rmap.c revision 2da28bfd9665f49d40abb4c7720b43135feaf79a
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 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 * (code doesn't rely on that order so it could be switched around) 41 * ->tasklist_lock 42 * anon_vma->lock (memory_failure, collect_procs_anon) 43 * pte map lock 44 */ 45 46#include <linux/mm.h> 47#include <linux/pagemap.h> 48#include <linux/swap.h> 49#include <linux/swapops.h> 50#include <linux/slab.h> 51#include <linux/init.h> 52#include <linux/ksm.h> 53#include <linux/rmap.h> 54#include <linux/rcupdate.h> 55#include <linux/module.h> 56#include <linux/memcontrol.h> 57#include <linux/mmu_notifier.h> 58#include <linux/migrate.h> 59#include <linux/hugetlb.h> 60 61#include <asm/tlbflush.h> 62 63#include "internal.h" 64 65static struct kmem_cache *anon_vma_cachep; 66static struct kmem_cache *anon_vma_chain_cachep; 67 68static inline struct anon_vma *anon_vma_alloc(void) 69{ 70 return kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); 71} 72 73void anon_vma_free(struct anon_vma *anon_vma) 74{ 75 kmem_cache_free(anon_vma_cachep, anon_vma); 76} 77 78static inline struct anon_vma_chain *anon_vma_chain_alloc(void) 79{ 80 return kmem_cache_alloc(anon_vma_chain_cachep, GFP_KERNEL); 81} 82 83static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) 84{ 85 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); 86} 87 88/** 89 * anon_vma_prepare - attach an anon_vma to a memory region 90 * @vma: the memory region in question 91 * 92 * This makes sure the memory mapping described by 'vma' has 93 * an 'anon_vma' attached to it, so that we can associate the 94 * anonymous pages mapped into it with that anon_vma. 95 * 96 * The common case will be that we already have one, but if 97 * not we either need to find an adjacent mapping that we 98 * can re-use the anon_vma from (very common when the only 99 * reason for splitting a vma has been mprotect()), or we 100 * allocate a new one. 101 * 102 * Anon-vma allocations are very subtle, because we may have 103 * optimistically looked up an anon_vma in page_lock_anon_vma() 104 * and that may actually touch the spinlock even in the newly 105 * allocated vma (it depends on RCU to make sure that the 106 * anon_vma isn't actually destroyed). 107 * 108 * As a result, we need to do proper anon_vma locking even 109 * for the new allocation. At the same time, we do not want 110 * to do any locking for the common case of already having 111 * an anon_vma. 112 * 113 * This must be called with the mmap_sem held for reading. 114 */ 115int anon_vma_prepare(struct vm_area_struct *vma) 116{ 117 struct anon_vma *anon_vma = vma->anon_vma; 118 struct anon_vma_chain *avc; 119 120 might_sleep(); 121 if (unlikely(!anon_vma)) { 122 struct mm_struct *mm = vma->vm_mm; 123 struct anon_vma *allocated; 124 125 avc = anon_vma_chain_alloc(); 126 if (!avc) 127 goto out_enomem; 128 129 anon_vma = find_mergeable_anon_vma(vma); 130 allocated = NULL; 131 if (!anon_vma) { 132 anon_vma = anon_vma_alloc(); 133 if (unlikely(!anon_vma)) 134 goto out_enomem_free_avc; 135 allocated = anon_vma; 136 /* 137 * This VMA had no anon_vma yet. This anon_vma is 138 * the root of any anon_vma tree that might form. 139 */ 140 anon_vma->root = anon_vma; 141 } 142 143 anon_vma_lock(anon_vma); 144 /* page_table_lock to protect against threads */ 145 spin_lock(&mm->page_table_lock); 146 if (likely(!vma->anon_vma)) { 147 vma->anon_vma = anon_vma; 148 avc->anon_vma = anon_vma; 149 avc->vma = vma; 150 list_add(&avc->same_vma, &vma->anon_vma_chain); 151 list_add_tail(&avc->same_anon_vma, &anon_vma->head); 152 allocated = NULL; 153 avc = NULL; 154 } 155 spin_unlock(&mm->page_table_lock); 156 anon_vma_unlock(anon_vma); 157 158 if (unlikely(allocated)) 159 anon_vma_free(allocated); 160 if (unlikely(avc)) 161 anon_vma_chain_free(avc); 162 } 163 return 0; 164 165 out_enomem_free_avc: 166 anon_vma_chain_free(avc); 167 out_enomem: 168 return -ENOMEM; 169} 170 171static void anon_vma_chain_link(struct vm_area_struct *vma, 172 struct anon_vma_chain *avc, 173 struct anon_vma *anon_vma) 174{ 175 avc->vma = vma; 176 avc->anon_vma = anon_vma; 177 list_add(&avc->same_vma, &vma->anon_vma_chain); 178 179 anon_vma_lock(anon_vma); 180 /* 181 * It's critical to add new vmas to the tail of the anon_vma, 182 * see comment in huge_memory.c:__split_huge_page(). 183 */ 184 list_add_tail(&avc->same_anon_vma, &anon_vma->head); 185 anon_vma_unlock(anon_vma); 186} 187 188/* 189 * Attach the anon_vmas from src to dst. 190 * Returns 0 on success, -ENOMEM on failure. 191 */ 192int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) 193{ 194 struct anon_vma_chain *avc, *pavc; 195 196 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { 197 avc = anon_vma_chain_alloc(); 198 if (!avc) 199 goto enomem_failure; 200 anon_vma_chain_link(dst, avc, pavc->anon_vma); 201 } 202 return 0; 203 204 enomem_failure: 205 unlink_anon_vmas(dst); 206 return -ENOMEM; 207} 208 209/* 210 * Attach vma to its own anon_vma, as well as to the anon_vmas that 211 * the corresponding VMA in the parent process is attached to. 212 * Returns 0 on success, non-zero on failure. 213 */ 214int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) 215{ 216 struct anon_vma_chain *avc; 217 struct anon_vma *anon_vma; 218 219 /* Don't bother if the parent process has no anon_vma here. */ 220 if (!pvma->anon_vma) 221 return 0; 222 223 /* 224 * First, attach the new VMA to the parent VMA's anon_vmas, 225 * so rmap can find non-COWed pages in child processes. 226 */ 227 if (anon_vma_clone(vma, pvma)) 228 return -ENOMEM; 229 230 /* Then add our own anon_vma. */ 231 anon_vma = anon_vma_alloc(); 232 if (!anon_vma) 233 goto out_error; 234 avc = anon_vma_chain_alloc(); 235 if (!avc) 236 goto out_error_free_anon_vma; 237 238 /* 239 * The root anon_vma's spinlock is the lock actually used when we 240 * lock any of the anon_vmas in this anon_vma tree. 241 */ 242 anon_vma->root = pvma->anon_vma->root; 243 /* 244 * With KSM refcounts, an anon_vma can stay around longer than the 245 * process it belongs to. The root anon_vma needs to be pinned 246 * until this anon_vma is freed, because the lock lives in the root. 247 */ 248 get_anon_vma(anon_vma->root); 249 /* Mark this anon_vma as the one where our new (COWed) pages go. */ 250 vma->anon_vma = anon_vma; 251 anon_vma_chain_link(vma, avc, anon_vma); 252 253 return 0; 254 255 out_error_free_anon_vma: 256 anon_vma_free(anon_vma); 257 out_error: 258 unlink_anon_vmas(vma); 259 return -ENOMEM; 260} 261 262static void anon_vma_unlink(struct anon_vma_chain *anon_vma_chain) 263{ 264 struct anon_vma *anon_vma = anon_vma_chain->anon_vma; 265 int empty; 266 267 /* If anon_vma_fork fails, we can get an empty anon_vma_chain. */ 268 if (!anon_vma) 269 return; 270 271 anon_vma_lock(anon_vma); 272 list_del(&anon_vma_chain->same_anon_vma); 273 274 /* We must garbage collect the anon_vma if it's empty */ 275 empty = list_empty(&anon_vma->head) && !anonvma_external_refcount(anon_vma); 276 anon_vma_unlock(anon_vma); 277 278 if (empty) { 279 /* We no longer need the root anon_vma */ 280 if (anon_vma->root != anon_vma) 281 drop_anon_vma(anon_vma->root); 282 anon_vma_free(anon_vma); 283 } 284} 285 286void unlink_anon_vmas(struct vm_area_struct *vma) 287{ 288 struct anon_vma_chain *avc, *next; 289 290 /* 291 * Unlink each anon_vma chained to the VMA. This list is ordered 292 * from newest to oldest, ensuring the root anon_vma gets freed last. 293 */ 294 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { 295 anon_vma_unlink(avc); 296 list_del(&avc->same_vma); 297 anon_vma_chain_free(avc); 298 } 299} 300 301static void anon_vma_ctor(void *data) 302{ 303 struct anon_vma *anon_vma = data; 304 305 spin_lock_init(&anon_vma->lock); 306 anonvma_external_refcount_init(anon_vma); 307 INIT_LIST_HEAD(&anon_vma->head); 308} 309 310void __init anon_vma_init(void) 311{ 312 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), 313 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC, anon_vma_ctor); 314 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, SLAB_PANIC); 315} 316 317/* 318 * Getting a lock on a stable anon_vma from a page off the LRU is 319 * tricky: page_lock_anon_vma rely on RCU to guard against the races. 320 */ 321struct anon_vma *__page_lock_anon_vma(struct page *page) 322{ 323 struct anon_vma *anon_vma, *root_anon_vma; 324 unsigned long anon_mapping; 325 326 rcu_read_lock(); 327 anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping); 328 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) 329 goto out; 330 if (!page_mapped(page)) 331 goto out; 332 333 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 334 root_anon_vma = ACCESS_ONCE(anon_vma->root); 335 spin_lock(&root_anon_vma->lock); 336 337 /* 338 * If this page is still mapped, then its anon_vma cannot have been 339 * freed. But if it has been unmapped, we have no security against 340 * the anon_vma structure being freed and reused (for another anon_vma: 341 * SLAB_DESTROY_BY_RCU guarantees that - so the spin_lock above cannot 342 * corrupt): with anon_vma_prepare() or anon_vma_fork() redirecting 343 * anon_vma->root before page_unlock_anon_vma() is called to unlock. 344 */ 345 if (page_mapped(page)) 346 return anon_vma; 347 348 spin_unlock(&root_anon_vma->lock); 349out: 350 rcu_read_unlock(); 351 return NULL; 352} 353 354void page_unlock_anon_vma(struct anon_vma *anon_vma) 355 __releases(&anon_vma->root->lock) 356 __releases(RCU) 357{ 358 anon_vma_unlock(anon_vma); 359 rcu_read_unlock(); 360} 361 362/* 363 * At what user virtual address is page expected in @vma? 364 * Returns virtual address or -EFAULT if page's index/offset is not 365 * within the range mapped the @vma. 366 */ 367inline unsigned long 368vma_address(struct page *page, struct vm_area_struct *vma) 369{ 370 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 371 unsigned long address; 372 373 if (unlikely(is_vm_hugetlb_page(vma))) 374 pgoff = page->index << huge_page_order(page_hstate(page)); 375 address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); 376 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) { 377 /* page should be within @vma mapping range */ 378 return -EFAULT; 379 } 380 return address; 381} 382 383/* 384 * At what user virtual address is page expected in vma? 385 * Caller should check the page is actually part of the vma. 386 */ 387unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) 388{ 389 if (PageAnon(page)) { 390 struct anon_vma *page__anon_vma = page_anon_vma(page); 391 /* 392 * Note: swapoff's unuse_vma() is more efficient with this 393 * check, and needs it to match anon_vma when KSM is active. 394 */ 395 if (!vma->anon_vma || !page__anon_vma || 396 vma->anon_vma->root != page__anon_vma->root) 397 return -EFAULT; 398 } else if (page->mapping && !(vma->vm_flags & VM_NONLINEAR)) { 399 if (!vma->vm_file || 400 vma->vm_file->f_mapping != page->mapping) 401 return -EFAULT; 402 } else 403 return -EFAULT; 404 return vma_address(page, vma); 405} 406 407/* 408 * Check that @page is mapped at @address into @mm. 409 * 410 * If @sync is false, page_check_address may perform a racy check to avoid 411 * the page table lock when the pte is not present (helpful when reclaiming 412 * highly shared pages). 413 * 414 * On success returns with pte mapped and locked. 415 */ 416pte_t *__page_check_address(struct page *page, struct mm_struct *mm, 417 unsigned long address, spinlock_t **ptlp, int sync) 418{ 419 pgd_t *pgd; 420 pud_t *pud; 421 pmd_t *pmd; 422 pte_t *pte; 423 spinlock_t *ptl; 424 425 if (unlikely(PageHuge(page))) { 426 pte = huge_pte_offset(mm, address); 427 ptl = &mm->page_table_lock; 428 goto check; 429 } 430 431 pgd = pgd_offset(mm, address); 432 if (!pgd_present(*pgd)) 433 return NULL; 434 435 pud = pud_offset(pgd, address); 436 if (!pud_present(*pud)) 437 return NULL; 438 439 pmd = pmd_offset(pud, address); 440 if (!pmd_present(*pmd)) 441 return NULL; 442 if (pmd_trans_huge(*pmd)) 443 return NULL; 444 445 pte = pte_offset_map(pmd, address); 446 /* Make a quick check before getting the lock */ 447 if (!sync && !pte_present(*pte)) { 448 pte_unmap(pte); 449 return NULL; 450 } 451 452 ptl = pte_lockptr(mm, pmd); 453check: 454 spin_lock(ptl); 455 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { 456 *ptlp = ptl; 457 return pte; 458 } 459 pte_unmap_unlock(pte, ptl); 460 return NULL; 461} 462 463/** 464 * page_mapped_in_vma - check whether a page is really mapped in a VMA 465 * @page: the page to test 466 * @vma: the VMA to test 467 * 468 * Returns 1 if the page is mapped into the page tables of the VMA, 0 469 * if the page is not mapped into the page tables of this VMA. Only 470 * valid for normal file or anonymous VMAs. 471 */ 472int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma) 473{ 474 unsigned long address; 475 pte_t *pte; 476 spinlock_t *ptl; 477 478 address = vma_address(page, vma); 479 if (address == -EFAULT) /* out of vma range */ 480 return 0; 481 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1); 482 if (!pte) /* the page is not in this mm */ 483 return 0; 484 pte_unmap_unlock(pte, ptl); 485 486 return 1; 487} 488 489/* 490 * Subfunctions of page_referenced: page_referenced_one called 491 * repeatedly from either page_referenced_anon or page_referenced_file. 492 */ 493int page_referenced_one(struct page *page, struct vm_area_struct *vma, 494 unsigned long address, unsigned int *mapcount, 495 unsigned long *vm_flags) 496{ 497 struct mm_struct *mm = vma->vm_mm; 498 int referenced = 0; 499 500 if (unlikely(PageTransHuge(page))) { 501 pmd_t *pmd; 502 503 spin_lock(&mm->page_table_lock); 504 /* 505 * rmap might return false positives; we must filter 506 * these out using page_check_address_pmd(). 507 */ 508 pmd = page_check_address_pmd(page, mm, address, 509 PAGE_CHECK_ADDRESS_PMD_FLAG); 510 if (!pmd) { 511 spin_unlock(&mm->page_table_lock); 512 goto out; 513 } 514 515 if (vma->vm_flags & VM_LOCKED) { 516 spin_unlock(&mm->page_table_lock); 517 *mapcount = 0; /* break early from loop */ 518 *vm_flags |= VM_LOCKED; 519 goto out; 520 } 521 522 /* go ahead even if the pmd is pmd_trans_splitting() */ 523 if (pmdp_clear_flush_young_notify(vma, address, pmd)) 524 referenced++; 525 spin_unlock(&mm->page_table_lock); 526 } else { 527 pte_t *pte; 528 spinlock_t *ptl; 529 530 /* 531 * rmap might return false positives; we must filter 532 * these out using page_check_address(). 533 */ 534 pte = page_check_address(page, mm, address, &ptl, 0); 535 if (!pte) 536 goto out; 537 538 if (vma->vm_flags & VM_LOCKED) { 539 pte_unmap_unlock(pte, ptl); 540 *mapcount = 0; /* break early from loop */ 541 *vm_flags |= VM_LOCKED; 542 goto out; 543 } 544 545 if (ptep_clear_flush_young_notify(vma, address, pte)) { 546 /* 547 * Don't treat a reference through a sequentially read 548 * mapping as such. If the page has been used in 549 * another mapping, we will catch it; if this other 550 * mapping is already gone, the unmap path will have 551 * set PG_referenced or activated the page. 552 */ 553 if (likely(!VM_SequentialReadHint(vma))) 554 referenced++; 555 } 556 pte_unmap_unlock(pte, ptl); 557 } 558 559 /* Pretend the page is referenced if the task has the 560 swap token and is in the middle of a page fault. */ 561 if (mm != current->mm && has_swap_token(mm) && 562 rwsem_is_locked(&mm->mmap_sem)) 563 referenced++; 564 565 (*mapcount)--; 566 567 if (referenced) 568 *vm_flags |= vma->vm_flags; 569out: 570 return referenced; 571} 572 573static int page_referenced_anon(struct page *page, 574 struct mem_cgroup *mem_cont, 575 unsigned long *vm_flags) 576{ 577 unsigned int mapcount; 578 struct anon_vma *anon_vma; 579 struct anon_vma_chain *avc; 580 int referenced = 0; 581 582 anon_vma = page_lock_anon_vma(page); 583 if (!anon_vma) 584 return referenced; 585 586 mapcount = page_mapcount(page); 587 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { 588 struct vm_area_struct *vma = avc->vma; 589 unsigned long address = vma_address(page, vma); 590 if (address == -EFAULT) 591 continue; 592 /* 593 * If we are reclaiming on behalf of a cgroup, skip 594 * counting on behalf of references from different 595 * cgroups 596 */ 597 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont)) 598 continue; 599 referenced += page_referenced_one(page, vma, address, 600 &mapcount, vm_flags); 601 if (!mapcount) 602 break; 603 } 604 605 page_unlock_anon_vma(anon_vma); 606 return referenced; 607} 608 609/** 610 * page_referenced_file - referenced check for object-based rmap 611 * @page: the page we're checking references on. 612 * @mem_cont: target memory controller 613 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page 614 * 615 * For an object-based mapped page, find all the places it is mapped and 616 * check/clear the referenced flag. This is done by following the page->mapping 617 * pointer, then walking the chain of vmas it holds. It returns the number 618 * of references it found. 619 * 620 * This function is only called from page_referenced for object-based pages. 621 */ 622static int page_referenced_file(struct page *page, 623 struct mem_cgroup *mem_cont, 624 unsigned long *vm_flags) 625{ 626 unsigned int mapcount; 627 struct address_space *mapping = page->mapping; 628 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 629 struct vm_area_struct *vma; 630 struct prio_tree_iter iter; 631 int referenced = 0; 632 633 /* 634 * The caller's checks on page->mapping and !PageAnon have made 635 * sure that this is a file page: the check for page->mapping 636 * excludes the case just before it gets set on an anon page. 637 */ 638 BUG_ON(PageAnon(page)); 639 640 /* 641 * The page lock not only makes sure that page->mapping cannot 642 * suddenly be NULLified by truncation, it makes sure that the 643 * structure at mapping cannot be freed and reused yet, 644 * so we can safely take mapping->i_mmap_lock. 645 */ 646 BUG_ON(!PageLocked(page)); 647 648 spin_lock(&mapping->i_mmap_lock); 649 650 /* 651 * i_mmap_lock does not stabilize mapcount at all, but mapcount 652 * is more likely to be accurate if we note it after spinning. 653 */ 654 mapcount = page_mapcount(page); 655 656 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 657 unsigned long address = vma_address(page, vma); 658 if (address == -EFAULT) 659 continue; 660 /* 661 * If we are reclaiming on behalf of a cgroup, skip 662 * counting on behalf of references from different 663 * cgroups 664 */ 665 if (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont)) 666 continue; 667 referenced += page_referenced_one(page, vma, address, 668 &mapcount, vm_flags); 669 if (!mapcount) 670 break; 671 } 672 673 spin_unlock(&mapping->i_mmap_lock); 674 return referenced; 675} 676 677/** 678 * page_referenced - test if the page was referenced 679 * @page: the page to test 680 * @is_locked: caller holds lock on the page 681 * @mem_cont: target memory controller 682 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page 683 * 684 * Quick test_and_clear_referenced for all mappings to a page, 685 * returns the number of ptes which referenced the page. 686 */ 687int page_referenced(struct page *page, 688 int is_locked, 689 struct mem_cgroup *mem_cont, 690 unsigned long *vm_flags) 691{ 692 int referenced = 0; 693 int we_locked = 0; 694 695 *vm_flags = 0; 696 if (page_mapped(page) && page_rmapping(page)) { 697 if (!is_locked && (!PageAnon(page) || PageKsm(page))) { 698 we_locked = trylock_page(page); 699 if (!we_locked) { 700 referenced++; 701 goto out; 702 } 703 } 704 if (unlikely(PageKsm(page))) 705 referenced += page_referenced_ksm(page, mem_cont, 706 vm_flags); 707 else if (PageAnon(page)) 708 referenced += page_referenced_anon(page, mem_cont, 709 vm_flags); 710 else if (page->mapping) 711 referenced += page_referenced_file(page, mem_cont, 712 vm_flags); 713 if (we_locked) 714 unlock_page(page); 715 } 716out: 717 if (page_test_and_clear_young(page)) 718 referenced++; 719 720 return referenced; 721} 722 723static int page_mkclean_one(struct page *page, struct vm_area_struct *vma, 724 unsigned long address) 725{ 726 struct mm_struct *mm = vma->vm_mm; 727 pte_t *pte; 728 spinlock_t *ptl; 729 int ret = 0; 730 731 pte = page_check_address(page, mm, address, &ptl, 1); 732 if (!pte) 733 goto out; 734 735 if (pte_dirty(*pte) || pte_write(*pte)) { 736 pte_t entry; 737 738 flush_cache_page(vma, address, pte_pfn(*pte)); 739 entry = ptep_clear_flush_notify(vma, address, pte); 740 entry = pte_wrprotect(entry); 741 entry = pte_mkclean(entry); 742 set_pte_at(mm, address, pte, entry); 743 ret = 1; 744 } 745 746 pte_unmap_unlock(pte, ptl); 747out: 748 return ret; 749} 750 751static int page_mkclean_file(struct address_space *mapping, struct page *page) 752{ 753 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 754 struct vm_area_struct *vma; 755 struct prio_tree_iter iter; 756 int ret = 0; 757 758 BUG_ON(PageAnon(page)); 759 760 spin_lock(&mapping->i_mmap_lock); 761 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 762 if (vma->vm_flags & VM_SHARED) { 763 unsigned long address = vma_address(page, vma); 764 if (address == -EFAULT) 765 continue; 766 ret += page_mkclean_one(page, vma, address); 767 } 768 } 769 spin_unlock(&mapping->i_mmap_lock); 770 return ret; 771} 772 773int page_mkclean(struct page *page) 774{ 775 int ret = 0; 776 777 BUG_ON(!PageLocked(page)); 778 779 if (page_mapped(page)) { 780 struct address_space *mapping = page_mapping(page); 781 if (mapping) { 782 ret = page_mkclean_file(mapping, page); 783 if (page_test_dirty(page)) { 784 page_clear_dirty(page, 1); 785 ret = 1; 786 } 787 } 788 } 789 790 return ret; 791} 792EXPORT_SYMBOL_GPL(page_mkclean); 793 794/** 795 * page_move_anon_rmap - move a page to our anon_vma 796 * @page: the page to move to our anon_vma 797 * @vma: the vma the page belongs to 798 * @address: the user virtual address mapped 799 * 800 * When a page belongs exclusively to one process after a COW event, 801 * that page can be moved into the anon_vma that belongs to just that 802 * process, so the rmap code will not search the parent or sibling 803 * processes. 804 */ 805void page_move_anon_rmap(struct page *page, 806 struct vm_area_struct *vma, unsigned long address) 807{ 808 struct anon_vma *anon_vma = vma->anon_vma; 809 810 VM_BUG_ON(!PageLocked(page)); 811 VM_BUG_ON(!anon_vma); 812 VM_BUG_ON(page->index != linear_page_index(vma, address)); 813 814 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 815 page->mapping = (struct address_space *) anon_vma; 816} 817 818/** 819 * __page_set_anon_rmap - set up new anonymous rmap 820 * @page: Page to add to rmap 821 * @vma: VM area to add page to. 822 * @address: User virtual address of the mapping 823 * @exclusive: the page is exclusively owned by the current process 824 */ 825static void __page_set_anon_rmap(struct page *page, 826 struct vm_area_struct *vma, unsigned long address, int exclusive) 827{ 828 struct anon_vma *anon_vma = vma->anon_vma; 829 830 BUG_ON(!anon_vma); 831 832 if (PageAnon(page)) 833 return; 834 835 /* 836 * If the page isn't exclusively mapped into this vma, 837 * we must use the _oldest_ possible anon_vma for the 838 * page mapping! 839 */ 840 if (!exclusive) 841 anon_vma = anon_vma->root; 842 843 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 844 page->mapping = (struct address_space *) anon_vma; 845 page->index = linear_page_index(vma, address); 846} 847 848/** 849 * __page_check_anon_rmap - sanity check anonymous rmap addition 850 * @page: the page to add the mapping to 851 * @vma: the vm area in which the mapping is added 852 * @address: the user virtual address mapped 853 */ 854static void __page_check_anon_rmap(struct page *page, 855 struct vm_area_struct *vma, unsigned long address) 856{ 857#ifdef CONFIG_DEBUG_VM 858 /* 859 * The page's anon-rmap details (mapping and index) are guaranteed to 860 * be set up correctly at this point. 861 * 862 * We have exclusion against page_add_anon_rmap because the caller 863 * always holds the page locked, except if called from page_dup_rmap, 864 * in which case the page is already known to be setup. 865 * 866 * We have exclusion against page_add_new_anon_rmap because those pages 867 * are initially only visible via the pagetables, and the pte is locked 868 * over the call to page_add_new_anon_rmap. 869 */ 870 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root); 871 BUG_ON(page->index != linear_page_index(vma, address)); 872#endif 873} 874 875/** 876 * page_add_anon_rmap - add pte mapping to an anonymous page 877 * @page: the page to add the mapping to 878 * @vma: the vm area in which the mapping is added 879 * @address: the user virtual address mapped 880 * 881 * The caller needs to hold the pte lock, and the page must be locked in 882 * the anon_vma case: to serialize mapping,index checking after setting, 883 * and to ensure that PageAnon is not being upgraded racily to PageKsm 884 * (but PageKsm is never downgraded to PageAnon). 885 */ 886void page_add_anon_rmap(struct page *page, 887 struct vm_area_struct *vma, unsigned long address) 888{ 889 do_page_add_anon_rmap(page, vma, address, 0); 890} 891 892/* 893 * Special version of the above for do_swap_page, which often runs 894 * into pages that are exclusively owned by the current process. 895 * Everybody else should continue to use page_add_anon_rmap above. 896 */ 897void do_page_add_anon_rmap(struct page *page, 898 struct vm_area_struct *vma, unsigned long address, int exclusive) 899{ 900 int first = atomic_inc_and_test(&page->_mapcount); 901 if (first) { 902 if (!PageTransHuge(page)) 903 __inc_zone_page_state(page, NR_ANON_PAGES); 904 else 905 __inc_zone_page_state(page, 906 NR_ANON_TRANSPARENT_HUGEPAGES); 907 } 908 if (unlikely(PageKsm(page))) 909 return; 910 911 VM_BUG_ON(!PageLocked(page)); 912 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); 913 if (first) 914 __page_set_anon_rmap(page, vma, address, exclusive); 915 else 916 __page_check_anon_rmap(page, vma, address); 917} 918 919/** 920 * page_add_new_anon_rmap - add pte mapping to a new anonymous page 921 * @page: the page to add the mapping to 922 * @vma: the vm area in which the mapping is added 923 * @address: the user virtual address mapped 924 * 925 * Same as page_add_anon_rmap but must only be called on *new* pages. 926 * This means the inc-and-test can be bypassed. 927 * Page does not have to be locked. 928 */ 929void page_add_new_anon_rmap(struct page *page, 930 struct vm_area_struct *vma, unsigned long address) 931{ 932 VM_BUG_ON(address < vma->vm_start || address >= vma->vm_end); 933 SetPageSwapBacked(page); 934 atomic_set(&page->_mapcount, 0); /* increment count (starts at -1) */ 935 if (!PageTransHuge(page)) 936 __inc_zone_page_state(page, NR_ANON_PAGES); 937 else 938 __inc_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES); 939 __page_set_anon_rmap(page, vma, address, 1); 940 if (page_evictable(page, vma)) 941 lru_cache_add_lru(page, LRU_ACTIVE_ANON); 942 else 943 add_page_to_unevictable_list(page); 944} 945 946/** 947 * page_add_file_rmap - add pte mapping to a file page 948 * @page: the page to add the mapping to 949 * 950 * The caller needs to hold the pte lock. 951 */ 952void page_add_file_rmap(struct page *page) 953{ 954 if (atomic_inc_and_test(&page->_mapcount)) { 955 __inc_zone_page_state(page, NR_FILE_MAPPED); 956 mem_cgroup_inc_page_stat(page, MEMCG_NR_FILE_MAPPED); 957 } 958} 959 960/** 961 * page_remove_rmap - take down pte mapping from a page 962 * @page: page to remove mapping from 963 * 964 * The caller needs to hold the pte lock. 965 */ 966void page_remove_rmap(struct page *page) 967{ 968 /* page still mapped by someone else? */ 969 if (!atomic_add_negative(-1, &page->_mapcount)) 970 return; 971 972 /* 973 * Now that the last pte has gone, s390 must transfer dirty 974 * flag from storage key to struct page. We can usually skip 975 * this if the page is anon, so about to be freed; but perhaps 976 * not if it's in swapcache - there might be another pte slot 977 * containing the swap entry, but page not yet written to swap. 978 */ 979 if ((!PageAnon(page) || PageSwapCache(page)) && page_test_dirty(page)) { 980 page_clear_dirty(page, 1); 981 set_page_dirty(page); 982 } 983 /* 984 * Hugepages are not counted in NR_ANON_PAGES nor NR_FILE_MAPPED 985 * and not charged by memcg for now. 986 */ 987 if (unlikely(PageHuge(page))) 988 return; 989 if (PageAnon(page)) { 990 mem_cgroup_uncharge_page(page); 991 if (!PageTransHuge(page)) 992 __dec_zone_page_state(page, NR_ANON_PAGES); 993 else 994 __dec_zone_page_state(page, 995 NR_ANON_TRANSPARENT_HUGEPAGES); 996 } else { 997 __dec_zone_page_state(page, NR_FILE_MAPPED); 998 mem_cgroup_dec_page_stat(page, MEMCG_NR_FILE_MAPPED); 999 } 1000 /* 1001 * It would be tidy to reset the PageAnon mapping here, 1002 * but that might overwrite a racing page_add_anon_rmap 1003 * which increments mapcount after us but sets mapping 1004 * before us: so leave the reset to free_hot_cold_page, 1005 * and remember that it's only reliable while mapped. 1006 * Leaving it set also helps swapoff to reinstate ptes 1007 * faster for those pages still in swapcache. 1008 */ 1009} 1010 1011/* 1012 * Subfunctions of try_to_unmap: try_to_unmap_one called 1013 * repeatedly from either try_to_unmap_anon or try_to_unmap_file. 1014 */ 1015int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, 1016 unsigned long address, enum ttu_flags flags) 1017{ 1018 struct mm_struct *mm = vma->vm_mm; 1019 pte_t *pte; 1020 pte_t pteval; 1021 spinlock_t *ptl; 1022 int ret = SWAP_AGAIN; 1023 1024 pte = page_check_address(page, mm, address, &ptl, 0); 1025 if (!pte) 1026 goto out; 1027 1028 /* 1029 * If the page is mlock()d, we cannot swap it out. 1030 * If it's recently referenced (perhaps page_referenced 1031 * skipped over this mm) then we should reactivate it. 1032 */ 1033 if (!(flags & TTU_IGNORE_MLOCK)) { 1034 if (vma->vm_flags & VM_LOCKED) 1035 goto out_mlock; 1036 1037 if (TTU_ACTION(flags) == TTU_MUNLOCK) 1038 goto out_unmap; 1039 } 1040 if (!(flags & TTU_IGNORE_ACCESS)) { 1041 if (ptep_clear_flush_young_notify(vma, address, pte)) { 1042 ret = SWAP_FAIL; 1043 goto out_unmap; 1044 } 1045 } 1046 1047 /* Nuke the page table entry. */ 1048 flush_cache_page(vma, address, page_to_pfn(page)); 1049 pteval = ptep_clear_flush_notify(vma, address, pte); 1050 1051 /* Move the dirty bit to the physical page now the pte is gone. */ 1052 if (pte_dirty(pteval)) 1053 set_page_dirty(page); 1054 1055 /* Update high watermark before we lower rss */ 1056 update_hiwater_rss(mm); 1057 1058 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { 1059 if (PageAnon(page)) 1060 dec_mm_counter(mm, MM_ANONPAGES); 1061 else 1062 dec_mm_counter(mm, MM_FILEPAGES); 1063 set_pte_at(mm, address, pte, 1064 swp_entry_to_pte(make_hwpoison_entry(page))); 1065 } else if (PageAnon(page)) { 1066 swp_entry_t entry = { .val = page_private(page) }; 1067 1068 if (PageSwapCache(page)) { 1069 /* 1070 * Store the swap location in the pte. 1071 * See handle_pte_fault() ... 1072 */ 1073 if (swap_duplicate(entry) < 0) { 1074 set_pte_at(mm, address, pte, pteval); 1075 ret = SWAP_FAIL; 1076 goto out_unmap; 1077 } 1078 if (list_empty(&mm->mmlist)) { 1079 spin_lock(&mmlist_lock); 1080 if (list_empty(&mm->mmlist)) 1081 list_add(&mm->mmlist, &init_mm.mmlist); 1082 spin_unlock(&mmlist_lock); 1083 } 1084 dec_mm_counter(mm, MM_ANONPAGES); 1085 inc_mm_counter(mm, MM_SWAPENTS); 1086 } else if (PAGE_MIGRATION) { 1087 /* 1088 * Store the pfn of the page in a special migration 1089 * pte. do_swap_page() will wait until the migration 1090 * pte is removed and then restart fault handling. 1091 */ 1092 BUG_ON(TTU_ACTION(flags) != TTU_MIGRATION); 1093 entry = make_migration_entry(page, pte_write(pteval)); 1094 } 1095 set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); 1096 BUG_ON(pte_file(*pte)); 1097 } else if (PAGE_MIGRATION && (TTU_ACTION(flags) == TTU_MIGRATION)) { 1098 /* Establish migration entry for a file page */ 1099 swp_entry_t entry; 1100 entry = make_migration_entry(page, pte_write(pteval)); 1101 set_pte_at(mm, address, pte, swp_entry_to_pte(entry)); 1102 } else 1103 dec_mm_counter(mm, MM_FILEPAGES); 1104 1105 page_remove_rmap(page); 1106 page_cache_release(page); 1107 1108out_unmap: 1109 pte_unmap_unlock(pte, ptl); 1110out: 1111 return ret; 1112 1113out_mlock: 1114 pte_unmap_unlock(pte, ptl); 1115 1116 1117 /* 1118 * We need mmap_sem locking, Otherwise VM_LOCKED check makes 1119 * unstable result and race. Plus, We can't wait here because 1120 * we now hold anon_vma->lock or mapping->i_mmap_lock. 1121 * if trylock failed, the page remain in evictable lru and later 1122 * vmscan could retry to move the page to unevictable lru if the 1123 * page is actually mlocked. 1124 */ 1125 if (down_read_trylock(&vma->vm_mm->mmap_sem)) { 1126 if (vma->vm_flags & VM_LOCKED) { 1127 mlock_vma_page(page); 1128 ret = SWAP_MLOCK; 1129 } 1130 up_read(&vma->vm_mm->mmap_sem); 1131 } 1132 return ret; 1133} 1134 1135/* 1136 * objrmap doesn't work for nonlinear VMAs because the assumption that 1137 * offset-into-file correlates with offset-into-virtual-addresses does not hold. 1138 * Consequently, given a particular page and its ->index, we cannot locate the 1139 * ptes which are mapping that page without an exhaustive linear search. 1140 * 1141 * So what this code does is a mini "virtual scan" of each nonlinear VMA which 1142 * maps the file to which the target page belongs. The ->vm_private_data field 1143 * holds the current cursor into that scan. Successive searches will circulate 1144 * around the vma's virtual address space. 1145 * 1146 * So as more replacement pressure is applied to the pages in a nonlinear VMA, 1147 * more scanning pressure is placed against them as well. Eventually pages 1148 * will become fully unmapped and are eligible for eviction. 1149 * 1150 * For very sparsely populated VMAs this is a little inefficient - chances are 1151 * there there won't be many ptes located within the scan cluster. In this case 1152 * maybe we could scan further - to the end of the pte page, perhaps. 1153 * 1154 * Mlocked pages: check VM_LOCKED under mmap_sem held for read, if we can 1155 * acquire it without blocking. If vma locked, mlock the pages in the cluster, 1156 * rather than unmapping them. If we encounter the "check_page" that vmscan is 1157 * trying to unmap, return SWAP_MLOCK, else default SWAP_AGAIN. 1158 */ 1159#define CLUSTER_SIZE min(32*PAGE_SIZE, PMD_SIZE) 1160#define CLUSTER_MASK (~(CLUSTER_SIZE - 1)) 1161 1162static int try_to_unmap_cluster(unsigned long cursor, unsigned int *mapcount, 1163 struct vm_area_struct *vma, struct page *check_page) 1164{ 1165 struct mm_struct *mm = vma->vm_mm; 1166 pgd_t *pgd; 1167 pud_t *pud; 1168 pmd_t *pmd; 1169 pte_t *pte; 1170 pte_t pteval; 1171 spinlock_t *ptl; 1172 struct page *page; 1173 unsigned long address; 1174 unsigned long end; 1175 int ret = SWAP_AGAIN; 1176 int locked_vma = 0; 1177 1178 address = (vma->vm_start + cursor) & CLUSTER_MASK; 1179 end = address + CLUSTER_SIZE; 1180 if (address < vma->vm_start) 1181 address = vma->vm_start; 1182 if (end > vma->vm_end) 1183 end = vma->vm_end; 1184 1185 pgd = pgd_offset(mm, address); 1186 if (!pgd_present(*pgd)) 1187 return ret; 1188 1189 pud = pud_offset(pgd, address); 1190 if (!pud_present(*pud)) 1191 return ret; 1192 1193 pmd = pmd_offset(pud, address); 1194 if (!pmd_present(*pmd)) 1195 return ret; 1196 1197 /* 1198 * If we can acquire the mmap_sem for read, and vma is VM_LOCKED, 1199 * keep the sem while scanning the cluster for mlocking pages. 1200 */ 1201 if (down_read_trylock(&vma->vm_mm->mmap_sem)) { 1202 locked_vma = (vma->vm_flags & VM_LOCKED); 1203 if (!locked_vma) 1204 up_read(&vma->vm_mm->mmap_sem); /* don't need it */ 1205 } 1206 1207 pte = pte_offset_map_lock(mm, pmd, address, &ptl); 1208 1209 /* Update high watermark before we lower rss */ 1210 update_hiwater_rss(mm); 1211 1212 for (; address < end; pte++, address += PAGE_SIZE) { 1213 if (!pte_present(*pte)) 1214 continue; 1215 page = vm_normal_page(vma, address, *pte); 1216 BUG_ON(!page || PageAnon(page)); 1217 1218 if (locked_vma) { 1219 mlock_vma_page(page); /* no-op if already mlocked */ 1220 if (page == check_page) 1221 ret = SWAP_MLOCK; 1222 continue; /* don't unmap */ 1223 } 1224 1225 if (ptep_clear_flush_young_notify(vma, address, pte)) 1226 continue; 1227 1228 /* Nuke the page table entry. */ 1229 flush_cache_page(vma, address, pte_pfn(*pte)); 1230 pteval = ptep_clear_flush_notify(vma, address, pte); 1231 1232 /* If nonlinear, store the file page offset in the pte. */ 1233 if (page->index != linear_page_index(vma, address)) 1234 set_pte_at(mm, address, pte, pgoff_to_pte(page->index)); 1235 1236 /* Move the dirty bit to the physical page now the pte is gone. */ 1237 if (pte_dirty(pteval)) 1238 set_page_dirty(page); 1239 1240 page_remove_rmap(page); 1241 page_cache_release(page); 1242 dec_mm_counter(mm, MM_FILEPAGES); 1243 (*mapcount)--; 1244 } 1245 pte_unmap_unlock(pte - 1, ptl); 1246 if (locked_vma) 1247 up_read(&vma->vm_mm->mmap_sem); 1248 return ret; 1249} 1250 1251bool is_vma_temporary_stack(struct vm_area_struct *vma) 1252{ 1253 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP); 1254 1255 if (!maybe_stack) 1256 return false; 1257 1258 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) == 1259 VM_STACK_INCOMPLETE_SETUP) 1260 return true; 1261 1262 return false; 1263} 1264 1265/** 1266 * try_to_unmap_anon - unmap or unlock anonymous page using the object-based 1267 * rmap method 1268 * @page: the page to unmap/unlock 1269 * @flags: action and flags 1270 * 1271 * Find all the mappings of a page using the mapping pointer and the vma chains 1272 * contained in the anon_vma struct it points to. 1273 * 1274 * This function is only called from try_to_unmap/try_to_munlock for 1275 * anonymous pages. 1276 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma 1277 * where the page was found will be held for write. So, we won't recheck 1278 * vm_flags for that VMA. That should be OK, because that vma shouldn't be 1279 * 'LOCKED. 1280 */ 1281static int try_to_unmap_anon(struct page *page, enum ttu_flags flags) 1282{ 1283 struct anon_vma *anon_vma; 1284 struct anon_vma_chain *avc; 1285 int ret = SWAP_AGAIN; 1286 1287 anon_vma = page_lock_anon_vma(page); 1288 if (!anon_vma) 1289 return ret; 1290 1291 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { 1292 struct vm_area_struct *vma = avc->vma; 1293 unsigned long address; 1294 1295 /* 1296 * During exec, a temporary VMA is setup and later moved. 1297 * The VMA is moved under the anon_vma lock but not the 1298 * page tables leading to a race where migration cannot 1299 * find the migration ptes. Rather than increasing the 1300 * locking requirements of exec(), migration skips 1301 * temporary VMAs until after exec() completes. 1302 */ 1303 if (PAGE_MIGRATION && (flags & TTU_MIGRATION) && 1304 is_vma_temporary_stack(vma)) 1305 continue; 1306 1307 address = vma_address(page, vma); 1308 if (address == -EFAULT) 1309 continue; 1310 ret = try_to_unmap_one(page, vma, address, flags); 1311 if (ret != SWAP_AGAIN || !page_mapped(page)) 1312 break; 1313 } 1314 1315 page_unlock_anon_vma(anon_vma); 1316 return ret; 1317} 1318 1319/** 1320 * try_to_unmap_file - unmap/unlock file page using the object-based rmap method 1321 * @page: the page to unmap/unlock 1322 * @flags: action and flags 1323 * 1324 * Find all the mappings of a page using the mapping pointer and the vma chains 1325 * contained in the address_space struct it points to. 1326 * 1327 * This function is only called from try_to_unmap/try_to_munlock for 1328 * object-based pages. 1329 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma 1330 * where the page was found will be held for write. So, we won't recheck 1331 * vm_flags for that VMA. That should be OK, because that vma shouldn't be 1332 * 'LOCKED. 1333 */ 1334static int try_to_unmap_file(struct page *page, enum ttu_flags flags) 1335{ 1336 struct address_space *mapping = page->mapping; 1337 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 1338 struct vm_area_struct *vma; 1339 struct prio_tree_iter iter; 1340 int ret = SWAP_AGAIN; 1341 unsigned long cursor; 1342 unsigned long max_nl_cursor = 0; 1343 unsigned long max_nl_size = 0; 1344 unsigned int mapcount; 1345 1346 spin_lock(&mapping->i_mmap_lock); 1347 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 1348 unsigned long address = vma_address(page, vma); 1349 if (address == -EFAULT) 1350 continue; 1351 ret = try_to_unmap_one(page, vma, address, flags); 1352 if (ret != SWAP_AGAIN || !page_mapped(page)) 1353 goto out; 1354 } 1355 1356 if (list_empty(&mapping->i_mmap_nonlinear)) 1357 goto out; 1358 1359 /* 1360 * We don't bother to try to find the munlocked page in nonlinears. 1361 * It's costly. Instead, later, page reclaim logic may call 1362 * try_to_unmap(TTU_MUNLOCK) and recover PG_mlocked lazily. 1363 */ 1364 if (TTU_ACTION(flags) == TTU_MUNLOCK) 1365 goto out; 1366 1367 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, 1368 shared.vm_set.list) { 1369 cursor = (unsigned long) vma->vm_private_data; 1370 if (cursor > max_nl_cursor) 1371 max_nl_cursor = cursor; 1372 cursor = vma->vm_end - vma->vm_start; 1373 if (cursor > max_nl_size) 1374 max_nl_size = cursor; 1375 } 1376 1377 if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */ 1378 ret = SWAP_FAIL; 1379 goto out; 1380 } 1381 1382 /* 1383 * We don't try to search for this page in the nonlinear vmas, 1384 * and page_referenced wouldn't have found it anyway. Instead 1385 * just walk the nonlinear vmas trying to age and unmap some. 1386 * The mapcount of the page we came in with is irrelevant, 1387 * but even so use it as a guide to how hard we should try? 1388 */ 1389 mapcount = page_mapcount(page); 1390 if (!mapcount) 1391 goto out; 1392 cond_resched_lock(&mapping->i_mmap_lock); 1393 1394 max_nl_size = (max_nl_size + CLUSTER_SIZE - 1) & CLUSTER_MASK; 1395 if (max_nl_cursor == 0) 1396 max_nl_cursor = CLUSTER_SIZE; 1397 1398 do { 1399 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, 1400 shared.vm_set.list) { 1401 cursor = (unsigned long) vma->vm_private_data; 1402 while ( cursor < max_nl_cursor && 1403 cursor < vma->vm_end - vma->vm_start) { 1404 if (try_to_unmap_cluster(cursor, &mapcount, 1405 vma, page) == SWAP_MLOCK) 1406 ret = SWAP_MLOCK; 1407 cursor += CLUSTER_SIZE; 1408 vma->vm_private_data = (void *) cursor; 1409 if ((int)mapcount <= 0) 1410 goto out; 1411 } 1412 vma->vm_private_data = (void *) max_nl_cursor; 1413 } 1414 cond_resched_lock(&mapping->i_mmap_lock); 1415 max_nl_cursor += CLUSTER_SIZE; 1416 } while (max_nl_cursor <= max_nl_size); 1417 1418 /* 1419 * Don't loop forever (perhaps all the remaining pages are 1420 * in locked vmas). Reset cursor on all unreserved nonlinear 1421 * vmas, now forgetting on which ones it had fallen behind. 1422 */ 1423 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list) 1424 vma->vm_private_data = NULL; 1425out: 1426 spin_unlock(&mapping->i_mmap_lock); 1427 return ret; 1428} 1429 1430/** 1431 * try_to_unmap - try to remove all page table mappings to a page 1432 * @page: the page to get unmapped 1433 * @flags: action and flags 1434 * 1435 * Tries to remove all the page table entries which are mapping this 1436 * page, used in the pageout path. Caller must hold the page lock. 1437 * Return values are: 1438 * 1439 * SWAP_SUCCESS - we succeeded in removing all mappings 1440 * SWAP_AGAIN - we missed a mapping, try again later 1441 * SWAP_FAIL - the page is unswappable 1442 * SWAP_MLOCK - page is mlocked. 1443 */ 1444int try_to_unmap(struct page *page, enum ttu_flags flags) 1445{ 1446 int ret; 1447 1448 BUG_ON(!PageLocked(page)); 1449 VM_BUG_ON(!PageHuge(page) && PageTransHuge(page)); 1450 1451 if (unlikely(PageKsm(page))) 1452 ret = try_to_unmap_ksm(page, flags); 1453 else if (PageAnon(page)) 1454 ret = try_to_unmap_anon(page, flags); 1455 else 1456 ret = try_to_unmap_file(page, flags); 1457 if (ret != SWAP_MLOCK && !page_mapped(page)) 1458 ret = SWAP_SUCCESS; 1459 return ret; 1460} 1461 1462/** 1463 * try_to_munlock - try to munlock a page 1464 * @page: the page to be munlocked 1465 * 1466 * Called from munlock code. Checks all of the VMAs mapping the page 1467 * to make sure nobody else has this page mlocked. The page will be 1468 * returned with PG_mlocked cleared if no other vmas have it mlocked. 1469 * 1470 * Return values are: 1471 * 1472 * SWAP_AGAIN - no vma is holding page mlocked, or, 1473 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem 1474 * SWAP_FAIL - page cannot be located at present 1475 * SWAP_MLOCK - page is now mlocked. 1476 */ 1477int try_to_munlock(struct page *page) 1478{ 1479 VM_BUG_ON(!PageLocked(page) || PageLRU(page)); 1480 1481 if (unlikely(PageKsm(page))) 1482 return try_to_unmap_ksm(page, TTU_MUNLOCK); 1483 else if (PageAnon(page)) 1484 return try_to_unmap_anon(page, TTU_MUNLOCK); 1485 else 1486 return try_to_unmap_file(page, TTU_MUNLOCK); 1487} 1488 1489#if defined(CONFIG_KSM) || defined(CONFIG_MIGRATION) 1490/* 1491 * Drop an anon_vma refcount, freeing the anon_vma and anon_vma->root 1492 * if necessary. Be careful to do all the tests under the lock. Once 1493 * we know we are the last user, nobody else can get a reference and we 1494 * can do the freeing without the lock. 1495 */ 1496void drop_anon_vma(struct anon_vma *anon_vma) 1497{ 1498 BUG_ON(atomic_read(&anon_vma->external_refcount) <= 0); 1499 if (atomic_dec_and_lock(&anon_vma->external_refcount, &anon_vma->root->lock)) { 1500 struct anon_vma *root = anon_vma->root; 1501 int empty = list_empty(&anon_vma->head); 1502 int last_root_user = 0; 1503 int root_empty = 0; 1504 1505 /* 1506 * The refcount on a non-root anon_vma got dropped. Drop 1507 * the refcount on the root and check if we need to free it. 1508 */ 1509 if (empty && anon_vma != root) { 1510 BUG_ON(atomic_read(&root->external_refcount) <= 0); 1511 last_root_user = atomic_dec_and_test(&root->external_refcount); 1512 root_empty = list_empty(&root->head); 1513 } 1514 anon_vma_unlock(anon_vma); 1515 1516 if (empty) { 1517 anon_vma_free(anon_vma); 1518 if (root_empty && last_root_user) 1519 anon_vma_free(root); 1520 } 1521 } 1522} 1523#endif 1524 1525#ifdef CONFIG_MIGRATION 1526/* 1527 * rmap_walk() and its helpers rmap_walk_anon() and rmap_walk_file(): 1528 * Called by migrate.c to remove migration ptes, but might be used more later. 1529 */ 1530static int rmap_walk_anon(struct page *page, int (*rmap_one)(struct page *, 1531 struct vm_area_struct *, unsigned long, void *), void *arg) 1532{ 1533 struct anon_vma *anon_vma; 1534 struct anon_vma_chain *avc; 1535 int ret = SWAP_AGAIN; 1536 1537 /* 1538 * Note: remove_migration_ptes() cannot use page_lock_anon_vma() 1539 * because that depends on page_mapped(); but not all its usages 1540 * are holding mmap_sem. Users without mmap_sem are required to 1541 * take a reference count to prevent the anon_vma disappearing 1542 */ 1543 anon_vma = page_anon_vma(page); 1544 if (!anon_vma) 1545 return ret; 1546 anon_vma_lock(anon_vma); 1547 list_for_each_entry(avc, &anon_vma->head, same_anon_vma) { 1548 struct vm_area_struct *vma = avc->vma; 1549 unsigned long address = vma_address(page, vma); 1550 if (address == -EFAULT) 1551 continue; 1552 ret = rmap_one(page, vma, address, arg); 1553 if (ret != SWAP_AGAIN) 1554 break; 1555 } 1556 anon_vma_unlock(anon_vma); 1557 return ret; 1558} 1559 1560static int rmap_walk_file(struct page *page, int (*rmap_one)(struct page *, 1561 struct vm_area_struct *, unsigned long, void *), void *arg) 1562{ 1563 struct address_space *mapping = page->mapping; 1564 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT); 1565 struct vm_area_struct *vma; 1566 struct prio_tree_iter iter; 1567 int ret = SWAP_AGAIN; 1568 1569 if (!mapping) 1570 return ret; 1571 spin_lock(&mapping->i_mmap_lock); 1572 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff, pgoff) { 1573 unsigned long address = vma_address(page, vma); 1574 if (address == -EFAULT) 1575 continue; 1576 ret = rmap_one(page, vma, address, arg); 1577 if (ret != SWAP_AGAIN) 1578 break; 1579 } 1580 /* 1581 * No nonlinear handling: being always shared, nonlinear vmas 1582 * never contain migration ptes. Decide what to do about this 1583 * limitation to linear when we need rmap_walk() on nonlinear. 1584 */ 1585 spin_unlock(&mapping->i_mmap_lock); 1586 return ret; 1587} 1588 1589int rmap_walk(struct page *page, int (*rmap_one)(struct page *, 1590 struct vm_area_struct *, unsigned long, void *), void *arg) 1591{ 1592 VM_BUG_ON(!PageLocked(page)); 1593 1594 if (unlikely(PageKsm(page))) 1595 return rmap_walk_ksm(page, rmap_one, arg); 1596 else if (PageAnon(page)) 1597 return rmap_walk_anon(page, rmap_one, arg); 1598 else 1599 return rmap_walk_file(page, rmap_one, arg); 1600} 1601#endif /* CONFIG_MIGRATION */ 1602 1603#ifdef CONFIG_HUGETLB_PAGE 1604/* 1605 * The following three functions are for anonymous (private mapped) hugepages. 1606 * Unlike common anonymous pages, anonymous hugepages have no accounting code 1607 * and no lru code, because we handle hugepages differently from common pages. 1608 */ 1609static void __hugepage_set_anon_rmap(struct page *page, 1610 struct vm_area_struct *vma, unsigned long address, int exclusive) 1611{ 1612 struct anon_vma *anon_vma = vma->anon_vma; 1613 1614 BUG_ON(!anon_vma); 1615 1616 if (PageAnon(page)) 1617 return; 1618 if (!exclusive) 1619 anon_vma = anon_vma->root; 1620 1621 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 1622 page->mapping = (struct address_space *) anon_vma; 1623 page->index = linear_page_index(vma, address); 1624} 1625 1626void hugepage_add_anon_rmap(struct page *page, 1627 struct vm_area_struct *vma, unsigned long address) 1628{ 1629 struct anon_vma *anon_vma = vma->anon_vma; 1630 int first; 1631 1632 BUG_ON(!PageLocked(page)); 1633 BUG_ON(!anon_vma); 1634 BUG_ON(address < vma->vm_start || address >= vma->vm_end); 1635 first = atomic_inc_and_test(&page->_mapcount); 1636 if (first) 1637 __hugepage_set_anon_rmap(page, vma, address, 0); 1638} 1639 1640void hugepage_add_new_anon_rmap(struct page *page, 1641 struct vm_area_struct *vma, unsigned long address) 1642{ 1643 BUG_ON(address < vma->vm_start || address >= vma->vm_end); 1644 atomic_set(&page->_mapcount, 0); 1645 __hugepage_set_anon_rmap(page, vma, address, 1); 1646} 1647#endif /* CONFIG_HUGETLB_PAGE */ 1648