rmap.c revision 3dc147414ccad81dc33edb80774b1fed12a38c08
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_mutex (while writing or truncating, not reading or faulting) 24 * inode->i_alloc_sem (vmtruncate_range) 25 * mm->mmap_sem 26 * page->flags PG_locked (lock_page) 27 * mapping->i_mmap_lock 28 * anon_vma->lock 29 * mm->page_table_lock or pte_lock 30 * zone->lru_lock (in mark_page_accessed, isolate_lru_page) 31 * swap_lock (in swap_duplicate, swap_info_get) 32 * mmlist_lock (in mmput, drain_mmlist and others) 33 * mapping->private_lock (in __set_page_dirty_buffers) 34 * inode_lock (in set_page_dirty's __mark_inode_dirty) 35 * sb_lock (within inode_lock in fs/fs-writeback.c) 36 * mapping->tree_lock (widely used, in set_page_dirty, 37 * in arch-dependent flush_dcache_mmap_lock, 38 * within inode_lock in __sync_single_inode) 39 */ 40 41#include <linux/mm.h> 42#include <linux/pagemap.h> 43#include <linux/swap.h> 44#include <linux/swapops.h> 45#include <linux/slab.h> 46#include <linux/init.h> 47#include <linux/rmap.h> 48#include <linux/rcupdate.h> 49#include <linux/module.h> 50#include <linux/memcontrol.h> 51#include <linux/mmu_notifier.h> 52#include <linux/migrate.h> 53 54#include <asm/tlbflush.h> 55 56#include "internal.h" 57 58static struct kmem_cache *anon_vma_cachep; 59 60static inline struct anon_vma *anon_vma_alloc(void) 61{ 62 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); 63} 64 65static inline void anon_vma_free(struct anon_vma *anon_vma) 66{ 67 kmem_cache_free(anon_vma_cachep, anon_vma); 68} 69 70/** 71 * anon_vma_prepare - attach an anon_vma to a memory region 72 * @vma: the memory region in question 73 * 74 * This makes sure the memory mapping described by 'vma' has 75 * an 'anon_vma' attached to it, so that we can associate the 76 * anonymous pages mapped into it with that anon_vma. 77 * 78 * The common case will be that we already have one, but if 79 * if not we either need to find an adjacent mapping that we 80 * can re-use the anon_vma from (very common when the only 81 * reason for splitting a vma has been mprotect()), or we 82 * allocate a new one. 83 * 84 * Anon-vma allocations are very subtle, because we may have 85 * optimistically looked up an anon_vma in page_lock_anon_vma() 86 * and that may actually touch the spinlock even in the newly 87 * allocated vma (it depends on RCU to make sure that the 88 * anon_vma isn't actually destroyed). 89 * 90 * As a result, we need to do proper anon_vma locking even 91 * for the new allocation. At the same time, we do not want 92 * to do any locking for the common case of already having 93 * an anon_vma. 94 * 95 * This must be called with the mmap_sem held for reading. 96 */ 97int anon_vma_prepare(struct vm_area_struct *vma) 98{ 99 struct anon_vma *anon_vma = vma->anon_vma; 100 101 might_sleep(); 102 if (unlikely(!anon_vma)) { 103 struct mm_struct *mm = vma->vm_mm; 104 struct anon_vma *allocated; 105 106 anon_vma = find_mergeable_anon_vma(vma); 107 allocated = NULL; 108 if (!anon_vma) { 109 anon_vma = anon_vma_alloc(); 110 if (unlikely(!anon_vma)) 111 return -ENOMEM; 112 allocated = anon_vma; 113 } 114 spin_lock(&anon_vma->lock); 115 116 /* page_table_lock to protect against threads */ 117 spin_lock(&mm->page_table_lock); 118 if (likely(!vma->anon_vma)) { 119 vma->anon_vma = anon_vma; 120 list_add_tail(&vma->anon_vma_node, &anon_vma->head); 121 allocated = NULL; 122 } 123 spin_unlock(&mm->page_table_lock); 124 125 spin_unlock(&anon_vma->lock); 126 if (unlikely(allocated)) 127 anon_vma_free(allocated); 128 } 129 return 0; 130} 131 132void __anon_vma_merge(struct vm_area_struct *vma, struct vm_area_struct *next) 133{ 134 BUG_ON(vma->anon_vma != next->anon_vma); 135 list_del(&next->anon_vma_node); 136} 137 138void __anon_vma_link(struct vm_area_struct *vma) 139{ 140 struct anon_vma *anon_vma = vma->anon_vma; 141 142 if (anon_vma) 143 list_add_tail(&vma->anon_vma_node, &anon_vma->head); 144} 145 146void anon_vma_link(struct vm_area_struct *vma) 147{ 148 struct anon_vma *anon_vma = vma->anon_vma; 149 150 if (anon_vma) { 151 spin_lock(&anon_vma->lock); 152 list_add_tail(&vma->anon_vma_node, &anon_vma->head); 153 spin_unlock(&anon_vma->lock); 154 } 155} 156 157void anon_vma_unlink(struct vm_area_struct *vma) 158{ 159 struct anon_vma *anon_vma = vma->anon_vma; 160 int empty; 161 162 if (!anon_vma) 163 return; 164 165 spin_lock(&anon_vma->lock); 166 list_del(&vma->anon_vma_node); 167 168 /* We must garbage collect the anon_vma if it's empty */ 169 empty = list_empty(&anon_vma->head); 170 spin_unlock(&anon_vma->lock); 171 172 if (empty) 173 anon_vma_free(anon_vma); 174} 175 176static void anon_vma_ctor(void *data) 177{ 178 struct anon_vma *anon_vma = data; 179 180 spin_lock_init(&anon_vma->lock); 181 INIT_LIST_HEAD(&anon_vma->head); 182} 183 184void __init anon_vma_init(void) 185{ 186 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), 187 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor); 188} 189 190/* 191 * Getting a lock on a stable anon_vma from a page off the LRU is 192 * tricky: page_lock_anon_vma rely on RCU to guard against the races. 193 */ 194static struct anon_vma *page_lock_anon_vma(struct page *page) 195{ 196 struct anon_vma *anon_vma; 197 unsigned long anon_mapping; 198 199 rcu_read_lock(); 200 anon_mapping = (unsigned long) page->mapping; 201 if (!(anon_mapping & PAGE_MAPPING_ANON)) 202 goto out; 203 if (!page_mapped(page)) 204 goto out; 205 206 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 207 spin_lock(&anon_vma->lock); 208 return anon_vma; 209out: 210 rcu_read_unlock(); 211 return NULL; 212} 213 214static void page_unlock_anon_vma(struct anon_vma *anon_vma) 215{ 216 spin_unlock(&anon_vma->lock); 217 rcu_read_unlock(); 218} 219 220/* 221 * At what user virtual address is page expected in @vma? 222 * Returns virtual address or -EFAULT if page's index/offset is not 223 * within the range mapped the @vma. 224 */ 225static inline unsigned long 226vma_address(struct page *page, struct vm_area_struct *vma) 227{ 228 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 229 unsigned long address; 230 231 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 232 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { 233 /* page should be within @vma mapping range */ 234 return -EFAULT; 235 } 236 return address; 237} 238 239/* 240 * At what user virtual address is page expected in vma? checking that the 241 * page matches the vma: currently only used on anon pages, by unuse_vma; 242 */ 243unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) 244{ 245 if (PageAnon(page)) { 246 if ((void *)vma->anon_vma != 247 (void *)page->mapping - PAGE_MAPPING_ANON) 248 return -EFAULT; 249 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { 250 if (!vma->vm_file || 251 vma->vm_file->f_mapping != page->mapping) 252 return -EFAULT; 253 } else 254 return -EFAULT; 255 return vma_address(page, vma); 256} 257 258/* 259 * Check that @page is mapped at @address into @mm. 260 * 261 * If @sync is false, page_check_address may perform a racy check to avoid 262 * the page table lock when the pte is not present (helpful when reclaiming 263 * highly shared pages). 264 * 265 * On success returns with pte mapped and locked. 266 */ 267pte_t *page_check_address(struct page *page, struct mm_struct *mm, 268 unsigned long address, spinlock_t **ptlp, int sync) 269{ 270 pgd_t *pgd; 271 pud_t *pud; 272 pmd_t *pmd; 273 pte_t *pte; 274 spinlock_t *ptl; 275 276 pgd = pgd_offset(mm, address); 277 if (!pgd_present(*pgd)) 278 return NULL; 279 280 pud = pud_offset(pgd, address); 281 if (!pud_present(*pud)) 282 return NULL; 283 284 pmd = pmd_offset(pud, address); 285 if (!pmd_present(*pmd)) 286 return NULL; 287 288 pte = pte_offset_map(pmd, address); 289 /* Make a quick check before getting the lock */ 290 if (!sync && !pte_present(*pte)) { 291 pte_unmap(pte); 292 return NULL; 293 } 294 295 ptl = pte_lockptr(mm, pmd); 296 spin_lock(ptl); 297 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { 298 *ptlp = ptl; 299 return pte; 300 } 301 pte_unmap_unlock(pte, ptl); 302 return NULL; 303} 304 305/** 306 * page_mapped_in_vma - check whether a page is really mapped in a VMA 307 * @page: the page to test 308 * @vma: the VMA to test 309 * 310 * Returns 1 if the page is mapped into the page tables of the VMA, 0 311 * if the page is not mapped into the page tables of this VMA. Only 312 * valid for normal file or anonymous VMAs. 313 */ 314static int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma) 315{ 316 unsigned long address; 317 pte_t *pte; 318 spinlock_t *ptl; 319 320 address = vma_address(page, vma); 321 if (address == -EFAULT) /* out of vma range */ 322 return 0; 323 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1); 324 if (!pte) /* the page is not in this mm */ 325 return 0; 326 pte_unmap_unlock(pte, ptl); 327 328 return 1; 329} 330 331/* 332 * Subfunctions of page_referenced: page_referenced_one called 333 * repeatedly from either page_referenced_anon or page_referenced_file. 334 */ 335static int page_referenced_one(struct page *page, 336 struct vm_area_struct *vma, unsigned int *mapcount) 337{ 338 struct mm_struct *mm = vma->vm_mm; 339 unsigned long address; 340 pte_t *pte; 341 spinlock_t *ptl; 342 int referenced = 0; 343 344 address = vma_address(page, vma); 345 if (address == -EFAULT) 346 goto out; 347 348 pte = page_check_address(page, mm, address, &ptl, 0); 349 if (!pte) 350 goto out; 351 352 /* 353 * Don't want to elevate referenced for mlocked page that gets this far, 354 * in order that it progresses to try_to_unmap and is moved to the 355 * unevictable list. 356 */ 357 if (vma->vm_flags & VM_LOCKED) { 358 *mapcount = 1; /* break early from loop */ 359 goto out_unmap; 360 } 361 362 if (ptep_clear_flush_young_notify(vma, address, pte)) { 363 /* 364 * Don't treat a reference through a sequentially read 365 * mapping as such. If the page has been used in 366 * another mapping, we will catch it; if this other 367 * mapping is already gone, the unmap path will have 368 * set PG_referenced or activated the page. 369 */ 370 if (likely(!VM_SequentialReadHint(vma))) 371 referenced++; 372 } 373 374 /* Pretend the page is referenced if the task has the 375 swap token and is in the middle of a page fault. */ 376 if (mm != current->mm && has_swap_token(mm) && 377 rwsem_is_locked(&mm->mmap_sem)) 378 referenced++; 379 380out_unmap: 381 (*mapcount)--; 382 pte_unmap_unlock(pte, ptl); 383out: 384 return referenced; 385} 386 387static int page_referenced_anon(struct page *page, 388 struct mem_cgroup *mem_cont) 389{ 390 unsigned int mapcount; 391 struct anon_vma *anon_vma; 392 struct vm_area_struct *vma; 393 int referenced = 0; 394 395 anon_vma = page_lock_anon_vma(page); 396 if (!anon_vma) 397 return referenced; 398 399 mapcount = page_mapcount(page); 400 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { 401 /* 402 * If we are reclaiming on behalf of a cgroup, skip 403 * counting on behalf of references from different 404 * cgroups 405 */ 406 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont)) 407 continue; 408 referenced += page_referenced_one(page, vma, &mapcount); 409 if (!mapcount) 410 break; 411 } 412 413 page_unlock_anon_vma(anon_vma); 414 return referenced; 415} 416 417/** 418 * page_referenced_file - referenced check for object-based rmap 419 * @page: the page we're checking references on. 420 * @mem_cont: target memory controller 421 * 422 * For an object-based mapped page, find all the places it is mapped and 423 * check/clear the referenced flag. This is done by following the page->mapping 424 * pointer, then walking the chain of vmas it holds. It returns the number 425 * of references it found. 426 * 427 * This function is only called from page_referenced for object-based pages. 428 */ 429static int page_referenced_file(struct page *page, 430 struct mem_cgroup *mem_cont) 431{ 432 unsigned int mapcount; 433 struct address_space *mapping = page->mapping; 434 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 435 struct vm_area_struct *vma; 436 struct prio_tree_iter iter; 437 int referenced = 0; 438 439 /* 440 * The caller's checks on page->mapping and !PageAnon have made 441 * sure that this is a file page: the check for page->mapping 442 * excludes the case just before it gets set on an anon page. 443 */ 444 BUG_ON(PageAnon(page)); 445 446 /* 447 * The page lock not only makes sure that page->mapping cannot 448 * suddenly be NULLified by truncation, it makes sure that the 449 * structure at mapping cannot be freed and reused yet, 450 * so we can safely take mapping->i_mmap_lock. 451 */ 452 BUG_ON(!PageLocked(page)); 453 454 spin_lock(&mapping->i_mmap_lock); 455 456 /* 457 * i_mmap_lock does not stabilize mapcount at all, but mapcount 458 * is more likely to be accurate if we note it after spinning. 459 */ 460 mapcount = page_mapcount(page); 461 462 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 463 /* 464 * If we are reclaiming on behalf of a cgroup, skip 465 * counting on behalf of references from different 466 * cgroups 467 */ 468 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont)) 469 continue; 470 referenced += page_referenced_one(page, vma, &mapcount); 471 if (!mapcount) 472 break; 473 } 474 475 spin_unlock(&mapping->i_mmap_lock); 476 return referenced; 477} 478 479/** 480 * page_referenced - test if the page was referenced 481 * @page: the page to test 482 * @is_locked: caller holds lock on the page 483 * @mem_cont: target memory controller 484 * 485 * Quick test_and_clear_referenced for all mappings to a page, 486 * returns the number of ptes which referenced the page. 487 */ 488int page_referenced(struct page *page, int is_locked, 489 struct mem_cgroup *mem_cont) 490{ 491 int referenced = 0; 492 493 if (TestClearPageReferenced(page)) 494 referenced++; 495 496 if (page_mapped(page) && page->mapping) { 497 if (PageAnon(page)) 498 referenced += page_referenced_anon(page, mem_cont); 499 else if (is_locked) 500 referenced += page_referenced_file(page, mem_cont); 501 else if (!trylock_page(page)) 502 referenced++; 503 else { 504 if (page->mapping) 505 referenced += 506 page_referenced_file(page, mem_cont); 507 unlock_page(page); 508 } 509 } 510 511 if (page_test_and_clear_young(page)) 512 referenced++; 513 514 return referenced; 515} 516 517static int page_mkclean_one(struct page *page, struct vm_area_struct *vma) 518{ 519 struct mm_struct *mm = vma->vm_mm; 520 unsigned long address; 521 pte_t *pte; 522 spinlock_t *ptl; 523 int ret = 0; 524 525 address = vma_address(page, vma); 526 if (address == -EFAULT) 527 goto out; 528 529 pte = page_check_address(page, mm, address, &ptl, 1); 530 if (!pte) 531 goto out; 532 533 if (pte_dirty(*pte) || pte_write(*pte)) { 534 pte_t entry; 535 536 flush_cache_page(vma, address, pte_pfn(*pte)); 537 entry = ptep_clear_flush_notify(vma, address, pte); 538 entry = pte_wrprotect(entry); 539 entry = pte_mkclean(entry); 540 set_pte_at(mm, address, pte, entry); 541 ret = 1; 542 } 543 544 pte_unmap_unlock(pte, ptl); 545out: 546 return ret; 547} 548 549static int page_mkclean_file(struct address_space *mapping, struct page *page) 550{ 551 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 552 struct vm_area_struct *vma; 553 struct prio_tree_iter iter; 554 int ret = 0; 555 556 BUG_ON(PageAnon(page)); 557 558 spin_lock(&mapping->i_mmap_lock); 559 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 560 if (vma->vm_flags & VM_SHARED) 561 ret += page_mkclean_one(page, vma); 562 } 563 spin_unlock(&mapping->i_mmap_lock); 564 return ret; 565} 566 567int page_mkclean(struct page *page) 568{ 569 int ret = 0; 570 571 BUG_ON(!PageLocked(page)); 572 573 if (page_mapped(page)) { 574 struct address_space *mapping = page_mapping(page); 575 if (mapping) { 576 ret = page_mkclean_file(mapping, page); 577 if (page_test_dirty(page)) { 578 page_clear_dirty(page); 579 ret = 1; 580 } 581 } 582 } 583 584 return ret; 585} 586EXPORT_SYMBOL_GPL(page_mkclean); 587 588/** 589 * __page_set_anon_rmap - setup new anonymous rmap 590 * @page: the page to add the mapping to 591 * @vma: the vm area in which the mapping is added 592 * @address: the user virtual address mapped 593 */ 594static void __page_set_anon_rmap(struct page *page, 595 struct vm_area_struct *vma, unsigned long address) 596{ 597 struct anon_vma *anon_vma = vma->anon_vma; 598 599 BUG_ON(!anon_vma); 600 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 601 page->mapping = (struct address_space *) anon_vma; 602 603 page->index = linear_page_index(vma, address); 604 605 /* 606 * nr_mapped state can be updated without turning off 607 * interrupts because it is not modified via interrupt. 608 */ 609 __inc_zone_page_state(page, NR_ANON_PAGES); 610} 611 612/** 613 * __page_check_anon_rmap - sanity check anonymous rmap addition 614 * @page: the page to add the mapping to 615 * @vma: the vm area in which the mapping is added 616 * @address: the user virtual address mapped 617 */ 618static void __page_check_anon_rmap(struct page *page, 619 struct vm_area_struct *vma, unsigned long address) 620{ 621#ifdef CONFIG_DEBUG_VM 622 /* 623 * The page's anon-rmap details (mapping and index) are guaranteed to 624 * be set up correctly at this point. 625 * 626 * We have exclusion against page_add_anon_rmap because the caller 627 * always holds the page locked, except if called from page_dup_rmap, 628 * in which case the page is already known to be setup. 629 * 630 * We have exclusion against page_add_new_anon_rmap because those pages 631 * are initially only visible via the pagetables, and the pte is locked 632 * over the call to page_add_new_anon_rmap. 633 */ 634 struct anon_vma *anon_vma = vma->anon_vma; 635 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 636 BUG_ON(page->mapping != (struct address_space *)anon_vma); 637 BUG_ON(page->index != linear_page_index(vma, address)); 638#endif 639} 640 641/** 642 * page_add_anon_rmap - add pte mapping to an anonymous page 643 * @page: the page to add the mapping to 644 * @vma: the vm area in which the mapping is added 645 * @address: the user virtual address mapped 646 * 647 * The caller needs to hold the pte lock and the page must be locked. 648 */ 649void page_add_anon_rmap(struct page *page, 650 struct vm_area_struct *vma, unsigned long address) 651{ 652 VM_BUG_ON(!PageLocked(page)); 653 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); 654 if (atomic_inc_and_test(&page->_mapcount)) 655 __page_set_anon_rmap(page, vma, address); 656 else 657 __page_check_anon_rmap(page, vma, address); 658} 659 660/** 661 * page_add_new_anon_rmap - add pte mapping to a new anonymous page 662 * @page: the page to add the mapping to 663 * @vma: the vm area in which the mapping is added 664 * @address: the user virtual address mapped 665 * 666 * Same as page_add_anon_rmap but must only be called on *new* pages. 667 * This means the inc-and-test can be bypassed. 668 * Page does not have to be locked. 669 */ 670void page_add_new_anon_rmap(struct page *page, 671 struct vm_area_struct *vma, unsigned long address) 672{ 673 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); 674 SetPageSwapBacked(page); 675 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */ 676 __page_set_anon_rmap(page, vma, address); 677 if (page_evictable(page, vma)) 678 lru_cache_add_lru(page, LRU_ACTIVE_ANON); 679 else 680 add_page_to_unevictable_list(page); 681} 682 683/** 684 * page_add_file_rmap - add pte mapping to a file page 685 * @page: the page to add the mapping to 686 * 687 * The caller needs to hold the pte lock. 688 */ 689void page_add_file_rmap(struct page *page) 690{ 691 if (atomic_inc_and_test(&page->_mapcount)) 692 __inc_zone_page_state(page, NR_FILE_MAPPED); 693} 694 695#ifdef CONFIG_DEBUG_VM 696/** 697 * page_dup_rmap - duplicate pte mapping to a page 698 * @page: the page to add the mapping to 699 * @vma: the vm area being duplicated 700 * @address: the user virtual address mapped 701 * 702 * For copy_page_range only: minimal extract from page_add_file_rmap / 703 * page_add_anon_rmap, avoiding unnecessary tests (already checked) so it's 704 * quicker. 705 * 706 * The caller needs to hold the pte lock. 707 */ 708void page_dup_rmap(struct page *page, struct vm_area_struct *vma, unsigned long address) 709{ 710 BUG_ON(page_mapcount(page) == 0); 711 if (PageAnon(page)) 712 __page_check_anon_rmap(page, vma, address); 713 atomic_inc(&page->_mapcount); 714} 715#endif 716 717/** 718 * page_remove_rmap - take down pte mapping from a page 719 * @page: page to remove mapping from 720 * @vma: the vm area in which the mapping is removed 721 * 722 * The caller needs to hold the pte lock. 723 */ 724void page_remove_rmap(struct page *page, struct vm_area_struct *vma) 725{ 726 if (atomic_add_negative(-1, &page->_mapcount)) { 727 /* 728 * Now that the last pte has gone, s390 must transfer dirty 729 * flag from storage key to struct page. We can usually skip 730 * this if the page is anon, so about to be freed; but perhaps 731 * not if it's in swapcache - there might be another pte slot 732 * containing the swap entry, but page not yet written to swap. 733 */ 734 if ((!PageAnon(page) || PageSwapCache(page)) && 735 page_test_dirty(page)) { 736 page_clear_dirty(page); 737 set_page_dirty(page); 738 } 739 if (PageAnon(page)) 740 mem_cgroup_uncharge_page(page); 741 __dec_zone_page_state(page, 742 PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED); 743 /* 744 * It would be tidy to reset the PageAnon mapping here, 745 * but that might overwrite a racing page_add_anon_rmap 746 * which increments mapcount after us but sets mapping 747 * before us: so leave the reset to free_hot_cold_page, 748 * and remember that it's only reliable while mapped. 749 * Leaving it set also helps swapoff to reinstate ptes 750 * faster for those pages still in swapcache. 751 */ 752 } 753} 754 755/* 756 * Subfunctions of try_to_unmap: try_to_unmap_one called 757 * repeatedly from either try_to_unmap_anon or try_to_unmap_file. 758 */ 759static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, 760 int migration) 761{ 762 struct mm_struct *mm = vma->vm_mm; 763 unsigned long address; 764 pte_t *pte; 765 pte_t pteval; 766 spinlock_t *ptl; 767 int ret = SWAP_AGAIN; 768 769 address = vma_address(page, vma); 770 if (address == -EFAULT) 771 goto out; 772 773 pte = page_check_address(page, mm, address, &ptl, 0); 774 if (!pte) 775 goto out; 776 777 /* 778 * If the page is mlock()d, we cannot swap it out. 779 * If it's recently referenced (perhaps page_referenced 780 * skipped over this mm) then we should reactivate it. 781 */ 782 if (!migration) { 783 if (vma->vm_flags & VM_LOCKED) { 784 ret = SWAP_MLOCK; 785 goto out_unmap; 786 } 787 if (ptep_clear_flush_young_notify(vma, address, pte)) { 788 ret = SWAP_FAIL; 789 goto out_unmap; 790 } 791 } 792 793 /* Nuke the page table entry. */ 794 flush_cache_page(vma, address, page_to_pfn(page)); 795 pteval = ptep_clear_flush_notify(vma, address, pte); 796 797 /* Move the dirty bit to the physical page now the pte is gone. */ 798 if (pte_dirty(pteval)) 799 set_page_dirty(page); 800 801 /* Update high watermark before we lower rss */ 802 update_hiwater_rss(mm); 803 804 if (PageAnon(page)) { 805 swp_entry_t entry = { .val = page_private(page) }; 806 807 if (PageSwapCache(page)) { 808 /* 809 * Store the swap location in the pte. 810 * See handle_pte_fault() ... 811 */ 812 swap_duplicate(entry); 813 if (list_empty(&mm->mmlist)) { 814 spin_lock(&mmlist_lock); 815 if (list_empty(&mm->mmlist)) 816 list_add(&mm->mmlist, &init_mm.mmlist); 817 spin_unlock(&mmlist_lock); 818 } 819 dec_mm_counter(mm, anon_rss); 820 } else if (PAGE_MIGRATION) { 821 /* 822 * Store the pfn of the page in a special migration 823 * pte. do_swap_page() will wait until the migration 824 * pte is removed and then restart fault handling. 825 */ 826 BUG_ON(!migration); 827 entry = make_migration_entry(page, pte_write(pteval)); 828 } 829 set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); 830 BUG_ON(pte_file(*pte)); 831 } else if (PAGE_MIGRATION && migration) { 832 /* Establish migration entry for a file page */ 833 swp_entry_t entry; 834 entry = make_migration_entry(page, pte_write(pteval)); 835 set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); 836 } else 837 dec_mm_counter(mm, file_rss); 838 839 840 page_remove_rmap(page, vma); 841 page_cache_release(page); 842 843out_unmap: 844 pte_unmap_unlock(pte, ptl); 845out: 846 return ret; 847} 848 849/* 850 * objrmap doesn't work for nonlinear VMAs because the assumption that 851 * offset-into-file correlates with offset-into-virtual-addresses does not hold. 852 * Consequently, given a particular page and its ->index, we cannot locate the 853 * ptes which are mapping that page without an exhaustive linear search. 854 * 855 * So what this code does is a mini "virtual scan" of each nonlinear VMA which 856 * maps the file to which the target page belongs. The ->vm_private_data field 857 * holds the current cursor into that scan. Successive searches will circulate 858 * around the vma's virtual address space. 859 * 860 * So as more replacement pressure is applied to the pages in a nonlinear VMA, 861 * more scanning pressure is placed against them as well. Eventually pages 862 * will become fully unmapped and are eligible for eviction. 863 * 864 * For very sparsely populated VMAs this is a little inefficient - chances are 865 * there there won't be many ptes located within the scan cluster. In this case 866 * maybe we could scan further - to the end of the pte page, perhaps. 867 * 868 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can 869 * acquire it without blocking. If vma locked, mlock the pages in the cluster, 870 * rather than unmapping them. If we encounter the "check_page" that vmscan is 871 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN. 872 */ 873#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) 874#define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) 875 876static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount, 877 struct vm_area_struct *vma, struct page *check_page) 878{ 879 struct mm_struct *mm = vma->vm_mm; 880 pgd_t *pgd; 881 pud_t *pud; 882 pmd_t *pmd; 883 pte_t *pte; 884 pte_t pteval; 885 spinlock_t *ptl; 886 struct page *page; 887 unsigned long address; 888 unsigned long end; 889 int ret = SWAP_AGAIN; 890 int locked_vma = 0; 891 892 address = (vma->vm_start + cursor) & CLUSTER_MASK; 893 end = address + CLUSTER_SIZE; 894 if (address < vma->vm_start) 895 address = vma->vm_start; 896 if (end > vma->vm_end) 897 end = vma->vm_end; 898 899 pgd = pgd_offset(mm, address); 900 if (!pgd_present(*pgd)) 901 return ret; 902 903 pud = pud_offset(pgd, address); 904 if (!pud_present(*pud)) 905 return ret; 906 907 pmd = pmd_offset(pud, address); 908 if (!pmd_present(*pmd)) 909 return ret; 910 911 /* 912 * MLOCK_PAGES => feature is configured. 913 * if we can acquire the mmap_sem for read, and vma is VM_LOCKED, 914 * keep the sem while scanning the cluster for mlocking pages. 915 */ 916 if (MLOCK_PAGES && down_read_trylock(&vma->vm_mm->mmap_sem)) { 917 locked_vma = (vma->vm_flags & VM_LOCKED); 918 if (!locked_vma) 919 up_read(&vma->vm_mm->mmap_sem); /* don't need it */ 920 } 921 922 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 923 924 /* Update high watermark before we lower rss */ 925 update_hiwater_rss(mm); 926 927 for (; address < end; pte++, address += PAGE_SIZE) { 928 if (!pte_present(*pte)) 929 continue; 930 page = vm_normal_page(vma, address, *pte); 931 BUG_ON(!page || PageAnon(page)); 932 933 if (locked_vma) { 934 mlock_vma_page(page); /* no-op if already mlocked */ 935 if (page == check_page) 936 ret = SWAP_MLOCK; 937 continue; /* don't unmap */ 938 } 939 940 if (ptep_clear_flush_young_notify(vma, address, pte)) 941 continue; 942 943 /* Nuke the page table entry. */ 944 flush_cache_page(vma, address, pte_pfn(*pte)); 945 pteval = ptep_clear_flush_notify(vma, address, pte); 946 947 /* If nonlinear, store the file page offset in the pte. */ 948 if (page->index != linear_page_index(vma, address)) 949 set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); 950 951 /* Move the dirty bit to the physical page now the pte is gone. */ 952 if (pte_dirty(pteval)) 953 set_page_dirty(page); 954 955 page_remove_rmap(page, vma); 956 page_cache_release(page); 957 dec_mm_counter(mm, file_rss); 958 (*mapcount)--; 959 } 960 pte_unmap_unlock(pte - 1, ptl); 961 if (locked_vma) 962 up_read(&vma->vm_mm->mmap_sem); 963 return ret; 964} 965 966/* 967 * common handling for pages mapped in VM_LOCKED vmas 968 */ 969static int try_to_mlock_page(struct page *page, struct vm_area_struct *vma) 970{ 971 int mlocked = 0; 972 973 if (down_read_trylock(&vma->vm_mm->mmap_sem)) { 974 if (vma->vm_flags & VM_LOCKED) { 975 mlock_vma_page(page); 976 mlocked++; /* really mlocked the page */ 977 } 978 up_read(&vma->vm_mm->mmap_sem); 979 } 980 return mlocked; 981} 982 983/** 984 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based 985 * rmap method 986 * @page: the page to unmap/unlock 987 * @unlock: request for unlock rather than unmap [unlikely] 988 * @migration: unmapping for migration - ignored if @unlock 989 * 990 * Find all the mappings of a page using the mapping pointer and the vma chains 991 * contained in the anon_vma struct it points to. 992 * 993 * This function is only called from try_to_unmap/try_to_munlock for 994 * anonymous pages. 995 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma 996 * where the page was found will be held for write. So, we won't recheck 997 * vm_flags for that VMA. That should be OK, because that vma shouldn't be 998 * 'LOCKED. 999 */ 1000static int try_to_unmap_anon(struct page *page, int unlock, int migration) 1001{ 1002 struct anon_vma *anon_vma; 1003 struct vm_area_struct *vma; 1004 unsigned int mlocked = 0; 1005 int ret = SWAP_AGAIN; 1006 1007 if (MLOCK_PAGES && unlikely(unlock)) 1008 ret = SWAP_SUCCESS; /* default for try_to_munlock() */ 1009 1010 anon_vma = page_lock_anon_vma(page); 1011 if (!anon_vma) 1012 return ret; 1013 1014 list_for_each_entry(vma, &anon_vma->head, anon_vma_node) { 1015 if (MLOCK_PAGES && unlikely(unlock)) { 1016 if (!((vma->vm_flags & VM_LOCKED) && 1017 page_mapped_in_vma(page, vma))) 1018 continue; /* must visit all unlocked vmas */ 1019 ret = SWAP_MLOCK; /* saw at least one mlocked vma */ 1020 } else { 1021 ret = try_to_unmap_one(page, vma, migration); 1022 if (ret == SWAP_FAIL || !page_mapped(page)) 1023 break; 1024 } 1025 if (ret == SWAP_MLOCK) { 1026 mlocked = try_to_mlock_page(page, vma); 1027 if (mlocked) 1028 break; /* stop if actually mlocked page */ 1029 } 1030 } 1031 1032 page_unlock_anon_vma(anon_vma); 1033 1034 if (mlocked) 1035 ret = SWAP_MLOCK; /* actually mlocked the page */ 1036 else if (ret == SWAP_MLOCK) 1037 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */ 1038 1039 return ret; 1040} 1041 1042/** 1043 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method 1044 * @page: the page to unmap/unlock 1045 * @unlock: request for unlock rather than unmap [unlikely] 1046 * @migration: unmapping for migration - ignored if @unlock 1047 * 1048 * Find all the mappings of a page using the mapping pointer and the vma chains 1049 * contained in the address_space struct it points to. 1050 * 1051 * This function is only called from try_to_unmap/try_to_munlock for 1052 * object-based pages. 1053 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma 1054 * where the page was found will be held for write. So, we won't recheck 1055 * vm_flags for that VMA. That should be OK, because that vma shouldn't be 1056 * 'LOCKED. 1057 */ 1058static int try_to_unmap_file(struct page *page, int unlock, int migration) 1059{ 1060 struct address_space *mapping = page->mapping; 1061 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 1062 struct vm_area_struct *vma; 1063 struct prio_tree_iter iter; 1064 int ret = SWAP_AGAIN; 1065 unsigned long cursor; 1066 unsigned long max_nl_cursor = 0; 1067 unsigned long max_nl_size = 0; 1068 unsigned int mapcount; 1069 unsigned int mlocked = 0; 1070 1071 if (MLOCK_PAGES && unlikely(unlock)) 1072 ret = SWAP_SUCCESS; /* default for try_to_munlock() */ 1073 1074 spin_lock(&mapping->i_mmap_lock); 1075 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 1076 if (MLOCK_PAGES && unlikely(unlock)) { 1077 if (!(vma->vm_flags & VM_LOCKED)) 1078 continue; /* must visit all vmas */ 1079 ret = SWAP_MLOCK; 1080 } else { 1081 ret = try_to_unmap_one(page, vma, migration); 1082 if (ret == SWAP_FAIL || !page_mapped(page)) 1083 goto out; 1084 } 1085 if (ret == SWAP_MLOCK) { 1086 mlocked = try_to_mlock_page(page, vma); 1087 if (mlocked) 1088 break; /* stop if actually mlocked page */ 1089 } 1090 } 1091 1092 if (mlocked) 1093 goto out; 1094 1095 if (list_empty(&mapping->i_mmap_nonlinear)) 1096 goto out; 1097 1098 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, 1099 shared.vm_set.list) { 1100 if (MLOCK_PAGES && unlikely(unlock)) { 1101 if (!(vma->vm_flags & VM_LOCKED)) 1102 continue; /* must visit all vmas */ 1103 ret = SWAP_MLOCK; /* leave mlocked == 0 */ 1104 goto out; /* no need to look further */ 1105 } 1106 if (!MLOCK_PAGES && !migration && (vma->vm_flags & VM_LOCKED)) 1107 continue; 1108 cursor = (unsigned long) vma->vm_private_data; 1109 if (cursor > max_nl_cursor) 1110 max_nl_cursor = cursor; 1111 cursor = vma->vm_end - vma->vm_start; 1112 if (cursor > max_nl_size) 1113 max_nl_size = cursor; 1114 } 1115 1116 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */ 1117 ret = SWAP_FAIL; 1118 goto out; 1119 } 1120 1121 /* 1122 * We don't try to search for this page in the nonlinear vmas, 1123 * and page_referenced wouldn't have found it anyway. Instead 1124 * just walk the nonlinear vmas trying to age and unmap some. 1125 * The mapcount of the page we came in with is irrelevant, 1126 * but even so use it as a guide to how hard we should try? 1127 */ 1128 mapcount = page_mapcount(page); 1129 if (!mapcount) 1130 goto out; 1131 cond_resched_lock(&mapping->i_mmap_lock); 1132 1133 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; 1134 if (max_nl_cursor == 0) 1135 max_nl_cursor = CLUSTER_SIZE; 1136 1137 do { 1138 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, 1139 shared.vm_set.list) { 1140 if (!MLOCK_PAGES && !migration && 1141 (vma->vm_flags & VM_LOCKED)) 1142 continue; 1143 cursor = (unsigned long) vma->vm_private_data; 1144 while ( cursor < max_nl_cursor && 1145 cursor < vma->vm_end - vma->vm_start) { 1146 ret = try_to_unmap_cluster(cursor, &mapcount, 1147 vma, page); 1148 if (ret == SWAP_MLOCK) 1149 mlocked = 2; /* to return below */ 1150 cursor += CLUSTER_SIZE; 1151 vma->vm_private_data = (void *) cursor; 1152 if ((int)mapcount <= 0) 1153 goto out; 1154 } 1155 vma->vm_private_data = (void *) max_nl_cursor; 1156 } 1157 cond_resched_lock(&mapping->i_mmap_lock); 1158 max_nl_cursor += CLUSTER_SIZE; 1159 } while (max_nl_cursor <= max_nl_size); 1160 1161 /* 1162 * Don't loop forever (perhaps all the remaining pages are 1163 * in locked vmas). Reset cursor on all unreserved nonlinear 1164 * vmas, now forgetting on which ones it had fallen behind. 1165 */ 1166 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) 1167 vma->vm_private_data = NULL; 1168out: 1169 spin_unlock(&mapping->i_mmap_lock); 1170 if (mlocked) 1171 ret = SWAP_MLOCK; /* actually mlocked the page */ 1172 else if (ret == SWAP_MLOCK) 1173 ret = SWAP_AGAIN; /* saw VM_LOCKED vma */ 1174 return ret; 1175} 1176 1177/** 1178 * try_to_unmap - try to remove all page table mappings to a page 1179 * @page: the page to get unmapped 1180 * @migration: migration flag 1181 * 1182 * Tries to remove all the page table entries which are mapping this 1183 * page, used in the pageout path. Caller must hold the page lock. 1184 * Return values are: 1185 * 1186 * SWAP_SUCCESS - we succeeded in removing all mappings 1187 * SWAP_AGAIN - we missed a mapping, try again later 1188 * SWAP_FAIL - the page is unswappable 1189 * SWAP_MLOCK - page is mlocked. 1190 */ 1191int try_to_unmap(struct page *page, int migration) 1192{ 1193 int ret; 1194 1195 BUG_ON(!PageLocked(page)); 1196 1197 if (PageAnon(page)) 1198 ret = try_to_unmap_anon(page, 0, migration); 1199 else 1200 ret = try_to_unmap_file(page, 0, migration); 1201 if (ret != SWAP_MLOCK && !page_mapped(page)) 1202 ret = SWAP_SUCCESS; 1203 return ret; 1204} 1205 1206#ifdef CONFIG_UNEVICTABLE_LRU 1207/** 1208 * try_to_munlock - try to munlock a page 1209 * @page: the page to be munlocked 1210 * 1211 * Called from munlock code. Checks all of the VMAs mapping the page 1212 * to make sure nobody else has this page mlocked. The page will be 1213 * returned with PG_mlocked cleared if no other vmas have it mlocked. 1214 * 1215 * Return values are: 1216 * 1217 * SWAP_SUCCESS - no vma's holding page mlocked. 1218 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem 1219 * SWAP_MLOCK - page is now mlocked. 1220 */ 1221int try_to_munlock(struct page *page) 1222{ 1223 VM_BUG_ON(!PageLocked(page) || PageLRU(page)); 1224 1225 if (PageAnon(page)) 1226 return try_to_unmap_anon(page, 1, 0); 1227 else 1228 return try_to_unmap_file(page, 1, 0); 1229} 1230#endif 1231