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 * mm->mmap_sem 25 * page->flags PG_locked (lock_page) 26 * hugetlbfs_i_mmap_rwsem_key (in huge_pmd_share) 27 * mapping->i_mmap_rwsem 28 * anon_vma->rwsem 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 * mem_cgroup_{begin,end}_page_stat (memcg->move_lock) 35 * mapping->tree_lock (widely used) 36 * inode->i_lock (in set_page_dirty's __mark_inode_dirty) 37 * bdi.wb->list_lock (in set_page_dirty's __mark_inode_dirty) 38 * sb_lock (within inode_lock in fs/fs-writeback.c) 39 * mapping->tree_lock (widely used, in set_page_dirty, 40 * in arch-dependent flush_dcache_mmap_lock, 41 * within bdi.wb->list_lock in __sync_single_inode) 42 * 43 * anon_vma->rwsem,mapping->i_mutex (memory_failure, collect_procs_anon) 44 * ->tasklist_lock 45 * pte map lock 46 */ 47 48 #include <linux/mm.h> 49 #include <linux/pagemap.h> 50 #include <linux/swap.h> 51 #include <linux/swapops.h> 52 #include <linux/slab.h> 53 #include <linux/init.h> 54 #include <linux/ksm.h> 55 #include <linux/rmap.h> 56 #include <linux/rcupdate.h> 57 #include <linux/export.h> 58 #include <linux/memcontrol.h> 59 #include <linux/mmu_notifier.h> 60 #include <linux/migrate.h> 61 #include <linux/hugetlb.h> 62 #include <linux/backing-dev.h> 63 #include <linux/page_idle.h> 64 65 #include <asm/tlbflush.h> 66 67 #include <trace/events/tlb.h> 68 69 #include "internal.h" 70 71 static struct kmem_cache *anon_vma_cachep; 72 static struct kmem_cache *anon_vma_chain_cachep; 73 74 static inline struct anon_vma *anon_vma_alloc(void) 75 { 76 struct anon_vma *anon_vma; 77 78 anon_vma = kmem_cache_alloc(anon_vma_cachep, GFP_KERNEL); 79 if (anon_vma) { 80 atomic_set(&anon_vma->refcount, 1); 81 anon_vma->degree = 1; /* Reference for first vma */ 82 anon_vma->parent = anon_vma; 83 /* 84 * Initialise the anon_vma root to point to itself. If called 85 * from fork, the root will be reset to the parents anon_vma. 86 */ 87 anon_vma->root = anon_vma; 88 } 89 90 return anon_vma; 91 } 92 93 static inline void anon_vma_free(struct anon_vma *anon_vma) 94 { 95 VM_BUG_ON(atomic_read(&anon_vma->refcount)); 96 97 /* 98 * Synchronize against page_lock_anon_vma_read() such that 99 * we can safely hold the lock without the anon_vma getting 100 * freed. 101 * 102 * Relies on the full mb implied by the atomic_dec_and_test() from 103 * put_anon_vma() against the acquire barrier implied by 104 * down_read_trylock() from page_lock_anon_vma_read(). This orders: 105 * 106 * page_lock_anon_vma_read() VS put_anon_vma() 107 * down_read_trylock() atomic_dec_and_test() 108 * LOCK MB 109 * atomic_read() rwsem_is_locked() 110 * 111 * LOCK should suffice since the actual taking of the lock must 112 * happen _before_ what follows. 113 */ 114 might_sleep(); 115 if (rwsem_is_locked(&anon_vma->root->rwsem)) { 116 anon_vma_lock_write(anon_vma); 117 anon_vma_unlock_write(anon_vma); 118 } 119 120 kmem_cache_free(anon_vma_cachep, anon_vma); 121 } 122 123 static inline struct anon_vma_chain *anon_vma_chain_alloc(gfp_t gfp) 124 { 125 return kmem_cache_alloc(anon_vma_chain_cachep, gfp); 126 } 127 128 static void anon_vma_chain_free(struct anon_vma_chain *anon_vma_chain) 129 { 130 kmem_cache_free(anon_vma_chain_cachep, anon_vma_chain); 131 } 132 133 static void anon_vma_chain_link(struct vm_area_struct *vma, 134 struct anon_vma_chain *avc, 135 struct anon_vma *anon_vma) 136 { 137 avc->vma = vma; 138 avc->anon_vma = anon_vma; 139 list_add(&avc->same_vma, &vma->anon_vma_chain); 140 anon_vma_interval_tree_insert(avc, &anon_vma->rb_root); 141 } 142 143 /** 144 * __anon_vma_prepare - attach an anon_vma to a memory region 145 * @vma: the memory region in question 146 * 147 * This makes sure the memory mapping described by 'vma' has 148 * an 'anon_vma' attached to it, so that we can associate the 149 * anonymous pages mapped into it with that anon_vma. 150 * 151 * The common case will be that we already have one, which 152 * is handled inline by anon_vma_prepare(). But if 153 * not we either need to find an adjacent mapping that we 154 * can re-use the anon_vma from (very common when the only 155 * reason for splitting a vma has been mprotect()), or we 156 * allocate a new one. 157 * 158 * Anon-vma allocations are very subtle, because we may have 159 * optimistically looked up an anon_vma in page_lock_anon_vma_read() 160 * and that may actually touch the spinlock even in the newly 161 * allocated vma (it depends on RCU to make sure that the 162 * anon_vma isn't actually destroyed). 163 * 164 * As a result, we need to do proper anon_vma locking even 165 * for the new allocation. At the same time, we do not want 166 * to do any locking for the common case of already having 167 * an anon_vma. 168 * 169 * This must be called with the mmap_sem held for reading. 170 */ 171 int __anon_vma_prepare(struct vm_area_struct *vma) 172 { 173 struct mm_struct *mm = vma->vm_mm; 174 struct anon_vma *anon_vma, *allocated; 175 struct anon_vma_chain *avc; 176 177 might_sleep(); 178 179 avc = anon_vma_chain_alloc(GFP_KERNEL); 180 if (!avc) 181 goto out_enomem; 182 183 anon_vma = find_mergeable_anon_vma(vma); 184 allocated = NULL; 185 if (!anon_vma) { 186 anon_vma = anon_vma_alloc(); 187 if (unlikely(!anon_vma)) 188 goto out_enomem_free_avc; 189 allocated = anon_vma; 190 } 191 192 anon_vma_lock_write(anon_vma); 193 /* page_table_lock to protect against threads */ 194 spin_lock(&mm->page_table_lock); 195 if (likely(!vma->anon_vma)) { 196 vma->anon_vma = anon_vma; 197 anon_vma_chain_link(vma, avc, anon_vma); 198 /* vma reference or self-parent link for new root */ 199 anon_vma->degree++; 200 allocated = NULL; 201 avc = NULL; 202 } 203 spin_unlock(&mm->page_table_lock); 204 anon_vma_unlock_write(anon_vma); 205 206 if (unlikely(allocated)) 207 put_anon_vma(allocated); 208 if (unlikely(avc)) 209 anon_vma_chain_free(avc); 210 211 return 0; 212 213 out_enomem_free_avc: 214 anon_vma_chain_free(avc); 215 out_enomem: 216 return -ENOMEM; 217 } 218 219 /* 220 * This is a useful helper function for locking the anon_vma root as 221 * we traverse the vma->anon_vma_chain, looping over anon_vma's that 222 * have the same vma. 223 * 224 * Such anon_vma's should have the same root, so you'd expect to see 225 * just a single mutex_lock for the whole traversal. 226 */ 227 static inline struct anon_vma *lock_anon_vma_root(struct anon_vma *root, struct anon_vma *anon_vma) 228 { 229 struct anon_vma *new_root = anon_vma->root; 230 if (new_root != root) { 231 if (WARN_ON_ONCE(root)) 232 up_write(&root->rwsem); 233 root = new_root; 234 down_write(&root->rwsem); 235 } 236 return root; 237 } 238 239 static inline void unlock_anon_vma_root(struct anon_vma *root) 240 { 241 if (root) 242 up_write(&root->rwsem); 243 } 244 245 /* 246 * Attach the anon_vmas from src to dst. 247 * Returns 0 on success, -ENOMEM on failure. 248 * 249 * If dst->anon_vma is NULL this function tries to find and reuse existing 250 * anon_vma which has no vmas and only one child anon_vma. This prevents 251 * degradation of anon_vma hierarchy to endless linear chain in case of 252 * constantly forking task. On the other hand, an anon_vma with more than one 253 * child isn't reused even if there was no alive vma, thus rmap walker has a 254 * good chance of avoiding scanning the whole hierarchy when it searches where 255 * page is mapped. 256 */ 257 int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src) 258 { 259 struct anon_vma_chain *avc, *pavc; 260 struct anon_vma *root = NULL; 261 262 list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) { 263 struct anon_vma *anon_vma; 264 265 avc = anon_vma_chain_alloc(GFP_NOWAIT | __GFP_NOWARN); 266 if (unlikely(!avc)) { 267 unlock_anon_vma_root(root); 268 root = NULL; 269 avc = anon_vma_chain_alloc(GFP_KERNEL); 270 if (!avc) 271 goto enomem_failure; 272 } 273 anon_vma = pavc->anon_vma; 274 root = lock_anon_vma_root(root, anon_vma); 275 anon_vma_chain_link(dst, avc, anon_vma); 276 277 /* 278 * Reuse existing anon_vma if its degree lower than two, 279 * that means it has no vma and only one anon_vma child. 280 * 281 * Do not chose parent anon_vma, otherwise first child 282 * will always reuse it. Root anon_vma is never reused: 283 * it has self-parent reference and at least one child. 284 */ 285 if (!dst->anon_vma && anon_vma != src->anon_vma && 286 anon_vma->degree < 2) 287 dst->anon_vma = anon_vma; 288 } 289 if (dst->anon_vma) 290 dst->anon_vma->degree++; 291 unlock_anon_vma_root(root); 292 return 0; 293 294 enomem_failure: 295 /* 296 * dst->anon_vma is dropped here otherwise its degree can be incorrectly 297 * decremented in unlink_anon_vmas(). 298 * We can safely do this because callers of anon_vma_clone() don't care 299 * about dst->anon_vma if anon_vma_clone() failed. 300 */ 301 dst->anon_vma = NULL; 302 unlink_anon_vmas(dst); 303 return -ENOMEM; 304 } 305 306 /* 307 * Attach vma to its own anon_vma, as well as to the anon_vmas that 308 * the corresponding VMA in the parent process is attached to. 309 * Returns 0 on success, non-zero on failure. 310 */ 311 int anon_vma_fork(struct vm_area_struct *vma, struct vm_area_struct *pvma) 312 { 313 struct anon_vma_chain *avc; 314 struct anon_vma *anon_vma; 315 int error; 316 317 /* Don't bother if the parent process has no anon_vma here. */ 318 if (!pvma->anon_vma) 319 return 0; 320 321 /* Drop inherited anon_vma, we'll reuse existing or allocate new. */ 322 vma->anon_vma = NULL; 323 324 /* 325 * First, attach the new VMA to the parent VMA's anon_vmas, 326 * so rmap can find non-COWed pages in child processes. 327 */ 328 error = anon_vma_clone(vma, pvma); 329 if (error) 330 return error; 331 332 /* An existing anon_vma has been reused, all done then. */ 333 if (vma->anon_vma) 334 return 0; 335 336 /* Then add our own anon_vma. */ 337 anon_vma = anon_vma_alloc(); 338 if (!anon_vma) 339 goto out_error; 340 avc = anon_vma_chain_alloc(GFP_KERNEL); 341 if (!avc) 342 goto out_error_free_anon_vma; 343 344 /* 345 * The root anon_vma's spinlock is the lock actually used when we 346 * lock any of the anon_vmas in this anon_vma tree. 347 */ 348 anon_vma->root = pvma->anon_vma->root; 349 anon_vma->parent = pvma->anon_vma; 350 /* 351 * With refcounts, an anon_vma can stay around longer than the 352 * process it belongs to. The root anon_vma needs to be pinned until 353 * this anon_vma is freed, because the lock lives in the root. 354 */ 355 get_anon_vma(anon_vma->root); 356 /* Mark this anon_vma as the one where our new (COWed) pages go. */ 357 vma->anon_vma = anon_vma; 358 anon_vma_lock_write(anon_vma); 359 anon_vma_chain_link(vma, avc, anon_vma); 360 anon_vma->parent->degree++; 361 anon_vma_unlock_write(anon_vma); 362 363 return 0; 364 365 out_error_free_anon_vma: 366 put_anon_vma(anon_vma); 367 out_error: 368 unlink_anon_vmas(vma); 369 return -ENOMEM; 370 } 371 372 void unlink_anon_vmas(struct vm_area_struct *vma) 373 { 374 struct anon_vma_chain *avc, *next; 375 struct anon_vma *root = NULL; 376 377 /* 378 * Unlink each anon_vma chained to the VMA. This list is ordered 379 * from newest to oldest, ensuring the root anon_vma gets freed last. 380 */ 381 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { 382 struct anon_vma *anon_vma = avc->anon_vma; 383 384 root = lock_anon_vma_root(root, anon_vma); 385 anon_vma_interval_tree_remove(avc, &anon_vma->rb_root); 386 387 /* 388 * Leave empty anon_vmas on the list - we'll need 389 * to free them outside the lock. 390 */ 391 if (RB_EMPTY_ROOT(&anon_vma->rb_root)) { 392 anon_vma->parent->degree--; 393 continue; 394 } 395 396 list_del(&avc->same_vma); 397 anon_vma_chain_free(avc); 398 } 399 if (vma->anon_vma) 400 vma->anon_vma->degree--; 401 unlock_anon_vma_root(root); 402 403 /* 404 * Iterate the list once more, it now only contains empty and unlinked 405 * anon_vmas, destroy them. Could not do before due to __put_anon_vma() 406 * needing to write-acquire the anon_vma->root->rwsem. 407 */ 408 list_for_each_entry_safe(avc, next, &vma->anon_vma_chain, same_vma) { 409 struct anon_vma *anon_vma = avc->anon_vma; 410 411 VM_WARN_ON(anon_vma->degree); 412 put_anon_vma(anon_vma); 413 414 list_del(&avc->same_vma); 415 anon_vma_chain_free(avc); 416 } 417 } 418 419 static void anon_vma_ctor(void *data) 420 { 421 struct anon_vma *anon_vma = data; 422 423 init_rwsem(&anon_vma->rwsem); 424 atomic_set(&anon_vma->refcount, 0); 425 anon_vma->rb_root = RB_ROOT; 426 } 427 428 void __init anon_vma_init(void) 429 { 430 anon_vma_cachep = kmem_cache_create("anon_vma", sizeof(struct anon_vma), 431 0, SLAB_DESTROY_BY_RCU|SLAB_PANIC|SLAB_ACCOUNT, 432 anon_vma_ctor); 433 anon_vma_chain_cachep = KMEM_CACHE(anon_vma_chain, 434 SLAB_PANIC|SLAB_ACCOUNT); 435 } 436 437 /* 438 * Getting a lock on a stable anon_vma from a page off the LRU is tricky! 439 * 440 * Since there is no serialization what so ever against page_remove_rmap() 441 * the best this function can do is return a locked anon_vma that might 442 * have been relevant to this page. 443 * 444 * The page might have been remapped to a different anon_vma or the anon_vma 445 * returned may already be freed (and even reused). 446 * 447 * In case it was remapped to a different anon_vma, the new anon_vma will be a 448 * child of the old anon_vma, and the anon_vma lifetime rules will therefore 449 * ensure that any anon_vma obtained from the page will still be valid for as 450 * long as we observe page_mapped() [ hence all those page_mapped() tests ]. 451 * 452 * All users of this function must be very careful when walking the anon_vma 453 * chain and verify that the page in question is indeed mapped in it 454 * [ something equivalent to page_mapped_in_vma() ]. 455 * 456 * Since anon_vma's slab is DESTROY_BY_RCU and we know from page_remove_rmap() 457 * that the anon_vma pointer from page->mapping is valid if there is a 458 * mapcount, we can dereference the anon_vma after observing those. 459 */ 460 struct anon_vma *page_get_anon_vma(struct page *page) 461 { 462 struct anon_vma *anon_vma = NULL; 463 unsigned long anon_mapping; 464 465 rcu_read_lock(); 466 anon_mapping = (unsigned long)READ_ONCE(page->mapping); 467 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) 468 goto out; 469 if (!page_mapped(page)) 470 goto out; 471 472 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 473 if (!atomic_inc_not_zero(&anon_vma->refcount)) { 474 anon_vma = NULL; 475 goto out; 476 } 477 478 /* 479 * If this page is still mapped, then its anon_vma cannot have been 480 * freed. But if it has been unmapped, we have no security against the 481 * anon_vma structure being freed and reused (for another anon_vma: 482 * SLAB_DESTROY_BY_RCU guarantees that - so the atomic_inc_not_zero() 483 * above cannot corrupt). 484 */ 485 if (!page_mapped(page)) { 486 rcu_read_unlock(); 487 put_anon_vma(anon_vma); 488 return NULL; 489 } 490 out: 491 rcu_read_unlock(); 492 493 return anon_vma; 494 } 495 496 /* 497 * Similar to page_get_anon_vma() except it locks the anon_vma. 498 * 499 * Its a little more complex as it tries to keep the fast path to a single 500 * atomic op -- the trylock. If we fail the trylock, we fall back to getting a 501 * reference like with page_get_anon_vma() and then block on the mutex. 502 */ 503 struct anon_vma *page_lock_anon_vma_read(struct page *page) 504 { 505 struct anon_vma *anon_vma = NULL; 506 struct anon_vma *root_anon_vma; 507 unsigned long anon_mapping; 508 509 rcu_read_lock(); 510 anon_mapping = (unsigned long)READ_ONCE(page->mapping); 511 if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON) 512 goto out; 513 if (!page_mapped(page)) 514 goto out; 515 516 anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON); 517 root_anon_vma = READ_ONCE(anon_vma->root); 518 if (down_read_trylock(&root_anon_vma->rwsem)) { 519 /* 520 * If the page is still mapped, then this anon_vma is still 521 * its anon_vma, and holding the mutex ensures that it will 522 * not go away, see anon_vma_free(). 523 */ 524 if (!page_mapped(page)) { 525 up_read(&root_anon_vma->rwsem); 526 anon_vma = NULL; 527 } 528 goto out; 529 } 530 531 /* trylock failed, we got to sleep */ 532 if (!atomic_inc_not_zero(&anon_vma->refcount)) { 533 anon_vma = NULL; 534 goto out; 535 } 536 537 if (!page_mapped(page)) { 538 rcu_read_unlock(); 539 put_anon_vma(anon_vma); 540 return NULL; 541 } 542 543 /* we pinned the anon_vma, its safe to sleep */ 544 rcu_read_unlock(); 545 anon_vma_lock_read(anon_vma); 546 547 if (atomic_dec_and_test(&anon_vma->refcount)) { 548 /* 549 * Oops, we held the last refcount, release the lock 550 * and bail -- can't simply use put_anon_vma() because 551 * we'll deadlock on the anon_vma_lock_write() recursion. 552 */ 553 anon_vma_unlock_read(anon_vma); 554 __put_anon_vma(anon_vma); 555 anon_vma = NULL; 556 } 557 558 return anon_vma; 559 560 out: 561 rcu_read_unlock(); 562 return anon_vma; 563 } 564 565 void page_unlock_anon_vma_read(struct anon_vma *anon_vma) 566 { 567 anon_vma_unlock_read(anon_vma); 568 } 569 570 #ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH 571 /* 572 * Flush TLB entries for recently unmapped pages from remote CPUs. It is 573 * important if a PTE was dirty when it was unmapped that it's flushed 574 * before any IO is initiated on the page to prevent lost writes. Similarly, 575 * it must be flushed before freeing to prevent data leakage. 576 */ 577 void try_to_unmap_flush(void) 578 { 579 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 580 int cpu; 581 582 if (!tlb_ubc->flush_required) 583 return; 584 585 cpu = get_cpu(); 586 587 if (cpumask_test_cpu(cpu, &tlb_ubc->cpumask)) { 588 count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL); 589 local_flush_tlb(); 590 trace_tlb_flush(TLB_LOCAL_SHOOTDOWN, TLB_FLUSH_ALL); 591 } 592 593 if (cpumask_any_but(&tlb_ubc->cpumask, cpu) < nr_cpu_ids) 594 flush_tlb_others(&tlb_ubc->cpumask, NULL, 0, TLB_FLUSH_ALL); 595 cpumask_clear(&tlb_ubc->cpumask); 596 tlb_ubc->flush_required = false; 597 tlb_ubc->writable = false; 598 put_cpu(); 599 } 600 601 /* Flush iff there are potentially writable TLB entries that can race with IO */ 602 void try_to_unmap_flush_dirty(void) 603 { 604 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 605 606 if (tlb_ubc->writable) 607 try_to_unmap_flush(); 608 } 609 610 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, 611 struct page *page, bool writable) 612 { 613 struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc; 614 615 cpumask_or(&tlb_ubc->cpumask, &tlb_ubc->cpumask, mm_cpumask(mm)); 616 tlb_ubc->flush_required = true; 617 618 /* 619 * If the PTE was dirty then it's best to assume it's writable. The 620 * caller must use try_to_unmap_flush_dirty() or try_to_unmap_flush() 621 * before the page is queued for IO. 622 */ 623 if (writable) 624 tlb_ubc->writable = true; 625 } 626 627 /* 628 * Returns true if the TLB flush should be deferred to the end of a batch of 629 * unmap operations to reduce IPIs. 630 */ 631 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) 632 { 633 bool should_defer = false; 634 635 if (!(flags & TTU_BATCH_FLUSH)) 636 return false; 637 638 /* If remote CPUs need to be flushed then defer batch the flush */ 639 if (cpumask_any_but(mm_cpumask(mm), get_cpu()) < nr_cpu_ids) 640 should_defer = true; 641 put_cpu(); 642 643 return should_defer; 644 } 645 #else 646 static void set_tlb_ubc_flush_pending(struct mm_struct *mm, 647 struct page *page, bool writable) 648 { 649 } 650 651 static bool should_defer_flush(struct mm_struct *mm, enum ttu_flags flags) 652 { 653 return false; 654 } 655 #endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */ 656 657 /* 658 * At what user virtual address is page expected in vma? 659 * Caller should check the page is actually part of the vma. 660 */ 661 unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma) 662 { 663 unsigned long address; 664 if (PageAnon(page)) { 665 struct anon_vma *page__anon_vma = page_anon_vma(page); 666 /* 667 * Note: swapoff's unuse_vma() is more efficient with this 668 * check, and needs it to match anon_vma when KSM is active. 669 */ 670 if (!vma->anon_vma || !page__anon_vma || 671 vma->anon_vma->root != page__anon_vma->root) 672 return -EFAULT; 673 } else if (page->mapping) { 674 if (!vma->vm_file || vma->vm_file->f_mapping != page->mapping) 675 return -EFAULT; 676 } else 677 return -EFAULT; 678 address = __vma_address(page, vma); 679 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) 680 return -EFAULT; 681 return address; 682 } 683 684 pmd_t *mm_find_pmd(struct mm_struct *mm, unsigned long address) 685 { 686 pgd_t *pgd; 687 pud_t *pud; 688 pmd_t *pmd = NULL; 689 pmd_t pmde; 690 691 pgd = pgd_offset(mm, address); 692 if (!pgd_present(*pgd)) 693 goto out; 694 695 pud = pud_offset(pgd, address); 696 if (!pud_present(*pud)) 697 goto out; 698 699 pmd = pmd_offset(pud, address); 700 /* 701 * Some THP functions use the sequence pmdp_huge_clear_flush(), set_pmd_at() 702 * without holding anon_vma lock for write. So when looking for a 703 * genuine pmde (in which to find pte), test present and !THP together. 704 */ 705 pmde = *pmd; 706 barrier(); 707 if (!pmd_present(pmde) || pmd_trans_huge(pmde)) 708 pmd = NULL; 709 out: 710 return pmd; 711 } 712 713 /* 714 * Check that @page is mapped at @address into @mm. 715 * 716 * If @sync is false, page_check_address may perform a racy check to avoid 717 * the page table lock when the pte is not present (helpful when reclaiming 718 * highly shared pages). 719 * 720 * On success returns with pte mapped and locked. 721 */ 722 pte_t *__page_check_address(struct page *page, struct mm_struct *mm, 723 unsigned long address, spinlock_t **ptlp, int sync) 724 { 725 pmd_t *pmd; 726 pte_t *pte; 727 spinlock_t *ptl; 728 729 if (unlikely(PageHuge(page))) { 730 /* when pud is not present, pte will be NULL */ 731 pte = huge_pte_offset(mm, address); 732 if (!pte) 733 return NULL; 734 735 ptl = huge_pte_lockptr(page_hstate(page), mm, pte); 736 goto check; 737 } 738 739 pmd = mm_find_pmd(mm, address); 740 if (!pmd) 741 return NULL; 742 743 pte = pte_offset_map(pmd, address); 744 /* Make a quick check before getting the lock */ 745 if (!sync && !pte_present(*pte)) { 746 pte_unmap(pte); 747 return NULL; 748 } 749 750 ptl = pte_lockptr(mm, pmd); 751 check: 752 spin_lock(ptl); 753 if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) { 754 *ptlp = ptl; 755 return pte; 756 } 757 pte_unmap_unlock(pte, ptl); 758 return NULL; 759 } 760 761 /** 762 * page_mapped_in_vma - check whether a page is really mapped in a VMA 763 * @page: the page to test 764 * @vma: the VMA to test 765 * 766 * Returns 1 if the page is mapped into the page tables of the VMA, 0 767 * if the page is not mapped into the page tables of this VMA. Only 768 * valid for normal file or anonymous VMAs. 769 */ 770 int page_mapped_in_vma(struct page *page, struct vm_area_struct *vma) 771 { 772 unsigned long address; 773 pte_t *pte; 774 spinlock_t *ptl; 775 776 address = __vma_address(page, vma); 777 if (unlikely(address < vma->vm_start || address >= vma->vm_end)) 778 return 0; 779 pte = page_check_address(page, vma->vm_mm, address, &ptl, 1); 780 if (!pte) /* the page is not in this mm */ 781 return 0; 782 pte_unmap_unlock(pte, ptl); 783 784 return 1; 785 } 786 787 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 788 /* 789 * Check that @page is mapped at @address into @mm. In contrast to 790 * page_check_address(), this function can handle transparent huge pages. 791 * 792 * On success returns true with pte mapped and locked. For PMD-mapped 793 * transparent huge pages *@ptep is set to NULL. 794 */ 795 bool page_check_address_transhuge(struct page *page, struct mm_struct *mm, 796 unsigned long address, pmd_t **pmdp, 797 pte_t **ptep, spinlock_t **ptlp) 798 { 799 pgd_t *pgd; 800 pud_t *pud; 801 pmd_t *pmd; 802 pte_t *pte; 803 spinlock_t *ptl; 804 805 if (unlikely(PageHuge(page))) { 806 /* when pud is not present, pte will be NULL */ 807 pte = huge_pte_offset(mm, address); 808 if (!pte) 809 return false; 810 811 ptl = huge_pte_lockptr(page_hstate(page), mm, pte); 812 pmd = NULL; 813 goto check_pte; 814 } 815 816 pgd = pgd_offset(mm, address); 817 if (!pgd_present(*pgd)) 818 return false; 819 pud = pud_offset(pgd, address); 820 if (!pud_present(*pud)) 821 return false; 822 pmd = pmd_offset(pud, address); 823 824 if (pmd_trans_huge(*pmd)) { 825 ptl = pmd_lock(mm, pmd); 826 if (!pmd_present(*pmd)) 827 goto unlock_pmd; 828 if (unlikely(!pmd_trans_huge(*pmd))) { 829 spin_unlock(ptl); 830 goto map_pte; 831 } 832 833 if (pmd_page(*pmd) != page) 834 goto unlock_pmd; 835 836 pte = NULL; 837 goto found; 838 unlock_pmd: 839 spin_unlock(ptl); 840 return false; 841 } else { 842 pmd_t pmde = *pmd; 843 844 barrier(); 845 if (!pmd_present(pmde) || pmd_trans_huge(pmde)) 846 return false; 847 } 848 map_pte: 849 pte = pte_offset_map(pmd, address); 850 if (!pte_present(*pte)) { 851 pte_unmap(pte); 852 return false; 853 } 854 855 ptl = pte_lockptr(mm, pmd); 856 check_pte: 857 spin_lock(ptl); 858 859 if (!pte_present(*pte)) { 860 pte_unmap_unlock(pte, ptl); 861 return false; 862 } 863 864 /* THP can be referenced by any subpage */ 865 if (pte_pfn(*pte) - page_to_pfn(page) >= hpage_nr_pages(page)) { 866 pte_unmap_unlock(pte, ptl); 867 return false; 868 } 869 found: 870 *ptep = pte; 871 *pmdp = pmd; 872 *ptlp = ptl; 873 return true; 874 } 875 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 876 877 struct page_referenced_arg { 878 int mapcount; 879 int referenced; 880 unsigned long vm_flags; 881 struct mem_cgroup *memcg; 882 }; 883 /* 884 * arg: page_referenced_arg will be passed 885 */ 886 static int page_referenced_one(struct page *page, struct vm_area_struct *vma, 887 unsigned long address, void *arg) 888 { 889 struct mm_struct *mm = vma->vm_mm; 890 struct page_referenced_arg *pra = arg; 891 pmd_t *pmd; 892 pte_t *pte; 893 spinlock_t *ptl; 894 int referenced = 0; 895 896 if (!page_check_address_transhuge(page, mm, address, &pmd, &pte, &ptl)) 897 return SWAP_AGAIN; 898 899 if (vma->vm_flags & VM_LOCKED) { 900 if (pte) 901 pte_unmap(pte); 902 spin_unlock(ptl); 903 pra->vm_flags |= VM_LOCKED; 904 return SWAP_FAIL; /* To break the loop */ 905 } 906 907 if (pte) { 908 if (ptep_clear_flush_young_notify(vma, address, pte)) { 909 /* 910 * Don't treat a reference through a sequentially read 911 * mapping as such. If the page has been used in 912 * another mapping, we will catch it; if this other 913 * mapping is already gone, the unmap path will have 914 * set PG_referenced or activated the page. 915 */ 916 if (likely(!(vma->vm_flags & VM_SEQ_READ))) 917 referenced++; 918 } 919 pte_unmap(pte); 920 } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) { 921 if (pmdp_clear_flush_young_notify(vma, address, pmd)) 922 referenced++; 923 } else { 924 /* unexpected pmd-mapped page? */ 925 WARN_ON_ONCE(1); 926 } 927 spin_unlock(ptl); 928 929 if (referenced) 930 clear_page_idle(page); 931 if (test_and_clear_page_young(page)) 932 referenced++; 933 934 if (referenced) { 935 pra->referenced++; 936 pra->vm_flags |= vma->vm_flags; 937 } 938 939 pra->mapcount--; 940 if (!pra->mapcount) 941 return SWAP_SUCCESS; /* To break the loop */ 942 943 return SWAP_AGAIN; 944 } 945 946 static bool invalid_page_referenced_vma(struct vm_area_struct *vma, void *arg) 947 { 948 struct page_referenced_arg *pra = arg; 949 struct mem_cgroup *memcg = pra->memcg; 950 951 if (!mm_match_cgroup(vma->vm_mm, memcg)) 952 return true; 953 954 return false; 955 } 956 957 /** 958 * page_referenced - test if the page was referenced 959 * @page: the page to test 960 * @is_locked: caller holds lock on the page 961 * @memcg: target memory cgroup 962 * @vm_flags: collect encountered vma->vm_flags who actually referenced the page 963 * 964 * Quick test_and_clear_referenced for all mappings to a page, 965 * returns the number of ptes which referenced the page. 966 */ 967 int page_referenced(struct page *page, 968 int is_locked, 969 struct mem_cgroup *memcg, 970 unsigned long *vm_flags) 971 { 972 int ret; 973 int we_locked = 0; 974 struct page_referenced_arg pra = { 975 .mapcount = total_mapcount(page), 976 .memcg = memcg, 977 }; 978 struct rmap_walk_control rwc = { 979 .rmap_one = page_referenced_one, 980 .arg = (void *)&pra, 981 .anon_lock = page_lock_anon_vma_read, 982 }; 983 984 *vm_flags = 0; 985 if (!page_mapped(page)) 986 return 0; 987 988 if (!page_rmapping(page)) 989 return 0; 990 991 if (!is_locked && (!PageAnon(page) || PageKsm(page))) { 992 we_locked = trylock_page(page); 993 if (!we_locked) 994 return 1; 995 } 996 997 /* 998 * If we are reclaiming on behalf of a cgroup, skip 999 * counting on behalf of references from different 1000 * cgroups 1001 */ 1002 if (memcg) { 1003 rwc.invalid_vma = invalid_page_referenced_vma; 1004 } 1005 1006 ret = rmap_walk(page, &rwc); 1007 *vm_flags = pra.vm_flags; 1008 1009 if (we_locked) 1010 unlock_page(page); 1011 1012 return pra.referenced; 1013 } 1014 1015 static int page_mkclean_one(struct page *page, struct vm_area_struct *vma, 1016 unsigned long address, void *arg) 1017 { 1018 struct mm_struct *mm = vma->vm_mm; 1019 pte_t *pte; 1020 spinlock_t *ptl; 1021 int ret = 0; 1022 int *cleaned = arg; 1023 1024 pte = page_check_address(page, mm, address, &ptl, 1); 1025 if (!pte) 1026 goto out; 1027 1028 if (pte_dirty(*pte) || pte_write(*pte)) { 1029 pte_t entry; 1030 1031 flush_cache_page(vma, address, pte_pfn(*pte)); 1032 entry = ptep_clear_flush(vma, address, pte); 1033 entry = pte_wrprotect(entry); 1034 entry = pte_mkclean(entry); 1035 set_pte_at(mm, address, pte, entry); 1036 ret = 1; 1037 } 1038 1039 pte_unmap_unlock(pte, ptl); 1040 1041 if (ret) { 1042 mmu_notifier_invalidate_page(mm, address); 1043 (*cleaned)++; 1044 } 1045 out: 1046 return SWAP_AGAIN; 1047 } 1048 1049 static bool invalid_mkclean_vma(struct vm_area_struct *vma, void *arg) 1050 { 1051 if (vma->vm_flags & VM_SHARED) 1052 return false; 1053 1054 return true; 1055 } 1056 1057 int page_mkclean(struct page *page) 1058 { 1059 int cleaned = 0; 1060 struct address_space *mapping; 1061 struct rmap_walk_control rwc = { 1062 .arg = (void *)&cleaned, 1063 .rmap_one = page_mkclean_one, 1064 .invalid_vma = invalid_mkclean_vma, 1065 }; 1066 1067 BUG_ON(!PageLocked(page)); 1068 1069 if (!page_mapped(page)) 1070 return 0; 1071 1072 mapping = page_mapping(page); 1073 if (!mapping) 1074 return 0; 1075 1076 rmap_walk(page, &rwc); 1077 1078 return cleaned; 1079 } 1080 EXPORT_SYMBOL_GPL(page_mkclean); 1081 1082 /** 1083 * page_move_anon_rmap - move a page to our anon_vma 1084 * @page: the page to move to our anon_vma 1085 * @vma: the vma the page belongs to 1086 * 1087 * When a page belongs exclusively to one process after a COW event, 1088 * that page can be moved into the anon_vma that belongs to just that 1089 * process, so the rmap code will not search the parent or sibling 1090 * processes. 1091 */ 1092 void page_move_anon_rmap(struct page *page, struct vm_area_struct *vma) 1093 { 1094 struct anon_vma *anon_vma = vma->anon_vma; 1095 1096 page = compound_head(page); 1097 1098 VM_BUG_ON_PAGE(!PageLocked(page), page); 1099 VM_BUG_ON_VMA(!anon_vma, vma); 1100 1101 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 1102 /* 1103 * Ensure that anon_vma and the PAGE_MAPPING_ANON bit are written 1104 * simultaneously, so a concurrent reader (eg page_referenced()'s 1105 * PageAnon()) will not see one without the other. 1106 */ 1107 WRITE_ONCE(page->mapping, (struct address_space *) anon_vma); 1108 } 1109 1110 /** 1111 * __page_set_anon_rmap - set up new anonymous rmap 1112 * @page: Page to add to rmap 1113 * @vma: VM area to add page to. 1114 * @address: User virtual address of the mapping 1115 * @exclusive: the page is exclusively owned by the current process 1116 */ 1117 static void __page_set_anon_rmap(struct page *page, 1118 struct vm_area_struct *vma, unsigned long address, int exclusive) 1119 { 1120 struct anon_vma *anon_vma = vma->anon_vma; 1121 1122 BUG_ON(!anon_vma); 1123 1124 if (PageAnon(page)) 1125 return; 1126 1127 /* 1128 * If the page isn't exclusively mapped into this vma, 1129 * we must use the _oldest_ possible anon_vma for the 1130 * page mapping! 1131 */ 1132 if (!exclusive) 1133 anon_vma = anon_vma->root; 1134 1135 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 1136 page->mapping = (struct address_space *) anon_vma; 1137 page->index = linear_page_index(vma, address); 1138 } 1139 1140 /** 1141 * __page_check_anon_rmap - sanity check anonymous rmap addition 1142 * @page: the page to add the mapping to 1143 * @vma: the vm area in which the mapping is added 1144 * @address: the user virtual address mapped 1145 */ 1146 static void __page_check_anon_rmap(struct page *page, 1147 struct vm_area_struct *vma, unsigned long address) 1148 { 1149 #ifdef CONFIG_DEBUG_VM 1150 /* 1151 * The page's anon-rmap details (mapping and index) are guaranteed to 1152 * be set up correctly at this point. 1153 * 1154 * We have exclusion against page_add_anon_rmap because the caller 1155 * always holds the page locked, except if called from page_dup_rmap, 1156 * in which case the page is already known to be setup. 1157 * 1158 * We have exclusion against page_add_new_anon_rmap because those pages 1159 * are initially only visible via the pagetables, and the pte is locked 1160 * over the call to page_add_new_anon_rmap. 1161 */ 1162 BUG_ON(page_anon_vma(page)->root != vma->anon_vma->root); 1163 BUG_ON(page_to_pgoff(page) != linear_page_index(vma, address)); 1164 #endif 1165 } 1166 1167 /** 1168 * page_add_anon_rmap - add pte mapping to an anonymous page 1169 * @page: the page to add the mapping to 1170 * @vma: the vm area in which the mapping is added 1171 * @address: the user virtual address mapped 1172 * @compound: charge the page as compound or small page 1173 * 1174 * The caller needs to hold the pte lock, and the page must be locked in 1175 * the anon_vma case: to serialize mapping,index checking after setting, 1176 * and to ensure that PageAnon is not being upgraded racily to PageKsm 1177 * (but PageKsm is never downgraded to PageAnon). 1178 */ 1179 void page_add_anon_rmap(struct page *page, 1180 struct vm_area_struct *vma, unsigned long address, bool compound) 1181 { 1182 do_page_add_anon_rmap(page, vma, address, compound ? RMAP_COMPOUND : 0); 1183 } 1184 1185 /* 1186 * Special version of the above for do_swap_page, which often runs 1187 * into pages that are exclusively owned by the current process. 1188 * Everybody else should continue to use page_add_anon_rmap above. 1189 */ 1190 void do_page_add_anon_rmap(struct page *page, 1191 struct vm_area_struct *vma, unsigned long address, int flags) 1192 { 1193 bool compound = flags & RMAP_COMPOUND; 1194 bool first; 1195 1196 if (compound) { 1197 atomic_t *mapcount; 1198 VM_BUG_ON_PAGE(!PageLocked(page), page); 1199 VM_BUG_ON_PAGE(!PageTransHuge(page), page); 1200 mapcount = compound_mapcount_ptr(page); 1201 first = atomic_inc_and_test(mapcount); 1202 } else { 1203 first = atomic_inc_and_test(&page->_mapcount); 1204 } 1205 1206 if (first) { 1207 int nr = compound ? hpage_nr_pages(page) : 1; 1208 /* 1209 * We use the irq-unsafe __{inc|mod}_zone_page_stat because 1210 * these counters are not modified in interrupt context, and 1211 * pte lock(a spinlock) is held, which implies preemption 1212 * disabled. 1213 */ 1214 if (compound) 1215 __inc_node_page_state(page, NR_ANON_THPS); 1216 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr); 1217 } 1218 if (unlikely(PageKsm(page))) 1219 return; 1220 1221 VM_BUG_ON_PAGE(!PageLocked(page), page); 1222 1223 /* address might be in next vma when migration races vma_adjust */ 1224 if (first) 1225 __page_set_anon_rmap(page, vma, address, 1226 flags & RMAP_EXCLUSIVE); 1227 else 1228 __page_check_anon_rmap(page, vma, address); 1229 } 1230 1231 /** 1232 * page_add_new_anon_rmap - add pte mapping to a new anonymous page 1233 * @page: the page to add the mapping to 1234 * @vma: the vm area in which the mapping is added 1235 * @address: the user virtual address mapped 1236 * @compound: charge the page as compound or small page 1237 * 1238 * Same as page_add_anon_rmap but must only be called on *new* pages. 1239 * This means the inc-and-test can be bypassed. 1240 * Page does not have to be locked. 1241 */ 1242 void page_add_new_anon_rmap(struct page *page, 1243 struct vm_area_struct *vma, unsigned long address, bool compound) 1244 { 1245 int nr = compound ? hpage_nr_pages(page) : 1; 1246 1247 VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma); 1248 __SetPageSwapBacked(page); 1249 if (compound) { 1250 VM_BUG_ON_PAGE(!PageTransHuge(page), page); 1251 /* increment count (starts at -1) */ 1252 atomic_set(compound_mapcount_ptr(page), 0); 1253 __inc_node_page_state(page, NR_ANON_THPS); 1254 } else { 1255 /* Anon THP always mapped first with PMD */ 1256 VM_BUG_ON_PAGE(PageTransCompound(page), page); 1257 /* increment count (starts at -1) */ 1258 atomic_set(&page->_mapcount, 0); 1259 } 1260 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, nr); 1261 __page_set_anon_rmap(page, vma, address, 1); 1262 } 1263 1264 /** 1265 * page_add_file_rmap - add pte mapping to a file page 1266 * @page: the page to add the mapping to 1267 * 1268 * The caller needs to hold the pte lock. 1269 */ 1270 void page_add_file_rmap(struct page *page, bool compound) 1271 { 1272 int i, nr = 1; 1273 1274 VM_BUG_ON_PAGE(compound && !PageTransHuge(page), page); 1275 lock_page_memcg(page); 1276 if (compound && PageTransHuge(page)) { 1277 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { 1278 if (atomic_inc_and_test(&page[i]._mapcount)) 1279 nr++; 1280 } 1281 if (!atomic_inc_and_test(compound_mapcount_ptr(page))) 1282 goto out; 1283 VM_BUG_ON_PAGE(!PageSwapBacked(page), page); 1284 __inc_node_page_state(page, NR_SHMEM_PMDMAPPED); 1285 } else { 1286 if (PageTransCompound(page) && page_mapping(page)) { 1287 VM_WARN_ON_ONCE(!PageLocked(page)); 1288 1289 SetPageDoubleMap(compound_head(page)); 1290 if (PageMlocked(page)) 1291 clear_page_mlock(compound_head(page)); 1292 } 1293 if (!atomic_inc_and_test(&page->_mapcount)) 1294 goto out; 1295 } 1296 __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, nr); 1297 mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED); 1298 out: 1299 unlock_page_memcg(page); 1300 } 1301 1302 static void page_remove_file_rmap(struct page *page, bool compound) 1303 { 1304 int i, nr = 1; 1305 1306 VM_BUG_ON_PAGE(compound && !PageHead(page), page); 1307 lock_page_memcg(page); 1308 1309 /* Hugepages are not counted in NR_FILE_MAPPED for now. */ 1310 if (unlikely(PageHuge(page))) { 1311 /* hugetlb pages are always mapped with pmds */ 1312 atomic_dec(compound_mapcount_ptr(page)); 1313 goto out; 1314 } 1315 1316 /* page still mapped by someone else? */ 1317 if (compound && PageTransHuge(page)) { 1318 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { 1319 if (atomic_add_negative(-1, &page[i]._mapcount)) 1320 nr++; 1321 } 1322 if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) 1323 goto out; 1324 VM_BUG_ON_PAGE(!PageSwapBacked(page), page); 1325 __dec_node_page_state(page, NR_SHMEM_PMDMAPPED); 1326 } else { 1327 if (!atomic_add_negative(-1, &page->_mapcount)) 1328 goto out; 1329 } 1330 1331 /* 1332 * We use the irq-unsafe __{inc|mod}_zone_page_state because 1333 * these counters are not modified in interrupt context, and 1334 * pte lock(a spinlock) is held, which implies preemption disabled. 1335 */ 1336 __mod_node_page_state(page_pgdat(page), NR_FILE_MAPPED, -nr); 1337 mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_FILE_MAPPED); 1338 1339 if (unlikely(PageMlocked(page))) 1340 clear_page_mlock(page); 1341 out: 1342 unlock_page_memcg(page); 1343 } 1344 1345 static void page_remove_anon_compound_rmap(struct page *page) 1346 { 1347 int i, nr; 1348 1349 if (!atomic_add_negative(-1, compound_mapcount_ptr(page))) 1350 return; 1351 1352 /* Hugepages are not counted in NR_ANON_PAGES for now. */ 1353 if (unlikely(PageHuge(page))) 1354 return; 1355 1356 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) 1357 return; 1358 1359 __dec_node_page_state(page, NR_ANON_THPS); 1360 1361 if (TestClearPageDoubleMap(page)) { 1362 /* 1363 * Subpages can be mapped with PTEs too. Check how many of 1364 * themi are still mapped. 1365 */ 1366 for (i = 0, nr = 0; i < HPAGE_PMD_NR; i++) { 1367 if (atomic_add_negative(-1, &page[i]._mapcount)) 1368 nr++; 1369 } 1370 } else { 1371 nr = HPAGE_PMD_NR; 1372 } 1373 1374 if (unlikely(PageMlocked(page))) 1375 clear_page_mlock(page); 1376 1377 if (nr) { 1378 __mod_node_page_state(page_pgdat(page), NR_ANON_MAPPED, -nr); 1379 deferred_split_huge_page(page); 1380 } 1381 } 1382 1383 /** 1384 * page_remove_rmap - take down pte mapping from a page 1385 * @page: page to remove mapping from 1386 * @compound: uncharge the page as compound or small page 1387 * 1388 * The caller needs to hold the pte lock. 1389 */ 1390 void page_remove_rmap(struct page *page, bool compound) 1391 { 1392 if (!PageAnon(page)) 1393 return page_remove_file_rmap(page, compound); 1394 1395 if (compound) 1396 return page_remove_anon_compound_rmap(page); 1397 1398 /* page still mapped by someone else? */ 1399 if (!atomic_add_negative(-1, &page->_mapcount)) 1400 return; 1401 1402 /* 1403 * We use the irq-unsafe __{inc|mod}_zone_page_stat because 1404 * these counters are not modified in interrupt context, and 1405 * pte lock(a spinlock) is held, which implies preemption disabled. 1406 */ 1407 __dec_node_page_state(page, NR_ANON_MAPPED); 1408 1409 if (unlikely(PageMlocked(page))) 1410 clear_page_mlock(page); 1411 1412 if (PageTransCompound(page)) 1413 deferred_split_huge_page(compound_head(page)); 1414 1415 /* 1416 * It would be tidy to reset the PageAnon mapping here, 1417 * but that might overwrite a racing page_add_anon_rmap 1418 * which increments mapcount after us but sets mapping 1419 * before us: so leave the reset to free_hot_cold_page, 1420 * and remember that it's only reliable while mapped. 1421 * Leaving it set also helps swapoff to reinstate ptes 1422 * faster for those pages still in swapcache. 1423 */ 1424 } 1425 1426 struct rmap_private { 1427 enum ttu_flags flags; 1428 int lazyfreed; 1429 }; 1430 1431 /* 1432 * @arg: enum ttu_flags will be passed to this argument 1433 */ 1434 static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma, 1435 unsigned long address, void *arg) 1436 { 1437 struct mm_struct *mm = vma->vm_mm; 1438 pte_t *pte; 1439 pte_t pteval; 1440 spinlock_t *ptl; 1441 int ret = SWAP_AGAIN; 1442 struct rmap_private *rp = arg; 1443 enum ttu_flags flags = rp->flags; 1444 1445 /* munlock has nothing to gain from examining un-locked vmas */ 1446 if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED)) 1447 goto out; 1448 1449 if (flags & TTU_SPLIT_HUGE_PMD) { 1450 split_huge_pmd_address(vma, address, 1451 flags & TTU_MIGRATION, page); 1452 /* check if we have anything to do after split */ 1453 if (page_mapcount(page) == 0) 1454 goto out; 1455 } 1456 1457 pte = page_check_address(page, mm, address, &ptl, 1458 PageTransCompound(page)); 1459 if (!pte) 1460 goto out; 1461 1462 /* 1463 * If the page is mlock()d, we cannot swap it out. 1464 * If it's recently referenced (perhaps page_referenced 1465 * skipped over this mm) then we should reactivate it. 1466 */ 1467 if (!(flags & TTU_IGNORE_MLOCK)) { 1468 if (vma->vm_flags & VM_LOCKED) { 1469 /* PTE-mapped THP are never mlocked */ 1470 if (!PageTransCompound(page)) { 1471 /* 1472 * Holding pte lock, we do *not* need 1473 * mmap_sem here 1474 */ 1475 mlock_vma_page(page); 1476 } 1477 ret = SWAP_MLOCK; 1478 goto out_unmap; 1479 } 1480 if (flags & TTU_MUNLOCK) 1481 goto out_unmap; 1482 } 1483 if (!(flags & TTU_IGNORE_ACCESS)) { 1484 if (ptep_clear_flush_young_notify(vma, address, pte)) { 1485 ret = SWAP_FAIL; 1486 goto out_unmap; 1487 } 1488 } 1489 1490 /* Nuke the page table entry. */ 1491 flush_cache_page(vma, address, page_to_pfn(page)); 1492 if (should_defer_flush(mm, flags)) { 1493 /* 1494 * We clear the PTE but do not flush so potentially a remote 1495 * CPU could still be writing to the page. If the entry was 1496 * previously clean then the architecture must guarantee that 1497 * a clear->dirty transition on a cached TLB entry is written 1498 * through and traps if the PTE is unmapped. 1499 */ 1500 pteval = ptep_get_and_clear(mm, address, pte); 1501 1502 set_tlb_ubc_flush_pending(mm, page, pte_dirty(pteval)); 1503 } else { 1504 pteval = ptep_clear_flush(vma, address, pte); 1505 } 1506 1507 /* Move the dirty bit to the physical page now the pte is gone. */ 1508 if (pte_dirty(pteval)) 1509 set_page_dirty(page); 1510 1511 /* Update high watermark before we lower rss */ 1512 update_hiwater_rss(mm); 1513 1514 if (PageHWPoison(page) && !(flags & TTU_IGNORE_HWPOISON)) { 1515 if (PageHuge(page)) { 1516 hugetlb_count_sub(1 << compound_order(page), mm); 1517 } else { 1518 dec_mm_counter(mm, mm_counter(page)); 1519 } 1520 set_pte_at(mm, address, pte, 1521 swp_entry_to_pte(make_hwpoison_entry(page))); 1522 } else if (pte_unused(pteval)) { 1523 /* 1524 * The guest indicated that the page content is of no 1525 * interest anymore. Simply discard the pte, vmscan 1526 * will take care of the rest. 1527 */ 1528 dec_mm_counter(mm, mm_counter(page)); 1529 } else if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION)) { 1530 swp_entry_t entry; 1531 pte_t swp_pte; 1532 /* 1533 * Store the pfn of the page in a special migration 1534 * pte. do_swap_page() will wait until the migration 1535 * pte is removed and then restart fault handling. 1536 */ 1537 entry = make_migration_entry(page, pte_write(pteval)); 1538 swp_pte = swp_entry_to_pte(entry); 1539 if (pte_soft_dirty(pteval)) 1540 swp_pte = pte_swp_mksoft_dirty(swp_pte); 1541 set_pte_at(mm, address, pte, swp_pte); 1542 } else if (PageAnon(page)) { 1543 swp_entry_t entry = { .val = page_private(page) }; 1544 pte_t swp_pte; 1545 /* 1546 * Store the swap location in the pte. 1547 * See handle_pte_fault() ... 1548 */ 1549 VM_BUG_ON_PAGE(!PageSwapCache(page), page); 1550 1551 if (!PageDirty(page) && (flags & TTU_LZFREE)) { 1552 /* It's a freeable page by MADV_FREE */ 1553 dec_mm_counter(mm, MM_ANONPAGES); 1554 rp->lazyfreed++; 1555 goto discard; 1556 } 1557 1558 if (swap_duplicate(entry) < 0) { 1559 set_pte_at(mm, address, pte, pteval); 1560 ret = SWAP_FAIL; 1561 goto out_unmap; 1562 } 1563 if (list_empty(&mm->mmlist)) { 1564 spin_lock(&mmlist_lock); 1565 if (list_empty(&mm->mmlist)) 1566 list_add(&mm->mmlist, &init_mm.mmlist); 1567 spin_unlock(&mmlist_lock); 1568 } 1569 dec_mm_counter(mm, MM_ANONPAGES); 1570 inc_mm_counter(mm, MM_SWAPENTS); 1571 swp_pte = swp_entry_to_pte(entry); 1572 if (pte_soft_dirty(pteval)) 1573 swp_pte = pte_swp_mksoft_dirty(swp_pte); 1574 set_pte_at(mm, address, pte, swp_pte); 1575 } else 1576 dec_mm_counter(mm, mm_counter_file(page)); 1577 1578 discard: 1579 page_remove_rmap(page, PageHuge(page)); 1580 put_page(page); 1581 1582 out_unmap: 1583 pte_unmap_unlock(pte, ptl); 1584 if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK)) 1585 mmu_notifier_invalidate_page(mm, address); 1586 out: 1587 return ret; 1588 } 1589 1590 bool is_vma_temporary_stack(struct vm_area_struct *vma) 1591 { 1592 int maybe_stack = vma->vm_flags & (VM_GROWSDOWN | VM_GROWSUP); 1593 1594 if (!maybe_stack) 1595 return false; 1596 1597 if ((vma->vm_flags & VM_STACK_INCOMPLETE_SETUP) == 1598 VM_STACK_INCOMPLETE_SETUP) 1599 return true; 1600 1601 return false; 1602 } 1603 1604 static bool invalid_migration_vma(struct vm_area_struct *vma, void *arg) 1605 { 1606 return is_vma_temporary_stack(vma); 1607 } 1608 1609 static int page_mapcount_is_zero(struct page *page) 1610 { 1611 return !page_mapcount(page); 1612 } 1613 1614 /** 1615 * try_to_unmap - try to remove all page table mappings to a page 1616 * @page: the page to get unmapped 1617 * @flags: action and flags 1618 * 1619 * Tries to remove all the page table entries which are mapping this 1620 * page, used in the pageout path. Caller must hold the page lock. 1621 * Return values are: 1622 * 1623 * SWAP_SUCCESS - we succeeded in removing all mappings 1624 * SWAP_AGAIN - we missed a mapping, try again later 1625 * SWAP_FAIL - the page is unswappable 1626 * SWAP_MLOCK - page is mlocked. 1627 */ 1628 int try_to_unmap(struct page *page, enum ttu_flags flags) 1629 { 1630 int ret; 1631 struct rmap_private rp = { 1632 .flags = flags, 1633 .lazyfreed = 0, 1634 }; 1635 1636 struct rmap_walk_control rwc = { 1637 .rmap_one = try_to_unmap_one, 1638 .arg = &rp, 1639 .done = page_mapcount_is_zero, 1640 .anon_lock = page_lock_anon_vma_read, 1641 }; 1642 1643 /* 1644 * During exec, a temporary VMA is setup and later moved. 1645 * The VMA is moved under the anon_vma lock but not the 1646 * page tables leading to a race where migration cannot 1647 * find the migration ptes. Rather than increasing the 1648 * locking requirements of exec(), migration skips 1649 * temporary VMAs until after exec() completes. 1650 */ 1651 if ((flags & TTU_MIGRATION) && !PageKsm(page) && PageAnon(page)) 1652 rwc.invalid_vma = invalid_migration_vma; 1653 1654 if (flags & TTU_RMAP_LOCKED) 1655 ret = rmap_walk_locked(page, &rwc); 1656 else 1657 ret = rmap_walk(page, &rwc); 1658 1659 if (ret != SWAP_MLOCK && !page_mapcount(page)) { 1660 ret = SWAP_SUCCESS; 1661 if (rp.lazyfreed && !PageDirty(page)) 1662 ret = SWAP_LZFREE; 1663 } 1664 return ret; 1665 } 1666 1667 static int page_not_mapped(struct page *page) 1668 { 1669 return !page_mapped(page); 1670 }; 1671 1672 /** 1673 * try_to_munlock - try to munlock a page 1674 * @page: the page to be munlocked 1675 * 1676 * Called from munlock code. Checks all of the VMAs mapping the page 1677 * to make sure nobody else has this page mlocked. The page will be 1678 * returned with PG_mlocked cleared if no other vmas have it mlocked. 1679 * 1680 * Return values are: 1681 * 1682 * SWAP_AGAIN - no vma is holding page mlocked, or, 1683 * SWAP_AGAIN - page mapped in mlocked vma -- couldn't acquire mmap sem 1684 * SWAP_FAIL - page cannot be located at present 1685 * SWAP_MLOCK - page is now mlocked. 1686 */ 1687 int try_to_munlock(struct page *page) 1688 { 1689 int ret; 1690 struct rmap_private rp = { 1691 .flags = TTU_MUNLOCK, 1692 .lazyfreed = 0, 1693 }; 1694 1695 struct rmap_walk_control rwc = { 1696 .rmap_one = try_to_unmap_one, 1697 .arg = &rp, 1698 .done = page_not_mapped, 1699 .anon_lock = page_lock_anon_vma_read, 1700 1701 }; 1702 1703 VM_BUG_ON_PAGE(!PageLocked(page) || PageLRU(page), page); 1704 1705 ret = rmap_walk(page, &rwc); 1706 return ret; 1707 } 1708 1709 void __put_anon_vma(struct anon_vma *anon_vma) 1710 { 1711 struct anon_vma *root = anon_vma->root; 1712 1713 anon_vma_free(anon_vma); 1714 if (root != anon_vma && atomic_dec_and_test(&root->refcount)) 1715 anon_vma_free(root); 1716 } 1717 1718 static struct anon_vma *rmap_walk_anon_lock(struct page *page, 1719 struct rmap_walk_control *rwc) 1720 { 1721 struct anon_vma *anon_vma; 1722 1723 if (rwc->anon_lock) 1724 return rwc->anon_lock(page); 1725 1726 /* 1727 * Note: remove_migration_ptes() cannot use page_lock_anon_vma_read() 1728 * because that depends on page_mapped(); but not all its usages 1729 * are holding mmap_sem. Users without mmap_sem are required to 1730 * take a reference count to prevent the anon_vma disappearing 1731 */ 1732 anon_vma = page_anon_vma(page); 1733 if (!anon_vma) 1734 return NULL; 1735 1736 anon_vma_lock_read(anon_vma); 1737 return anon_vma; 1738 } 1739 1740 /* 1741 * rmap_walk_anon - do something to anonymous page using the object-based 1742 * rmap method 1743 * @page: the page to be handled 1744 * @rwc: control variable according to each walk type 1745 * 1746 * Find all the mappings of a page using the mapping pointer and the vma chains 1747 * contained in the anon_vma struct it points to. 1748 * 1749 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma 1750 * where the page was found will be held for write. So, we won't recheck 1751 * vm_flags for that VMA. That should be OK, because that vma shouldn't be 1752 * LOCKED. 1753 */ 1754 static int rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc, 1755 bool locked) 1756 { 1757 struct anon_vma *anon_vma; 1758 pgoff_t pgoff; 1759 struct anon_vma_chain *avc; 1760 int ret = SWAP_AGAIN; 1761 1762 if (locked) { 1763 anon_vma = page_anon_vma(page); 1764 /* anon_vma disappear under us? */ 1765 VM_BUG_ON_PAGE(!anon_vma, page); 1766 } else { 1767 anon_vma = rmap_walk_anon_lock(page, rwc); 1768 } 1769 if (!anon_vma) 1770 return ret; 1771 1772 pgoff = page_to_pgoff(page); 1773 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) { 1774 struct vm_area_struct *vma = avc->vma; 1775 unsigned long address = vma_address(page, vma); 1776 1777 cond_resched(); 1778 1779 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 1780 continue; 1781 1782 ret = rwc->rmap_one(page, vma, address, rwc->arg); 1783 if (ret != SWAP_AGAIN) 1784 break; 1785 if (rwc->done && rwc->done(page)) 1786 break; 1787 } 1788 1789 if (!locked) 1790 anon_vma_unlock_read(anon_vma); 1791 return ret; 1792 } 1793 1794 /* 1795 * rmap_walk_file - do something to file page using the object-based rmap method 1796 * @page: the page to be handled 1797 * @rwc: control variable according to each walk type 1798 * 1799 * Find all the mappings of a page using the mapping pointer and the vma chains 1800 * contained in the address_space struct it points to. 1801 * 1802 * When called from try_to_munlock(), the mmap_sem of the mm containing the vma 1803 * where the page was found will be held for write. So, we won't recheck 1804 * vm_flags for that VMA. That should be OK, because that vma shouldn't be 1805 * LOCKED. 1806 */ 1807 static int rmap_walk_file(struct page *page, struct rmap_walk_control *rwc, 1808 bool locked) 1809 { 1810 struct address_space *mapping = page_mapping(page); 1811 pgoff_t pgoff; 1812 struct vm_area_struct *vma; 1813 int ret = SWAP_AGAIN; 1814 1815 /* 1816 * The page lock not only makes sure that page->mapping cannot 1817 * suddenly be NULLified by truncation, it makes sure that the 1818 * structure at mapping cannot be freed and reused yet, 1819 * so we can safely take mapping->i_mmap_rwsem. 1820 */ 1821 VM_BUG_ON_PAGE(!PageLocked(page), page); 1822 1823 if (!mapping) 1824 return ret; 1825 1826 pgoff = page_to_pgoff(page); 1827 if (!locked) 1828 i_mmap_lock_read(mapping); 1829 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { 1830 unsigned long address = vma_address(page, vma); 1831 1832 cond_resched(); 1833 1834 if (rwc->invalid_vma && rwc->invalid_vma(vma, rwc->arg)) 1835 continue; 1836 1837 ret = rwc->rmap_one(page, vma, address, rwc->arg); 1838 if (ret != SWAP_AGAIN) 1839 goto done; 1840 if (rwc->done && rwc->done(page)) 1841 goto done; 1842 } 1843 1844 done: 1845 if (!locked) 1846 i_mmap_unlock_read(mapping); 1847 return ret; 1848 } 1849 1850 int rmap_walk(struct page *page, struct rmap_walk_control *rwc) 1851 { 1852 if (unlikely(PageKsm(page))) 1853 return rmap_walk_ksm(page, rwc); 1854 else if (PageAnon(page)) 1855 return rmap_walk_anon(page, rwc, false); 1856 else 1857 return rmap_walk_file(page, rwc, false); 1858 } 1859 1860 /* Like rmap_walk, but caller holds relevant rmap lock */ 1861 int rmap_walk_locked(struct page *page, struct rmap_walk_control *rwc) 1862 { 1863 /* no ksm support for now */ 1864 VM_BUG_ON_PAGE(PageKsm(page), page); 1865 if (PageAnon(page)) 1866 return rmap_walk_anon(page, rwc, true); 1867 else 1868 return rmap_walk_file(page, rwc, true); 1869 } 1870 1871 #ifdef CONFIG_HUGETLB_PAGE 1872 /* 1873 * The following three functions are for anonymous (private mapped) hugepages. 1874 * Unlike common anonymous pages, anonymous hugepages have no accounting code 1875 * and no lru code, because we handle hugepages differently from common pages. 1876 */ 1877 static void __hugepage_set_anon_rmap(struct page *page, 1878 struct vm_area_struct *vma, unsigned long address, int exclusive) 1879 { 1880 struct anon_vma *anon_vma = vma->anon_vma; 1881 1882 BUG_ON(!anon_vma); 1883 1884 if (PageAnon(page)) 1885 return; 1886 if (!exclusive) 1887 anon_vma = anon_vma->root; 1888 1889 anon_vma = (void *) anon_vma + PAGE_MAPPING_ANON; 1890 page->mapping = (struct address_space *) anon_vma; 1891 page->index = linear_page_index(vma, address); 1892 } 1893 1894 void hugepage_add_anon_rmap(struct page *page, 1895 struct vm_area_struct *vma, unsigned long address) 1896 { 1897 struct anon_vma *anon_vma = vma->anon_vma; 1898 int first; 1899 1900 BUG_ON(!PageLocked(page)); 1901 BUG_ON(!anon_vma); 1902 /* address might be in next vma when migration races vma_adjust */ 1903 first = atomic_inc_and_test(compound_mapcount_ptr(page)); 1904 if (first) 1905 __hugepage_set_anon_rmap(page, vma, address, 0); 1906 } 1907 1908 void hugepage_add_new_anon_rmap(struct page *page, 1909 struct vm_area_struct *vma, unsigned long address) 1910 { 1911 BUG_ON(address < vma->vm_start || address >= vma->vm_end); 1912 atomic_set(compound_mapcount_ptr(page), 0); 1913 __hugepage_set_anon_rmap(page, vma, address, 1); 1914 } 1915 #endif /* CONFIG_HUGETLB_PAGE */ 1916