| 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/config.h> |
| 18 | #include <linux/stddef.h> |
| 19 | #include <linux/mm.h> |
| 20 | #include <linux/swap.h> |
| 21 | #include <linux/interrupt.h> |
| 22 | #include <linux/pagemap.h> |
| 23 | #include <linux/bootmem.h> |
| 24 | #include <linux/compiler.h> |
| 25 | #include <linux/kernel.h> |
| 26 | #include <linux/module.h> |
| 27 | #include <linux/suspend.h> |
| 28 | #include <linux/pagevec.h> |
| 29 | #include <linux/blkdev.h> |
| 30 | #include <linux/slab.h> |
| 31 | #include <linux/notifier.h> |
| 32 | #include <linux/topology.h> |
| 33 | #include <linux/sysctl.h> |
| 34 | #include <linux/cpu.h> |
| 35 | #include <linux/cpuset.h> |
| 36 | #include <linux/memory_hotplug.h> |
| 37 | #include <linux/nodemask.h> |
| 38 | #include <linux/vmalloc.h> |
| 39 | #include <linux/mempolicy.h> |
| 40 | |
| 41 | #include <asm/tlbflush.h> |
| 42 | #include "internal.h" |
| 43 | |
| 44 | /* |
| 45 | * MCD - HACK: Find somewhere to initialize this EARLY, or make this |
| 46 | * initializer cleaner |
| 47 | */ |
| 48 | nodemask_t node_online_map __read_mostly = { { [0] = 1UL } }; |
| 49 | EXPORT_SYMBOL(node_online_map); |
| 50 | nodemask_t node_possible_map __read_mostly = NODE_MASK_ALL; |
| 51 | EXPORT_SYMBOL(node_possible_map); |
| 52 | struct pglist_data *pgdat_list __read_mostly; |
| 53 | unsigned long totalram_pages __read_mostly; |
| 54 | unsigned long totalhigh_pages __read_mostly; |
| 55 | long nr_swap_pages; |
| 56 | int percpu_pagelist_fraction; |
| 57 | |
| 58 | static void fastcall free_hot_cold_page(struct page *page, int cold); |
| 59 | static void __free_pages_ok(struct page *page, unsigned int order); |
| 60 | |
| 61 | /* |
| 62 | * results with 256, 32 in the lowmem_reserve sysctl: |
| 63 | * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) |
| 64 | * 1G machine -> (16M dma, 784M normal, 224M high) |
| 65 | * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA |
| 66 | * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL |
| 67 | * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA |
| 68 | * |
| 69 | * TBD: should special case ZONE_DMA32 machines here - in those we normally |
| 70 | * don't need any ZONE_NORMAL reservation |
| 71 | */ |
| 72 | int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 256, 32 }; |
| 73 | |
| 74 | EXPORT_SYMBOL(totalram_pages); |
| 75 | |
| 76 | /* |
| 77 | * Used by page_zone() to look up the address of the struct zone whose |
| 78 | * id is encoded in the upper bits of page->flags |
| 79 | */ |
| 80 | struct zone *zone_table[1 << ZONETABLE_SHIFT] __read_mostly; |
| 81 | EXPORT_SYMBOL(zone_table); |
| 82 | |
| 83 | static char *zone_names[MAX_NR_ZONES] = { "DMA", "DMA32", "Normal", "HighMem" }; |
| 84 | int min_free_kbytes = 1024; |
| 85 | |
| 86 | unsigned long __initdata nr_kernel_pages; |
| 87 | unsigned long __initdata nr_all_pages; |
| 88 | |
| 89 | #ifdef CONFIG_DEBUG_VM |
| 90 | static int page_outside_zone_boundaries(struct zone *zone, struct page *page) |
| 91 | { |
| 92 | int ret = 0; |
| 93 | unsigned seq; |
| 94 | unsigned long pfn = page_to_pfn(page); |
| 95 | |
| 96 | do { |
| 97 | seq = zone_span_seqbegin(zone); |
| 98 | if (pfn >= zone->zone_start_pfn + zone->spanned_pages) |
| 99 | ret = 1; |
| 100 | else if (pfn < zone->zone_start_pfn) |
| 101 | ret = 1; |
| 102 | } while (zone_span_seqretry(zone, seq)); |
| 103 | |
| 104 | return ret; |
| 105 | } |
| 106 | |
| 107 | static int page_is_consistent(struct zone *zone, struct page *page) |
| 108 | { |
| 109 | #ifdef CONFIG_HOLES_IN_ZONE |
| 110 | if (!pfn_valid(page_to_pfn(page))) |
| 111 | return 0; |
| 112 | #endif |
| 113 | if (zone != page_zone(page)) |
| 114 | return 0; |
| 115 | |
| 116 | return 1; |
| 117 | } |
| 118 | /* |
| 119 | * Temporary debugging check for pages not lying within a given zone. |
| 120 | */ |
| 121 | static int bad_range(struct zone *zone, struct page *page) |
| 122 | { |
| 123 | if (page_outside_zone_boundaries(zone, page)) |
| 124 | return 1; |
| 125 | if (!page_is_consistent(zone, page)) |
| 126 | return 1; |
| 127 | |
| 128 | return 0; |
| 129 | } |
| 130 | |
| 131 | #else |
| 132 | static inline int bad_range(struct zone *zone, struct page *page) |
| 133 | { |
| 134 | return 0; |
| 135 | } |
| 136 | #endif |
| 137 | |
| 138 | static void bad_page(struct page *page) |
| 139 | { |
| 140 | printk(KERN_EMERG "Bad page state in process '%s'\n" |
| 141 | KERN_EMERG "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n" |
| 142 | KERN_EMERG "Trying to fix it up, but a reboot is needed\n" |
| 143 | KERN_EMERG "Backtrace:\n", |
| 144 | current->comm, page, (int)(2*sizeof(unsigned long)), |
| 145 | (unsigned long)page->flags, page->mapping, |
| 146 | page_mapcount(page), page_count(page)); |
| 147 | dump_stack(); |
| 148 | page->flags &= ~(1 << PG_lru | |
| 149 | 1 << PG_private | |
| 150 | 1 << PG_locked | |
| 151 | 1 << PG_active | |
| 152 | 1 << PG_dirty | |
| 153 | 1 << PG_reclaim | |
| 154 | 1 << PG_slab | |
| 155 | 1 << PG_swapcache | |
| 156 | 1 << PG_writeback ); |
| 157 | set_page_count(page, 0); |
| 158 | reset_page_mapcount(page); |
| 159 | page->mapping = NULL; |
| 160 | add_taint(TAINT_BAD_PAGE); |
| 161 | } |
| 162 | |
| 163 | /* |
| 164 | * Higher-order pages are called "compound pages". They are structured thusly: |
| 165 | * |
| 166 | * The first PAGE_SIZE page is called the "head page". |
| 167 | * |
| 168 | * The remaining PAGE_SIZE pages are called "tail pages". |
| 169 | * |
| 170 | * All pages have PG_compound set. All pages have their ->private pointing at |
| 171 | * the head page (even the head page has this). |
| 172 | * |
| 173 | * The first tail page's ->lru.next holds the address of the compound page's |
| 174 | * put_page() function. Its ->lru.prev holds the order of allocation. |
| 175 | * This usage means that zero-order pages may not be compound. |
| 176 | */ |
| 177 | |
| 178 | static void free_compound_page(struct page *page) |
| 179 | { |
| 180 | __free_pages_ok(page, (unsigned long)page[1].lru.prev); |
| 181 | } |
| 182 | |
| 183 | static void prep_compound_page(struct page *page, unsigned long order) |
| 184 | { |
| 185 | int i; |
| 186 | int nr_pages = 1 << order; |
| 187 | |
| 188 | page[1].lru.next = (void *)free_compound_page; /* set dtor */ |
| 189 | page[1].lru.prev = (void *)order; |
| 190 | for (i = 0; i < nr_pages; i++) { |
| 191 | struct page *p = page + i; |
| 192 | |
| 193 | SetPageCompound(p); |
| 194 | set_page_private(p, (unsigned long)page); |
| 195 | } |
| 196 | } |
| 197 | |
| 198 | static void destroy_compound_page(struct page *page, unsigned long order) |
| 199 | { |
| 200 | int i; |
| 201 | int nr_pages = 1 << order; |
| 202 | |
| 203 | if (unlikely((unsigned long)page[1].lru.prev != order)) |
| 204 | bad_page(page); |
| 205 | |
| 206 | for (i = 0; i < nr_pages; i++) { |
| 207 | struct page *p = page + i; |
| 208 | |
| 209 | if (unlikely(!PageCompound(p) | |
| 210 | (page_private(p) != (unsigned long)page))) |
| 211 | bad_page(page); |
| 212 | ClearPageCompound(p); |
| 213 | } |
| 214 | } |
| 215 | |
| 216 | /* |
| 217 | * function for dealing with page's order in buddy system. |
| 218 | * zone->lock is already acquired when we use these. |
| 219 | * So, we don't need atomic page->flags operations here. |
| 220 | */ |
| 221 | static inline unsigned long page_order(struct page *page) { |
| 222 | return page_private(page); |
| 223 | } |
| 224 | |
| 225 | static inline void set_page_order(struct page *page, int order) { |
| 226 | set_page_private(page, order); |
| 227 | __SetPagePrivate(page); |
| 228 | } |
| 229 | |
| 230 | static inline void rmv_page_order(struct page *page) |
| 231 | { |
| 232 | __ClearPagePrivate(page); |
| 233 | set_page_private(page, 0); |
| 234 | } |
| 235 | |
| 236 | /* |
| 237 | * Locate the struct page for both the matching buddy in our |
| 238 | * pair (buddy1) and the combined O(n+1) page they form (page). |
| 239 | * |
| 240 | * 1) Any buddy B1 will have an order O twin B2 which satisfies |
| 241 | * the following equation: |
| 242 | * B2 = B1 ^ (1 << O) |
| 243 | * For example, if the starting buddy (buddy2) is #8 its order |
| 244 | * 1 buddy is #10: |
| 245 | * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 |
| 246 | * |
| 247 | * 2) Any buddy B will have an order O+1 parent P which |
| 248 | * satisfies the following equation: |
| 249 | * P = B & ~(1 << O) |
| 250 | * |
| 251 | * Assumption: *_mem_map is contigious at least up to MAX_ORDER |
| 252 | */ |
| 253 | static inline struct page * |
| 254 | __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) |
| 255 | { |
| 256 | unsigned long buddy_idx = page_idx ^ (1 << order); |
| 257 | |
| 258 | return page + (buddy_idx - page_idx); |
| 259 | } |
| 260 | |
| 261 | static inline unsigned long |
| 262 | __find_combined_index(unsigned long page_idx, unsigned int order) |
| 263 | { |
| 264 | return (page_idx & ~(1 << order)); |
| 265 | } |
| 266 | |
| 267 | /* |
| 268 | * This function checks whether a page is free && is the buddy |
| 269 | * we can do coalesce a page and its buddy if |
| 270 | * (a) the buddy is not in a hole && |
| 271 | * (b) the buddy is free && |
| 272 | * (c) the buddy is on the buddy system && |
| 273 | * (d) a page and its buddy have the same order. |
| 274 | * for recording page's order, we use page_private(page) and PG_private. |
| 275 | * |
| 276 | */ |
| 277 | static inline int page_is_buddy(struct page *page, int order) |
| 278 | { |
| 279 | #ifdef CONFIG_HOLES_IN_ZONE |
| 280 | if (!pfn_valid(page_to_pfn(page))) |
| 281 | return 0; |
| 282 | #endif |
| 283 | |
| 284 | if (PagePrivate(page) && |
| 285 | (page_order(page) == order) && |
| 286 | page_count(page) == 0) |
| 287 | return 1; |
| 288 | return 0; |
| 289 | } |
| 290 | |
| 291 | /* |
| 292 | * Freeing function for a buddy system allocator. |
| 293 | * |
| 294 | * The concept of a buddy system is to maintain direct-mapped table |
| 295 | * (containing bit values) for memory blocks of various "orders". |
| 296 | * The bottom level table contains the map for the smallest allocatable |
| 297 | * units of memory (here, pages), and each level above it describes |
| 298 | * pairs of units from the levels below, hence, "buddies". |
| 299 | * At a high level, all that happens here is marking the table entry |
| 300 | * at the bottom level available, and propagating the changes upward |
| 301 | * as necessary, plus some accounting needed to play nicely with other |
| 302 | * parts of the VM system. |
| 303 | * At each level, we keep a list of pages, which are heads of continuous |
| 304 | * free pages of length of (1 << order) and marked with PG_Private.Page's |
| 305 | * order is recorded in page_private(page) field. |
| 306 | * So when we are allocating or freeing one, we can derive the state of the |
| 307 | * other. That is, if we allocate a small block, and both were |
| 308 | * free, the remainder of the region must be split into blocks. |
| 309 | * If a block is freed, and its buddy is also free, then this |
| 310 | * triggers coalescing into a block of larger size. |
| 311 | * |
| 312 | * -- wli |
| 313 | */ |
| 314 | |
| 315 | static inline void __free_one_page(struct page *page, |
| 316 | struct zone *zone, unsigned int order) |
| 317 | { |
| 318 | unsigned long page_idx; |
| 319 | int order_size = 1 << order; |
| 320 | |
| 321 | if (unlikely(PageCompound(page))) |
| 322 | destroy_compound_page(page, order); |
| 323 | |
| 324 | page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); |
| 325 | |
| 326 | BUG_ON(page_idx & (order_size - 1)); |
| 327 | BUG_ON(bad_range(zone, page)); |
| 328 | |
| 329 | zone->free_pages += order_size; |
| 330 | while (order < MAX_ORDER-1) { |
| 331 | unsigned long combined_idx; |
| 332 | struct free_area *area; |
| 333 | struct page *buddy; |
| 334 | |
| 335 | buddy = __page_find_buddy(page, page_idx, order); |
| 336 | if (!page_is_buddy(buddy, order)) |
| 337 | break; /* Move the buddy up one level. */ |
| 338 | |
| 339 | list_del(&buddy->lru); |
| 340 | area = zone->free_area + order; |
| 341 | area->nr_free--; |
| 342 | rmv_page_order(buddy); |
| 343 | combined_idx = __find_combined_index(page_idx, order); |
| 344 | page = page + (combined_idx - page_idx); |
| 345 | page_idx = combined_idx; |
| 346 | order++; |
| 347 | } |
| 348 | set_page_order(page, order); |
| 349 | list_add(&page->lru, &zone->free_area[order].free_list); |
| 350 | zone->free_area[order].nr_free++; |
| 351 | } |
| 352 | |
| 353 | static inline int free_pages_check(struct page *page) |
| 354 | { |
| 355 | if (unlikely(page_mapcount(page) | |
| 356 | (page->mapping != NULL) | |
| 357 | (page_count(page) != 0) | |
| 358 | (page->flags & ( |
| 359 | 1 << PG_lru | |
| 360 | 1 << PG_private | |
| 361 | 1 << PG_locked | |
| 362 | 1 << PG_active | |
| 363 | 1 << PG_reclaim | |
| 364 | 1 << PG_slab | |
| 365 | 1 << PG_swapcache | |
| 366 | 1 << PG_writeback | |
| 367 | 1 << PG_reserved )))) |
| 368 | bad_page(page); |
| 369 | if (PageDirty(page)) |
| 370 | __ClearPageDirty(page); |
| 371 | /* |
| 372 | * For now, we report if PG_reserved was found set, but do not |
| 373 | * clear it, and do not free the page. But we shall soon need |
| 374 | * to do more, for when the ZERO_PAGE count wraps negative. |
| 375 | */ |
| 376 | return PageReserved(page); |
| 377 | } |
| 378 | |
| 379 | /* |
| 380 | * Frees a list of pages. |
| 381 | * Assumes all pages on list are in same zone, and of same order. |
| 382 | * count is the number of pages to free. |
| 383 | * |
| 384 | * If the zone was previously in an "all pages pinned" state then look to |
| 385 | * see if this freeing clears that state. |
| 386 | * |
| 387 | * And clear the zone's pages_scanned counter, to hold off the "all pages are |
| 388 | * pinned" detection logic. |
| 389 | */ |
| 390 | static void free_pages_bulk(struct zone *zone, int count, |
| 391 | struct list_head *list, int order) |
| 392 | { |
| 393 | spin_lock(&zone->lock); |
| 394 | zone->all_unreclaimable = 0; |
| 395 | zone->pages_scanned = 0; |
| 396 | while (count--) { |
| 397 | struct page *page; |
| 398 | |
| 399 | BUG_ON(list_empty(list)); |
| 400 | page = list_entry(list->prev, struct page, lru); |
| 401 | /* have to delete it as __free_one_page list manipulates */ |
| 402 | list_del(&page->lru); |
| 403 | __free_one_page(page, zone, order); |
| 404 | } |
| 405 | spin_unlock(&zone->lock); |
| 406 | } |
| 407 | |
| 408 | static void free_one_page(struct zone *zone, struct page *page, int order) |
| 409 | { |
| 410 | LIST_HEAD(list); |
| 411 | list_add(&page->lru, &list); |
| 412 | free_pages_bulk(zone, 1, &list, order); |
| 413 | } |
| 414 | |
| 415 | static void __free_pages_ok(struct page *page, unsigned int order) |
| 416 | { |
| 417 | unsigned long flags; |
| 418 | int i; |
| 419 | int reserved = 0; |
| 420 | |
| 421 | arch_free_page(page, order); |
| 422 | if (!PageHighMem(page)) |
| 423 | mutex_debug_check_no_locks_freed(page_address(page), |
| 424 | PAGE_SIZE<<order); |
| 425 | |
| 426 | #ifndef CONFIG_MMU |
| 427 | for (i = 1 ; i < (1 << order) ; ++i) |
| 428 | __put_page(page + i); |
| 429 | #endif |
| 430 | |
| 431 | for (i = 0 ; i < (1 << order) ; ++i) |
| 432 | reserved += free_pages_check(page + i); |
| 433 | if (reserved) |
| 434 | return; |
| 435 | |
| 436 | kernel_map_pages(page, 1 << order, 0); |
| 437 | local_irq_save(flags); |
| 438 | __mod_page_state(pgfree, 1 << order); |
| 439 | free_one_page(page_zone(page), page, order); |
| 440 | local_irq_restore(flags); |
| 441 | } |
| 442 | |
| 443 | /* |
| 444 | * permit the bootmem allocator to evade page validation on high-order frees |
| 445 | */ |
| 446 | void fastcall __init __free_pages_bootmem(struct page *page, unsigned int order) |
| 447 | { |
| 448 | if (order == 0) { |
| 449 | __ClearPageReserved(page); |
| 450 | set_page_count(page, 0); |
| 451 | |
| 452 | free_hot_cold_page(page, 0); |
| 453 | } else { |
| 454 | LIST_HEAD(list); |
| 455 | int loop; |
| 456 | |
| 457 | for (loop = 0; loop < BITS_PER_LONG; loop++) { |
| 458 | struct page *p = &page[loop]; |
| 459 | |
| 460 | if (loop + 16 < BITS_PER_LONG) |
| 461 | prefetchw(p + 16); |
| 462 | __ClearPageReserved(p); |
| 463 | set_page_count(p, 0); |
| 464 | } |
| 465 | |
| 466 | arch_free_page(page, order); |
| 467 | |
| 468 | mod_page_state(pgfree, 1 << order); |
| 469 | |
| 470 | list_add(&page->lru, &list); |
| 471 | kernel_map_pages(page, 1 << order, 0); |
| 472 | free_pages_bulk(page_zone(page), 1, &list, order); |
| 473 | } |
| 474 | } |
| 475 | |
| 476 | |
| 477 | /* |
| 478 | * The order of subdivision here is critical for the IO subsystem. |
| 479 | * Please do not alter this order without good reasons and regression |
| 480 | * testing. Specifically, as large blocks of memory are subdivided, |
| 481 | * the order in which smaller blocks are delivered depends on the order |
| 482 | * they're subdivided in this function. This is the primary factor |
| 483 | * influencing the order in which pages are delivered to the IO |
| 484 | * subsystem according to empirical testing, and this is also justified |
| 485 | * by considering the behavior of a buddy system containing a single |
| 486 | * large block of memory acted on by a series of small allocations. |
| 487 | * This behavior is a critical factor in sglist merging's success. |
| 488 | * |
| 489 | * -- wli |
| 490 | */ |
| 491 | static inline void expand(struct zone *zone, struct page *page, |
| 492 | int low, int high, struct free_area *area) |
| 493 | { |
| 494 | unsigned long size = 1 << high; |
| 495 | |
| 496 | while (high > low) { |
| 497 | area--; |
| 498 | high--; |
| 499 | size >>= 1; |
| 500 | BUG_ON(bad_range(zone, &page[size])); |
| 501 | list_add(&page[size].lru, &area->free_list); |
| 502 | area->nr_free++; |
| 503 | set_page_order(&page[size], high); |
| 504 | } |
| 505 | } |
| 506 | |
| 507 | /* |
| 508 | * This page is about to be returned from the page allocator |
| 509 | */ |
| 510 | static int prep_new_page(struct page *page, int order) |
| 511 | { |
| 512 | if (unlikely(page_mapcount(page) | |
| 513 | (page->mapping != NULL) | |
| 514 | (page_count(page) != 0) | |
| 515 | (page->flags & ( |
| 516 | 1 << PG_lru | |
| 517 | 1 << PG_private | |
| 518 | 1 << PG_locked | |
| 519 | 1 << PG_active | |
| 520 | 1 << PG_dirty | |
| 521 | 1 << PG_reclaim | |
| 522 | 1 << PG_slab | |
| 523 | 1 << PG_swapcache | |
| 524 | 1 << PG_writeback | |
| 525 | 1 << PG_reserved )))) |
| 526 | bad_page(page); |
| 527 | |
| 528 | /* |
| 529 | * For now, we report if PG_reserved was found set, but do not |
| 530 | * clear it, and do not allocate the page: as a safety net. |
| 531 | */ |
| 532 | if (PageReserved(page)) |
| 533 | return 1; |
| 534 | |
| 535 | page->flags &= ~(1 << PG_uptodate | 1 << PG_error | |
| 536 | 1 << PG_referenced | 1 << PG_arch_1 | |
| 537 | 1 << PG_checked | 1 << PG_mappedtodisk); |
| 538 | set_page_private(page, 0); |
| 539 | set_page_refs(page, order); |
| 540 | kernel_map_pages(page, 1 << order, 1); |
| 541 | return 0; |
| 542 | } |
| 543 | |
| 544 | /* |
| 545 | * Do the hard work of removing an element from the buddy allocator. |
| 546 | * Call me with the zone->lock already held. |
| 547 | */ |
| 548 | static struct page *__rmqueue(struct zone *zone, unsigned int order) |
| 549 | { |
| 550 | struct free_area * area; |
| 551 | unsigned int current_order; |
| 552 | struct page *page; |
| 553 | |
| 554 | for (current_order = order; current_order < MAX_ORDER; ++current_order) { |
| 555 | area = zone->free_area + current_order; |
| 556 | if (list_empty(&area->free_list)) |
| 557 | continue; |
| 558 | |
| 559 | page = list_entry(area->free_list.next, struct page, lru); |
| 560 | list_del(&page->lru); |
| 561 | rmv_page_order(page); |
| 562 | area->nr_free--; |
| 563 | zone->free_pages -= 1UL << order; |
| 564 | expand(zone, page, order, current_order, area); |
| 565 | return page; |
| 566 | } |
| 567 | |
| 568 | return NULL; |
| 569 | } |
| 570 | |
| 571 | /* |
| 572 | * Obtain a specified number of elements from the buddy allocator, all under |
| 573 | * a single hold of the lock, for efficiency. Add them to the supplied list. |
| 574 | * Returns the number of new pages which were placed at *list. |
| 575 | */ |
| 576 | static int rmqueue_bulk(struct zone *zone, unsigned int order, |
| 577 | unsigned long count, struct list_head *list) |
| 578 | { |
| 579 | int i; |
| 580 | |
| 581 | spin_lock(&zone->lock); |
| 582 | for (i = 0; i < count; ++i) { |
| 583 | struct page *page = __rmqueue(zone, order); |
| 584 | if (unlikely(page == NULL)) |
| 585 | break; |
| 586 | list_add_tail(&page->lru, list); |
| 587 | } |
| 588 | spin_unlock(&zone->lock); |
| 589 | return i; |
| 590 | } |
| 591 | |
| 592 | #ifdef CONFIG_NUMA |
| 593 | /* Called from the slab reaper to drain remote pagesets */ |
| 594 | void drain_remote_pages(void) |
| 595 | { |
| 596 | struct zone *zone; |
| 597 | int i; |
| 598 | unsigned long flags; |
| 599 | |
| 600 | local_irq_save(flags); |
| 601 | for_each_zone(zone) { |
| 602 | struct per_cpu_pageset *pset; |
| 603 | |
| 604 | /* Do not drain local pagesets */ |
| 605 | if (zone->zone_pgdat->node_id == numa_node_id()) |
| 606 | continue; |
| 607 | |
| 608 | pset = zone_pcp(zone, smp_processor_id()); |
| 609 | for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) { |
| 610 | struct per_cpu_pages *pcp; |
| 611 | |
| 612 | pcp = &pset->pcp[i]; |
| 613 | free_pages_bulk(zone, pcp->count, &pcp->list, 0); |
| 614 | pcp->count = 0; |
| 615 | } |
| 616 | } |
| 617 | local_irq_restore(flags); |
| 618 | } |
| 619 | #endif |
| 620 | |
| 621 | #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU) |
| 622 | static void __drain_pages(unsigned int cpu) |
| 623 | { |
| 624 | unsigned long flags; |
| 625 | struct zone *zone; |
| 626 | int i; |
| 627 | |
| 628 | for_each_zone(zone) { |
| 629 | struct per_cpu_pageset *pset; |
| 630 | |
| 631 | pset = zone_pcp(zone, cpu); |
| 632 | for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) { |
| 633 | struct per_cpu_pages *pcp; |
| 634 | |
| 635 | pcp = &pset->pcp[i]; |
| 636 | local_irq_save(flags); |
| 637 | free_pages_bulk(zone, pcp->count, &pcp->list, 0); |
| 638 | pcp->count = 0; |
| 639 | local_irq_restore(flags); |
| 640 | } |
| 641 | } |
| 642 | } |
| 643 | #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */ |
| 644 | |
| 645 | #ifdef CONFIG_PM |
| 646 | |
| 647 | void mark_free_pages(struct zone *zone) |
| 648 | { |
| 649 | unsigned long zone_pfn, flags; |
| 650 | int order; |
| 651 | struct list_head *curr; |
| 652 | |
| 653 | if (!zone->spanned_pages) |
| 654 | return; |
| 655 | |
| 656 | spin_lock_irqsave(&zone->lock, flags); |
| 657 | for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn) |
| 658 | ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn)); |
| 659 | |
| 660 | for (order = MAX_ORDER - 1; order >= 0; --order) |
| 661 | list_for_each(curr, &zone->free_area[order].free_list) { |
| 662 | unsigned long start_pfn, i; |
| 663 | |
| 664 | start_pfn = page_to_pfn(list_entry(curr, struct page, lru)); |
| 665 | |
| 666 | for (i=0; i < (1<<order); i++) |
| 667 | SetPageNosaveFree(pfn_to_page(start_pfn+i)); |
| 668 | } |
| 669 | spin_unlock_irqrestore(&zone->lock, flags); |
| 670 | } |
| 671 | |
| 672 | /* |
| 673 | * Spill all of this CPU's per-cpu pages back into the buddy allocator. |
| 674 | */ |
| 675 | void drain_local_pages(void) |
| 676 | { |
| 677 | unsigned long flags; |
| 678 | |
| 679 | local_irq_save(flags); |
| 680 | __drain_pages(smp_processor_id()); |
| 681 | local_irq_restore(flags); |
| 682 | } |
| 683 | #endif /* CONFIG_PM */ |
| 684 | |
| 685 | static void zone_statistics(struct zonelist *zonelist, struct zone *z, int cpu) |
| 686 | { |
| 687 | #ifdef CONFIG_NUMA |
| 688 | pg_data_t *pg = z->zone_pgdat; |
| 689 | pg_data_t *orig = zonelist->zones[0]->zone_pgdat; |
| 690 | struct per_cpu_pageset *p; |
| 691 | |
| 692 | p = zone_pcp(z, cpu); |
| 693 | if (pg == orig) { |
| 694 | p->numa_hit++; |
| 695 | } else { |
| 696 | p->numa_miss++; |
| 697 | zone_pcp(zonelist->zones[0], cpu)->numa_foreign++; |
| 698 | } |
| 699 | if (pg == NODE_DATA(numa_node_id())) |
| 700 | p->local_node++; |
| 701 | else |
| 702 | p->other_node++; |
| 703 | #endif |
| 704 | } |
| 705 | |
| 706 | /* |
| 707 | * Free a 0-order page |
| 708 | */ |
| 709 | static void fastcall free_hot_cold_page(struct page *page, int cold) |
| 710 | { |
| 711 | struct zone *zone = page_zone(page); |
| 712 | struct per_cpu_pages *pcp; |
| 713 | unsigned long flags; |
| 714 | |
| 715 | arch_free_page(page, 0); |
| 716 | |
| 717 | if (PageAnon(page)) |
| 718 | page->mapping = NULL; |
| 719 | if (free_pages_check(page)) |
| 720 | return; |
| 721 | |
| 722 | kernel_map_pages(page, 1, 0); |
| 723 | |
| 724 | pcp = &zone_pcp(zone, get_cpu())->pcp[cold]; |
| 725 | local_irq_save(flags); |
| 726 | __inc_page_state(pgfree); |
| 727 | list_add(&page->lru, &pcp->list); |
| 728 | pcp->count++; |
| 729 | if (pcp->count >= pcp->high) { |
| 730 | free_pages_bulk(zone, pcp->batch, &pcp->list, 0); |
| 731 | pcp->count -= pcp->batch; |
| 732 | } |
| 733 | local_irq_restore(flags); |
| 734 | put_cpu(); |
| 735 | } |
| 736 | |
| 737 | void fastcall free_hot_page(struct page *page) |
| 738 | { |
| 739 | free_hot_cold_page(page, 0); |
| 740 | } |
| 741 | |
| 742 | void fastcall free_cold_page(struct page *page) |
| 743 | { |
| 744 | free_hot_cold_page(page, 1); |
| 745 | } |
| 746 | |
| 747 | static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) |
| 748 | { |
| 749 | int i; |
| 750 | |
| 751 | BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM); |
| 752 | for(i = 0; i < (1 << order); i++) |
| 753 | clear_highpage(page + i); |
| 754 | } |
| 755 | |
| 756 | /* |
| 757 | * Really, prep_compound_page() should be called from __rmqueue_bulk(). But |
| 758 | * we cheat by calling it from here, in the order > 0 path. Saves a branch |
| 759 | * or two. |
| 760 | */ |
| 761 | static struct page *buffered_rmqueue(struct zonelist *zonelist, |
| 762 | struct zone *zone, int order, gfp_t gfp_flags) |
| 763 | { |
| 764 | unsigned long flags; |
| 765 | struct page *page; |
| 766 | int cold = !!(gfp_flags & __GFP_COLD); |
| 767 | int cpu; |
| 768 | |
| 769 | again: |
| 770 | cpu = get_cpu(); |
| 771 | if (likely(order == 0)) { |
| 772 | struct per_cpu_pages *pcp; |
| 773 | |
| 774 | pcp = &zone_pcp(zone, cpu)->pcp[cold]; |
| 775 | local_irq_save(flags); |
| 776 | if (!pcp->count) { |
| 777 | pcp->count += rmqueue_bulk(zone, 0, |
| 778 | pcp->batch, &pcp->list); |
| 779 | if (unlikely(!pcp->count)) |
| 780 | goto failed; |
| 781 | } |
| 782 | page = list_entry(pcp->list.next, struct page, lru); |
| 783 | list_del(&page->lru); |
| 784 | pcp->count--; |
| 785 | } else { |
| 786 | spin_lock_irqsave(&zone->lock, flags); |
| 787 | page = __rmqueue(zone, order); |
| 788 | spin_unlock(&zone->lock); |
| 789 | if (!page) |
| 790 | goto failed; |
| 791 | } |
| 792 | |
| 793 | __mod_page_state_zone(zone, pgalloc, 1 << order); |
| 794 | zone_statistics(zonelist, zone, cpu); |
| 795 | local_irq_restore(flags); |
| 796 | put_cpu(); |
| 797 | |
| 798 | BUG_ON(bad_range(zone, page)); |
| 799 | if (prep_new_page(page, order)) |
| 800 | goto again; |
| 801 | |
| 802 | if (gfp_flags & __GFP_ZERO) |
| 803 | prep_zero_page(page, order, gfp_flags); |
| 804 | |
| 805 | if (order && (gfp_flags & __GFP_COMP)) |
| 806 | prep_compound_page(page, order); |
| 807 | return page; |
| 808 | |
| 809 | failed: |
| 810 | local_irq_restore(flags); |
| 811 | put_cpu(); |
| 812 | return NULL; |
| 813 | } |
| 814 | |
| 815 | #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */ |
| 816 | #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */ |
| 817 | #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */ |
| 818 | #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */ |
| 819 | #define ALLOC_HARDER 0x10 /* try to alloc harder */ |
| 820 | #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ |
| 821 | #define ALLOC_CPUSET 0x40 /* check for correct cpuset */ |
| 822 | |
| 823 | /* |
| 824 | * Return 1 if free pages are above 'mark'. This takes into account the order |
| 825 | * of the allocation. |
| 826 | */ |
| 827 | int zone_watermark_ok(struct zone *z, int order, unsigned long mark, |
| 828 | int classzone_idx, int alloc_flags) |
| 829 | { |
| 830 | /* free_pages my go negative - that's OK */ |
| 831 | long min = mark, free_pages = z->free_pages - (1 << order) + 1; |
| 832 | int o; |
| 833 | |
| 834 | if (alloc_flags & ALLOC_HIGH) |
| 835 | min -= min / 2; |
| 836 | if (alloc_flags & ALLOC_HARDER) |
| 837 | min -= min / 4; |
| 838 | |
| 839 | if (free_pages <= min + z->lowmem_reserve[classzone_idx]) |
| 840 | return 0; |
| 841 | for (o = 0; o < order; o++) { |
| 842 | /* At the next order, this order's pages become unavailable */ |
| 843 | free_pages -= z->free_area[o].nr_free << o; |
| 844 | |
| 845 | /* Require fewer higher order pages to be free */ |
| 846 | min >>= 1; |
| 847 | |
| 848 | if (free_pages <= min) |
| 849 | return 0; |
| 850 | } |
| 851 | return 1; |
| 852 | } |
| 853 | |
| 854 | /* |
| 855 | * get_page_from_freeliest goes through the zonelist trying to allocate |
| 856 | * a page. |
| 857 | */ |
| 858 | static struct page * |
| 859 | get_page_from_freelist(gfp_t gfp_mask, unsigned int order, |
| 860 | struct zonelist *zonelist, int alloc_flags) |
| 861 | { |
| 862 | struct zone **z = zonelist->zones; |
| 863 | struct page *page = NULL; |
| 864 | int classzone_idx = zone_idx(*z); |
| 865 | |
| 866 | /* |
| 867 | * Go through the zonelist once, looking for a zone with enough free. |
| 868 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. |
| 869 | */ |
| 870 | do { |
| 871 | if ((alloc_flags & ALLOC_CPUSET) && |
| 872 | !cpuset_zone_allowed(*z, gfp_mask)) |
| 873 | continue; |
| 874 | |
| 875 | if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { |
| 876 | unsigned long mark; |
| 877 | if (alloc_flags & ALLOC_WMARK_MIN) |
| 878 | mark = (*z)->pages_min; |
| 879 | else if (alloc_flags & ALLOC_WMARK_LOW) |
| 880 | mark = (*z)->pages_low; |
| 881 | else |
| 882 | mark = (*z)->pages_high; |
| 883 | if (!zone_watermark_ok(*z, order, mark, |
| 884 | classzone_idx, alloc_flags)) |
| 885 | if (!zone_reclaim_mode || |
| 886 | !zone_reclaim(*z, gfp_mask, order)) |
| 887 | continue; |
| 888 | } |
| 889 | |
| 890 | page = buffered_rmqueue(zonelist, *z, order, gfp_mask); |
| 891 | if (page) { |
| 892 | break; |
| 893 | } |
| 894 | } while (*(++z) != NULL); |
| 895 | return page; |
| 896 | } |
| 897 | |
| 898 | /* |
| 899 | * This is the 'heart' of the zoned buddy allocator. |
| 900 | */ |
| 901 | struct page * fastcall |
| 902 | __alloc_pages(gfp_t gfp_mask, unsigned int order, |
| 903 | struct zonelist *zonelist) |
| 904 | { |
| 905 | const gfp_t wait = gfp_mask & __GFP_WAIT; |
| 906 | struct zone **z; |
| 907 | struct page *page; |
| 908 | struct reclaim_state reclaim_state; |
| 909 | struct task_struct *p = current; |
| 910 | int do_retry; |
| 911 | int alloc_flags; |
| 912 | int did_some_progress; |
| 913 | |
| 914 | might_sleep_if(wait); |
| 915 | |
| 916 | restart: |
| 917 | z = zonelist->zones; /* the list of zones suitable for gfp_mask */ |
| 918 | |
| 919 | if (unlikely(*z == NULL)) { |
| 920 | /* Should this ever happen?? */ |
| 921 | return NULL; |
| 922 | } |
| 923 | |
| 924 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order, |
| 925 | zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET); |
| 926 | if (page) |
| 927 | goto got_pg; |
| 928 | |
| 929 | do { |
| 930 | wakeup_kswapd(*z, order); |
| 931 | } while (*(++z)); |
| 932 | |
| 933 | /* |
| 934 | * OK, we're below the kswapd watermark and have kicked background |
| 935 | * reclaim. Now things get more complex, so set up alloc_flags according |
| 936 | * to how we want to proceed. |
| 937 | * |
| 938 | * The caller may dip into page reserves a bit more if the caller |
| 939 | * cannot run direct reclaim, or if the caller has realtime scheduling |
| 940 | * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will |
| 941 | * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). |
| 942 | */ |
| 943 | alloc_flags = ALLOC_WMARK_MIN; |
| 944 | if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait) |
| 945 | alloc_flags |= ALLOC_HARDER; |
| 946 | if (gfp_mask & __GFP_HIGH) |
| 947 | alloc_flags |= ALLOC_HIGH; |
| 948 | alloc_flags |= ALLOC_CPUSET; |
| 949 | |
| 950 | /* |
| 951 | * Go through the zonelist again. Let __GFP_HIGH and allocations |
| 952 | * coming from realtime tasks go deeper into reserves. |
| 953 | * |
| 954 | * This is the last chance, in general, before the goto nopage. |
| 955 | * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. |
| 956 | * See also cpuset_zone_allowed() comment in kernel/cpuset.c. |
| 957 | */ |
| 958 | page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags); |
| 959 | if (page) |
| 960 | goto got_pg; |
| 961 | |
| 962 | /* This allocation should allow future memory freeing. */ |
| 963 | |
| 964 | if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE))) |
| 965 | && !in_interrupt()) { |
| 966 | if (!(gfp_mask & __GFP_NOMEMALLOC)) { |
| 967 | nofail_alloc: |
| 968 | /* go through the zonelist yet again, ignoring mins */ |
| 969 | page = get_page_from_freelist(gfp_mask, order, |
| 970 | zonelist, ALLOC_NO_WATERMARKS); |
| 971 | if (page) |
| 972 | goto got_pg; |
| 973 | if (gfp_mask & __GFP_NOFAIL) { |
| 974 | blk_congestion_wait(WRITE, HZ/50); |
| 975 | goto nofail_alloc; |
| 976 | } |
| 977 | } |
| 978 | goto nopage; |
| 979 | } |
| 980 | |
| 981 | /* Atomic allocations - we can't balance anything */ |
| 982 | if (!wait) |
| 983 | goto nopage; |
| 984 | |
| 985 | rebalance: |
| 986 | cond_resched(); |
| 987 | |
| 988 | /* We now go into synchronous reclaim */ |
| 989 | cpuset_memory_pressure_bump(); |
| 990 | p->flags |= PF_MEMALLOC; |
| 991 | reclaim_state.reclaimed_slab = 0; |
| 992 | p->reclaim_state = &reclaim_state; |
| 993 | |
| 994 | did_some_progress = try_to_free_pages(zonelist->zones, gfp_mask); |
| 995 | |
| 996 | p->reclaim_state = NULL; |
| 997 | p->flags &= ~PF_MEMALLOC; |
| 998 | |
| 999 | cond_resched(); |
| 1000 | |
| 1001 | if (likely(did_some_progress)) { |
| 1002 | page = get_page_from_freelist(gfp_mask, order, |
| 1003 | zonelist, alloc_flags); |
| 1004 | if (page) |
| 1005 | goto got_pg; |
| 1006 | } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { |
| 1007 | /* |
| 1008 | * Go through the zonelist yet one more time, keep |
| 1009 | * very high watermark here, this is only to catch |
| 1010 | * a parallel oom killing, we must fail if we're still |
| 1011 | * under heavy pressure. |
| 1012 | */ |
| 1013 | page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order, |
| 1014 | zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET); |
| 1015 | if (page) |
| 1016 | goto got_pg; |
| 1017 | |
| 1018 | out_of_memory(zonelist, gfp_mask, order); |
| 1019 | goto restart; |
| 1020 | } |
| 1021 | |
| 1022 | /* |
| 1023 | * Don't let big-order allocations loop unless the caller explicitly |
| 1024 | * requests that. Wait for some write requests to complete then retry. |
| 1025 | * |
| 1026 | * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order |
| 1027 | * <= 3, but that may not be true in other implementations. |
| 1028 | */ |
| 1029 | do_retry = 0; |
| 1030 | if (!(gfp_mask & __GFP_NORETRY)) { |
| 1031 | if ((order <= 3) || (gfp_mask & __GFP_REPEAT)) |
| 1032 | do_retry = 1; |
| 1033 | if (gfp_mask & __GFP_NOFAIL) |
| 1034 | do_retry = 1; |
| 1035 | } |
| 1036 | if (do_retry) { |
| 1037 | blk_congestion_wait(WRITE, HZ/50); |
| 1038 | goto rebalance; |
| 1039 | } |
| 1040 | |
| 1041 | nopage: |
| 1042 | if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { |
| 1043 | printk(KERN_WARNING "%s: page allocation failure." |
| 1044 | " order:%d, mode:0x%x\n", |
| 1045 | p->comm, order, gfp_mask); |
| 1046 | dump_stack(); |
| 1047 | show_mem(); |
| 1048 | } |
| 1049 | got_pg: |
| 1050 | return page; |
| 1051 | } |
| 1052 | |
| 1053 | EXPORT_SYMBOL(__alloc_pages); |
| 1054 | |
| 1055 | /* |
| 1056 | * Common helper functions. |
| 1057 | */ |
| 1058 | fastcall unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) |
| 1059 | { |
| 1060 | struct page * page; |
| 1061 | page = alloc_pages(gfp_mask, order); |
| 1062 | if (!page) |
| 1063 | return 0; |
| 1064 | return (unsigned long) page_address(page); |
| 1065 | } |
| 1066 | |
| 1067 | EXPORT_SYMBOL(__get_free_pages); |
| 1068 | |
| 1069 | fastcall unsigned long get_zeroed_page(gfp_t gfp_mask) |
| 1070 | { |
| 1071 | struct page * page; |
| 1072 | |
| 1073 | /* |
| 1074 | * get_zeroed_page() returns a 32-bit address, which cannot represent |
| 1075 | * a highmem page |
| 1076 | */ |
| 1077 | BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); |
| 1078 | |
| 1079 | page = alloc_pages(gfp_mask | __GFP_ZERO, 0); |
| 1080 | if (page) |
| 1081 | return (unsigned long) page_address(page); |
| 1082 | return 0; |
| 1083 | } |
| 1084 | |
| 1085 | EXPORT_SYMBOL(get_zeroed_page); |
| 1086 | |
| 1087 | void __pagevec_free(struct pagevec *pvec) |
| 1088 | { |
| 1089 | int i = pagevec_count(pvec); |
| 1090 | |
| 1091 | while (--i >= 0) |
| 1092 | free_hot_cold_page(pvec->pages[i], pvec->cold); |
| 1093 | } |
| 1094 | |
| 1095 | fastcall void __free_pages(struct page *page, unsigned int order) |
| 1096 | { |
| 1097 | if (put_page_testzero(page)) { |
| 1098 | if (order == 0) |
| 1099 | free_hot_page(page); |
| 1100 | else |
| 1101 | __free_pages_ok(page, order); |
| 1102 | } |
| 1103 | } |
| 1104 | |
| 1105 | EXPORT_SYMBOL(__free_pages); |
| 1106 | |
| 1107 | fastcall void free_pages(unsigned long addr, unsigned int order) |
| 1108 | { |
| 1109 | if (addr != 0) { |
| 1110 | BUG_ON(!virt_addr_valid((void *)addr)); |
| 1111 | __free_pages(virt_to_page((void *)addr), order); |
| 1112 | } |
| 1113 | } |
| 1114 | |
| 1115 | EXPORT_SYMBOL(free_pages); |
| 1116 | |
| 1117 | /* |
| 1118 | * Total amount of free (allocatable) RAM: |
| 1119 | */ |
| 1120 | unsigned int nr_free_pages(void) |
| 1121 | { |
| 1122 | unsigned int sum = 0; |
| 1123 | struct zone *zone; |
| 1124 | |
| 1125 | for_each_zone(zone) |
| 1126 | sum += zone->free_pages; |
| 1127 | |
| 1128 | return sum; |
| 1129 | } |
| 1130 | |
| 1131 | EXPORT_SYMBOL(nr_free_pages); |
| 1132 | |
| 1133 | #ifdef CONFIG_NUMA |
| 1134 | unsigned int nr_free_pages_pgdat(pg_data_t *pgdat) |
| 1135 | { |
| 1136 | unsigned int i, sum = 0; |
| 1137 | |
| 1138 | for (i = 0; i < MAX_NR_ZONES; i++) |
| 1139 | sum += pgdat->node_zones[i].free_pages; |
| 1140 | |
| 1141 | return sum; |
| 1142 | } |
| 1143 | #endif |
| 1144 | |
| 1145 | static unsigned int nr_free_zone_pages(int offset) |
| 1146 | { |
| 1147 | /* Just pick one node, since fallback list is circular */ |
| 1148 | pg_data_t *pgdat = NODE_DATA(numa_node_id()); |
| 1149 | unsigned int sum = 0; |
| 1150 | |
| 1151 | struct zonelist *zonelist = pgdat->node_zonelists + offset; |
| 1152 | struct zone **zonep = zonelist->zones; |
| 1153 | struct zone *zone; |
| 1154 | |
| 1155 | for (zone = *zonep++; zone; zone = *zonep++) { |
| 1156 | unsigned long size = zone->present_pages; |
| 1157 | unsigned long high = zone->pages_high; |
| 1158 | if (size > high) |
| 1159 | sum += size - high; |
| 1160 | } |
| 1161 | |
| 1162 | return sum; |
| 1163 | } |
| 1164 | |
| 1165 | /* |
| 1166 | * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL |
| 1167 | */ |
| 1168 | unsigned int nr_free_buffer_pages(void) |
| 1169 | { |
| 1170 | return nr_free_zone_pages(gfp_zone(GFP_USER)); |
| 1171 | } |
| 1172 | |
| 1173 | /* |
| 1174 | * Amount of free RAM allocatable within all zones |
| 1175 | */ |
| 1176 | unsigned int nr_free_pagecache_pages(void) |
| 1177 | { |
| 1178 | return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER)); |
| 1179 | } |
| 1180 | |
| 1181 | #ifdef CONFIG_HIGHMEM |
| 1182 | unsigned int nr_free_highpages (void) |
| 1183 | { |
| 1184 | pg_data_t *pgdat; |
| 1185 | unsigned int pages = 0; |
| 1186 | |
| 1187 | for_each_pgdat(pgdat) |
| 1188 | pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages; |
| 1189 | |
| 1190 | return pages; |
| 1191 | } |
| 1192 | #endif |
| 1193 | |
| 1194 | #ifdef CONFIG_NUMA |
| 1195 | static void show_node(struct zone *zone) |
| 1196 | { |
| 1197 | printk("Node %d ", zone->zone_pgdat->node_id); |
| 1198 | } |
| 1199 | #else |
| 1200 | #define show_node(zone) do { } while (0) |
| 1201 | #endif |
| 1202 | |
| 1203 | /* |
| 1204 | * Accumulate the page_state information across all CPUs. |
| 1205 | * The result is unavoidably approximate - it can change |
| 1206 | * during and after execution of this function. |
| 1207 | */ |
| 1208 | static DEFINE_PER_CPU(struct page_state, page_states) = {0}; |
| 1209 | |
| 1210 | atomic_t nr_pagecache = ATOMIC_INIT(0); |
| 1211 | EXPORT_SYMBOL(nr_pagecache); |
| 1212 | #ifdef CONFIG_SMP |
| 1213 | DEFINE_PER_CPU(long, nr_pagecache_local) = 0; |
| 1214 | #endif |
| 1215 | |
| 1216 | static void __get_page_state(struct page_state *ret, int nr, cpumask_t *cpumask) |
| 1217 | { |
| 1218 | int cpu = 0; |
| 1219 | |
| 1220 | memset(ret, 0, nr * sizeof(unsigned long)); |
| 1221 | cpus_and(*cpumask, *cpumask, cpu_online_map); |
| 1222 | |
| 1223 | cpu = first_cpu(*cpumask); |
| 1224 | while (cpu < NR_CPUS) { |
| 1225 | unsigned long *in, *out, off; |
| 1226 | |
| 1227 | if (!cpu_isset(cpu, *cpumask)) |
| 1228 | continue; |
| 1229 | |
| 1230 | in = (unsigned long *)&per_cpu(page_states, cpu); |
| 1231 | |
| 1232 | cpu = next_cpu(cpu, *cpumask); |
| 1233 | |
| 1234 | if (likely(cpu < NR_CPUS)) |
| 1235 | prefetch(&per_cpu(page_states, cpu)); |
| 1236 | |
| 1237 | out = (unsigned long *)ret; |
| 1238 | for (off = 0; off < nr; off++) |
| 1239 | *out++ += *in++; |
| 1240 | } |
| 1241 | } |
| 1242 | |
| 1243 | void get_page_state_node(struct page_state *ret, int node) |
| 1244 | { |
| 1245 | int nr; |
| 1246 | cpumask_t mask = node_to_cpumask(node); |
| 1247 | |
| 1248 | nr = offsetof(struct page_state, GET_PAGE_STATE_LAST); |
| 1249 | nr /= sizeof(unsigned long); |
| 1250 | |
| 1251 | __get_page_state(ret, nr+1, &mask); |
| 1252 | } |
| 1253 | |
| 1254 | void get_page_state(struct page_state *ret) |
| 1255 | { |
| 1256 | int nr; |
| 1257 | cpumask_t mask = CPU_MASK_ALL; |
| 1258 | |
| 1259 | nr = offsetof(struct page_state, GET_PAGE_STATE_LAST); |
| 1260 | nr /= sizeof(unsigned long); |
| 1261 | |
| 1262 | __get_page_state(ret, nr + 1, &mask); |
| 1263 | } |
| 1264 | |
| 1265 | void get_full_page_state(struct page_state *ret) |
| 1266 | { |
| 1267 | cpumask_t mask = CPU_MASK_ALL; |
| 1268 | |
| 1269 | __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long), &mask); |
| 1270 | } |
| 1271 | |
| 1272 | unsigned long read_page_state_offset(unsigned long offset) |
| 1273 | { |
| 1274 | unsigned long ret = 0; |
| 1275 | int cpu; |
| 1276 | |
| 1277 | for_each_online_cpu(cpu) { |
| 1278 | unsigned long in; |
| 1279 | |
| 1280 | in = (unsigned long)&per_cpu(page_states, cpu) + offset; |
| 1281 | ret += *((unsigned long *)in); |
| 1282 | } |
| 1283 | return ret; |
| 1284 | } |
| 1285 | |
| 1286 | void __mod_page_state_offset(unsigned long offset, unsigned long delta) |
| 1287 | { |
| 1288 | void *ptr; |
| 1289 | |
| 1290 | ptr = &__get_cpu_var(page_states); |
| 1291 | *(unsigned long *)(ptr + offset) += delta; |
| 1292 | } |
| 1293 | EXPORT_SYMBOL(__mod_page_state_offset); |
| 1294 | |
| 1295 | void mod_page_state_offset(unsigned long offset, unsigned long delta) |
| 1296 | { |
| 1297 | unsigned long flags; |
| 1298 | void *ptr; |
| 1299 | |
| 1300 | local_irq_save(flags); |
| 1301 | ptr = &__get_cpu_var(page_states); |
| 1302 | *(unsigned long *)(ptr + offset) += delta; |
| 1303 | local_irq_restore(flags); |
| 1304 | } |
| 1305 | EXPORT_SYMBOL(mod_page_state_offset); |
| 1306 | |
| 1307 | void __get_zone_counts(unsigned long *active, unsigned long *inactive, |
| 1308 | unsigned long *free, struct pglist_data *pgdat) |
| 1309 | { |
| 1310 | struct zone *zones = pgdat->node_zones; |
| 1311 | int i; |
| 1312 | |
| 1313 | *active = 0; |
| 1314 | *inactive = 0; |
| 1315 | *free = 0; |
| 1316 | for (i = 0; i < MAX_NR_ZONES; i++) { |
| 1317 | *active += zones[i].nr_active; |
| 1318 | *inactive += zones[i].nr_inactive; |
| 1319 | *free += zones[i].free_pages; |
| 1320 | } |
| 1321 | } |
| 1322 | |
| 1323 | void get_zone_counts(unsigned long *active, |
| 1324 | unsigned long *inactive, unsigned long *free) |
| 1325 | { |
| 1326 | struct pglist_data *pgdat; |
| 1327 | |
| 1328 | *active = 0; |
| 1329 | *inactive = 0; |
| 1330 | *free = 0; |
| 1331 | for_each_pgdat(pgdat) { |
| 1332 | unsigned long l, m, n; |
| 1333 | __get_zone_counts(&l, &m, &n, pgdat); |
| 1334 | *active += l; |
| 1335 | *inactive += m; |
| 1336 | *free += n; |
| 1337 | } |
| 1338 | } |
| 1339 | |
| 1340 | void si_meminfo(struct sysinfo *val) |
| 1341 | { |
| 1342 | val->totalram = totalram_pages; |
| 1343 | val->sharedram = 0; |
| 1344 | val->freeram = nr_free_pages(); |
| 1345 | val->bufferram = nr_blockdev_pages(); |
| 1346 | #ifdef CONFIG_HIGHMEM |
| 1347 | val->totalhigh = totalhigh_pages; |
| 1348 | val->freehigh = nr_free_highpages(); |
| 1349 | #else |
| 1350 | val->totalhigh = 0; |
| 1351 | val->freehigh = 0; |
| 1352 | #endif |
| 1353 | val->mem_unit = PAGE_SIZE; |
| 1354 | } |
| 1355 | |
| 1356 | EXPORT_SYMBOL(si_meminfo); |
| 1357 | |
| 1358 | #ifdef CONFIG_NUMA |
| 1359 | void si_meminfo_node(struct sysinfo *val, int nid) |
| 1360 | { |
| 1361 | pg_data_t *pgdat = NODE_DATA(nid); |
| 1362 | |
| 1363 | val->totalram = pgdat->node_present_pages; |
| 1364 | val->freeram = nr_free_pages_pgdat(pgdat); |
| 1365 | val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; |
| 1366 | val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages; |
| 1367 | val->mem_unit = PAGE_SIZE; |
| 1368 | } |
| 1369 | #endif |
| 1370 | |
| 1371 | #define K(x) ((x) << (PAGE_SHIFT-10)) |
| 1372 | |
| 1373 | /* |
| 1374 | * Show free area list (used inside shift_scroll-lock stuff) |
| 1375 | * We also calculate the percentage fragmentation. We do this by counting the |
| 1376 | * memory on each free list with the exception of the first item on the list. |
| 1377 | */ |
| 1378 | void show_free_areas(void) |
| 1379 | { |
| 1380 | struct page_state ps; |
| 1381 | int cpu, temperature; |
| 1382 | unsigned long active; |
| 1383 | unsigned long inactive; |
| 1384 | unsigned long free; |
| 1385 | struct zone *zone; |
| 1386 | |
| 1387 | for_each_zone(zone) { |
| 1388 | show_node(zone); |
| 1389 | printk("%s per-cpu:", zone->name); |
| 1390 | |
| 1391 | if (!populated_zone(zone)) { |
| 1392 | printk(" empty\n"); |
| 1393 | continue; |
| 1394 | } else |
| 1395 | printk("\n"); |
| 1396 | |
| 1397 | for_each_online_cpu(cpu) { |
| 1398 | struct per_cpu_pageset *pageset; |
| 1399 | |
| 1400 | pageset = zone_pcp(zone, cpu); |
| 1401 | |
| 1402 | for (temperature = 0; temperature < 2; temperature++) |
| 1403 | printk("cpu %d %s: high %d, batch %d used:%d\n", |
| 1404 | cpu, |
| 1405 | temperature ? "cold" : "hot", |
| 1406 | pageset->pcp[temperature].high, |
| 1407 | pageset->pcp[temperature].batch, |
| 1408 | pageset->pcp[temperature].count); |
| 1409 | } |
| 1410 | } |
| 1411 | |
| 1412 | get_page_state(&ps); |
| 1413 | get_zone_counts(&active, &inactive, &free); |
| 1414 | |
| 1415 | printk("Free pages: %11ukB (%ukB HighMem)\n", |
| 1416 | K(nr_free_pages()), |
| 1417 | K(nr_free_highpages())); |
| 1418 | |
| 1419 | printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu " |
| 1420 | "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n", |
| 1421 | active, |
| 1422 | inactive, |
| 1423 | ps.nr_dirty, |
| 1424 | ps.nr_writeback, |
| 1425 | ps.nr_unstable, |
| 1426 | nr_free_pages(), |
| 1427 | ps.nr_slab, |
| 1428 | ps.nr_mapped, |
| 1429 | ps.nr_page_table_pages); |
| 1430 | |
| 1431 | for_each_zone(zone) { |
| 1432 | int i; |
| 1433 | |
| 1434 | show_node(zone); |
| 1435 | printk("%s" |
| 1436 | " free:%lukB" |
| 1437 | " min:%lukB" |
| 1438 | " low:%lukB" |
| 1439 | " high:%lukB" |
| 1440 | " active:%lukB" |
| 1441 | " inactive:%lukB" |
| 1442 | " present:%lukB" |
| 1443 | " pages_scanned:%lu" |
| 1444 | " all_unreclaimable? %s" |
| 1445 | "\n", |
| 1446 | zone->name, |
| 1447 | K(zone->free_pages), |
| 1448 | K(zone->pages_min), |
| 1449 | K(zone->pages_low), |
| 1450 | K(zone->pages_high), |
| 1451 | K(zone->nr_active), |
| 1452 | K(zone->nr_inactive), |
| 1453 | K(zone->present_pages), |
| 1454 | zone->pages_scanned, |
| 1455 | (zone->all_unreclaimable ? "yes" : "no") |
| 1456 | ); |
| 1457 | printk("lowmem_reserve[]:"); |
| 1458 | for (i = 0; i < MAX_NR_ZONES; i++) |
| 1459 | printk(" %lu", zone->lowmem_reserve[i]); |
| 1460 | printk("\n"); |
| 1461 | } |
| 1462 | |
| 1463 | for_each_zone(zone) { |
| 1464 | unsigned long nr, flags, order, total = 0; |
| 1465 | |
| 1466 | show_node(zone); |
| 1467 | printk("%s: ", zone->name); |
| 1468 | if (!populated_zone(zone)) { |
| 1469 | printk("empty\n"); |
| 1470 | continue; |
| 1471 | } |
| 1472 | |
| 1473 | spin_lock_irqsave(&zone->lock, flags); |
| 1474 | for (order = 0; order < MAX_ORDER; order++) { |
| 1475 | nr = zone->free_area[order].nr_free; |
| 1476 | total += nr << order; |
| 1477 | printk("%lu*%lukB ", nr, K(1UL) << order); |
| 1478 | } |
| 1479 | spin_unlock_irqrestore(&zone->lock, flags); |
| 1480 | printk("= %lukB\n", K(total)); |
| 1481 | } |
| 1482 | |
| 1483 | show_swap_cache_info(); |
| 1484 | } |
| 1485 | |
| 1486 | /* |
| 1487 | * Builds allocation fallback zone lists. |
| 1488 | * |
| 1489 | * Add all populated zones of a node to the zonelist. |
| 1490 | */ |
| 1491 | static int __init build_zonelists_node(pg_data_t *pgdat, |
| 1492 | struct zonelist *zonelist, int nr_zones, int zone_type) |
| 1493 | { |
| 1494 | struct zone *zone; |
| 1495 | |
| 1496 | BUG_ON(zone_type > ZONE_HIGHMEM); |
| 1497 | |
| 1498 | do { |
| 1499 | zone = pgdat->node_zones + zone_type; |
| 1500 | if (populated_zone(zone)) { |
| 1501 | #ifndef CONFIG_HIGHMEM |
| 1502 | BUG_ON(zone_type > ZONE_NORMAL); |
| 1503 | #endif |
| 1504 | zonelist->zones[nr_zones++] = zone; |
| 1505 | check_highest_zone(zone_type); |
| 1506 | } |
| 1507 | zone_type--; |
| 1508 | |
| 1509 | } while (zone_type >= 0); |
| 1510 | return nr_zones; |
| 1511 | } |
| 1512 | |
| 1513 | static inline int highest_zone(int zone_bits) |
| 1514 | { |
| 1515 | int res = ZONE_NORMAL; |
| 1516 | if (zone_bits & (__force int)__GFP_HIGHMEM) |
| 1517 | res = ZONE_HIGHMEM; |
| 1518 | if (zone_bits & (__force int)__GFP_DMA32) |
| 1519 | res = ZONE_DMA32; |
| 1520 | if (zone_bits & (__force int)__GFP_DMA) |
| 1521 | res = ZONE_DMA; |
| 1522 | return res; |
| 1523 | } |
| 1524 | |
| 1525 | #ifdef CONFIG_NUMA |
| 1526 | #define MAX_NODE_LOAD (num_online_nodes()) |
| 1527 | static int __initdata node_load[MAX_NUMNODES]; |
| 1528 | /** |
| 1529 | * find_next_best_node - find the next node that should appear in a given node's fallback list |
| 1530 | * @node: node whose fallback list we're appending |
| 1531 | * @used_node_mask: nodemask_t of already used nodes |
| 1532 | * |
| 1533 | * We use a number of factors to determine which is the next node that should |
| 1534 | * appear on a given node's fallback list. The node should not have appeared |
| 1535 | * already in @node's fallback list, and it should be the next closest node |
| 1536 | * according to the distance array (which contains arbitrary distance values |
| 1537 | * from each node to each node in the system), and should also prefer nodes |
| 1538 | * with no CPUs, since presumably they'll have very little allocation pressure |
| 1539 | * on them otherwise. |
| 1540 | * It returns -1 if no node is found. |
| 1541 | */ |
| 1542 | static int __init find_next_best_node(int node, nodemask_t *used_node_mask) |
| 1543 | { |
| 1544 | int n, val; |
| 1545 | int min_val = INT_MAX; |
| 1546 | int best_node = -1; |
| 1547 | |
| 1548 | /* Use the local node if we haven't already */ |
| 1549 | if (!node_isset(node, *used_node_mask)) { |
| 1550 | node_set(node, *used_node_mask); |
| 1551 | return node; |
| 1552 | } |
| 1553 | |
| 1554 | for_each_online_node(n) { |
| 1555 | cpumask_t tmp; |
| 1556 | |
| 1557 | /* Don't want a node to appear more than once */ |
| 1558 | if (node_isset(n, *used_node_mask)) |
| 1559 | continue; |
| 1560 | |
| 1561 | /* Use the distance array to find the distance */ |
| 1562 | val = node_distance(node, n); |
| 1563 | |
| 1564 | /* Penalize nodes under us ("prefer the next node") */ |
| 1565 | val += (n < node); |
| 1566 | |
| 1567 | /* Give preference to headless and unused nodes */ |
| 1568 | tmp = node_to_cpumask(n); |
| 1569 | if (!cpus_empty(tmp)) |
| 1570 | val += PENALTY_FOR_NODE_WITH_CPUS; |
| 1571 | |
| 1572 | /* Slight preference for less loaded node */ |
| 1573 | val *= (MAX_NODE_LOAD*MAX_NUMNODES); |
| 1574 | val += node_load[n]; |
| 1575 | |
| 1576 | if (val < min_val) { |
| 1577 | min_val = val; |
| 1578 | best_node = n; |
| 1579 | } |
| 1580 | } |
| 1581 | |
| 1582 | if (best_node >= 0) |
| 1583 | node_set(best_node, *used_node_mask); |
| 1584 | |
| 1585 | return best_node; |
| 1586 | } |
| 1587 | |
| 1588 | static void __init build_zonelists(pg_data_t *pgdat) |
| 1589 | { |
| 1590 | int i, j, k, node, local_node; |
| 1591 | int prev_node, load; |
| 1592 | struct zonelist *zonelist; |
| 1593 | nodemask_t used_mask; |
| 1594 | |
| 1595 | /* initialize zonelists */ |
| 1596 | for (i = 0; i < GFP_ZONETYPES; i++) { |
| 1597 | zonelist = pgdat->node_zonelists + i; |
| 1598 | zonelist->zones[0] = NULL; |
| 1599 | } |
| 1600 | |
| 1601 | /* NUMA-aware ordering of nodes */ |
| 1602 | local_node = pgdat->node_id; |
| 1603 | load = num_online_nodes(); |
| 1604 | prev_node = local_node; |
| 1605 | nodes_clear(used_mask); |
| 1606 | while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { |
| 1607 | int distance = node_distance(local_node, node); |
| 1608 | |
| 1609 | /* |
| 1610 | * If another node is sufficiently far away then it is better |
| 1611 | * to reclaim pages in a zone before going off node. |
| 1612 | */ |
| 1613 | if (distance > RECLAIM_DISTANCE) |
| 1614 | zone_reclaim_mode = 1; |
| 1615 | |
| 1616 | /* |
| 1617 | * We don't want to pressure a particular node. |
| 1618 | * So adding penalty to the first node in same |
| 1619 | * distance group to make it round-robin. |
| 1620 | */ |
| 1621 | |
| 1622 | if (distance != node_distance(local_node, prev_node)) |
| 1623 | node_load[node] += load; |
| 1624 | prev_node = node; |
| 1625 | load--; |
| 1626 | for (i = 0; i < GFP_ZONETYPES; i++) { |
| 1627 | zonelist = pgdat->node_zonelists + i; |
| 1628 | for (j = 0; zonelist->zones[j] != NULL; j++); |
| 1629 | |
| 1630 | k = highest_zone(i); |
| 1631 | |
| 1632 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, k); |
| 1633 | zonelist->zones[j] = NULL; |
| 1634 | } |
| 1635 | } |
| 1636 | } |
| 1637 | |
| 1638 | #else /* CONFIG_NUMA */ |
| 1639 | |
| 1640 | static void __init build_zonelists(pg_data_t *pgdat) |
| 1641 | { |
| 1642 | int i, j, k, node, local_node; |
| 1643 | |
| 1644 | local_node = pgdat->node_id; |
| 1645 | for (i = 0; i < GFP_ZONETYPES; i++) { |
| 1646 | struct zonelist *zonelist; |
| 1647 | |
| 1648 | zonelist = pgdat->node_zonelists + i; |
| 1649 | |
| 1650 | j = 0; |
| 1651 | k = highest_zone(i); |
| 1652 | j = build_zonelists_node(pgdat, zonelist, j, k); |
| 1653 | /* |
| 1654 | * Now we build the zonelist so that it contains the zones |
| 1655 | * of all the other nodes. |
| 1656 | * We don't want to pressure a particular node, so when |
| 1657 | * building the zones for node N, we make sure that the |
| 1658 | * zones coming right after the local ones are those from |
| 1659 | * node N+1 (modulo N) |
| 1660 | */ |
| 1661 | for (node = local_node + 1; node < MAX_NUMNODES; node++) { |
| 1662 | if (!node_online(node)) |
| 1663 | continue; |
| 1664 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, k); |
| 1665 | } |
| 1666 | for (node = 0; node < local_node; node++) { |
| 1667 | if (!node_online(node)) |
| 1668 | continue; |
| 1669 | j = build_zonelists_node(NODE_DATA(node), zonelist, j, k); |
| 1670 | } |
| 1671 | |
| 1672 | zonelist->zones[j] = NULL; |
| 1673 | } |
| 1674 | } |
| 1675 | |
| 1676 | #endif /* CONFIG_NUMA */ |
| 1677 | |
| 1678 | void __init build_all_zonelists(void) |
| 1679 | { |
| 1680 | int i; |
| 1681 | |
| 1682 | for_each_online_node(i) |
| 1683 | build_zonelists(NODE_DATA(i)); |
| 1684 | printk("Built %i zonelists\n", num_online_nodes()); |
| 1685 | cpuset_init_current_mems_allowed(); |
| 1686 | } |
| 1687 | |
| 1688 | /* |
| 1689 | * Helper functions to size the waitqueue hash table. |
| 1690 | * Essentially these want to choose hash table sizes sufficiently |
| 1691 | * large so that collisions trying to wait on pages are rare. |
| 1692 | * But in fact, the number of active page waitqueues on typical |
| 1693 | * systems is ridiculously low, less than 200. So this is even |
| 1694 | * conservative, even though it seems large. |
| 1695 | * |
| 1696 | * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to |
| 1697 | * waitqueues, i.e. the size of the waitq table given the number of pages. |
| 1698 | */ |
| 1699 | #define PAGES_PER_WAITQUEUE 256 |
| 1700 | |
| 1701 | static inline unsigned long wait_table_size(unsigned long pages) |
| 1702 | { |
| 1703 | unsigned long size = 1; |
| 1704 | |
| 1705 | pages /= PAGES_PER_WAITQUEUE; |
| 1706 | |
| 1707 | while (size < pages) |
| 1708 | size <<= 1; |
| 1709 | |
| 1710 | /* |
| 1711 | * Once we have dozens or even hundreds of threads sleeping |
| 1712 | * on IO we've got bigger problems than wait queue collision. |
| 1713 | * Limit the size of the wait table to a reasonable size. |
| 1714 | */ |
| 1715 | size = min(size, 4096UL); |
| 1716 | |
| 1717 | return max(size, 4UL); |
| 1718 | } |
| 1719 | |
| 1720 | /* |
| 1721 | * This is an integer logarithm so that shifts can be used later |
| 1722 | * to extract the more random high bits from the multiplicative |
| 1723 | * hash function before the remainder is taken. |
| 1724 | */ |
| 1725 | static inline unsigned long wait_table_bits(unsigned long size) |
| 1726 | { |
| 1727 | return ffz(~size); |
| 1728 | } |
| 1729 | |
| 1730 | #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) |
| 1731 | |
| 1732 | static void __init calculate_zone_totalpages(struct pglist_data *pgdat, |
| 1733 | unsigned long *zones_size, unsigned long *zholes_size) |
| 1734 | { |
| 1735 | unsigned long realtotalpages, totalpages = 0; |
| 1736 | int i; |
| 1737 | |
| 1738 | for (i = 0; i < MAX_NR_ZONES; i++) |
| 1739 | totalpages += zones_size[i]; |
| 1740 | pgdat->node_spanned_pages = totalpages; |
| 1741 | |
| 1742 | realtotalpages = totalpages; |
| 1743 | if (zholes_size) |
| 1744 | for (i = 0; i < MAX_NR_ZONES; i++) |
| 1745 | realtotalpages -= zholes_size[i]; |
| 1746 | pgdat->node_present_pages = realtotalpages; |
| 1747 | printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages); |
| 1748 | } |
| 1749 | |
| 1750 | |
| 1751 | /* |
| 1752 | * Initially all pages are reserved - free ones are freed |
| 1753 | * up by free_all_bootmem() once the early boot process is |
| 1754 | * done. Non-atomic initialization, single-pass. |
| 1755 | */ |
| 1756 | void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, |
| 1757 | unsigned long start_pfn) |
| 1758 | { |
| 1759 | struct page *page; |
| 1760 | unsigned long end_pfn = start_pfn + size; |
| 1761 | unsigned long pfn; |
| 1762 | |
| 1763 | for (pfn = start_pfn; pfn < end_pfn; pfn++) { |
| 1764 | if (!early_pfn_valid(pfn)) |
| 1765 | continue; |
| 1766 | page = pfn_to_page(pfn); |
| 1767 | set_page_links(page, zone, nid, pfn); |
| 1768 | set_page_count(page, 1); |
| 1769 | reset_page_mapcount(page); |
| 1770 | SetPageReserved(page); |
| 1771 | INIT_LIST_HEAD(&page->lru); |
| 1772 | #ifdef WANT_PAGE_VIRTUAL |
| 1773 | /* The shift won't overflow because ZONE_NORMAL is below 4G. */ |
| 1774 | if (!is_highmem_idx(zone)) |
| 1775 | set_page_address(page, __va(pfn << PAGE_SHIFT)); |
| 1776 | #endif |
| 1777 | } |
| 1778 | } |
| 1779 | |
| 1780 | void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone, |
| 1781 | unsigned long size) |
| 1782 | { |
| 1783 | int order; |
| 1784 | for (order = 0; order < MAX_ORDER ; order++) { |
| 1785 | INIT_LIST_HEAD(&zone->free_area[order].free_list); |
| 1786 | zone->free_area[order].nr_free = 0; |
| 1787 | } |
| 1788 | } |
| 1789 | |
| 1790 | #define ZONETABLE_INDEX(x, zone_nr) ((x << ZONES_SHIFT) | zone_nr) |
| 1791 | void zonetable_add(struct zone *zone, int nid, int zid, unsigned long pfn, |
| 1792 | unsigned long size) |
| 1793 | { |
| 1794 | unsigned long snum = pfn_to_section_nr(pfn); |
| 1795 | unsigned long end = pfn_to_section_nr(pfn + size); |
| 1796 | |
| 1797 | if (FLAGS_HAS_NODE) |
| 1798 | zone_table[ZONETABLE_INDEX(nid, zid)] = zone; |
| 1799 | else |
| 1800 | for (; snum <= end; snum++) |
| 1801 | zone_table[ZONETABLE_INDEX(snum, zid)] = zone; |
| 1802 | } |
| 1803 | |
| 1804 | #ifndef __HAVE_ARCH_MEMMAP_INIT |
| 1805 | #define memmap_init(size, nid, zone, start_pfn) \ |
| 1806 | memmap_init_zone((size), (nid), (zone), (start_pfn)) |
| 1807 | #endif |
| 1808 | |
| 1809 | static int __cpuinit zone_batchsize(struct zone *zone) |
| 1810 | { |
| 1811 | int batch; |
| 1812 | |
| 1813 | /* |
| 1814 | * The per-cpu-pages pools are set to around 1000th of the |
| 1815 | * size of the zone. But no more than 1/2 of a meg. |
| 1816 | * |
| 1817 | * OK, so we don't know how big the cache is. So guess. |
| 1818 | */ |
| 1819 | batch = zone->present_pages / 1024; |
| 1820 | if (batch * PAGE_SIZE > 512 * 1024) |
| 1821 | batch = (512 * 1024) / PAGE_SIZE; |
| 1822 | batch /= 4; /* We effectively *= 4 below */ |
| 1823 | if (batch < 1) |
| 1824 | batch = 1; |
| 1825 | |
| 1826 | /* |
| 1827 | * Clamp the batch to a 2^n - 1 value. Having a power |
| 1828 | * of 2 value was found to be more likely to have |
| 1829 | * suboptimal cache aliasing properties in some cases. |
| 1830 | * |
| 1831 | * For example if 2 tasks are alternately allocating |
| 1832 | * batches of pages, one task can end up with a lot |
| 1833 | * of pages of one half of the possible page colors |
| 1834 | * and the other with pages of the other colors. |
| 1835 | */ |
| 1836 | batch = (1 << (fls(batch + batch/2)-1)) - 1; |
| 1837 | |
| 1838 | return batch; |
| 1839 | } |
| 1840 | |
| 1841 | inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) |
| 1842 | { |
| 1843 | struct per_cpu_pages *pcp; |
| 1844 | |
| 1845 | memset(p, 0, sizeof(*p)); |
| 1846 | |
| 1847 | pcp = &p->pcp[0]; /* hot */ |
| 1848 | pcp->count = 0; |
| 1849 | pcp->high = 6 * batch; |
| 1850 | pcp->batch = max(1UL, 1 * batch); |
| 1851 | INIT_LIST_HEAD(&pcp->list); |
| 1852 | |
| 1853 | pcp = &p->pcp[1]; /* cold*/ |
| 1854 | pcp->count = 0; |
| 1855 | pcp->high = 2 * batch; |
| 1856 | pcp->batch = max(1UL, batch/2); |
| 1857 | INIT_LIST_HEAD(&pcp->list); |
| 1858 | } |
| 1859 | |
| 1860 | /* |
| 1861 | * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist |
| 1862 | * to the value high for the pageset p. |
| 1863 | */ |
| 1864 | |
| 1865 | static void setup_pagelist_highmark(struct per_cpu_pageset *p, |
| 1866 | unsigned long high) |
| 1867 | { |
| 1868 | struct per_cpu_pages *pcp; |
| 1869 | |
| 1870 | pcp = &p->pcp[0]; /* hot list */ |
| 1871 | pcp->high = high; |
| 1872 | pcp->batch = max(1UL, high/4); |
| 1873 | if ((high/4) > (PAGE_SHIFT * 8)) |
| 1874 | pcp->batch = PAGE_SHIFT * 8; |
| 1875 | } |
| 1876 | |
| 1877 | |
| 1878 | #ifdef CONFIG_NUMA |
| 1879 | /* |
| 1880 | * Boot pageset table. One per cpu which is going to be used for all |
| 1881 | * zones and all nodes. The parameters will be set in such a way |
| 1882 | * that an item put on a list will immediately be handed over to |
| 1883 | * the buddy list. This is safe since pageset manipulation is done |
| 1884 | * with interrupts disabled. |
| 1885 | * |
| 1886 | * Some NUMA counter updates may also be caught by the boot pagesets. |
| 1887 | * |
| 1888 | * The boot_pagesets must be kept even after bootup is complete for |
| 1889 | * unused processors and/or zones. They do play a role for bootstrapping |
| 1890 | * hotplugged processors. |
| 1891 | * |
| 1892 | * zoneinfo_show() and maybe other functions do |
| 1893 | * not check if the processor is online before following the pageset pointer. |
| 1894 | * Other parts of the kernel may not check if the zone is available. |
| 1895 | */ |
| 1896 | static struct per_cpu_pageset boot_pageset[NR_CPUS]; |
| 1897 | |
| 1898 | /* |
| 1899 | * Dynamically allocate memory for the |
| 1900 | * per cpu pageset array in struct zone. |
| 1901 | */ |
| 1902 | static int __cpuinit process_zones(int cpu) |
| 1903 | { |
| 1904 | struct zone *zone, *dzone; |
| 1905 | |
| 1906 | for_each_zone(zone) { |
| 1907 | |
| 1908 | zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset), |
| 1909 | GFP_KERNEL, cpu_to_node(cpu)); |
| 1910 | if (!zone_pcp(zone, cpu)) |
| 1911 | goto bad; |
| 1912 | |
| 1913 | setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone)); |
| 1914 | |
| 1915 | if (percpu_pagelist_fraction) |
| 1916 | setup_pagelist_highmark(zone_pcp(zone, cpu), |
| 1917 | (zone->present_pages / percpu_pagelist_fraction)); |
| 1918 | } |
| 1919 | |
| 1920 | return 0; |
| 1921 | bad: |
| 1922 | for_each_zone(dzone) { |
| 1923 | if (dzone == zone) |
| 1924 | break; |
| 1925 | kfree(zone_pcp(dzone, cpu)); |
| 1926 | zone_pcp(dzone, cpu) = NULL; |
| 1927 | } |
| 1928 | return -ENOMEM; |
| 1929 | } |
| 1930 | |
| 1931 | static inline void free_zone_pagesets(int cpu) |
| 1932 | { |
| 1933 | struct zone *zone; |
| 1934 | |
| 1935 | for_each_zone(zone) { |
| 1936 | struct per_cpu_pageset *pset = zone_pcp(zone, cpu); |
| 1937 | |
| 1938 | zone_pcp(zone, cpu) = NULL; |
| 1939 | kfree(pset); |
| 1940 | } |
| 1941 | } |
| 1942 | |
| 1943 | static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb, |
| 1944 | unsigned long action, |
| 1945 | void *hcpu) |
| 1946 | { |
| 1947 | int cpu = (long)hcpu; |
| 1948 | int ret = NOTIFY_OK; |
| 1949 | |
| 1950 | switch (action) { |
| 1951 | case CPU_UP_PREPARE: |
| 1952 | if (process_zones(cpu)) |
| 1953 | ret = NOTIFY_BAD; |
| 1954 | break; |
| 1955 | case CPU_UP_CANCELED: |
| 1956 | case CPU_DEAD: |
| 1957 | free_zone_pagesets(cpu); |
| 1958 | break; |
| 1959 | default: |
| 1960 | break; |
| 1961 | } |
| 1962 | return ret; |
| 1963 | } |
| 1964 | |
| 1965 | static struct notifier_block pageset_notifier = |
| 1966 | { &pageset_cpuup_callback, NULL, 0 }; |
| 1967 | |
| 1968 | void __init setup_per_cpu_pageset(void) |
| 1969 | { |
| 1970 | int err; |
| 1971 | |
| 1972 | /* Initialize per_cpu_pageset for cpu 0. |
| 1973 | * A cpuup callback will do this for every cpu |
| 1974 | * as it comes online |
| 1975 | */ |
| 1976 | err = process_zones(smp_processor_id()); |
| 1977 | BUG_ON(err); |
| 1978 | register_cpu_notifier(&pageset_notifier); |
| 1979 | } |
| 1980 | |
| 1981 | #endif |
| 1982 | |
| 1983 | static __meminit |
| 1984 | void zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) |
| 1985 | { |
| 1986 | int i; |
| 1987 | struct pglist_data *pgdat = zone->zone_pgdat; |
| 1988 | |
| 1989 | /* |
| 1990 | * The per-page waitqueue mechanism uses hashed waitqueues |
| 1991 | * per zone. |
| 1992 | */ |
| 1993 | zone->wait_table_size = wait_table_size(zone_size_pages); |
| 1994 | zone->wait_table_bits = wait_table_bits(zone->wait_table_size); |
| 1995 | zone->wait_table = (wait_queue_head_t *) |
| 1996 | alloc_bootmem_node(pgdat, zone->wait_table_size |
| 1997 | * sizeof(wait_queue_head_t)); |
| 1998 | |
| 1999 | for(i = 0; i < zone->wait_table_size; ++i) |
| 2000 | init_waitqueue_head(zone->wait_table + i); |
| 2001 | } |
| 2002 | |
| 2003 | static __meminit void zone_pcp_init(struct zone *zone) |
| 2004 | { |
| 2005 | int cpu; |
| 2006 | unsigned long batch = zone_batchsize(zone); |
| 2007 | |
| 2008 | for (cpu = 0; cpu < NR_CPUS; cpu++) { |
| 2009 | #ifdef CONFIG_NUMA |
| 2010 | /* Early boot. Slab allocator not functional yet */ |
| 2011 | zone_pcp(zone, cpu) = &boot_pageset[cpu]; |
| 2012 | setup_pageset(&boot_pageset[cpu],0); |
| 2013 | #else |
| 2014 | setup_pageset(zone_pcp(zone,cpu), batch); |
| 2015 | #endif |
| 2016 | } |
| 2017 | printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n", |
| 2018 | zone->name, zone->present_pages, batch); |
| 2019 | } |
| 2020 | |
| 2021 | static __meminit void init_currently_empty_zone(struct zone *zone, |
| 2022 | unsigned long zone_start_pfn, unsigned long size) |
| 2023 | { |
| 2024 | struct pglist_data *pgdat = zone->zone_pgdat; |
| 2025 | |
| 2026 | zone_wait_table_init(zone, size); |
| 2027 | pgdat->nr_zones = zone_idx(zone) + 1; |
| 2028 | |
| 2029 | zone->zone_mem_map = pfn_to_page(zone_start_pfn); |
| 2030 | zone->zone_start_pfn = zone_start_pfn; |
| 2031 | |
| 2032 | memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn); |
| 2033 | |
| 2034 | zone_init_free_lists(pgdat, zone, zone->spanned_pages); |
| 2035 | } |
| 2036 | |
| 2037 | /* |
| 2038 | * Set up the zone data structures: |
| 2039 | * - mark all pages reserved |
| 2040 | * - mark all memory queues empty |
| 2041 | * - clear the memory bitmaps |
| 2042 | */ |
| 2043 | static void __init free_area_init_core(struct pglist_data *pgdat, |
| 2044 | unsigned long *zones_size, unsigned long *zholes_size) |
| 2045 | { |
| 2046 | unsigned long j; |
| 2047 | int nid = pgdat->node_id; |
| 2048 | unsigned long zone_start_pfn = pgdat->node_start_pfn; |
| 2049 | |
| 2050 | pgdat_resize_init(pgdat); |
| 2051 | pgdat->nr_zones = 0; |
| 2052 | init_waitqueue_head(&pgdat->kswapd_wait); |
| 2053 | pgdat->kswapd_max_order = 0; |
| 2054 | |
| 2055 | for (j = 0; j < MAX_NR_ZONES; j++) { |
| 2056 | struct zone *zone = pgdat->node_zones + j; |
| 2057 | unsigned long size, realsize; |
| 2058 | |
| 2059 | realsize = size = zones_size[j]; |
| 2060 | if (zholes_size) |
| 2061 | realsize -= zholes_size[j]; |
| 2062 | |
| 2063 | if (j < ZONE_HIGHMEM) |
| 2064 | nr_kernel_pages += realsize; |
| 2065 | nr_all_pages += realsize; |
| 2066 | |
| 2067 | zone->spanned_pages = size; |
| 2068 | zone->present_pages = realsize; |
| 2069 | zone->name = zone_names[j]; |
| 2070 | spin_lock_init(&zone->lock); |
| 2071 | spin_lock_init(&zone->lru_lock); |
| 2072 | zone_seqlock_init(zone); |
| 2073 | zone->zone_pgdat = pgdat; |
| 2074 | zone->free_pages = 0; |
| 2075 | |
| 2076 | zone->temp_priority = zone->prev_priority = DEF_PRIORITY; |
| 2077 | |
| 2078 | zone_pcp_init(zone); |
| 2079 | INIT_LIST_HEAD(&zone->active_list); |
| 2080 | INIT_LIST_HEAD(&zone->inactive_list); |
| 2081 | zone->nr_scan_active = 0; |
| 2082 | zone->nr_scan_inactive = 0; |
| 2083 | zone->nr_active = 0; |
| 2084 | zone->nr_inactive = 0; |
| 2085 | atomic_set(&zone->reclaim_in_progress, 0); |
| 2086 | if (!size) |
| 2087 | continue; |
| 2088 | |
| 2089 | zonetable_add(zone, nid, j, zone_start_pfn, size); |
| 2090 | init_currently_empty_zone(zone, zone_start_pfn, size); |
| 2091 | zone_start_pfn += size; |
| 2092 | } |
| 2093 | } |
| 2094 | |
| 2095 | static void __init alloc_node_mem_map(struct pglist_data *pgdat) |
| 2096 | { |
| 2097 | /* Skip empty nodes */ |
| 2098 | if (!pgdat->node_spanned_pages) |
| 2099 | return; |
| 2100 | |
| 2101 | #ifdef CONFIG_FLAT_NODE_MEM_MAP |
| 2102 | /* ia64 gets its own node_mem_map, before this, without bootmem */ |
| 2103 | if (!pgdat->node_mem_map) { |
| 2104 | unsigned long size; |
| 2105 | struct page *map; |
| 2106 | |
| 2107 | size = (pgdat->node_spanned_pages + 1) * sizeof(struct page); |
| 2108 | map = alloc_remap(pgdat->node_id, size); |
| 2109 | if (!map) |
| 2110 | map = alloc_bootmem_node(pgdat, size); |
| 2111 | pgdat->node_mem_map = map; |
| 2112 | } |
| 2113 | #ifdef CONFIG_FLATMEM |
| 2114 | /* |
| 2115 | * With no DISCONTIG, the global mem_map is just set as node 0's |
| 2116 | */ |
| 2117 | if (pgdat == NODE_DATA(0)) |
| 2118 | mem_map = NODE_DATA(0)->node_mem_map; |
| 2119 | #endif |
| 2120 | #endif /* CONFIG_FLAT_NODE_MEM_MAP */ |
| 2121 | } |
| 2122 | |
| 2123 | void __init free_area_init_node(int nid, struct pglist_data *pgdat, |
| 2124 | unsigned long *zones_size, unsigned long node_start_pfn, |
| 2125 | unsigned long *zholes_size) |
| 2126 | { |
| 2127 | pgdat->node_id = nid; |
| 2128 | pgdat->node_start_pfn = node_start_pfn; |
| 2129 | calculate_zone_totalpages(pgdat, zones_size, zholes_size); |
| 2130 | |
| 2131 | alloc_node_mem_map(pgdat); |
| 2132 | |
| 2133 | free_area_init_core(pgdat, zones_size, zholes_size); |
| 2134 | } |
| 2135 | |
| 2136 | #ifndef CONFIG_NEED_MULTIPLE_NODES |
| 2137 | static bootmem_data_t contig_bootmem_data; |
| 2138 | struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data }; |
| 2139 | |
| 2140 | EXPORT_SYMBOL(contig_page_data); |
| 2141 | #endif |
| 2142 | |
| 2143 | void __init free_area_init(unsigned long *zones_size) |
| 2144 | { |
| 2145 | free_area_init_node(0, NODE_DATA(0), zones_size, |
| 2146 | __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); |
| 2147 | } |
| 2148 | |
| 2149 | #ifdef CONFIG_PROC_FS |
| 2150 | |
| 2151 | #include <linux/seq_file.h> |
| 2152 | |
| 2153 | static void *frag_start(struct seq_file *m, loff_t *pos) |
| 2154 | { |
| 2155 | pg_data_t *pgdat; |
| 2156 | loff_t node = *pos; |
| 2157 | |
| 2158 | for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next) |
| 2159 | --node; |
| 2160 | |
| 2161 | return pgdat; |
| 2162 | } |
| 2163 | |
| 2164 | static void *frag_next(struct seq_file *m, void *arg, loff_t *pos) |
| 2165 | { |
| 2166 | pg_data_t *pgdat = (pg_data_t *)arg; |
| 2167 | |
| 2168 | (*pos)++; |
| 2169 | return pgdat->pgdat_next; |
| 2170 | } |
| 2171 | |
| 2172 | static void frag_stop(struct seq_file *m, void *arg) |
| 2173 | { |
| 2174 | } |
| 2175 | |
| 2176 | /* |
| 2177 | * This walks the free areas for each zone. |
| 2178 | */ |
| 2179 | static int frag_show(struct seq_file *m, void *arg) |
| 2180 | { |
| 2181 | pg_data_t *pgdat = (pg_data_t *)arg; |
| 2182 | struct zone *zone; |
| 2183 | struct zone *node_zones = pgdat->node_zones; |
| 2184 | unsigned long flags; |
| 2185 | int order; |
| 2186 | |
| 2187 | for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) { |
| 2188 | if (!populated_zone(zone)) |
| 2189 | continue; |
| 2190 | |
| 2191 | spin_lock_irqsave(&zone->lock, flags); |
| 2192 | seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name); |
| 2193 | for (order = 0; order < MAX_ORDER; ++order) |
| 2194 | seq_printf(m, "%6lu ", zone->free_area[order].nr_free); |
| 2195 | spin_unlock_irqrestore(&zone->lock, flags); |
| 2196 | seq_putc(m, '\n'); |
| 2197 | } |
| 2198 | return 0; |
| 2199 | } |
| 2200 | |
| 2201 | struct seq_operations fragmentation_op = { |
| 2202 | .start = frag_start, |
| 2203 | .next = frag_next, |
| 2204 | .stop = frag_stop, |
| 2205 | .show = frag_show, |
| 2206 | }; |
| 2207 | |
| 2208 | /* |
| 2209 | * Output information about zones in @pgdat. |
| 2210 | */ |
| 2211 | static int zoneinfo_show(struct seq_file *m, void *arg) |
| 2212 | { |
| 2213 | pg_data_t *pgdat = arg; |
| 2214 | struct zone *zone; |
| 2215 | struct zone *node_zones = pgdat->node_zones; |
| 2216 | unsigned long flags; |
| 2217 | |
| 2218 | for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) { |
| 2219 | int i; |
| 2220 | |
| 2221 | if (!populated_zone(zone)) |
| 2222 | continue; |
| 2223 | |
| 2224 | spin_lock_irqsave(&zone->lock, flags); |
| 2225 | seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name); |
| 2226 | seq_printf(m, |
| 2227 | "\n pages free %lu" |
| 2228 | "\n min %lu" |
| 2229 | "\n low %lu" |
| 2230 | "\n high %lu" |
| 2231 | "\n active %lu" |
| 2232 | "\n inactive %lu" |
| 2233 | "\n scanned %lu (a: %lu i: %lu)" |
| 2234 | "\n spanned %lu" |
| 2235 | "\n present %lu", |
| 2236 | zone->free_pages, |
| 2237 | zone->pages_min, |
| 2238 | zone->pages_low, |
| 2239 | zone->pages_high, |
| 2240 | zone->nr_active, |
| 2241 | zone->nr_inactive, |
| 2242 | zone->pages_scanned, |
| 2243 | zone->nr_scan_active, zone->nr_scan_inactive, |
| 2244 | zone->spanned_pages, |
| 2245 | zone->present_pages); |
| 2246 | seq_printf(m, |
| 2247 | "\n protection: (%lu", |
| 2248 | zone->lowmem_reserve[0]); |
| 2249 | for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++) |
| 2250 | seq_printf(m, ", %lu", zone->lowmem_reserve[i]); |
| 2251 | seq_printf(m, |
| 2252 | ")" |
| 2253 | "\n pagesets"); |
| 2254 | for_each_online_cpu(i) { |
| 2255 | struct per_cpu_pageset *pageset; |
| 2256 | int j; |
| 2257 | |
| 2258 | pageset = zone_pcp(zone, i); |
| 2259 | for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) { |
| 2260 | if (pageset->pcp[j].count) |
| 2261 | break; |
| 2262 | } |
| 2263 | if (j == ARRAY_SIZE(pageset->pcp)) |
| 2264 | continue; |
| 2265 | for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) { |
| 2266 | seq_printf(m, |
| 2267 | "\n cpu: %i pcp: %i" |
| 2268 | "\n count: %i" |
| 2269 | "\n high: %i" |
| 2270 | "\n batch: %i", |
| 2271 | i, j, |
| 2272 | pageset->pcp[j].count, |
| 2273 | pageset->pcp[j].high, |
| 2274 | pageset->pcp[j].batch); |
| 2275 | } |
| 2276 | #ifdef CONFIG_NUMA |
| 2277 | seq_printf(m, |
| 2278 | "\n numa_hit: %lu" |
| 2279 | "\n numa_miss: %lu" |
| 2280 | "\n numa_foreign: %lu" |
| 2281 | "\n interleave_hit: %lu" |
| 2282 | "\n local_node: %lu" |
| 2283 | "\n other_node: %lu", |
| 2284 | pageset->numa_hit, |
| 2285 | pageset->numa_miss, |
| 2286 | pageset->numa_foreign, |
| 2287 | pageset->interleave_hit, |
| 2288 | pageset->local_node, |
| 2289 | pageset->other_node); |
| 2290 | #endif |
| 2291 | } |
| 2292 | seq_printf(m, |
| 2293 | "\n all_unreclaimable: %u" |
| 2294 | "\n prev_priority: %i" |
| 2295 | "\n temp_priority: %i" |
| 2296 | "\n start_pfn: %lu", |
| 2297 | zone->all_unreclaimable, |
| 2298 | zone->prev_priority, |
| 2299 | zone->temp_priority, |
| 2300 | zone->zone_start_pfn); |
| 2301 | spin_unlock_irqrestore(&zone->lock, flags); |
| 2302 | seq_putc(m, '\n'); |
| 2303 | } |
| 2304 | return 0; |
| 2305 | } |
| 2306 | |
| 2307 | struct seq_operations zoneinfo_op = { |
| 2308 | .start = frag_start, /* iterate over all zones. The same as in |
| 2309 | * fragmentation. */ |
| 2310 | .next = frag_next, |
| 2311 | .stop = frag_stop, |
| 2312 | .show = zoneinfo_show, |
| 2313 | }; |
| 2314 | |
| 2315 | static char *vmstat_text[] = { |
| 2316 | "nr_dirty", |
| 2317 | "nr_writeback", |
| 2318 | "nr_unstable", |
| 2319 | "nr_page_table_pages", |
| 2320 | "nr_mapped", |
| 2321 | "nr_slab", |
| 2322 | |
| 2323 | "pgpgin", |
| 2324 | "pgpgout", |
| 2325 | "pswpin", |
| 2326 | "pswpout", |
| 2327 | |
| 2328 | "pgalloc_high", |
| 2329 | "pgalloc_normal", |
| 2330 | "pgalloc_dma32", |
| 2331 | "pgalloc_dma", |
| 2332 | |
| 2333 | "pgfree", |
| 2334 | "pgactivate", |
| 2335 | "pgdeactivate", |
| 2336 | |
| 2337 | "pgfault", |
| 2338 | "pgmajfault", |
| 2339 | |
| 2340 | "pgrefill_high", |
| 2341 | "pgrefill_normal", |
| 2342 | "pgrefill_dma32", |
| 2343 | "pgrefill_dma", |
| 2344 | |
| 2345 | "pgsteal_high", |
| 2346 | "pgsteal_normal", |
| 2347 | "pgsteal_dma32", |
| 2348 | "pgsteal_dma", |
| 2349 | |
| 2350 | "pgscan_kswapd_high", |
| 2351 | "pgscan_kswapd_normal", |
| 2352 | "pgscan_kswapd_dma32", |
| 2353 | "pgscan_kswapd_dma", |
| 2354 | |
| 2355 | "pgscan_direct_high", |
| 2356 | "pgscan_direct_normal", |
| 2357 | "pgscan_direct_dma32", |
| 2358 | "pgscan_direct_dma", |
| 2359 | |
| 2360 | "pginodesteal", |
| 2361 | "slabs_scanned", |
| 2362 | "kswapd_steal", |
| 2363 | "kswapd_inodesteal", |
| 2364 | "pageoutrun", |
| 2365 | "allocstall", |
| 2366 | |
| 2367 | "pgrotated", |
| 2368 | "nr_bounce", |
| 2369 | }; |
| 2370 | |
| 2371 | static void *vmstat_start(struct seq_file *m, loff_t *pos) |
| 2372 | { |
| 2373 | struct page_state *ps; |
| 2374 | |
| 2375 | if (*pos >= ARRAY_SIZE(vmstat_text)) |
| 2376 | return NULL; |
| 2377 | |
| 2378 | ps = kmalloc(sizeof(*ps), GFP_KERNEL); |
| 2379 | m->private = ps; |
| 2380 | if (!ps) |
| 2381 | return ERR_PTR(-ENOMEM); |
| 2382 | get_full_page_state(ps); |
| 2383 | ps->pgpgin /= 2; /* sectors -> kbytes */ |
| 2384 | ps->pgpgout /= 2; |
| 2385 | return (unsigned long *)ps + *pos; |
| 2386 | } |
| 2387 | |
| 2388 | static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos) |
| 2389 | { |
| 2390 | (*pos)++; |
| 2391 | if (*pos >= ARRAY_SIZE(vmstat_text)) |
| 2392 | return NULL; |
| 2393 | return (unsigned long *)m->private + *pos; |
| 2394 | } |
| 2395 | |
| 2396 | static int vmstat_show(struct seq_file *m, void *arg) |
| 2397 | { |
| 2398 | unsigned long *l = arg; |
| 2399 | unsigned long off = l - (unsigned long *)m->private; |
| 2400 | |
| 2401 | seq_printf(m, "%s %lu\n", vmstat_text[off], *l); |
| 2402 | return 0; |
| 2403 | } |
| 2404 | |
| 2405 | static void vmstat_stop(struct seq_file *m, void *arg) |
| 2406 | { |
| 2407 | kfree(m->private); |
| 2408 | m->private = NULL; |
| 2409 | } |
| 2410 | |
| 2411 | struct seq_operations vmstat_op = { |
| 2412 | .start = vmstat_start, |
| 2413 | .next = vmstat_next, |
| 2414 | .stop = vmstat_stop, |
| 2415 | .show = vmstat_show, |
| 2416 | }; |
| 2417 | |
| 2418 | #endif /* CONFIG_PROC_FS */ |
| 2419 | |
| 2420 | #ifdef CONFIG_HOTPLUG_CPU |
| 2421 | static int page_alloc_cpu_notify(struct notifier_block *self, |
| 2422 | unsigned long action, void *hcpu) |
| 2423 | { |
| 2424 | int cpu = (unsigned long)hcpu; |
| 2425 | long *count; |
| 2426 | unsigned long *src, *dest; |
| 2427 | |
| 2428 | if (action == CPU_DEAD) { |
| 2429 | int i; |
| 2430 | |
| 2431 | /* Drain local pagecache count. */ |
| 2432 | count = &per_cpu(nr_pagecache_local, cpu); |
| 2433 | atomic_add(*count, &nr_pagecache); |
| 2434 | *count = 0; |
| 2435 | local_irq_disable(); |
| 2436 | __drain_pages(cpu); |
| 2437 | |
| 2438 | /* Add dead cpu's page_states to our own. */ |
| 2439 | dest = (unsigned long *)&__get_cpu_var(page_states); |
| 2440 | src = (unsigned long *)&per_cpu(page_states, cpu); |
| 2441 | |
| 2442 | for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long); |
| 2443 | i++) { |
| 2444 | dest[i] += src[i]; |
| 2445 | src[i] = 0; |
| 2446 | } |
| 2447 | |
| 2448 | local_irq_enable(); |
| 2449 | } |
| 2450 | return NOTIFY_OK; |
| 2451 | } |
| 2452 | #endif /* CONFIG_HOTPLUG_CPU */ |
| 2453 | |
| 2454 | void __init page_alloc_init(void) |
| 2455 | { |
| 2456 | hotcpu_notifier(page_alloc_cpu_notify, 0); |
| 2457 | } |
| 2458 | |
| 2459 | /* |
| 2460 | * setup_per_zone_lowmem_reserve - called whenever |
| 2461 | * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone |
| 2462 | * has a correct pages reserved value, so an adequate number of |
| 2463 | * pages are left in the zone after a successful __alloc_pages(). |
| 2464 | */ |
| 2465 | static void setup_per_zone_lowmem_reserve(void) |
| 2466 | { |
| 2467 | struct pglist_data *pgdat; |
| 2468 | int j, idx; |
| 2469 | |
| 2470 | for_each_pgdat(pgdat) { |
| 2471 | for (j = 0; j < MAX_NR_ZONES; j++) { |
| 2472 | struct zone *zone = pgdat->node_zones + j; |
| 2473 | unsigned long present_pages = zone->present_pages; |
| 2474 | |
| 2475 | zone->lowmem_reserve[j] = 0; |
| 2476 | |
| 2477 | for (idx = j-1; idx >= 0; idx--) { |
| 2478 | struct zone *lower_zone; |
| 2479 | |
| 2480 | if (sysctl_lowmem_reserve_ratio[idx] < 1) |
| 2481 | sysctl_lowmem_reserve_ratio[idx] = 1; |
| 2482 | |
| 2483 | lower_zone = pgdat->node_zones + idx; |
| 2484 | lower_zone->lowmem_reserve[j] = present_pages / |
| 2485 | sysctl_lowmem_reserve_ratio[idx]; |
| 2486 | present_pages += lower_zone->present_pages; |
| 2487 | } |
| 2488 | } |
| 2489 | } |
| 2490 | } |
| 2491 | |
| 2492 | /* |
| 2493 | * setup_per_zone_pages_min - called when min_free_kbytes changes. Ensures |
| 2494 | * that the pages_{min,low,high} values for each zone are set correctly |
| 2495 | * with respect to min_free_kbytes. |
| 2496 | */ |
| 2497 | void setup_per_zone_pages_min(void) |
| 2498 | { |
| 2499 | unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); |
| 2500 | unsigned long lowmem_pages = 0; |
| 2501 | struct zone *zone; |
| 2502 | unsigned long flags; |
| 2503 | |
| 2504 | /* Calculate total number of !ZONE_HIGHMEM pages */ |
| 2505 | for_each_zone(zone) { |
| 2506 | if (!is_highmem(zone)) |
| 2507 | lowmem_pages += zone->present_pages; |
| 2508 | } |
| 2509 | |
| 2510 | for_each_zone(zone) { |
| 2511 | unsigned long tmp; |
| 2512 | spin_lock_irqsave(&zone->lru_lock, flags); |
| 2513 | tmp = (pages_min * zone->present_pages) / lowmem_pages; |
| 2514 | if (is_highmem(zone)) { |
| 2515 | /* |
| 2516 | * __GFP_HIGH and PF_MEMALLOC allocations usually don't |
| 2517 | * need highmem pages, so cap pages_min to a small |
| 2518 | * value here. |
| 2519 | * |
| 2520 | * The (pages_high-pages_low) and (pages_low-pages_min) |
| 2521 | * deltas controls asynch page reclaim, and so should |
| 2522 | * not be capped for highmem. |
| 2523 | */ |
| 2524 | int min_pages; |
| 2525 | |
| 2526 | min_pages = zone->present_pages / 1024; |
| 2527 | if (min_pages < SWAP_CLUSTER_MAX) |
| 2528 | min_pages = SWAP_CLUSTER_MAX; |
| 2529 | if (min_pages > 128) |
| 2530 | min_pages = 128; |
| 2531 | zone->pages_min = min_pages; |
| 2532 | } else { |
| 2533 | /* |
| 2534 | * If it's a lowmem zone, reserve a number of pages |
| 2535 | * proportionate to the zone's size. |
| 2536 | */ |
| 2537 | zone->pages_min = tmp; |
| 2538 | } |
| 2539 | |
| 2540 | zone->pages_low = zone->pages_min + tmp / 4; |
| 2541 | zone->pages_high = zone->pages_min + tmp / 2; |
| 2542 | spin_unlock_irqrestore(&zone->lru_lock, flags); |
| 2543 | } |
| 2544 | } |
| 2545 | |
| 2546 | /* |
| 2547 | * Initialise min_free_kbytes. |
| 2548 | * |
| 2549 | * For small machines we want it small (128k min). For large machines |
| 2550 | * we want it large (64MB max). But it is not linear, because network |
| 2551 | * bandwidth does not increase linearly with machine size. We use |
| 2552 | * |
| 2553 | * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: |
| 2554 | * min_free_kbytes = sqrt(lowmem_kbytes * 16) |
| 2555 | * |
| 2556 | * which yields |
| 2557 | * |
| 2558 | * 16MB: 512k |
| 2559 | * 32MB: 724k |
| 2560 | * 64MB: 1024k |
| 2561 | * 128MB: 1448k |
| 2562 | * 256MB: 2048k |
| 2563 | * 512MB: 2896k |
| 2564 | * 1024MB: 4096k |
| 2565 | * 2048MB: 5792k |
| 2566 | * 4096MB: 8192k |
| 2567 | * 8192MB: 11584k |
| 2568 | * 16384MB: 16384k |
| 2569 | */ |
| 2570 | static int __init init_per_zone_pages_min(void) |
| 2571 | { |
| 2572 | unsigned long lowmem_kbytes; |
| 2573 | |
| 2574 | lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); |
| 2575 | |
| 2576 | min_free_kbytes = int_sqrt(lowmem_kbytes * 16); |
| 2577 | if (min_free_kbytes < 128) |
| 2578 | min_free_kbytes = 128; |
| 2579 | if (min_free_kbytes > 65536) |
| 2580 | min_free_kbytes = 65536; |
| 2581 | setup_per_zone_pages_min(); |
| 2582 | setup_per_zone_lowmem_reserve(); |
| 2583 | return 0; |
| 2584 | } |
| 2585 | module_init(init_per_zone_pages_min) |
| 2586 | |
| 2587 | /* |
| 2588 | * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so |
| 2589 | * that we can call two helper functions whenever min_free_kbytes |
| 2590 | * changes. |
| 2591 | */ |
| 2592 | int min_free_kbytes_sysctl_handler(ctl_table *table, int write, |
| 2593 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) |
| 2594 | { |
| 2595 | proc_dointvec(table, write, file, buffer, length, ppos); |
| 2596 | setup_per_zone_pages_min(); |
| 2597 | return 0; |
| 2598 | } |
| 2599 | |
| 2600 | /* |
| 2601 | * lowmem_reserve_ratio_sysctl_handler - just a wrapper around |
| 2602 | * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() |
| 2603 | * whenever sysctl_lowmem_reserve_ratio changes. |
| 2604 | * |
| 2605 | * The reserve ratio obviously has absolutely no relation with the |
| 2606 | * pages_min watermarks. The lowmem reserve ratio can only make sense |
| 2607 | * if in function of the boot time zone sizes. |
| 2608 | */ |
| 2609 | int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, |
| 2610 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) |
| 2611 | { |
| 2612 | proc_dointvec_minmax(table, write, file, buffer, length, ppos); |
| 2613 | setup_per_zone_lowmem_reserve(); |
| 2614 | return 0; |
| 2615 | } |
| 2616 | |
| 2617 | /* |
| 2618 | * percpu_pagelist_fraction - changes the pcp->high for each zone on each |
| 2619 | * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist |
| 2620 | * can have before it gets flushed back to buddy allocator. |
| 2621 | */ |
| 2622 | |
| 2623 | int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, |
| 2624 | struct file *file, void __user *buffer, size_t *length, loff_t *ppos) |
| 2625 | { |
| 2626 | struct zone *zone; |
| 2627 | unsigned int cpu; |
| 2628 | int ret; |
| 2629 | |
| 2630 | ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos); |
| 2631 | if (!write || (ret == -EINVAL)) |
| 2632 | return ret; |
| 2633 | for_each_zone(zone) { |
| 2634 | for_each_online_cpu(cpu) { |
| 2635 | unsigned long high; |
| 2636 | high = zone->present_pages / percpu_pagelist_fraction; |
| 2637 | setup_pagelist_highmark(zone_pcp(zone, cpu), high); |
| 2638 | } |
| 2639 | } |
| 2640 | return 0; |
| 2641 | } |
| 2642 | |
| 2643 | __initdata int hashdist = HASHDIST_DEFAULT; |
| 2644 | |
| 2645 | #ifdef CONFIG_NUMA |
| 2646 | static int __init set_hashdist(char *str) |
| 2647 | { |
| 2648 | if (!str) |
| 2649 | return 0; |
| 2650 | hashdist = simple_strtoul(str, &str, 0); |
| 2651 | return 1; |
| 2652 | } |
| 2653 | __setup("hashdist=", set_hashdist); |
| 2654 | #endif |
| 2655 | |
| 2656 | /* |
| 2657 | * allocate a large system hash table from bootmem |
| 2658 | * - it is assumed that the hash table must contain an exact power-of-2 |
| 2659 | * quantity of entries |
| 2660 | * - limit is the number of hash buckets, not the total allocation size |
| 2661 | */ |
| 2662 | void *__init alloc_large_system_hash(const char *tablename, |
| 2663 | unsigned long bucketsize, |
| 2664 | unsigned long numentries, |
| 2665 | int scale, |
| 2666 | int flags, |
| 2667 | unsigned int *_hash_shift, |
| 2668 | unsigned int *_hash_mask, |
| 2669 | unsigned long limit) |
| 2670 | { |
| 2671 | unsigned long long max = limit; |
| 2672 | unsigned long log2qty, size; |
| 2673 | void *table = NULL; |
| 2674 | |
| 2675 | /* allow the kernel cmdline to have a say */ |
| 2676 | if (!numentries) { |
| 2677 | /* round applicable memory size up to nearest megabyte */ |
| 2678 | numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages; |
| 2679 | numentries += (1UL << (20 - PAGE_SHIFT)) - 1; |
| 2680 | numentries >>= 20 - PAGE_SHIFT; |
| 2681 | numentries <<= 20 - PAGE_SHIFT; |
| 2682 | |
| 2683 | /* limit to 1 bucket per 2^scale bytes of low memory */ |
| 2684 | if (scale > PAGE_SHIFT) |
| 2685 | numentries >>= (scale - PAGE_SHIFT); |
| 2686 | else |
| 2687 | numentries <<= (PAGE_SHIFT - scale); |
| 2688 | } |
| 2689 | /* rounded up to nearest power of 2 in size */ |
| 2690 | numentries = 1UL << (long_log2(numentries) + 1); |
| 2691 | |
| 2692 | /* limit allocation size to 1/16 total memory by default */ |
| 2693 | if (max == 0) { |
| 2694 | max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; |
| 2695 | do_div(max, bucketsize); |
| 2696 | } |
| 2697 | |
| 2698 | if (numentries > max) |
| 2699 | numentries = max; |
| 2700 | |
| 2701 | log2qty = long_log2(numentries); |
| 2702 | |
| 2703 | do { |
| 2704 | size = bucketsize << log2qty; |
| 2705 | if (flags & HASH_EARLY) |
| 2706 | table = alloc_bootmem(size); |
| 2707 | else if (hashdist) |
| 2708 | table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); |
| 2709 | else { |
| 2710 | unsigned long order; |
| 2711 | for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++) |
| 2712 | ; |
| 2713 | table = (void*) __get_free_pages(GFP_ATOMIC, order); |
| 2714 | } |
| 2715 | } while (!table && size > PAGE_SIZE && --log2qty); |
| 2716 | |
| 2717 | if (!table) |
| 2718 | panic("Failed to allocate %s hash table\n", tablename); |
| 2719 | |
| 2720 | printk("%s hash table entries: %d (order: %d, %lu bytes)\n", |
| 2721 | tablename, |
| 2722 | (1U << log2qty), |
| 2723 | long_log2(size) - PAGE_SHIFT, |
| 2724 | size); |
| 2725 | |
| 2726 | if (_hash_shift) |
| 2727 | *_hash_shift = log2qty; |
| 2728 | if (_hash_mask) |
| 2729 | *_hash_mask = (1 << log2qty) - 1; |
| 2730 | |
| 2731 | return table; |
| 2732 | } |