_freepages_range(&cc, outer_start, end); |
| 6372 | if (!outer_end) { |
| 6373 | ret = -EBUSY; |
| 6374 | goto done; |
| 6375 | } |
| 6376 | |
| 6377 | /* Free head and tail (if any) */ |
| 6378 | if (start != outer_start) |
| 6379 | free_contig_range(outer_start, start - outer_start); |
| 6380 | if (end != outer_end) |
| 6381 | free_contig_range(end, outer_end - end); |
| 6382 | |
| 6383 | done: |
| 6384 | undo_isolate_page_range(pfn_max_align_down(start), |
| 6385 | pfn_max_align_up(end), migratetype); |
| 6386 | return ret; |
| 6387 | } |
| 6388 | |
| 6389 | void free_contig_range(unsigned long pfn, unsigned nr_pages) |
| 6390 | { |
| 6391 | unsigned int count = 0; |
| 6392 | |
| 6393 | for (; nr_pages--; pfn++) { |
| 6394 | struct page *page = pfn_to_page(pfn); |
| 6395 | |
| 6396 | count += page_count(page) != 1; |
| 6397 | __free_page(page); |
| 6398 | } |
| 6399 | WARN(count != 0, "%d pages are still in use!\n", count); |
| 6400 | } |
| 6401 | #endif |
| 6402 | |
| 6403 | #ifdef CONFIG_MEMORY_HOTPLUG |
| 6404 | /* |
| 6405 | * The zone indicated has a new number of managed_pages; batch sizes and percpu |
| 6406 | * page high values need to be recalulated. |
| 6407 | */ |
| 6408 | void __meminit zone_pcp_update(struct zone *zone) |
| 6409 | { |
| 6410 | unsigned cpu; |
| 6411 | mutex_lock(&pcp_batch_high_lock); |
| 6412 | for_each_possible_cpu(cpu) |
| 6413 | pageset_set_high_and_batch(zone, |
| 6414 | per_cpu_ptr(zone->pageset, cpu)); |
| 6415 | mutex_unlock(&pcp_batch_high_lock); |
| 6416 | } |
| 6417 | #endif |
| 6418 | |
| 6419 | void zone_pcp_reset(struct zone *zone) |
| 6420 | { |
| 6421 | unsigned long flags; |
| 6422 | int cpu; |
| 6423 | struct per_cpu_pageset *pset; |
| 6424 | |
| 6425 | /* avoid races with drain_pages() */ |
| 6426 | local_irq_save(flags); |
| 6427 | if (zone->pageset != &boot_pageset) { |
| 6428 | for_each_online_cpu(cpu) { |
| 6429 | pset = per_cpu_ptr(zone->pageset, cpu); |
| 6430 | drain_zonestat(zone, pset); |
| 6431 | } |
| 6432 | free_percpu(zone->pageset); |
| 6433 | zone->pageset = &boot_pageset; |
| 6434 | } |
| 6435 | local_irq_restore(flags); |
| 6436 | } |
| 6437 | |
| 6438 | #ifdef CONFIG_MEMORY_HOTREMOVE |
| 6439 | /* |
| 6440 | * All pages in the range must be isolated before calling this. |
| 6441 | */ |
| 6442 | void |
| 6443 | __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) |
| 6444 | { |
| 6445 | struct page *page; |
| 6446 | struct zone *zone; |
| 6447 | int order, i; |
| 6448 | unsigned long pfn; |
| 6449 | unsigned long flags; |
| 6450 | /* find the first valid pfn */ |
| 6451 | for (pfn = start_pfn; pfn < end_pfn; pfn++) |
| 6452 | if (pfn_valid(pfn)) |
| 6453 | break; |
| 6454 | if (pfn == end_pfn) |
| 6455 | return; |
| 6456 | zone = page_zone(pfn_to_page(pfn)); |
| 6457 | spin_lock_irqsave(&zone->lock, flags); |
| 6458 | pfn = start_pfn; |
| 6459 | while (pfn < end_pfn) { |
| 6460 | if (!pfn_valid(pfn)) { |
| 6461 | pfn++; |
| 6462 | continue; |
| 6463 | } |
| 6464 | page = pfn_to_page(pfn); |
| 6465 | /* |
| 6466 | * The HWPoisoned page may be not in buddy system, and |
| 6467 | * page_count() is not 0. |
| 6468 | */ |
| 6469 | if (unlikely(!PageBuddy(page) && PageHWPoison(page))) { |
| 6470 | pfn++; |
| 6471 | SetPageReserved(page); |
| 6472 | continue; |
| 6473 | } |
| 6474 | |
| 6475 | BUG_ON(page_count(page)); |
| 6476 | BUG_ON(!PageBuddy(page)); |
| 6477 | order = page_order(page); |
| 6478 | #ifdef CONFIG_DEBUG_VM |
| 6479 | printk(KERN_INFO "remove from free list %lx %d %lx\n", |
| 6480 | pfn, 1 << order, end_pfn); |
| 6481 | #endif |
| 6482 | list_del(&page->lru); |
| 6483 | rmv_page_order(page); |
| 6484 | zone->free_area[order].nr_free--; |
| 6485 | for (i = 0; i < (1 << order); i++) |
| 6486 | SetPageReserved((page+i)); |
| 6487 | pfn += (1 << order); |
| 6488 | } |
| 6489 | spin_unlock_irqrestore(&zone->lock, flags); |
| 6490 | } |
| 6491 | #endif |
| 6492 | |
| 6493 | #ifdef CONFIG_MEMORY_FAILURE |
| 6494 | bool is_free_buddy_page(struct page *page) |
| 6495 | { |
| 6496 | struct zone *zone = page_zone(page); |
| 6497 | unsigned long pfn = page_to_pfn(page); |
| 6498 | unsigned long flags; |
| 6499 | int order; |
| 6500 | |
| 6501 | spin_lock_irqsave(&zone->lock, flags); |
| 6502 | for (order = 0; order < MAX_ORDER; order++) { |
| 6503 | struct page *page_head = page - (pfn & ((1 << order) - 1)); |
| 6504 | |
| 6505 | if (PageBuddy(page_head) && page_order(page_head) >= order) |
| 6506 | break; |
| 6507 | } |
| 6508 | spin_unlock_irqrestore(&zone->lock, flags); |
| 6509 | |
| 6510 | return order < MAX_ORDER; |
| 6511 | } |
| 6512 | #endif |
| 6513 | |
| 6514 | static const struct trace_print_flags pageflag_names[] = { |
| 6515 | {1UL << PG_locked, "locked" }, |
| 6516 | {1UL << PG_error, "error" }, |
| 6517 | {1UL << PG_referenced, "referenced" }, |
| 6518 | {1UL << PG_uptodate, "uptodate" }, |
| 6519 | {1UL << PG_dirty, "dirty" }, |
| 6520 | {1UL << PG_lru, "lru" }, |
| 6521 | {1UL << PG_active, "active" }, |
| 6522 | {1UL << PG_slab, "slab" }, |
| 6523 | {1UL << PG_owner_priv_1, "owner_priv_1" }, |
| 6524 | {1UL << PG_arch_1, "arch_1" }, |
| 6525 | {1UL << PG_reserved, "reserved" }, |
| 6526 | {1UL << PG_private, "private" }, |
| 6527 | {1UL << PG_private_2, "private_2" }, |
| 6528 | {1UL << PG_writeback, "writeback" }, |
| 6529 | #ifdef CONFIG_PAGEFLAGS_EXTENDED |
| 6530 | {1UL << PG_head, "head" }, |
| 6531 | {1UL << PG_tail, "tail" }, |
| 6532 | #else |
| 6533 | {1UL << PG_compound, "compound" }, |
| 6534 | #endif |
| 6535 | {1UL << PG_swapcache, "swapcache" }, |
| 6536 | {1UL << PG_mappedtodisk, "mappedtodisk" }, |
| 6537 | {1UL << PG_reclaim, "reclaim" }, |
| 6538 | {1UL << PG_swapbacked, "swapbacked" }, |
| 6539 | {1UL << PG_unevictable, "unevictable" }, |
| 6540 | #ifdef CONFIG_MMU |
| 6541 | {1UL << PG_mlocked, "mlocked" }, |
| 6542 | #endif |
| 6543 | #ifdef CONFIG_ARCH_USES_PG_UNCACHED |
| 6544 | {1UL << PG_uncached, "uncached" }, |
| 6545 | #endif |
| 6546 | #ifdef CONFIG_MEMORY_FAILURE |
| 6547 | {1UL << PG_hwpoison, "hwpoison" }, |
| 6548 | #endif |
| 6549 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
| 6550 | {1UL << PG_compound_lock, "compound_lock" }, |
| 6551 | #endif |
| 6552 | }; |
| 6553 | |
| 6554 | static void dump_page_flags(unsigned long flags) |
| 6555 | { |
| 6556 | const char *delim = ""; |
| 6557 | unsigned long mask; |
| 6558 | int i; |
| 6559 | |
| 6560 | BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS); |
| 6561 | |
| 6562 | printk(KERN_ALERT "page flags: %#lx(", flags); |
| 6563 | |
| 6564 | /* remove zone id */ |
| 6565 | flags &= (1UL << NR_PAGEFLAGS) - 1; |
| 6566 | |
| 6567 | for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) { |
| 6568 | |
| 6569 | mask = pageflag_names[i].mask; |
| 6570 | if ((flags & mask) != mask) |
| 6571 | continue; |
| 6572 | |
| 6573 | flags &= ~mask; |
| 6574 | printk("%s%s", delim, pageflag_names[i].name); |
| 6575 | delim = "|"; |
| 6576 | } |
| 6577 | |
| 6578 | /* check for left over flags */ |
| 6579 | if (flags) |
| 6580 | printk("%s%#lx", delim, flags); |
| 6581 | |
| 6582 | printk(")\n"); |
| 6583 | } |
| 6584 | |
| 6585 | void dump_page_badflags(struct page *page, const char *reason, |
| 6586 | unsigned long badflags) |
| 6587 | { |
| 6588 | printk(KERN_ALERT |
| 6589 | "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n", |
| 6590 | page, atomic_read(&page->_count), page_mapcount(page), |
| 6591 | page->mapping, page->index); |
| 6592 | dump_page_flags(page->flags); |
| 6593 | if (reason) |
| 6594 | pr_alert("page dumped because: %s\n", reason); |
| 6595 | if (page->flags & badflags) { |
| 6596 | pr_alert("bad because of flags:\n"); |
| 6597 | dump_page_flags(page->flags & badflags); |
| 6598 | } |
| 6599 | mem_cgroup_print_bad_page(page); |
| 6600 | } |
| 6601 | |
| 6602 | void dump_page(struct page *page, const char *reason) |
| 6603 | { |
| 6604 | dump_page_badflags(page, reason, 0); |
| 6605 | } |
| 6606 | EXPORT_SYMBOL(dump_page); |