Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * mm/rmap.c - physical to virtual reverse mappings |
| 3 | * |
| 4 | * Copyright 2001, Rik van Riel <riel@conectiva.com.br> |
| 5 | * Released under the General Public License (GPL). |
| 6 | * |
| 7 | * Simple, low overhead reverse mapping scheme. |
| 8 | * Please try to keep this thing as modular as possible. |
| 9 | * |
| 10 | * Provides methods for unmapping each kind of mapped page: |
| 11 | * the anon methods track anonymous pages, and |
| 12 | * the file methods track pages belonging to an inode. |
| 13 | * |
| 14 | * Original design by Rik van Riel <riel@conectiva.com.br> 2001 |
| 15 | * File methods by Dave McCracken <dmccr@us.ibm.com> 2003, 2004 |
| 16 | * Anonymous methods by Andrea Arcangeli <andrea@suse.de> 2004 |
| 17 | * Contributions by Hugh Dickins <hugh@veritas.com> 2003, 2004 |
| 18 | */ |
| 19 | |
| 20 | /* |
| 21 | * Lock ordering in mm: |
| 22 | * |
| 23 | * inode->i_sem (while writing or truncating, not reading or faulting) |
| 24 | * inode->i_alloc_sem |
| 25 | * |
| 26 | * When a page fault occurs in writing from user to file, down_read |
| 27 | * of mmap_sem nests within i_sem; in sys_msync, i_sem nests within |
| 28 | * down_read of mmap_sem; i_sem and down_write of mmap_sem are never |
| 29 | * taken together; in truncation, i_sem is taken outermost. |
| 30 | * |
| 31 | * mm->mmap_sem |
| 32 | * page->flags PG_locked (lock_page) |
| 33 | * mapping->i_mmap_lock |
| 34 | * anon_vma->lock |
Hugh Dickins | b8072f0 | 2005-10-29 18:16:41 -0700 | [diff] [blame] | 35 | * mm->page_table_lock or pte_lock |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 36 | * zone->lru_lock (in mark_page_accessed) |
Hugh Dickins | 5d337b9 | 2005-09-03 15:54:41 -0700 | [diff] [blame] | 37 | * swap_lock (in swap_duplicate, swap_info_get) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 38 | * mmlist_lock (in mmput, drain_mmlist and others) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 39 | * mapping->private_lock (in __set_page_dirty_buffers) |
| 40 | * inode_lock (in set_page_dirty's __mark_inode_dirty) |
| 41 | * sb_lock (within inode_lock in fs/fs-writeback.c) |
| 42 | * mapping->tree_lock (widely used, in set_page_dirty, |
| 43 | * in arch-dependent flush_dcache_mmap_lock, |
| 44 | * within inode_lock in __sync_single_inode) |
| 45 | */ |
| 46 | |
| 47 | #include <linux/mm.h> |
| 48 | #include <linux/pagemap.h> |
| 49 | #include <linux/swap.h> |
| 50 | #include <linux/swapops.h> |
| 51 | #include <linux/slab.h> |
| 52 | #include <linux/init.h> |
| 53 | #include <linux/rmap.h> |
| 54 | #include <linux/rcupdate.h> |
| 55 | |
| 56 | #include <asm/tlbflush.h> |
| 57 | |
| 58 | //#define RMAP_DEBUG /* can be enabled only for debugging */ |
| 59 | |
| 60 | kmem_cache_t *anon_vma_cachep; |
| 61 | |
| 62 | static inline void validate_anon_vma(struct vm_area_struct *find_vma) |
| 63 | { |
| 64 | #ifdef RMAP_DEBUG |
| 65 | struct anon_vma *anon_vma = find_vma->anon_vma; |
| 66 | struct vm_area_struct *vma; |
| 67 | unsigned int mapcount = 0; |
| 68 | int found = 0; |
| 69 | |
| 70 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { |
| 71 | mapcount++; |
| 72 | BUG_ON(mapcount > 100000); |
| 73 | if (vma == find_vma) |
| 74 | found = 1; |
| 75 | } |
| 76 | BUG_ON(!found); |
| 77 | #endif |
| 78 | } |
| 79 | |
| 80 | /* This must be called under the mmap_sem. */ |
| 81 | int anon_vma_prepare(struct vm_area_struct *vma) |
| 82 | { |
| 83 | struct anon_vma *anon_vma = vma->anon_vma; |
| 84 | |
| 85 | might_sleep(); |
| 86 | if (unlikely(!anon_vma)) { |
| 87 | struct mm_struct *mm = vma->vm_mm; |
| 88 | struct anon_vma *allocated, *locked; |
| 89 | |
| 90 | anon_vma = find_mergeable_anon_vma(vma); |
| 91 | if (anon_vma) { |
| 92 | allocated = NULL; |
| 93 | locked = anon_vma; |
| 94 | spin_lock(&locked->lock); |
| 95 | } else { |
| 96 | anon_vma = anon_vma_alloc(); |
| 97 | if (unlikely(!anon_vma)) |
| 98 | return -ENOMEM; |
| 99 | allocated = anon_vma; |
| 100 | locked = NULL; |
| 101 | } |
| 102 | |
| 103 | /* page_table_lock to protect against threads */ |
| 104 | spin_lock(&mm->page_table_lock); |
| 105 | if (likely(!vma->anon_vma)) { |
| 106 | vma->anon_vma = anon_vma; |
| 107 | list_add(&vma->anon_vma_node, &anon_vma->head); |
| 108 | allocated = NULL; |
| 109 | } |
| 110 | spin_unlock(&mm->page_table_lock); |
| 111 | |
| 112 | if (locked) |
| 113 | spin_unlock(&locked->lock); |
| 114 | if (unlikely(allocated)) |
| 115 | anon_vma_free(allocated); |
| 116 | } |
| 117 | return 0; |
| 118 | } |
| 119 | |
| 120 | void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next) |
| 121 | { |
| 122 | BUG_ON(vma->anon_vma != next->anon_vma); |
| 123 | list_del(&next->anon_vma_node); |
| 124 | } |
| 125 | |
| 126 | void __anon_vma_link(struct vm_area_struct *vma) |
| 127 | { |
| 128 | struct anon_vma *anon_vma = vma->anon_vma; |
| 129 | |
| 130 | if (anon_vma) { |
| 131 | list_add(&vma->anon_vma_node, &anon_vma->head); |
| 132 | validate_anon_vma(vma); |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | void anon_vma_link(struct vm_area_struct *vma) |
| 137 | { |
| 138 | struct anon_vma *anon_vma = vma->anon_vma; |
| 139 | |
| 140 | if (anon_vma) { |
| 141 | spin_lock(&anon_vma->lock); |
| 142 | list_add(&vma->anon_vma_node, &anon_vma->head); |
| 143 | validate_anon_vma(vma); |
| 144 | spin_unlock(&anon_vma->lock); |
| 145 | } |
| 146 | } |
| 147 | |
| 148 | void anon_vma_unlink(struct vm_area_struct *vma) |
| 149 | { |
| 150 | struct anon_vma *anon_vma = vma->anon_vma; |
| 151 | int empty; |
| 152 | |
| 153 | if (!anon_vma) |
| 154 | return; |
| 155 | |
| 156 | spin_lock(&anon_vma->lock); |
| 157 | validate_anon_vma(vma); |
| 158 | list_del(&vma->anon_vma_node); |
| 159 | |
| 160 | /* We must garbage collect the anon_vma if it's empty */ |
| 161 | empty = list_empty(&anon_vma->head); |
| 162 | spin_unlock(&anon_vma->lock); |
| 163 | |
| 164 | if (empty) |
| 165 | anon_vma_free(anon_vma); |
| 166 | } |
| 167 | |
| 168 | static void anon_vma_ctor(void *data, kmem_cache_t *cachep, unsigned long flags) |
| 169 | { |
| 170 | if ((flags & (SLAB_CTOR_VERIFY|SLAB_CTOR_CONSTRUCTOR)) == |
| 171 | SLAB_CTOR_CONSTRUCTOR) { |
| 172 | struct anon_vma *anon_vma = data; |
| 173 | |
| 174 | spin_lock_init(&anon_vma->lock); |
| 175 | INIT_LIST_HEAD(&anon_vma->head); |
| 176 | } |
| 177 | } |
| 178 | |
| 179 | void __init anon_vma_init(void) |
| 180 | { |
| 181 | anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), |
| 182 | 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor, NULL); |
| 183 | } |
| 184 | |
| 185 | /* |
| 186 | * Getting a lock on a stable anon_vma from a page off the LRU is |
| 187 | * tricky: page_lock_anon_vma rely on RCU to guard against the races. |
| 188 | */ |
| 189 | static struct anon_vma *page_lock_anon_vma(struct page *page) |
| 190 | { |
| 191 | struct anon_vma *anon_vma = NULL; |
| 192 | unsigned long anon_mapping; |
| 193 | |
| 194 | rcu_read_lock(); |
| 195 | anon_mapping = (unsigned long) page->mapping; |
| 196 | if (!(anon_mapping & PAGE_MAPPING_ANON)) |
| 197 | goto out; |
| 198 | if (!page_mapped(page)) |
| 199 | goto out; |
| 200 | |
| 201 | anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); |
| 202 | spin_lock(&anon_vma->lock); |
| 203 | out: |
| 204 | rcu_read_unlock(); |
| 205 | return anon_vma; |
| 206 | } |
| 207 | |
| 208 | /* |
| 209 | * At what user virtual address is page expected in vma? |
| 210 | */ |
| 211 | static inline unsigned long |
| 212 | vma_address(struct page *page, struct vm_area_struct *vma) |
| 213 | { |
| 214 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 215 | unsigned long address; |
| 216 | |
| 217 | address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); |
| 218 | if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { |
| 219 | /* page should be within any vma from prio_tree_next */ |
| 220 | BUG_ON(!PageAnon(page)); |
| 221 | return -EFAULT; |
| 222 | } |
| 223 | return address; |
| 224 | } |
| 225 | |
| 226 | /* |
| 227 | * At what user virtual address is page expected in vma? checking that the |
Hugh Dickins | ee498ed | 2005-11-21 21:32:18 -0800 | [diff] [blame] | 228 | * page matches the vma: currently only used on anon pages, by unuse_vma; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 229 | */ |
| 230 | unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) |
| 231 | { |
| 232 | if (PageAnon(page)) { |
| 233 | if ((void *)vma->anon_vma != |
| 234 | (void *)page->mapping - PAGE_MAPPING_ANON) |
| 235 | return -EFAULT; |
| 236 | } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { |
Hugh Dickins | ee498ed | 2005-11-21 21:32:18 -0800 | [diff] [blame] | 237 | if (!vma->vm_file || |
| 238 | vma->vm_file->f_mapping != page->mapping) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 239 | return -EFAULT; |
| 240 | } else |
| 241 | return -EFAULT; |
| 242 | return vma_address(page, vma); |
| 243 | } |
| 244 | |
| 245 | /* |
Nikita Danilov | 81b4082 | 2005-05-01 08:58:36 -0700 | [diff] [blame] | 246 | * Check that @page is mapped at @address into @mm. |
| 247 | * |
Hugh Dickins | b8072f0 | 2005-10-29 18:16:41 -0700 | [diff] [blame] | 248 | * On success returns with pte mapped and locked. |
Nikita Danilov | 81b4082 | 2005-05-01 08:58:36 -0700 | [diff] [blame] | 249 | */ |
Carsten Otte | ceffc07 | 2005-06-23 22:05:25 -0700 | [diff] [blame] | 250 | pte_t *page_check_address(struct page *page, struct mm_struct *mm, |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 251 | unsigned long address, spinlock_t **ptlp) |
Nikita Danilov | 81b4082 | 2005-05-01 08:58:36 -0700 | [diff] [blame] | 252 | { |
| 253 | pgd_t *pgd; |
| 254 | pud_t *pud; |
| 255 | pmd_t *pmd; |
| 256 | pte_t *pte; |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 257 | spinlock_t *ptl; |
Nikita Danilov | 81b4082 | 2005-05-01 08:58:36 -0700 | [diff] [blame] | 258 | |
Nikita Danilov | 81b4082 | 2005-05-01 08:58:36 -0700 | [diff] [blame] | 259 | pgd = pgd_offset(mm, address); |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 260 | if (!pgd_present(*pgd)) |
| 261 | return NULL; |
| 262 | |
| 263 | pud = pud_offset(pgd, address); |
| 264 | if (!pud_present(*pud)) |
| 265 | return NULL; |
| 266 | |
| 267 | pmd = pmd_offset(pud, address); |
| 268 | if (!pmd_present(*pmd)) |
| 269 | return NULL; |
| 270 | |
| 271 | pte = pte_offset_map(pmd, address); |
| 272 | /* Make a quick check before getting the lock */ |
| 273 | if (!pte_present(*pte)) { |
| 274 | pte_unmap(pte); |
| 275 | return NULL; |
Nikita Danilov | 81b4082 | 2005-05-01 08:58:36 -0700 | [diff] [blame] | 276 | } |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 277 | |
Hugh Dickins | 4c21e2f | 2005-10-29 18:16:40 -0700 | [diff] [blame] | 278 | ptl = pte_lockptr(mm, pmd); |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 279 | spin_lock(ptl); |
| 280 | if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { |
| 281 | *ptlp = ptl; |
| 282 | return pte; |
| 283 | } |
| 284 | pte_unmap_unlock(pte, ptl); |
| 285 | return NULL; |
Nikita Danilov | 81b4082 | 2005-05-01 08:58:36 -0700 | [diff] [blame] | 286 | } |
| 287 | |
| 288 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 289 | * Subfunctions of page_referenced: page_referenced_one called |
| 290 | * repeatedly from either page_referenced_anon or page_referenced_file. |
| 291 | */ |
| 292 | static int page_referenced_one(struct page *page, |
| 293 | struct vm_area_struct *vma, unsigned int *mapcount, int ignore_token) |
| 294 | { |
| 295 | struct mm_struct *mm = vma->vm_mm; |
| 296 | unsigned long address; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 297 | pte_t *pte; |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 298 | spinlock_t *ptl; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 299 | int referenced = 0; |
| 300 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 301 | address = vma_address(page, vma); |
| 302 | if (address == -EFAULT) |
| 303 | goto out; |
| 304 | |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 305 | pte = page_check_address(page, mm, address, &ptl); |
| 306 | if (!pte) |
| 307 | goto out; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 308 | |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 309 | if (ptep_clear_flush_young(vma, address, pte)) |
| 310 | referenced++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 311 | |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 312 | /* Pretend the page is referenced if the task has the |
| 313 | swap token and is in the middle of a page fault. */ |
| 314 | if (mm != current->mm && !ignore_token && has_swap_token(mm) && |
| 315 | rwsem_is_locked(&mm->mmap_sem)) |
| 316 | referenced++; |
| 317 | |
| 318 | (*mapcount)--; |
| 319 | pte_unmap_unlock(pte, ptl); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 320 | out: |
| 321 | return referenced; |
| 322 | } |
| 323 | |
| 324 | static int page_referenced_anon(struct page *page, int ignore_token) |
| 325 | { |
| 326 | unsigned int mapcount; |
| 327 | struct anon_vma *anon_vma; |
| 328 | struct vm_area_struct *vma; |
| 329 | int referenced = 0; |
| 330 | |
| 331 | anon_vma = page_lock_anon_vma(page); |
| 332 | if (!anon_vma) |
| 333 | return referenced; |
| 334 | |
| 335 | mapcount = page_mapcount(page); |
| 336 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { |
| 337 | referenced += page_referenced_one(page, vma, &mapcount, |
| 338 | ignore_token); |
| 339 | if (!mapcount) |
| 340 | break; |
| 341 | } |
| 342 | spin_unlock(&anon_vma->lock); |
| 343 | return referenced; |
| 344 | } |
| 345 | |
| 346 | /** |
| 347 | * page_referenced_file - referenced check for object-based rmap |
| 348 | * @page: the page we're checking references on. |
| 349 | * |
| 350 | * For an object-based mapped page, find all the places it is mapped and |
| 351 | * check/clear the referenced flag. This is done by following the page->mapping |
| 352 | * pointer, then walking the chain of vmas it holds. It returns the number |
| 353 | * of references it found. |
| 354 | * |
| 355 | * This function is only called from page_referenced for object-based pages. |
| 356 | */ |
| 357 | static int page_referenced_file(struct page *page, int ignore_token) |
| 358 | { |
| 359 | unsigned int mapcount; |
| 360 | struct address_space *mapping = page->mapping; |
| 361 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 362 | struct vm_area_struct *vma; |
| 363 | struct prio_tree_iter iter; |
| 364 | int referenced = 0; |
| 365 | |
| 366 | /* |
| 367 | * The caller's checks on page->mapping and !PageAnon have made |
| 368 | * sure that this is a file page: the check for page->mapping |
| 369 | * excludes the case just before it gets set on an anon page. |
| 370 | */ |
| 371 | BUG_ON(PageAnon(page)); |
| 372 | |
| 373 | /* |
| 374 | * The page lock not only makes sure that page->mapping cannot |
| 375 | * suddenly be NULLified by truncation, it makes sure that the |
| 376 | * structure at mapping cannot be freed and reused yet, |
| 377 | * so we can safely take mapping->i_mmap_lock. |
| 378 | */ |
| 379 | BUG_ON(!PageLocked(page)); |
| 380 | |
| 381 | spin_lock(&mapping->i_mmap_lock); |
| 382 | |
| 383 | /* |
| 384 | * i_mmap_lock does not stabilize mapcount at all, but mapcount |
| 385 | * is more likely to be accurate if we note it after spinning. |
| 386 | */ |
| 387 | mapcount = page_mapcount(page); |
| 388 | |
| 389 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { |
| 390 | if ((vma->vm_flags & (VM_LOCKED|VM_MAYSHARE)) |
| 391 | == (VM_LOCKED|VM_MAYSHARE)) { |
| 392 | referenced++; |
| 393 | break; |
| 394 | } |
| 395 | referenced += page_referenced_one(page, vma, &mapcount, |
| 396 | ignore_token); |
| 397 | if (!mapcount) |
| 398 | break; |
| 399 | } |
| 400 | |
| 401 | spin_unlock(&mapping->i_mmap_lock); |
| 402 | return referenced; |
| 403 | } |
| 404 | |
| 405 | /** |
| 406 | * page_referenced - test if the page was referenced |
| 407 | * @page: the page to test |
| 408 | * @is_locked: caller holds lock on the page |
| 409 | * |
| 410 | * Quick test_and_clear_referenced for all mappings to a page, |
| 411 | * returns the number of ptes which referenced the page. |
| 412 | */ |
| 413 | int page_referenced(struct page *page, int is_locked, int ignore_token) |
| 414 | { |
| 415 | int referenced = 0; |
| 416 | |
| 417 | if (!swap_token_default_timeout) |
| 418 | ignore_token = 1; |
| 419 | |
| 420 | if (page_test_and_clear_young(page)) |
| 421 | referenced++; |
| 422 | |
| 423 | if (TestClearPageReferenced(page)) |
| 424 | referenced++; |
| 425 | |
| 426 | if (page_mapped(page) && page->mapping) { |
| 427 | if (PageAnon(page)) |
| 428 | referenced += page_referenced_anon(page, ignore_token); |
| 429 | else if (is_locked) |
| 430 | referenced += page_referenced_file(page, ignore_token); |
| 431 | else if (TestSetPageLocked(page)) |
| 432 | referenced++; |
| 433 | else { |
| 434 | if (page->mapping) |
| 435 | referenced += page_referenced_file(page, |
| 436 | ignore_token); |
| 437 | unlock_page(page); |
| 438 | } |
| 439 | } |
| 440 | return referenced; |
| 441 | } |
| 442 | |
| 443 | /** |
| 444 | * page_add_anon_rmap - add pte mapping to an anonymous page |
| 445 | * @page: the page to add the mapping to |
| 446 | * @vma: the vm area in which the mapping is added |
| 447 | * @address: the user virtual address mapped |
| 448 | * |
Hugh Dickins | b8072f0 | 2005-10-29 18:16:41 -0700 | [diff] [blame] | 449 | * The caller needs to hold the pte lock. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 450 | */ |
| 451 | void page_add_anon_rmap(struct page *page, |
| 452 | struct vm_area_struct *vma, unsigned long address) |
| 453 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 454 | if (atomic_inc_and_test(&page->_mapcount)) { |
Nick Piggin | 2822c1a | 2005-09-03 15:54:47 -0700 | [diff] [blame] | 455 | struct anon_vma *anon_vma = vma->anon_vma; |
Nick Piggin | 2822c1a | 2005-09-03 15:54:47 -0700 | [diff] [blame] | 456 | |
| 457 | BUG_ON(!anon_vma); |
| 458 | anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 459 | page->mapping = (struct address_space *) anon_vma; |
Nick Piggin | 2822c1a | 2005-09-03 15:54:47 -0700 | [diff] [blame] | 460 | |
Nick Piggin | 4d7670e | 2005-09-03 15:54:48 -0700 | [diff] [blame] | 461 | page->index = linear_page_index(vma, address); |
Nick Piggin | 2822c1a | 2005-09-03 15:54:47 -0700 | [diff] [blame] | 462 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 463 | inc_page_state(nr_mapped); |
| 464 | } |
| 465 | /* else checking page index and mapping is racy */ |
| 466 | } |
| 467 | |
| 468 | /** |
| 469 | * page_add_file_rmap - add pte mapping to a file page |
| 470 | * @page: the page to add the mapping to |
| 471 | * |
Hugh Dickins | b8072f0 | 2005-10-29 18:16:41 -0700 | [diff] [blame] | 472 | * The caller needs to hold the pte lock. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 473 | */ |
| 474 | void page_add_file_rmap(struct page *page) |
| 475 | { |
| 476 | BUG_ON(PageAnon(page)); |
Nick Piggin | b581003 | 2005-10-29 18:16:12 -0700 | [diff] [blame] | 477 | BUG_ON(!pfn_valid(page_to_pfn(page))); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 478 | |
| 479 | if (atomic_inc_and_test(&page->_mapcount)) |
| 480 | inc_page_state(nr_mapped); |
| 481 | } |
| 482 | |
| 483 | /** |
| 484 | * page_remove_rmap - take down pte mapping from a page |
| 485 | * @page: page to remove mapping from |
| 486 | * |
Hugh Dickins | b8072f0 | 2005-10-29 18:16:41 -0700 | [diff] [blame] | 487 | * The caller needs to hold the pte lock. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 488 | */ |
| 489 | void page_remove_rmap(struct page *page) |
| 490 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 491 | if (atomic_add_negative(-1, &page->_mapcount)) { |
| 492 | BUG_ON(page_mapcount(page) < 0); |
| 493 | /* |
| 494 | * It would be tidy to reset the PageAnon mapping here, |
| 495 | * but that might overwrite a racing page_add_anon_rmap |
| 496 | * which increments mapcount after us but sets mapping |
| 497 | * before us: so leave the reset to free_hot_cold_page, |
| 498 | * and remember that it's only reliable while mapped. |
| 499 | * Leaving it set also helps swapoff to reinstate ptes |
| 500 | * faster for those pages still in swapcache. |
| 501 | */ |
| 502 | if (page_test_and_clear_dirty(page)) |
| 503 | set_page_dirty(page); |
| 504 | dec_page_state(nr_mapped); |
| 505 | } |
| 506 | } |
| 507 | |
| 508 | /* |
| 509 | * Subfunctions of try_to_unmap: try_to_unmap_one called |
| 510 | * repeatedly from either try_to_unmap_anon or try_to_unmap_file. |
| 511 | */ |
| 512 | static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma) |
| 513 | { |
| 514 | struct mm_struct *mm = vma->vm_mm; |
| 515 | unsigned long address; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 516 | pte_t *pte; |
| 517 | pte_t pteval; |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 518 | spinlock_t *ptl; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 519 | int ret = SWAP_AGAIN; |
| 520 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 521 | address = vma_address(page, vma); |
| 522 | if (address == -EFAULT) |
| 523 | goto out; |
| 524 | |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 525 | pte = page_check_address(page, mm, address, &ptl); |
| 526 | if (!pte) |
Nikita Danilov | 81b4082 | 2005-05-01 08:58:36 -0700 | [diff] [blame] | 527 | goto out; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 528 | |
| 529 | /* |
| 530 | * If the page is mlock()d, we cannot swap it out. |
| 531 | * If it's recently referenced (perhaps page_referenced |
| 532 | * skipped over this mm) then we should reactivate it. |
| 533 | */ |
Hugh Dickins | 101d2be | 2005-11-21 21:32:16 -0800 | [diff] [blame] | 534 | if ((vma->vm_flags & VM_LOCKED) || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 535 | ptep_clear_flush_young(vma, address, pte)) { |
| 536 | ret = SWAP_FAIL; |
| 537 | goto out_unmap; |
| 538 | } |
| 539 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 540 | /* Nuke the page table entry. */ |
| 541 | flush_cache_page(vma, address, page_to_pfn(page)); |
| 542 | pteval = ptep_clear_flush(vma, address, pte); |
| 543 | |
| 544 | /* Move the dirty bit to the physical page now the pte is gone. */ |
| 545 | if (pte_dirty(pteval)) |
| 546 | set_page_dirty(page); |
| 547 | |
Hugh Dickins | 365e9c87 | 2005-10-29 18:16:18 -0700 | [diff] [blame] | 548 | /* Update high watermark before we lower rss */ |
| 549 | update_hiwater_rss(mm); |
| 550 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 551 | if (PageAnon(page)) { |
Hugh Dickins | 4c21e2f | 2005-10-29 18:16:40 -0700 | [diff] [blame] | 552 | swp_entry_t entry = { .val = page_private(page) }; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 553 | /* |
| 554 | * Store the swap location in the pte. |
| 555 | * See handle_pte_fault() ... |
| 556 | */ |
| 557 | BUG_ON(!PageSwapCache(page)); |
| 558 | swap_duplicate(entry); |
| 559 | if (list_empty(&mm->mmlist)) { |
| 560 | spin_lock(&mmlist_lock); |
Hugh Dickins | f412ac0 | 2005-10-29 18:16:41 -0700 | [diff] [blame] | 561 | if (list_empty(&mm->mmlist)) |
| 562 | list_add(&mm->mmlist, &init_mm.mmlist); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 563 | spin_unlock(&mmlist_lock); |
| 564 | } |
| 565 | set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); |
| 566 | BUG_ON(pte_file(*pte)); |
| 567 | dec_mm_counter(mm, anon_rss); |
Hugh Dickins | 4294621 | 2005-10-29 18:16:05 -0700 | [diff] [blame] | 568 | } else |
| 569 | dec_mm_counter(mm, file_rss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 570 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 571 | page_remove_rmap(page); |
| 572 | page_cache_release(page); |
| 573 | |
| 574 | out_unmap: |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 575 | pte_unmap_unlock(pte, ptl); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 576 | out: |
| 577 | return ret; |
| 578 | } |
| 579 | |
| 580 | /* |
| 581 | * objrmap doesn't work for nonlinear VMAs because the assumption that |
| 582 | * offset-into-file correlates with offset-into-virtual-addresses does not hold. |
| 583 | * Consequently, given a particular page and its ->index, we cannot locate the |
| 584 | * ptes which are mapping that page without an exhaustive linear search. |
| 585 | * |
| 586 | * So what this code does is a mini "virtual scan" of each nonlinear VMA which |
| 587 | * maps the file to which the target page belongs. The ->vm_private_data field |
| 588 | * holds the current cursor into that scan. Successive searches will circulate |
| 589 | * around the vma's virtual address space. |
| 590 | * |
| 591 | * So as more replacement pressure is applied to the pages in a nonlinear VMA, |
| 592 | * more scanning pressure is placed against them as well. Eventually pages |
| 593 | * will become fully unmapped and are eligible for eviction. |
| 594 | * |
| 595 | * For very sparsely populated VMAs this is a little inefficient - chances are |
| 596 | * there there won't be many ptes located within the scan cluster. In this case |
| 597 | * maybe we could scan further - to the end of the pte page, perhaps. |
| 598 | */ |
| 599 | #define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) |
| 600 | #define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) |
| 601 | |
| 602 | static void try_to_unmap_cluster(unsigned long cursor, |
| 603 | unsigned int *mapcount, struct vm_area_struct *vma) |
| 604 | { |
| 605 | struct mm_struct *mm = vma->vm_mm; |
| 606 | pgd_t *pgd; |
| 607 | pud_t *pud; |
| 608 | pmd_t *pmd; |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 609 | pte_t *pte; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 610 | pte_t pteval; |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 611 | spinlock_t *ptl; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 612 | struct page *page; |
| 613 | unsigned long address; |
| 614 | unsigned long end; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 615 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 616 | address = (vma->vm_start + cursor) & CLUSTER_MASK; |
| 617 | end = address + CLUSTER_SIZE; |
| 618 | if (address < vma->vm_start) |
| 619 | address = vma->vm_start; |
| 620 | if (end > vma->vm_end) |
| 621 | end = vma->vm_end; |
| 622 | |
| 623 | pgd = pgd_offset(mm, address); |
| 624 | if (!pgd_present(*pgd)) |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 625 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 626 | |
| 627 | pud = pud_offset(pgd, address); |
| 628 | if (!pud_present(*pud)) |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 629 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 630 | |
| 631 | pmd = pmd_offset(pud, address); |
| 632 | if (!pmd_present(*pmd)) |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 633 | return; |
| 634 | |
| 635 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 636 | |
Hugh Dickins | 365e9c87 | 2005-10-29 18:16:18 -0700 | [diff] [blame] | 637 | /* Update high watermark before we lower rss */ |
| 638 | update_hiwater_rss(mm); |
| 639 | |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 640 | for (; address < end; pte++, address += PAGE_SIZE) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 641 | if (!pte_present(*pte)) |
| 642 | continue; |
Linus Torvalds | 6aab341 | 2005-11-28 14:34:23 -0800 | [diff] [blame^] | 643 | page = vm_normal_page(vma, address, *pte); |
| 644 | BUG_ON(!page || PageAnon(page)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 645 | |
| 646 | if (ptep_clear_flush_young(vma, address, pte)) |
| 647 | continue; |
| 648 | |
| 649 | /* Nuke the page table entry. */ |
| 650 | flush_cache_page(vma, address, pfn); |
| 651 | pteval = ptep_clear_flush(vma, address, pte); |
| 652 | |
| 653 | /* If nonlinear, store the file page offset in the pte. */ |
| 654 | if (page->index != linear_page_index(vma, address)) |
| 655 | set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); |
| 656 | |
| 657 | /* Move the dirty bit to the physical page now the pte is gone. */ |
| 658 | if (pte_dirty(pteval)) |
| 659 | set_page_dirty(page); |
| 660 | |
| 661 | page_remove_rmap(page); |
| 662 | page_cache_release(page); |
Hugh Dickins | 4294621 | 2005-10-29 18:16:05 -0700 | [diff] [blame] | 663 | dec_mm_counter(mm, file_rss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 664 | (*mapcount)--; |
| 665 | } |
Hugh Dickins | c071880 | 2005-10-29 18:16:31 -0700 | [diff] [blame] | 666 | pte_unmap_unlock(pte - 1, ptl); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 667 | } |
| 668 | |
| 669 | static int try_to_unmap_anon(struct page *page) |
| 670 | { |
| 671 | struct anon_vma *anon_vma; |
| 672 | struct vm_area_struct *vma; |
| 673 | int ret = SWAP_AGAIN; |
| 674 | |
| 675 | anon_vma = page_lock_anon_vma(page); |
| 676 | if (!anon_vma) |
| 677 | return ret; |
| 678 | |
| 679 | list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { |
| 680 | ret = try_to_unmap_one(page, vma); |
| 681 | if (ret == SWAP_FAIL || !page_mapped(page)) |
| 682 | break; |
| 683 | } |
| 684 | spin_unlock(&anon_vma->lock); |
| 685 | return ret; |
| 686 | } |
| 687 | |
| 688 | /** |
| 689 | * try_to_unmap_file - unmap file page using the object-based rmap method |
| 690 | * @page: the page to unmap |
| 691 | * |
| 692 | * Find all the mappings of a page using the mapping pointer and the vma chains |
| 693 | * contained in the address_space struct it points to. |
| 694 | * |
| 695 | * This function is only called from try_to_unmap for object-based pages. |
| 696 | */ |
| 697 | static int try_to_unmap_file(struct page *page) |
| 698 | { |
| 699 | struct address_space *mapping = page->mapping; |
| 700 | pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); |
| 701 | struct vm_area_struct *vma; |
| 702 | struct prio_tree_iter iter; |
| 703 | int ret = SWAP_AGAIN; |
| 704 | unsigned long cursor; |
| 705 | unsigned long max_nl_cursor = 0; |
| 706 | unsigned long max_nl_size = 0; |
| 707 | unsigned int mapcount; |
| 708 | |
| 709 | spin_lock(&mapping->i_mmap_lock); |
| 710 | vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { |
| 711 | ret = try_to_unmap_one(page, vma); |
| 712 | if (ret == SWAP_FAIL || !page_mapped(page)) |
| 713 | goto out; |
| 714 | } |
| 715 | |
| 716 | if (list_empty(&mapping->i_mmap_nonlinear)) |
| 717 | goto out; |
| 718 | |
| 719 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, |
| 720 | shared.vm_set.list) { |
Hugh Dickins | 101d2be | 2005-11-21 21:32:16 -0800 | [diff] [blame] | 721 | if (vma->vm_flags & VM_LOCKED) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 722 | continue; |
| 723 | cursor = (unsigned long) vma->vm_private_data; |
| 724 | if (cursor > max_nl_cursor) |
| 725 | max_nl_cursor = cursor; |
| 726 | cursor = vma->vm_end - vma->vm_start; |
| 727 | if (cursor > max_nl_size) |
| 728 | max_nl_size = cursor; |
| 729 | } |
| 730 | |
| 731 | if (max_nl_size == 0) { /* any nonlinears locked or reserved */ |
| 732 | ret = SWAP_FAIL; |
| 733 | goto out; |
| 734 | } |
| 735 | |
| 736 | /* |
| 737 | * We don't try to search for this page in the nonlinear vmas, |
| 738 | * and page_referenced wouldn't have found it anyway. Instead |
| 739 | * just walk the nonlinear vmas trying to age and unmap some. |
| 740 | * The mapcount of the page we came in with is irrelevant, |
| 741 | * but even so use it as a guide to how hard we should try? |
| 742 | */ |
| 743 | mapcount = page_mapcount(page); |
| 744 | if (!mapcount) |
| 745 | goto out; |
| 746 | cond_resched_lock(&mapping->i_mmap_lock); |
| 747 | |
| 748 | max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; |
| 749 | if (max_nl_cursor == 0) |
| 750 | max_nl_cursor = CLUSTER_SIZE; |
| 751 | |
| 752 | do { |
| 753 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, |
| 754 | shared.vm_set.list) { |
Hugh Dickins | 101d2be | 2005-11-21 21:32:16 -0800 | [diff] [blame] | 755 | if (vma->vm_flags & VM_LOCKED) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 756 | continue; |
| 757 | cursor = (unsigned long) vma->vm_private_data; |
Hugh Dickins | 839b968 | 2005-09-03 15:54:43 -0700 | [diff] [blame] | 758 | while ( cursor < max_nl_cursor && |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 759 | cursor < vma->vm_end - vma->vm_start) { |
| 760 | try_to_unmap_cluster(cursor, &mapcount, vma); |
| 761 | cursor += CLUSTER_SIZE; |
| 762 | vma->vm_private_data = (void *) cursor; |
| 763 | if ((int)mapcount <= 0) |
| 764 | goto out; |
| 765 | } |
| 766 | vma->vm_private_data = (void *) max_nl_cursor; |
| 767 | } |
| 768 | cond_resched_lock(&mapping->i_mmap_lock); |
| 769 | max_nl_cursor += CLUSTER_SIZE; |
| 770 | } while (max_nl_cursor <= max_nl_size); |
| 771 | |
| 772 | /* |
| 773 | * Don't loop forever (perhaps all the remaining pages are |
| 774 | * in locked vmas). Reset cursor on all unreserved nonlinear |
| 775 | * vmas, now forgetting on which ones it had fallen behind. |
| 776 | */ |
Hugh Dickins | 101d2be | 2005-11-21 21:32:16 -0800 | [diff] [blame] | 777 | list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) |
| 778 | vma->vm_private_data = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 779 | out: |
| 780 | spin_unlock(&mapping->i_mmap_lock); |
| 781 | return ret; |
| 782 | } |
| 783 | |
| 784 | /** |
| 785 | * try_to_unmap - try to remove all page table mappings to a page |
| 786 | * @page: the page to get unmapped |
| 787 | * |
| 788 | * Tries to remove all the page table entries which are mapping this |
| 789 | * page, used in the pageout path. Caller must hold the page lock. |
| 790 | * Return values are: |
| 791 | * |
| 792 | * SWAP_SUCCESS - we succeeded in removing all mappings |
| 793 | * SWAP_AGAIN - we missed a mapping, try again later |
| 794 | * SWAP_FAIL - the page is unswappable |
| 795 | */ |
| 796 | int try_to_unmap(struct page *page) |
| 797 | { |
| 798 | int ret; |
| 799 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 800 | BUG_ON(!PageLocked(page)); |
| 801 | |
| 802 | if (PageAnon(page)) |
| 803 | ret = try_to_unmap_anon(page); |
| 804 | else |
| 805 | ret = try_to_unmap_file(page); |
| 806 | |
| 807 | if (!page_mapped(page)) |
| 808 | ret = SWAP_SUCCESS; |
| 809 | return ret; |
| 810 | } |
Nikita Danilov | 81b4082 | 2005-05-01 08:58:36 -0700 | [diff] [blame] | 811 | |