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/prefetch.h> 57 #include <linux/mm_inline.h> 58 #include <linux/migrate.h> 59 #include <linux/page-debug-flags.h> 60 #include <linux/hugetlb.h> 61 #include <linux/sched/rt.h> 62 63 #include <asm/sections.h> 64 #include <asm/tlbflush.h> 65 #include <asm/div64.h> 66 #include "internal.h" 67 68 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */ 69 static DEFINE_MUTEX(pcp_batch_high_lock); 70 #define MIN_PERCPU_PAGELIST_FRACTION (8) 71 72 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID 73 DEFINE_PER_CPU(int, numa_node); 74 EXPORT_PER_CPU_SYMBOL(numa_node); 75 #endif 76 77 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 78 /* 79 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. 80 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. 81 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() 82 * defined in <linux/topology.h>. 83 */ 84 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ 85 EXPORT_PER_CPU_SYMBOL(_numa_mem_); 86 int _node_numa_mem_[MAX_NUMNODES]; 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, unsigned int order, 413 gfp_t gfp_flags) 414 { 415 int i; 416 417 /* 418 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 419 * and __GFP_HIGHMEM from hard or soft interrupt context. 420 */ 421 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 422 for (i = 0; i < (1 << order); i++) 423 clear_highpage(page + i); 424 } 425 426 #ifdef CONFIG_DEBUG_PAGEALLOC 427 unsigned int _debug_guardpage_minorder; 428 429 static int __init debug_guardpage_minorder_setup(char *buf) 430 { 431 unsigned long res; 432 433 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) { 434 printk(KERN_ERR "Bad debug_guardpage_minorder value\n"); 435 return 0; 436 } 437 _debug_guardpage_minorder = res; 438 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res); 439 return 0; 440 } 441 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup); 442 443 static inline void set_page_guard_flag(struct page *page) 444 { 445 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); 446 } 447 448 static inline void clear_page_guard_flag(struct page *page) 449 { 450 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); 451 } 452 #else 453 static inline void set_page_guard_flag(struct page *page) { } 454 static inline void clear_page_guard_flag(struct page *page) { } 455 #endif 456 457 static inline void set_page_order(struct page *page, unsigned int order) 458 { 459 set_page_private(page, order); 460 __SetPageBuddy(page); 461 } 462 463 static inline void rmv_page_order(struct page *page) 464 { 465 __ClearPageBuddy(page); 466 set_page_private(page, 0); 467 } 468 469 /* 470 * Locate the struct page for both the matching buddy in our 471 * pair (buddy1) and the combined O(n+1) page they form (page). 472 * 473 * 1) Any buddy B1 will have an order O twin B2 which satisfies 474 * the following equation: 475 * B2 = B1 ^ (1 << O) 476 * For example, if the starting buddy (buddy2) is #8 its order 477 * 1 buddy is #10: 478 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 479 * 480 * 2) Any buddy B will have an order O+1 parent P which 481 * satisfies the following equation: 482 * P = B & ~(1 << O) 483 * 484 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 485 */ 486 static inline unsigned long 487 __find_buddy_index(unsigned long page_idx, unsigned int order) 488 { 489 return page_idx ^ (1 << order); 490 } 491 492 /* 493 * This function checks whether a page is free && is the buddy 494 * we can do coalesce a page and its buddy if 495 * (a) the buddy is not in a hole && 496 * (b) the buddy is in the buddy system && 497 * (c) a page and its buddy have the same order && 498 * (d) a page and its buddy are in the same zone. 499 * 500 * For recording whether a page is in the buddy system, we set ->_mapcount 501 * PAGE_BUDDY_MAPCOUNT_VALUE. 502 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is 503 * serialized by zone->lock. 504 * 505 * For recording page's order, we use page_private(page). 506 */ 507 static inline int page_is_buddy(struct page *page, struct page *buddy, 508 unsigned int order) 509 { 510 if (!pfn_valid_within(page_to_pfn(buddy))) 511 return 0; 512 513 if (page_is_guard(buddy) && page_order(buddy) == order) { 514 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); 515 516 if (page_zone_id(page) != page_zone_id(buddy)) 517 return 0; 518 519 return 1; 520 } 521 522 if (PageBuddy(buddy) && page_order(buddy) == order) { 523 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); 524 525 /* 526 * zone check is done late to avoid uselessly 527 * calculating zone/node ids for pages that could 528 * never merge. 529 */ 530 if (page_zone_id(page) != page_zone_id(buddy)) 531 return 0; 532 533 return 1; 534 } 535 return 0; 536 } 537 538 /* 539 * Freeing function for a buddy system allocator. 540 * 541 * The concept of a buddy system is to maintain direct-mapped table 542 * (containing bit values) for memory blocks of various "orders". 543 * The bottom level table contains the map for the smallest allocatable 544 * units of memory (here, pages), and each level above it describes 545 * pairs of units from the levels below, hence, "buddies". 546 * At a high level, all that happens here is marking the table entry 547 * at the bottom level available, and propagating the changes upward 548 * as necessary, plus some accounting needed to play nicely with other 549 * parts of the VM system. 550 * At each level, we keep a list of pages, which are heads of continuous 551 * free pages of length of (1 << order) and marked with _mapcount 552 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page) 553 * field. 554 * So when we are allocating or freeing one, we can derive the state of the 555 * other. That is, if we allocate a small block, and both were 556 * free, the remainder of the region must be split into blocks. 557 * If a block is freed, and its buddy is also free, then this 558 * triggers coalescing into a block of larger size. 559 * 560 * -- nyc 561 */ 562 563 static inline void __free_one_page(struct page *page, 564 unsigned long pfn, 565 struct zone *zone, unsigned int order, 566 int migratetype) 567 { 568 unsigned long page_idx; 569 unsigned long combined_idx; 570 unsigned long uninitialized_var(buddy_idx); 571 struct page *buddy; 572 573 VM_BUG_ON(!zone_is_initialized(zone)); 574 575 if (unlikely(PageCompound(page))) 576 if (unlikely(destroy_compound_page(page, order))) 577 return; 578 579 VM_BUG_ON(migratetype == -1); 580 581 page_idx = pfn & ((1 << MAX_ORDER) - 1); 582 583 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page); 584 VM_BUG_ON_PAGE(bad_range(zone, page), page); 585 586 while (order < MAX_ORDER-1) { 587 buddy_idx = __find_buddy_index(page_idx, order); 588 buddy = page + (buddy_idx - page_idx); 589 if (!page_is_buddy(page, buddy, order)) 590 break; 591 /* 592 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, 593 * merge with it and move up one order. 594 */ 595 if (page_is_guard(buddy)) { 596 clear_page_guard_flag(buddy); 597 set_page_private(page, 0); 598 __mod_zone_freepage_state(zone, 1 << order, 599 migratetype); 600 } else { 601 list_del(&buddy->lru); 602 zone->free_area[order].nr_free--; 603 rmv_page_order(buddy); 604 } 605 combined_idx = buddy_idx & page_idx; 606 page = page + (combined_idx - page_idx); 607 page_idx = combined_idx; 608 order++; 609 } 610 set_page_order(page, order); 611 612 /* 613 * If this is not the largest possible page, check if the buddy 614 * of the next-highest order is free. If it is, it's possible 615 * that pages are being freed that will coalesce soon. In case, 616 * that is happening, add the free page to the tail of the list 617 * so it's less likely to be used soon and more likely to be merged 618 * as a higher order page 619 */ 620 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { 621 struct page *higher_page, *higher_buddy; 622 combined_idx = buddy_idx & page_idx; 623 higher_page = page + (combined_idx - page_idx); 624 buddy_idx = __find_buddy_index(combined_idx, order + 1); 625 higher_buddy = higher_page + (buddy_idx - combined_idx); 626 if (page_is_buddy(higher_page, higher_buddy, order + 1)) { 627 list_add_tail(&page->lru, 628 &zone->free_area[order].free_list[migratetype]); 629 goto out; 630 } 631 } 632 633 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 634 out: 635 zone->free_area[order].nr_free++; 636 } 637 638 static inline int free_pages_check(struct page *page) 639 { 640 const char *bad_reason = NULL; 641 unsigned long bad_flags = 0; 642 643 if (unlikely(page_mapcount(page))) 644 bad_reason = "nonzero mapcount"; 645 if (unlikely(page->mapping != NULL)) 646 bad_reason = "non-NULL mapping"; 647 if (unlikely(atomic_read(&page->_count) != 0)) 648 bad_reason = "nonzero _count"; 649 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) { 650 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set"; 651 bad_flags = PAGE_FLAGS_CHECK_AT_FREE; 652 } 653 if (unlikely(mem_cgroup_bad_page_check(page))) 654 bad_reason = "cgroup check failed"; 655 if (unlikely(bad_reason)) { 656 bad_page(page, bad_reason, bad_flags); 657 return 1; 658 } 659 page_cpupid_reset_last(page); 660 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 661 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 662 return 0; 663 } 664 665 /* 666 * Frees a number of pages from the PCP lists 667 * Assumes all pages on list are in same zone, and of same order. 668 * count is the number of pages to free. 669 * 670 * If the zone was previously in an "all pages pinned" state then look to 671 * see if this freeing clears that state. 672 * 673 * And clear the zone's pages_scanned counter, to hold off the "all pages are 674 * pinned" detection logic. 675 */ 676 static void free_pcppages_bulk(struct zone *zone, int count, 677 struct per_cpu_pages *pcp) 678 { 679 int migratetype = 0; 680 int batch_free = 0; 681 int to_free = count; 682 unsigned long nr_scanned; 683 684 spin_lock(&zone->lock); 685 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED); 686 if (nr_scanned) 687 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned); 688 689 while (to_free) { 690 struct page *page; 691 struct list_head *list; 692 693 /* 694 * Remove pages from lists in a round-robin fashion. A 695 * batch_free count is maintained that is incremented when an 696 * empty list is encountered. This is so more pages are freed 697 * off fuller lists instead of spinning excessively around empty 698 * lists 699 */ 700 do { 701 batch_free++; 702 if (++migratetype == MIGRATE_PCPTYPES) 703 migratetype = 0; 704 list = &pcp->lists[migratetype]; 705 } while (list_empty(list)); 706 707 /* This is the only non-empty list. Free them all. */ 708 if (batch_free == MIGRATE_PCPTYPES) 709 batch_free = to_free; 710 711 do { 712 int mt; /* migratetype of the to-be-freed page */ 713 714 page = list_entry(list->prev, struct page, lru); 715 /* must delete as __free_one_page list manipulates */ 716 list_del(&page->lru); 717 mt = get_freepage_migratetype(page); 718 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 719 __free_one_page(page, page_to_pfn(page), zone, 0, mt); 720 trace_mm_page_pcpu_drain(page, 0, mt); 721 if (likely(!is_migrate_isolate_page(page))) { 722 __mod_zone_page_state(zone, NR_FREE_PAGES, 1); 723 if (is_migrate_cma(mt)) 724 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1); 725 } 726 } while (--to_free && --batch_free && !list_empty(list)); 727 } 728 spin_unlock(&zone->lock); 729 } 730 731 static void free_one_page(struct zone *zone, 732 struct page *page, unsigned long pfn, 733 unsigned int order, 734 int migratetype) 735 { 736 unsigned long nr_scanned; 737 spin_lock(&zone->lock); 738 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED); 739 if (nr_scanned) 740 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned); 741 742 __free_one_page(page, pfn, zone, order, migratetype); 743 if (unlikely(!is_migrate_isolate(migratetype))) 744 __mod_zone_freepage_state(zone, 1 << order, migratetype); 745 spin_unlock(&zone->lock); 746 } 747 748 static bool free_pages_prepare(struct page *page, unsigned int order) 749 { 750 int i; 751 int bad = 0; 752 753 trace_mm_page_free(page, order); 754 kmemcheck_free_shadow(page, order); 755 756 if (PageAnon(page)) 757 page->mapping = NULL; 758 for (i = 0; i < (1 << order); i++) 759 bad += free_pages_check(page + i); 760 if (bad) 761 return false; 762 763 if (!PageHighMem(page)) { 764 debug_check_no_locks_freed(page_address(page), 765 PAGE_SIZE << order); 766 debug_check_no_obj_freed(page_address(page), 767 PAGE_SIZE << order); 768 } 769 arch_free_page(page, order); 770 kernel_map_pages(page, 1 << order, 0); 771 772 return true; 773 } 774 775 static void __free_pages_ok(struct page *page, unsigned int order) 776 { 777 unsigned long flags; 778 int migratetype; 779 unsigned long pfn = page_to_pfn(page); 780 781 if (!free_pages_prepare(page, order)) 782 return; 783 784 migratetype = get_pfnblock_migratetype(page, pfn); 785 local_irq_save(flags); 786 __count_vm_events(PGFREE, 1 << order); 787 set_freepage_migratetype(page, migratetype); 788 free_one_page(page_zone(page), page, pfn, order, migratetype); 789 local_irq_restore(flags); 790 } 791 792 void __init __free_pages_bootmem(struct page *page, unsigned int order) 793 { 794 unsigned int nr_pages = 1 << order; 795 struct page *p = page; 796 unsigned int loop; 797 798 prefetchw(p); 799 for (loop = 0; loop < (nr_pages - 1); loop++, p++) { 800 prefetchw(p + 1); 801 __ClearPageReserved(p); 802 set_page_count(p, 0); 803 } 804 __ClearPageReserved(p); 805 set_page_count(p, 0); 806 807 page_zone(page)->managed_pages += nr_pages; 808 set_page_refcounted(page); 809 __free_pages(page, order); 810 } 811 812 #ifdef CONFIG_CMA 813 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */ 814 void __init init_cma_reserved_pageblock(struct page *page) 815 { 816 unsigned i = pageblock_nr_pages; 817 struct page *p = page; 818 819 do { 820 __ClearPageReserved(p); 821 set_page_count(p, 0); 822 } while (++p, --i); 823 824 set_pageblock_migratetype(page, MIGRATE_CMA); 825 826 if (pageblock_order >= MAX_ORDER) { 827 i = pageblock_nr_pages; 828 p = page; 829 do { 830 set_page_refcounted(p); 831 __free_pages(p, MAX_ORDER - 1); 832 p += MAX_ORDER_NR_PAGES; 833 } while (i -= MAX_ORDER_NR_PAGES); 834 } else { 835 set_page_refcounted(page); 836 __free_pages(page, pageblock_order); 837 } 838 839 adjust_managed_page_count(page, pageblock_nr_pages); 840 } 841 #endif 842 843 /* 844 * The order of subdivision here is critical for the IO subsystem. 845 * Please do not alter this order without good reasons and regression 846 * testing. Specifically, as large blocks of memory are subdivided, 847 * the order in which smaller blocks are delivered depends on the order 848 * they're subdivided in this function. This is the primary factor 849 * influencing the order in which pages are delivered to the IO 850 * subsystem according to empirical testing, and this is also justified 851 * by considering the behavior of a buddy system containing a single 852 * large block of memory acted on by a series of small allocations. 853 * This behavior is a critical factor in sglist merging's success. 854 * 855 * -- nyc 856 */ 857 static inline void expand(struct zone *zone, struct page *page, 858 int low, int high, struct free_area *area, 859 int migratetype) 860 { 861 unsigned long size = 1 << high; 862 863 while (high > low) { 864 area--; 865 high--; 866 size >>= 1; 867 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]); 868 869 #ifdef CONFIG_DEBUG_PAGEALLOC 870 if (high < debug_guardpage_minorder()) { 871 /* 872 * Mark as guard pages (or page), that will allow to 873 * merge back to allocator when buddy will be freed. 874 * Corresponding page table entries will not be touched, 875 * pages will stay not present in virtual address space 876 */ 877 INIT_LIST_HEAD(&page[size].lru); 878 set_page_guard_flag(&page[size]); 879 set_page_private(&page[size], high); 880 /* Guard pages are not available for any usage */ 881 __mod_zone_freepage_state(zone, -(1 << high), 882 migratetype); 883 continue; 884 } 885 #endif 886 list_add(&page[size].lru, &area->free_list[migratetype]); 887 area->nr_free++; 888 set_page_order(&page[size], high); 889 } 890 } 891 892 /* 893 * This page is about to be returned from the page allocator 894 */ 895 static inline int check_new_page(struct page *page) 896 { 897 const char *bad_reason = NULL; 898 unsigned long bad_flags = 0; 899 900 if (unlikely(page_mapcount(page))) 901 bad_reason = "nonzero mapcount"; 902 if (unlikely(page->mapping != NULL)) 903 bad_reason = "non-NULL mapping"; 904 if (unlikely(atomic_read(&page->_count) != 0)) 905 bad_reason = "nonzero _count"; 906 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) { 907 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set"; 908 bad_flags = PAGE_FLAGS_CHECK_AT_PREP; 909 } 910 if (unlikely(mem_cgroup_bad_page_check(page))) 911 bad_reason = "cgroup check failed"; 912 if (unlikely(bad_reason)) { 913 bad_page(page, bad_reason, bad_flags); 914 return 1; 915 } 916 return 0; 917 } 918 919 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags) 920 { 921 int i; 922 923 for (i = 0; i < (1 << order); i++) { 924 struct page *p = page + i; 925 if (unlikely(check_new_page(p))) 926 return 1; 927 } 928 929 set_page_private(page, 0); 930 set_page_refcounted(page); 931 932 arch_alloc_page(page, order); 933 kernel_map_pages(page, 1 << order, 1); 934 935 if (gfp_flags & __GFP_ZERO) 936 prep_zero_page(page, order, gfp_flags); 937 938 if (order && (gfp_flags & __GFP_COMP)) 939 prep_compound_page(page, order); 940 941 return 0; 942 } 943 944 /* 945 * Go through the free lists for the given migratetype and remove 946 * the smallest available page from the freelists 947 */ 948 static inline 949 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 950 int migratetype) 951 { 952 unsigned int current_order; 953 struct free_area *area; 954 struct page *page; 955 956 /* Find a page of the appropriate size in the preferred list */ 957 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 958 area = &(zone->free_area[current_order]); 959 if (list_empty(&area->free_list[migratetype])) 960 continue; 961 962 page = list_entry(area->free_list[migratetype].next, 963 struct page, lru); 964 list_del(&page->lru); 965 rmv_page_order(page); 966 area->nr_free--; 967 expand(zone, page, order, current_order, area, migratetype); 968 set_freepage_migratetype(page, migratetype); 969 return page; 970 } 971 972 return NULL; 973 } 974 975 976 /* 977 * This array describes the order lists are fallen back to when 978 * the free lists for the desirable migrate type are depleted 979 */ 980 static int fallbacks[MIGRATE_TYPES][4] = { 981 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 982 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 983 #ifdef CONFIG_CMA 984 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 985 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */ 986 #else 987 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 988 #endif 989 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */ 990 #ifdef CONFIG_MEMORY_ISOLATION 991 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */ 992 #endif 993 }; 994 995 /* 996 * Move the free pages in a range to the free lists of the requested type. 997 * Note that start_page and end_pages are not aligned on a pageblock 998 * boundary. If alignment is required, use move_freepages_block() 999 */ 1000 int move_freepages(struct zone *zone, 1001 struct page *start_page, struct page *end_page, 1002 int migratetype) 1003 { 1004 struct page *page; 1005 unsigned long order; 1006 int pages_moved = 0; 1007 1008 #ifndef CONFIG_HOLES_IN_ZONE 1009 /* 1010 * page_zone is not safe to call in this context when 1011 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 1012 * anyway as we check zone boundaries in move_freepages_block(). 1013 * Remove at a later date when no bug reports exist related to 1014 * grouping pages by mobility 1015 */ 1016 VM_BUG_ON(page_zone(start_page) != page_zone(end_page)); 1017 #endif 1018 1019 for (page = start_page; page <= end_page;) { 1020 /* Make sure we are not inadvertently changing nodes */ 1021 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page); 1022 1023 if (!pfn_valid_within(page_to_pfn(page))) { 1024 page++; 1025 continue; 1026 } 1027 1028 if (!PageBuddy(page)) { 1029 page++; 1030 continue; 1031 } 1032 1033 order = page_order(page); 1034 list_move(&page->lru, 1035 &zone->free_area[order].free_list[migratetype]); 1036 set_freepage_migratetype(page, migratetype); 1037 page += 1 << order; 1038 pages_moved += 1 << order; 1039 } 1040 1041 return pages_moved; 1042 } 1043 1044 int move_freepages_block(struct zone *zone, struct page *page, 1045 int migratetype) 1046 { 1047 unsigned long start_pfn, end_pfn; 1048 struct page *start_page, *end_page; 1049 1050 start_pfn = page_to_pfn(page); 1051 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 1052 start_page = pfn_to_page(start_pfn); 1053 end_page = start_page + pageblock_nr_pages - 1; 1054 end_pfn = start_pfn + pageblock_nr_pages - 1; 1055 1056 /* Do not cross zone boundaries */ 1057 if (!zone_spans_pfn(zone, start_pfn)) 1058 start_page = page; 1059 if (!zone_spans_pfn(zone, end_pfn)) 1060 return 0; 1061 1062 return move_freepages(zone, start_page, end_page, migratetype); 1063 } 1064 1065 static void change_pageblock_range(struct page *pageblock_page, 1066 int start_order, int migratetype) 1067 { 1068 int nr_pageblocks = 1 << (start_order - pageblock_order); 1069 1070 while (nr_pageblocks--) { 1071 set_pageblock_migratetype(pageblock_page, migratetype); 1072 pageblock_page += pageblock_nr_pages; 1073 } 1074 } 1075 1076 /* 1077 * If breaking a large block of pages, move all free pages to the preferred 1078 * allocation list. If falling back for a reclaimable kernel allocation, be 1079 * more aggressive about taking ownership of free pages. 1080 * 1081 * On the other hand, never change migration type of MIGRATE_CMA pageblocks 1082 * nor move CMA pages to different free lists. We don't want unmovable pages 1083 * to be allocated from MIGRATE_CMA areas. 1084 * 1085 * Returns the new migratetype of the pageblock (or the same old migratetype 1086 * if it was unchanged). 1087 */ 1088 static int try_to_steal_freepages(struct zone *zone, struct page *page, 1089 int start_type, int fallback_type) 1090 { 1091 int current_order = page_order(page); 1092 1093 /* 1094 * When borrowing from MIGRATE_CMA, we need to release the excess 1095 * buddy pages to CMA itself. We also ensure the freepage_migratetype 1096 * is set to CMA so it is returned to the correct freelist in case 1097 * the page ends up being not actually allocated from the pcp lists. 1098 */ 1099 if (is_migrate_cma(fallback_type)) 1100 return fallback_type; 1101 1102 /* Take ownership for orders >= pageblock_order */ 1103 if (current_order >= pageblock_order) { 1104 change_pageblock_range(page, current_order, start_type); 1105 return start_type; 1106 } 1107 1108 if (current_order >= pageblock_order / 2 || 1109 start_type == MIGRATE_RECLAIMABLE || 1110 page_group_by_mobility_disabled) { 1111 int pages; 1112 1113 pages = move_freepages_block(zone, page, start_type); 1114 1115 /* Claim the whole block if over half of it is free */ 1116 if (pages >= (1 << (pageblock_order-1)) || 1117 page_group_by_mobility_disabled) { 1118 1119 set_pageblock_migratetype(page, start_type); 1120 return start_type; 1121 } 1122 1123 } 1124 1125 return fallback_type; 1126 } 1127 1128 /* Remove an element from the buddy allocator from the fallback list */ 1129 static inline struct page * 1130 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype) 1131 { 1132 struct free_area *area; 1133 unsigned int current_order; 1134 struct page *page; 1135 int migratetype, new_type, i; 1136 1137 /* Find the largest possible block of pages in the other list */ 1138 for (current_order = MAX_ORDER-1; 1139 current_order >= order && current_order <= MAX_ORDER-1; 1140 --current_order) { 1141 for (i = 0;; i++) { 1142 migratetype = fallbacks[start_migratetype][i]; 1143 1144 /* MIGRATE_RESERVE handled later if necessary */ 1145 if (migratetype == MIGRATE_RESERVE) 1146 break; 1147 1148 area = &(zone->free_area[current_order]); 1149 if (list_empty(&area->free_list[migratetype])) 1150 continue; 1151 1152 page = list_entry(area->free_list[migratetype].next, 1153 struct page, lru); 1154 area->nr_free--; 1155 1156 new_type = try_to_steal_freepages(zone, page, 1157 start_migratetype, 1158 migratetype); 1159 1160 /* Remove the page from the freelists */ 1161 list_del(&page->lru); 1162 rmv_page_order(page); 1163 1164 expand(zone, page, order, current_order, area, 1165 new_type); 1166 /* The freepage_migratetype may differ from pageblock's 1167 * migratetype depending on the decisions in 1168 * try_to_steal_freepages. This is OK as long as it does 1169 * not differ for MIGRATE_CMA type. 1170 */ 1171 set_freepage_migratetype(page, new_type); 1172 1173 trace_mm_page_alloc_extfrag(page, order, current_order, 1174 start_migratetype, migratetype, new_type); 1175 1176 return page; 1177 } 1178 } 1179 1180 return NULL; 1181 } 1182 1183 /* 1184 * Do the hard work of removing an element from the buddy allocator. 1185 * Call me with the zone->lock already held. 1186 */ 1187 static struct page *__rmqueue(struct zone *zone, unsigned int order, 1188 int migratetype) 1189 { 1190 struct page *page; 1191 1192 retry_reserve: 1193 page = __rmqueue_smallest(zone, order, migratetype); 1194 1195 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 1196 page = __rmqueue_fallback(zone, order, migratetype); 1197 1198 /* 1199 * Use MIGRATE_RESERVE rather than fail an allocation. goto 1200 * is used because __rmqueue_smallest is an inline function 1201 * and we want just one call site 1202 */ 1203 if (!page) { 1204 migratetype = MIGRATE_RESERVE; 1205 goto retry_reserve; 1206 } 1207 } 1208 1209 trace_mm_page_alloc_zone_locked(page, order, migratetype); 1210 return page; 1211 } 1212 1213 /* 1214 * Obtain a specified number of elements from the buddy allocator, all under 1215 * a single hold of the lock, for efficiency. Add them to the supplied list. 1216 * Returns the number of new pages which were placed at *list. 1217 */ 1218 static int rmqueue_bulk(struct zone *zone, unsigned int order, 1219 unsigned long count, struct list_head *list, 1220 int migratetype, bool cold) 1221 { 1222 int i; 1223 1224 spin_lock(&zone->lock); 1225 for (i = 0; i < count; ++i) { 1226 struct page *page = __rmqueue(zone, order, migratetype); 1227 if (unlikely(page == NULL)) 1228 break; 1229 1230 /* 1231 * Split buddy pages returned by expand() are received here 1232 * in physical page order. The page is added to the callers and 1233 * list and the list head then moves forward. From the callers 1234 * perspective, the linked list is ordered by page number in 1235 * some conditions. This is useful for IO devices that can 1236 * merge IO requests if the physical pages are ordered 1237 * properly. 1238 */ 1239 if (likely(!cold)) 1240 list_add(&page->lru, list); 1241 else 1242 list_add_tail(&page->lru, list); 1243 list = &page->lru; 1244 if (is_migrate_cma(get_freepage_migratetype(page))) 1245 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1246 -(1 << order)); 1247 } 1248 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 1249 spin_unlock(&zone->lock); 1250 return i; 1251 } 1252 1253 #ifdef CONFIG_NUMA 1254 /* 1255 * Called from the vmstat counter updater to drain pagesets of this 1256 * currently executing processor on remote nodes after they have 1257 * expired. 1258 * 1259 * Note that this function must be called with the thread pinned to 1260 * a single processor. 1261 */ 1262 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1263 { 1264 unsigned long flags; 1265 int to_drain, batch; 1266 1267 local_irq_save(flags); 1268 batch = ACCESS_ONCE(pcp->batch); 1269 to_drain = min(pcp->count, batch); 1270 if (to_drain > 0) { 1271 free_pcppages_bulk(zone, to_drain, pcp); 1272 pcp->count -= to_drain; 1273 } 1274 local_irq_restore(flags); 1275 } 1276 #endif 1277 1278 /* 1279 * Drain pages of the indicated processor. 1280 * 1281 * The processor must either be the current processor and the 1282 * thread pinned to the current processor or a processor that 1283 * is not online. 1284 */ 1285 static void drain_pages(unsigned int cpu) 1286 { 1287 unsigned long flags; 1288 struct zone *zone; 1289 1290 for_each_populated_zone(zone) { 1291 struct per_cpu_pageset *pset; 1292 struct per_cpu_pages *pcp; 1293 1294 local_irq_save(flags); 1295 pset = per_cpu_ptr(zone->pageset, cpu); 1296 1297 pcp = &pset->pcp; 1298 if (pcp->count) { 1299 free_pcppages_bulk(zone, pcp->count, pcp); 1300 pcp->count = 0; 1301 } 1302 local_irq_restore(flags); 1303 } 1304 } 1305 1306 /* 1307 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1308 */ 1309 void drain_local_pages(void *arg) 1310 { 1311 drain_pages(smp_processor_id()); 1312 } 1313 1314 /* 1315 * Spill all the per-cpu pages from all CPUs back into the buddy allocator. 1316 * 1317 * Note that this code is protected against sending an IPI to an offline 1318 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs: 1319 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but 1320 * nothing keeps CPUs from showing up after we populated the cpumask and 1321 * before the call to on_each_cpu_mask(). 1322 */ 1323 void drain_all_pages(void) 1324 { 1325 int cpu; 1326 struct per_cpu_pageset *pcp; 1327 struct zone *zone; 1328 1329 /* 1330 * Allocate in the BSS so we wont require allocation in 1331 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y 1332 */ 1333 static cpumask_t cpus_with_pcps; 1334 1335 /* 1336 * We don't care about racing with CPU hotplug event 1337 * as offline notification will cause the notified 1338 * cpu to drain that CPU pcps and on_each_cpu_mask 1339 * disables preemption as part of its processing 1340 */ 1341 for_each_online_cpu(cpu) { 1342 bool has_pcps = false; 1343 for_each_populated_zone(zone) { 1344 pcp = per_cpu_ptr(zone->pageset, cpu); 1345 if (pcp->pcp.count) { 1346 has_pcps = true; 1347 break; 1348 } 1349 } 1350 if (has_pcps) 1351 cpumask_set_cpu(cpu, &cpus_with_pcps); 1352 else 1353 cpumask_clear_cpu(cpu, &cpus_with_pcps); 1354 } 1355 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1); 1356 } 1357 1358 #ifdef CONFIG_HIBERNATION 1359 1360 void mark_free_pages(struct zone *zone) 1361 { 1362 unsigned long pfn, max_zone_pfn; 1363 unsigned long flags; 1364 unsigned int order, t; 1365 struct list_head *curr; 1366 1367 if (zone_is_empty(zone)) 1368 return; 1369 1370 spin_lock_irqsave(&zone->lock, flags); 1371 1372 max_zone_pfn = zone_end_pfn(zone); 1373 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1374 if (pfn_valid(pfn)) { 1375 struct page *page = pfn_to_page(pfn); 1376 1377 if (!swsusp_page_is_forbidden(page)) 1378 swsusp_unset_page_free(page); 1379 } 1380 1381 for_each_migratetype_order(order, t) { 1382 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1383 unsigned long i; 1384 1385 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1386 for (i = 0; i < (1UL << order); i++) 1387 swsusp_set_page_free(pfn_to_page(pfn + i)); 1388 } 1389 } 1390 spin_unlock_irqrestore(&zone->lock, flags); 1391 } 1392 #endif /* CONFIG_PM */ 1393 1394 /* 1395 * Free a 0-order page 1396 * cold == true ? free a cold page : free a hot page 1397 */ 1398 void free_hot_cold_page(struct page *page, bool cold) 1399 { 1400 struct zone *zone = page_zone(page); 1401 struct per_cpu_pages *pcp; 1402 unsigned long flags; 1403 unsigned long pfn = page_to_pfn(page); 1404 int migratetype; 1405 1406 if (!free_pages_prepare(page, 0)) 1407 return; 1408 1409 migratetype = get_pfnblock_migratetype(page, pfn); 1410 set_freepage_migratetype(page, migratetype); 1411 local_irq_save(flags); 1412 __count_vm_event(PGFREE); 1413 1414 /* 1415 * We only track unmovable, reclaimable and movable on pcp lists. 1416 * Free ISOLATE pages back to the allocator because they are being 1417 * offlined but treat RESERVE as movable pages so we can get those 1418 * areas back if necessary. Otherwise, we may have to free 1419 * excessively into the page allocator 1420 */ 1421 if (migratetype >= MIGRATE_PCPTYPES) { 1422 if (unlikely(is_migrate_isolate(migratetype))) { 1423 free_one_page(zone, page, pfn, 0, migratetype); 1424 goto out; 1425 } 1426 migratetype = MIGRATE_MOVABLE; 1427 } 1428 1429 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1430 if (!cold) 1431 list_add(&page->lru, &pcp->lists[migratetype]); 1432 else 1433 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1434 pcp->count++; 1435 if (pcp->count >= pcp->high) { 1436 unsigned long batch = ACCESS_ONCE(pcp->batch); 1437 free_pcppages_bulk(zone, batch, pcp); 1438 pcp->count -= batch; 1439 } 1440 1441 out: 1442 local_irq_restore(flags); 1443 } 1444 1445 /* 1446 * Free a list of 0-order pages 1447 */ 1448 void free_hot_cold_page_list(struct list_head *list, bool cold) 1449 { 1450 struct page *page, *next; 1451 1452 list_for_each_entry_safe(page, next, list, lru) { 1453 trace_mm_page_free_batched(page, cold); 1454 free_hot_cold_page(page, cold); 1455 } 1456 } 1457 1458 /* 1459 * split_page takes a non-compound higher-order page, and splits it into 1460 * n (1<<order) sub-pages: page[0..n] 1461 * Each sub-page must be freed individually. 1462 * 1463 * Note: this is probably too low level an operation for use in drivers. 1464 * Please consult with lkml before using this in your driver. 1465 */ 1466 void split_page(struct page *page, unsigned int order) 1467 { 1468 int i; 1469 1470 VM_BUG_ON_PAGE(PageCompound(page), page); 1471 VM_BUG_ON_PAGE(!page_count(page), page); 1472 1473 #ifdef CONFIG_KMEMCHECK 1474 /* 1475 * Split shadow pages too, because free(page[0]) would 1476 * otherwise free the whole shadow. 1477 */ 1478 if (kmemcheck_page_is_tracked(page)) 1479 split_page(virt_to_page(page[0].shadow), order); 1480 #endif 1481 1482 for (i = 1; i < (1 << order); i++) 1483 set_page_refcounted(page + i); 1484 } 1485 EXPORT_SYMBOL_GPL(split_page); 1486 1487 static int __isolate_free_page(struct page *page, unsigned int order) 1488 { 1489 unsigned long watermark; 1490 struct zone *zone; 1491 int mt; 1492 1493 BUG_ON(!PageBuddy(page)); 1494 1495 zone = page_zone(page); 1496 mt = get_pageblock_migratetype(page); 1497 1498 if (!is_migrate_isolate(mt)) { 1499 /* Obey watermarks as if the page was being allocated */ 1500 watermark = low_wmark_pages(zone) + (1 << order); 1501 if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) 1502 return 0; 1503 1504 __mod_zone_freepage_state(zone, -(1UL << order), mt); 1505 } 1506 1507 /* Remove page from free list */ 1508 list_del(&page->lru); 1509 zone->free_area[order].nr_free--; 1510 rmv_page_order(page); 1511 1512 /* Set the pageblock if the isolated page is at least a pageblock */ 1513 if (order >= pageblock_order - 1) { 1514 struct page *endpage = page + (1 << order) - 1; 1515 for (; page < endpage; page += pageblock_nr_pages) { 1516 int mt = get_pageblock_migratetype(page); 1517 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)) 1518 set_pageblock_migratetype(page, 1519 MIGRATE_MOVABLE); 1520 } 1521 } 1522 1523 return 1UL << order; 1524 } 1525 1526 /* 1527 * Similar to split_page except the page is already free. As this is only 1528 * being used for migration, the migratetype of the block also changes. 1529 * As this is called with interrupts disabled, the caller is responsible 1530 * for calling arch_alloc_page() and kernel_map_page() after interrupts 1531 * are enabled. 1532 * 1533 * Note: this is probably too low level an operation for use in drivers. 1534 * Please consult with lkml before using this in your driver. 1535 */ 1536 int split_free_page(struct page *page) 1537 { 1538 unsigned int order; 1539 int nr_pages; 1540 1541 order = page_order(page); 1542 1543 nr_pages = __isolate_free_page(page, order); 1544 if (!nr_pages) 1545 return 0; 1546 1547 /* Split into individual pages */ 1548 set_page_refcounted(page); 1549 split_page(page, order); 1550 return nr_pages; 1551 } 1552 1553 /* 1554 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1555 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1556 * or two. 1557 */ 1558 static inline 1559 struct page *buffered_rmqueue(struct zone *preferred_zone, 1560 struct zone *zone, unsigned int order, 1561 gfp_t gfp_flags, int migratetype) 1562 { 1563 unsigned long flags; 1564 struct page *page; 1565 bool cold = ((gfp_flags & __GFP_COLD) != 0); 1566 1567 again: 1568 if (likely(order == 0)) { 1569 struct per_cpu_pages *pcp; 1570 struct list_head *list; 1571 1572 local_irq_save(flags); 1573 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1574 list = &pcp->lists[migratetype]; 1575 if (list_empty(list)) { 1576 pcp->count += rmqueue_bulk(zone, 0, 1577 pcp->batch, list, 1578 migratetype, cold); 1579 if (unlikely(list_empty(list))) 1580 goto failed; 1581 } 1582 1583 if (cold) 1584 page = list_entry(list->prev, struct page, lru); 1585 else 1586 page = list_entry(list->next, struct page, lru); 1587 1588 list_del(&page->lru); 1589 pcp->count--; 1590 } else { 1591 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1592 /* 1593 * __GFP_NOFAIL is not to be used in new code. 1594 * 1595 * All __GFP_NOFAIL callers should be fixed so that they 1596 * properly detect and handle allocation failures. 1597 * 1598 * We most definitely don't want callers attempting to 1599 * allocate greater than order-1 page units with 1600 * __GFP_NOFAIL. 1601 */ 1602 WARN_ON_ONCE(order > 1); 1603 } 1604 spin_lock_irqsave(&zone->lock, flags); 1605 page = __rmqueue(zone, order, migratetype); 1606 spin_unlock(&zone->lock); 1607 if (!page) 1608 goto failed; 1609 __mod_zone_freepage_state(zone, -(1 << order), 1610 get_freepage_migratetype(page)); 1611 } 1612 1613 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order)); 1614 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 && 1615 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) 1616 set_bit(ZONE_FAIR_DEPLETED, &zone->flags); 1617 1618 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1619 zone_statistics(preferred_zone, zone, gfp_flags); 1620 local_irq_restore(flags); 1621 1622 VM_BUG_ON_PAGE(bad_range(zone, page), page); 1623 if (prep_new_page(page, order, gfp_flags)) 1624 goto again; 1625 return page; 1626 1627 failed: 1628 local_irq_restore(flags); 1629 return NULL; 1630 } 1631 1632 #ifdef CONFIG_FAIL_PAGE_ALLOC 1633 1634 static struct { 1635 struct fault_attr attr; 1636 1637 u32 ignore_gfp_highmem; 1638 u32 ignore_gfp_wait; 1639 u32 min_order; 1640 } fail_page_alloc = { 1641 .attr = FAULT_ATTR_INITIALIZER, 1642 .ignore_gfp_wait = 1, 1643 .ignore_gfp_highmem = 1, 1644 .min_order = 1, 1645 }; 1646 1647 static int __init setup_fail_page_alloc(char *str) 1648 { 1649 return setup_fault_attr(&fail_page_alloc.attr, str); 1650 } 1651 __setup("fail_page_alloc=", setup_fail_page_alloc); 1652 1653 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1654 { 1655 if (order < fail_page_alloc.min_order) 1656 return false; 1657 if (gfp_mask & __GFP_NOFAIL) 1658 return false; 1659 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1660 return false; 1661 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1662 return false; 1663 1664 return should_fail(&fail_page_alloc.attr, 1 << order); 1665 } 1666 1667 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1668 1669 static int __init fail_page_alloc_debugfs(void) 1670 { 1671 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1672 struct dentry *dir; 1673 1674 dir = fault_create_debugfs_attr("fail_page_alloc", NULL, 1675 &fail_page_alloc.attr); 1676 if (IS_ERR(dir)) 1677 return PTR_ERR(dir); 1678 1679 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, 1680 &fail_page_alloc.ignore_gfp_wait)) 1681 goto fail; 1682 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1683 &fail_page_alloc.ignore_gfp_highmem)) 1684 goto fail; 1685 if (!debugfs_create_u32("min-order", mode, dir, 1686 &fail_page_alloc.min_order)) 1687 goto fail; 1688 1689 return 0; 1690 fail: 1691 debugfs_remove_recursive(dir); 1692 1693 return -ENOMEM; 1694 } 1695 1696 late_initcall(fail_page_alloc_debugfs); 1697 1698 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1699 1700 #else /* CONFIG_FAIL_PAGE_ALLOC */ 1701 1702 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1703 { 1704 return false; 1705 } 1706 1707 #endif /* CONFIG_FAIL_PAGE_ALLOC */ 1708 1709 /* 1710 * Return true if free pages are above 'mark'. This takes into account the order 1711 * of the allocation. 1712 */ 1713 static bool __zone_watermark_ok(struct zone *z, unsigned int order, 1714 unsigned long mark, int classzone_idx, int alloc_flags, 1715 long free_pages) 1716 { 1717 /* free_pages my go negative - that's OK */ 1718 long min = mark; 1719 int o; 1720 long free_cma = 0; 1721 1722 free_pages -= (1 << order) - 1; 1723 if (alloc_flags & ALLOC_HIGH) 1724 min -= min / 2; 1725 if (alloc_flags & ALLOC_HARDER) 1726 min -= min / 4; 1727 #ifdef CONFIG_CMA 1728 /* If allocation can't use CMA areas don't use free CMA pages */ 1729 if (!(alloc_flags & ALLOC_CMA)) 1730 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES); 1731 #endif 1732 1733 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx]) 1734 return false; 1735 for (o = 0; o < order; o++) { 1736 /* At the next order, this order's pages become unavailable */ 1737 free_pages -= z->free_area[o].nr_free << o; 1738 1739 /* Require fewer higher order pages to be free */ 1740 min >>= 1; 1741 1742 if (free_pages <= min) 1743 return false; 1744 } 1745 return true; 1746 } 1747 1748 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark, 1749 int classzone_idx, int alloc_flags) 1750 { 1751 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1752 zone_page_state(z, NR_FREE_PAGES)); 1753 } 1754 1755 bool zone_watermark_ok_safe(struct zone *z, unsigned int order, 1756 unsigned long mark, int classzone_idx, int alloc_flags) 1757 { 1758 long free_pages = zone_page_state(z, NR_FREE_PAGES); 1759 1760 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) 1761 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); 1762 1763 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1764 free_pages); 1765 } 1766 1767 #ifdef CONFIG_NUMA 1768 /* 1769 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1770 * skip over zones that are not allowed by the cpuset, or that have 1771 * been recently (in last second) found to be nearly full. See further 1772 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1773 * that have to skip over a lot of full or unallowed zones. 1774 * 1775 * If the zonelist cache is present in the passed zonelist, then 1776 * returns a pointer to the allowed node mask (either the current 1777 * tasks mems_allowed, or node_states[N_MEMORY].) 1778 * 1779 * If the zonelist cache is not available for this zonelist, does 1780 * nothing and returns NULL. 1781 * 1782 * If the fullzones BITMAP in the zonelist cache is stale (more than 1783 * a second since last zap'd) then we zap it out (clear its bits.) 1784 * 1785 * We hold off even calling zlc_setup, until after we've checked the 1786 * first zone in the zonelist, on the theory that most allocations will 1787 * be satisfied from that first zone, so best to examine that zone as 1788 * quickly as we can. 1789 */ 1790 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1791 { 1792 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1793 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1794 1795 zlc = zonelist->zlcache_ptr; 1796 if (!zlc) 1797 return NULL; 1798 1799 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1800 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1801 zlc->last_full_zap = jiffies; 1802 } 1803 1804 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1805 &cpuset_current_mems_allowed : 1806 &node_states[N_MEMORY]; 1807 return allowednodes; 1808 } 1809 1810 /* 1811 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1812 * if it is worth looking at further for free memory: 1813 * 1) Check that the zone isn't thought to be full (doesn't have its 1814 * bit set in the zonelist_cache fullzones BITMAP). 1815 * 2) Check that the zones node (obtained from the zonelist_cache 1816 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1817 * Return true (non-zero) if zone is worth looking at further, or 1818 * else return false (zero) if it is not. 1819 * 1820 * This check -ignores- the distinction between various watermarks, 1821 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1822 * found to be full for any variation of these watermarks, it will 1823 * be considered full for up to one second by all requests, unless 1824 * we are so low on memory on all allowed nodes that we are forced 1825 * into the second scan of the zonelist. 1826 * 1827 * In the second scan we ignore this zonelist cache and exactly 1828 * apply the watermarks to all zones, even it is slower to do so. 1829 * We are low on memory in the second scan, and should leave no stone 1830 * unturned looking for a free page. 1831 */ 1832 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1833 nodemask_t *allowednodes) 1834 { 1835 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1836 int i; /* index of *z in zonelist zones */ 1837 int n; /* node that zone *z is on */ 1838 1839 zlc = zonelist->zlcache_ptr; 1840 if (!zlc) 1841 return 1; 1842 1843 i = z - zonelist->_zonerefs; 1844 n = zlc->z_to_n[i]; 1845 1846 /* This zone is worth trying if it is allowed but not full */ 1847 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1848 } 1849 1850 /* 1851 * Given 'z' scanning a zonelist, set the corresponding bit in 1852 * zlc->fullzones, so that subsequent attempts to allocate a page 1853 * from that zone don't waste time re-examining it. 1854 */ 1855 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1856 { 1857 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1858 int i; /* index of *z in zonelist zones */ 1859 1860 zlc = zonelist->zlcache_ptr; 1861 if (!zlc) 1862 return; 1863 1864 i = z - zonelist->_zonerefs; 1865 1866 set_bit(i, zlc->fullzones); 1867 } 1868 1869 /* 1870 * clear all zones full, called after direct reclaim makes progress so that 1871 * a zone that was recently full is not skipped over for up to a second 1872 */ 1873 static void zlc_clear_zones_full(struct zonelist *zonelist) 1874 { 1875 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1876 1877 zlc = zonelist->zlcache_ptr; 1878 if (!zlc) 1879 return; 1880 1881 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1882 } 1883 1884 static bool zone_local(struct zone *local_zone, struct zone *zone) 1885 { 1886 return local_zone->node == zone->node; 1887 } 1888 1889 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) 1890 { 1891 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) < 1892 RECLAIM_DISTANCE; 1893 } 1894 1895 #else /* CONFIG_NUMA */ 1896 1897 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1898 { 1899 return NULL; 1900 } 1901 1902 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1903 nodemask_t *allowednodes) 1904 { 1905 return 1; 1906 } 1907 1908 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1909 { 1910 } 1911 1912 static void zlc_clear_zones_full(struct zonelist *zonelist) 1913 { 1914 } 1915 1916 static bool zone_local(struct zone *local_zone, struct zone *zone) 1917 { 1918 return true; 1919 } 1920 1921 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) 1922 { 1923 return true; 1924 } 1925 1926 #endif /* CONFIG_NUMA */ 1927 1928 static void reset_alloc_batches(struct zone *preferred_zone) 1929 { 1930 struct zone *zone = preferred_zone->zone_pgdat->node_zones; 1931 1932 do { 1933 mod_zone_page_state(zone, NR_ALLOC_BATCH, 1934 high_wmark_pages(zone) - low_wmark_pages(zone) - 1935 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH])); 1936 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags); 1937 } while (zone++ != preferred_zone); 1938 } 1939 1940 /* 1941 * get_page_from_freelist goes through the zonelist trying to allocate 1942 * a page. 1943 */ 1944 static struct page * 1945 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1946 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1947 struct zone *preferred_zone, int classzone_idx, int migratetype) 1948 { 1949 struct zoneref *z; 1950 struct page *page = NULL; 1951 struct zone *zone; 1952 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1953 int zlc_active = 0; /* set if using zonelist_cache */ 1954 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1955 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) && 1956 (gfp_mask & __GFP_WRITE); 1957 int nr_fair_skipped = 0; 1958 bool zonelist_rescan; 1959 1960 zonelist_scan: 1961 zonelist_rescan = false; 1962 1963 /* 1964 * Scan zonelist, looking for a zone with enough free. 1965 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c. 1966 */ 1967 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1968 high_zoneidx, nodemask) { 1969 unsigned long mark; 1970 1971 if (IS_ENABLED(CONFIG_NUMA) && zlc_active && 1972 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1973 continue; 1974 if (cpusets_enabled() && 1975 (alloc_flags & ALLOC_CPUSET) && 1976 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1977 continue; 1978 /* 1979 * Distribute pages in proportion to the individual 1980 * zone size to ensure fair page aging. The zone a 1981 * page was allocated in should have no effect on the 1982 * time the page has in memory before being reclaimed. 1983 */ 1984 if (alloc_flags & ALLOC_FAIR) { 1985 if (!zone_local(preferred_zone, zone)) 1986 break; 1987 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) { 1988 nr_fair_skipped++; 1989 continue; 1990 } 1991 } 1992 /* 1993 * When allocating a page cache page for writing, we 1994 * want to get it from a zone that is within its dirty 1995 * limit, such that no single zone holds more than its 1996 * proportional share of globally allowed dirty pages. 1997 * The dirty limits take into account the zone's 1998 * lowmem reserves and high watermark so that kswapd 1999 * should be able to balance it without having to 2000 * write pages from its LRU list. 2001 * 2002 * This may look like it could increase pressure on 2003 * lower zones by failing allocations in higher zones 2004 * before they are full. But the pages that do spill 2005 * over are limited as the lower zones are protected 2006 * by this very same mechanism. It should not become 2007 * a practical burden to them. 2008 * 2009 * XXX: For now, allow allocations to potentially 2010 * exceed the per-zone dirty limit in the slowpath 2011 * (ALLOC_WMARK_LOW unset) before going into reclaim, 2012 * which is important when on a NUMA setup the allowed 2013 * zones are together not big enough to reach the 2014 * global limit. The proper fix for these situations 2015 * will require awareness of zones in the 2016 * dirty-throttling and the flusher threads. 2017 */ 2018 if (consider_zone_dirty && !zone_dirty_ok(zone)) 2019 continue; 2020 2021 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 2022 if (!zone_watermark_ok(zone, order, mark, 2023 classzone_idx, alloc_flags)) { 2024 int ret; 2025 2026 /* Checked here to keep the fast path fast */ 2027 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 2028 if (alloc_flags & ALLOC_NO_WATERMARKS) 2029 goto try_this_zone; 2030 2031 if (IS_ENABLED(CONFIG_NUMA) && 2032 !did_zlc_setup && nr_online_nodes > 1) { 2033 /* 2034 * we do zlc_setup if there are multiple nodes 2035 * and before considering the first zone allowed 2036 * by the cpuset. 2037 */ 2038 allowednodes = zlc_setup(zonelist, alloc_flags); 2039 zlc_active = 1; 2040 did_zlc_setup = 1; 2041 } 2042 2043 if (zone_reclaim_mode == 0 || 2044 !zone_allows_reclaim(preferred_zone, zone)) 2045 goto this_zone_full; 2046 2047 /* 2048 * As we may have just activated ZLC, check if the first 2049 * eligible zone has failed zone_reclaim recently. 2050 */ 2051 if (IS_ENABLED(CONFIG_NUMA) && zlc_active && 2052 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 2053 continue; 2054 2055 ret = zone_reclaim(zone, gfp_mask, order); 2056 switch (ret) { 2057 case ZONE_RECLAIM_NOSCAN: 2058 /* did not scan */ 2059 continue; 2060 case ZONE_RECLAIM_FULL: 2061 /* scanned but unreclaimable */ 2062 continue; 2063 default: 2064 /* did we reclaim enough */ 2065 if (zone_watermark_ok(zone, order, mark, 2066 classzone_idx, alloc_flags)) 2067 goto try_this_zone; 2068 2069 /* 2070 * Failed to reclaim enough to meet watermark. 2071 * Only mark the zone full if checking the min 2072 * watermark or if we failed to reclaim just 2073 * 1<<order pages or else the page allocator 2074 * fastpath will prematurely mark zones full 2075 * when the watermark is between the low and 2076 * min watermarks. 2077 */ 2078 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) || 2079 ret == ZONE_RECLAIM_SOME) 2080 goto this_zone_full; 2081 2082 continue; 2083 } 2084 } 2085 2086 try_this_zone: 2087 page = buffered_rmqueue(preferred_zone, zone, order, 2088 gfp_mask, migratetype); 2089 if (page) 2090 break; 2091 this_zone_full: 2092 if (IS_ENABLED(CONFIG_NUMA) && zlc_active) 2093 zlc_mark_zone_full(zonelist, z); 2094 } 2095 2096 if (page) { 2097 /* 2098 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was 2099 * necessary to allocate the page. The expectation is 2100 * that the caller is taking steps that will free more 2101 * memory. The caller should avoid the page being used 2102 * for !PFMEMALLOC purposes. 2103 */ 2104 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS); 2105 return page; 2106 } 2107 2108 /* 2109 * The first pass makes sure allocations are spread fairly within the 2110 * local node. However, the local node might have free pages left 2111 * after the fairness batches are exhausted, and remote zones haven't 2112 * even been considered yet. Try once more without fairness, and 2113 * include remote zones now, before entering the slowpath and waking 2114 * kswapd: prefer spilling to a remote zone over swapping locally. 2115 */ 2116 if (alloc_flags & ALLOC_FAIR) { 2117 alloc_flags &= ~ALLOC_FAIR; 2118 if (nr_fair_skipped) { 2119 zonelist_rescan = true; 2120 reset_alloc_batches(preferred_zone); 2121 } 2122 if (nr_online_nodes > 1) 2123 zonelist_rescan = true; 2124 } 2125 2126 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) { 2127 /* Disable zlc cache for second zonelist scan */ 2128 zlc_active = 0; 2129 zonelist_rescan = true; 2130 } 2131 2132 if (zonelist_rescan) 2133 goto zonelist_scan; 2134 2135 return NULL; 2136 } 2137 2138 /* 2139 * Large machines with many possible nodes should not always dump per-node 2140 * meminfo in irq context. 2141 */ 2142 static inline bool should_suppress_show_mem(void) 2143 { 2144 bool ret = false; 2145 2146 #if NODES_SHIFT > 8 2147 ret = in_interrupt(); 2148 #endif 2149 return ret; 2150 } 2151 2152 static DEFINE_RATELIMIT_STATE(nopage_rs, 2153 DEFAULT_RATELIMIT_INTERVAL, 2154 DEFAULT_RATELIMIT_BURST); 2155 2156 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) 2157 { 2158 unsigned int filter = SHOW_MEM_FILTER_NODES; 2159 2160 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) || 2161 debug_guardpage_minorder() > 0) 2162 return; 2163 2164 /* 2165 * This documents exceptions given to allocations in certain 2166 * contexts that are allowed to allocate outside current's set 2167 * of allowed nodes. 2168 */ 2169 if (!(gfp_mask & __GFP_NOMEMALLOC)) 2170 if (test_thread_flag(TIF_MEMDIE) || 2171 (current->flags & (PF_MEMALLOC | PF_EXITING))) 2172 filter &= ~SHOW_MEM_FILTER_NODES; 2173 if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) 2174 filter &= ~SHOW_MEM_FILTER_NODES; 2175 2176 if (fmt) { 2177 struct va_format vaf; 2178 va_list args; 2179 2180 va_start(args, fmt); 2181 2182 vaf.fmt = fmt; 2183 vaf.va = &args; 2184 2185 pr_warn("%pV", &vaf); 2186 2187 va_end(args); 2188 } 2189 2190 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n", 2191 current->comm, order, gfp_mask); 2192 2193 dump_stack(); 2194 if (!should_suppress_show_mem()) 2195 show_mem(filter); 2196 } 2197 2198 static inline int 2199 should_alloc_retry(gfp_t gfp_mask, unsigned int order, 2200 unsigned long did_some_progress, 2201 unsigned long pages_reclaimed) 2202 { 2203 /* Do not loop if specifically requested */ 2204 if (gfp_mask & __GFP_NORETRY) 2205 return 0; 2206 2207 /* Always retry if specifically requested */ 2208 if (gfp_mask & __GFP_NOFAIL) 2209 return 1; 2210 2211 /* 2212 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim 2213 * making forward progress without invoking OOM. Suspend also disables 2214 * storage devices so kswapd will not help. Bail if we are suspending. 2215 */ 2216 if (!did_some_progress && pm_suspended_storage()) 2217 return 0; 2218 2219 /* 2220 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 2221 * means __GFP_NOFAIL, but that may not be true in other 2222 * implementations. 2223 */ 2224 if (order <= PAGE_ALLOC_COSTLY_ORDER) 2225 return 1; 2226 2227 /* 2228 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 2229 * specified, then we retry until we no longer reclaim any pages 2230 * (above), or we've reclaimed an order of pages at least as 2231 * large as the allocation's order. In both cases, if the 2232 * allocation still fails, we stop retrying. 2233 */ 2234 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 2235 return 1; 2236 2237 return 0; 2238 } 2239 2240 static inline struct page * 2241 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 2242 struct zonelist *zonelist, enum zone_type high_zoneidx, 2243 nodemask_t *nodemask, struct zone *preferred_zone, 2244 int classzone_idx, int migratetype) 2245 { 2246 struct page *page; 2247 2248 /* Acquire the per-zone oom lock for each zone */ 2249 if (!oom_zonelist_trylock(zonelist, gfp_mask)) { 2250 schedule_timeout_uninterruptible(1); 2251 return NULL; 2252 } 2253 2254 /* 2255 * PM-freezer should be notified that there might be an OOM killer on 2256 * its way to kill and wake somebody up. This is too early and we might 2257 * end up not killing anything but false positives are acceptable. 2258 * See freeze_processes. 2259 */ 2260 note_oom_kill(); 2261 2262 /* 2263 * Go through the zonelist yet one more time, keep very high watermark 2264 * here, this is only to catch a parallel oom killing, we must fail if 2265 * we're still under heavy pressure. 2266 */ 2267 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 2268 order, zonelist, high_zoneidx, 2269 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 2270 preferred_zone, classzone_idx, migratetype); 2271 if (page) 2272 goto out; 2273 2274 if (!(gfp_mask & __GFP_NOFAIL)) { 2275 /* The OOM killer will not help higher order allocs */ 2276 if (order > PAGE_ALLOC_COSTLY_ORDER) 2277 goto out; 2278 /* The OOM killer does not needlessly kill tasks for lowmem */ 2279 if (high_zoneidx < ZONE_NORMAL) 2280 goto out; 2281 /* 2282 * GFP_THISNODE contains __GFP_NORETRY and we never hit this. 2283 * Sanity check for bare calls of __GFP_THISNODE, not real OOM. 2284 * The caller should handle page allocation failure by itself if 2285 * it specifies __GFP_THISNODE. 2286 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. 2287 */ 2288 if (gfp_mask & __GFP_THISNODE) 2289 goto out; 2290 } 2291 /* Exhausted what can be done so it's blamo time */ 2292 out_of_memory(zonelist, gfp_mask, order, nodemask, false); 2293 2294 out: 2295 oom_zonelist_unlock(zonelist, gfp_mask); 2296 return page; 2297 } 2298 2299 #ifdef CONFIG_COMPACTION 2300 /* Try memory compaction for high-order allocations before reclaim */ 2301 static struct page * 2302 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 2303 struct zonelist *zonelist, enum zone_type high_zoneidx, 2304 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 2305 int classzone_idx, int migratetype, enum migrate_mode mode, 2306 int *contended_compaction, bool *deferred_compaction) 2307 { 2308 struct zone *last_compact_zone = NULL; 2309 unsigned long compact_result; 2310 struct page *page; 2311 2312 if (!order) 2313 return NULL; 2314 2315 current->flags |= PF_MEMALLOC; 2316 compact_result = try_to_compact_pages(zonelist, order, gfp_mask, 2317 nodemask, mode, 2318 contended_compaction, 2319 &last_compact_zone); 2320 current->flags &= ~PF_MEMALLOC; 2321 2322 switch (compact_result) { 2323 case COMPACT_DEFERRED: 2324 *deferred_compaction = true; 2325 /* fall-through */ 2326 case COMPACT_SKIPPED: 2327 return NULL; 2328 default: 2329 break; 2330 } 2331 2332 /* 2333 * At least in one zone compaction wasn't deferred or skipped, so let's 2334 * count a compaction stall 2335 */ 2336 count_vm_event(COMPACTSTALL); 2337 2338 /* Page migration frees to the PCP lists but we want merging */ 2339 drain_pages(get_cpu()); 2340 put_cpu(); 2341 2342 page = get_page_from_freelist(gfp_mask, nodemask, 2343 order, zonelist, high_zoneidx, 2344 alloc_flags & ~ALLOC_NO_WATERMARKS, 2345 preferred_zone, classzone_idx, migratetype); 2346 2347 if (page) { 2348 struct zone *zone = page_zone(page); 2349 2350 zone->compact_blockskip_flush = false; 2351 compaction_defer_reset(zone, order, true); 2352 count_vm_event(COMPACTSUCCESS); 2353 return page; 2354 } 2355 2356 /* 2357 * last_compact_zone is where try_to_compact_pages thought allocation 2358 * should succeed, so it did not defer compaction. But here we know 2359 * that it didn't succeed, so we do the defer. 2360 */ 2361 if (last_compact_zone && mode != MIGRATE_ASYNC) 2362 defer_compaction(last_compact_zone, order); 2363 2364 /* 2365 * It's bad if compaction run occurs and fails. The most likely reason 2366 * is that pages exist, but not enough to satisfy watermarks. 2367 */ 2368 count_vm_event(COMPACTFAIL); 2369 2370 cond_resched(); 2371 2372 return NULL; 2373 } 2374 #else 2375 static inline struct page * 2376 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 2377 struct zonelist *zonelist, enum zone_type high_zoneidx, 2378 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 2379 int classzone_idx, int migratetype, enum migrate_mode mode, 2380 int *contended_compaction, bool *deferred_compaction) 2381 { 2382 return NULL; 2383 } 2384 #endif /* CONFIG_COMPACTION */ 2385 2386 /* Perform direct synchronous page reclaim */ 2387 static int 2388 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist, 2389 nodemask_t *nodemask) 2390 { 2391 struct reclaim_state reclaim_state; 2392 int progress; 2393 2394 cond_resched(); 2395 2396 /* We now go into synchronous reclaim */ 2397 cpuset_memory_pressure_bump(); 2398 current->flags |= PF_MEMALLOC; 2399 lockdep_set_current_reclaim_state(gfp_mask); 2400 reclaim_state.reclaimed_slab = 0; 2401 current->reclaim_state = &reclaim_state; 2402 2403 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 2404 2405 current->reclaim_state = NULL; 2406 lockdep_clear_current_reclaim_state(); 2407 current->flags &= ~PF_MEMALLOC; 2408 2409 cond_resched(); 2410 2411 return progress; 2412 } 2413 2414 /* The really slow allocator path where we enter direct reclaim */ 2415 static inline struct page * 2416 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 2417 struct zonelist *zonelist, enum zone_type high_zoneidx, 2418 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 2419 int classzone_idx, int migratetype, unsigned long *did_some_progress) 2420 { 2421 struct page *page = NULL; 2422 bool drained = false; 2423 2424 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist, 2425 nodemask); 2426 if (unlikely(!(*did_some_progress))) 2427 return NULL; 2428 2429 /* After successful reclaim, reconsider all zones for allocation */ 2430 if (IS_ENABLED(CONFIG_NUMA)) 2431 zlc_clear_zones_full(zonelist); 2432 2433 retry: 2434 page = get_page_from_freelist(gfp_mask, nodemask, order, 2435 zonelist, high_zoneidx, 2436 alloc_flags & ~ALLOC_NO_WATERMARKS, 2437 preferred_zone, classzone_idx, 2438 migratetype); 2439 2440 /* 2441 * If an allocation failed after direct reclaim, it could be because 2442 * pages are pinned on the per-cpu lists. Drain them and try again 2443 */ 2444 if (!page && !drained) { 2445 drain_all_pages(); 2446 drained = true; 2447 goto retry; 2448 } 2449 2450 return page; 2451 } 2452 2453 /* 2454 * This is called in the allocator slow-path if the allocation request is of 2455 * sufficient urgency to ignore watermarks and take other desperate measures 2456 */ 2457 static inline struct page * 2458 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 2459 struct zonelist *zonelist, enum zone_type high_zoneidx, 2460 nodemask_t *nodemask, struct zone *preferred_zone, 2461 int classzone_idx, int migratetype) 2462 { 2463 struct page *page; 2464 2465 do { 2466 page = get_page_from_freelist(gfp_mask, nodemask, order, 2467 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 2468 preferred_zone, classzone_idx, migratetype); 2469 2470 if (!page && gfp_mask & __GFP_NOFAIL) 2471 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2472 } while (!page && (gfp_mask & __GFP_NOFAIL)); 2473 2474 return page; 2475 } 2476 2477 static void wake_all_kswapds(unsigned int order, 2478 struct zonelist *zonelist, 2479 enum zone_type high_zoneidx, 2480 struct zone *preferred_zone, 2481 nodemask_t *nodemask) 2482 { 2483 struct zoneref *z; 2484 struct zone *zone; 2485 2486 for_each_zone_zonelist_nodemask(zone, z, zonelist, 2487 high_zoneidx, nodemask) 2488 wakeup_kswapd(zone, order, zone_idx(preferred_zone)); 2489 } 2490 2491 static inline int 2492 gfp_to_alloc_flags(gfp_t gfp_mask) 2493 { 2494 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 2495 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD)); 2496 2497 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 2498 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); 2499 2500 /* 2501 * The caller may dip into page reserves a bit more if the caller 2502 * cannot run direct reclaim, or if the caller has realtime scheduling 2503 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 2504 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH). 2505 */ 2506 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); 2507 2508 if (atomic) { 2509 /* 2510 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even 2511 * if it can't schedule. 2512 */ 2513 if (!(gfp_mask & __GFP_NOMEMALLOC)) 2514 alloc_flags |= ALLOC_HARDER; 2515 /* 2516 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the 2517 * comment for __cpuset_node_allowed_softwall(). 2518 */ 2519 alloc_flags &= ~ALLOC_CPUSET; 2520 } else if (unlikely(rt_task(current)) && !in_interrupt()) 2521 alloc_flags |= ALLOC_HARDER; 2522 2523 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 2524 if (gfp_mask & __GFP_MEMALLOC) 2525 alloc_flags |= ALLOC_NO_WATERMARKS; 2526 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC)) 2527 alloc_flags |= ALLOC_NO_WATERMARKS; 2528 else if (!in_interrupt() && 2529 ((current->flags & PF_MEMALLOC) || 2530 unlikely(test_thread_flag(TIF_MEMDIE)))) 2531 alloc_flags |= ALLOC_NO_WATERMARKS; 2532 } 2533 #ifdef CONFIG_CMA 2534 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE) 2535 alloc_flags |= ALLOC_CMA; 2536 #endif 2537 return alloc_flags; 2538 } 2539 2540 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask) 2541 { 2542 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS); 2543 } 2544 2545 static inline struct page * 2546 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 2547 struct zonelist *zonelist, enum zone_type high_zoneidx, 2548 nodemask_t *nodemask, struct zone *preferred_zone, 2549 int classzone_idx, int migratetype) 2550 { 2551 const gfp_t wait = gfp_mask & __GFP_WAIT; 2552 struct page *page = NULL; 2553 int alloc_flags; 2554 unsigned long pages_reclaimed = 0; 2555 unsigned long did_some_progress; 2556 enum migrate_mode migration_mode = MIGRATE_ASYNC; 2557 bool deferred_compaction = false; 2558 int contended_compaction = COMPACT_CONTENDED_NONE; 2559 2560 /* 2561 * In the slowpath, we sanity check order to avoid ever trying to 2562 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 2563 * be using allocators in order of preference for an area that is 2564 * too large. 2565 */ 2566 if (order >= MAX_ORDER) { 2567 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 2568 return NULL; 2569 } 2570 2571 /* 2572 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 2573 * __GFP_NOWARN set) should not cause reclaim since the subsystem 2574 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 2575 * using a larger set of nodes after it has established that the 2576 * allowed per node queues are empty and that nodes are 2577 * over allocated. 2578 */ 2579 if (IS_ENABLED(CONFIG_NUMA) && 2580 (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 2581 goto nopage; 2582 2583 restart: 2584 if (!(gfp_mask & __GFP_NO_KSWAPD)) 2585 wake_all_kswapds(order, zonelist, high_zoneidx, 2586 preferred_zone, nodemask); 2587 2588 /* 2589 * OK, we're below the kswapd watermark and have kicked background 2590 * reclaim. Now things get more complex, so set up alloc_flags according 2591 * to how we want to proceed. 2592 */ 2593 alloc_flags = gfp_to_alloc_flags(gfp_mask); 2594 2595 /* 2596 * Find the true preferred zone if the allocation is unconstrained by 2597 * cpusets. 2598 */ 2599 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) { 2600 struct zoneref *preferred_zoneref; 2601 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx, 2602 NULL, &preferred_zone); 2603 classzone_idx = zonelist_zone_idx(preferred_zoneref); 2604 } 2605 2606 rebalance: 2607 /* This is the last chance, in general, before the goto nopage. */ 2608 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 2609 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 2610 preferred_zone, classzone_idx, migratetype); 2611 if (page) 2612 goto got_pg; 2613 2614 /* Allocate without watermarks if the context allows */ 2615 if (alloc_flags & ALLOC_NO_WATERMARKS) { 2616 /* 2617 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds 2618 * the allocation is high priority and these type of 2619 * allocations are system rather than user orientated 2620 */ 2621 zonelist = node_zonelist(numa_node_id(), gfp_mask); 2622 2623 page = __alloc_pages_high_priority(gfp_mask, order, 2624 zonelist, high_zoneidx, nodemask, 2625 preferred_zone, classzone_idx, migratetype); 2626 if (page) { 2627 goto got_pg; 2628 } 2629 } 2630 2631 /* Atomic allocations - we can't balance anything */ 2632 if (!wait) { 2633 /* 2634 * All existing users of the deprecated __GFP_NOFAIL are 2635 * blockable, so warn of any new users that actually allow this 2636 * type of allocation to fail. 2637 */ 2638 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL); 2639 goto nopage; 2640 } 2641 2642 /* Avoid recursion of direct reclaim */ 2643 if (current->flags & PF_MEMALLOC) 2644 goto nopage; 2645 2646 /* Avoid allocations with no watermarks from looping endlessly */ 2647 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 2648 goto nopage; 2649 2650 /* 2651 * Try direct compaction. The first pass is asynchronous. Subsequent 2652 * attempts after direct reclaim are synchronous 2653 */ 2654 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist, 2655 high_zoneidx, nodemask, alloc_flags, 2656 preferred_zone, 2657 classzone_idx, migratetype, 2658 migration_mode, &contended_compaction, 2659 &deferred_compaction); 2660 if (page) 2661 goto got_pg; 2662 2663 /* Checks for THP-specific high-order allocations */ 2664 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) { 2665 /* 2666 * If compaction is deferred for high-order allocations, it is 2667 * because sync compaction recently failed. If this is the case 2668 * and the caller requested a THP allocation, we do not want 2669 * to heavily disrupt the system, so we fail the allocation 2670 * instead of entering direct reclaim. 2671 */ 2672 if (deferred_compaction) 2673 goto nopage; 2674 2675 /* 2676 * In all zones where compaction was attempted (and not 2677 * deferred or skipped), lock contention has been detected. 2678 * For THP allocation we do not want to disrupt the others 2679 * so we fallback to base pages instead. 2680 */ 2681 if (contended_compaction == COMPACT_CONTENDED_LOCK) 2682 goto nopage; 2683 2684 /* 2685 * If compaction was aborted due to need_resched(), we do not 2686 * want to further increase allocation latency, unless it is 2687 * khugepaged trying to collapse. 2688 */ 2689 if (contended_compaction == COMPACT_CONTENDED_SCHED 2690 && !(current->flags & PF_KTHREAD)) 2691 goto nopage; 2692 } 2693 2694 /* 2695 * It can become very expensive to allocate transparent hugepages at 2696 * fault, so use asynchronous memory compaction for THP unless it is 2697 * khugepaged trying to collapse. 2698 */ 2699 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE || 2700 (current->flags & PF_KTHREAD)) 2701 migration_mode = MIGRATE_SYNC_LIGHT; 2702 2703 /* Try direct reclaim and then allocating */ 2704 page = __alloc_pages_direct_reclaim(gfp_mask, order, 2705 zonelist, high_zoneidx, 2706 nodemask, 2707 alloc_flags, preferred_zone, 2708 classzone_idx, migratetype, 2709 &did_some_progress); 2710 if (page) 2711 goto got_pg; 2712 2713 /* 2714 * If we failed to make any progress reclaiming, then we are 2715 * running out of options and have to consider going OOM 2716 */ 2717 if (!did_some_progress) { 2718 if (oom_gfp_allowed(gfp_mask)) { 2719 if (oom_killer_disabled) 2720 goto nopage; 2721 /* Coredumps can quickly deplete all memory reserves */ 2722 if ((current->flags & PF_DUMPCORE) && 2723 !(gfp_mask & __GFP_NOFAIL)) 2724 goto nopage; 2725 page = __alloc_pages_may_oom(gfp_mask, order, 2726 zonelist, high_zoneidx, 2727 nodemask, preferred_zone, 2728 classzone_idx, migratetype); 2729 if (page) 2730 goto got_pg; 2731 2732 if (!(gfp_mask & __GFP_NOFAIL)) { 2733 /* 2734 * The oom killer is not called for high-order 2735 * allocations that may fail, so if no progress 2736 * is being made, there are no other options and 2737 * retrying is unlikely to help. 2738 */ 2739 if (order > PAGE_ALLOC_COSTLY_ORDER) 2740 goto nopage; 2741 /* 2742 * The oom killer is not called for lowmem 2743 * allocations to prevent needlessly killing 2744 * innocent tasks. 2745 */ 2746 if (high_zoneidx < ZONE_NORMAL) 2747 goto nopage; 2748 } 2749 2750 goto restart; 2751 } 2752 } 2753 2754 /* Check if we should retry the allocation */ 2755 pages_reclaimed += did_some_progress; 2756 if (should_alloc_retry(gfp_mask, order, did_some_progress, 2757 pages_reclaimed)) { 2758 /* Wait for some write requests to complete then retry */ 2759 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2760 goto rebalance; 2761 } else { 2762 /* 2763 * High-order allocations do not necessarily loop after 2764 * direct reclaim and reclaim/compaction depends on compaction 2765 * being called after reclaim so call directly if necessary 2766 */ 2767 page = __alloc_pages_direct_compact(gfp_mask, order, zonelist, 2768 high_zoneidx, nodemask, alloc_flags, 2769 preferred_zone, 2770 classzone_idx, migratetype, 2771 migration_mode, &contended_compaction, 2772 &deferred_compaction); 2773 if (page) 2774 goto got_pg; 2775 } 2776 2777 nopage: 2778 warn_alloc_failed(gfp_mask, order, NULL); 2779 return page; 2780 got_pg: 2781 if (kmemcheck_enabled) 2782 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 2783 2784 return page; 2785 } 2786 2787 /* 2788 * This is the 'heart' of the zoned buddy allocator. 2789 */ 2790 struct page * 2791 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 2792 struct zonelist *zonelist, nodemask_t *nodemask) 2793 { 2794 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 2795 struct zone *preferred_zone; 2796 struct zoneref *preferred_zoneref; 2797 struct page *page = NULL; 2798 int migratetype = gfpflags_to_migratetype(gfp_mask); 2799 unsigned int cpuset_mems_cookie; 2800 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR; 2801 int classzone_idx; 2802 2803 gfp_mask &= gfp_allowed_mask; 2804 2805 lockdep_trace_alloc(gfp_mask); 2806 2807 might_sleep_if(gfp_mask & __GFP_WAIT); 2808 2809 if (should_fail_alloc_page(gfp_mask, order)) 2810 return NULL; 2811 2812 /* 2813 * Check the zones suitable for the gfp_mask contain at least one 2814 * valid zone. It's possible to have an empty zonelist as a result 2815 * of GFP_THISNODE and a memoryless node 2816 */ 2817 if (unlikely(!zonelist->_zonerefs->zone)) 2818 return NULL; 2819 2820 if (IS_ENABLED(CONFIG_CMA) && migratetype == MIGRATE_MOVABLE) 2821 alloc_flags |= ALLOC_CMA; 2822 2823 retry_cpuset: 2824 cpuset_mems_cookie = read_mems_allowed_begin(); 2825 2826 /* The preferred zone is used for statistics later */ 2827 preferred_zoneref = first_zones_zonelist(zonelist, high_zoneidx, 2828 nodemask ? : &cpuset_current_mems_allowed, 2829 &preferred_zone); 2830 if (!preferred_zone) 2831 goto out; 2832 classzone_idx = zonelist_zone_idx(preferred_zoneref); 2833 2834 /* First allocation attempt */ 2835 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 2836 zonelist, high_zoneidx, alloc_flags, 2837 preferred_zone, classzone_idx, migratetype); 2838 if (unlikely(!page)) { 2839 /* 2840 * Runtime PM, block IO and its error handling path 2841 * can deadlock because I/O on the device might not 2842 * complete. 2843 */ 2844 gfp_mask = memalloc_noio_flags(gfp_mask); 2845 page = __alloc_pages_slowpath(gfp_mask, order, 2846 zonelist, high_zoneidx, nodemask, 2847 preferred_zone, classzone_idx, migratetype); 2848 } 2849 2850 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 2851 2852 out: 2853 /* 2854 * When updating a task's mems_allowed, it is possible to race with 2855 * parallel threads in such a way that an allocation can fail while 2856 * the mask is being updated. If a page allocation is about to fail, 2857 * check if the cpuset changed during allocation and if so, retry. 2858 */ 2859 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 2860 goto retry_cpuset; 2861 2862 return page; 2863 } 2864 EXPORT_SYMBOL(__alloc_pages_nodemask); 2865 2866 /* 2867 * Common helper functions. 2868 */ 2869 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 2870 { 2871 struct page *page; 2872 2873 /* 2874 * __get_free_pages() returns a 32-bit address, which cannot represent 2875 * a highmem page 2876 */ 2877 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 2878 2879 page = alloc_pages(gfp_mask, order); 2880 if (!page) 2881 return 0; 2882 return (unsigned long) page_address(page); 2883 } 2884 EXPORT_SYMBOL(__get_free_pages); 2885 2886 unsigned long get_zeroed_page(gfp_t gfp_mask) 2887 { 2888 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 2889 } 2890 EXPORT_SYMBOL(get_zeroed_page); 2891 2892 void __free_pages(struct page *page, unsigned int order) 2893 { 2894 if (put_page_testzero(page)) { 2895 if (order == 0) 2896 free_hot_cold_page(page, false); 2897 else 2898 __free_pages_ok(page, order); 2899 } 2900 } 2901 2902 EXPORT_SYMBOL(__free_pages); 2903 2904 void free_pages(unsigned long addr, unsigned int order) 2905 { 2906 if (addr != 0) { 2907 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2908 __free_pages(virt_to_page((void *)addr), order); 2909 } 2910 } 2911 2912 EXPORT_SYMBOL(free_pages); 2913 2914 /* 2915 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter 2916 * of the current memory cgroup. 2917 * 2918 * It should be used when the caller would like to use kmalloc, but since the 2919 * allocation is large, it has to fall back to the page allocator. 2920 */ 2921 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order) 2922 { 2923 struct page *page; 2924 struct mem_cgroup *memcg = NULL; 2925 2926 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) 2927 return NULL; 2928 page = alloc_pages(gfp_mask, order); 2929 memcg_kmem_commit_charge(page, memcg, order); 2930 return page; 2931 } 2932 2933 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order) 2934 { 2935 struct page *page; 2936 struct mem_cgroup *memcg = NULL; 2937 2938 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) 2939 return NULL; 2940 page = alloc_pages_node(nid, gfp_mask, order); 2941 memcg_kmem_commit_charge(page, memcg, order); 2942 return page; 2943 } 2944 2945 /* 2946 * __free_kmem_pages and free_kmem_pages will free pages allocated with 2947 * alloc_kmem_pages. 2948 */ 2949 void __free_kmem_pages(struct page *page, unsigned int order) 2950 { 2951 memcg_kmem_uncharge_pages(page, order); 2952 __free_pages(page, order); 2953 } 2954 2955 void free_kmem_pages(unsigned long addr, unsigned int order) 2956 { 2957 if (addr != 0) { 2958 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2959 __free_kmem_pages(virt_to_page((void *)addr), order); 2960 } 2961 } 2962 2963 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) 2964 { 2965 if (addr) { 2966 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2967 unsigned long used = addr + PAGE_ALIGN(size); 2968 2969 split_page(virt_to_page((void *)addr), order); 2970 while (used < alloc_end) { 2971 free_page(used); 2972 used += PAGE_SIZE; 2973 } 2974 } 2975 return (void *)addr; 2976 } 2977 2978 /** 2979 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2980 * @size: the number of bytes to allocate 2981 * @gfp_mask: GFP flags for the allocation 2982 * 2983 * This function is similar to alloc_pages(), except that it allocates the 2984 * minimum number of pages to satisfy the request. alloc_pages() can only 2985 * allocate memory in power-of-two pages. 2986 * 2987 * This function is also limited by MAX_ORDER. 2988 * 2989 * Memory allocated by this function must be released by free_pages_exact(). 2990 */ 2991 void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2992 { 2993 unsigned int order = get_order(size); 2994 unsigned long addr; 2995 2996 addr = __get_free_pages(gfp_mask, order); 2997 return make_alloc_exact(addr, order, size); 2998 } 2999 EXPORT_SYMBOL(alloc_pages_exact); 3000 3001 /** 3002 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous 3003 * pages on a node. 3004 * @nid: the preferred node ID where memory should be allocated 3005 * @size: the number of bytes to allocate 3006 * @gfp_mask: GFP flags for the allocation 3007 * 3008 * Like alloc_pages_exact(), but try to allocate on node nid first before falling 3009 * back. 3010 * Note this is not alloc_pages_exact_node() which allocates on a specific node, 3011 * but is not exact. 3012 */ 3013 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) 3014 { 3015 unsigned order = get_order(size); 3016 struct page *p = alloc_pages_node(nid, gfp_mask, order); 3017 if (!p) 3018 return NULL; 3019 return make_alloc_exact((unsigned long)page_address(p), order, size); 3020 } 3021 3022 /** 3023 * free_pages_exact - release memory allocated via alloc_pages_exact() 3024 * @virt: the value returned by alloc_pages_exact. 3025 * @size: size of allocation, same value as passed to alloc_pages_exact(). 3026 * 3027 * Release the memory allocated by a previous call to alloc_pages_exact. 3028 */ 3029 void free_pages_exact(void *virt, size_t size) 3030 { 3031 unsigned long addr = (unsigned long)virt; 3032 unsigned long end = addr + PAGE_ALIGN(size); 3033 3034 while (addr < end) { 3035 free_page(addr); 3036 addr += PAGE_SIZE; 3037 } 3038 } 3039 EXPORT_SYMBOL(free_pages_exact); 3040 3041 /** 3042 * nr_free_zone_pages - count number of pages beyond high watermark 3043 * @offset: The zone index of the highest zone 3044 * 3045 * nr_free_zone_pages() counts the number of counts pages which are beyond the 3046 * high watermark within all zones at or below a given zone index. For each 3047 * zone, the number of pages is calculated as: 3048 * managed_pages - high_pages 3049 */ 3050 static unsigned long nr_free_zone_pages(int offset) 3051 { 3052 struct zoneref *z; 3053 struct zone *zone; 3054 3055 /* Just pick one node, since fallback list is circular */ 3056 unsigned long sum = 0; 3057 3058 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 3059 3060 for_each_zone_zonelist(zone, z, zonelist, offset) { 3061 unsigned long size = zone->managed_pages; 3062 unsigned long high = high_wmark_pages(zone); 3063 if (size > high) 3064 sum += size - high; 3065 } 3066 3067 return sum; 3068 } 3069 3070 /** 3071 * nr_free_buffer_pages - count number of pages beyond high watermark 3072 * 3073 * nr_free_buffer_pages() counts the number of pages which are beyond the high 3074 * watermark within ZONE_DMA and ZONE_NORMAL. 3075 */ 3076 unsigned long nr_free_buffer_pages(void) 3077 { 3078 return nr_free_zone_pages(gfp_zone(GFP_USER)); 3079 } 3080 EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 3081 3082 /** 3083 * nr_free_pagecache_pages - count number of pages beyond high watermark 3084 * 3085 * nr_free_pagecache_pages() counts the number of pages which are beyond the 3086 * high watermark within all zones. 3087 */ 3088 unsigned long nr_free_pagecache_pages(void) 3089 { 3090 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 3091 } 3092 3093 static inline void show_node(struct zone *zone) 3094 { 3095 if (IS_ENABLED(CONFIG_NUMA)) 3096 printk("Node %d ", zone_to_nid(zone)); 3097 } 3098 3099 void si_meminfo(struct sysinfo *val) 3100 { 3101 val->totalram = totalram_pages; 3102 val->sharedram = global_page_state(NR_SHMEM); 3103 val->freeram = global_page_state(NR_FREE_PAGES); 3104 val->bufferram = nr_blockdev_pages(); 3105 val->totalhigh = totalhigh_pages; 3106 val->freehigh = nr_free_highpages(); 3107 val->mem_unit = PAGE_SIZE; 3108 } 3109 3110 EXPORT_SYMBOL(si_meminfo); 3111 3112 #ifdef CONFIG_NUMA 3113 void si_meminfo_node(struct sysinfo *val, int nid) 3114 { 3115 int zone_type; /* needs to be signed */ 3116 unsigned long managed_pages = 0; 3117 pg_data_t *pgdat = NODE_DATA(nid); 3118 3119 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) 3120 managed_pages += pgdat->node_zones[zone_type].managed_pages; 3121 val->totalram = managed_pages; 3122 val->sharedram = node_page_state(nid, NR_SHMEM); 3123 val->freeram = node_page_state(nid, NR_FREE_PAGES); 3124 #ifdef CONFIG_HIGHMEM 3125 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages; 3126 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 3127 NR_FREE_PAGES); 3128 #else 3129 val->totalhigh = 0; 3130 val->freehigh = 0; 3131 #endif 3132 val->mem_unit = PAGE_SIZE; 3133 } 3134 #endif 3135 3136 /* 3137 * Determine whether the node should be displayed or not, depending on whether 3138 * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). 3139 */ 3140 bool skip_free_areas_node(unsigned int flags, int nid) 3141 { 3142 bool ret = false; 3143 unsigned int cpuset_mems_cookie; 3144 3145 if (!(flags & SHOW_MEM_FILTER_NODES)) 3146 goto out; 3147 3148 do { 3149 cpuset_mems_cookie = read_mems_allowed_begin(); 3150 ret = !node_isset(nid, cpuset_current_mems_allowed); 3151 } while (read_mems_allowed_retry(cpuset_mems_cookie)); 3152 out: 3153 return ret; 3154 } 3155 3156 #define K(x) ((x) << (PAGE_SHIFT-10)) 3157 3158 static void show_migration_types(unsigned char type) 3159 { 3160 static const char types[MIGRATE_TYPES] = { 3161 [MIGRATE_UNMOVABLE] = 'U', 3162 [MIGRATE_RECLAIMABLE] = 'E', 3163 [MIGRATE_MOVABLE] = 'M', 3164 [MIGRATE_RESERVE] = 'R', 3165 #ifdef CONFIG_CMA 3166 [MIGRATE_CMA] = 'C', 3167 #endif 3168 #ifdef CONFIG_MEMORY_ISOLATION 3169 [MIGRATE_ISOLATE] = 'I', 3170 #endif 3171 }; 3172 char tmp[MIGRATE_TYPES + 1]; 3173 char *p = tmp; 3174 int i; 3175 3176 for (i = 0; i < MIGRATE_TYPES; i++) { 3177 if (type & (1 << i)) 3178 *p++ = types[i]; 3179 } 3180 3181 *p = '\0'; 3182 printk("(%s) ", tmp); 3183 } 3184 3185 /* 3186 * Show free area list (used inside shift_scroll-lock stuff) 3187 * We also calculate the percentage fragmentation. We do this by counting the 3188 * memory on each free list with the exception of the first item on the list. 3189 * Suppresses nodes that are not allowed by current's cpuset if 3190 * SHOW_MEM_FILTER_NODES is passed. 3191 */ 3192 void show_free_areas(unsigned int filter) 3193 { 3194 int cpu; 3195 struct zone *zone; 3196 3197 for_each_populated_zone(zone) { 3198 if (skip_free_areas_node(filter, zone_to_nid(zone))) 3199 continue; 3200 show_node(zone); 3201 printk("%s per-cpu:\n", zone->name); 3202 3203 for_each_online_cpu(cpu) { 3204 struct per_cpu_pageset *pageset; 3205 3206 pageset = per_cpu_ptr(zone->pageset, cpu); 3207 3208 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 3209 cpu, pageset->pcp.high, 3210 pageset->pcp.batch, pageset->pcp.count); 3211 } 3212 } 3213 3214 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 3215 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 3216 " unevictable:%lu" 3217 " dirty:%lu writeback:%lu unstable:%lu\n" 3218 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 3219 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n" 3220 " free_cma:%lu\n", 3221 global_page_state(NR_ACTIVE_ANON), 3222 global_page_state(NR_INACTIVE_ANON), 3223 global_page_state(NR_ISOLATED_ANON), 3224 global_page_state(NR_ACTIVE_FILE), 3225 global_page_state(NR_INACTIVE_FILE), 3226 global_page_state(NR_ISOLATED_FILE), 3227 global_page_state(NR_UNEVICTABLE), 3228 global_page_state(NR_FILE_DIRTY), 3229 global_page_state(NR_WRITEBACK), 3230 global_page_state(NR_UNSTABLE_NFS), 3231 global_page_state(NR_FREE_PAGES), 3232 global_page_state(NR_SLAB_RECLAIMABLE), 3233 global_page_state(NR_SLAB_UNRECLAIMABLE), 3234 global_page_state(NR_FILE_MAPPED), 3235 global_page_state(NR_SHMEM), 3236 global_page_state(NR_PAGETABLE), 3237 global_page_state(NR_BOUNCE), 3238 global_page_state(NR_FREE_CMA_PAGES)); 3239 3240 for_each_populated_zone(zone) { 3241 int i; 3242 3243 if (skip_free_areas_node(filter, zone_to_nid(zone))) 3244 continue; 3245 show_node(zone); 3246 printk("%s" 3247 " free:%lukB" 3248 " min:%lukB" 3249 " low:%lukB" 3250 " high:%lukB" 3251 " active_anon:%lukB" 3252 " inactive_anon:%lukB" 3253 " active_file:%lukB" 3254 " inactive_file:%lukB" 3255 " unevictable:%lukB" 3256 " isolated(anon):%lukB" 3257 " isolated(file):%lukB" 3258 " present:%lukB" 3259 " managed:%lukB" 3260 " mlocked:%lukB" 3261 " dirty:%lukB" 3262 " writeback:%lukB" 3263 " mapped:%lukB" 3264 " shmem:%lukB" 3265 " slab_reclaimable:%lukB" 3266 " slab_unreclaimable:%lukB" 3267 " kernel_stack:%lukB" 3268 " pagetables:%lukB" 3269 " unstable:%lukB" 3270 " bounce:%lukB" 3271 " free_cma:%lukB" 3272 " writeback_tmp:%lukB" 3273 " pages_scanned:%lu" 3274 " all_unreclaimable? %s" 3275 "\n", 3276 zone->name, 3277 K(zone_page_state(zone, NR_FREE_PAGES)), 3278 K(min_wmark_pages(zone)), 3279 K(low_wmark_pages(zone)), 3280 K(high_wmark_pages(zone)), 3281 K(zone_page_state(zone, NR_ACTIVE_ANON)), 3282 K(zone_page_state(zone, NR_INACTIVE_ANON)), 3283 K(zone_page_state(zone, NR_ACTIVE_FILE)), 3284 K(zone_page_state(zone, NR_INACTIVE_FILE)), 3285 K(zone_page_state(zone, NR_UNEVICTABLE)), 3286 K(zone_page_state(zone, NR_ISOLATED_ANON)), 3287 K(zone_page_state(zone, NR_ISOLATED_FILE)), 3288 K(zone->present_pages), 3289 K(zone->managed_pages), 3290 K(zone_page_state(zone, NR_MLOCK)), 3291 K(zone_page_state(zone, NR_FILE_DIRTY)), 3292 K(zone_page_state(zone, NR_WRITEBACK)), 3293 K(zone_page_state(zone, NR_FILE_MAPPED)), 3294 K(zone_page_state(zone, NR_SHMEM)), 3295 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 3296 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 3297 zone_page_state(zone, NR_KERNEL_STACK) * 3298 THREAD_SIZE / 1024, 3299 K(zone_page_state(zone, NR_PAGETABLE)), 3300 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 3301 K(zone_page_state(zone, NR_BOUNCE)), 3302 K(zone_page_state(zone, NR_FREE_CMA_PAGES)), 3303 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 3304 K(zone_page_state(zone, NR_PAGES_SCANNED)), 3305 (!zone_reclaimable(zone) ? "yes" : "no") 3306 ); 3307 printk("lowmem_reserve[]:"); 3308 for (i = 0; i < MAX_NR_ZONES; i++) 3309 printk(" %ld", zone->lowmem_reserve[i]); 3310 printk("\n"); 3311 } 3312 3313 for_each_populated_zone(zone) { 3314 unsigned long nr[MAX_ORDER], flags, order, total = 0; 3315 unsigned char types[MAX_ORDER]; 3316 3317 if (skip_free_areas_node(filter, zone_to_nid(zone))) 3318 continue; 3319 show_node(zone); 3320 printk("%s: ", zone->name); 3321 3322 spin_lock_irqsave(&zone->lock, flags); 3323 for (order = 0; order < MAX_ORDER; order++) { 3324 struct free_area *area = &zone->free_area[order]; 3325 int type; 3326 3327 nr[order] = area->nr_free; 3328 total += nr[order] << order; 3329 3330 types[order] = 0; 3331 for (type = 0; type < MIGRATE_TYPES; type++) { 3332 if (!list_empty(&area->free_list[type])) 3333 types[order] |= 1 << type; 3334 } 3335 } 3336 spin_unlock_irqrestore(&zone->lock, flags); 3337 for (order = 0; order < MAX_ORDER; order++) { 3338 printk("%lu*%lukB ", nr[order], K(1UL) << order); 3339 if (nr[order]) 3340 show_migration_types(types[order]); 3341 } 3342 printk("= %lukB\n", K(total)); 3343 } 3344 3345 hugetlb_show_meminfo(); 3346 3347 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 3348 3349 show_swap_cache_info(); 3350 } 3351 3352 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 3353 { 3354 zoneref->zone = zone; 3355 zoneref->zone_idx = zone_idx(zone); 3356 } 3357 3358 /* 3359 * Builds allocation fallback zone lists. 3360 * 3361 * Add all populated zones of a node to the zonelist. 3362 */ 3363 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 3364 int nr_zones) 3365 { 3366 struct zone *zone; 3367 enum zone_type zone_type = MAX_NR_ZONES; 3368 3369 do { 3370 zone_type--; 3371 zone = pgdat->node_zones + zone_type; 3372 if (populated_zone(zone)) { 3373 zoneref_set_zone(zone, 3374 &zonelist->_zonerefs[nr_zones++]); 3375 check_highest_zone(zone_type); 3376 } 3377 } while (zone_type); 3378 3379 return nr_zones; 3380 } 3381 3382 3383 /* 3384 * zonelist_order: 3385 * 0 = automatic detection of better ordering. 3386 * 1 = order by ([node] distance, -zonetype) 3387 * 2 = order by (-zonetype, [node] distance) 3388 * 3389 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 3390 * the same zonelist. So only NUMA can configure this param. 3391 */ 3392 #define ZONELIST_ORDER_DEFAULT 0 3393 #define ZONELIST_ORDER_NODE 1 3394 #define ZONELIST_ORDER_ZONE 2 3395 3396 /* zonelist order in the kernel. 3397 * set_zonelist_order() will set this to NODE or ZONE. 3398 */ 3399 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 3400 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 3401 3402 3403 #ifdef CONFIG_NUMA 3404 /* The value user specified ....changed by config */ 3405 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 3406 /* string for sysctl */ 3407 #define NUMA_ZONELIST_ORDER_LEN 16 3408 char numa_zonelist_order[16] = "default"; 3409 3410 /* 3411 * interface for configure zonelist ordering. 3412 * command line option "numa_zonelist_order" 3413 * = "[dD]efault - default, automatic configuration. 3414 * = "[nN]ode - order by node locality, then by zone within node 3415 * = "[zZ]one - order by zone, then by locality within zone 3416 */ 3417 3418 static int __parse_numa_zonelist_order(char *s) 3419 { 3420 if (*s == 'd' || *s == 'D') { 3421 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 3422 } else if (*s == 'n' || *s == 'N') { 3423 user_zonelist_order = ZONELIST_ORDER_NODE; 3424 } else if (*s == 'z' || *s == 'Z') { 3425 user_zonelist_order = ZONELIST_ORDER_ZONE; 3426 } else { 3427 printk(KERN_WARNING 3428 "Ignoring invalid numa_zonelist_order value: " 3429 "%s\n", s); 3430 return -EINVAL; 3431 } 3432 return 0; 3433 } 3434 3435 static __init int setup_numa_zonelist_order(char *s) 3436 { 3437 int ret; 3438 3439 if (!s) 3440 return 0; 3441 3442 ret = __parse_numa_zonelist_order(s); 3443 if (ret == 0) 3444 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); 3445 3446 return ret; 3447 } 3448 early_param("numa_zonelist_order", setup_numa_zonelist_order); 3449 3450 /* 3451 * sysctl handler for numa_zonelist_order 3452 */ 3453 int numa_zonelist_order_handler(struct ctl_table *table, int write, 3454 void __user *buffer, size_t *length, 3455 loff_t *ppos) 3456 { 3457 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 3458 int ret; 3459 static DEFINE_MUTEX(zl_order_mutex); 3460 3461 mutex_lock(&zl_order_mutex); 3462 if (write) { 3463 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) { 3464 ret = -EINVAL; 3465 goto out; 3466 } 3467 strcpy(saved_string, (char *)table->data); 3468 } 3469 ret = proc_dostring(table, write, buffer, length, ppos); 3470 if (ret) 3471 goto out; 3472 if (write) { 3473 int oldval = user_zonelist_order; 3474 3475 ret = __parse_numa_zonelist_order((char *)table->data); 3476 if (ret) { 3477 /* 3478 * bogus value. restore saved string 3479 */ 3480 strncpy((char *)table->data, saved_string, 3481 NUMA_ZONELIST_ORDER_LEN); 3482 user_zonelist_order = oldval; 3483 } else if (oldval != user_zonelist_order) { 3484 mutex_lock(&zonelists_mutex); 3485 build_all_zonelists(NULL, NULL); 3486 mutex_unlock(&zonelists_mutex); 3487 } 3488 } 3489 out: 3490 mutex_unlock(&zl_order_mutex); 3491 return ret; 3492 } 3493 3494 3495 #define MAX_NODE_LOAD (nr_online_nodes) 3496 static int node_load[MAX_NUMNODES]; 3497 3498 /** 3499 * find_next_best_node - find the next node that should appear in a given node's fallback list 3500 * @node: node whose fallback list we're appending 3501 * @used_node_mask: nodemask_t of already used nodes 3502 * 3503 * We use a number of factors to determine which is the next node that should 3504 * appear on a given node's fallback list. The node should not have appeared 3505 * already in @node's fallback list, and it should be the next closest node 3506 * according to the distance array (which contains arbitrary distance values 3507 * from each node to each node in the system), and should also prefer nodes 3508 * with no CPUs, since presumably they'll have very little allocation pressure 3509 * on them otherwise. 3510 * It returns -1 if no node is found. 3511 */ 3512 static int find_next_best_node(int node, nodemask_t *used_node_mask) 3513 { 3514 int n, val; 3515 int min_val = INT_MAX; 3516 int best_node = NUMA_NO_NODE; 3517 const struct cpumask *tmp = cpumask_of_node(0); 3518 3519 /* Use the local node if we haven't already */ 3520 if (!node_isset(node, *used_node_mask)) { 3521 node_set(node, *used_node_mask); 3522 return node; 3523 } 3524 3525 for_each_node_state(n, N_MEMORY) { 3526 3527 /* Don't want a node to appear more than once */ 3528 if (node_isset(n, *used_node_mask)) 3529 continue; 3530 3531 /* Use the distance array to find the distance */ 3532 val = node_distance(node, n); 3533 3534 /* Penalize nodes under us ("prefer the next node") */ 3535 val += (n < node); 3536 3537 /* Give preference to headless and unused nodes */ 3538 tmp = cpumask_of_node(n); 3539 if (!cpumask_empty(tmp)) 3540 val += PENALTY_FOR_NODE_WITH_CPUS; 3541 3542 /* Slight preference for less loaded node */ 3543 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 3544 val += node_load[n]; 3545 3546 if (val < min_val) { 3547 min_val = val; 3548 best_node = n; 3549 } 3550 } 3551 3552 if (best_node >= 0) 3553 node_set(best_node, *used_node_mask); 3554 3555 return best_node; 3556 } 3557 3558 3559 /* 3560 * Build zonelists ordered by node and zones within node. 3561 * This results in maximum locality--normal zone overflows into local 3562 * DMA zone, if any--but risks exhausting DMA zone. 3563 */ 3564 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 3565 { 3566 int j; 3567 struct zonelist *zonelist; 3568 3569 zonelist = &pgdat->node_zonelists[0]; 3570 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 3571 ; 3572 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 3573 zonelist->_zonerefs[j].zone = NULL; 3574 zonelist->_zonerefs[j].zone_idx = 0; 3575 } 3576 3577 /* 3578 * Build gfp_thisnode zonelists 3579 */ 3580 static void build_thisnode_zonelists(pg_data_t *pgdat) 3581 { 3582 int j; 3583 struct zonelist *zonelist; 3584 3585 zonelist = &pgdat->node_zonelists[1]; 3586 j = build_zonelists_node(pgdat, zonelist, 0); 3587 zonelist->_zonerefs[j].zone = NULL; 3588 zonelist->_zonerefs[j].zone_idx = 0; 3589 } 3590 3591 /* 3592 * Build zonelists ordered by zone and nodes within zones. 3593 * This results in conserving DMA zone[s] until all Normal memory is 3594 * exhausted, but results in overflowing to remote node while memory 3595 * may still exist in local DMA zone. 3596 */ 3597 static int node_order[MAX_NUMNODES]; 3598 3599 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 3600 { 3601 int pos, j, node; 3602 int zone_type; /* needs to be signed */ 3603 struct zone *z; 3604 struct zonelist *zonelist; 3605 3606 zonelist = &pgdat->node_zonelists[0]; 3607 pos = 0; 3608 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 3609 for (j = 0; j < nr_nodes; j++) { 3610 node = node_order[j]; 3611 z = &NODE_DATA(node)->node_zones[zone_type]; 3612 if (populated_zone(z)) { 3613 zoneref_set_zone(z, 3614 &zonelist->_zonerefs[pos++]); 3615 check_highest_zone(zone_type); 3616 } 3617 } 3618 } 3619 zonelist->_zonerefs[pos].zone = NULL; 3620 zonelist->_zonerefs[pos].zone_idx = 0; 3621 } 3622 3623 #if defined(CONFIG_64BIT) 3624 /* 3625 * Devices that require DMA32/DMA are relatively rare and do not justify a 3626 * penalty to every machine in case the specialised case applies. Default 3627 * to Node-ordering on 64-bit NUMA machines 3628 */ 3629 static int default_zonelist_order(void) 3630 { 3631 return ZONELIST_ORDER_NODE; 3632 } 3633 #else 3634 /* 3635 * On 32-bit, the Normal zone needs to be preserved for allocations accessible 3636 * by the kernel. If processes running on node 0 deplete the low memory zone 3637 * then reclaim will occur more frequency increasing stalls and potentially 3638 * be easier to OOM if a large percentage of the zone is under writeback or 3639 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set. 3640 * Hence, default to zone ordering on 32-bit. 3641 */ 3642 static int default_zonelist_order(void) 3643 { 3644 return ZONELIST_ORDER_ZONE; 3645 } 3646 #endif /* CONFIG_64BIT */ 3647 3648 static void set_zonelist_order(void) 3649 { 3650 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 3651 current_zonelist_order = default_zonelist_order(); 3652 else 3653 current_zonelist_order = user_zonelist_order; 3654 } 3655 3656 static void build_zonelists(pg_data_t *pgdat) 3657 { 3658 int j, node, load; 3659 enum zone_type i; 3660 nodemask_t used_mask; 3661 int local_node, prev_node; 3662 struct zonelist *zonelist; 3663 int order = current_zonelist_order; 3664 3665 /* initialize zonelists */ 3666 for (i = 0; i < MAX_ZONELISTS; i++) { 3667 zonelist = pgdat->node_zonelists + i; 3668 zonelist->_zonerefs[0].zone = NULL; 3669 zonelist->_zonerefs[0].zone_idx = 0; 3670 } 3671 3672 /* NUMA-aware ordering of nodes */ 3673 local_node = pgdat->node_id; 3674 load = nr_online_nodes; 3675 prev_node = local_node; 3676 nodes_clear(used_mask); 3677 3678 memset(node_order, 0, sizeof(node_order)); 3679 j = 0; 3680 3681 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 3682 /* 3683 * We don't want to pressure a particular node. 3684 * So adding penalty to the first node in same 3685 * distance group to make it round-robin. 3686 */ 3687 if (node_distance(local_node, node) != 3688 node_distance(local_node, prev_node)) 3689 node_load[node] = load; 3690 3691 prev_node = node; 3692 load--; 3693 if (order == ZONELIST_ORDER_NODE) 3694 build_zonelists_in_node_order(pgdat, node); 3695 else 3696 node_order[j++] = node; /* remember order */ 3697 } 3698 3699 if (order == ZONELIST_ORDER_ZONE) { 3700 /* calculate node order -- i.e., DMA last! */ 3701 build_zonelists_in_zone_order(pgdat, j); 3702 } 3703 3704 build_thisnode_zonelists(pgdat); 3705 } 3706 3707 /* Construct the zonelist performance cache - see further mmzone.h */ 3708 static void build_zonelist_cache(pg_data_t *pgdat) 3709 { 3710 struct zonelist *zonelist; 3711 struct zonelist_cache *zlc; 3712 struct zoneref *z; 3713 3714 zonelist = &pgdat->node_zonelists[0]; 3715 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 3716 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 3717 for (z = zonelist->_zonerefs; z->zone; z++) 3718 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 3719 } 3720 3721 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 3722 /* 3723 * Return node id of node used for "local" allocations. 3724 * I.e., first node id of first zone in arg node's generic zonelist. 3725 * Used for initializing percpu 'numa_mem', which is used primarily 3726 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. 3727 */ 3728 int local_memory_node(int node) 3729 { 3730 struct zone *zone; 3731 3732 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), 3733 gfp_zone(GFP_KERNEL), 3734 NULL, 3735 &zone); 3736 return zone->node; 3737 } 3738 #endif 3739 3740 #else /* CONFIG_NUMA */ 3741 3742 static void set_zonelist_order(void) 3743 { 3744 current_zonelist_order = ZONELIST_ORDER_ZONE; 3745 } 3746 3747 static void build_zonelists(pg_data_t *pgdat) 3748 { 3749 int node, local_node; 3750 enum zone_type j; 3751 struct zonelist *zonelist; 3752 3753 local_node = pgdat->node_id; 3754 3755 zonelist = &pgdat->node_zonelists[0]; 3756 j = build_zonelists_node(pgdat, zonelist, 0); 3757 3758 /* 3759 * Now we build the zonelist so that it contains the zones 3760 * of all the other nodes. 3761 * We don't want to pressure a particular node, so when 3762 * building the zones for node N, we make sure that the 3763 * zones coming right after the local ones are those from 3764 * node N+1 (modulo N) 3765 */ 3766 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 3767 if (!node_online(node)) 3768 continue; 3769 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 3770 } 3771 for (node = 0; node < local_node; node++) { 3772 if (!node_online(node)) 3773 continue; 3774 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 3775 } 3776 3777 zonelist->_zonerefs[j].zone = NULL; 3778 zonelist->_zonerefs[j].zone_idx = 0; 3779 } 3780 3781 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 3782 static void build_zonelist_cache(pg_data_t *pgdat) 3783 { 3784 pgdat->node_zonelists[0].zlcache_ptr = NULL; 3785 } 3786 3787 #endif /* CONFIG_NUMA */ 3788 3789 /* 3790 * Boot pageset table. One per cpu which is going to be used for all 3791 * zones and all nodes. The parameters will be set in such a way 3792 * that an item put on a list will immediately be handed over to 3793 * the buddy list. This is safe since pageset manipulation is done 3794 * with interrupts disabled. 3795 * 3796 * The boot_pagesets must be kept even after bootup is complete for 3797 * unused processors and/or zones. They do play a role for bootstrapping 3798 * hotplugged processors. 3799 * 3800 * zoneinfo_show() and maybe other functions do 3801 * not check if the processor is online before following the pageset pointer. 3802 * Other parts of the kernel may not check if the zone is available. 3803 */ 3804 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 3805 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 3806 static void setup_zone_pageset(struct zone *zone); 3807 3808 /* 3809 * Global mutex to protect against size modification of zonelists 3810 * as well as to serialize pageset setup for the new populated zone. 3811 */ 3812 DEFINE_MUTEX(zonelists_mutex); 3813 3814 /* return values int ....just for stop_machine() */ 3815 static int __build_all_zonelists(void *data) 3816 { 3817 int nid; 3818 int cpu; 3819 pg_data_t *self = data; 3820 3821 #ifdef CONFIG_NUMA 3822 memset(node_load, 0, sizeof(node_load)); 3823 #endif 3824 3825 if (self && !node_online(self->node_id)) { 3826 build_zonelists(self); 3827 build_zonelist_cache(self); 3828 } 3829 3830 for_each_online_node(nid) { 3831 pg_data_t *pgdat = NODE_DATA(nid); 3832 3833 build_zonelists(pgdat); 3834 build_zonelist_cache(pgdat); 3835 } 3836 3837 /* 3838 * Initialize the boot_pagesets that are going to be used 3839 * for bootstrapping processors. The real pagesets for 3840 * each zone will be allocated later when the per cpu 3841 * allocator is available. 3842 * 3843 * boot_pagesets are used also for bootstrapping offline 3844 * cpus if the system is already booted because the pagesets 3845 * are needed to initialize allocators on a specific cpu too. 3846 * F.e. the percpu allocator needs the page allocator which 3847 * needs the percpu allocator in order to allocate its pagesets 3848 * (a chicken-egg dilemma). 3849 */ 3850 for_each_possible_cpu(cpu) { 3851 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 3852 3853 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 3854 /* 3855 * We now know the "local memory node" for each node-- 3856 * i.e., the node of the first zone in the generic zonelist. 3857 * Set up numa_mem percpu variable for on-line cpus. During 3858 * boot, only the boot cpu should be on-line; we'll init the 3859 * secondary cpus' numa_mem as they come on-line. During 3860 * node/memory hotplug, we'll fixup all on-line cpus. 3861 */ 3862 if (cpu_online(cpu)) 3863 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); 3864 #endif 3865 } 3866 3867 return 0; 3868 } 3869 3870 /* 3871 * Called with zonelists_mutex held always 3872 * unless system_state == SYSTEM_BOOTING. 3873 */ 3874 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone) 3875 { 3876 set_zonelist_order(); 3877 3878 if (system_state == SYSTEM_BOOTING) { 3879 __build_all_zonelists(NULL); 3880 mminit_verify_zonelist(); 3881 cpuset_init_current_mems_allowed(); 3882 } else { 3883 #ifdef CONFIG_MEMORY_HOTPLUG 3884 if (zone) 3885 setup_zone_pageset(zone); 3886 #endif 3887 /* we have to stop all cpus to guarantee there is no user 3888 of zonelist */ 3889 stop_machine(__build_all_zonelists, pgdat, NULL); 3890 /* cpuset refresh routine should be here */ 3891 } 3892 vm_total_pages = nr_free_pagecache_pages(); 3893 /* 3894 * Disable grouping by mobility if the number of pages in the 3895 * system is too low to allow the mechanism to work. It would be 3896 * more accurate, but expensive to check per-zone. This check is 3897 * made on memory-hotadd so a system can start with mobility 3898 * disabled and enable it later 3899 */ 3900 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 3901 page_group_by_mobility_disabled = 1; 3902 else 3903 page_group_by_mobility_disabled = 0; 3904 3905 printk("Built %i zonelists in %s order, mobility grouping %s. " 3906 "Total pages: %ld\n", 3907 nr_online_nodes, 3908 zonelist_order_name[current_zonelist_order], 3909 page_group_by_mobility_disabled ? "off" : "on", 3910 vm_total_pages); 3911 #ifdef CONFIG_NUMA 3912 printk("Policy zone: %s\n", zone_names[policy_zone]); 3913 #endif 3914 } 3915 3916 /* 3917 * Helper functions to size the waitqueue hash table. 3918 * Essentially these want to choose hash table sizes sufficiently 3919 * large so that collisions trying to wait on pages are rare. 3920 * But in fact, the number of active page waitqueues on typical 3921 * systems is ridiculously low, less than 200. So this is even 3922 * conservative, even though it seems large. 3923 * 3924 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 3925 * waitqueues, i.e. the size of the waitq table given the number of pages. 3926 */ 3927 #define PAGES_PER_WAITQUEUE 256 3928 3929 #ifndef CONFIG_MEMORY_HOTPLUG 3930 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3931 { 3932 unsigned long size = 1; 3933 3934 pages /= PAGES_PER_WAITQUEUE; 3935 3936 while (size < pages) 3937 size <<= 1; 3938 3939 /* 3940 * Once we have dozens or even hundreds of threads sleeping 3941 * on IO we've got bigger problems than wait queue collision. 3942 * Limit the size of the wait table to a reasonable size. 3943 */ 3944 size = min(size, 4096UL); 3945 3946 return max(size, 4UL); 3947 } 3948 #else 3949 /* 3950 * A zone's size might be changed by hot-add, so it is not possible to determine 3951 * a suitable size for its wait_table. So we use the maximum size now. 3952 * 3953 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 3954 * 3955 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 3956 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 3957 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 3958 * 3959 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 3960 * or more by the traditional way. (See above). It equals: 3961 * 3962 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 3963 * ia64(16K page size) : = ( 8G + 4M)byte. 3964 * powerpc (64K page size) : = (32G +16M)byte. 3965 */ 3966 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3967 { 3968 return 4096UL; 3969 } 3970 #endif 3971 3972 /* 3973 * This is an integer logarithm so that shifts can be used later 3974 * to extract the more random high bits from the multiplicative 3975 * hash function before the remainder is taken. 3976 */ 3977 static inline unsigned long wait_table_bits(unsigned long size) 3978 { 3979 return ffz(~size); 3980 } 3981 3982 /* 3983 * Check if a pageblock contains reserved pages 3984 */ 3985 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) 3986 { 3987 unsigned long pfn; 3988 3989 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3990 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) 3991 return 1; 3992 } 3993 return 0; 3994 } 3995 3996 /* 3997 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 3998 * of blocks reserved is based on min_wmark_pages(zone). The memory within 3999 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 4000 * higher will lead to a bigger reserve which will get freed as contiguous 4001 * blocks as reclaim kicks in 4002 */ 4003 static void setup_zone_migrate_reserve(struct zone *zone) 4004 { 4005 unsigned long start_pfn, pfn, end_pfn, block_end_pfn; 4006 struct page *page; 4007 unsigned long block_migratetype; 4008 int reserve; 4009 int old_reserve; 4010 4011 /* 4012 * Get the start pfn, end pfn and the number of blocks to reserve 4013 * We have to be careful to be aligned to pageblock_nr_pages to 4014 * make sure that we always check pfn_valid for the first page in 4015 * the block. 4016 */ 4017 start_pfn = zone->zone_start_pfn; 4018 end_pfn = zone_end_pfn(zone); 4019 start_pfn = roundup(start_pfn, pageblock_nr_pages); 4020 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 4021 pageblock_order; 4022 4023 /* 4024 * Reserve blocks are generally in place to help high-order atomic 4025 * allocations that are short-lived. A min_free_kbytes value that 4026 * would result in more than 2 reserve blocks for atomic allocations 4027 * is assumed to be in place to help anti-fragmentation for the 4028 * future allocation of hugepages at runtime. 4029 */ 4030 reserve = min(2, reserve); 4031 old_reserve = zone->nr_migrate_reserve_block; 4032 4033 /* When memory hot-add, we almost always need to do nothing */ 4034 if (reserve == old_reserve) 4035 return; 4036 zone->nr_migrate_reserve_block = reserve; 4037 4038 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 4039 if (!pfn_valid(pfn)) 4040 continue; 4041 page = pfn_to_page(pfn); 4042 4043 /* Watch out for overlapping nodes */ 4044 if (page_to_nid(page) != zone_to_nid(zone)) 4045 continue; 4046 4047 block_migratetype = get_pageblock_migratetype(page); 4048 4049 /* Only test what is necessary when the reserves are not met */ 4050 if (reserve > 0) { 4051 /* 4052 * Blocks with reserved pages will never free, skip 4053 * them. 4054 */ 4055 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); 4056 if (pageblock_is_reserved(pfn, block_end_pfn)) 4057 continue; 4058 4059 /* If this block is reserved, account for it */ 4060 if (block_migratetype == MIGRATE_RESERVE) { 4061 reserve--; 4062 continue; 4063 } 4064 4065 /* Suitable for reserving if this block is movable */ 4066 if (block_migratetype == MIGRATE_MOVABLE) { 4067 set_pageblock_migratetype(page, 4068 MIGRATE_RESERVE); 4069 move_freepages_block(zone, page, 4070 MIGRATE_RESERVE); 4071 reserve--; 4072 continue; 4073 } 4074 } else if (!old_reserve) { 4075 /* 4076 * At boot time we don't need to scan the whole zone 4077 * for turning off MIGRATE_RESERVE. 4078 */ 4079 break; 4080 } 4081 4082 /* 4083 * If the reserve is met and this is a previous reserved block, 4084 * take it back 4085 */ 4086 if (block_migratetype == MIGRATE_RESERVE) { 4087 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 4088 move_freepages_block(zone, page, MIGRATE_MOVABLE); 4089 } 4090 } 4091 } 4092 4093 /* 4094 * Initially all pages are reserved - free ones are freed 4095 * up by free_all_bootmem() once the early boot process is 4096 * done. Non-atomic initialization, single-pass. 4097 */ 4098 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 4099 unsigned long start_pfn, enum memmap_context context) 4100 { 4101 struct page *page; 4102 unsigned long end_pfn = start_pfn + size; 4103 unsigned long pfn; 4104 struct zone *z; 4105 4106 if (highest_memmap_pfn < end_pfn - 1) 4107 highest_memmap_pfn = end_pfn - 1; 4108 4109 z = &NODE_DATA(nid)->node_zones[zone]; 4110 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 4111 /* 4112 * There can be holes in boot-time mem_map[]s 4113 * handed to this function. They do not 4114 * exist on hotplugged memory. 4115 */ 4116 if (context == MEMMAP_EARLY) { 4117 if (!early_pfn_valid(pfn)) 4118 continue; 4119 if (!early_pfn_in_nid(pfn, nid)) 4120 continue; 4121 } 4122 page = pfn_to_page(pfn); 4123 set_page_links(page, zone, nid, pfn); 4124 mminit_verify_page_links(page, zone, nid, pfn); 4125 init_page_count(page); 4126 page_mapcount_reset(page); 4127 page_cpupid_reset_last(page); 4128 SetPageReserved(page); 4129 /* 4130 * Mark the block movable so that blocks are reserved for 4131 * movable at startup. This will force kernel allocations 4132 * to reserve their blocks rather than leaking throughout 4133 * the address space during boot when many long-lived 4134 * kernel allocations are made. Later some blocks near 4135 * the start are marked MIGRATE_RESERVE by 4136 * setup_zone_migrate_reserve() 4137 * 4138 * bitmap is created for zone's valid pfn range. but memmap 4139 * can be created for invalid pages (for alignment) 4140 * check here not to call set_pageblock_migratetype() against 4141 * pfn out of zone. 4142 */ 4143 if ((z->zone_start_pfn <= pfn) 4144 && (pfn < zone_end_pfn(z)) 4145 && !(pfn & (pageblock_nr_pages - 1))) 4146 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 4147 4148 INIT_LIST_HEAD(&page->lru); 4149 #ifdef WANT_PAGE_VIRTUAL 4150 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 4151 if (!is_highmem_idx(zone)) 4152 set_page_address(page, __va(pfn << PAGE_SHIFT)); 4153 #endif 4154 } 4155 } 4156 4157 static void __meminit zone_init_free_lists(struct zone *zone) 4158 { 4159 unsigned int order, t; 4160 for_each_migratetype_order(order, t) { 4161 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 4162 zone->free_area[order].nr_free = 0; 4163 } 4164 } 4165 4166 #ifndef __HAVE_ARCH_MEMMAP_INIT 4167 #define memmap_init(size, nid, zone, start_pfn) \ 4168 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 4169 #endif 4170 4171 static int zone_batchsize(struct zone *zone) 4172 { 4173 #ifdef CONFIG_MMU 4174 int batch; 4175 4176 /* 4177 * The per-cpu-pages pools are set to around 1000th of the 4178 * size of the zone. But no more than 1/2 of a meg. 4179 * 4180 * OK, so we don't know how big the cache is. So guess. 4181 */ 4182 batch = zone->managed_pages / 1024; 4183 if (batch * PAGE_SIZE > 512 * 1024) 4184 batch = (512 * 1024) / PAGE_SIZE; 4185 batch /= 4; /* We effectively *= 4 below */ 4186 if (batch < 1) 4187 batch = 1; 4188 4189 /* 4190 * Clamp the batch to a 2^n - 1 value. Having a power 4191 * of 2 value was found to be more likely to have 4192 * suboptimal cache aliasing properties in some cases. 4193 * 4194 * For example if 2 tasks are alternately allocating 4195 * batches of pages, one task can end up with a lot 4196 * of pages of one half of the possible page colors 4197 * and the other with pages of the other colors. 4198 */ 4199 batch = rounddown_pow_of_two(batch + batch/2) - 1; 4200 4201 return batch; 4202 4203 #else 4204 /* The deferral and batching of frees should be suppressed under NOMMU 4205 * conditions. 4206 * 4207 * The problem is that NOMMU needs to be able to allocate large chunks 4208 * of contiguous memory as there's no hardware page translation to 4209 * assemble apparent contiguous memory from discontiguous pages. 4210 * 4211 * Queueing large contiguous runs of pages for batching, however, 4212 * causes the pages to actually be freed in smaller chunks. As there 4213 * can be a significant delay between the individual batches being 4214 * recycled, this leads to the once large chunks of space being 4215 * fragmented and becoming unavailable for high-order allocations. 4216 */ 4217 return 0; 4218 #endif 4219 } 4220 4221 /* 4222 * pcp->high and pcp->batch values are related and dependent on one another: 4223 * ->batch must never be higher then ->high. 4224 * The following function updates them in a safe manner without read side 4225 * locking. 4226 * 4227 * Any new users of pcp->batch and pcp->high should ensure they can cope with 4228 * those fields changing asynchronously (acording the the above rule). 4229 * 4230 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function 4231 * outside of boot time (or some other assurance that no concurrent updaters 4232 * exist). 4233 */ 4234 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high, 4235 unsigned long batch) 4236 { 4237 /* start with a fail safe value for batch */ 4238 pcp->batch = 1; 4239 smp_wmb(); 4240 4241 /* Update high, then batch, in order */ 4242 pcp->high = high; 4243 smp_wmb(); 4244 4245 pcp->batch = batch; 4246 } 4247 4248 /* a companion to pageset_set_high() */ 4249 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch) 4250 { 4251 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch)); 4252 } 4253 4254 static void pageset_init(struct per_cpu_pageset *p) 4255 { 4256 struct per_cpu_pages *pcp; 4257 int migratetype; 4258 4259 memset(p, 0, sizeof(*p)); 4260 4261 pcp = &p->pcp; 4262 pcp->count = 0; 4263 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 4264 INIT_LIST_HEAD(&pcp->lists[migratetype]); 4265 } 4266 4267 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 4268 { 4269 pageset_init(p); 4270 pageset_set_batch(p, batch); 4271 } 4272 4273 /* 4274 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist 4275 * to the value high for the pageset p. 4276 */ 4277 static void pageset_set_high(struct per_cpu_pageset *p, 4278 unsigned long high) 4279 { 4280 unsigned long batch = max(1UL, high / 4); 4281 if ((high / 4) > (PAGE_SHIFT * 8)) 4282 batch = PAGE_SHIFT * 8; 4283 4284 pageset_update(&p->pcp, high, batch); 4285 } 4286 4287 static void pageset_set_high_and_batch(struct zone *zone, 4288 struct per_cpu_pageset *pcp) 4289 { 4290 if (percpu_pagelist_fraction) 4291 pageset_set_high(pcp, 4292 (zone->managed_pages / 4293 percpu_pagelist_fraction)); 4294 else 4295 pageset_set_batch(pcp, zone_batchsize(zone)); 4296 } 4297 4298 static void __meminit zone_pageset_init(struct zone *zone, int cpu) 4299 { 4300 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 4301 4302 pageset_init(pcp); 4303 pageset_set_high_and_batch(zone, pcp); 4304 } 4305 4306 static void __meminit setup_zone_pageset(struct zone *zone) 4307 { 4308 int cpu; 4309 zone->pageset = alloc_percpu(struct per_cpu_pageset); 4310 for_each_possible_cpu(cpu) 4311 zone_pageset_init(zone, cpu); 4312 } 4313 4314 /* 4315 * Allocate per cpu pagesets and initialize them. 4316 * Before this call only boot pagesets were available. 4317 */ 4318 void __init setup_per_cpu_pageset(void) 4319 { 4320 struct zone *zone; 4321 4322 for_each_populated_zone(zone) 4323 setup_zone_pageset(zone); 4324 } 4325 4326 static noinline __init_refok 4327 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 4328 { 4329 int i; 4330 size_t alloc_size; 4331 4332 /* 4333 * The per-page waitqueue mechanism uses hashed waitqueues 4334 * per zone. 4335 */ 4336 zone->wait_table_hash_nr_entries = 4337 wait_table_hash_nr_entries(zone_size_pages); 4338 zone->wait_table_bits = 4339 wait_table_bits(zone->wait_table_hash_nr_entries); 4340 alloc_size = zone->wait_table_hash_nr_entries 4341 * sizeof(wait_queue_head_t); 4342 4343 if (!slab_is_available()) { 4344 zone->wait_table = (wait_queue_head_t *) 4345 memblock_virt_alloc_node_nopanic( 4346 alloc_size, zone->zone_pgdat->node_id); 4347 } else { 4348 /* 4349 * This case means that a zone whose size was 0 gets new memory 4350 * via memory hot-add. 4351 * But it may be the case that a new node was hot-added. In 4352 * this case vmalloc() will not be able to use this new node's 4353 * memory - this wait_table must be initialized to use this new 4354 * node itself as well. 4355 * To use this new node's memory, further consideration will be 4356 * necessary. 4357 */ 4358 zone->wait_table = vmalloc(alloc_size); 4359 } 4360 if (!zone->wait_table) 4361 return -ENOMEM; 4362 4363 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i) 4364 init_waitqueue_head(zone->wait_table + i); 4365 4366 return 0; 4367 } 4368 4369 static __meminit void zone_pcp_init(struct zone *zone) 4370 { 4371 /* 4372 * per cpu subsystem is not up at this point. The following code 4373 * relies on the ability of the linker to provide the 4374 * offset of a (static) per cpu variable into the per cpu area. 4375 */ 4376 zone->pageset = &boot_pageset; 4377 4378 if (populated_zone(zone)) 4379 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 4380 zone->name, zone->present_pages, 4381 zone_batchsize(zone)); 4382 } 4383 4384 int __meminit init_currently_empty_zone(struct zone *zone, 4385 unsigned long zone_start_pfn, 4386 unsigned long size, 4387 enum memmap_context context) 4388 { 4389 struct pglist_data *pgdat = zone->zone_pgdat; 4390 int ret; 4391 ret = zone_wait_table_init(zone, size); 4392 if (ret) 4393 return ret; 4394 pgdat->nr_zones = zone_idx(zone) + 1; 4395 4396 zone->zone_start_pfn = zone_start_pfn; 4397 4398 mminit_dprintk(MMINIT_TRACE, "memmap_init", 4399 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 4400 pgdat->node_id, 4401 (unsigned long)zone_idx(zone), 4402 zone_start_pfn, (zone_start_pfn + size)); 4403 4404 zone_init_free_lists(zone); 4405 4406 return 0; 4407 } 4408 4409 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 4410 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 4411 /* 4412 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 4413 */ 4414 int __meminit __early_pfn_to_nid(unsigned long pfn) 4415 { 4416 unsigned long start_pfn, end_pfn; 4417 int nid; 4418 /* 4419 * NOTE: The following SMP-unsafe globals are only used early in boot 4420 * when the kernel is running single-threaded. 4421 */ 4422 static unsigned long __meminitdata last_start_pfn, last_end_pfn; 4423 static int __meminitdata last_nid; 4424 4425 if (last_start_pfn <= pfn && pfn < last_end_pfn) 4426 return last_nid; 4427 4428 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn); 4429 if (nid != -1) { 4430 last_start_pfn = start_pfn; 4431 last_end_pfn = end_pfn; 4432 last_nid = nid; 4433 } 4434 4435 return nid; 4436 } 4437 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 4438 4439 int __meminit early_pfn_to_nid(unsigned long pfn) 4440 { 4441 int nid; 4442 4443 nid = __early_pfn_to_nid(pfn); 4444 if (nid >= 0) 4445 return nid; 4446 /* just returns 0 */ 4447 return 0; 4448 } 4449 4450 #ifdef CONFIG_NODES_SPAN_OTHER_NODES 4451 bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 4452 { 4453 int nid; 4454 4455 nid = __early_pfn_to_nid(pfn); 4456 if (nid >= 0 && nid != node) 4457 return false; 4458 return true; 4459 } 4460 #endif 4461 4462 /** 4463 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range 4464 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 4465 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid 4466 * 4467 * If an architecture guarantees that all ranges registered contain no holes 4468 * and may be freed, this this function may be used instead of calling 4469 * memblock_free_early_nid() manually. 4470 */ 4471 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn) 4472 { 4473 unsigned long start_pfn, end_pfn; 4474 int i, this_nid; 4475 4476 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) { 4477 start_pfn = min(start_pfn, max_low_pfn); 4478 end_pfn = min(end_pfn, max_low_pfn); 4479 4480 if (start_pfn < end_pfn) 4481 memblock_free_early_nid(PFN_PHYS(start_pfn), 4482 (end_pfn - start_pfn) << PAGE_SHIFT, 4483 this_nid); 4484 } 4485 } 4486 4487 /** 4488 * sparse_memory_present_with_active_regions - Call memory_present for each active range 4489 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 4490 * 4491 * If an architecture guarantees that all ranges registered contain no holes and may 4492 * be freed, this function may be used instead of calling memory_present() manually. 4493 */ 4494 void __init sparse_memory_present_with_active_regions(int nid) 4495 { 4496 unsigned long start_pfn, end_pfn; 4497 int i, this_nid; 4498 4499 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) 4500 memory_present(this_nid, start_pfn, end_pfn); 4501 } 4502 4503 /** 4504 * get_pfn_range_for_nid - Return the start and end page frames for a node 4505 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 4506 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 4507 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 4508 * 4509 * It returns the start and end page frame of a node based on information 4510 * provided by memblock_set_node(). If called for a node 4511 * with no available memory, a warning is printed and the start and end 4512 * PFNs will be 0. 4513 */ 4514 void __meminit get_pfn_range_for_nid(unsigned int nid, 4515 unsigned long *start_pfn, unsigned long *end_pfn) 4516 { 4517 unsigned long this_start_pfn, this_end_pfn; 4518 int i; 4519 4520 *start_pfn = -1UL; 4521 *end_pfn = 0; 4522 4523 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) { 4524 *start_pfn = min(*start_pfn, this_start_pfn); 4525 *end_pfn = max(*end_pfn, this_end_pfn); 4526 } 4527 4528 if (*start_pfn == -1UL) 4529 *start_pfn = 0; 4530 } 4531 4532 /* 4533 * This finds a zone that can be used for ZONE_MOVABLE pages. The 4534 * assumption is made that zones within a node are ordered in monotonic 4535 * increasing memory addresses so that the "highest" populated zone is used 4536 */ 4537 static void __init find_usable_zone_for_movable(void) 4538 { 4539 int zone_index; 4540 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 4541 if (zone_index == ZONE_MOVABLE) 4542 continue; 4543 4544 if (arch_zone_highest_possible_pfn[zone_index] > 4545 arch_zone_lowest_possible_pfn[zone_index]) 4546 break; 4547 } 4548 4549 VM_BUG_ON(zone_index == -1); 4550 movable_zone = zone_index; 4551 } 4552 4553 /* 4554 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 4555 * because it is sized independent of architecture. Unlike the other zones, 4556 * the starting point for ZONE_MOVABLE is not fixed. It may be different 4557 * in each node depending on the size of each node and how evenly kernelcore 4558 * is distributed. This helper function adjusts the zone ranges 4559 * provided by the architecture for a given node by using the end of the 4560 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 4561 * zones within a node are in order of monotonic increases memory addresses 4562 */ 4563 static void __meminit adjust_zone_range_for_zone_movable(int nid, 4564 unsigned long zone_type, 4565 unsigned long node_start_pfn, 4566 unsigned long node_end_pfn, 4567 unsigned long *zone_start_pfn, 4568 unsigned long *zone_end_pfn) 4569 { 4570 /* Only adjust if ZONE_MOVABLE is on this node */ 4571 if (zone_movable_pfn[nid]) { 4572 /* Size ZONE_MOVABLE */ 4573 if (zone_type == ZONE_MOVABLE) { 4574 *zone_start_pfn = zone_movable_pfn[nid]; 4575 *zone_end_pfn = min(node_end_pfn, 4576 arch_zone_highest_possible_pfn[movable_zone]); 4577 4578 /* Adjust for ZONE_MOVABLE starting within this range */ 4579 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 4580 *zone_end_pfn > zone_movable_pfn[nid]) { 4581 *zone_end_pfn = zone_movable_pfn[nid]; 4582 4583 /* Check if this whole range is within ZONE_MOVABLE */ 4584 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 4585 *zone_start_pfn = *zone_end_pfn; 4586 } 4587 } 4588 4589 /* 4590 * Return the number of pages a zone spans in a node, including holes 4591 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 4592 */ 4593 static unsigned long __meminit zone_spanned_pages_in_node(int nid, 4594 unsigned long zone_type, 4595 unsigned long node_start_pfn, 4596 unsigned long node_end_pfn, 4597 unsigned long *ignored) 4598 { 4599 unsigned long zone_start_pfn, zone_end_pfn; 4600 4601 /* Get the start and end of the zone */ 4602 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 4603 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 4604 adjust_zone_range_for_zone_movable(nid, zone_type, 4605 node_start_pfn, node_end_pfn, 4606 &zone_start_pfn, &zone_end_pfn); 4607 4608 /* Check that this node has pages within the zone's required range */ 4609 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 4610 return 0; 4611 4612 /* Move the zone boundaries inside the node if necessary */ 4613 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 4614 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 4615 4616 /* Return the spanned pages */ 4617 return zone_end_pfn - zone_start_pfn; 4618 } 4619 4620 /* 4621 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 4622 * then all holes in the requested range will be accounted for. 4623 */ 4624 unsigned long __meminit __absent_pages_in_range(int nid, 4625 unsigned long range_start_pfn, 4626 unsigned long range_end_pfn) 4627 { 4628 unsigned long nr_absent = range_end_pfn - range_start_pfn; 4629 unsigned long start_pfn, end_pfn; 4630 int i; 4631 4632 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 4633 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn); 4634 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn); 4635 nr_absent -= end_pfn - start_pfn; 4636 } 4637 return nr_absent; 4638 } 4639 4640 /** 4641 * absent_pages_in_range - Return number of page frames in holes within a range 4642 * @start_pfn: The start PFN to start searching for holes 4643 * @end_pfn: The end PFN to stop searching for holes 4644 * 4645 * It returns the number of pages frames in memory holes within a range. 4646 */ 4647 unsigned long __init absent_pages_in_range(unsigned long start_pfn, 4648 unsigned long end_pfn) 4649 { 4650 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 4651 } 4652 4653 /* Return the number of page frames in holes in a zone on a node */ 4654 static unsigned long __meminit zone_absent_pages_in_node(int nid, 4655 unsigned long zone_type, 4656 unsigned long node_start_pfn, 4657 unsigned long node_end_pfn, 4658 unsigned long *ignored) 4659 { 4660 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type]; 4661 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type]; 4662 unsigned long zone_start_pfn, zone_end_pfn; 4663 4664 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high); 4665 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high); 4666 4667 adjust_zone_range_for_zone_movable(nid, zone_type, 4668 node_start_pfn, node_end_pfn, 4669 &zone_start_pfn, &zone_end_pfn); 4670 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 4671 } 4672 4673 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4674 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 4675 unsigned long zone_type, 4676 unsigned long node_start_pfn, 4677 unsigned long node_end_pfn, 4678 unsigned long *zones_size) 4679 { 4680 return zones_size[zone_type]; 4681 } 4682 4683 static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 4684 unsigned long zone_type, 4685 unsigned long node_start_pfn, 4686 unsigned long node_end_pfn, 4687 unsigned long *zholes_size) 4688 { 4689 if (!zholes_size) 4690 return 0; 4691 4692 return zholes_size[zone_type]; 4693 } 4694 4695 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4696 4697 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 4698 unsigned long node_start_pfn, 4699 unsigned long node_end_pfn, 4700 unsigned long *zones_size, 4701 unsigned long *zholes_size) 4702 { 4703 unsigned long realtotalpages, totalpages = 0; 4704 enum zone_type i; 4705 4706 for (i = 0; i < MAX_NR_ZONES; i++) 4707 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 4708 node_start_pfn, 4709 node_end_pfn, 4710 zones_size); 4711 pgdat->node_spanned_pages = totalpages; 4712 4713 realtotalpages = totalpages; 4714 for (i = 0; i < MAX_NR_ZONES; i++) 4715 realtotalpages -= 4716 zone_absent_pages_in_node(pgdat->node_id, i, 4717 node_start_pfn, node_end_pfn, 4718 zholes_size); 4719 pgdat->node_present_pages = realtotalpages; 4720 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 4721 realtotalpages); 4722 } 4723 4724 #ifndef CONFIG_SPARSEMEM 4725 /* 4726 * Calculate the size of the zone->blockflags rounded to an unsigned long 4727 * Start by making sure zonesize is a multiple of pageblock_order by rounding 4728 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 4729 * round what is now in bits to nearest long in bits, then return it in 4730 * bytes. 4731 */ 4732 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize) 4733 { 4734 unsigned long usemapsize; 4735 4736 zonesize += zone_start_pfn & (pageblock_nr_pages-1); 4737 usemapsize = roundup(zonesize, pageblock_nr_pages); 4738 usemapsize = usemapsize >> pageblock_order; 4739 usemapsize *= NR_PAGEBLOCK_BITS; 4740 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 4741 4742 return usemapsize / 8; 4743 } 4744 4745 static void __init setup_usemap(struct pglist_data *pgdat, 4746 struct zone *zone, 4747 unsigned long zone_start_pfn, 4748 unsigned long zonesize) 4749 { 4750 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize); 4751 zone->pageblock_flags = NULL; 4752 if (usemapsize) 4753 zone->pageblock_flags = 4754 memblock_virt_alloc_node_nopanic(usemapsize, 4755 pgdat->node_id); 4756 } 4757 #else 4758 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone, 4759 unsigned long zone_start_pfn, unsigned long zonesize) {} 4760 #endif /* CONFIG_SPARSEMEM */ 4761 4762 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 4763 4764 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 4765 void __paginginit set_pageblock_order(void) 4766 { 4767 unsigned int order; 4768 4769 /* Check that pageblock_nr_pages has not already been setup */ 4770 if (pageblock_order) 4771 return; 4772 4773 if (HPAGE_SHIFT > PAGE_SHIFT) 4774 order = HUGETLB_PAGE_ORDER; 4775 else 4776 order = MAX_ORDER - 1; 4777 4778 /* 4779 * Assume the largest contiguous order of interest is a huge page. 4780 * This value may be variable depending on boot parameters on IA64 and 4781 * powerpc. 4782 */ 4783 pageblock_order = order; 4784 } 4785 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4786 4787 /* 4788 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 4789 * is unused as pageblock_order is set at compile-time. See 4790 * include/linux/pageblock-flags.h for the values of pageblock_order based on 4791 * the kernel config 4792 */ 4793 void __paginginit set_pageblock_order(void) 4794 { 4795 } 4796 4797 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4798 4799 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages, 4800 unsigned long present_pages) 4801 { 4802 unsigned long pages = spanned_pages; 4803 4804 /* 4805 * Provide a more accurate estimation if there are holes within 4806 * the zone and SPARSEMEM is in use. If there are holes within the 4807 * zone, each populated memory region may cost us one or two extra 4808 * memmap pages due to alignment because memmap pages for each 4809 * populated regions may not naturally algined on page boundary. 4810 * So the (present_pages >> 4) heuristic is a tradeoff for that. 4811 */ 4812 if (spanned_pages > present_pages + (present_pages >> 4) && 4813 IS_ENABLED(CONFIG_SPARSEMEM)) 4814 pages = present_pages; 4815 4816 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT; 4817 } 4818 4819 /* 4820 * Set up the zone data structures: 4821 * - mark all pages reserved 4822 * - mark all memory queues empty 4823 * - clear the memory bitmaps 4824 * 4825 * NOTE: pgdat should get zeroed by caller. 4826 */ 4827 static void __paginginit free_area_init_core(struct pglist_data *pgdat, 4828 unsigned long node_start_pfn, unsigned long node_end_pfn, 4829 unsigned long *zones_size, unsigned long *zholes_size) 4830 { 4831 enum zone_type j; 4832 int nid = pgdat->node_id; 4833 unsigned long zone_start_pfn = pgdat->node_start_pfn; 4834 int ret; 4835 4836 pgdat_resize_init(pgdat); 4837 #ifdef CONFIG_NUMA_BALANCING 4838 spin_lock_init(&pgdat->numabalancing_migrate_lock); 4839 pgdat->numabalancing_migrate_nr_pages = 0; 4840 pgdat->numabalancing_migrate_next_window = jiffies; 4841 #endif 4842 init_waitqueue_head(&pgdat->kswapd_wait); 4843 init_waitqueue_head(&pgdat->pfmemalloc_wait); 4844 pgdat_page_cgroup_init(pgdat); 4845 4846 for (j = 0; j < MAX_NR_ZONES; j++) { 4847 struct zone *zone = pgdat->node_zones + j; 4848 unsigned long size, realsize, freesize, memmap_pages; 4849 4850 size = zone_spanned_pages_in_node(nid, j, node_start_pfn, 4851 node_end_pfn, zones_size); 4852 realsize = freesize = size - zone_absent_pages_in_node(nid, j, 4853 node_start_pfn, 4854 node_end_pfn, 4855 zholes_size); 4856 4857 /* 4858 * Adjust freesize so that it accounts for how much memory 4859 * is used by this zone for memmap. This affects the watermark 4860 * and per-cpu initialisations 4861 */ 4862 memmap_pages = calc_memmap_size(size, realsize); 4863 if (freesize >= memmap_pages) { 4864 freesize -= memmap_pages; 4865 if (memmap_pages) 4866 printk(KERN_DEBUG 4867 " %s zone: %lu pages used for memmap\n", 4868 zone_names[j], memmap_pages); 4869 } else 4870 printk(KERN_WARNING 4871 " %s zone: %lu pages exceeds freesize %lu\n", 4872 zone_names[j], memmap_pages, freesize); 4873 4874 /* Account for reserved pages */ 4875 if (j == 0 && freesize > dma_reserve) { 4876 freesize -= dma_reserve; 4877 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 4878 zone_names[0], dma_reserve); 4879 } 4880 4881 if (!is_highmem_idx(j)) 4882 nr_kernel_pages += freesize; 4883 /* Charge for highmem memmap if there are enough kernel pages */ 4884 else if (nr_kernel_pages > memmap_pages * 2) 4885 nr_kernel_pages -= memmap_pages; 4886 nr_all_pages += freesize; 4887 4888 zone->spanned_pages = size; 4889 zone->present_pages = realsize; 4890 /* 4891 * Set an approximate value for lowmem here, it will be adjusted 4892 * when the bootmem allocator frees pages into the buddy system. 4893 * And all highmem pages will be managed by the buddy system. 4894 */ 4895 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize; 4896 #ifdef CONFIG_NUMA 4897 zone->node = nid; 4898 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio) 4899 / 100; 4900 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100; 4901 #endif 4902 zone->name = zone_names[j]; 4903 spin_lock_init(&zone->lock); 4904 spin_lock_init(&zone->lru_lock); 4905 zone_seqlock_init(zone); 4906 zone->zone_pgdat = pgdat; 4907 zone_pcp_init(zone); 4908 4909 /* For bootup, initialized properly in watermark setup */ 4910 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages); 4911 4912 lruvec_init(&zone->lruvec); 4913 if (!size) 4914 continue; 4915 4916 set_pageblock_order(); 4917 setup_usemap(pgdat, zone, zone_start_pfn, size); 4918 ret = init_currently_empty_zone(zone, zone_start_pfn, 4919 size, MEMMAP_EARLY); 4920 BUG_ON(ret); 4921 memmap_init(size, nid, j, zone_start_pfn); 4922 zone_start_pfn += size; 4923 } 4924 } 4925 4926 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 4927 { 4928 /* Skip empty nodes */ 4929 if (!pgdat->node_spanned_pages) 4930 return; 4931 4932 #ifdef CONFIG_FLAT_NODE_MEM_MAP 4933 /* ia64 gets its own node_mem_map, before this, without bootmem */ 4934 if (!pgdat->node_mem_map) { 4935 unsigned long size, start, end; 4936 struct page *map; 4937 4938 /* 4939 * The zone's endpoints aren't required to be MAX_ORDER 4940 * aligned but the node_mem_map endpoints must be in order 4941 * for the buddy allocator to function correctly. 4942 */ 4943 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 4944 end = pgdat_end_pfn(pgdat); 4945 end = ALIGN(end, MAX_ORDER_NR_PAGES); 4946 size = (end - start) * sizeof(struct page); 4947 map = alloc_remap(pgdat->node_id, size); 4948 if (!map) 4949 map = memblock_virt_alloc_node_nopanic(size, 4950 pgdat->node_id); 4951 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 4952 } 4953 #ifndef CONFIG_NEED_MULTIPLE_NODES 4954 /* 4955 * With no DISCONTIG, the global mem_map is just set as node 0's 4956 */ 4957 if (pgdat == NODE_DATA(0)) { 4958 mem_map = NODE_DATA(0)->node_mem_map; 4959 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 4960 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 4961 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 4962 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4963 } 4964 #endif 4965 #endif /* CONFIG_FLAT_NODE_MEM_MAP */ 4966 } 4967 4968 void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 4969 unsigned long node_start_pfn, unsigned long *zholes_size) 4970 { 4971 pg_data_t *pgdat = NODE_DATA(nid); 4972 unsigned long start_pfn = 0; 4973 unsigned long end_pfn = 0; 4974 4975 /* pg_data_t should be reset to zero when it's allocated */ 4976 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx); 4977 4978 pgdat->node_id = nid; 4979 pgdat->node_start_pfn = node_start_pfn; 4980 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 4981 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); 4982 printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid, 4983 (u64) start_pfn << PAGE_SHIFT, (u64) (end_pfn << PAGE_SHIFT) - 1); 4984 #endif 4985 calculate_node_totalpages(pgdat, start_pfn, end_pfn, 4986 zones_size, zholes_size); 4987 4988 alloc_node_mem_map(pgdat); 4989 #ifdef CONFIG_FLAT_NODE_MEM_MAP 4990 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 4991 nid, (unsigned long)pgdat, 4992 (unsigned long)pgdat->node_mem_map); 4993 #endif 4994 4995 free_area_init_core(pgdat, start_pfn, end_pfn, 4996 zones_size, zholes_size); 4997 } 4998 4999 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5000 5001 #if MAX_NUMNODES > 1 5002 /* 5003 * Figure out the number of possible node ids. 5004 */ 5005 void __init setup_nr_node_ids(void) 5006 { 5007 unsigned int node; 5008 unsigned int highest = 0; 5009 5010 for_each_node_mask(node, node_possible_map) 5011 highest = node; 5012 nr_node_ids = highest + 1; 5013 } 5014 #endif 5015 5016 /** 5017 * node_map_pfn_alignment - determine the maximum internode alignment 5018 * 5019 * This function should be called after node map is populated and sorted. 5020 * It calculates the maximum power of two alignment which can distinguish 5021 * all the nodes. 5022 * 5023 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value 5024 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the 5025 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is 5026 * shifted, 1GiB is enough and this function will indicate so. 5027 * 5028 * This is used to test whether pfn -> nid mapping of the chosen memory 5029 * model has fine enough granularity to avoid incorrect mapping for the 5030 * populated node map. 5031 * 5032 * Returns the determined alignment in pfn's. 0 if there is no alignment 5033 * requirement (single node). 5034 */ 5035 unsigned long __init node_map_pfn_alignment(void) 5036 { 5037 unsigned long accl_mask = 0, last_end = 0; 5038 unsigned long start, end, mask; 5039 int last_nid = -1; 5040 int i, nid; 5041 5042 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) { 5043 if (!start || last_nid < 0 || last_nid == nid) { 5044 last_nid = nid; 5045 last_end = end; 5046 continue; 5047 } 5048 5049 /* 5050 * Start with a mask granular enough to pin-point to the 5051 * start pfn and tick off bits one-by-one until it becomes 5052 * too coarse to separate the current node from the last. 5053 */ 5054 mask = ~((1 << __ffs(start)) - 1); 5055 while (mask && last_end <= (start & (mask << 1))) 5056 mask <<= 1; 5057 5058 /* accumulate all internode masks */ 5059 accl_mask |= mask; 5060 } 5061 5062 /* convert mask to number of pages */ 5063 return ~accl_mask + 1; 5064 } 5065 5066 /* Find the lowest pfn for a node */ 5067 static unsigned long __init find_min_pfn_for_node(int nid) 5068 { 5069 unsigned long min_pfn = ULONG_MAX; 5070 unsigned long start_pfn; 5071 int i; 5072 5073 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL) 5074 min_pfn = min(min_pfn, start_pfn); 5075 5076 if (min_pfn == ULONG_MAX) { 5077 printk(KERN_WARNING 5078 "Could not find start_pfn for node %d\n", nid); 5079 return 0; 5080 } 5081 5082 return min_pfn; 5083 } 5084 5085 /** 5086 * find_min_pfn_with_active_regions - Find the minimum PFN registered 5087 * 5088 * It returns the minimum PFN based on information provided via 5089 * memblock_set_node(). 5090 */ 5091 unsigned long __init find_min_pfn_with_active_regions(void) 5092 { 5093 return find_min_pfn_for_node(MAX_NUMNODES); 5094 } 5095 5096 /* 5097 * early_calculate_totalpages() 5098 * Sum pages in active regions for movable zone. 5099 * Populate N_MEMORY for calculating usable_nodes. 5100 */ 5101 static unsigned long __init early_calculate_totalpages(void) 5102 { 5103 unsigned long totalpages = 0; 5104 unsigned long start_pfn, end_pfn; 5105 int i, nid; 5106 5107 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { 5108 unsigned long pages = end_pfn - start_pfn; 5109 5110 totalpages += pages; 5111 if (pages) 5112 node_set_state(nid, N_MEMORY); 5113 } 5114 return totalpages; 5115 } 5116 5117 /* 5118 * Find the PFN the Movable zone begins in each node. Kernel memory 5119 * is spread evenly between nodes as long as the nodes have enough 5120 * memory. When they don't, some nodes will have more kernelcore than 5121 * others 5122 */ 5123 static void __init find_zone_movable_pfns_for_nodes(void) 5124 { 5125 int i, nid; 5126 unsigned long usable_startpfn; 5127 unsigned long kernelcore_node, kernelcore_remaining; 5128 /* save the state before borrow the nodemask */ 5129 nodemask_t saved_node_state = node_states[N_MEMORY]; 5130 unsigned long totalpages = early_calculate_totalpages(); 5131 int usable_nodes = nodes_weight(node_states[N_MEMORY]); 5132 struct memblock_region *r; 5133 5134 /* Need to find movable_zone earlier when movable_node is specified. */ 5135 find_usable_zone_for_movable(); 5136 5137 /* 5138 * If movable_node is specified, ignore kernelcore and movablecore 5139 * options. 5140 */ 5141 if (movable_node_is_enabled()) { 5142 for_each_memblock(memory, r) { 5143 if (!memblock_is_hotpluggable(r)) 5144 continue; 5145 5146 nid = r->nid; 5147 5148 usable_startpfn = PFN_DOWN(r->base); 5149 zone_movable_pfn[nid] = zone_movable_pfn[nid] ? 5150 min(usable_startpfn, zone_movable_pfn[nid]) : 5151 usable_startpfn; 5152 } 5153 5154 goto out2; 5155 } 5156 5157 /* 5158 * If movablecore=nn[KMG] was specified, calculate what size of 5159 * kernelcore that corresponds so that memory usable for 5160 * any allocation type is evenly spread. If both kernelcore 5161 * and movablecore are specified, then the value of kernelcore 5162 * will be used for required_kernelcore if it's greater than 5163 * what movablecore would have allowed. 5164 */ 5165 if (required_movablecore) { 5166 unsigned long corepages; 5167 5168 /* 5169 * Round-up so that ZONE_MOVABLE is at least as large as what 5170 * was requested by the user 5171 */ 5172 required_movablecore = 5173 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 5174 corepages = totalpages - required_movablecore; 5175 5176 required_kernelcore = max(required_kernelcore, corepages); 5177 } 5178 5179 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 5180 if (!required_kernelcore) 5181 goto out; 5182 5183 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 5184 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 5185 5186 restart: 5187 /* Spread kernelcore memory as evenly as possible throughout nodes */ 5188 kernelcore_node = required_kernelcore / usable_nodes; 5189 for_each_node_state(nid, N_MEMORY) { 5190 unsigned long start_pfn, end_pfn; 5191 5192 /* 5193 * Recalculate kernelcore_node if the division per node 5194 * now exceeds what is necessary to satisfy the requested 5195 * amount of memory for the kernel 5196 */ 5197 if (required_kernelcore < kernelcore_node) 5198 kernelcore_node = required_kernelcore / usable_nodes; 5199 5200 /* 5201 * As the map is walked, we track how much memory is usable 5202 * by the kernel using kernelcore_remaining. When it is 5203 * 0, the rest of the node is usable by ZONE_MOVABLE 5204 */ 5205 kernelcore_remaining = kernelcore_node; 5206 5207 /* Go through each range of PFNs within this node */ 5208 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 5209 unsigned long size_pages; 5210 5211 start_pfn = max(start_pfn, zone_movable_pfn[nid]); 5212 if (start_pfn >= end_pfn) 5213 continue; 5214 5215 /* Account for what is only usable for kernelcore */ 5216 if (start_pfn < usable_startpfn) { 5217 unsigned long kernel_pages; 5218 kernel_pages = min(end_pfn, usable_startpfn) 5219 - start_pfn; 5220 5221 kernelcore_remaining -= min(kernel_pages, 5222 kernelcore_remaining); 5223 required_kernelcore -= min(kernel_pages, 5224 required_kernelcore); 5225 5226 /* Continue if range is now fully accounted */ 5227 if (end_pfn <= usable_startpfn) { 5228 5229 /* 5230 * Push zone_movable_pfn to the end so 5231 * that if we have to rebalance 5232 * kernelcore across nodes, we will 5233 * not double account here 5234 */ 5235 zone_movable_pfn[nid] = end_pfn; 5236 continue; 5237 } 5238 start_pfn = usable_startpfn; 5239 } 5240 5241 /* 5242 * The usable PFN range for ZONE_MOVABLE is from 5243 * start_pfn->end_pfn. Calculate size_pages as the 5244 * number of pages used as kernelcore 5245 */ 5246 size_pages = end_pfn - start_pfn; 5247 if (size_pages > kernelcore_remaining) 5248 size_pages = kernelcore_remaining; 5249 zone_movable_pfn[nid] = start_pfn + size_pages; 5250 5251 /* 5252 * Some kernelcore has been met, update counts and 5253 * break if the kernelcore for this node has been 5254 * satisfied 5255 */ 5256 required_kernelcore -= min(required_kernelcore, 5257 size_pages); 5258 kernelcore_remaining -= size_pages; 5259 if (!kernelcore_remaining) 5260 break; 5261 } 5262 } 5263 5264 /* 5265 * If there is still required_kernelcore, we do another pass with one 5266 * less node in the count. This will push zone_movable_pfn[nid] further 5267 * along on the nodes that still have memory until kernelcore is 5268 * satisfied 5269 */ 5270 usable_nodes--; 5271 if (usable_nodes && required_kernelcore > usable_nodes) 5272 goto restart; 5273 5274 out2: 5275 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 5276 for (nid = 0; nid < MAX_NUMNODES; nid++) 5277 zone_movable_pfn[nid] = 5278 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 5279 5280 out: 5281 /* restore the node_state */ 5282 node_states[N_MEMORY] = saved_node_state; 5283 } 5284 5285 /* Any regular or high memory on that node ? */ 5286 static void check_for_memory(pg_data_t *pgdat, int nid) 5287 { 5288 enum zone_type zone_type; 5289 5290 if (N_MEMORY == N_NORMAL_MEMORY) 5291 return; 5292 5293 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) { 5294 struct zone *zone = &pgdat->node_zones[zone_type]; 5295 if (populated_zone(zone)) { 5296 node_set_state(nid, N_HIGH_MEMORY); 5297 if (N_NORMAL_MEMORY != N_HIGH_MEMORY && 5298 zone_type <= ZONE_NORMAL) 5299 node_set_state(nid, N_NORMAL_MEMORY); 5300 break; 5301 } 5302 } 5303 } 5304 5305 /** 5306 * free_area_init_nodes - Initialise all pg_data_t and zone data 5307 * @max_zone_pfn: an array of max PFNs for each zone 5308 * 5309 * This will call free_area_init_node() for each active node in the system. 5310 * Using the page ranges provided by memblock_set_node(), the size of each 5311 * zone in each node and their holes is calculated. If the maximum PFN 5312 * between two adjacent zones match, it is assumed that the zone is empty. 5313 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 5314 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 5315 * starts where the previous one ended. For example, ZONE_DMA32 starts 5316 * at arch_max_dma_pfn. 5317 */ 5318 void __init free_area_init_nodes(unsigned long *max_zone_pfn) 5319 { 5320 unsigned long start_pfn, end_pfn; 5321 int i, nid; 5322 5323 /* Record where the zone boundaries are */ 5324 memset(arch_zone_lowest_possible_pfn, 0, 5325 sizeof(arch_zone_lowest_possible_pfn)); 5326 memset(arch_zone_highest_possible_pfn, 0, 5327 sizeof(arch_zone_highest_possible_pfn)); 5328 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 5329 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 5330 for (i = 1; i < MAX_NR_ZONES; i++) { 5331 if (i == ZONE_MOVABLE) 5332 continue; 5333 arch_zone_lowest_possible_pfn[i] = 5334 arch_zone_highest_possible_pfn[i-1]; 5335 arch_zone_highest_possible_pfn[i] = 5336 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 5337 } 5338 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 5339 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 5340 5341 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 5342 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 5343 find_zone_movable_pfns_for_nodes(); 5344 5345 /* Print out the zone ranges */ 5346 printk("Zone ranges:\n"); 5347 for (i = 0; i < MAX_NR_ZONES; i++) { 5348 if (i == ZONE_MOVABLE) 5349 continue; 5350 printk(KERN_CONT " %-8s ", zone_names[i]); 5351 if (arch_zone_lowest_possible_pfn[i] == 5352 arch_zone_highest_possible_pfn[i]) 5353 printk(KERN_CONT "empty\n"); 5354 else 5355 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n", 5356 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT, 5357 (arch_zone_highest_possible_pfn[i] 5358 << PAGE_SHIFT) - 1); 5359 } 5360 5361 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 5362 printk("Movable zone start for each node\n"); 5363 for (i = 0; i < MAX_NUMNODES; i++) { 5364 if (zone_movable_pfn[i]) 5365 printk(" Node %d: %#010lx\n", i, 5366 zone_movable_pfn[i] << PAGE_SHIFT); 5367 } 5368 5369 /* Print out the early node map */ 5370 printk("Early memory node ranges\n"); 5371 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) 5372 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid, 5373 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1); 5374 5375 /* Initialise every node */ 5376 mminit_verify_pageflags_layout(); 5377 setup_nr_node_ids(); 5378 for_each_online_node(nid) { 5379 pg_data_t *pgdat = NODE_DATA(nid); 5380 free_area_init_node(nid, NULL, 5381 find_min_pfn_for_node(nid), NULL); 5382 5383 /* Any memory on that node */ 5384 if (pgdat->node_present_pages) 5385 node_set_state(nid, N_MEMORY); 5386 check_for_memory(pgdat, nid); 5387 } 5388 } 5389 5390 static int __init cmdline_parse_core(char *p, unsigned long *core) 5391 { 5392 unsigned long long coremem; 5393 if (!p) 5394 return -EINVAL; 5395 5396 coremem = memparse(p, &p); 5397 *core = coremem >> PAGE_SHIFT; 5398 5399 /* Paranoid check that UL is enough for the coremem value */ 5400 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 5401 5402 return 0; 5403 } 5404 5405 /* 5406 * kernelcore=size sets the amount of memory for use for allocations that 5407 * cannot be reclaimed or migrated. 5408 */ 5409 static int __init cmdline_parse_kernelcore(char *p) 5410 { 5411 return cmdline_parse_core(p, &required_kernelcore); 5412 } 5413 5414 /* 5415 * movablecore=size sets the amount of memory for use for allocations that 5416 * can be reclaimed or migrated. 5417 */ 5418 static int __init cmdline_parse_movablecore(char *p) 5419 { 5420 return cmdline_parse_core(p, &required_movablecore); 5421 } 5422 5423 early_param("kernelcore", cmdline_parse_kernelcore); 5424 early_param("movablecore", cmdline_parse_movablecore); 5425 5426 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5427 5428 void adjust_managed_page_count(struct page *page, long count) 5429 { 5430 spin_lock(&managed_page_count_lock); 5431 page_zone(page)->managed_pages += count; 5432 totalram_pages += count; 5433 #ifdef CONFIG_HIGHMEM 5434 if (PageHighMem(page)) 5435 totalhigh_pages += count; 5436 #endif 5437 spin_unlock(&managed_page_count_lock); 5438 } 5439 EXPORT_SYMBOL(adjust_managed_page_count); 5440 5441 unsigned long free_reserved_area(void *start, void *end, int poison, char *s) 5442 { 5443 void *pos; 5444 unsigned long pages = 0; 5445 5446 start = (void *)PAGE_ALIGN((unsigned long)start); 5447 end = (void *)((unsigned long)end & PAGE_MASK); 5448 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) { 5449 if ((unsigned int)poison <= 0xFF) 5450 memset(pos, poison, PAGE_SIZE); 5451 free_reserved_page(virt_to_page(pos)); 5452 } 5453 5454 if (pages && s) 5455 pr_info("Freeing %s memory: %ldK (%p - %p)\n", 5456 s, pages << (PAGE_SHIFT - 10), start, end); 5457 5458 return pages; 5459 } 5460 EXPORT_SYMBOL(free_reserved_area); 5461 5462 #ifdef CONFIG_HIGHMEM 5463 void free_highmem_page(struct page *page) 5464 { 5465 __free_reserved_page(page); 5466 totalram_pages++; 5467 page_zone(page)->managed_pages++; 5468 totalhigh_pages++; 5469 } 5470 #endif 5471 5472 5473 void __init mem_init_print_info(const char *str) 5474 { 5475 unsigned long physpages, codesize, datasize, rosize, bss_size; 5476 unsigned long init_code_size, init_data_size; 5477 5478 physpages = get_num_physpages(); 5479 codesize = _etext - _stext; 5480 datasize = _edata - _sdata; 5481 rosize = __end_rodata - __start_rodata; 5482 bss_size = __bss_stop - __bss_start; 5483 init_data_size = __init_end - __init_begin; 5484 init_code_size = _einittext - _sinittext; 5485 5486 /* 5487 * Detect special cases and adjust section sizes accordingly: 5488 * 1) .init.* may be embedded into .data sections 5489 * 2) .init.text.* may be out of [__init_begin, __init_end], 5490 * please refer to arch/tile/kernel/vmlinux.lds.S. 5491 * 3) .rodata.* may be embedded into .text or .data sections. 5492 */ 5493 #define adj_init_size(start, end, size, pos, adj) \ 5494 do { \ 5495 if (start <= pos && pos < end && size > adj) \ 5496 size -= adj; \ 5497 } while (0) 5498 5499 adj_init_size(__init_begin, __init_end, init_data_size, 5500 _sinittext, init_code_size); 5501 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size); 5502 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size); 5503 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize); 5504 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize); 5505 5506 #undef adj_init_size 5507 5508 printk("Memory: %luK/%luK available " 5509 "(%luK kernel code, %luK rwdata, %luK rodata, " 5510 "%luK init, %luK bss, %luK reserved" 5511 #ifdef CONFIG_HIGHMEM 5512 ", %luK highmem" 5513 #endif 5514 "%s%s)\n", 5515 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10), 5516 codesize >> 10, datasize >> 10, rosize >> 10, 5517 (init_data_size + init_code_size) >> 10, bss_size >> 10, 5518 (physpages - totalram_pages) << (PAGE_SHIFT-10), 5519 #ifdef CONFIG_HIGHMEM 5520 totalhigh_pages << (PAGE_SHIFT-10), 5521 #endif 5522 str ? ", " : "", str ? str : ""); 5523 } 5524 5525 /** 5526 * set_dma_reserve - set the specified number of pages reserved in the first zone 5527 * @new_dma_reserve: The number of pages to mark reserved 5528 * 5529 * The per-cpu batchsize and zone watermarks are determined by present_pages. 5530 * In the DMA zone, a significant percentage may be consumed by kernel image 5531 * and other unfreeable allocations which can skew the watermarks badly. This 5532 * function may optionally be used to account for unfreeable pages in the 5533 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 5534 * smaller per-cpu batchsize. 5535 */ 5536 void __init set_dma_reserve(unsigned long new_dma_reserve) 5537 { 5538 dma_reserve = new_dma_reserve; 5539 } 5540 5541 void __init free_area_init(unsigned long *zones_size) 5542 { 5543 free_area_init_node(0, zones_size, 5544 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 5545 } 5546 5547 static int page_alloc_cpu_notify(struct notifier_block *self, 5548 unsigned long action, void *hcpu) 5549 { 5550 int cpu = (unsigned long)hcpu; 5551 5552 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 5553 lru_add_drain_cpu(cpu); 5554 drain_pages(cpu); 5555 5556 /* 5557 * Spill the event counters of the dead processor 5558 * into the current processors event counters. 5559 * This artificially elevates the count of the current 5560 * processor. 5561 */ 5562 vm_events_fold_cpu(cpu); 5563 5564 /* 5565 * Zero the differential counters of the dead processor 5566 * so that the vm statistics are consistent. 5567 * 5568 * This is only okay since the processor is dead and cannot 5569 * race with what we are doing. 5570 */ 5571 cpu_vm_stats_fold(cpu); 5572 } 5573 return NOTIFY_OK; 5574 } 5575 5576 void __init page_alloc_init(void) 5577 { 5578 hotcpu_notifier(page_alloc_cpu_notify, 0); 5579 } 5580 5581 /* 5582 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 5583 * or min_free_kbytes changes. 5584 */ 5585 static void calculate_totalreserve_pages(void) 5586 { 5587 struct pglist_data *pgdat; 5588 unsigned long reserve_pages = 0; 5589 enum zone_type i, j; 5590 5591 for_each_online_pgdat(pgdat) { 5592 for (i = 0; i < MAX_NR_ZONES; i++) { 5593 struct zone *zone = pgdat->node_zones + i; 5594 long max = 0; 5595 5596 /* Find valid and maximum lowmem_reserve in the zone */ 5597 for (j = i; j < MAX_NR_ZONES; j++) { 5598 if (zone->lowmem_reserve[j] > max) 5599 max = zone->lowmem_reserve[j]; 5600 } 5601 5602 /* we treat the high watermark as reserved pages. */ 5603 max += high_wmark_pages(zone); 5604 5605 if (max > zone->managed_pages) 5606 max = zone->managed_pages; 5607 reserve_pages += max; 5608 /* 5609 * Lowmem reserves are not available to 5610 * GFP_HIGHUSER page cache allocations and 5611 * kswapd tries to balance zones to their high 5612 * watermark. As a result, neither should be 5613 * regarded as dirtyable memory, to prevent a 5614 * situation where reclaim has to clean pages 5615 * in order to balance the zones. 5616 */ 5617 zone->dirty_balance_reserve = max; 5618 } 5619 } 5620 dirty_balance_reserve = reserve_pages; 5621 totalreserve_pages = reserve_pages; 5622 } 5623 5624 /* 5625 * setup_per_zone_lowmem_reserve - called whenever 5626 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 5627 * has a correct pages reserved value, so an adequate number of 5628 * pages are left in the zone after a successful __alloc_pages(). 5629 */ 5630 static void setup_per_zone_lowmem_reserve(void) 5631 { 5632 struct pglist_data *pgdat; 5633 enum zone_type j, idx; 5634 5635 for_each_online_pgdat(pgdat) { 5636 for (j = 0; j < MAX_NR_ZONES; j++) { 5637 struct zone *zone = pgdat->node_zones + j; 5638 unsigned long managed_pages = zone->managed_pages; 5639 5640 zone->lowmem_reserve[j] = 0; 5641 5642 idx = j; 5643 while (idx) { 5644 struct zone *lower_zone; 5645 5646 idx--; 5647 5648 if (sysctl_lowmem_reserve_ratio[idx] < 1) 5649 sysctl_lowmem_reserve_ratio[idx] = 1; 5650 5651 lower_zone = pgdat->node_zones + idx; 5652 lower_zone->lowmem_reserve[j] = managed_pages / 5653 sysctl_lowmem_reserve_ratio[idx]; 5654 managed_pages += lower_zone->managed_pages; 5655 } 5656 } 5657 } 5658 5659 /* update totalreserve_pages */ 5660 calculate_totalreserve_pages(); 5661 } 5662 5663 static void __setup_per_zone_wmarks(void) 5664 { 5665 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 5666 unsigned long lowmem_pages = 0; 5667 struct zone *zone; 5668 unsigned long flags; 5669 5670 /* Calculate total number of !ZONE_HIGHMEM pages */ 5671 for_each_zone(zone) { 5672 if (!is_highmem(zone)) 5673 lowmem_pages += zone->managed_pages; 5674 } 5675 5676 for_each_zone(zone) { 5677 u64 tmp; 5678 5679 spin_lock_irqsave(&zone->lock, flags); 5680 tmp = (u64)pages_min * zone->managed_pages; 5681 do_div(tmp, lowmem_pages); 5682 if (is_highmem(zone)) { 5683 /* 5684 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 5685 * need highmem pages, so cap pages_min to a small 5686 * value here. 5687 * 5688 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 5689 * deltas controls asynch page reclaim, and so should 5690 * not be capped for highmem. 5691 */ 5692 unsigned long min_pages; 5693 5694 min_pages = zone->managed_pages / 1024; 5695 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL); 5696 zone->watermark[WMARK_MIN] = min_pages; 5697 } else { 5698 /* 5699 * If it's a lowmem zone, reserve a number of pages 5700 * proportionate to the zone's size. 5701 */ 5702 zone->watermark[WMARK_MIN] = tmp; 5703 } 5704 5705 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 5706 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 5707 5708 __mod_zone_page_state(zone, NR_ALLOC_BATCH, 5709 high_wmark_pages(zone) - low_wmark_pages(zone) - 5710 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH])); 5711 5712 setup_zone_migrate_reserve(zone); 5713 spin_unlock_irqrestore(&zone->lock, flags); 5714 } 5715 5716 /* update totalreserve_pages */ 5717 calculate_totalreserve_pages(); 5718 } 5719 5720 /** 5721 * setup_per_zone_wmarks - called when min_free_kbytes changes 5722 * or when memory is hot-{added|removed} 5723 * 5724 * Ensures that the watermark[min,low,high] values for each zone are set 5725 * correctly with respect to min_free_kbytes. 5726 */ 5727 void setup_per_zone_wmarks(void) 5728 { 5729 mutex_lock(&zonelists_mutex); 5730 __setup_per_zone_wmarks(); 5731 mutex_unlock(&zonelists_mutex); 5732 } 5733 5734 /* 5735 * The inactive anon list should be small enough that the VM never has to 5736 * do too much work, but large enough that each inactive page has a chance 5737 * to be referenced again before it is swapped out. 5738 * 5739 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 5740 * INACTIVE_ANON pages on this zone's LRU, maintained by the 5741 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 5742 * the anonymous pages are kept on the inactive list. 5743 * 5744 * total target max 5745 * memory ratio inactive anon 5746 * ------------------------------------- 5747 * 10MB 1 5MB 5748 * 100MB 1 50MB 5749 * 1GB 3 250MB 5750 * 10GB 10 0.9GB 5751 * 100GB 31 3GB 5752 * 1TB 101 10GB 5753 * 10TB 320 32GB 5754 */ 5755 static void __meminit calculate_zone_inactive_ratio(struct zone *zone) 5756 { 5757 unsigned int gb, ratio; 5758 5759 /* Zone size in gigabytes */ 5760 gb = zone->managed_pages >> (30 - PAGE_SHIFT); 5761 if (gb) 5762 ratio = int_sqrt(10 * gb); 5763 else 5764 ratio = 1; 5765 5766 zone->inactive_ratio = ratio; 5767 } 5768 5769 static void __meminit setup_per_zone_inactive_ratio(void) 5770 { 5771 struct zone *zone; 5772 5773 for_each_zone(zone) 5774 calculate_zone_inactive_ratio(zone); 5775 } 5776 5777 /* 5778 * Initialise min_free_kbytes. 5779 * 5780 * For small machines we want it small (128k min). For large machines 5781 * we want it large (64MB max). But it is not linear, because network 5782 * bandwidth does not increase linearly with machine size. We use 5783 * 5784 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 5785 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 5786 * 5787 * which yields 5788 * 5789 * 16MB: 512k 5790 * 32MB: 724k 5791 * 64MB: 1024k 5792 * 128MB: 1448k 5793 * 256MB: 2048k 5794 * 512MB: 2896k 5795 * 1024MB: 4096k 5796 * 2048MB: 5792k 5797 * 4096MB: 8192k 5798 * 8192MB: 11584k 5799 * 16384MB: 16384k 5800 */ 5801 int __meminit init_per_zone_wmark_min(void) 5802 { 5803 unsigned long lowmem_kbytes; 5804 int new_min_free_kbytes; 5805 5806 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 5807 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 5808 5809 if (new_min_free_kbytes > user_min_free_kbytes) { 5810 min_free_kbytes = new_min_free_kbytes; 5811 if (min_free_kbytes < 128) 5812 min_free_kbytes = 128; 5813 if (min_free_kbytes > 65536) 5814 min_free_kbytes = 65536; 5815 } else { 5816 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n", 5817 new_min_free_kbytes, user_min_free_kbytes); 5818 } 5819 setup_per_zone_wmarks(); 5820 refresh_zone_stat_thresholds(); 5821 setup_per_zone_lowmem_reserve(); 5822 setup_per_zone_inactive_ratio(); 5823 return 0; 5824 } 5825 module_init(init_per_zone_wmark_min) 5826 5827 /* 5828 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 5829 * that we can call two helper functions whenever min_free_kbytes 5830 * changes. 5831 */ 5832 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write, 5833 void __user *buffer, size_t *length, loff_t *ppos) 5834 { 5835 int rc; 5836 5837 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5838 if (rc) 5839 return rc; 5840 5841 if (write) { 5842 user_min_free_kbytes = min_free_kbytes; 5843 setup_per_zone_wmarks(); 5844 } 5845 return 0; 5846 } 5847 5848 #ifdef CONFIG_NUMA 5849 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write, 5850 void __user *buffer, size_t *length, loff_t *ppos) 5851 { 5852 struct zone *zone; 5853 int rc; 5854 5855 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5856 if (rc) 5857 return rc; 5858 5859 for_each_zone(zone) 5860 zone->min_unmapped_pages = (zone->managed_pages * 5861 sysctl_min_unmapped_ratio) / 100; 5862 return 0; 5863 } 5864 5865 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write, 5866 void __user *buffer, size_t *length, loff_t *ppos) 5867 { 5868 struct zone *zone; 5869 int rc; 5870 5871 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5872 if (rc) 5873 return rc; 5874 5875 for_each_zone(zone) 5876 zone->min_slab_pages = (zone->managed_pages * 5877 sysctl_min_slab_ratio) / 100; 5878 return 0; 5879 } 5880 #endif 5881 5882 /* 5883 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 5884 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 5885 * whenever sysctl_lowmem_reserve_ratio changes. 5886 * 5887 * The reserve ratio obviously has absolutely no relation with the 5888 * minimum watermarks. The lowmem reserve ratio can only make sense 5889 * if in function of the boot time zone sizes. 5890 */ 5891 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write, 5892 void __user *buffer, size_t *length, loff_t *ppos) 5893 { 5894 proc_dointvec_minmax(table, write, buffer, length, ppos); 5895 setup_per_zone_lowmem_reserve(); 5896 return 0; 5897 } 5898 5899 /* 5900 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 5901 * cpu. It is the fraction of total pages in each zone that a hot per cpu 5902 * pagelist can have before it gets flushed back to buddy allocator. 5903 */ 5904 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write, 5905 void __user *buffer, size_t *length, loff_t *ppos) 5906 { 5907 struct zone *zone; 5908 int old_percpu_pagelist_fraction; 5909 int ret; 5910 5911 mutex_lock(&pcp_batch_high_lock); 5912 old_percpu_pagelist_fraction = percpu_pagelist_fraction; 5913 5914 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 5915 if (!write || ret < 0) 5916 goto out; 5917 5918 /* Sanity checking to avoid pcp imbalance */ 5919 if (percpu_pagelist_fraction && 5920 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) { 5921 percpu_pagelist_fraction = old_percpu_pagelist_fraction; 5922 ret = -EINVAL; 5923 goto out; 5924 } 5925 5926 /* No change? */ 5927 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction) 5928 goto out; 5929 5930 for_each_populated_zone(zone) { 5931 unsigned int cpu; 5932 5933 for_each_possible_cpu(cpu) 5934 pageset_set_high_and_batch(zone, 5935 per_cpu_ptr(zone->pageset, cpu)); 5936 } 5937 out: 5938 mutex_unlock(&pcp_batch_high_lock); 5939 return ret; 5940 } 5941 5942 int hashdist = HASHDIST_DEFAULT; 5943 5944 #ifdef CONFIG_NUMA 5945 static int __init set_hashdist(char *str) 5946 { 5947 if (!str) 5948 return 0; 5949 hashdist = simple_strtoul(str, &str, 0); 5950 return 1; 5951 } 5952 __setup("hashdist=", set_hashdist); 5953 #endif 5954 5955 /* 5956 * allocate a large system hash table from bootmem 5957 * - it is assumed that the hash table must contain an exact power-of-2 5958 * quantity of entries 5959 * - limit is the number of hash buckets, not the total allocation size 5960 */ 5961 void *__init alloc_large_system_hash(const char *tablename, 5962 unsigned long bucketsize, 5963 unsigned long numentries, 5964 int scale, 5965 int flags, 5966 unsigned int *_hash_shift, 5967 unsigned int *_hash_mask, 5968 unsigned long low_limit, 5969 unsigned long high_limit) 5970 { 5971 unsigned long long max = high_limit; 5972 unsigned long log2qty, size; 5973 void *table = NULL; 5974 5975 /* allow the kernel cmdline to have a say */ 5976 if (!numentries) { 5977 /* round applicable memory size up to nearest megabyte */ 5978 numentries = nr_kernel_pages; 5979 5980 /* It isn't necessary when PAGE_SIZE >= 1MB */ 5981 if (PAGE_SHIFT < 20) 5982 numentries = round_up(numentries, (1<<20)/PAGE_SIZE); 5983 5984 /* limit to 1 bucket per 2^scale bytes of low memory */ 5985 if (scale > PAGE_SHIFT) 5986 numentries >>= (scale - PAGE_SHIFT); 5987 else 5988 numentries <<= (PAGE_SHIFT - scale); 5989 5990 /* Make sure we've got at least a 0-order allocation.. */ 5991 if (unlikely(flags & HASH_SMALL)) { 5992 /* Makes no sense without HASH_EARLY */ 5993 WARN_ON(!(flags & HASH_EARLY)); 5994 if (!(numentries >> *_hash_shift)) { 5995 numentries = 1UL << *_hash_shift; 5996 BUG_ON(!numentries); 5997 } 5998 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 5999 numentries = PAGE_SIZE / bucketsize; 6000 } 6001 numentries = roundup_pow_of_two(numentries); 6002 6003 /* limit allocation size to 1/16 total memory by default */ 6004 if (max == 0) { 6005 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 6006 do_div(max, bucketsize); 6007 } 6008 max = min(max, 0x80000000ULL); 6009 6010 if (numentries < low_limit) 6011 numentries = low_limit; 6012 if (numentries > max) 6013 numentries = max; 6014 6015 log2qty = ilog2(numentries); 6016 6017 do { 6018 size = bucketsize << log2qty; 6019 if (flags & HASH_EARLY) 6020 table = memblock_virt_alloc_nopanic(size, 0); 6021 else if (hashdist) 6022 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 6023 else { 6024 /* 6025 * If bucketsize is not a power-of-two, we may free 6026 * some pages at the end of hash table which 6027 * alloc_pages_exact() automatically does 6028 */ 6029 if (get_order(size) < MAX_ORDER) { 6030 table = alloc_pages_exact(size, GFP_ATOMIC); 6031 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 6032 } 6033 } 6034 } while (!table && size > PAGE_SIZE && --log2qty); 6035 6036 if (!table) 6037 panic("Failed to allocate %s hash table\n", tablename); 6038 6039 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", 6040 tablename, 6041 (1UL << log2qty), 6042 ilog2(size) - PAGE_SHIFT, 6043 size); 6044 6045 if (_hash_shift) 6046 *_hash_shift = log2qty; 6047 if (_hash_mask) 6048 *_hash_mask = (1 << log2qty) - 1; 6049 6050 return table; 6051 } 6052 6053 /* Return a pointer to the bitmap storing bits affecting a block of pages */ 6054 static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 6055 unsigned long pfn) 6056 { 6057 #ifdef CONFIG_SPARSEMEM 6058 return __pfn_to_section(pfn)->pageblock_flags; 6059 #else 6060 return zone->pageblock_flags; 6061 #endif /* CONFIG_SPARSEMEM */ 6062 } 6063 6064 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 6065 { 6066 #ifdef CONFIG_SPARSEMEM 6067 pfn &= (PAGES_PER_SECTION-1); 6068 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 6069 #else 6070 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages); 6071 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 6072 #endif /* CONFIG_SPARSEMEM */ 6073 } 6074 6075 /** 6076 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages 6077 * @page: The page within the block of interest 6078 * @pfn: The target page frame number 6079 * @end_bitidx: The last bit of interest to retrieve 6080 * @mask: mask of bits that the caller is interested in 6081 * 6082 * Return: pageblock_bits flags 6083 */ 6084 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn, 6085 unsigned long end_bitidx, 6086 unsigned long mask) 6087 { 6088 struct zone *zone; 6089 unsigned long *bitmap; 6090 unsigned long bitidx, word_bitidx; 6091 unsigned long word; 6092 6093 zone = page_zone(page); 6094 bitmap = get_pageblock_bitmap(zone, pfn); 6095 bitidx = pfn_to_bitidx(zone, pfn); 6096 word_bitidx = bitidx / BITS_PER_LONG; 6097 bitidx &= (BITS_PER_LONG-1); 6098 6099 word = bitmap[word_bitidx]; 6100 bitidx += end_bitidx; 6101 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask; 6102 } 6103 6104 /** 6105 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages 6106 * @page: The page within the block of interest 6107 * @flags: The flags to set 6108 * @pfn: The target page frame number 6109 * @end_bitidx: The last bit of interest 6110 * @mask: mask of bits that the caller is interested in 6111 */ 6112 void set_pfnblock_flags_mask(struct page *page, unsigned long flags, 6113 unsigned long pfn, 6114 unsigned long end_bitidx, 6115 unsigned long mask) 6116 { 6117 struct zone *zone; 6118 unsigned long *bitmap; 6119 unsigned long bitidx, word_bitidx; 6120 unsigned long old_word, word; 6121 6122 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4); 6123 6124 zone = page_zone(page); 6125 bitmap = get_pageblock_bitmap(zone, pfn); 6126 bitidx = pfn_to_bitidx(zone, pfn); 6127 word_bitidx = bitidx / BITS_PER_LONG; 6128 bitidx &= (BITS_PER_LONG-1); 6129 6130 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page); 6131 6132 bitidx += end_bitidx; 6133 mask <<= (BITS_PER_LONG - bitidx - 1); 6134 flags <<= (BITS_PER_LONG - bitidx - 1); 6135 6136 word = ACCESS_ONCE(bitmap[word_bitidx]); 6137 for (;;) { 6138 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags); 6139 if (word == old_word) 6140 break; 6141 word = old_word; 6142 } 6143 } 6144 6145 /* 6146 * This function checks whether pageblock includes unmovable pages or not. 6147 * If @count is not zero, it is okay to include less @count unmovable pages 6148 * 6149 * PageLRU check without isolation or lru_lock could race so that 6150 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't 6151 * expect this function should be exact. 6152 */ 6153 bool has_unmovable_pages(struct zone *zone, struct page *page, int count, 6154 bool skip_hwpoisoned_pages) 6155 { 6156 unsigned long pfn, iter, found; 6157 int mt; 6158 6159 /* 6160 * For avoiding noise data, lru_add_drain_all() should be called 6161 * If ZONE_MOVABLE, the zone never contains unmovable pages 6162 */ 6163 if (zone_idx(zone) == ZONE_MOVABLE) 6164 return false; 6165 mt = get_pageblock_migratetype(page); 6166 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt)) 6167 return false; 6168 6169 pfn = page_to_pfn(page); 6170 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { 6171 unsigned long check = pfn + iter; 6172 6173 if (!pfn_valid_within(check)) 6174 continue; 6175 6176 page = pfn_to_page(check); 6177 6178 /* 6179 * Hugepages are not in LRU lists, but they're movable. 6180 * We need not scan over tail pages bacause we don't 6181 * handle each tail page individually in migration. 6182 */ 6183 if (PageHuge(page)) { 6184 iter = round_up(iter + 1, 1<<compound_order(page)) - 1; 6185 continue; 6186 } 6187 6188 /* 6189 * We can't use page_count without pin a page 6190 * because another CPU can free compound page. 6191 * This check already skips compound tails of THP 6192 * because their page->_count is zero at all time. 6193 */ 6194 if (!atomic_read(&page->_count)) { 6195 if (PageBuddy(page)) 6196 iter += (1 << page_order(page)) - 1; 6197 continue; 6198 } 6199 6200 /* 6201 * The HWPoisoned page may be not in buddy system, and 6202 * page_count() is not 0. 6203 */ 6204 if (skip_hwpoisoned_pages && PageHWPoison(page)) 6205 continue; 6206 6207 if (!PageLRU(page)) 6208 found++; 6209 /* 6210 * If there are RECLAIMABLE pages, we need to check it. 6211 * But now, memory offline itself doesn't call shrink_slab() 6212 * and it still to be fixed. 6213 */ 6214 /* 6215 * If the page is not RAM, page_count()should be 0. 6216 * we don't need more check. This is an _used_ not-movable page. 6217 * 6218 * The problematic thing here is PG_reserved pages. PG_reserved 6219 * is set to both of a memory hole page and a _used_ kernel 6220 * page at boot. 6221 */ 6222 if (found > count) 6223 return true; 6224 } 6225 return false; 6226 } 6227 6228 bool is_pageblock_removable_nolock(struct page *page) 6229 { 6230 struct zone *zone; 6231 unsigned long pfn; 6232 6233 /* 6234 * We have to be careful here because we are iterating over memory 6235 * sections which are not zone aware so we might end up outside of 6236 * the zone but still within the section. 6237 * We have to take care about the node as well. If the node is offline 6238 * its NODE_DATA will be NULL - see page_zone. 6239 */ 6240 if (!node_online(page_to_nid(page))) 6241 return false; 6242 6243 zone = page_zone(page); 6244 pfn = page_to_pfn(page); 6245 if (!zone_spans_pfn(zone, pfn)) 6246 return false; 6247 6248 return !has_unmovable_pages(zone, page, 0, true); 6249 } 6250 6251 #ifdef CONFIG_CMA 6252 6253 static unsigned long pfn_max_align_down(unsigned long pfn) 6254 { 6255 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES, 6256 pageblock_nr_pages) - 1); 6257 } 6258 6259 static unsigned long pfn_max_align_up(unsigned long pfn) 6260 { 6261 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES, 6262 pageblock_nr_pages)); 6263 } 6264 6265 /* [start, end) must belong to a single zone. */ 6266 static int __alloc_contig_migrate_range(struct compact_control *cc, 6267 unsigned long start, unsigned long end) 6268 { 6269 /* This function is based on compact_zone() from compaction.c. */ 6270 unsigned long nr_reclaimed; 6271 unsigned long pfn = start; 6272 unsigned int tries = 0; 6273 int ret = 0; 6274 6275 migrate_prep(); 6276 6277 while (pfn < end || !list_empty(&cc->migratepages)) { 6278 if (fatal_signal_pending(current)) { 6279 ret = -EINTR; 6280 break; 6281 } 6282 6283 if (list_empty(&cc->migratepages)) { 6284 cc->nr_migratepages = 0; 6285 pfn = isolate_migratepages_range(cc, pfn, end); 6286 if (!pfn) { 6287 ret = -EINTR; 6288 break; 6289 } 6290 tries = 0; 6291 } else if (++tries == 5) { 6292 ret = ret < 0 ? ret : -EBUSY; 6293 break; 6294 } 6295 6296 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone, 6297 &cc->migratepages); 6298 cc->nr_migratepages -= nr_reclaimed; 6299 6300 ret = migrate_pages(&cc->migratepages, alloc_migrate_target, 6301 NULL, 0, cc->mode, MR_CMA); 6302 } 6303 if (ret < 0) { 6304 putback_movable_pages(&cc->migratepages); 6305 return ret; 6306 } 6307 return 0; 6308 } 6309 6310 /** 6311 * alloc_contig_range() -- tries to allocate given range of pages 6312 * @start: start PFN to allocate 6313 * @end: one-past-the-last PFN to allocate 6314 * @migratetype: migratetype of the underlaying pageblocks (either 6315 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks 6316 * in range must have the same migratetype and it must 6317 * be either of the two. 6318 * 6319 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES 6320 * aligned, however it's the caller's responsibility to guarantee that 6321 * we are the only thread that changes migrate type of pageblocks the 6322 * pages fall in. 6323 * 6324 * The PFN range must belong to a single zone. 6325 * 6326 * Returns zero on success or negative error code. On success all 6327 * pages which PFN is in [start, end) are allocated for the caller and 6328 * need to be freed with free_contig_range(). 6329 */ 6330 int alloc_contig_range(unsigned long start, unsigned long end, 6331 unsigned migratetype) 6332 { 6333 unsigned long outer_start, outer_end; 6334 int ret = 0, order; 6335 6336 struct compact_control cc = { 6337 .nr_migratepages = 0, 6338 .order = -1, 6339 .zone = page_zone(pfn_to_page(start)), 6340 .mode = MIGRATE_SYNC, 6341 .ignore_skip_hint = true, 6342 }; 6343 INIT_LIST_HEAD(&cc.migratepages); 6344 6345 /* 6346 * What we do here is we mark all pageblocks in range as 6347 * MIGRATE_ISOLATE. Because pageblock and max order pages may 6348 * have different sizes, and due to the way page allocator 6349 * work, we align the range to biggest of the two pages so 6350 * that page allocator won't try to merge buddies from 6351 * different pageblocks and change MIGRATE_ISOLATE to some 6352 * other migration type. 6353 * 6354 * Once the pageblocks are marked as MIGRATE_ISOLATE, we 6355 * migrate the pages from an unaligned range (ie. pages that 6356 * we are interested in). This will put all the pages in 6357 * range back to page allocator as MIGRATE_ISOLATE. 6358 * 6359 * When this is done, we take the pages in range from page 6360 * allocator removing them from the buddy system. This way 6361 * page allocator will never consider using them. 6362 * 6363 * This lets us mark the pageblocks back as 6364 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the 6365 * aligned range but not in the unaligned, original range are 6366 * put back to page allocator so that buddy can use them. 6367 */ 6368 6369 ret = start_isolate_page_range(pfn_max_align_down(start), 6370 pfn_max_align_up(end), migratetype, 6371 false); 6372 if (ret) 6373 return ret; 6374 6375 ret = __alloc_contig_migrate_range(&cc, start, end); 6376 if (ret) 6377 goto done; 6378 6379 /* 6380 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES 6381 * aligned blocks that are marked as MIGRATE_ISOLATE. What's 6382 * more, all pages in [start, end) are free in page allocator. 6383 * What we are going to do is to allocate all pages from 6384 * [start, end) (that is remove them from page allocator). 6385 * 6386 * The only problem is that pages at the beginning and at the 6387 * end of interesting range may be not aligned with pages that 6388 * page allocator holds, ie. they can be part of higher order 6389 * pages. Because of this, we reserve the bigger range and 6390 * once this is done free the pages we are not interested in. 6391 * 6392 * We don't have to hold zone->lock here because the pages are 6393 * isolated thus they won't get removed from buddy. 6394 */ 6395 6396 lru_add_drain_all(); 6397 drain_all_pages(); 6398 6399 order = 0; 6400 outer_start = start; 6401 while (!PageBuddy(pfn_to_page(outer_start))) { 6402 if (++order >= MAX_ORDER) { 6403 ret = -EBUSY; 6404 goto done; 6405 } 6406 outer_start &= ~0UL << order; 6407 } 6408 6409 /* Make sure the range is really isolated. */ 6410 if (test_pages_isolated(outer_start, end, false)) { 6411 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n", 6412 outer_start, end); 6413 ret = -EBUSY; 6414 goto done; 6415 } 6416 6417 6418 /* Grab isolated pages from freelists. */ 6419 outer_end = isolate_freepages_range(&cc, outer_start, end); 6420 if (!outer_end) { 6421 ret = -EBUSY; 6422 goto done; 6423 } 6424 6425 /* Free head and tail (if any) */ 6426 if (start != outer_start) 6427 free_contig_range(outer_start, start - outer_start); 6428 if (end != outer_end) 6429 free_contig_range(end, outer_end - end); 6430 6431 done: 6432 undo_isolate_page_range(pfn_max_align_down(start), 6433 pfn_max_align_up(end), migratetype); 6434 return ret; 6435 } 6436 6437 void free_contig_range(unsigned long pfn, unsigned nr_pages) 6438 { 6439 unsigned int count = 0; 6440 6441 for (; nr_pages--; pfn++) { 6442 struct page *page = pfn_to_page(pfn); 6443 6444 count += page_count(page) != 1; 6445 __free_page(page); 6446 } 6447 WARN(count != 0, "%d pages are still in use!\n", count); 6448 } 6449 #endif 6450 6451 #ifdef CONFIG_MEMORY_HOTPLUG 6452 /* 6453 * The zone indicated has a new number of managed_pages; batch sizes and percpu 6454 * page high values need to be recalulated. 6455 */ 6456 void __meminit zone_pcp_update(struct zone *zone) 6457 { 6458 unsigned cpu; 6459 mutex_lock(&pcp_batch_high_lock); 6460 for_each_possible_cpu(cpu) 6461 pageset_set_high_and_batch(zone, 6462 per_cpu_ptr(zone->pageset, cpu)); 6463 mutex_unlock(&pcp_batch_high_lock); 6464 } 6465 #endif 6466 6467 void zone_pcp_reset(struct zone *zone) 6468 { 6469 unsigned long flags; 6470 int cpu; 6471 struct per_cpu_pageset *pset; 6472 6473 /* avoid races with drain_pages() */ 6474 local_irq_save(flags); 6475 if (zone->pageset != &boot_pageset) { 6476 for_each_online_cpu(cpu) { 6477 pset = per_cpu_ptr(zone->pageset, cpu); 6478 drain_zonestat(zone, pset); 6479 } 6480 free_percpu(zone->pageset); 6481 zone->pageset = &boot_pageset; 6482 } 6483 local_irq_restore(flags); 6484 } 6485 6486 #ifdef CONFIG_MEMORY_HOTREMOVE 6487 /* 6488 * All pages in the range must be isolated before calling this. 6489 */ 6490 void 6491 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 6492 { 6493 struct page *page; 6494 struct zone *zone; 6495 unsigned int order, i; 6496 unsigned long pfn; 6497 unsigned long flags; 6498 /* find the first valid pfn */ 6499 for (pfn = start_pfn; pfn < end_pfn; pfn++) 6500 if (pfn_valid(pfn)) 6501 break; 6502 if (pfn == end_pfn) 6503 return; 6504 zone = page_zone(pfn_to_page(pfn)); 6505 spin_lock_irqsave(&zone->lock, flags); 6506 pfn = start_pfn; 6507 while (pfn < end_pfn) { 6508 if (!pfn_valid(pfn)) { 6509 pfn++; 6510 continue; 6511 } 6512 page = pfn_to_page(pfn); 6513 /* 6514 * The HWPoisoned page may be not in buddy system, and 6515 * page_count() is not 0. 6516 */ 6517 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) { 6518 pfn++; 6519 SetPageReserved(page); 6520 continue; 6521 } 6522 6523 BUG_ON(page_count(page)); 6524 BUG_ON(!PageBuddy(page)); 6525 order = page_order(page); 6526 #ifdef CONFIG_DEBUG_VM 6527 printk(KERN_INFO "remove from free list %lx %d %lx\n", 6528 pfn, 1 << order, end_pfn); 6529 #endif 6530 list_del(&page->lru); 6531 rmv_page_order(page); 6532 zone->free_area[order].nr_free--; 6533 for (i = 0; i < (1 << order); i++) 6534 SetPageReserved((page+i)); 6535 pfn += (1 << order); 6536 } 6537 spin_unlock_irqrestore(&zone->lock, flags); 6538 } 6539 #endif 6540 6541 #ifdef CONFIG_MEMORY_FAILURE 6542 bool is_free_buddy_page(struct page *page) 6543 { 6544 struct zone *zone = page_zone(page); 6545 unsigned long pfn = page_to_pfn(page); 6546 unsigned long flags; 6547 unsigned int order; 6548 6549 spin_lock_irqsave(&zone->lock, flags); 6550 for (order = 0; order < MAX_ORDER; order++) { 6551 struct page *page_head = page - (pfn & ((1 << order) - 1)); 6552 6553 if (PageBuddy(page_head) && page_order(page_head) >= order) 6554 break; 6555 } 6556 spin_unlock_irqrestore(&zone->lock, flags); 6557 6558 return order < MAX_ORDER; 6559 } 6560 #endif 6561