Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 1 | /* |
| 2 | * Memory Migration functionality - linux/mm/migration.c |
| 3 | * |
| 4 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter |
| 5 | * |
| 6 | * Page migration was first developed in the context of the memory hotplug |
| 7 | * project. The main authors of the migration code are: |
| 8 | * |
| 9 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> |
| 10 | * Hirokazu Takahashi <taka@valinux.co.jp> |
| 11 | * Dave Hansen <haveblue@us.ibm.com> |
| 12 | * Christoph Lameter <clameter@sgi.com> |
| 13 | */ |
| 14 | |
| 15 | #include <linux/migrate.h> |
| 16 | #include <linux/module.h> |
| 17 | #include <linux/swap.h> |
Christoph Lameter | 0697212 | 2006-06-23 02:03:35 -0700 | [diff] [blame] | 18 | #include <linux/swapops.h> |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 19 | #include <linux/pagemap.h> |
Christoph Lameter | e23ca00 | 2006-04-10 22:52:57 -0700 | [diff] [blame] | 20 | #include <linux/buffer_head.h> |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 21 | #include <linux/mm_inline.h> |
| 22 | #include <linux/pagevec.h> |
| 23 | #include <linux/rmap.h> |
| 24 | #include <linux/topology.h> |
| 25 | #include <linux/cpu.h> |
| 26 | #include <linux/cpuset.h> |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 27 | #include <linux/writeback.h> |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 28 | |
| 29 | #include "internal.h" |
| 30 | |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 31 | #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru)) |
| 32 | |
| 33 | /* |
| 34 | * Isolate one page from the LRU lists. If successful put it onto |
| 35 | * the indicated list with elevated page count. |
| 36 | * |
| 37 | * Result: |
| 38 | * -EBUSY: page not on LRU list |
| 39 | * 0: page removed from LRU list and added to the specified list. |
| 40 | */ |
| 41 | int isolate_lru_page(struct page *page, struct list_head *pagelist) |
| 42 | { |
| 43 | int ret = -EBUSY; |
| 44 | |
| 45 | if (PageLRU(page)) { |
| 46 | struct zone *zone = page_zone(page); |
| 47 | |
| 48 | spin_lock_irq(&zone->lru_lock); |
| 49 | if (PageLRU(page)) { |
| 50 | ret = 0; |
| 51 | get_page(page); |
| 52 | ClearPageLRU(page); |
| 53 | if (PageActive(page)) |
| 54 | del_page_from_active_list(zone, page); |
| 55 | else |
| 56 | del_page_from_inactive_list(zone, page); |
| 57 | list_add_tail(&page->lru, pagelist); |
| 58 | } |
| 59 | spin_unlock_irq(&zone->lru_lock); |
| 60 | } |
| 61 | return ret; |
| 62 | } |
| 63 | |
| 64 | /* |
| 65 | * migrate_prep() needs to be called after we have compiled the list of pages |
| 66 | * to be migrated using isolate_lru_page() but before we begin a series of calls |
| 67 | * to migrate_pages(). |
| 68 | */ |
| 69 | int migrate_prep(void) |
| 70 | { |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 71 | /* |
| 72 | * Clear the LRU lists so pages can be isolated. |
| 73 | * Note that pages may be moved off the LRU after we have |
| 74 | * drained them. Those pages will fail to migrate like other |
| 75 | * pages that may be busy. |
| 76 | */ |
| 77 | lru_add_drain_all(); |
| 78 | |
| 79 | return 0; |
| 80 | } |
| 81 | |
| 82 | static inline void move_to_lru(struct page *page) |
| 83 | { |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 84 | if (PageActive(page)) { |
| 85 | /* |
| 86 | * lru_cache_add_active checks that |
| 87 | * the PG_active bit is off. |
| 88 | */ |
| 89 | ClearPageActive(page); |
| 90 | lru_cache_add_active(page); |
| 91 | } else { |
| 92 | lru_cache_add(page); |
| 93 | } |
| 94 | put_page(page); |
| 95 | } |
| 96 | |
| 97 | /* |
| 98 | * Add isolated pages on the list back to the LRU. |
| 99 | * |
| 100 | * returns the number of pages put back. |
| 101 | */ |
| 102 | int putback_lru_pages(struct list_head *l) |
| 103 | { |
| 104 | struct page *page; |
| 105 | struct page *page2; |
| 106 | int count = 0; |
| 107 | |
| 108 | list_for_each_entry_safe(page, page2, l, lru) { |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 109 | list_del(&page->lru); |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 110 | move_to_lru(page); |
| 111 | count++; |
| 112 | } |
| 113 | return count; |
| 114 | } |
| 115 | |
Christoph Lameter | 0697212 | 2006-06-23 02:03:35 -0700 | [diff] [blame] | 116 | static inline int is_swap_pte(pte_t pte) |
| 117 | { |
| 118 | return !pte_none(pte) && !pte_present(pte) && !pte_file(pte); |
| 119 | } |
| 120 | |
| 121 | /* |
| 122 | * Restore a potential migration pte to a working pte entry |
| 123 | */ |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 124 | static void remove_migration_pte(struct vm_area_struct *vma, |
Christoph Lameter | 0697212 | 2006-06-23 02:03:35 -0700 | [diff] [blame] | 125 | struct page *old, struct page *new) |
| 126 | { |
| 127 | struct mm_struct *mm = vma->vm_mm; |
| 128 | swp_entry_t entry; |
| 129 | pgd_t *pgd; |
| 130 | pud_t *pud; |
| 131 | pmd_t *pmd; |
| 132 | pte_t *ptep, pte; |
| 133 | spinlock_t *ptl; |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 134 | unsigned long addr = page_address_in_vma(new, vma); |
| 135 | |
| 136 | if (addr == -EFAULT) |
| 137 | return; |
Christoph Lameter | 0697212 | 2006-06-23 02:03:35 -0700 | [diff] [blame] | 138 | |
| 139 | pgd = pgd_offset(mm, addr); |
| 140 | if (!pgd_present(*pgd)) |
| 141 | return; |
| 142 | |
| 143 | pud = pud_offset(pgd, addr); |
| 144 | if (!pud_present(*pud)) |
| 145 | return; |
| 146 | |
| 147 | pmd = pmd_offset(pud, addr); |
| 148 | if (!pmd_present(*pmd)) |
| 149 | return; |
| 150 | |
| 151 | ptep = pte_offset_map(pmd, addr); |
| 152 | |
| 153 | if (!is_swap_pte(*ptep)) { |
| 154 | pte_unmap(ptep); |
| 155 | return; |
| 156 | } |
| 157 | |
| 158 | ptl = pte_lockptr(mm, pmd); |
| 159 | spin_lock(ptl); |
| 160 | pte = *ptep; |
| 161 | if (!is_swap_pte(pte)) |
| 162 | goto out; |
| 163 | |
| 164 | entry = pte_to_swp_entry(pte); |
| 165 | |
| 166 | if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old) |
| 167 | goto out; |
| 168 | |
Christoph Lameter | 0697212 | 2006-06-23 02:03:35 -0700 | [diff] [blame] | 169 | get_page(new); |
| 170 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); |
| 171 | if (is_write_migration_entry(entry)) |
| 172 | pte = pte_mkwrite(pte); |
| 173 | set_pte_at(mm, addr, ptep, pte); |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 174 | |
| 175 | if (PageAnon(new)) |
| 176 | page_add_anon_rmap(new, vma, addr); |
| 177 | else |
| 178 | page_add_file_rmap(new); |
| 179 | |
| 180 | /* No need to invalidate - it was non-present before */ |
| 181 | update_mmu_cache(vma, addr, pte); |
| 182 | lazy_mmu_prot_update(pte); |
| 183 | |
Christoph Lameter | 0697212 | 2006-06-23 02:03:35 -0700 | [diff] [blame] | 184 | out: |
| 185 | pte_unmap_unlock(ptep, ptl); |
| 186 | } |
| 187 | |
| 188 | /* |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 189 | * Note that remove_file_migration_ptes will only work on regular mappings, |
| 190 | * Nonlinear mappings do not use migration entries. |
| 191 | */ |
| 192 | static void remove_file_migration_ptes(struct page *old, struct page *new) |
| 193 | { |
| 194 | struct vm_area_struct *vma; |
| 195 | struct address_space *mapping = page_mapping(new); |
| 196 | struct prio_tree_iter iter; |
| 197 | pgoff_t pgoff = new->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 198 | |
| 199 | if (!mapping) |
| 200 | return; |
| 201 | |
| 202 | spin_lock(&mapping->i_mmap_lock); |
| 203 | |
| 204 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) |
| 205 | remove_migration_pte(vma, old, new); |
| 206 | |
| 207 | spin_unlock(&mapping->i_mmap_lock); |
| 208 | } |
| 209 | |
| 210 | /* |
Christoph Lameter | 0697212 | 2006-06-23 02:03:35 -0700 | [diff] [blame] | 211 | * Must hold mmap_sem lock on at least one of the vmas containing |
| 212 | * the page so that the anon_vma cannot vanish. |
| 213 | */ |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 214 | static void remove_anon_migration_ptes(struct page *old, struct page *new) |
Christoph Lameter | 0697212 | 2006-06-23 02:03:35 -0700 | [diff] [blame] | 215 | { |
| 216 | struct anon_vma *anon_vma; |
| 217 | struct vm_area_struct *vma; |
| 218 | unsigned long mapping; |
| 219 | |
| 220 | mapping = (unsigned long)new->mapping; |
| 221 | |
| 222 | if (!mapping || (mapping & PAGE_MAPPING_ANON) == 0) |
| 223 | return; |
| 224 | |
| 225 | /* |
| 226 | * We hold the mmap_sem lock. So no need to call page_lock_anon_vma. |
| 227 | */ |
| 228 | anon_vma = (struct anon_vma *) (mapping - PAGE_MAPPING_ANON); |
| 229 | spin_lock(&anon_vma->lock); |
| 230 | |
| 231 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 232 | remove_migration_pte(vma, old, new); |
Christoph Lameter | 0697212 | 2006-06-23 02:03:35 -0700 | [diff] [blame] | 233 | |
| 234 | spin_unlock(&anon_vma->lock); |
| 235 | } |
| 236 | |
| 237 | /* |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 238 | * Get rid of all migration entries and replace them by |
| 239 | * references to the indicated page. |
| 240 | */ |
| 241 | static void remove_migration_ptes(struct page *old, struct page *new) |
| 242 | { |
| 243 | if (PageAnon(new)) |
| 244 | remove_anon_migration_ptes(old, new); |
| 245 | else |
| 246 | remove_file_migration_ptes(old, new); |
| 247 | } |
| 248 | |
| 249 | /* |
Christoph Lameter | 0697212 | 2006-06-23 02:03:35 -0700 | [diff] [blame] | 250 | * Something used the pte of a page under migration. We need to |
| 251 | * get to the page and wait until migration is finished. |
| 252 | * When we return from this function the fault will be retried. |
| 253 | * |
| 254 | * This function is called from do_swap_page(). |
| 255 | */ |
| 256 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
| 257 | unsigned long address) |
| 258 | { |
| 259 | pte_t *ptep, pte; |
| 260 | spinlock_t *ptl; |
| 261 | swp_entry_t entry; |
| 262 | struct page *page; |
| 263 | |
| 264 | ptep = pte_offset_map_lock(mm, pmd, address, &ptl); |
| 265 | pte = *ptep; |
| 266 | if (!is_swap_pte(pte)) |
| 267 | goto out; |
| 268 | |
| 269 | entry = pte_to_swp_entry(pte); |
| 270 | if (!is_migration_entry(entry)) |
| 271 | goto out; |
| 272 | |
| 273 | page = migration_entry_to_page(entry); |
| 274 | |
| 275 | get_page(page); |
| 276 | pte_unmap_unlock(ptep, ptl); |
| 277 | wait_on_page_locked(page); |
| 278 | put_page(page); |
| 279 | return; |
| 280 | out: |
| 281 | pte_unmap_unlock(ptep, ptl); |
| 282 | } |
| 283 | |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 284 | /* |
Christoph Lameter | c3fcf8a | 2006-06-23 02:03:32 -0700 | [diff] [blame] | 285 | * Replace the page in the mapping. |
Christoph Lameter | 5b5c712 | 2006-06-23 02:03:29 -0700 | [diff] [blame] | 286 | * |
| 287 | * The number of remaining references must be: |
| 288 | * 1 for anonymous pages without a mapping |
| 289 | * 2 for pages with a mapping |
| 290 | * 3 for pages with a mapping and PagePrivate set. |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 291 | */ |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 292 | static int migrate_page_move_mapping(struct address_space *mapping, |
| 293 | struct page *newpage, struct page *page) |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 294 | { |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 295 | struct page **radix_pointer; |
| 296 | |
Christoph Lameter | 6c5240a | 2006-06-23 02:03:37 -0700 | [diff] [blame] | 297 | if (!mapping) { |
| 298 | /* Anonymous page */ |
| 299 | if (page_count(page) != 1) |
| 300 | return -EAGAIN; |
| 301 | return 0; |
| 302 | } |
| 303 | |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 304 | write_lock_irq(&mapping->tree_lock); |
| 305 | |
| 306 | radix_pointer = (struct page **)radix_tree_lookup_slot( |
| 307 | &mapping->page_tree, |
| 308 | page_index(page)); |
| 309 | |
Christoph Lameter | 6c5240a | 2006-06-23 02:03:37 -0700 | [diff] [blame] | 310 | if (page_count(page) != 2 + !!PagePrivate(page) || |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 311 | *radix_pointer != page) { |
| 312 | write_unlock_irq(&mapping->tree_lock); |
Christoph Lameter | e23ca00 | 2006-04-10 22:52:57 -0700 | [diff] [blame] | 313 | return -EAGAIN; |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 314 | } |
| 315 | |
| 316 | /* |
| 317 | * Now we know that no one else is looking at the page. |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 318 | */ |
| 319 | get_page(newpage); |
Christoph Lameter | 6c5240a | 2006-06-23 02:03:37 -0700 | [diff] [blame] | 320 | #ifdef CONFIG_SWAP |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 321 | if (PageSwapCache(page)) { |
| 322 | SetPageSwapCache(newpage); |
| 323 | set_page_private(newpage, page_private(page)); |
| 324 | } |
Christoph Lameter | 6c5240a | 2006-06-23 02:03:37 -0700 | [diff] [blame] | 325 | #endif |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 326 | |
| 327 | *radix_pointer = newpage; |
| 328 | __put_page(page); |
| 329 | write_unlock_irq(&mapping->tree_lock); |
| 330 | |
| 331 | return 0; |
| 332 | } |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 333 | |
| 334 | /* |
| 335 | * Copy the page to its new location |
| 336 | */ |
Christoph Lameter | e7340f7 | 2006-06-23 02:03:29 -0700 | [diff] [blame] | 337 | static void migrate_page_copy(struct page *newpage, struct page *page) |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 338 | { |
| 339 | copy_highpage(newpage, page); |
| 340 | |
| 341 | if (PageError(page)) |
| 342 | SetPageError(newpage); |
| 343 | if (PageReferenced(page)) |
| 344 | SetPageReferenced(newpage); |
| 345 | if (PageUptodate(page)) |
| 346 | SetPageUptodate(newpage); |
| 347 | if (PageActive(page)) |
| 348 | SetPageActive(newpage); |
| 349 | if (PageChecked(page)) |
| 350 | SetPageChecked(newpage); |
| 351 | if (PageMappedToDisk(page)) |
| 352 | SetPageMappedToDisk(newpage); |
| 353 | |
| 354 | if (PageDirty(page)) { |
| 355 | clear_page_dirty_for_io(page); |
| 356 | set_page_dirty(newpage); |
| 357 | } |
| 358 | |
Christoph Lameter | 6c5240a | 2006-06-23 02:03:37 -0700 | [diff] [blame] | 359 | #ifdef CONFIG_SWAP |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 360 | ClearPageSwapCache(page); |
Christoph Lameter | 6c5240a | 2006-06-23 02:03:37 -0700 | [diff] [blame] | 361 | #endif |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 362 | ClearPageActive(page); |
| 363 | ClearPagePrivate(page); |
| 364 | set_page_private(page, 0); |
| 365 | page->mapping = NULL; |
| 366 | |
| 367 | /* |
| 368 | * If any waiters have accumulated on the new page then |
| 369 | * wake them up. |
| 370 | */ |
| 371 | if (PageWriteback(newpage)) |
| 372 | end_page_writeback(newpage); |
| 373 | } |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 374 | |
Christoph Lameter | 1d8b85c | 2006-06-23 02:03:28 -0700 | [diff] [blame] | 375 | /************************************************************ |
| 376 | * Migration functions |
| 377 | ***********************************************************/ |
| 378 | |
| 379 | /* Always fail migration. Used for mappings that are not movable */ |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 380 | int fail_migrate_page(struct address_space *mapping, |
| 381 | struct page *newpage, struct page *page) |
Christoph Lameter | 1d8b85c | 2006-06-23 02:03:28 -0700 | [diff] [blame] | 382 | { |
| 383 | return -EIO; |
| 384 | } |
| 385 | EXPORT_SYMBOL(fail_migrate_page); |
| 386 | |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 387 | /* |
| 388 | * Common logic to directly migrate a single page suitable for |
| 389 | * pages that do not use PagePrivate. |
| 390 | * |
| 391 | * Pages are locked upon entry and exit. |
| 392 | */ |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 393 | int migrate_page(struct address_space *mapping, |
| 394 | struct page *newpage, struct page *page) |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 395 | { |
| 396 | int rc; |
| 397 | |
| 398 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ |
| 399 | |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 400 | rc = migrate_page_move_mapping(mapping, newpage, page); |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 401 | |
| 402 | if (rc) |
| 403 | return rc; |
| 404 | |
| 405 | migrate_page_copy(newpage, page); |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 406 | return 0; |
| 407 | } |
| 408 | EXPORT_SYMBOL(migrate_page); |
| 409 | |
| 410 | /* |
Christoph Lameter | 1d8b85c | 2006-06-23 02:03:28 -0700 | [diff] [blame] | 411 | * Migration function for pages with buffers. This function can only be used |
| 412 | * if the underlying filesystem guarantees that no other references to "page" |
| 413 | * exist. |
| 414 | */ |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 415 | int buffer_migrate_page(struct address_space *mapping, |
| 416 | struct page *newpage, struct page *page) |
Christoph Lameter | 1d8b85c | 2006-06-23 02:03:28 -0700 | [diff] [blame] | 417 | { |
Christoph Lameter | 1d8b85c | 2006-06-23 02:03:28 -0700 | [diff] [blame] | 418 | struct buffer_head *bh, *head; |
| 419 | int rc; |
| 420 | |
Christoph Lameter | 1d8b85c | 2006-06-23 02:03:28 -0700 | [diff] [blame] | 421 | if (!page_has_buffers(page)) |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 422 | return migrate_page(mapping, newpage, page); |
Christoph Lameter | 1d8b85c | 2006-06-23 02:03:28 -0700 | [diff] [blame] | 423 | |
| 424 | head = page_buffers(page); |
| 425 | |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 426 | rc = migrate_page_move_mapping(mapping, newpage, page); |
Christoph Lameter | 1d8b85c | 2006-06-23 02:03:28 -0700 | [diff] [blame] | 427 | |
| 428 | if (rc) |
| 429 | return rc; |
| 430 | |
| 431 | bh = head; |
| 432 | do { |
| 433 | get_bh(bh); |
| 434 | lock_buffer(bh); |
| 435 | bh = bh->b_this_page; |
| 436 | |
| 437 | } while (bh != head); |
| 438 | |
| 439 | ClearPagePrivate(page); |
| 440 | set_page_private(newpage, page_private(page)); |
| 441 | set_page_private(page, 0); |
| 442 | put_page(page); |
| 443 | get_page(newpage); |
| 444 | |
| 445 | bh = head; |
| 446 | do { |
| 447 | set_bh_page(bh, newpage, bh_offset(bh)); |
| 448 | bh = bh->b_this_page; |
| 449 | |
| 450 | } while (bh != head); |
| 451 | |
| 452 | SetPagePrivate(newpage); |
| 453 | |
| 454 | migrate_page_copy(newpage, page); |
| 455 | |
| 456 | bh = head; |
| 457 | do { |
| 458 | unlock_buffer(bh); |
| 459 | put_bh(bh); |
| 460 | bh = bh->b_this_page; |
| 461 | |
| 462 | } while (bh != head); |
| 463 | |
| 464 | return 0; |
| 465 | } |
| 466 | EXPORT_SYMBOL(buffer_migrate_page); |
| 467 | |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 468 | /* |
| 469 | * Writeback a page to clean the dirty state |
| 470 | */ |
| 471 | static int writeout(struct address_space *mapping, struct page *page) |
| 472 | { |
| 473 | struct writeback_control wbc = { |
| 474 | .sync_mode = WB_SYNC_NONE, |
| 475 | .nr_to_write = 1, |
| 476 | .range_start = 0, |
| 477 | .range_end = LLONG_MAX, |
| 478 | .nonblocking = 1, |
| 479 | .for_reclaim = 1 |
| 480 | }; |
| 481 | int rc; |
| 482 | |
| 483 | if (!mapping->a_ops->writepage) |
| 484 | /* No write method for the address space */ |
| 485 | return -EINVAL; |
| 486 | |
| 487 | if (!clear_page_dirty_for_io(page)) |
| 488 | /* Someone else already triggered a write */ |
| 489 | return -EAGAIN; |
| 490 | |
| 491 | /* |
| 492 | * A dirty page may imply that the underlying filesystem has |
| 493 | * the page on some queue. So the page must be clean for |
| 494 | * migration. Writeout may mean we loose the lock and the |
| 495 | * page state is no longer what we checked for earlier. |
| 496 | * At this point we know that the migration attempt cannot |
| 497 | * be successful. |
| 498 | */ |
| 499 | remove_migration_ptes(page, page); |
| 500 | |
| 501 | rc = mapping->a_ops->writepage(page, &wbc); |
| 502 | if (rc < 0) |
| 503 | /* I/O Error writing */ |
| 504 | return -EIO; |
| 505 | |
| 506 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
| 507 | /* unlocked. Relock */ |
| 508 | lock_page(page); |
| 509 | |
| 510 | return -EAGAIN; |
| 511 | } |
| 512 | |
| 513 | /* |
| 514 | * Default handling if a filesystem does not provide a migration function. |
| 515 | */ |
Christoph Lameter | 8351a6e | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 516 | static int fallback_migrate_page(struct address_space *mapping, |
| 517 | struct page *newpage, struct page *page) |
| 518 | { |
Christoph Lameter | 04e62a2 | 2006-06-23 02:03:38 -0700 | [diff] [blame] | 519 | if (PageDirty(page)) |
| 520 | return writeout(mapping, page); |
Christoph Lameter | 8351a6e | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 521 | |
| 522 | /* |
| 523 | * Buffers may be managed in a filesystem specific way. |
| 524 | * We must have no buffers or drop them. |
| 525 | */ |
| 526 | if (page_has_buffers(page) && |
| 527 | !try_to_release_page(page, GFP_KERNEL)) |
| 528 | return -EAGAIN; |
| 529 | |
| 530 | return migrate_page(mapping, newpage, page); |
| 531 | } |
| 532 | |
Christoph Lameter | 1d8b85c | 2006-06-23 02:03:28 -0700 | [diff] [blame] | 533 | /* |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 534 | * Move a page to a newly allocated page |
| 535 | * The page is locked and all ptes have been successfully removed. |
| 536 | * |
| 537 | * The new page will have replaced the old page if this function |
| 538 | * is successful. |
| 539 | */ |
| 540 | static int move_to_new_page(struct page *newpage, struct page *page) |
| 541 | { |
| 542 | struct address_space *mapping; |
| 543 | int rc; |
| 544 | |
| 545 | /* |
| 546 | * Block others from accessing the page when we get around to |
| 547 | * establishing additional references. We are the only one |
| 548 | * holding a reference to the new page at this point. |
| 549 | */ |
| 550 | if (TestSetPageLocked(newpage)) |
| 551 | BUG(); |
| 552 | |
| 553 | /* Prepare mapping for the new page.*/ |
| 554 | newpage->index = page->index; |
| 555 | newpage->mapping = page->mapping; |
| 556 | |
| 557 | mapping = page_mapping(page); |
| 558 | if (!mapping) |
| 559 | rc = migrate_page(mapping, newpage, page); |
| 560 | else if (mapping->a_ops->migratepage) |
| 561 | /* |
| 562 | * Most pages have a mapping and most filesystems |
| 563 | * should provide a migration function. Anonymous |
| 564 | * pages are part of swap space which also has its |
| 565 | * own migration function. This is the most common |
| 566 | * path for page migration. |
| 567 | */ |
| 568 | rc = mapping->a_ops->migratepage(mapping, |
| 569 | newpage, page); |
| 570 | else |
| 571 | rc = fallback_migrate_page(mapping, newpage, page); |
| 572 | |
| 573 | if (!rc) |
| 574 | remove_migration_ptes(page, newpage); |
| 575 | else |
| 576 | newpage->mapping = NULL; |
| 577 | |
| 578 | unlock_page(newpage); |
| 579 | |
| 580 | return rc; |
| 581 | } |
| 582 | |
| 583 | /* |
| 584 | * Obtain the lock on page, remove all ptes and migrate the page |
| 585 | * to the newly allocated page in newpage. |
| 586 | */ |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 587 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, |
| 588 | struct page *page, int force) |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 589 | { |
| 590 | int rc = 0; |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 591 | struct page *newpage = get_new_page(page, private); |
| 592 | |
| 593 | if (!newpage) |
| 594 | return -ENOMEM; |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 595 | |
| 596 | if (page_count(page) == 1) |
| 597 | /* page was freed from under us. So we are done. */ |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 598 | goto move_newpage; |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 599 | |
| 600 | rc = -EAGAIN; |
| 601 | if (TestSetPageLocked(page)) { |
| 602 | if (!force) |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 603 | goto move_newpage; |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 604 | lock_page(page); |
| 605 | } |
| 606 | |
| 607 | if (PageWriteback(page)) { |
| 608 | if (!force) |
| 609 | goto unlock; |
| 610 | wait_on_page_writeback(page); |
| 611 | } |
| 612 | |
| 613 | /* |
| 614 | * Establish migration ptes or remove ptes |
| 615 | */ |
| 616 | if (try_to_unmap(page, 1) != SWAP_FAIL) { |
| 617 | if (!page_mapped(page)) |
| 618 | rc = move_to_new_page(newpage, page); |
| 619 | } else |
| 620 | /* A vma has VM_LOCKED set -> permanent failure */ |
| 621 | rc = -EPERM; |
| 622 | |
| 623 | if (rc) |
| 624 | remove_migration_ptes(page, page); |
| 625 | unlock: |
| 626 | unlock_page(page); |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 627 | |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 628 | if (rc != -EAGAIN) { |
Christoph Lameter | aaa994b | 2006-06-23 02:03:52 -0700 | [diff] [blame] | 629 | /* |
| 630 | * A page that has been migrated has all references |
| 631 | * removed and will be freed. A page that has not been |
| 632 | * migrated will have kepts its references and be |
| 633 | * restored. |
| 634 | */ |
| 635 | list_del(&page->lru); |
| 636 | move_to_lru(page); |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 637 | } |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 638 | |
| 639 | move_newpage: |
| 640 | /* |
| 641 | * Move the new page to the LRU. If migration was not successful |
| 642 | * then this will free the page. |
| 643 | */ |
| 644 | move_to_lru(newpage); |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 645 | return rc; |
| 646 | } |
| 647 | |
| 648 | /* |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 649 | * migrate_pages |
| 650 | * |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 651 | * The function takes one list of pages to migrate and a function |
| 652 | * that determines from the page to be migrated and the private data |
| 653 | * the target of the move and allocates the page. |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 654 | * |
| 655 | * The function returns after 10 attempts or if no pages |
| 656 | * are movable anymore because to has become empty |
Christoph Lameter | aaa994b | 2006-06-23 02:03:52 -0700 | [diff] [blame] | 657 | * or no retryable pages exist anymore. All pages will be |
| 658 | * retruned to the LRU or freed. |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 659 | * |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 660 | * Return: Number of pages not migrated or error code. |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 661 | */ |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 662 | int migrate_pages(struct list_head *from, |
| 663 | new_page_t get_new_page, unsigned long private) |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 664 | { |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 665 | int retry = 1; |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 666 | int nr_failed = 0; |
| 667 | int pass = 0; |
| 668 | struct page *page; |
| 669 | struct page *page2; |
| 670 | int swapwrite = current->flags & PF_SWAPWRITE; |
| 671 | int rc; |
| 672 | |
| 673 | if (!swapwrite) |
| 674 | current->flags |= PF_SWAPWRITE; |
| 675 | |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 676 | for(pass = 0; pass < 10 && retry; pass++) { |
| 677 | retry = 0; |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 678 | |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 679 | list_for_each_entry_safe(page, page2, from, lru) { |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 680 | cond_resched(); |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 681 | |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 682 | rc = unmap_and_move(get_new_page, private, |
| 683 | page, pass > 2); |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 684 | |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 685 | switch(rc) { |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 686 | case -ENOMEM: |
| 687 | goto out; |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 688 | case -EAGAIN: |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 689 | retry++; |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 690 | break; |
| 691 | case 0: |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 692 | break; |
| 693 | default: |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 694 | /* Permanent failure */ |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 695 | nr_failed++; |
Christoph Lameter | e24f0b8 | 2006-06-23 02:03:51 -0700 | [diff] [blame] | 696 | break; |
Christoph Lameter | 2d1db3b | 2006-06-23 02:03:33 -0700 | [diff] [blame] | 697 | } |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 698 | } |
| 699 | } |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 700 | rc = 0; |
| 701 | out: |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 702 | if (!swapwrite) |
| 703 | current->flags &= ~PF_SWAPWRITE; |
| 704 | |
Christoph Lameter | aaa994b | 2006-06-23 02:03:52 -0700 | [diff] [blame] | 705 | putback_lru_pages(from); |
Christoph Lameter | 95a402c | 2006-06-23 02:03:53 -0700 | [diff] [blame^] | 706 | |
| 707 | if (rc) |
| 708 | return rc; |
| 709 | |
Christoph Lameter | b20a350 | 2006-03-22 00:09:12 -0800 | [diff] [blame] | 710 | return nr_failed + retry; |
| 711 | } |
| 712 | |