개념 설명 전체 · v6.18.37 / mm/filemap.c
1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * linux/mm/filemap.c 4 * 5 * Copyright (C) 1994-1999 Linus Torvalds 6 */ 7 8 /* 9 * This file handles the generic file mmap semantics used by 10 * most "normal" filesystems (but you don't /have/ to use this: 11 * the NFS filesystem used to do this differently, for example) 12 */ 13 #include <linux/export.h> 14 #include <linux/compiler.h> 15 #include <linux/dax.h> 16 #include <linux/fs.h> 17 #include <linux/sched/signal.h> 18 #include <linux/uaccess.h> 19 #include <linux/capability.h> 20 #include <linux/kernel_stat.h> 21 #include <linux/gfp.h> 22 #include <linux/mm.h> 23 #include <linux/swap.h> 24 #include <linux/swapops.h> 25 #include <linux/syscalls.h> 26 #include <linux/mman.h> 27 #include <linux/pagemap.h> 28 #include <linux/file.h> 29 #include <linux/uio.h> 30 #include <linux/error-injection.h> 31 #include <linux/hash.h> 32 #include <linux/writeback.h> 33 #include <linux/backing-dev.h> 34 #include <linux/pagevec.h> 35 #include <linux/security.h> 36 #include <linux/cpuset.h> 37 #include <linux/hugetlb.h> 38 #include <linux/memcontrol.h> 39 #include <linux/shmem_fs.h> 40 #include <linux/rmap.h> 41 #include <linux/delayacct.h> 42 #include <linux/psi.h> 43 #include <linux/ramfs.h> 44 #include <linux/page_idle.h> 45 #include <linux/migrate.h> 46 #include <linux/pipe_fs_i.h> 47 #include <linux/splice.h> 48 #include <linux/rcupdate_wait.h> 49 #include <linux/sched/mm.h> 50 #include <linux/sysctl.h> 51 #include <asm/pgalloc.h> 52 #include <asm/tlbflush.h> 53 #include "internal.h" 54 55 #define CREATE_TRACE_POINTS 56 #include <trace/events/filemap.h> 57 58 /* 59 * FIXME: remove all knowledge of the buffer layer from the core VM 60 */ 61 #include <linux/buffer_head.h> /* for try_to_free_buffers */ 62 63 #include <asm/mman.h> 64 65 #include "swap.h" 66 67 /* 68 * Shared mappings implemented 30.11.1994. It's not fully working yet, 69 * though. 70 * 71 * Shared mappings now work. 15.8.1995 Bruno. 72 * 73 * finished 'unifying' the page and buffer cache and SMP-threaded the 74 * page-cache, 21.05.1999, Ingo Molnar <[email protected]> 75 * 76 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <[email protected]> 77 */ 78 79 /* 80 * Lock ordering: 81 * 82 * ->i_mmap_rwsem (truncate_pagecache) 83 * ->private_lock (__free_pte->block_dirty_folio) 84 * ->swap_lock (exclusive_swap_page, others) 85 * ->i_pages lock 86 * 87 * ->i_rwsem 88 * ->invalidate_lock (acquired by fs in truncate path) 89 * ->i_mmap_rwsem (truncate->unmap_mapping_range) 90 * 91 * ->mmap_lock 92 * ->i_mmap_rwsem 93 * ->page_table_lock or pte_lock (various, mainly in memory.c) 94 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock) 95 * 96 * ->mmap_lock 97 * ->invalidate_lock (filemap_fault) 98 * ->lock_page (filemap_fault, access_process_vm) 99 * 100 * ->i_rwsem (generic_perform_write) 101 * ->mmap_lock (fault_in_readable->do_page_fault) 102 * 103 * bdi->wb.list_lock 104 * sb_lock (fs/fs-writeback.c) 105 * ->i_pages lock (__sync_single_inode) 106 * 107 * ->i_mmap_rwsem 108 * ->anon_vma.lock (vma_merge) 109 * 110 * ->anon_vma.lock 111 * ->page_table_lock or pte_lock (anon_vma_prepare and various) 112 * 113 * ->page_table_lock or pte_lock 114 * ->swap_lock (try_to_unmap_one) 115 * ->private_lock (try_to_unmap_one) 116 * ->i_pages lock (try_to_unmap_one) 117 * ->lruvec->lru_lock (follow_page_mask->mark_page_accessed) 118 * ->lruvec->lru_lock (check_pte_range->folio_isolate_lru) 119 * ->private_lock (folio_remove_rmap_pte->set_page_dirty) 120 * ->i_pages lock (folio_remove_rmap_pte->set_page_dirty) 121 * bdi.wb->list_lock (folio_remove_rmap_pte->set_page_dirty) 122 * ->inode->i_lock (folio_remove_rmap_pte->set_page_dirty) 123 * bdi.wb->list_lock (zap_pte_range->set_page_dirty) 124 * ->inode->i_lock (zap_pte_range->set_page_dirty) 125 * ->private_lock (zap_pte_range->block_dirty_folio) 126 */ 127 128 static void page_cache_delete(struct address_space *mapping, 129 struct folio *folio, void *shadow) 130 { 131 XA_STATE(xas, &mapping->i_pages, folio->index); 132 long nr = 1; 133 134 mapping_set_update(&xas, mapping); 135 136 xas_set_order(&xas, folio->index, folio_order(folio)); 137 nr = folio_nr_pages(folio); 138 139 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 140 141 xas_store(&xas, shadow); 142 xas_init_marks(&xas); 143 144 folio->mapping = NULL; 145 /* Leave folio->index set: truncation lookup relies upon it */ 146 mapping->nrpages -= nr; 147 } 148 149 static void filemap_unaccount_folio(struct address_space *mapping, 150 struct folio *folio) 151 { 152 long nr; 153 154 VM_BUG_ON_FOLIO(folio_mapped(folio), folio); 155 if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(folio_mapped(folio))) { 156 pr_alert("BUG: Bad page cache in process %s pfn:%05lx\n", 157 current->comm, folio_pfn(folio)); 158 dump_page(&folio->page, "still mapped when deleted"); 159 dump_stack(); 160 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 161 162 if (mapping_exiting(mapping) && !folio_test_large(folio)) { 163 int mapcount = folio_mapcount(folio); 164 165 if (folio_ref_count(folio) >= mapcount + 2) { 166 /* 167 * All vmas have already been torn down, so it's 168 * a good bet that actually the page is unmapped 169 * and we'd rather not leak it: if we're wrong, 170 * another bad page check should catch it later. 171 */ 172 atomic_set(&folio->_mapcount, -1); 173 folio_ref_sub(folio, mapcount); 174 } 175 } 176 } 177 178 /* hugetlb folios do not participate in page cache accounting. */ 179 if (folio_test_hugetlb(folio)) 180 return; 181 182 nr = folio_nr_pages(folio); 183 184 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr); 185 if (folio_test_swapbacked(folio)) { 186 __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr); 187 if (folio_test_pmd_mappable(folio)) 188 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, -nr); 189 } else if (folio_test_pmd_mappable(folio)) { 190 __lruvec_stat_mod_folio(folio, NR_FILE_THPS, -nr); 191 filemap_nr_thps_dec(mapping); 192 } 193 if (test_bit(AS_KERNEL_FILE, &folio->mapping->flags)) 194 mod_node_page_state(folio_pgdat(folio), 195 NR_KERNEL_FILE_PAGES, -nr); 196 197 /* 198 * At this point folio must be either written or cleaned by 199 * truncate. Dirty folio here signals a bug and loss of 200 * unwritten data - on ordinary filesystems. 201 * 202 * But it's harmless on in-memory filesystems like tmpfs; and can 203 * occur when a driver which did get_user_pages() sets page dirty 204 * before putting it, while the inode is being finally evicted. 205 * 206 * Below fixes dirty accounting after removing the folio entirely 207 * but leaves the dirty flag set: it has no effect for truncated 208 * folio and anyway will be cleared before returning folio to 209 * buddy allocator. 210 */ 211 if (WARN_ON_ONCE(folio_test_dirty(folio) && 212 mapping_can_writeback(mapping))) 213 folio_account_cleaned(folio, inode_to_wb(mapping->host)); 214 } 215 216 /* 217 * Delete a page from the page cache and free it. Caller has to make 218 * sure the page is locked and that nobody else uses it - or that usage 219 * is safe. The caller must hold the i_pages lock. 220 */ 221 void __filemap_remove_folio(struct folio *folio, void *shadow) 222 { 223 struct address_space *mapping = folio->mapping; 224 225 trace_mm_filemap_delete_from_page_cache(folio); 226 filemap_unaccount_folio(mapping, folio); 227 page_cache_delete(mapping, folio, shadow); 228 } 229 230 static void filemap_free_folio(const struct address_space *mapping, 231 struct folio *folio) 232 { 233 void (*free_folio)(struct folio *); 234 235 free_folio = mapping->a_ops->free_folio; 236 if (free_folio) 237 free_folio(folio); 238 239 folio_put_refs(folio, folio_nr_pages(folio)); 240 } 241 242 /** 243 * filemap_remove_folio - Remove folio from page cache. 244 * @folio: The folio. 245 * 246 * This must be called only on folios that are locked and have been 247 * verified to be in the page cache. It will never put the folio into 248 * the free list because the caller has a reference on the page. 249 */ 250 void filemap_remove_folio(struct folio *folio) 251 { 252 struct address_space *mapping = folio->mapping; 253 254 BUG_ON(!folio_test_locked(folio)); 255 spin_lock(&mapping->host->i_lock); 256 xa_lock_irq(&mapping->i_pages); 257 __filemap_remove_folio(folio, NULL); 258 xa_unlock_irq(&mapping->i_pages); 259 if (mapping_shrinkable(mapping)) 260 inode_add_lru(mapping->host); 261 spin_unlock(&mapping->host->i_lock); 262 263 filemap_free_folio(mapping, folio); 264 } 265 266 /* 267 * page_cache_delete_batch - delete several folios from page cache 268 * @mapping: the mapping to which folios belong 269 * @fbatch: batch of folios to delete 270 * 271 * The function walks over mapping->i_pages and removes folios passed in 272 * @fbatch from the mapping. The function expects @fbatch to be sorted 273 * by page index and is optimised for it to be dense. 274 * It tolerates holes in @fbatch (mapping entries at those indices are not 275 * modified). 276 * 277 * The function expects the i_pages lock to be held. 278 */ 279 static void page_cache_delete_batch(struct address_space *mapping, 280 struct folio_batch *fbatch) 281 { 282 XA_STATE(xas, &mapping->i_pages, fbatch->folios[0]->index); 283 long total_pages = 0; 284 int i = 0; 285 struct folio *folio; 286 287 mapping_set_update(&xas, mapping); 288 xas_for_each(&xas, folio, ULONG_MAX) { 289 if (i >= folio_batch_count(fbatch)) 290 break; 291 292 /* A swap/dax/shadow entry got inserted? Skip it. */ 293 if (xa_is_value(folio)) 294 continue; 295 /* 296 * A page got inserted in our range? Skip it. We have our 297 * pages locked so they are protected from being removed. 298 * If we see a page whose index is higher than ours, it 299 * means our page has been removed, which shouldn't be 300 * possible because we're holding the PageLock. 301 */ 302 if (folio != fbatch->folios[i]) { 303 VM_BUG_ON_FOLIO(folio->index > 304 fbatch->folios[i]->index, folio); 305 continue; 306 } 307 308 WARN_ON_ONCE(!folio_test_locked(folio)); 309 310 folio->mapping = NULL; 311 /* Leave folio->index set: truncation lookup relies on it */ 312 313 i++; 314 xas_store(&xas, NULL); 315 total_pages += folio_nr_pages(folio); 316 } 317 mapping->nrpages -= total_pages; 318 } 319 320 void delete_from_page_cache_batch(struct address_space *mapping, 321 struct folio_batch *fbatch) 322 { 323 int i; 324 325 if (!folio_batch_count(fbatch)) 326 return; 327 328 spin_lock(&mapping->host->i_lock); 329 xa_lock_irq(&mapping->i_pages); 330 for (i = 0; i < folio_batch_count(fbatch); i++) { 331 struct folio *folio = fbatch->folios[i]; 332 333 trace_mm_filemap_delete_from_page_cache(folio); 334 filemap_unaccount_folio(mapping, folio); 335 } 336 page_cache_delete_batch(mapping, fbatch); 337 xa_unlock_irq(&mapping->i_pages); 338 if (mapping_shrinkable(mapping)) 339 inode_add_lru(mapping->host); 340 spin_unlock(&mapping->host->i_lock); 341 342 for (i = 0; i < folio_batch_count(fbatch); i++) 343 filemap_free_folio(mapping, fbatch->folios[i]); 344 } 345 346 int filemap_check_errors(struct address_space *mapping) 347 { 348 int ret = 0; 349 /* Check for outstanding write errors */ 350 if (test_bit(AS_ENOSPC, &mapping->flags) && 351 test_and_clear_bit(AS_ENOSPC, &mapping->flags)) 352 ret = -ENOSPC; 353 if (test_bit(AS_EIO, &mapping->flags) && 354 test_and_clear_bit(AS_EIO, &mapping->flags)) 355 ret = -EIO; 356 return ret; 357 } 358 EXPORT_SYMBOL(filemap_check_errors); 359 360 static int filemap_check_and_keep_errors(struct address_space *mapping) 361 { 362 /* Check for outstanding write errors */ 363 if (test_bit(AS_EIO, &mapping->flags)) 364 return -EIO; 365 if (test_bit(AS_ENOSPC, &mapping->flags)) 366 return -ENOSPC; 367 return 0; 368 } 369 370 /** 371 * filemap_fdatawrite_wbc - start writeback on mapping dirty pages in range 372 * @mapping: address space structure to write 373 * @wbc: the writeback_control controlling the writeout 374 * 375 * Call writepages on the mapping using the provided wbc to control the 376 * writeout. 377 * 378 * Return: %0 on success, negative error code otherwise. 379 */ 380 int filemap_fdatawrite_wbc(struct address_space *mapping, 381 struct writeback_control *wbc) 382 { 383 int ret; 384 385 if (!mapping_can_writeback(mapping) || 386 !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) 387 return 0; 388 389 wbc_attach_fdatawrite_inode(wbc, mapping->host); 390 ret = do_writepages(mapping, wbc); 391 wbc_detach_inode(wbc); 392 return ret; 393 } 394 EXPORT_SYMBOL(filemap_fdatawrite_wbc); 395 396 /** 397 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range 398 * @mapping: address space structure to write 399 * @start: offset in bytes where the range starts 400 * @end: offset in bytes where the range ends (inclusive) 401 * @sync_mode: enable synchronous operation 402 * 403 * Start writeback against all of a mapping's dirty pages that lie 404 * within the byte offsets <start, end> inclusive. 405 * 406 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as 407 * opposed to a regular memory cleansing writeback. The difference between 408 * these two operations is that if a dirty page/buffer is encountered, it must 409 * be waited upon, and not just skipped over. 410 * 411 * Return: %0 on success, negative error code otherwise. 412 */ 413 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start, 414 loff_t end, int sync_mode) 415 { 416 struct writeback_control wbc = { 417 .sync_mode = sync_mode, 418 .nr_to_write = LONG_MAX, 419 .range_start = start, 420 .range_end = end, 421 }; 422 423 return filemap_fdatawrite_wbc(mapping, &wbc); 424 } 425 426 static inline int __filemap_fdatawrite(struct address_space *mapping, 427 int sync_mode) 428 { 429 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode); 430 } 431 432 int filemap_fdatawrite(struct address_space *mapping) 433 { 434 return __filemap_fdatawrite(mapping, WB_SYNC_ALL); 435 } 436 EXPORT_SYMBOL(filemap_fdatawrite); 437 438 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start, 439 loff_t end) 440 { 441 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL); 442 } 443 EXPORT_SYMBOL(filemap_fdatawrite_range); 444 445 /** 446 * filemap_fdatawrite_range_kick - start writeback on a range 447 * @mapping: target address_space 448 * @start: index to start writeback on 449 * @end: last (inclusive) index for writeback 450 * 451 * This is a non-integrity writeback helper, to start writing back folios 452 * for the indicated range. 453 * 454 * Return: %0 on success, negative error code otherwise. 455 */ 456 int filemap_fdatawrite_range_kick(struct address_space *mapping, loff_t start, 457 loff_t end) 458 { 459 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_NONE); 460 } 461 EXPORT_SYMBOL_GPL(filemap_fdatawrite_range_kick); 462 463 /** 464 * filemap_flush - mostly a non-blocking flush 465 * @mapping: target address_space 466 * 467 * This is a mostly non-blocking flush. Not suitable for data-integrity 468 * purposes - I/O may not be started against all dirty pages. 469 * 470 * Return: %0 on success, negative error code otherwise. 471 */ 472 int filemap_flush(struct address_space *mapping) 473 { 474 return __filemap_fdatawrite(mapping, WB_SYNC_NONE); 475 } 476 EXPORT_SYMBOL(filemap_flush); 477 478 /** 479 * filemap_range_has_page - check if a page exists in range. 480 * @mapping: address space within which to check 481 * @start_byte: offset in bytes where the range starts 482 * @end_byte: offset in bytes where the range ends (inclusive) 483 * 484 * Find at least one page in the range supplied, usually used to check if 485 * direct writing in this range will trigger a writeback. 486 * 487 * Return: %true if at least one page exists in the specified range, 488 * %false otherwise. 489 */ 490 bool filemap_range_has_page(struct address_space *mapping, 491 loff_t start_byte, loff_t end_byte) 492 { 493 struct folio *folio; 494 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT); 495 pgoff_t max = end_byte >> PAGE_SHIFT; 496 497 if (end_byte < start_byte) 498 return false; 499 500 rcu_read_lock(); 501 for (;;) { 502 folio = xas_find(&xas, max); 503 if (xas_retry(&xas, folio)) 504 continue; 505 /* Shadow entries don't count */ 506 if (xa_is_value(folio)) 507 continue; 508 /* 509 * We don't need to try to pin this page; we're about to 510 * release the RCU lock anyway. It is enough to know that 511 * there was a page here recently. 512 */ 513 break; 514 } 515 rcu_read_unlock(); 516 517 return folio != NULL; 518 } 519 EXPORT_SYMBOL(filemap_range_has_page); 520 521 static void __filemap_fdatawait_range(struct address_space *mapping, 522 loff_t start_byte, loff_t end_byte) 523 { 524 pgoff_t index = start_byte >> PAGE_SHIFT; 525 pgoff_t end = end_byte >> PAGE_SHIFT; 526 struct folio_batch fbatch; 527 unsigned nr_folios; 528 529 folio_batch_init(&fbatch); 530 531 while (index <= end) { 532 unsigned i; 533 534 nr_folios = filemap_get_folios_tag(mapping, &index, end, 535 PAGECACHE_TAG_WRITEBACK, &fbatch); 536 537 if (!nr_folios) 538 break; 539 540 for (i = 0; i < nr_folios; i++) { 541 struct folio *folio = fbatch.folios[i]; 542 543 folio_wait_writeback(folio); 544 } 545 folio_batch_release(&fbatch); 546 cond_resched(); 547 } 548 } 549 550 /** 551 * filemap_fdatawait_range - wait for writeback to complete 552 * @mapping: address space structure to wait for 553 * @start_byte: offset in bytes where the range starts 554 * @end_byte: offset in bytes where the range ends (inclusive) 555 * 556 * Walk the list of under-writeback pages of the given address space 557 * in the given range and wait for all of them. Check error status of 558 * the address space and return it. 559 * 560 * Since the error status of the address space is cleared by this function, 561 * callers are responsible for checking the return value and handling and/or 562 * reporting the error. 563 * 564 * Return: error status of the address space. 565 */ 566 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte, 567 loff_t end_byte) 568 { 569 __filemap_fdatawait_range(mapping, start_byte, end_byte); 570 return filemap_check_errors(mapping); 571 } 572 EXPORT_SYMBOL(filemap_fdatawait_range); 573 574 /** 575 * filemap_fdatawait_range_keep_errors - wait for writeback to complete 576 * @mapping: address space structure to wait for 577 * @start_byte: offset in bytes where the range starts 578 * @end_byte: offset in bytes where the range ends (inclusive) 579 * 580 * Walk the list of under-writeback pages of the given address space in the 581 * given range and wait for all of them. Unlike filemap_fdatawait_range(), 582 * this function does not clear error status of the address space. 583 * 584 * Use this function if callers don't handle errors themselves. Expected 585 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2), 586 * fsfreeze(8) 587 */ 588 int filemap_fdatawait_range_keep_errors(struct address_space *mapping, 589 loff_t start_byte, loff_t end_byte) 590 { 591 __filemap_fdatawait_range(mapping, start_byte, end_byte); 592 return filemap_check_and_keep_errors(mapping); 593 } 594 EXPORT_SYMBOL(filemap_fdatawait_range_keep_errors); 595 596 /** 597 * file_fdatawait_range - wait for writeback to complete 598 * @file: file pointing to address space structure to wait for 599 * @start_byte: offset in bytes where the range starts 600 * @end_byte: offset in bytes where the range ends (inclusive) 601 * 602 * Walk the list of under-writeback pages of the address space that file 603 * refers to, in the given range and wait for all of them. Check error 604 * status of the address space vs. the file->f_wb_err cursor and return it. 605 * 606 * Since the error status of the file is advanced by this function, 607 * callers are responsible for checking the return value and handling and/or 608 * reporting the error. 609 * 610 * Return: error status of the address space vs. the file->f_wb_err cursor. 611 */ 612 int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte) 613 { 614 struct address_space *mapping = file->f_mapping; 615 616 __filemap_fdatawait_range(mapping, start_byte, end_byte); 617 return file_check_and_advance_wb_err(file); 618 } 619 EXPORT_SYMBOL(file_fdatawait_range); 620 621 /** 622 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors 623 * @mapping: address space structure to wait for 624 * 625 * Walk the list of under-writeback pages of the given address space 626 * and wait for all of them. Unlike filemap_fdatawait(), this function 627 * does not clear error status of the address space. 628 * 629 * Use this function if callers don't handle errors themselves. Expected 630 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2), 631 * fsfreeze(8) 632 * 633 * Return: error status of the address space. 634 */ 635 int filemap_fdatawait_keep_errors(struct address_space *mapping) 636 { 637 __filemap_fdatawait_range(mapping, 0, LLONG_MAX); 638 return filemap_check_and_keep_errors(mapping); 639 } 640 EXPORT_SYMBOL(filemap_fdatawait_keep_errors); 641 642 /* Returns true if writeback might be needed or already in progress. */ 643 static bool mapping_needs_writeback(struct address_space *mapping) 644 { 645 return mapping->nrpages; 646 } 647 648 bool filemap_range_has_writeback(struct address_space *mapping, 649 loff_t start_byte, loff_t end_byte) 650 { 651 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT); 652 pgoff_t max = end_byte >> PAGE_SHIFT; 653 struct folio *folio; 654 655 if (end_byte < start_byte) 656 return false; 657 658 rcu_read_lock(); 659 xas_for_each(&xas, folio, max) { 660 if (xas_retry(&xas, folio)) 661 continue; 662 if (xa_is_value(folio)) 663 continue; 664 if (folio_test_dirty(folio) || folio_test_locked(folio) || 665 folio_test_writeback(folio)) 666 break; 667 } 668 rcu_read_unlock(); 669 return folio != NULL; 670 } 671 EXPORT_SYMBOL_GPL(filemap_range_has_writeback); 672 673 /** 674 * filemap_write_and_wait_range - write out & wait on a file range 675 * @mapping: the address_space for the pages 676 * @lstart: offset in bytes where the range starts 677 * @lend: offset in bytes where the range ends (inclusive) 678 * 679 * Write out and wait upon file offsets lstart->lend, inclusive. 680 * 681 * Note that @lend is inclusive (describes the last byte to be written) so 682 * that this function can be used to write to the very end-of-file (end = -1). 683 * 684 * Return: error status of the address space. 685 */ 686 int filemap_write_and_wait_range(struct address_space *mapping, 687 loff_t lstart, loff_t lend) 688 { 689 int err = 0, err2; 690 691 if (lend < lstart) 692 return 0; 693 694 if (mapping_needs_writeback(mapping)) { 695 err = __filemap_fdatawrite_range(mapping, lstart, lend, 696 WB_SYNC_ALL); 697 /* 698 * Even if the above returned error, the pages may be 699 * written partially (e.g. -ENOSPC), so we wait for it. 700 * But the -EIO is special case, it may indicate the worst 701 * thing (e.g. bug) happened, so we avoid waiting for it. 702 */ 703 if (err != -EIO) 704 __filemap_fdatawait_range(mapping, lstart, lend); 705 } 706 err2 = filemap_check_errors(mapping); 707 if (!err) 708 err = err2; 709 return err; 710 } 711 EXPORT_SYMBOL(filemap_write_and_wait_range); 712 713 void __filemap_set_wb_err(struct address_space *mapping, int err) 714 { 715 errseq_t eseq = errseq_set(&mapping->wb_err, err); 716 717 trace_filemap_set_wb_err(mapping, eseq); 718 } 719 EXPORT_SYMBOL(__filemap_set_wb_err); 720 721 /** 722 * file_check_and_advance_wb_err - report wb error (if any) that was previously 723 * and advance wb_err to current one 724 * @file: struct file on which the error is being reported 725 * 726 * When userland calls fsync (or something like nfsd does the equivalent), we 727 * want to report any writeback errors that occurred since the last fsync (or 728 * since the file was opened if there haven't been any). 729 * 730 * Grab the wb_err from the mapping. If it matches what we have in the file, 731 * then just quickly return 0. The file is all caught up. 732 * 733 * If it doesn't match, then take the mapping value, set the "seen" flag in 734 * it and try to swap it into place. If it works, or another task beat us 735 * to it with the new value, then update the f_wb_err and return the error 736 * portion. The error at this point must be reported via proper channels 737 * (a'la fsync, or NFS COMMIT operation, etc.). 738 * 739 * While we handle mapping->wb_err with atomic operations, the f_wb_err 740 * value is protected by the f_lock since we must ensure that it reflects 741 * the latest value swapped in for this file descriptor. 742 * 743 * Return: %0 on success, negative error code otherwise. 744 */ 745 int file_check_and_advance_wb_err(struct file *file) 746 { 747 int err = 0; 748 errseq_t old = READ_ONCE(file->f_wb_err); 749 struct address_space *mapping = file->f_mapping; 750 751 /* Locklessly handle the common case where nothing has changed */ 752 if (errseq_check(&mapping->wb_err, old)) { 753 /* Something changed, must use slow path */ 754 spin_lock(&file->f_lock); 755 old = file->f_wb_err; 756 err = errseq_check_and_advance(&mapping->wb_err, 757 &file->f_wb_err); 758 trace_file_check_and_advance_wb_err(file, old); 759 spin_unlock(&file->f_lock); 760 } 761 762 /* 763 * We're mostly using this function as a drop in replacement for 764 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect 765 * that the legacy code would have had on these flags. 766 */ 767 clear_bit(AS_EIO, &mapping->flags); 768 clear_bit(AS_ENOSPC, &mapping->flags); 769 return err; 770 } 771 EXPORT_SYMBOL(file_check_and_advance_wb_err); 772 773 /** 774 * file_write_and_wait_range - write out & wait on a file range 775 * @file: file pointing to address_space with pages 776 * @lstart: offset in bytes where the range starts 777 * @lend: offset in bytes where the range ends (inclusive) 778 * 779 * Write out and wait upon file offsets lstart->lend, inclusive. 780 * 781 * Note that @lend is inclusive (describes the last byte to be written) so 782 * that this function can be used to write to the very end-of-file (end = -1). 783 * 784 * After writing out and waiting on the data, we check and advance the 785 * f_wb_err cursor to the latest value, and return any errors detected there. 786 * 787 * Return: %0 on success, negative error code otherwise. 788 */ 789 int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend) 790 { 791 int err = 0, err2; 792 struct address_space *mapping = file->f_mapping; 793 794 if (lend < lstart) 795 return 0; 796 797 if (mapping_needs_writeback(mapping)) { 798 err = __filemap_fdatawrite_range(mapping, lstart, lend, 799 WB_SYNC_ALL); 800 /* See comment of filemap_write_and_wait() */ 801 if (err != -EIO) 802 __filemap_fdatawait_range(mapping, lstart, lend); 803 } 804 err2 = file_check_and_advance_wb_err(file); 805 if (!err) 806 err = err2; 807 return err; 808 } 809 EXPORT_SYMBOL(file_write_and_wait_range); 810 811 /** 812 * replace_page_cache_folio - replace a pagecache folio with a new one 813 * @old: folio to be replaced 814 * @new: folio to replace with 815 * 816 * This function replaces a folio in the pagecache with a new one. On 817 * success it acquires the pagecache reference for the new folio and 818 * drops it for the old folio. Both the old and new folios must be 819 * locked. This function does not add the new folio to the LRU, the 820 * caller must do that. 821 * 822 * The remove + add is atomic. This function cannot fail. 823 */ 824 void replace_page_cache_folio(struct folio *old, struct folio *new) 825 { 826 struct address_space *mapping = old->mapping; 827 void (*free_folio)(struct folio *) = mapping->a_ops->free_folio; 828 pgoff_t offset = old->index; 829 XA_STATE(xas, &mapping->i_pages, offset); 830 831 VM_BUG_ON_FOLIO(!folio_test_locked(old), old); 832 VM_BUG_ON_FOLIO(!folio_test_locked(new), new); 833 VM_BUG_ON_FOLIO(new->mapping, new); 834 835 folio_get(new); 836 new->mapping = mapping; 837 new->index = offset; 838 839 mem_cgroup_replace_folio(old, new); 840 841 xas_lock_irq(&xas); 842 xas_store(&xas, new); 843 844 old->mapping = NULL; 845 /* hugetlb pages do not participate in page cache accounting. */ 846 if (!folio_test_hugetlb(old)) 847 __lruvec_stat_sub_folio(old, NR_FILE_PAGES); 848 if (!folio_test_hugetlb(new)) 849 __lruvec_stat_add_folio(new, NR_FILE_PAGES); 850 if (folio_test_swapbacked(old)) 851 __lruvec_stat_sub_folio(old, NR_SHMEM); 852 if (folio_test_swapbacked(new)) 853 __lruvec_stat_add_folio(new, NR_SHMEM); 854 xas_unlock_irq(&xas); 855 if (free_folio) 856 free_folio(old); 857 folio_put(old); 858 } 859 EXPORT_SYMBOL_GPL(replace_page_cache_folio); 860 861 noinline int __filemap_add_folio(struct address_space *mapping, 862 struct folio *folio, pgoff_t index, gfp_t gfp, void **shadowp) 863 { 864 XA_STATE_ORDER(xas, &mapping->i_pages, index, folio_order(folio)); 865 bool huge; 866 long nr; 867 unsigned int forder = folio_order(folio); 868 869 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 870 VM_BUG_ON_FOLIO(folio_test_swapbacked(folio), folio); 871 VM_BUG_ON_FOLIO(folio_order(folio) < mapping_min_folio_order(mapping), 872 folio); 873 mapping_set_update(&xas, mapping); 874 875 VM_BUG_ON_FOLIO(index & (folio_nr_pages(folio) - 1), folio); 876 huge = folio_test_hugetlb(folio); 877 nr = folio_nr_pages(folio); 878 879 gfp &= GFP_RECLAIM_MASK; 880 folio_ref_add(folio, nr); 881 folio->mapping = mapping; 882 folio->index = xas.xa_index; 883 884 for (;;) { 885 int order = -1; 886 void *entry, *old = NULL; 887 888 xas_lock_irq(&xas); 889 xas_for_each_conflict(&xas, entry) { 890 old = entry; 891 if (!xa_is_value(entry)) { 892 xas_set_err(&xas, -EEXIST); 893 goto unlock; 894 } 895 /* 896 * If a larger entry exists, 897 * it will be the first and only entry iterated. 898 */ 899 if (order == -1) 900 order = xas_get_order(&xas); 901 } 902 903 if (old) { 904 if (order > 0 && order > forder) { 905 unsigned int split_order = max(forder, 906 xas_try_split_min_order(order)); 907 908 /* How to handle large swap entries? */ 909 BUG_ON(shmem_mapping(mapping)); 910 911 while (order > forder) { 912 xas_set_order(&xas, index, split_order); 913 xas_try_split(&xas, old, order); 914 if (xas_error(&xas)) 915 goto unlock; 916 order = split_order; 917 split_order = 918 max(xas_try_split_min_order( 919 split_order), 920 forder); 921 } 922 xas_reset(&xas); 923 } 924 if (shadowp) 925 *shadowp = old; 926 } 927 928 xas_store(&xas, folio); 929 if (xas_error(&xas)) 930 goto unlock; 931 932 mapping->nrpages += nr; 933 934 /* hugetlb pages do not participate in page cache accounting */ 935 if (!huge) { 936 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, nr); 937 if (folio_test_pmd_mappable(folio)) 938 __lruvec_stat_mod_folio(folio, 939 NR_FILE_THPS, nr); 940 } 941 942 unlock: 943 xas_unlock_irq(&xas); 944 945 if (!xas_nomem(&xas, gfp)) 946 break; 947 } 948 949 if (xas_error(&xas)) 950 goto error; 951 952 trace_mm_filemap_add_to_page_cache(folio); 953 return 0; 954 error: 955 folio->mapping = NULL; 956 /* Leave folio->index set: truncation relies upon it */ 957 folio_put_refs(folio, nr); 958 return xas_error(&xas); 959 } 960 ALLOW_ERROR_INJECTION(__filemap_add_folio, ERRNO); 961 962 int filemap_add_folio(struct address_space *mapping, struct folio *folio, 963 pgoff_t index, gfp_t gfp) 964 { 965 void *shadow = NULL; 966 int ret; 967 struct mem_cgroup *tmp; 968 bool kernel_file = test_bit(AS_KERNEL_FILE, &mapping->flags); 969 970 if (kernel_file) 971 tmp = set_active_memcg(root_mem_cgroup); 972 ret = mem_cgroup_charge(folio, NULL, gfp); 973 if (kernel_file) 974 set_active_memcg(tmp); 975 if (ret) 976 return ret; 977 978 __folio_set_locked(folio); 979 ret = __filemap_add_folio(mapping, folio, index, gfp, &shadow); 980 if (unlikely(ret)) { 981 mem_cgroup_uncharge(folio); 982 __folio_clear_locked(folio); 983 } else { 984 /* 985 * The folio might have been evicted from cache only 986 * recently, in which case it should be activated like 987 * any other repeatedly accessed folio. 988 * The exception is folios getting rewritten; evicting other 989 * data from the working set, only to cache data that will 990 * get overwritten with something else, is a waste of memory. 991 */ 992 WARN_ON_ONCE(folio_test_active(folio)); 993 if (!(gfp & __GFP_WRITE) && shadow) 994 workingset_refault(folio, shadow); 995 folio_add_lru(folio); 996 if (kernel_file) 997 mod_node_page_state(folio_pgdat(folio), 998 NR_KERNEL_FILE_PAGES, 999 folio_nr_pages(folio)); 1000 } 1001 return ret; 1002 } 1003 EXPORT_SYMBOL_GPL(filemap_add_folio); 1004 1005 #ifdef CONFIG_NUMA 1006 struct folio *filemap_alloc_folio_noprof(gfp_t gfp, unsigned int order) 1007 { 1008 int n; 1009 struct folio *folio; 1010 1011 if (cpuset_do_page_mem_spread()) { 1012 unsigned int cpuset_mems_cookie; 1013 do { 1014 cpuset_mems_cookie = read_mems_allowed_begin(); 1015 n = cpuset_mem_spread_node(); 1016 folio = __folio_alloc_node_noprof(gfp, order, n); 1017 } while (!folio && read_mems_allowed_retry(cpuset_mems_cookie)); 1018 1019 return folio; 1020 } 1021 return folio_alloc_noprof(gfp, order); 1022 } 1023 EXPORT_SYMBOL(filemap_alloc_folio_noprof); 1024 #endif 1025 1026 /* 1027 * filemap_invalidate_lock_two - lock invalidate_lock for two mappings 1028 * 1029 * Lock exclusively invalidate_lock of any passed mapping that is not NULL. 1030 * 1031 * @mapping1: the first mapping to lock 1032 * @mapping2: the second mapping to lock 1033 */ 1034 void filemap_invalidate_lock_two(struct address_space *mapping1, 1035 struct address_space *mapping2) 1036 { 1037 if (mapping1 > mapping2) 1038 swap(mapping1, mapping2); 1039 if (mapping1) 1040 down_write(&mapping1->invalidate_lock); 1041 if (mapping2 && mapping1 != mapping2) 1042 down_write_nested(&mapping2->invalidate_lock, 1); 1043 } 1044 EXPORT_SYMBOL(filemap_invalidate_lock_two); 1045 1046 /* 1047 * filemap_invalidate_unlock_two - unlock invalidate_lock for two mappings 1048 * 1049 * Unlock exclusive invalidate_lock of any passed mapping that is not NULL. 1050 * 1051 * @mapping1: the first mapping to unlock 1052 * @mapping2: the second mapping to unlock 1053 */ 1054 void filemap_invalidate_unlock_two(struct address_space *mapping1, 1055 struct address_space *mapping2) 1056 { 1057 if (mapping1) 1058 up_write(&mapping1->invalidate_lock); 1059 if (mapping2 && mapping1 != mapping2) 1060 up_write(&mapping2->invalidate_lock); 1061 } 1062 EXPORT_SYMBOL(filemap_invalidate_unlock_two); 1063 1064 /* 1065 * In order to wait for pages to become available there must be 1066 * waitqueues associated with pages. By using a hash table of 1067 * waitqueues where the bucket discipline is to maintain all 1068 * waiters on the same queue and wake all when any of the pages 1069 * become available, and for the woken contexts to check to be 1070 * sure the appropriate page became available, this saves space 1071 * at a cost of "thundering herd" phenomena during rare hash 1072 * collisions. 1073 */ 1074 #define PAGE_WAIT_TABLE_BITS 8 1075 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS) 1076 static wait_queue_head_t folio_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned; 1077 1078 static wait_queue_head_t *folio_waitqueue(struct folio *folio) 1079 { 1080 return &folio_wait_table[hash_ptr(folio, PAGE_WAIT_TABLE_BITS)]; 1081 } 1082 1083 /* How many times do we accept lock stealing from under a waiter? */ 1084 static int sysctl_page_lock_unfairness = 5; 1085 static const struct ctl_table filemap_sysctl_table[] = { 1086 { 1087 .procname = "page_lock_unfairness", 1088 .data = &sysctl_page_lock_unfairness, 1089 .maxlen = sizeof(sysctl_page_lock_unfairness), 1090 .mode = 0644, 1091 .proc_handler = proc_dointvec_minmax, 1092 .extra1 = SYSCTL_ZERO, 1093 } 1094 }; 1095 1096 void __init pagecache_init(void) 1097 { 1098 int i; 1099 1100 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++) 1101 init_waitqueue_head(&folio_wait_table[i]); 1102 1103 page_writeback_init(); 1104 register_sysctl_init("vm", filemap_sysctl_table); 1105 } 1106 1107 /* 1108 * The page wait code treats the "wait->flags" somewhat unusually, because 1109 * we have multiple different kinds of waits, not just the usual "exclusive" 1110 * one. 1111 * 1112 * We have: 1113 * 1114 * (a) no special bits set: 1115 * 1116 * We're just waiting for the bit to be released, and when a waker 1117 * calls the wakeup function, we set WQ_FLAG_WOKEN and wake it up, 1118 * and remove it from the wait queue. 1119 * 1120 * Simple and straightforward. 1121 * 1122 * (b) WQ_FLAG_EXCLUSIVE: 1123 * 1124 * The waiter is waiting to get the lock, and only one waiter should 1125 * be woken up to avoid any thundering herd behavior. We'll set the 1126 * WQ_FLAG_WOKEN bit, wake it up, and remove it from the wait queue. 1127 * 1128 * This is the traditional exclusive wait. 1129 * 1130 * (c) WQ_FLAG_EXCLUSIVE | WQ_FLAG_CUSTOM: 1131 * 1132 * The waiter is waiting to get the bit, and additionally wants the 1133 * lock to be transferred to it for fair lock behavior. If the lock 1134 * cannot be taken, we stop walking the wait queue without waking 1135 * the waiter. 1136 * 1137 * This is the "fair lock handoff" case, and in addition to setting 1138 * WQ_FLAG_WOKEN, we set WQ_FLAG_DONE to let the waiter easily see 1139 * that it now has the lock. 1140 */ 1141 static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg) 1142 { 1143 unsigned int flags; 1144 struct wait_page_key *key = arg; 1145 struct wait_page_queue *wait_page 1146 = container_of(wait, struct wait_page_queue, wait); 1147 1148 if (!wake_page_match(wait_page, key)) 1149 return 0; 1150 1151 /* 1152 * If it's a lock handoff wait, we get the bit for it, and 1153 * stop walking (and do not wake it up) if we can't. 1154 */ 1155 flags = wait->flags; 1156 if (flags & WQ_FLAG_EXCLUSIVE) { 1157 if (test_bit(key->bit_nr, &key->folio->flags.f)) 1158 return -1; 1159 if (flags & WQ_FLAG_CUSTOM) { 1160 if (test_and_set_bit(key->bit_nr, &key->folio->flags.f)) 1161 return -1; 1162 flags |= WQ_FLAG_DONE; 1163 } 1164 } 1165 1166 /* 1167 * We are holding the wait-queue lock, but the waiter that 1168 * is waiting for this will be checking the flags without 1169 * any locking. 1170 * 1171 * So update the flags atomically, and wake up the waiter 1172 * afterwards to avoid any races. This store-release pairs 1173 * with the load-acquire in folio_wait_bit_common(). 1174 */ 1175 smp_store_release(&wait->flags, flags | WQ_FLAG_WOKEN); 1176 wake_up_state(wait->private, mode); 1177 1178 /* 1179 * Ok, we have successfully done what we're waiting for, 1180 * and we can unconditionally remove the wait entry. 1181 * 1182 * Note that this pairs with the "finish_wait()" in the 1183 * waiter, and has to be the absolute last thing we do. 1184 * After this list_del_init(&wait->entry) the wait entry 1185 * might be de-allocated and the process might even have 1186 * exited. 1187 */ 1188 list_del_init_careful(&wait->entry); 1189 return (flags & WQ_FLAG_EXCLUSIVE) != 0; 1190 } 1191 1192 static void folio_wake_bit(struct folio *folio, int bit_nr) 1193 { 1194 wait_queue_head_t *q = folio_waitqueue(folio); 1195 struct wait_page_key key; 1196 unsigned long flags; 1197 1198 key.folio = folio; 1199 key.bit_nr = bit_nr; 1200 key.page_match = 0; 1201 1202 spin_lock_irqsave(&q->lock, flags); 1203 __wake_up_locked_key(q, TASK_NORMAL, &key); 1204 1205 /* 1206 * It's possible to miss clearing waiters here, when we woke our page 1207 * waiters, but the hashed waitqueue has waiters for other pages on it. 1208 * That's okay, it's a rare case. The next waker will clear it. 1209 * 1210 * Note that, depending on the page pool (buddy, hugetlb, ZONE_DEVICE, 1211 * other), the flag may be cleared in the course of freeing the page; 1212 * but that is not required for correctness. 1213 */ 1214 if (!waitqueue_active(q) || !key.page_match) 1215 folio_clear_waiters(folio); 1216 1217 spin_unlock_irqrestore(&q->lock, flags); 1218 } 1219 1220 /* 1221 * A choice of three behaviors for folio_wait_bit_common(): 1222 */ 1223 enum behavior { 1224 EXCLUSIVE, /* Hold ref to page and take the bit when woken, like 1225 * __folio_lock() waiting on then setting PG_locked. 1226 */ 1227 SHARED, /* Hold ref to page and check the bit when woken, like 1228 * folio_wait_writeback() waiting on PG_writeback. 1229 */ 1230 DROP, /* Drop ref to page before wait, no check when woken, 1231 * like folio_put_wait_locked() on PG_locked. 1232 */ 1233 }; 1234 1235 /* 1236 * Attempt to check (or get) the folio flag, and mark us done 1237 * if successful. 1238 */ 1239 static inline bool folio_trylock_flag(struct folio *folio, int bit_nr, 1240 struct wait_queue_entry *wait) 1241 { 1242 if (wait->flags & WQ_FLAG_EXCLUSIVE) { 1243 if (test_and_set_bit(bit_nr, &folio->flags.f)) 1244 return false; 1245 } else if (test_bit(bit_nr, &folio->flags.f)) 1246 return false; 1247 1248 wait->flags |= WQ_FLAG_WOKEN | WQ_FLAG_DONE; 1249 return true; 1250 } 1251 1252 static inline int folio_wait_bit_common(struct folio *folio, int bit_nr, 1253 int state, enum behavior behavior) 1254 { 1255 wait_queue_head_t *q = folio_waitqueue(folio); 1256 int unfairness = sysctl_page_lock_unfairness; 1257 struct wait_page_queue wait_page; 1258 wait_queue_entry_t *wait = &wait_page.wait; 1259 bool thrashing = false; 1260 unsigned long pflags; 1261 bool in_thrashing; 1262 1263 if (bit_nr == PG_locked && 1264 !folio_test_uptodate(folio) && folio_test_workingset(folio)) { 1265 delayacct_thrashing_start(&in_thrashing); 1266 psi_memstall_enter(&pflags); 1267 thrashing = true; 1268 } 1269 1270 init_wait(wait); 1271 wait->func = wake_page_function; 1272 wait_page.folio = folio; 1273 wait_page.bit_nr = bit_nr; 1274 1275 repeat: 1276 wait->flags = 0; 1277 if (behavior == EXCLUSIVE) { 1278 wait->flags = WQ_FLAG_EXCLUSIVE; 1279 if (--unfairness < 0) 1280 wait->flags |= WQ_FLAG_CUSTOM; 1281 } 1282 1283 /* 1284 * Do one last check whether we can get the 1285 * page bit synchronously. 1286 * 1287 * Do the folio_set_waiters() marking before that 1288 * to let any waker we _just_ missed know they 1289 * need to wake us up (otherwise they'll never 1290 * even go to the slow case that looks at the 1291 * page queue), and add ourselves to the wait 1292 * queue if we need to sleep. 1293 * 1294 * This part needs to be done under the queue 1295 * lock to avoid races. 1296 */ 1297 spin_lock_irq(&q->lock); 1298 folio_set_waiters(folio); 1299 if (!folio_trylock_flag(folio, bit_nr, wait)) 1300 __add_wait_queue_entry_tail(q, wait); 1301 spin_unlock_irq(&q->lock); 1302 1303 /* 1304 * From now on, all the logic will be based on 1305 * the WQ_FLAG_WOKEN and WQ_FLAG_DONE flag, to 1306 * see whether the page bit testing has already 1307 * been done by the wake function. 1308 * 1309 * We can drop our reference to the folio. 1310 */ 1311 if (behavior == DROP) 1312 folio_put(folio); 1313 1314 /* 1315 * Note that until the "finish_wait()", or until 1316 * we see the WQ_FLAG_WOKEN flag, we need to 1317 * be very careful with the 'wait->flags', because 1318 * we may race with a waker that sets them. 1319 */ 1320 for (;;) { 1321 unsigned int flags; 1322 1323 set_current_state(state); 1324 1325 /* Loop until we've been woken or interrupted */ 1326 flags = smp_load_acquire(&wait->flags); 1327 if (!(flags & WQ_FLAG_WOKEN)) { 1328 if (signal_pending_state(state, current)) 1329 break; 1330 1331 io_schedule(); 1332 continue; 1333 } 1334 1335 /* If we were non-exclusive, we're done */ 1336 if (behavior != EXCLUSIVE) 1337 break; 1338 1339 /* If the waker got the lock for us, we're done */ 1340 if (flags & WQ_FLAG_DONE) 1341 break; 1342 1343 /* 1344 * Otherwise, if we're getting the lock, we need to 1345 * try to get it ourselves. 1346 * 1347 * And if that fails, we'll have to retry this all. 1348 */ 1349 if (unlikely(test_and_set_bit(bit_nr, folio_flags(folio, 0)))) 1350 goto repeat; 1351 1352 wait->flags |= WQ_FLAG_DONE; 1353 break; 1354 } 1355 1356 /* 1357 * If a signal happened, this 'finish_wait()' may remove the last 1358 * waiter from the wait-queues, but the folio waiters bit will remain 1359 * set. That's ok. The next wakeup will take care of it, and trying 1360 * to do it here would be difficult and prone to races. 1361 */ 1362 finish_wait(q, wait); 1363 1364 if (thrashing) { 1365 delayacct_thrashing_end(&in_thrashing); 1366 psi_memstall_leave(&pflags); 1367 } 1368 1369 /* 1370 * NOTE! The wait->flags weren't stable until we've done the 1371 * 'finish_wait()', and we could have exited the loop above due 1372 * to a signal, and had a wakeup event happen after the signal 1373 * test but before the 'finish_wait()'. 1374 * 1375 * So only after the finish_wait() can we reliably determine 1376 * if we got woken up or not, so we can now figure out the final 1377 * return value based on that state without races. 1378 * 1379 * Also note that WQ_FLAG_WOKEN is sufficient for a non-exclusive 1380 * waiter, but an exclusive one requires WQ_FLAG_DONE. 1381 */ 1382 if (behavior == EXCLUSIVE) 1383 return wait->flags & WQ_FLAG_DONE ? 0 : -EINTR; 1384 1385 return wait->flags & WQ_FLAG_WOKEN ? 0 : -EINTR; 1386 } 1387 1388 #ifdef CONFIG_MIGRATION 1389 /** 1390 * migration_entry_wait_on_locked - Wait for a migration entry or 1391 * device_private entry to be removed. 1392 * @entry: migration or device_private swap entry. 1393 * @ptl: already locked ptl. This function will drop the lock. 1394 * 1395 * Wait for a migration entry referencing the given page, or device_private 1396 * entry referencing a dvice_private page to be unlocked. This is 1397 * equivalent to folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE) except 1398 * this can be called without taking a reference on the page. Instead this 1399 * should be called while holding the ptl for @entry referencing 1400 * the page. 1401 * 1402 * Returns after unlocking the ptl. 1403 * 1404 * This follows the same logic as folio_wait_bit_common() so see the comments 1405 * there. 1406 */ 1407 void migration_entry_wait_on_locked(swp_entry_t entry, spinlock_t *ptl) 1408 __releases(ptl) 1409 { 1410 struct wait_page_queue wait_page; 1411 wait_queue_entry_t *wait = &wait_page.wait; 1412 bool thrashing = false; 1413 unsigned long pflags; 1414 bool in_thrashing; 1415 wait_queue_head_t *q; 1416 struct folio *folio = pfn_swap_entry_folio(entry); 1417 1418 q = folio_waitqueue(folio); 1419 if (!folio_test_uptodate(folio) && folio_test_workingset(folio)) { 1420 delayacct_thrashing_start(&in_thrashing); 1421 psi_memstall_enter(&pflags); 1422 thrashing = true; 1423 } 1424 1425 init_wait(wait); 1426 wait->func = wake_page_function; 1427 wait_page.folio = folio; 1428 wait_page.bit_nr = PG_locked; 1429 wait->flags = 0; 1430 1431 spin_lock_irq(&q->lock); 1432 folio_set_waiters(folio); 1433 if (!folio_trylock_flag(folio, PG_locked, wait)) 1434 __add_wait_queue_entry_tail(q, wait); 1435 spin_unlock_irq(&q->lock); 1436 1437 /* 1438 * If a migration entry exists for the page the migration path must hold 1439 * a valid reference to the page, and it must take the ptl to remove the 1440 * migration entry. So the page is valid until the ptl is dropped. 1441 * Similarly any path attempting to drop the last reference to a 1442 * device-private page needs to grab the ptl to remove the device-private 1443 * entry. 1444 */ 1445 spin_unlock(ptl); 1446 1447 for (;;) { 1448 unsigned int flags; 1449 1450 set_current_state(TASK_UNINTERRUPTIBLE); 1451 1452 /* Loop until we've been woken or interrupted */ 1453 flags = smp_load_acquire(&wait->flags); 1454 if (!(flags & WQ_FLAG_WOKEN)) { 1455 if (signal_pending_state(TASK_UNINTERRUPTIBLE, current)) 1456 break; 1457 1458 io_schedule(); 1459 continue; 1460 } 1461 break; 1462 } 1463 1464 finish_wait(q, wait); 1465 1466 if (thrashing) { 1467 delayacct_thrashing_end(&in_thrashing); 1468 psi_memstall_leave(&pflags); 1469 } 1470 } 1471 #endif 1472 1473 void folio_wait_bit(struct folio *folio, int bit_nr) 1474 { 1475 folio_wait_bit_common(folio, bit_nr, TASK_UNINTERRUPTIBLE, SHARED); 1476 } 1477 EXPORT_SYMBOL(folio_wait_bit); 1478 1479 int folio_wait_bit_killable(struct folio *folio, int bit_nr) 1480 { 1481 return folio_wait_bit_common(folio, bit_nr, TASK_KILLABLE, SHARED); 1482 } 1483 EXPORT_SYMBOL(folio_wait_bit_killable); 1484 1485 /** 1486 * folio_put_wait_locked - Drop a reference and wait for it to be unlocked 1487 * @folio: The folio to wait for. 1488 * @state: The sleep state (TASK_KILLABLE, TASK_UNINTERRUPTIBLE, etc). 1489 * 1490 * The caller should hold a reference on @folio. They expect the page to 1491 * become unlocked relatively soon, but do not wish to hold up migration 1492 * (for example) by holding the reference while waiting for the folio to 1493 * come unlocked. After this function returns, the caller should not 1494 * dereference @folio. 1495 * 1496 * Return: 0 if the folio was unlocked or -EINTR if interrupted by a signal. 1497 */ 1498 static int folio_put_wait_locked(struct folio *folio, int state) 1499 { 1500 return folio_wait_bit_common(folio, PG_locked, state, DROP); 1501 } 1502 1503 /** 1504 * folio_unlock - Unlock a locked folio. 1505 * @folio: The folio. 1506 * 1507 * Unlocks the folio and wakes up any thread sleeping on the page lock. 1508 * 1509 * Context: May be called from interrupt or process context. May not be 1510 * called from NMI context. 1511 */ 1512 void folio_unlock(struct folio *folio) 1513 { 1514 /* Bit 7 allows x86 to check the byte's sign bit */ 1515 BUILD_BUG_ON(PG_waiters != 7); 1516 BUILD_BUG_ON(PG_locked > 7); 1517 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 1518 if (folio_xor_flags_has_waiters(folio, 1 << PG_locked)) 1519 folio_wake_bit(folio, PG_locked); 1520 } 1521 EXPORT_SYMBOL(folio_unlock); 1522 1523 /** 1524 * folio_end_read - End read on a folio. 1525 * @folio: The folio. 1526 * @success: True if all reads completed successfully. 1527 * 1528 * When all reads against a folio have completed, filesystems should 1529 * call this function to let the pagecache know that no more reads 1530 * are outstanding. This will unlock the folio and wake up any thread 1531 * sleeping on the lock. The folio will also be marked uptodate if all 1532 * reads succeeded. 1533 * 1534 * Context: May be called from interrupt or process context. May not be 1535 * called from NMI context. 1536 */ 1537 void folio_end_read(struct folio *folio, bool success) 1538 { 1539 unsigned long mask = 1 << PG_locked; 1540 1541 /* Must be in bottom byte for x86 to work */ 1542 BUILD_BUG_ON(PG_uptodate > 7); 1543 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 1544 VM_BUG_ON_FOLIO(success && folio_test_uptodate(folio), folio); 1545 1546 if (likely(success)) 1547 mask |= 1 << PG_uptodate; 1548 if (folio_xor_flags_has_waiters(folio, mask)) 1549 folio_wake_bit(folio, PG_locked); 1550 } 1551 EXPORT_SYMBOL(folio_end_read); 1552 1553 /** 1554 * folio_end_private_2 - Clear PG_private_2 and wake any waiters. 1555 * @folio: The folio. 1556 * 1557 * Clear the PG_private_2 bit on a folio and wake up any sleepers waiting for 1558 * it. The folio reference held for PG_private_2 being set is released. 1559 * 1560 * This is, for example, used when a netfs folio is being written to a local 1561 * disk cache, thereby allowing writes to the cache for the same folio to be 1562 * serialised. 1563 */ 1564 void folio_end_private_2(struct folio *folio) 1565 { 1566 VM_BUG_ON_FOLIO(!folio_test_private_2(folio), folio); 1567 clear_bit_unlock(PG_private_2, folio_flags(folio, 0)); 1568 folio_wake_bit(folio, PG_private_2); 1569 folio_put(folio); 1570 } 1571 EXPORT_SYMBOL(folio_end_private_2); 1572 1573 /** 1574 * folio_wait_private_2 - Wait for PG_private_2 to be cleared on a folio. 1575 * @folio: The folio to wait on. 1576 * 1577 * Wait for PG_private_2 to be cleared on a folio. 1578 */ 1579 void folio_wait_private_2(struct folio *folio) 1580 { 1581 while (folio_test_private_2(folio)) 1582 folio_wait_bit(folio, PG_private_2); 1583 } 1584 EXPORT_SYMBOL(folio_wait_private_2); 1585 1586 /** 1587 * folio_wait_private_2_killable - Wait for PG_private_2 to be cleared on a folio. 1588 * @folio: The folio to wait on. 1589 * 1590 * Wait for PG_private_2 to be cleared on a folio or until a fatal signal is 1591 * received by the calling task. 1592 * 1593 * Return: 1594 * - 0 if successful. 1595 * - -EINTR if a fatal signal was encountered. 1596 */ 1597 int folio_wait_private_2_killable(struct folio *folio) 1598 { 1599 int ret = 0; 1600 1601 while (folio_test_private_2(folio)) { 1602 ret = folio_wait_bit_killable(folio, PG_private_2); 1603 if (ret < 0) 1604 break; 1605 } 1606 1607 return ret; 1608 } 1609 EXPORT_SYMBOL(folio_wait_private_2_killable); 1610 1611 static void filemap_end_dropbehind(struct folio *folio) 1612 { 1613 struct address_space *mapping = folio->mapping; 1614 1615 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 1616 1617 if (folio_test_writeback(folio) || folio_test_dirty(folio)) 1618 return; 1619 if (!folio_test_clear_dropbehind(folio)) 1620 return; 1621 if (mapping) 1622 folio_unmap_invalidate(mapping, folio, 0); 1623 } 1624 1625 /* 1626 * If folio was marked as dropbehind, then pages should be dropped when writeback 1627 * completes. Do that now. If we fail, it's likely because of a big folio - 1628 * just reset dropbehind for that case and latter completions should invalidate. 1629 */ 1630 void folio_end_dropbehind(struct folio *folio) 1631 { 1632 if (!folio_test_dropbehind(folio)) 1633 return; 1634 1635 /* 1636 * Hitting !in_task() should not happen off RWF_DONTCACHE writeback, 1637 * but can happen if normal writeback just happens to find dirty folios 1638 * that were created as part of uncached writeback, and that writeback 1639 * would otherwise not need non-IRQ handling. Just skip the 1640 * invalidation in that case. 1641 */ 1642 if (in_task() && folio_trylock(folio)) { 1643 filemap_end_dropbehind(folio); 1644 folio_unlock(folio); 1645 } 1646 } 1647 EXPORT_SYMBOL_GPL(folio_end_dropbehind); 1648 1649 /** 1650 * folio_end_writeback_no_dropbehind - End writeback against a folio. 1651 * @folio: The folio. 1652 * 1653 * The folio must actually be under writeback. 1654 * This call is intended for filesystems that need to defer dropbehind. 1655 * 1656 * Context: May be called from process or interrupt context. 1657 */ 1658 void folio_end_writeback_no_dropbehind(struct folio *folio) 1659 { 1660 VM_BUG_ON_FOLIO(!folio_test_writeback(folio), folio); 1661 1662 /* 1663 * folio_test_clear_reclaim() could be used here but it is an 1664 * atomic operation and overkill in this particular case. Failing 1665 * to shuffle a folio marked for immediate reclaim is too mild 1666 * a gain to justify taking an atomic operation penalty at the 1667 * end of every folio writeback. 1668 */ 1669 if (folio_test_reclaim(folio)) { 1670 folio_clear_reclaim(folio); 1671 folio_rotate_reclaimable(folio); 1672 } 1673 1674 if (__folio_end_writeback(folio)) 1675 folio_wake_bit(folio, PG_writeback); 1676 1677 acct_reclaim_writeback(folio); 1678 } 1679 EXPORT_SYMBOL_GPL(folio_end_writeback_no_dropbehind); 1680 1681 /** 1682 * folio_end_writeback - End writeback against a folio. 1683 * @folio: The folio. 1684 * 1685 * The folio must actually be under writeback. 1686 * 1687 * Context: May be called from process or interrupt context. 1688 */ 1689 void folio_end_writeback(struct folio *folio) 1690 { 1691 VM_BUG_ON_FOLIO(!folio_test_writeback(folio), folio); 1692 1693 /* 1694 * Writeback does not hold a folio reference of its own, relying 1695 * on truncation to wait for the clearing of PG_writeback. 1696 * But here we must make sure that the folio is not freed and 1697 * reused before the folio_wake_bit(). 1698 */ 1699 folio_get(folio); 1700 folio_end_writeback_no_dropbehind(folio); 1701 folio_end_dropbehind(folio); 1702 folio_put(folio); 1703 } 1704 EXPORT_SYMBOL(folio_end_writeback); 1705 1706 /** 1707 * __folio_lock - Get a lock on the folio, assuming we need to sleep to get it. 1708 * @folio: The folio to lock 1709 */ 1710 void __folio_lock(struct folio *folio) 1711 { 1712 folio_wait_bit_common(folio, PG_locked, TASK_UNINTERRUPTIBLE, 1713 EXCLUSIVE); 1714 } 1715 EXPORT_SYMBOL(__folio_lock); 1716 1717 int __folio_lock_killable(struct folio *folio) 1718 { 1719 return folio_wait_bit_common(folio, PG_locked, TASK_KILLABLE, 1720 EXCLUSIVE); 1721 } 1722 EXPORT_SYMBOL_GPL(__folio_lock_killable); 1723 1724 static int __folio_lock_async(struct folio *folio, struct wait_page_queue *wait) 1725 { 1726 struct wait_queue_head *q = folio_waitqueue(folio); 1727 int ret; 1728 1729 wait->folio = folio; 1730 wait->bit_nr = PG_locked; 1731 1732 spin_lock_irq(&q->lock); 1733 __add_wait_queue_entry_tail(q, &wait->wait); 1734 folio_set_waiters(folio); 1735 ret = !folio_trylock(folio); 1736 /* 1737 * If we were successful now, we know we're still on the 1738 * waitqueue as we're still under the lock. This means it's 1739 * safe to remove and return success, we know the callback 1740 * isn't going to trigger. 1741 */ 1742 if (!ret) 1743 __remove_wait_queue(q, &wait->wait); 1744 else 1745 ret = -EIOCBQUEUED; 1746 spin_unlock_irq(&q->lock); 1747 return ret; 1748 } 1749 1750 /* 1751 * Return values: 1752 * 0 - folio is locked. 1753 * non-zero - folio is not locked. 1754 * mmap_lock or per-VMA lock has been released (mmap_read_unlock() or 1755 * vma_end_read()), unless flags had both FAULT_FLAG_ALLOW_RETRY and 1756 * FAULT_FLAG_RETRY_NOWAIT set, in which case the lock is still held. 1757 * 1758 * If neither ALLOW_RETRY nor KILLABLE are set, will always return 0 1759 * with the folio locked and the mmap_lock/per-VMA lock is left unperturbed. 1760 */ 1761 vm_fault_t __folio_lock_or_retry(struct folio *folio, struct vm_fault *vmf) 1762 { 1763 unsigned int flags = vmf->flags; 1764 1765 if (fault_flag_allow_retry_first(flags)) { 1766 /* 1767 * CAUTION! In this case, mmap_lock/per-VMA lock is not 1768 * released even though returning VM_FAULT_RETRY. 1769 */ 1770 if (flags & FAULT_FLAG_RETRY_NOWAIT) 1771 return VM_FAULT_RETRY; 1772 1773 release_fault_lock(vmf); 1774 if (flags & FAULT_FLAG_KILLABLE) 1775 folio_wait_locked_killable(folio); 1776 else 1777 folio_wait_locked(folio); 1778 return VM_FAULT_RETRY; 1779 } 1780 if (flags & FAULT_FLAG_KILLABLE) { 1781 bool ret; 1782 1783 ret = __folio_lock_killable(folio); 1784 if (ret) { 1785 release_fault_lock(vmf); 1786 return VM_FAULT_RETRY; 1787 } 1788 } else { 1789 __folio_lock(folio); 1790 } 1791 1792 return 0; 1793 } 1794 1795 /** 1796 * page_cache_next_miss() - Find the next gap in the page cache. 1797 * @mapping: Mapping. 1798 * @index: Index. 1799 * @max_scan: Maximum range to search. 1800 * 1801 * Search the range [index, min(index + max_scan - 1, ULONG_MAX)] for the 1802 * gap with the lowest index. 1803 * 1804 * This function may be called under the rcu_read_lock. However, this will 1805 * not atomically search a snapshot of the cache at a single point in time. 1806 * For example, if a gap is created at index 5, then subsequently a gap is 1807 * created at index 10, page_cache_next_miss covering both indices may 1808 * return 10 if called under the rcu_read_lock. 1809 * 1810 * Return: The index of the gap if found, otherwise an index outside the 1811 * range specified (in which case 'return - index >= max_scan' will be true). 1812 * In the rare case of index wrap-around, 0 will be returned. 1813 */ 1814 pgoff_t page_cache_next_miss(struct address_space *mapping, 1815 pgoff_t index, unsigned long max_scan) 1816 { 1817 XA_STATE(xas, &mapping->i_pages, index); 1818 unsigned long nr = max_scan; 1819 1820 while (nr--) { 1821 void *entry = xas_next(&xas); 1822 if (!entry || xa_is_value(entry)) 1823 return xas.xa_index; 1824 if (xas.xa_index == 0) 1825 return 0; 1826 } 1827 1828 return index + max_scan; 1829 } 1830 EXPORT_SYMBOL(page_cache_next_miss); 1831 1832 /** 1833 * page_cache_prev_miss() - Find the previous gap in the page cache. 1834 * @mapping: Mapping. 1835 * @index: Index. 1836 * @max_scan: Maximum range to search. 1837 * 1838 * Search the range [max(index - max_scan + 1, 0), index] for the 1839 * gap with the highest index. 1840 * 1841 * This function may be called under the rcu_read_lock. However, this will 1842 * not atomically search a snapshot of the cache at a single point in time. 1843 * For example, if a gap is created at index 10, then subsequently a gap is 1844 * created at index 5, page_cache_prev_miss() covering both indices may 1845 * return 5 if called under the rcu_read_lock. 1846 * 1847 * Return: The index of the gap if found, otherwise an index outside the 1848 * range specified (in which case 'index - return >= max_scan' will be true). 1849 * In the rare case of wrap-around, ULONG_MAX will be returned. 1850 */ 1851 pgoff_t page_cache_prev_miss(struct address_space *mapping, 1852 pgoff_t index, unsigned long max_scan) 1853 { 1854 XA_STATE(xas, &mapping->i_pages, index); 1855 1856 while (max_scan--) { 1857 void *entry = xas_prev(&xas); 1858 if (!entry || xa_is_value(entry)) 1859 break; 1860 if (xas.xa_index == ULONG_MAX) 1861 break; 1862 } 1863 1864 return xas.xa_index; 1865 } 1866 EXPORT_SYMBOL(page_cache_prev_miss); 1867 1868 /* 1869 * Lockless page cache protocol: 1870 * On the lookup side: 1871 * 1. Load the folio from i_pages 1872 * 2. Increment the refcount if it's not zero 1873 * 3. If the folio is not found by xas_reload(), put the refcount and retry 1874 * 1875 * On the removal side: 1876 * A. Freeze the page (by zeroing the refcount if nobody else has a reference) 1877 * B. Remove the page from i_pages 1878 * C. Return the page to the page allocator 1879 * 1880 * This means that any page may have its reference count temporarily 1881 * increased by a speculative page cache (or GUP-fast) lookup as it can 1882 * be allocated by another user before the RCU grace period expires. 1883 * Because the refcount temporarily acquired here may end up being the 1884 * last refcount on the page, any page allocation must be freeable by 1885 * folio_put(). 1886 */ 1887 1888 /* 1889 * filemap_get_entry - Get a page cache entry. 1890 * @mapping: the address_space to search 1891 * @index: The page cache index. 1892 * 1893 * Looks up the page cache entry at @mapping & @index. If it is a folio, 1894 * it is returned with an increased refcount. If it is a shadow entry 1895 * of a previously evicted folio, or a swap entry from shmem/tmpfs, 1896 * it is returned without further action. 1897 * 1898 * Return: The folio, swap or shadow entry, %NULL if nothing is found. 1899 */ 1900 void *filemap_get_entry(struct address_space *mapping, pgoff_t index) 1901 { 1902 XA_STATE(xas, &mapping->i_pages, index); 1903 struct folio *folio; 1904 1905 rcu_read_lock(); 1906 repeat: 1907 xas_reset(&xas); 1908 folio = xas_load(&xas); 1909 if (xas_retry(&xas, folio)) 1910 goto repeat; 1911 /* 1912 * A shadow entry of a recently evicted page, or a swap entry from 1913 * shmem/tmpfs. Return it without attempting to raise page count. 1914 */ 1915 if (!folio || xa_is_value(folio)) 1916 goto out; 1917 1918 if (!folio_try_get(folio)) 1919 goto repeat; 1920 1921 if (unlikely(folio != xas_reload(&xas))) { 1922 folio_put(folio); 1923 goto repeat; 1924 } 1925 out: 1926 rcu_read_unlock(); 1927 1928 return folio; 1929 } 1930 1931 /** 1932 * __filemap_get_folio - Find and get a reference to a folio. 1933 * @mapping: The address_space to search. 1934 * @index: The page index. 1935 * @fgp_flags: %FGP flags modify how the folio is returned. 1936 * @gfp: Memory allocation flags to use if %FGP_CREAT is specified. 1937 * 1938 * Looks up the page cache entry at @mapping & @index. 1939 * 1940 * If %FGP_LOCK or %FGP_CREAT are specified then the function may sleep even 1941 * if the %GFP flags specified for %FGP_CREAT are atomic. 1942 * 1943 * If this function returns a folio, it is returned with an increased refcount. 1944 * 1945 * Return: The found folio or an ERR_PTR() otherwise. 1946 */ 1947 struct folio *__filemap_get_folio(struct address_space *mapping, pgoff_t index, 1948 fgf_t fgp_flags, gfp_t gfp) 1949 { 1950 struct folio *folio; 1951 1952 repeat: 1953 folio = filemap_get_entry(mapping, index); 1954 if (xa_is_value(folio)) 1955 folio = NULL; 1956 if (!folio) 1957 goto no_page; 1958 1959 if (fgp_flags & FGP_LOCK) { 1960 if (fgp_flags & FGP_NOWAIT) { 1961 if (!folio_trylock(folio)) { 1962 folio_put(folio); 1963 return ERR_PTR(-EAGAIN); 1964 } 1965 } else { 1966 folio_lock(folio); 1967 } 1968 1969 /* Has the page been truncated? */ 1970 if (unlikely(folio->mapping != mapping)) { 1971 folio_unlock(folio); 1972 folio_put(folio); 1973 goto repeat; 1974 } 1975 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio); 1976 } 1977 1978 if (fgp_flags & FGP_ACCESSED) 1979 folio_mark_accessed(folio); 1980 else if (fgp_flags & FGP_WRITE) { 1981 /* Clear idle flag for buffer write */ 1982 if (folio_test_idle(folio)) 1983 folio_clear_idle(folio); 1984 } 1985 1986 if (fgp_flags & FGP_STABLE) 1987 folio_wait_stable(folio); 1988 no_page: 1989 if (!folio && (fgp_flags & FGP_CREAT)) { 1990 unsigned int min_order = mapping_min_folio_order(mapping); 1991 unsigned int order = max(min_order, FGF_GET_ORDER(fgp_flags)); 1992 int err; 1993 index = mapping_align_index(mapping, index); 1994 1995 if ((fgp_flags & FGP_WRITE) && mapping_can_writeback(mapping)) 1996 gfp |= __GFP_WRITE; 1997 if (fgp_flags & FGP_NOFS) 1998 gfp &= ~__GFP_FS; 1999 if (fgp_flags & FGP_NOWAIT) { 2000 gfp &= ~GFP_KERNEL; 2001 gfp |= GFP_NOWAIT; 2002 } 2003 if (WARN_ON_ONCE(!(fgp_flags & (FGP_LOCK | FGP_FOR_MMAP)))) 2004 fgp_flags |= FGP_LOCK; 2005 2006 if (order > mapping_max_folio_order(mapping)) 2007 order = mapping_max_folio_order(mapping); 2008 /* If we're not aligned, allocate a smaller folio */ 2009 if (index & ((1UL << order) - 1)) 2010 order = __ffs(index); 2011 2012 do { 2013 gfp_t alloc_gfp = gfp; 2014 2015 err = -ENOMEM; 2016 if (order > min_order) 2017 alloc_gfp |= __GFP_NORETRY | __GFP_NOWARN; 2018 folio = filemap_alloc_folio(alloc_gfp, order); 2019 if (!folio) 2020 continue; 2021 2022 /* Init accessed so avoid atomic mark_page_accessed later */ 2023 if (fgp_flags & FGP_ACCESSED) 2024 __folio_set_referenced(folio); 2025 if (fgp_flags & FGP_DONTCACHE) 2026 __folio_set_dropbehind(folio); 2027 2028 err = filemap_add_folio(mapping, folio, index, gfp); 2029 if (!err) 2030 break; 2031 folio_put(folio); 2032 folio = NULL; 2033 } while (order-- > min_order); 2034 2035 if (err == -EEXIST) 2036 goto repeat; 2037 if (err) { 2038 /* 2039 * When NOWAIT I/O fails to allocate folios this could 2040 * be due to a nonblocking memory allocation and not 2041 * because the system actually is out of memory. 2042 * Return -EAGAIN so that there caller retries in a 2043 * blocking fashion instead of propagating -ENOMEM 2044 * to the application. 2045 */ 2046 if ((fgp_flags & FGP_NOWAIT) && err == -ENOMEM) 2047 err = -EAGAIN; 2048 return ERR_PTR(err); 2049 } 2050 /* 2051 * filemap_add_folio locks the page, and for mmap 2052 * we expect an unlocked page. 2053 */ 2054 if (folio && (fgp_flags & FGP_FOR_MMAP)) 2055 folio_unlock(folio); 2056 } 2057 2058 if (!folio) 2059 return ERR_PTR(-ENOENT); 2060 /* not an uncached lookup, clear uncached if set */ 2061 if (folio_test_dropbehind(folio) && !(fgp_flags & FGP_DONTCACHE)) 2062 folio_clear_dropbehind(folio); 2063 return folio; 2064 } 2065 EXPORT_SYMBOL(__filemap_get_folio); 2066 2067 static inline struct folio *find_get_entry(struct xa_state *xas, pgoff_t max, 2068 xa_mark_t mark) 2069 { 2070 struct folio *folio; 2071 2072 retry: 2073 if (mark == XA_PRESENT) 2074 folio = xas_find(xas, max); 2075 else 2076 folio = xas_find_marked(xas, max, mark); 2077 2078 if (xas_retry(xas, folio)) 2079 goto retry; 2080 /* 2081 * A shadow entry of a recently evicted page, a swap 2082 * entry from shmem/tmpfs or a DAX entry. Return it 2083 * without attempting to raise page count. 2084 */ 2085 if (!folio || xa_is_value(folio)) 2086 return folio; 2087 2088 if (!folio_try_get(folio)) 2089 goto reset; 2090 2091 if (unlikely(folio != xas_reload(xas))) { 2092 folio_put(folio); 2093 goto reset; 2094 } 2095 2096 return folio; 2097 reset: 2098 xas_reset(xas); 2099 goto retry; 2100 } 2101 2102 /** 2103 * find_get_entries - gang pagecache lookup 2104 * @mapping: The address_space to search 2105 * @start: The starting page cache index 2106 * @end: The final page index (inclusive). 2107 * @fbatch: Where the resulting entries are placed. 2108 * @indices: The cache indices corresponding to the entries in @entries 2109 * 2110 * find_get_entries() will search for and return a batch of entries in 2111 * the mapping. The entries are placed in @fbatch. find_get_entries() 2112 * takes a reference on any actual folios it returns. 2113 * 2114 * The entries have ascending indexes. The indices may not be consecutive 2115 * due to not-present entries or large folios. 2116 * 2117 * Any shadow entries of evicted folios, or swap entries from 2118 * shmem/tmpfs, are included in the returned array. 2119 * 2120 * Return: The number of entries which were found. 2121 */ 2122 unsigned find_get_entries(struct address_space *mapping, pgoff_t *start, 2123 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices) 2124 { 2125 XA_STATE(xas, &mapping->i_pages, *start); 2126 struct folio *folio; 2127 2128 rcu_read_lock(); 2129 while ((folio = find_get_entry(&xas, end, XA_PRESENT)) != NULL) { 2130 indices[fbatch->nr] = xas.xa_index; 2131 if (!folio_batch_add(fbatch, folio)) 2132 break; 2133 } 2134 2135 if (folio_batch_count(fbatch)) { 2136 unsigned long nr; 2137 int idx = folio_batch_count(fbatch) - 1; 2138 2139 folio = fbatch->folios[idx]; 2140 if (!xa_is_value(folio)) 2141 nr = folio_nr_pages(folio); 2142 else 2143 nr = 1 << xa_get_order(&mapping->i_pages, indices[idx]); 2144 *start = round_down(indices[idx] + nr, nr); 2145 } 2146 rcu_read_unlock(); 2147 2148 return folio_batch_count(fbatch); 2149 } 2150 2151 /** 2152 * find_lock_entries - Find a batch of pagecache entries. 2153 * @mapping: The address_space to search. 2154 * @start: The starting page cache index. 2155 * @end: The final page index (inclusive). 2156 * @fbatch: Where the resulting entries are placed. 2157 * @indices: The cache indices of the entries in @fbatch. 2158 * 2159 * find_lock_entries() will return a batch of entries from @mapping. 2160 * Swap, shadow and DAX entries are included. Folios are returned 2161 * locked and with an incremented refcount. Folios which are locked 2162 * by somebody else or under writeback are skipped. Folios which are 2163 * partially outside the range are not returned. 2164 * 2165 * The entries have ascending indexes. The indices may not be consecutive 2166 * due to not-present entries, large folios, folios which could not be 2167 * locked or folios under writeback. 2168 * 2169 * Return: The number of entries which were found. 2170 */ 2171 unsigned find_lock_entries(struct address_space *mapping, pgoff_t *start, 2172 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices) 2173 { 2174 XA_STATE(xas, &mapping->i_pages, *start); 2175 struct folio *folio; 2176 2177 rcu_read_lock(); 2178 while ((folio = find_get_entry(&xas, end, XA_PRESENT))) { 2179 unsigned long base; 2180 unsigned long nr; 2181 2182 if (!xa_is_value(folio)) { 2183 nr = folio_nr_pages(folio); 2184 base = folio->index; 2185 /* Omit large folio which begins before the start */ 2186 if (base < *start) 2187 goto put; 2188 /* Omit large folio which extends beyond the end */ 2189 if (base + nr - 1 > end) 2190 goto put; 2191 if (!folio_trylock(folio)) 2192 goto put; 2193 if (folio->mapping != mapping || 2194 folio_test_writeback(folio)) 2195 goto unlock; 2196 VM_BUG_ON_FOLIO(!folio_contains(folio, xas.xa_index), 2197 folio); 2198 } else { 2199 nr = 1 << xas_get_order(&xas); 2200 base = xas.xa_index & ~(nr - 1); 2201 /* Omit order>0 value which begins before the start */ 2202 if (base < *start) 2203 continue; 2204 /* Omit order>0 value which extends beyond the end */ 2205 if (base + nr - 1 > end) 2206 break; 2207 } 2208 2209 /* Update start now so that last update is correct on return */ 2210 *start = base + nr; 2211 indices[fbatch->nr] = xas.xa_index; 2212 if (!folio_batch_add(fbatch, folio)) 2213 break; 2214 continue; 2215 unlock: 2216 folio_unlock(folio); 2217 put: 2218 folio_put(folio); 2219 } 2220 rcu_read_unlock(); 2221 2222 return folio_batch_count(fbatch); 2223 } 2224 2225 /** 2226 * filemap_get_folios - Get a batch of folios 2227 * @mapping: The address_space to search 2228 * @start: The starting page index 2229 * @end: The final page index (inclusive) 2230 * @fbatch: The batch to fill. 2231 * 2232 * Search for and return a batch of folios in the mapping starting at 2233 * index @start and up to index @end (inclusive). The folios are returned 2234 * in @fbatch with an elevated reference count. 2235 * 2236 * Return: The number of folios which were found. 2237 * We also update @start to index the next folio for the traversal. 2238 */ 2239 unsigned filemap_get_folios(struct address_space *mapping, pgoff_t *start, 2240 pgoff_t end, struct folio_batch *fbatch) 2241 { 2242 return filemap_get_folios_tag(mapping, start, end, XA_PRESENT, fbatch); 2243 } 2244 EXPORT_SYMBOL(filemap_get_folios); 2245 2246 /** 2247 * filemap_get_folios_contig - Get a batch of contiguous folios 2248 * @mapping: The address_space to search 2249 * @start: The starting page index 2250 * @end: The final page index (inclusive) 2251 * @fbatch: The batch to fill 2252 * 2253 * filemap_get_folios_contig() works exactly like filemap_get_folios(), 2254 * except the returned folios are guaranteed to be contiguous. This may 2255 * not return all contiguous folios if the batch gets filled up. 2256 * 2257 * Return: The number of folios found. 2258 * Also update @start to be positioned for traversal of the next folio. 2259 */ 2260 2261 unsigned filemap_get_folios_contig(struct address_space *mapping, 2262 pgoff_t *start, pgoff_t end, struct folio_batch *fbatch) 2263 { 2264 XA_STATE(xas, &mapping->i_pages, *start); 2265 unsigned long nr; 2266 struct folio *folio; 2267 2268 rcu_read_lock(); 2269 2270 for (folio = xas_load(&xas); folio && xas.xa_index <= end; 2271 folio = xas_next(&xas)) { 2272 if (xas_retry(&xas, folio)) 2273 continue; 2274 /* 2275 * If the entry has been swapped out, we can stop looking. 2276 * No current caller is looking for DAX entries. 2277 */ 2278 if (xa_is_value(folio)) 2279 goto update_start; 2280 2281 /* If we landed in the middle of a THP, continue at its end. */ 2282 if (xa_is_sibling(folio)) 2283 goto update_start; 2284 2285 if (!folio_try_get(folio)) 2286 goto retry; 2287 2288 if (unlikely(folio != xas_reload(&xas))) 2289 goto put_folio; 2290 2291 if (!folio_batch_add(fbatch, folio)) { 2292 nr = folio_nr_pages(folio); 2293 *start = folio->index + nr; 2294 goto out; 2295 } 2296 xas_advance(&xas, folio_next_index(folio) - 1); 2297 continue; 2298 put_folio: 2299 folio_put(folio); 2300 2301 retry: 2302 xas_reset(&xas); 2303 } 2304 2305 update_start: 2306 nr = folio_batch_count(fbatch); 2307 2308 if (nr) { 2309 folio = fbatch->folios[nr - 1]; 2310 *start = folio_next_index(folio); 2311 } 2312 out: 2313 rcu_read_unlock(); 2314 return folio_batch_count(fbatch); 2315 } 2316 EXPORT_SYMBOL(filemap_get_folios_contig); 2317 2318 /** 2319 * filemap_get_folios_tag - Get a batch of folios matching @tag 2320 * @mapping: The address_space to search 2321 * @start: The starting page index 2322 * @end: The final page index (inclusive) 2323 * @tag: The tag index 2324 * @fbatch: The batch to fill 2325 * 2326 * The first folio may start before @start; if it does, it will contain 2327 * @start. The final folio may extend beyond @end; if it does, it will 2328 * contain @end. The folios have ascending indices. There may be gaps 2329 * between the folios if there are indices which have no folio in the 2330 * page cache. If folios are added to or removed from the page cache 2331 * while this is running, they may or may not be found by this call. 2332 * Only returns folios that are tagged with @tag. 2333 * 2334 * Return: The number of folios found. 2335 * Also update @start to index the next folio for traversal. 2336 */ 2337 unsigned filemap_get_folios_tag(struct address_space *mapping, pgoff_t *start, 2338 pgoff_t end, xa_mark_t tag, struct folio_batch *fbatch) 2339 { 2340 XA_STATE(xas, &mapping->i_pages, *start); 2341 struct folio *folio; 2342 2343 rcu_read_lock(); 2344 while ((folio = find_get_entry(&xas, end, tag)) != NULL) { 2345 /* 2346 * Shadow entries should never be tagged, but this iteration 2347 * is lockless so there is a window for page reclaim to evict 2348 * a page we saw tagged. Skip over it. 2349 */ 2350 if (xa_is_value(folio)) 2351 continue; 2352 if (!folio_batch_add(fbatch, folio)) { 2353 unsigned long nr = folio_nr_pages(folio); 2354 *start = folio->index + nr; 2355 goto out; 2356 } 2357 } 2358 /* 2359 * We come here when there is no page beyond @end. We take care to not 2360 * overflow the index @start as it confuses some of the callers. This 2361 * breaks the iteration when there is a page at index -1 but that is 2362 * already broke anyway. 2363 */ 2364 if (end == (pgoff_t)-1) 2365 *start = (pgoff_t)-1; 2366 else 2367 *start = end + 1; 2368 out: 2369 rcu_read_unlock(); 2370 2371 return folio_batch_count(fbatch); 2372 } 2373 EXPORT_SYMBOL(filemap_get_folios_tag); 2374 2375 /* 2376 * CD/DVDs are error prone. When a medium error occurs, the driver may fail 2377 * a _large_ part of the i/o request. Imagine the worst scenario: 2378 * 2379 * ---R__________________________________________B__________ 2380 * ^ reading here ^ bad block(assume 4k) 2381 * 2382 * read(R) => miss => readahead(R...B) => media error => frustrating retries 2383 * => failing the whole request => read(R) => read(R+1) => 2384 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) => 2385 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) => 2386 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ...... 2387 * 2388 * It is going insane. Fix it by quickly scaling down the readahead size. 2389 */ 2390 static void shrink_readahead_size_eio(struct file_ra_state *ra) 2391 { 2392 ra->ra_pages /= 4; 2393 } 2394 2395 /* 2396 * filemap_get_read_batch - Get a batch of folios for read 2397 * 2398 * Get a batch of folios which represent a contiguous range of bytes in 2399 * the file. No exceptional entries will be returned. If @index is in 2400 * the middle of a folio, the entire folio will be returned. The last 2401 * folio in the batch may have the readahead flag set or the uptodate flag 2402 * clear so that the caller can take the appropriate action. 2403 */ 2404 static void filemap_get_read_batch(struct address_space *mapping, 2405 pgoff_t index, pgoff_t max, struct folio_batch *fbatch) 2406 { 2407 XA_STATE(xas, &mapping->i_pages, index); 2408 struct folio *folio; 2409 2410 rcu_read_lock(); 2411 for (folio = xas_load(&xas); folio; folio = xas_next(&xas)) { 2412 if (xas_retry(&xas, folio)) 2413 continue; 2414 if (xas.xa_index > max || xa_is_value(folio)) 2415 break; 2416 if (xa_is_sibling(folio)) 2417 break; 2418 if (!folio_try_get(folio)) 2419 goto retry; 2420 2421 if (unlikely(folio != xas_reload(&xas))) 2422 goto put_folio; 2423 2424 if (!folio_batch_add(fbatch, folio)) 2425 break; 2426 if (!folio_test_uptodate(folio)) 2427 break; 2428 if (folio_test_readahead(folio)) 2429 break; 2430 xas_advance(&xas, folio_next_index(folio) - 1); 2431 continue; 2432 put_folio: 2433 folio_put(folio); 2434 retry: 2435 xas_reset(&xas); 2436 } 2437 rcu_read_unlock(); 2438 } 2439 2440 static int filemap_read_folio(struct file *file, filler_t filler, 2441 struct folio *folio) 2442 { 2443 bool workingset = folio_test_workingset(folio); 2444 unsigned long pflags; 2445 int error; 2446 2447 /* Start the actual read. The read will unlock the page. */ 2448 if (unlikely(workingset)) 2449 psi_memstall_enter(&pflags); 2450 error = filler(file, folio); 2451 if (unlikely(workingset)) 2452 psi_memstall_leave(&pflags); 2453 if (error) 2454 return error; 2455 2456 error = folio_wait_locked_killable(folio); 2457 if (error) 2458 return error; 2459 if (folio_test_uptodate(folio)) 2460 return 0; 2461 if (file) 2462 shrink_readahead_size_eio(&file->f_ra); 2463 return -EIO; 2464 } 2465 2466 static bool filemap_range_uptodate(struct address_space *mapping, 2467 loff_t pos, size_t count, struct folio *folio, 2468 bool need_uptodate) 2469 { 2470 if (folio_test_uptodate(folio)) 2471 return true; 2472 /* pipes can't handle partially uptodate pages */ 2473 if (need_uptodate) 2474 return false; 2475 if (!mapping->a_ops->is_partially_uptodate) 2476 return false; 2477 if (mapping->host->i_blkbits >= folio_shift(folio)) 2478 return false; 2479 2480 if (folio_pos(folio) > pos) { 2481 count -= folio_pos(folio) - pos; 2482 pos = 0; 2483 } else { 2484 pos -= folio_pos(folio); 2485 } 2486 2487 if (pos == 0 && count >= folio_size(folio)) 2488 return false; 2489 2490 return mapping->a_ops->is_partially_uptodate(folio, pos, count); 2491 } 2492 2493 static int filemap_update_page(struct kiocb *iocb, 2494 struct address_space *mapping, size_t count, 2495 struct folio *folio, bool need_uptodate) 2496 { 2497 int error; 2498 2499 if (iocb->ki_flags & IOCB_NOWAIT) { 2500 if (!filemap_invalidate_trylock_shared(mapping)) 2501 return -EAGAIN; 2502 } else { 2503 filemap_invalidate_lock_shared(mapping); 2504 } 2505 2506 if (!folio_trylock(folio)) { 2507 error = -EAGAIN; 2508 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_NOIO)) 2509 goto unlock_mapping; 2510 if (!(iocb->ki_flags & IOCB_WAITQ)) { 2511 filemap_invalidate_unlock_shared(mapping); 2512 /* 2513 * This is where we usually end up waiting for a 2514 * previously submitted readahead to finish. 2515 */ 2516 folio_put_wait_locked(folio, TASK_KILLABLE); 2517 return AOP_TRUNCATED_PAGE; 2518 } 2519 error = __folio_lock_async(folio, iocb->ki_waitq); 2520 if (error) 2521 goto unlock_mapping; 2522 } 2523 2524 error = AOP_TRUNCATED_PAGE; 2525 if (!folio->mapping) 2526 goto unlock; 2527 2528 error = 0; 2529 if (filemap_range_uptodate(mapping, iocb->ki_pos, count, folio, 2530 need_uptodate)) 2531 goto unlock; 2532 2533 error = -EAGAIN; 2534 if (iocb->ki_flags & (IOCB_NOIO | IOCB_NOWAIT | IOCB_WAITQ)) 2535 goto unlock; 2536 2537 error = filemap_read_folio(iocb->ki_filp, mapping->a_ops->read_folio, 2538 folio); 2539 goto unlock_mapping; 2540 unlock: 2541 folio_unlock(folio); 2542 unlock_mapping: 2543 filemap_invalidate_unlock_shared(mapping); 2544 if (error == AOP_TRUNCATED_PAGE) 2545 folio_put(folio); 2546 return error; 2547 } 2548 2549 static int filemap_create_folio(struct kiocb *iocb, struct folio_batch *fbatch) 2550 { 2551 struct address_space *mapping = iocb->ki_filp->f_mapping; 2552 struct folio *folio; 2553 int error; 2554 unsigned int min_order = mapping_min_folio_order(mapping); 2555 pgoff_t index; 2556 2557 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_WAITQ)) 2558 return -EAGAIN; 2559 2560 folio = filemap_alloc_folio(mapping_gfp_mask(mapping), min_order); 2561 if (!folio) 2562 return -ENOMEM; 2563 if (iocb->ki_flags & IOCB_DONTCACHE) 2564 __folio_set_dropbehind(folio); 2565 2566 /* 2567 * Protect against truncate / hole punch. Grabbing invalidate_lock 2568 * here assures we cannot instantiate and bring uptodate new 2569 * pagecache folios after evicting page cache during truncate 2570 * and before actually freeing blocks. Note that we could 2571 * release invalidate_lock after inserting the folio into 2572 * the page cache as the locked folio would then be enough to 2573 * synchronize with hole punching. But there are code paths 2574 * such as filemap_update_page() filling in partially uptodate 2575 * pages or ->readahead() that need to hold invalidate_lock 2576 * while mapping blocks for IO so let's hold the lock here as 2577 * well to keep locking rules simple. 2578 */ 2579 filemap_invalidate_lock_shared(mapping); 2580 index = (iocb->ki_pos >> (PAGE_SHIFT + min_order)) << min_order; 2581 error = filemap_add_folio(mapping, folio, index, 2582 mapping_gfp_constraint(mapping, GFP_KERNEL)); 2583 if (error == -EEXIST) 2584 error = AOP_TRUNCATED_PAGE; 2585 if (error) 2586 goto error; 2587 2588 error = filemap_read_folio(iocb->ki_filp, mapping->a_ops->read_folio, 2589 folio); 2590 if (error) 2591 goto error; 2592 2593 filemap_invalidate_unlock_shared(mapping); 2594 folio_batch_add(fbatch, folio); 2595 return 0; 2596 error: 2597 filemap_invalidate_unlock_shared(mapping); 2598 folio_put(folio); 2599 return error; 2600 } 2601 2602 static int filemap_readahead(struct kiocb *iocb, struct file *file, 2603 struct address_space *mapping, struct folio *folio, 2604 pgoff_t last_index) 2605 { 2606 DEFINE_READAHEAD(ractl, file, &file->f_ra, mapping, folio->index); 2607 2608 if (iocb->ki_flags & IOCB_NOIO) 2609 return -EAGAIN; 2610 if (iocb->ki_flags & IOCB_DONTCACHE) 2611 ractl.dropbehind = 1; 2612 page_cache_async_ra(&ractl, folio, last_index - folio->index); 2613 return 0; 2614 } 2615 2616 static int filemap_get_pages(struct kiocb *iocb, size_t count, 2617 struct folio_batch *fbatch, bool need_uptodate) 2618 { 2619 struct file *filp = iocb->ki_filp; 2620 struct address_space *mapping = filp->f_mapping; 2621 pgoff_t index = iocb->ki_pos >> PAGE_SHIFT; 2622 pgoff_t last_index; 2623 struct folio *folio; 2624 unsigned int flags; 2625 int err = 0; 2626 2627 /* "last_index" is the index of the folio beyond the end of the read */ 2628 last_index = round_up(iocb->ki_pos + count, 2629 mapping_min_folio_nrbytes(mapping)) >> PAGE_SHIFT; 2630 retry: 2631 if (fatal_signal_pending(current)) 2632 return -EINTR; 2633 2634 filemap_get_read_batch(mapping, index, last_index - 1, fbatch); 2635 if (!folio_batch_count(fbatch)) { 2636 DEFINE_READAHEAD(ractl, filp, &filp->f_ra, mapping, index); 2637 2638 if (iocb->ki_flags & IOCB_NOIO) 2639 return -EAGAIN; 2640 if (iocb->ki_flags & IOCB_NOWAIT) 2641 flags = memalloc_noio_save(); 2642 if (iocb->ki_flags & IOCB_DONTCACHE) 2643 ractl.dropbehind = 1; 2644 page_cache_sync_ra(&ractl, last_index - index); 2645 if (iocb->ki_flags & IOCB_NOWAIT) 2646 memalloc_noio_restore(flags); 2647 filemap_get_read_batch(mapping, index, last_index - 1, fbatch); 2648 } 2649 if (!folio_batch_count(fbatch)) { 2650 err = filemap_create_folio(iocb, fbatch); 2651 if (err == AOP_TRUNCATED_PAGE) 2652 goto retry; 2653 return err; 2654 } 2655 2656 folio = fbatch->folios[folio_batch_count(fbatch) - 1]; 2657 if (folio_test_readahead(folio)) { 2658 err = filemap_readahead(iocb, filp, mapping, folio, last_index); 2659 if (err) 2660 goto err; 2661 } 2662 if (!folio_test_uptodate(folio)) { 2663 if (folio_batch_count(fbatch) > 1) { 2664 err = -EAGAIN; 2665 goto err; 2666 } 2667 err = filemap_update_page(iocb, mapping, count, folio, 2668 need_uptodate); 2669 if (err) 2670 goto err; 2671 } 2672 2673 trace_mm_filemap_get_pages(mapping, index, last_index - 1); 2674 return 0; 2675 err: 2676 if (err < 0) 2677 folio_put(folio); 2678 if (likely(--fbatch->nr)) 2679 return 0; 2680 if (err == AOP_TRUNCATED_PAGE) 2681 goto retry; 2682 return err; 2683 } 2684 2685 static inline bool pos_same_folio(loff_t pos1, loff_t pos2, struct folio *folio) 2686 { 2687 unsigned int shift = folio_shift(folio); 2688 2689 return (pos1 >> shift == pos2 >> shift); 2690 } 2691 2692 static void filemap_end_dropbehind_read(struct folio *folio) 2693 { 2694 if (!folio_test_dropbehind(folio)) 2695 return; 2696 if (folio_test_writeback(folio) || folio_test_dirty(folio)) 2697 return; 2698 if (folio_trylock(folio)) { 2699 filemap_end_dropbehind(folio); 2700 folio_unlock(folio); 2701 } 2702 } 2703 2704 /** 2705 * filemap_read - Read data from the page cache. 2706 * @iocb: The iocb to read. 2707 * @iter: Destination for the data. 2708 * @already_read: Number of bytes already read by the caller. 2709 * 2710 * Copies data from the page cache. If the data is not currently present, 2711 * uses the readahead and read_folio address_space operations to fetch it. 2712 * 2713 * Return: Total number of bytes copied, including those already read by 2714 * the caller. If an error happens before any bytes are copied, returns 2715 * a negative error number. 2716 */ 2717 ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter, 2718 ssize_t already_read) 2719 { 2720 struct file *filp = iocb->ki_filp; 2721 struct file_ra_state *ra = &filp->f_ra; 2722 struct address_space *mapping = filp->f_mapping; 2723 struct inode *inode = mapping->host; 2724 struct folio_batch fbatch; 2725 int i, error = 0; 2726 bool writably_mapped; 2727 loff_t isize, end_offset; 2728 loff_t last_pos = ra->prev_pos; 2729 2730 if (unlikely(iocb->ki_pos < 0)) 2731 return -EINVAL; 2732 if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes)) 2733 return 0; 2734 if (unlikely(!iov_iter_count(iter))) 2735 return 0; 2736 2737 iov_iter_truncate(iter, inode->i_sb->s_maxbytes - iocb->ki_pos); 2738 folio_batch_init(&fbatch); 2739 2740 do { 2741 cond_resched(); 2742 2743 /* 2744 * If we've already successfully copied some data, then we 2745 * can no longer safely return -EIOCBQUEUED. Hence mark 2746 * an async read NOWAIT at that point. 2747 */ 2748 if ((iocb->ki_flags & IOCB_WAITQ) && already_read) 2749 iocb->ki_flags |= IOCB_NOWAIT; 2750 2751 if (unlikely(iocb->ki_pos >= i_size_read(inode))) 2752 break; 2753 2754 error = filemap_get_pages(iocb, iter->count, &fbatch, false); 2755 if (error < 0) 2756 break; 2757 2758 /* 2759 * i_size must be checked after we know the pages are Uptodate. 2760 * 2761 * Checking i_size after the check allows us to calculate 2762 * the correct value for "nr", which means the zero-filled 2763 * part of the page is not copied back to userspace (unless 2764 * another truncate extends the file - this is desired though). 2765 */ 2766 isize = i_size_read(inode); 2767 if (unlikely(iocb->ki_pos >= isize)) 2768 goto put_folios; 2769 end_offset = min_t(loff_t, isize, iocb->ki_pos + iter->count); 2770 2771 /* 2772 * Once we start copying data, we don't want to be touching any 2773 * cachelines that might be contended: 2774 */ 2775 writably_mapped = mapping_writably_mapped(mapping); 2776 2777 /* 2778 * When a read accesses the same folio several times, only 2779 * mark it as accessed the first time. 2780 */ 2781 if (!pos_same_folio(iocb->ki_pos, last_pos - 1, 2782 fbatch.folios[0])) 2783 folio_mark_accessed(fbatch.folios[0]); 2784 2785 for (i = 0; i < folio_batch_count(&fbatch); i++) { 2786 struct folio *folio = fbatch.folios[i]; 2787 size_t fsize = folio_size(folio); 2788 size_t offset = iocb->ki_pos & (fsize - 1); 2789 size_t bytes = min_t(loff_t, end_offset - iocb->ki_pos, 2790 fsize - offset); 2791 size_t copied; 2792 2793 if (end_offset < folio_pos(folio)) 2794 break; 2795 if (i > 0) 2796 folio_mark_accessed(folio); 2797 /* 2798 * If users can be writing to this folio using arbitrary 2799 * virtual addresses, take care of potential aliasing 2800 * before reading the folio on the kernel side. 2801 */ 2802 if (writably_mapped) 2803 flush_dcache_folio(folio); 2804 2805 copied = copy_folio_to_iter(folio, offset, bytes, iter); 2806 2807 already_read += copied; 2808 iocb->ki_pos += copied; 2809 last_pos = iocb->ki_pos; 2810 2811 if (copied < bytes) { 2812 error = -EFAULT; 2813 break; 2814 } 2815 } 2816 put_folios: 2817 for (i = 0; i < folio_batch_count(&fbatch); i++) { 2818 struct folio *folio = fbatch.folios[i]; 2819 2820 filemap_end_dropbehind_read(folio); 2821 folio_put(folio); 2822 } 2823 folio_batch_init(&fbatch); 2824 } while (iov_iter_count(iter) && iocb->ki_pos < isize && !error); 2825 2826 file_accessed(filp); 2827 ra->prev_pos = last_pos; 2828 return already_read ? already_read : error; 2829 } 2830 EXPORT_SYMBOL_GPL(filemap_read); 2831 2832 int kiocb_write_and_wait(struct kiocb *iocb, size_t count) 2833 { 2834 struct address_space *mapping = iocb->ki_filp->f_mapping; 2835 loff_t pos = iocb->ki_pos; 2836 loff_t end = pos + count - 1; 2837 2838 if (iocb->ki_flags & IOCB_NOWAIT) { 2839 if (filemap_range_needs_writeback(mapping, pos, end)) 2840 return -EAGAIN; 2841 return 0; 2842 } 2843 2844 return filemap_write_and_wait_range(mapping, pos, end); 2845 } 2846 EXPORT_SYMBOL_GPL(kiocb_write_and_wait); 2847 2848 int filemap_invalidate_pages(struct address_space *mapping, 2849 loff_t pos, loff_t end, bool nowait) 2850 { 2851 int ret; 2852 2853 if (nowait) { 2854 /* we could block if there are any pages in the range */ 2855 if (filemap_range_has_page(mapping, pos, end)) 2856 return -EAGAIN; 2857 } else { 2858 ret = filemap_write_and_wait_range(mapping, pos, end); 2859 if (ret) 2860 return ret; 2861 } 2862 2863 /* 2864 * After a write we want buffered reads to be sure to go to disk to get 2865 * the new data. We invalidate clean cached page from the region we're 2866 * about to write. We do this *before* the write so that we can return 2867 * without clobbering -EIOCBQUEUED from ->direct_IO(). 2868 */ 2869 return invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT, 2870 end >> PAGE_SHIFT); 2871 } 2872 2873 int kiocb_invalidate_pages(struct kiocb *iocb, size_t count) 2874 { 2875 struct address_space *mapping = iocb->ki_filp->f_mapping; 2876 2877 return filemap_invalidate_pages(mapping, iocb->ki_pos, 2878 iocb->ki_pos + count - 1, 2879 iocb->ki_flags & IOCB_NOWAIT); 2880 } 2881 EXPORT_SYMBOL_GPL(kiocb_invalidate_pages); 2882 2883 /** 2884 * generic_file_read_iter - generic filesystem read routine 2885 * @iocb: kernel I/O control block 2886 * @iter: destination for the data read 2887 * 2888 * This is the "read_iter()" routine for all filesystems 2889 * that can use the page cache directly. 2890 * 2891 * The IOCB_NOWAIT flag in iocb->ki_flags indicates that -EAGAIN shall 2892 * be returned when no data can be read without waiting for I/O requests 2893 * to complete; it doesn't prevent readahead. 2894 * 2895 * The IOCB_NOIO flag in iocb->ki_flags indicates that no new I/O 2896 * requests shall be made for the read or for readahead. When no data 2897 * can be read, -EAGAIN shall be returned. When readahead would be 2898 * triggered, a partial, possibly empty read shall be returned. 2899 * 2900 * Return: 2901 * * number of bytes copied, even for partial reads 2902 * * negative error code (or 0 if IOCB_NOIO) if nothing was read 2903 */ 2904 ssize_t 2905 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter) 2906 { 2907 size_t count = iov_iter_count(iter); 2908 ssize_t retval = 0; 2909 2910 if (!count) 2911 return 0; /* skip atime */ 2912 2913 if (iocb->ki_flags & IOCB_DIRECT) { 2914 struct file *file = iocb->ki_filp; 2915 struct address_space *mapping = file->f_mapping; 2916 struct inode *inode = mapping->host; 2917 2918 retval = kiocb_write_and_wait(iocb, count); 2919 if (retval < 0) 2920 return retval; 2921 file_accessed(file); 2922 2923 retval = mapping->a_ops->direct_IO(iocb, iter); 2924 if (retval >= 0) { 2925 iocb->ki_pos += retval; 2926 count -= retval; 2927 } 2928 if (retval != -EIOCBQUEUED) 2929 iov_iter_revert(iter, count - iov_iter_count(iter)); 2930 2931 /* 2932 * Btrfs can have a short DIO read if we encounter 2933 * compressed extents, so if there was an error, or if 2934 * we've already read everything we wanted to, or if 2935 * there was a short read because we hit EOF, go ahead 2936 * and return. Otherwise fallthrough to buffered io for 2937 * the rest of the read. Buffered reads will not work for 2938 * DAX files, so don't bother trying. 2939 */ 2940 if (retval < 0 || !count || IS_DAX(inode)) 2941 return retval; 2942 if (iocb->ki_pos >= i_size_read(inode)) 2943 return retval; 2944 } 2945 2946 return filemap_read(iocb, iter, retval); 2947 } 2948 EXPORT_SYMBOL(generic_file_read_iter); 2949 2950 /* 2951 * Splice subpages from a folio into a pipe. 2952 */ 2953 size_t splice_folio_into_pipe(struct pipe_inode_info *pipe, 2954 struct folio *folio, loff_t fpos, size_t size) 2955 { 2956 struct page *page; 2957 size_t spliced = 0, offset = offset_in_folio(folio, fpos); 2958 2959 page = folio_page(folio, offset / PAGE_SIZE); 2960 size = min(size, folio_size(folio) - offset); 2961 offset %= PAGE_SIZE; 2962 2963 while (spliced < size && !pipe_is_full(pipe)) { 2964 struct pipe_buffer *buf = pipe_head_buf(pipe); 2965 size_t part = min_t(size_t, PAGE_SIZE - offset, size - spliced); 2966 2967 *buf = (struct pipe_buffer) { 2968 .ops = &page_cache_pipe_buf_ops, 2969 .page = page, 2970 .offset = offset, 2971 .len = part, 2972 }; 2973 folio_get(folio); 2974 pipe->head++; 2975 page++; 2976 spliced += part; 2977 offset = 0; 2978 } 2979 2980 return spliced; 2981 } 2982 2983 /** 2984 * filemap_splice_read - Splice data from a file's pagecache into a pipe 2985 * @in: The file to read from 2986 * @ppos: Pointer to the file position to read from 2987 * @pipe: The pipe to splice into 2988 * @len: The amount to splice 2989 * @flags: The SPLICE_F_* flags 2990 * 2991 * This function gets folios from a file's pagecache and splices them into the 2992 * pipe. Readahead will be called as necessary to fill more folios. This may 2993 * be used for blockdevs also. 2994 * 2995 * Return: On success, the number of bytes read will be returned and *@ppos 2996 * will be updated if appropriate; 0 will be returned if there is no more data 2997 * to be read; -EAGAIN will be returned if the pipe had no space, and some 2998 * other negative error code will be returned on error. A short read may occur 2999 * if the pipe has insufficient space, we reach the end of the data or we hit a 3000 * hole. 3001 */ 3002 ssize_t filemap_splice_read(struct file *in, loff_t *ppos, 3003 struct pipe_inode_info *pipe, 3004 size_t len, unsigned int flags) 3005 { 3006 struct folio_batch fbatch; 3007 struct kiocb iocb; 3008 size_t total_spliced = 0, used, npages; 3009 loff_t isize, end_offset; 3010 bool writably_mapped; 3011 int i, error = 0; 3012 3013 if (unlikely(*ppos >= in->f_mapping->host->i_sb->s_maxbytes)) 3014 return 0; 3015 3016 init_sync_kiocb(&iocb, in); 3017 iocb.ki_pos = *ppos; 3018 3019 /* Work out how much data we can actually add into the pipe */ 3020 used = pipe_buf_usage(pipe); 3021 npages = max_t(ssize_t, pipe->max_usage - used, 0); 3022 len = min_t(size_t, len, npages * PAGE_SIZE); 3023 3024 folio_batch_init(&fbatch); 3025 3026 do { 3027 cond_resched(); 3028 3029 if (*ppos >= i_size_read(in->f_mapping->host)) 3030 break; 3031 3032 iocb.ki_pos = *ppos; 3033 error = filemap_get_pages(&iocb, len, &fbatch, true); 3034 if (error < 0) 3035 break; 3036 3037 /* 3038 * i_size must be checked after we know the pages are Uptodate. 3039 * 3040 * Checking i_size after the check allows us to calculate 3041 * the correct value for "nr", which means the zero-filled 3042 * part of the page is not copied back to userspace (unless 3043 * another truncate extends the file - this is desired though). 3044 */ 3045 isize = i_size_read(in->f_mapping->host); 3046 if (unlikely(*ppos >= isize)) 3047 break; 3048 end_offset = min_t(loff_t, isize, *ppos + len); 3049 3050 /* 3051 * Once we start copying data, we don't want to be touching any 3052 * cachelines that might be contended: 3053 */ 3054 writably_mapped = mapping_writably_mapped(in->f_mapping); 3055 3056 for (i = 0; i < folio_batch_count(&fbatch); i++) { 3057 struct folio *folio = fbatch.folios[i]; 3058 size_t n; 3059 3060 if (folio_pos(folio) >= end_offset) 3061 goto out; 3062 folio_mark_accessed(folio); 3063 3064 /* 3065 * If users can be writing to this folio using arbitrary 3066 * virtual addresses, take care of potential aliasing 3067 * before reading the folio on the kernel side. 3068 */ 3069 if (writably_mapped) 3070 flush_dcache_folio(folio); 3071 3072 n = min_t(loff_t, len, isize - *ppos); 3073 n = splice_folio_into_pipe(pipe, folio, *ppos, n); 3074 if (!n) 3075 goto out; 3076 len -= n; 3077 total_spliced += n; 3078 *ppos += n; 3079 in->f_ra.prev_pos = *ppos; 3080 if (pipe_is_full(pipe)) 3081 goto out; 3082 } 3083 3084 folio_batch_release(&fbatch); 3085 } while (len); 3086 3087 out: 3088 folio_batch_release(&fbatch); 3089 file_accessed(in); 3090 3091 return total_spliced ? total_spliced : error; 3092 } 3093 EXPORT_SYMBOL(filemap_splice_read); 3094 3095 static inline loff_t folio_seek_hole_data(struct xa_state *xas, 3096 struct address_space *mapping, struct folio *folio, 3097 loff_t start, loff_t end, bool seek_data) 3098 { 3099 const struct address_space_operations *ops = mapping->a_ops; 3100 size_t offset, bsz = i_blocksize(mapping->host); 3101 3102 if (xa_is_value(folio) || folio_test_uptodate(folio)) 3103 return seek_data ? start : end; 3104 if (!ops->is_partially_uptodate) 3105 return seek_data ? end : start; 3106 3107 xas_pause(xas); 3108 rcu_read_unlock(); 3109 folio_lock(folio); 3110 if (unlikely(folio->mapping != mapping)) 3111 goto unlock; 3112 3113 offset = offset_in_folio(folio, start) & ~(bsz - 1); 3114 3115 do { 3116 if (ops->is_partially_uptodate(folio, offset, bsz) == 3117 seek_data) 3118 break; 3119 start = (start + bsz) & ~((u64)bsz - 1); 3120 offset += bsz; 3121 } while (offset < folio_size(folio)); 3122 unlock: 3123 folio_unlock(folio); 3124 rcu_read_lock(); 3125 return start; 3126 } 3127 3128 static inline size_t seek_folio_size(struct xa_state *xas, struct folio *folio) 3129 { 3130 if (xa_is_value(folio)) 3131 return PAGE_SIZE << xas_get_order(xas); 3132 return folio_size(folio); 3133 } 3134 3135 /** 3136 * mapping_seek_hole_data - Seek for SEEK_DATA / SEEK_HOLE in the page cache. 3137 * @mapping: Address space to search. 3138 * @start: First byte to consider. 3139 * @end: Limit of search (exclusive). 3140 * @whence: Either SEEK_HOLE or SEEK_DATA. 3141 * 3142 * If the page cache knows which blocks contain holes and which blocks 3143 * contain data, your filesystem can use this function to implement 3144 * SEEK_HOLE and SEEK_DATA. This is useful for filesystems which are 3145 * entirely memory-based such as tmpfs, and filesystems which support 3146 * unwritten extents. 3147 * 3148 * Return: The requested offset on success, or -ENXIO if @whence specifies 3149 * SEEK_DATA and there is no data after @start. There is an implicit hole 3150 * after @end - 1, so SEEK_HOLE returns @end if all the bytes between @start 3151 * and @end contain data. 3152 */ 3153 loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start, 3154 loff_t end, int whence) 3155 { 3156 XA_STATE(xas, &mapping->i_pages, start >> PAGE_SHIFT); 3157 pgoff_t max = (end - 1) >> PAGE_SHIFT; 3158 bool seek_data = (whence == SEEK_DATA); 3159 struct folio *folio; 3160 3161 if (end <= start) 3162 return -ENXIO; 3163 3164 rcu_read_lock(); 3165 while ((folio = find_get_entry(&xas, max, XA_PRESENT))) { 3166 loff_t pos = (u64)xas.xa_index << PAGE_SHIFT; 3167 size_t seek_size; 3168 3169 if (start < pos) { 3170 if (!seek_data) 3171 goto unlock; 3172 start = pos; 3173 } 3174 3175 seek_size = seek_folio_size(&xas, folio); 3176 pos = round_up((u64)pos + 1, seek_size); 3177 start = folio_seek_hole_data(&xas, mapping, folio, start, pos, 3178 seek_data); 3179 if (start < pos) 3180 goto unlock; 3181 if (start >= end) 3182 break; 3183 if (seek_size > PAGE_SIZE) 3184 xas_set(&xas, pos >> PAGE_SHIFT); 3185 if (!xa_is_value(folio)) 3186 folio_put(folio); 3187 } 3188 if (seek_data) 3189 start = -ENXIO; 3190 unlock: 3191 rcu_read_unlock(); 3192 if (folio && !xa_is_value(folio)) 3193 folio_put(folio); 3194 if (start > end) 3195 return end; 3196 return start; 3197 } 3198 3199 #ifdef CONFIG_MMU 3200 #define MMAP_LOTSAMISS (100) 3201 /* 3202 * lock_folio_maybe_drop_mmap - lock the page, possibly dropping the mmap_lock 3203 * @vmf - the vm_fault for this fault. 3204 * @folio - the folio to lock. 3205 * @fpin - the pointer to the file we may pin (or is already pinned). 3206 * 3207 * This works similar to lock_folio_or_retry in that it can drop the 3208 * mmap_lock. It differs in that it actually returns the folio locked 3209 * if it returns 1 and 0 if it couldn't lock the folio. If we did have 3210 * to drop the mmap_lock then fpin will point to the pinned file and 3211 * needs to be fput()'ed at a later point. 3212 */ 3213 static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio, 3214 struct file **fpin) 3215 { 3216 if (folio_trylock(folio)) 3217 return 1; 3218 3219 /* 3220 * NOTE! This will make us return with VM_FAULT_RETRY, but with 3221 * the fault lock still held. That's how FAULT_FLAG_RETRY_NOWAIT 3222 * is supposed to work. We have way too many special cases.. 3223 */ 3224 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) 3225 return 0; 3226 3227 *fpin = maybe_unlock_mmap_for_io(vmf, *fpin); 3228 if (vmf->flags & FAULT_FLAG_KILLABLE) { 3229 if (__folio_lock_killable(folio)) { 3230 /* 3231 * We didn't have the right flags to drop the 3232 * fault lock, but all fault_handlers only check 3233 * for fatal signals if we return VM_FAULT_RETRY, 3234 * so we need to drop the fault lock here and 3235 * return 0 if we don't have a fpin. 3236 */ 3237 if (*fpin == NULL) 3238 release_fault_lock(vmf); 3239 return 0; 3240 } 3241 } else 3242 __folio_lock(folio); 3243 3244 return 1; 3245 } 3246 3247 /* 3248 * Synchronous readahead happens when we don't even find a page in the page 3249 * cache at all. We don't want to perform IO under the mmap sem, so if we have 3250 * to drop the mmap sem we return the file that was pinned in order for us to do 3251 * that. If we didn't pin a file then we return NULL. The file that is 3252 * returned needs to be fput()'ed when we're done with it. 3253 */ 3254 static struct file *do_sync_mmap_readahead(struct vm_fault *vmf) 3255 { 3256 struct file *file = vmf->vma->vm_file; 3257 struct file_ra_state *ra = &file->f_ra; 3258 struct address_space *mapping = file->f_mapping; 3259 DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff); 3260 struct file *fpin = NULL; 3261 vm_flags_t vm_flags = vmf->vma->vm_flags; 3262 unsigned short mmap_miss; 3263 3264 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 3265 /* Use the readahead code, even if readahead is disabled */ 3266 if ((vm_flags & VM_HUGEPAGE) && HPAGE_PMD_ORDER <= MAX_PAGECACHE_ORDER) { 3267 fpin = maybe_unlock_mmap_for_io(vmf, fpin); 3268 ractl._index &= ~((unsigned long)HPAGE_PMD_NR - 1); 3269 ra->size = HPAGE_PMD_NR; 3270 /* 3271 * Fetch two PMD folios, so we get the chance to actually 3272 * readahead, unless we've been told not to. 3273 */ 3274 if (!(vm_flags & VM_RAND_READ)) 3275 ra->size *= 2; 3276 ra->async_size = HPAGE_PMD_NR; 3277 ra->order = HPAGE_PMD_ORDER; 3278 page_cache_ra_order(&ractl, ra); 3279 return fpin; 3280 } 3281 #endif 3282 3283 /* 3284 * If we don't want any read-ahead, don't bother. VM_EXEC case below is 3285 * already intended for random access. 3286 */ 3287 if ((vm_flags & (VM_RAND_READ | VM_EXEC)) == VM_RAND_READ) 3288 return fpin; 3289 if (!ra->ra_pages) 3290 return fpin; 3291 3292 if (vm_flags & VM_SEQ_READ) { 3293 fpin = maybe_unlock_mmap_for_io(vmf, fpin); 3294 page_cache_sync_ra(&ractl, ra->ra_pages); 3295 return fpin; 3296 } 3297 3298 /* Avoid banging the cache line if not needed */ 3299 mmap_miss = READ_ONCE(ra->mmap_miss); 3300 if (mmap_miss < MMAP_LOTSAMISS * 10) 3301 WRITE_ONCE(ra->mmap_miss, ++mmap_miss); 3302 3303 /* 3304 * Do we miss much more than hit in this file? If so, 3305 * stop bothering with read-ahead. It will only hurt. 3306 */ 3307 if (mmap_miss > MMAP_LOTSAMISS) 3308 return fpin; 3309 3310 if (vm_flags & VM_EXEC) { 3311 /* 3312 * Allow arch to request a preferred minimum folio order for 3313 * executable memory. This can often be beneficial to 3314 * performance if (e.g.) arm64 can contpte-map the folio. 3315 * Executable memory rarely benefits from readahead, due to its 3316 * random access nature, so set async_size to 0. 3317 * 3318 * Limit to the boundaries of the VMA to avoid reading in any 3319 * pad that might exist between sections, which would be a waste 3320 * of memory. 3321 */ 3322 struct vm_area_struct *vma = vmf->vma; 3323 unsigned long start = vma->vm_pgoff; 3324 unsigned long end = start + vma_pages(vma); 3325 unsigned long ra_end; 3326 3327 ra->order = exec_folio_order(); 3328 ra->start = round_down(vmf->pgoff, 1UL << ra->order); 3329 ra->start = max(ra->start, start); 3330 ra_end = round_up(ra->start + ra->ra_pages, 1UL << ra->order); 3331 ra_end = min(ra_end, end); 3332 ra->size = ra_end - ra->start; 3333 ra->async_size = 0; 3334 } else { 3335 /* 3336 * mmap read-around 3337 */ 3338 ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2); 3339 ra->size = ra->ra_pages; 3340 ra->async_size = ra->ra_pages / 4; 3341 ra->order = 0; 3342 } 3343 3344 fpin = maybe_unlock_mmap_for_io(vmf, fpin); 3345 ractl._index = ra->start; 3346 page_cache_ra_order(&ractl, ra); 3347 return fpin; 3348 } 3349 3350 /* 3351 * Asynchronous readahead happens when we find the page and PG_readahead, 3352 * so we want to possibly extend the readahead further. We return the file that 3353 * was pinned if we have to drop the mmap_lock in order to do IO. 3354 */ 3355 static struct file *do_async_mmap_readahead(struct vm_fault *vmf, 3356 struct folio *folio) 3357 { 3358 struct file *file = vmf->vma->vm_file; 3359 struct file_ra_state *ra = &file->f_ra; 3360 DEFINE_READAHEAD(ractl, file, ra, file->f_mapping, vmf->pgoff); 3361 struct file *fpin = NULL; 3362 unsigned short mmap_miss; 3363 3364 /* If we don't want any read-ahead, don't bother */ 3365 if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages) 3366 return fpin; 3367 3368 /* 3369 * If the folio is locked, we're likely racing against another fault. 3370 * Don't touch the mmap_miss counter to avoid decreasing it multiple 3371 * times for a single folio and break the balance with mmap_miss 3372 * increase in do_sync_mmap_readahead(). 3373 */ 3374 if (likely(!folio_test_locked(folio))) { 3375 mmap_miss = READ_ONCE(ra->mmap_miss); 3376 if (mmap_miss) 3377 WRITE_ONCE(ra->mmap_miss, --mmap_miss); 3378 } 3379 3380 if (folio_test_readahead(folio)) { 3381 fpin = maybe_unlock_mmap_for_io(vmf, fpin); 3382 page_cache_async_ra(&ractl, folio, ra->ra_pages); 3383 } 3384 return fpin; 3385 } 3386 3387 static vm_fault_t filemap_fault_recheck_pte_none(struct vm_fault *vmf) 3388 { 3389 struct vm_area_struct *vma = vmf->vma; 3390 vm_fault_t ret = 0; 3391 pte_t *ptep; 3392 3393 /* 3394 * We might have COW'ed a pagecache folio and might now have an mlocked 3395 * anon folio mapped. The original pagecache folio is not mlocked and 3396 * might have been evicted. During a read+clear/modify/write update of 3397 * the PTE, such as done in do_numa_page()/change_pte_range(), we 3398 * temporarily clear the PTE under PT lock and might detect it here as 3399 * "none" when not holding the PT lock. 3400 * 3401 * Not rechecking the PTE under PT lock could result in an unexpected 3402 * major fault in an mlock'ed region. Recheck only for this special 3403 * scenario while holding the PT lock, to not degrade non-mlocked 3404 * scenarios. Recheck the PTE without PT lock firstly, thereby reducing 3405 * the number of times we hold PT lock. 3406 */ 3407 if (!(vma->vm_flags & VM_LOCKED)) 3408 return 0; 3409 3410 if (!(vmf->flags & FAULT_FLAG_ORIG_PTE_VALID)) 3411 return 0; 3412 3413 ptep = pte_offset_map_ro_nolock(vma->vm_mm, vmf->pmd, vmf->address, 3414 &vmf->ptl); 3415 if (unlikely(!ptep)) 3416 return VM_FAULT_NOPAGE; 3417 3418 if (unlikely(!pte_none(ptep_get_lockless(ptep)))) { 3419 ret = VM_FAULT_NOPAGE; 3420 } else { 3421 spin_lock(vmf->ptl); 3422 if (unlikely(!pte_none(ptep_get(ptep)))) 3423 ret = VM_FAULT_NOPAGE; 3424 spin_unlock(vmf->ptl); 3425 } 3426 pte_unmap(ptep); 3427 return ret; 3428 } 3429 3430 /** 3431 * filemap_fault - read in file data for page fault handling 3432 * @vmf: struct vm_fault containing details of the fault 3433 * 3434 * filemap_fault() is invoked via the vma operations vector for a 3435 * mapped memory region to read in file data during a page fault. 3436 * 3437 * The goto's are kind of ugly, but this streamlines the normal case of having 3438 * it in the page cache, and handles the special cases reasonably without 3439 * having a lot of duplicated code. 3440 * 3441 * vma->vm_mm->mmap_lock must be held on entry. 3442 * 3443 * If our return value has VM_FAULT_RETRY set, it's because the mmap_lock 3444 * may be dropped before doing I/O or by lock_folio_maybe_drop_mmap(). 3445 * 3446 * If our return value does not have VM_FAULT_RETRY set, the mmap_lock 3447 * has not been released. 3448 * 3449 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set. 3450 * 3451 * Return: bitwise-OR of %VM_FAULT_ codes. 3452 */ 3453 vm_fault_t filemap_fault(struct vm_fault *vmf) 3454 { 3455 int error; 3456 struct file *file = vmf->vma->vm_file; 3457 struct file *fpin = NULL; 3458 struct address_space *mapping = file->f_mapping; 3459 struct inode *inode = mapping->host; 3460 pgoff_t max_idx, index = vmf->pgoff; 3461 struct folio *folio; 3462 vm_fault_t ret = 0; 3463 bool mapping_locked = false; 3464 3465 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE); 3466 if (unlikely(index >= max_idx)) 3467 return VM_FAULT_SIGBUS; 3468 3469 trace_mm_filemap_fault(mapping, index); 3470 3471 /* 3472 * Do we have something in the page cache already? 3473 */ 3474 folio = filemap_get_folio(mapping, index); 3475 if (likely(!IS_ERR(folio))) { 3476 /* 3477 * We found the page, so try async readahead before waiting for 3478 * the lock. 3479 */ 3480 if (!(vmf->flags & FAULT_FLAG_TRIED)) 3481 fpin = do_async_mmap_readahead(vmf, folio); 3482 if (unlikely(!folio_test_uptodate(folio))) { 3483 filemap_invalidate_lock_shared(mapping); 3484 mapping_locked = true; 3485 } 3486 } else { 3487 ret = filemap_fault_recheck_pte_none(vmf); 3488 if (unlikely(ret)) 3489 return ret; 3490 3491 /* No page in the page cache at all */ 3492 count_vm_event(PGMAJFAULT); 3493 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT); 3494 ret = VM_FAULT_MAJOR; 3495 fpin = do_sync_mmap_readahead(vmf); 3496 retry_find: 3497 /* 3498 * See comment in filemap_create_folio() why we need 3499 * invalidate_lock 3500 */ 3501 if (!mapping_locked) { 3502 filemap_invalidate_lock_shared(mapping); 3503 mapping_locked = true; 3504 } 3505 folio = __filemap_get_folio(mapping, index, 3506 FGP_CREAT|FGP_FOR_MMAP, 3507 vmf->gfp_mask); 3508 if (IS_ERR(folio)) { 3509 if (fpin) 3510 goto out_retry; 3511 filemap_invalidate_unlock_shared(mapping); 3512 return VM_FAULT_OOM; 3513 } 3514 } 3515 3516 if (!lock_folio_maybe_drop_mmap(vmf, folio, &fpin)) 3517 goto out_retry; 3518 3519 /* Did it get truncated? */ 3520 if (unlikely(folio->mapping != mapping)) { 3521 folio_unlock(folio); 3522 folio_put(folio); 3523 goto retry_find; 3524 } 3525 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio); 3526 3527 /* 3528 * We have a locked folio in the page cache, now we need to check 3529 * that it's up-to-date. If not, it is going to be due to an error, 3530 * or because readahead was otherwise unable to retrieve it. 3531 */ 3532 if (unlikely(!folio_test_uptodate(folio))) { 3533 /* 3534 * If the invalidate lock is not held, the folio was in cache 3535 * and uptodate and now it is not. Strange but possible since we 3536 * didn't hold the page lock all the time. Let's drop 3537 * everything, get the invalidate lock and try again. 3538 */ 3539 if (!mapping_locked) { 3540 folio_unlock(folio); 3541 folio_put(folio); 3542 goto retry_find; 3543 } 3544 3545 /* 3546 * OK, the folio is really not uptodate. This can be because the 3547 * VMA has the VM_RAND_READ flag set, or because an error 3548 * arose. Let's read it in directly. 3549 */ 3550 goto page_not_uptodate; 3551 } 3552 3553 /* 3554 * We've made it this far and we had to drop our mmap_lock, now is the 3555 * time to return to the upper layer and have it re-find the vma and 3556 * redo the fault. 3557 */ 3558 if (fpin) { 3559 folio_unlock(folio); 3560 goto out_retry; 3561 } 3562 if (mapping_locked) 3563 filemap_invalidate_unlock_shared(mapping); 3564 3565 /* 3566 * Found the page and have a reference on it. 3567 * We must recheck i_size under page lock. 3568 */ 3569 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE); 3570 if (unlikely(index >= max_idx)) { 3571 folio_unlock(folio); 3572 folio_put(folio); 3573 return VM_FAULT_SIGBUS; 3574 } 3575 3576 vmf->page = folio_file_page(folio, index); 3577 return ret | VM_FAULT_LOCKED; 3578 3579 page_not_uptodate: 3580 /* 3581 * Umm, take care of errors if the page isn't up-to-date. 3582 * Try to re-read it _once_. We do this synchronously, 3583 * because there really aren't any performance issues here 3584 * and we need to check for errors. 3585 */ 3586 fpin = maybe_unlock_mmap_for_io(vmf, fpin); 3587 error = filemap_read_folio(file, mapping->a_ops->read_folio, folio); 3588 if (fpin) 3589 goto out_retry; 3590 folio_put(folio); 3591 3592 if (!error || error == AOP_TRUNCATED_PAGE) 3593 goto retry_find; 3594 filemap_invalidate_unlock_shared(mapping); 3595 3596 return VM_FAULT_SIGBUS; 3597 3598 out_retry: 3599 /* 3600 * We dropped the mmap_lock, we need to return to the fault handler to 3601 * re-find the vma and come back and find our hopefully still populated 3602 * page. 3603 */ 3604 if (!IS_ERR(folio)) 3605 folio_put(folio); 3606 if (mapping_locked) 3607 filemap_invalidate_unlock_shared(mapping); 3608 if (fpin) 3609 fput(fpin); 3610 return ret | VM_FAULT_RETRY; 3611 } 3612 EXPORT_SYMBOL(filemap_fault); 3613 3614 static bool filemap_map_pmd(struct vm_fault *vmf, struct folio *folio, 3615 pgoff_t start) 3616 { 3617 struct mm_struct *mm = vmf->vma->vm_mm; 3618 3619 /* Huge page is mapped? No need to proceed. */ 3620 if (pmd_trans_huge(*vmf->pmd)) { 3621 folio_unlock(folio); 3622 folio_put(folio); 3623 return true; 3624 } 3625 3626 if (pmd_none(*vmf->pmd) && folio_test_pmd_mappable(folio)) { 3627 struct page *page = folio_file_page(folio, start); 3628 vm_fault_t ret = do_set_pmd(vmf, folio, page); 3629 if (!ret) { 3630 /* The page is mapped successfully, reference consumed. */ 3631 folio_unlock(folio); 3632 return true; 3633 } 3634 } 3635 3636 if (pmd_none(*vmf->pmd) && vmf->prealloc_pte) 3637 pmd_install(mm, vmf->pmd, &vmf->prealloc_pte); 3638 3639 return false; 3640 } 3641 3642 static struct folio *next_uptodate_folio(struct xa_state *xas, 3643 struct address_space *mapping, pgoff_t end_pgoff) 3644 { 3645 struct folio *folio = xas_next_entry(xas, end_pgoff); 3646 unsigned long max_idx; 3647 3648 do { 3649 if (!folio) 3650 return NULL; 3651 if (xas_retry(xas, folio)) 3652 continue; 3653 if (xa_is_value(folio)) 3654 continue; 3655 if (!folio_try_get(folio)) 3656 continue; 3657 if (folio_test_locked(folio)) 3658 goto skip; 3659 /* Has the page moved or been split? */ 3660 if (unlikely(folio != xas_reload(xas))) 3661 goto skip; 3662 if (!folio_test_uptodate(folio) || folio_test_readahead(folio)) 3663 goto skip; 3664 if (!folio_trylock(folio)) 3665 goto skip; 3666 if (folio->mapping != mapping) 3667 goto unlock; 3668 if (!folio_test_uptodate(folio)) 3669 goto unlock; 3670 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 3671 if (xas->xa_index >= max_idx) 3672 goto unlock; 3673 return folio; 3674 unlock: 3675 folio_unlock(folio); 3676 skip: 3677 folio_put(folio); 3678 } while ((folio = xas_next_entry(xas, end_pgoff)) != NULL); 3679 3680 return NULL; 3681 } 3682 3683 /* 3684 * Map page range [start_page, start_page + nr_pages) of folio. 3685 * start_page is gotten from start by folio_page(folio, start) 3686 */ 3687 static vm_fault_t filemap_map_folio_range(struct vm_fault *vmf, 3688 struct folio *folio, unsigned long start, 3689 unsigned long addr, unsigned int nr_pages, 3690 unsigned long *rss, unsigned short *mmap_miss, 3691 pgoff_t file_end) 3692 { 3693 struct address_space *mapping = folio->mapping; 3694 unsigned int ref_from_caller = 1; 3695 vm_fault_t ret = 0; 3696 struct page *page = folio_page(folio, start); 3697 unsigned int count = 0; 3698 pte_t *old_ptep = vmf->pte; 3699 unsigned long addr0; 3700 3701 /* 3702 * Map the large folio fully where possible: 3703 * 3704 * - The folio is fully within size of the file or belong 3705 * to shmem/tmpfs; 3706 * - The folio doesn't cross VMA boundary; 3707 * - The folio doesn't cross page table boundary; 3708 */ 3709 addr0 = addr - start * PAGE_SIZE; 3710 if ((file_end >= folio_next_index(folio) || shmem_mapping(mapping)) && 3711 folio_within_vma(folio, vmf->vma) && 3712 (addr0 & PMD_MASK) == ((addr0 + folio_size(folio) - 1) & PMD_MASK)) { 3713 vmf->pte -= start; 3714 page -= start; 3715 addr = addr0; 3716 nr_pages = folio_nr_pages(folio); 3717 } 3718 3719 do { 3720 if (PageHWPoison(page + count)) 3721 goto skip; 3722 3723 /* 3724 * If there are too many folios that are recently evicted 3725 * in a file, they will probably continue to be evicted. 3726 * In such situation, read-ahead is only a waste of IO. 3727 * Don't decrease mmap_miss in this scenario to make sure 3728 * we can stop read-ahead. 3729 */ 3730 if (!folio_test_workingset(folio)) 3731 (*mmap_miss)++; 3732 3733 /* 3734 * NOTE: If there're PTE markers, we'll leave them to be 3735 * handled in the specific fault path, and it'll prohibit the 3736 * fault-around logic. 3737 */ 3738 if (!pte_none(ptep_get(&vmf->pte[count]))) 3739 goto skip; 3740 3741 count++; 3742 continue; 3743 skip: 3744 if (count) { 3745 set_pte_range(vmf, folio, page, count, addr); 3746 *rss += count; 3747 folio_ref_add(folio, count - ref_from_caller); 3748 ref_from_caller = 0; 3749 if (in_range(vmf->address, addr, count * PAGE_SIZE)) 3750 ret = VM_FAULT_NOPAGE; 3751 } 3752 3753 count++; 3754 page += count; 3755 vmf->pte += count; 3756 addr += count * PAGE_SIZE; 3757 count = 0; 3758 } while (--nr_pages > 0); 3759 3760 if (count) { 3761 set_pte_range(vmf, folio, page, count, addr); 3762 *rss += count; 3763 folio_ref_add(folio, count - ref_from_caller); 3764 ref_from_caller = 0; 3765 if (in_range(vmf->address, addr, count * PAGE_SIZE)) 3766 ret = VM_FAULT_NOPAGE; 3767 } 3768 3769 vmf->pte = old_ptep; 3770 if (ref_from_caller) 3771 /* Locked folios cannot get truncated. */ 3772 folio_ref_dec(folio); 3773 3774 return ret; 3775 } 3776 3777 static vm_fault_t filemap_map_order0_folio(struct vm_fault *vmf, 3778 struct folio *folio, unsigned long addr, 3779 unsigned long *rss, unsigned short *mmap_miss) 3780 { 3781 vm_fault_t ret = 0; 3782 struct page *page = &folio->page; 3783 3784 if (PageHWPoison(page)) 3785 goto out; 3786 3787 /* See comment of filemap_map_folio_range() */ 3788 if (!folio_test_workingset(folio)) 3789 (*mmap_miss)++; 3790 3791 /* 3792 * NOTE: If there're PTE markers, we'll leave them to be 3793 * handled in the specific fault path, and it'll prohibit 3794 * the fault-around logic. 3795 */ 3796 if (!pte_none(ptep_get(vmf->pte))) 3797 goto out; 3798 3799 if (vmf->address == addr) 3800 ret = VM_FAULT_NOPAGE; 3801 3802 set_pte_range(vmf, folio, page, 1, addr); 3803 (*rss)++; 3804 return ret; 3805 3806 out: 3807 /* Locked folios cannot get truncated. */ 3808 folio_ref_dec(folio); 3809 return ret; 3810 } 3811 3812 vm_fault_t filemap_map_pages(struct vm_fault *vmf, 3813 pgoff_t start_pgoff, pgoff_t end_pgoff) 3814 { 3815 struct vm_area_struct *vma = vmf->vma; 3816 struct file *file = vma->vm_file; 3817 struct address_space *mapping = file->f_mapping; 3818 pgoff_t file_end, last_pgoff = start_pgoff; 3819 unsigned long addr; 3820 XA_STATE(xas, &mapping->i_pages, start_pgoff); 3821 struct folio *folio; 3822 vm_fault_t ret = 0; 3823 unsigned long rss = 0; 3824 unsigned int nr_pages = 0, folio_type; 3825 unsigned short mmap_miss = 0, mmap_miss_saved; 3826 3827 /* 3828 * Recalculate end_pgoff based on file_end before calling 3829 * next_uptodate_folio() to avoid races with concurrent 3830 * truncation. 3831 */ 3832 file_end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE) - 1; 3833 end_pgoff = min(end_pgoff, file_end); 3834 3835 rcu_read_lock(); 3836 folio = next_uptodate_folio(&xas, mapping, end_pgoff); 3837 if (!folio) 3838 goto out; 3839 3840 /* 3841 * Do not allow to map with PMD across i_size to preserve 3842 * SIGBUS semantics. 3843 * 3844 * Make an exception for shmem/tmpfs that for long time 3845 * intentionally mapped with PMDs across i_size. 3846 */ 3847 if ((file_end >= folio_next_index(folio) || shmem_mapping(mapping)) && 3848 filemap_map_pmd(vmf, folio, start_pgoff)) { 3849 ret = VM_FAULT_NOPAGE; 3850 goto out; 3851 } 3852 3853 addr = vma->vm_start + ((start_pgoff - vma->vm_pgoff) << PAGE_SHIFT); 3854 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl); 3855 if (!vmf->pte) { 3856 folio_unlock(folio); 3857 folio_put(folio); 3858 goto out; 3859 } 3860 3861 folio_type = mm_counter_file(folio); 3862 do { 3863 unsigned long end; 3864 3865 addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT; 3866 vmf->pte += xas.xa_index - last_pgoff; 3867 last_pgoff = xas.xa_index; 3868 end = folio_next_index(folio) - 1; 3869 nr_pages = min(end, end_pgoff) - xas.xa_index + 1; 3870 3871 if (!folio_test_large(folio)) 3872 ret |= filemap_map_order0_folio(vmf, 3873 folio, addr, &rss, &mmap_miss); 3874 else 3875 ret |= filemap_map_folio_range(vmf, folio, 3876 xas.xa_index - folio->index, addr, 3877 nr_pages, &rss, &mmap_miss, file_end); 3878 3879 folio_unlock(folio); 3880 } while ((folio = next_uptodate_folio(&xas, mapping, end_pgoff)) != NULL); 3881 add_mm_counter(vma->vm_mm, folio_type, rss); 3882 pte_unmap_unlock(vmf->pte, vmf->ptl); 3883 trace_mm_filemap_map_pages(mapping, start_pgoff, end_pgoff); 3884 out: 3885 rcu_read_unlock(); 3886 3887 mmap_miss_saved = READ_ONCE(file->f_ra.mmap_miss); 3888 if (mmap_miss >= mmap_miss_saved) 3889 WRITE_ONCE(file->f_ra.mmap_miss, 0); 3890 else 3891 WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss_saved - mmap_miss); 3892 3893 return ret; 3894 } 3895 EXPORT_SYMBOL(filemap_map_pages); 3896 3897 vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf) 3898 { 3899 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 3900 struct folio *folio = page_folio(vmf->page); 3901 vm_fault_t ret = VM_FAULT_LOCKED; 3902 3903 sb_start_pagefault(mapping->host->i_sb); 3904 file_update_time(vmf->vma->vm_file); 3905 folio_lock(folio); 3906 if (folio->mapping != mapping) { 3907 folio_unlock(folio); 3908 ret = VM_FAULT_NOPAGE; 3909 goto out; 3910 } 3911 /* 3912 * We mark the folio dirty already here so that when freeze is in 3913 * progress, we are guaranteed that writeback during freezing will 3914 * see the dirty folio and writeprotect it again. 3915 */ 3916 folio_mark_dirty(folio); 3917 folio_wait_stable(folio); 3918 out: 3919 sb_end_pagefault(mapping->host->i_sb); 3920 return ret; 3921 } 3922 3923 const struct vm_operations_struct generic_file_vm_ops = { 3924 .fault = filemap_fault, 3925 .map_pages = filemap_map_pages, 3926 .page_mkwrite = filemap_page_mkwrite, 3927 }; 3928 3929 /* This is used for a general mmap of a disk file */ 3930 3931 int generic_file_mmap(struct file *file, struct vm_area_struct *vma) 3932 { 3933 struct address_space *mapping = file->f_mapping; 3934 3935 if (!mapping->a_ops->read_folio) 3936 return -ENOEXEC; 3937 file_accessed(file); 3938 vma->vm_ops = &generic_file_vm_ops; 3939 return 0; 3940 } 3941 3942 int generic_file_mmap_prepare(struct vm_area_desc *desc) 3943 { 3944 struct file *file = desc->file; 3945 struct address_space *mapping = file->f_mapping; 3946 3947 if (!mapping->a_ops->read_folio) 3948 return -ENOEXEC; 3949 file_accessed(file); 3950 desc->vm_ops = &generic_file_vm_ops; 3951 return 0; 3952 } 3953 3954 /* 3955 * This is for filesystems which do not implement ->writepage. 3956 */ 3957 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma) 3958 { 3959 if (vma_is_shared_maywrite(vma)) 3960 return -EINVAL; 3961 return generic_file_mmap(file, vma); 3962 } 3963 3964 int generic_file_readonly_mmap_prepare(struct vm_area_desc *desc) 3965 { 3966 if (is_shared_maywrite(desc->vm_flags)) 3967 return -EINVAL; 3968 return generic_file_mmap_prepare(desc); 3969 } 3970 #else 3971 vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf) 3972 { 3973 return VM_FAULT_SIGBUS; 3974 } 3975 int generic_file_mmap(struct file *file, struct vm_area_struct *vma) 3976 { 3977 return -ENOSYS; 3978 } 3979 int generic_file_mmap_prepare(struct vm_area_desc *desc) 3980 { 3981 return -ENOSYS; 3982 } 3983 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma) 3984 { 3985 return -ENOSYS; 3986 } 3987 int generic_file_readonly_mmap_prepare(struct vm_area_desc *desc) 3988 { 3989 return -ENOSYS; 3990 } 3991 #endif /* CONFIG_MMU */ 3992 3993 EXPORT_SYMBOL(filemap_page_mkwrite); 3994 EXPORT_SYMBOL(generic_file_mmap); 3995 EXPORT_SYMBOL(generic_file_mmap_prepare); 3996 EXPORT_SYMBOL(generic_file_readonly_mmap); 3997 EXPORT_SYMBOL(generic_file_readonly_mmap_prepare); 3998 3999 static struct folio *do_read_cache_folio(struct address_space *mapping, 4000 pgoff_t index, filler_t filler, struct file *file, gfp_t gfp) 4001 { 4002 struct folio *folio; 4003 int err; 4004 4005 if (!filler) 4006 filler = mapping->a_ops->read_folio; 4007 repeat: 4008 folio = filemap_get_folio(mapping, index); 4009 if (IS_ERR(folio)) { 4010 folio = filemap_alloc_folio(gfp, 4011 mapping_min_folio_order(mapping)); 4012 if (!folio) 4013 return ERR_PTR(-ENOMEM); 4014 index = mapping_align_index(mapping, index); 4015 err = filemap_add_folio(mapping, folio, index, gfp); 4016 if (unlikely(err)) { 4017 folio_put(folio); 4018 if (err == -EEXIST) 4019 goto repeat; 4020 /* Presumably ENOMEM for xarray node */ 4021 return ERR_PTR(err); 4022 } 4023 4024 goto filler; 4025 } 4026 if (folio_test_uptodate(folio)) 4027 goto out; 4028 4029 if (!folio_trylock(folio)) { 4030 folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE); 4031 goto repeat; 4032 } 4033 4034 /* Folio was truncated from mapping */ 4035 if (!folio->mapping) { 4036 folio_unlock(folio); 4037 folio_put(folio); 4038 goto repeat; 4039 } 4040 4041 /* Someone else locked and filled the page in a very small window */ 4042 if (folio_test_uptodate(folio)) { 4043 folio_unlock(folio); 4044 goto out; 4045 } 4046 4047 filler: 4048 err = filemap_read_folio(file, filler, folio); 4049 if (err) { 4050 folio_put(folio); 4051 if (err == AOP_TRUNCATED_PAGE) 4052 goto repeat; 4053 return ERR_PTR(err); 4054 } 4055 4056 out: 4057 folio_mark_accessed(folio); 4058 return folio; 4059 } 4060 4061 /** 4062 * read_cache_folio - Read into page cache, fill it if needed. 4063 * @mapping: The address_space to read from. 4064 * @index: The index to read. 4065 * @filler: Function to perform the read, or NULL to use aops->read_folio(). 4066 * @file: Passed to filler function, may be NULL if not required. 4067 * 4068 * Read one page into the page cache. If it succeeds, the folio returned 4069 * will contain @index, but it may not be the first page of the folio. 4070 * 4071 * If the filler function returns an error, it will be returned to the 4072 * caller. 4073 * 4074 * Context: May sleep. Expects mapping->invalidate_lock to be held. 4075 * Return: An uptodate folio on success, ERR_PTR() on failure. 4076 */ 4077 struct folio *read_cache_folio(struct address_space *mapping, pgoff_t index, 4078 filler_t filler, struct file *file) 4079 { 4080 return do_read_cache_folio(mapping, index, filler, file, 4081 mapping_gfp_mask(mapping)); 4082 } 4083 EXPORT_SYMBOL(read_cache_folio); 4084 4085 /** 4086 * mapping_read_folio_gfp - Read into page cache, using specified allocation flags. 4087 * @mapping: The address_space for the folio. 4088 * @index: The index that the allocated folio will contain. 4089 * @gfp: The page allocator flags to use if allocating. 4090 * 4091 * This is the same as "read_cache_folio(mapping, index, NULL, NULL)", but with 4092 * any new memory allocations done using the specified allocation flags. 4093 * 4094 * The most likely error from this function is EIO, but ENOMEM is 4095 * possible and so is EINTR. If ->read_folio returns another error, 4096 * that will be returned to the caller. 4097 * 4098 * The function expects mapping->invalidate_lock to be already held. 4099 * 4100 * Return: Uptodate folio on success, ERR_PTR() on failure. 4101 */ 4102 struct folio *mapping_read_folio_gfp(struct address_space *mapping, 4103 pgoff_t index, gfp_t gfp) 4104 { 4105 return do_read_cache_folio(mapping, index, NULL, NULL, gfp); 4106 } 4107 EXPORT_SYMBOL(mapping_read_folio_gfp); 4108 4109 static struct page *do_read_cache_page(struct address_space *mapping, 4110 pgoff_t index, filler_t *filler, struct file *file, gfp_t gfp) 4111 { 4112 struct folio *folio; 4113 4114 folio = do_read_cache_folio(mapping, index, filler, file, gfp); 4115 if (IS_ERR(folio)) 4116 return &folio->page; 4117 return folio_file_page(folio, index); 4118 } 4119 4120 struct page *read_cache_page(struct address_space *mapping, 4121 pgoff_t index, filler_t *filler, struct file *file) 4122 { 4123 return do_read_cache_page(mapping, index, filler, file, 4124 mapping_gfp_mask(mapping)); 4125 } 4126 EXPORT_SYMBOL(read_cache_page); 4127 4128 /** 4129 * read_cache_page_gfp - read into page cache, using specified page allocation flags. 4130 * @mapping: the page's address_space 4131 * @index: the page index 4132 * @gfp: the page allocator flags to use if allocating 4133 * 4134 * This is the same as "read_mapping_page(mapping, index, NULL)", but with 4135 * any new page allocations done using the specified allocation flags. 4136 * 4137 * If the page does not get brought uptodate, return -EIO. 4138 * 4139 * The function expects mapping->invalidate_lock to be already held. 4140 * 4141 * Return: up to date page on success, ERR_PTR() on failure. 4142 */ 4143 struct page *read_cache_page_gfp(struct address_space *mapping, 4144 pgoff_t index, 4145 gfp_t gfp) 4146 { 4147 return do_read_cache_page(mapping, index, NULL, NULL, gfp); 4148 } 4149 EXPORT_SYMBOL(read_cache_page_gfp); 4150 4151 /* 4152 * Warn about a page cache invalidation failure during a direct I/O write. 4153 */ 4154 static void dio_warn_stale_pagecache(struct file *filp) 4155 { 4156 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST); 4157 char pathname[128]; 4158 char *path; 4159 4160 errseq_set(&filp->f_mapping->wb_err, -EIO); 4161 if (__ratelimit(&_rs)) { 4162 path = file_path(filp, pathname, sizeof(pathname)); 4163 if (IS_ERR(path)) 4164 path = "(unknown)"; 4165 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n"); 4166 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid, 4167 current->comm); 4168 } 4169 } 4170 4171 void kiocb_invalidate_post_direct_write(struct kiocb *iocb, size_t count) 4172 { 4173 struct address_space *mapping = iocb->ki_filp->f_mapping; 4174 4175 if (mapping->nrpages && 4176 invalidate_inode_pages2_range(mapping, 4177 iocb->ki_pos >> PAGE_SHIFT, 4178 (iocb->ki_pos + count - 1) >> PAGE_SHIFT)) 4179 dio_warn_stale_pagecache(iocb->ki_filp); 4180 } 4181 4182 ssize_t 4183 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from) 4184 { 4185 struct address_space *mapping = iocb->ki_filp->f_mapping; 4186 size_t write_len = iov_iter_count(from); 4187 ssize_t written; 4188 4189 /* 4190 * If a page can not be invalidated, return 0 to fall back 4191 * to buffered write. 4192 */ 4193 written = kiocb_invalidate_pages(iocb, write_len); 4194 if (written) { 4195 if (written == -EBUSY) 4196 return 0; 4197 return written; 4198 } 4199 4200 written = mapping->a_ops->direct_IO(iocb, from); 4201 4202 /* 4203 * Finally, try again to invalidate clean pages which might have been 4204 * cached by non-direct readahead, or faulted in by get_user_pages() 4205 * if the source of the write was an mmap'ed region of the file 4206 * we're writing. Either one is a pretty crazy thing to do, 4207 * so we don't support it 100%. If this invalidation 4208 * fails, tough, the write still worked... 4209 * 4210 * Most of the time we do not need this since dio_complete() will do 4211 * the invalidation for us. However there are some file systems that 4212 * do not end up with dio_complete() being called, so let's not break 4213 * them by removing it completely. 4214 * 4215 * Noticeable example is a blkdev_direct_IO(). 4216 * 4217 * Skip invalidation for async writes or if mapping has no pages. 4218 */ 4219 if (written > 0) { 4220 struct inode *inode = mapping->host; 4221 loff_t pos = iocb->ki_pos; 4222 4223 kiocb_invalidate_post_direct_write(iocb, written); 4224 pos += written; 4225 write_len -= written; 4226 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) { 4227 i_size_write(inode, pos); 4228 mark_inode_dirty(inode); 4229 } 4230 iocb->ki_pos = pos; 4231 } 4232 if (written != -EIOCBQUEUED) 4233 iov_iter_revert(from, write_len - iov_iter_count(from)); 4234 return written; 4235 } 4236 EXPORT_SYMBOL(generic_file_direct_write); 4237 4238 ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i) 4239 { 4240 struct file *file = iocb->ki_filp; 4241 loff_t pos = iocb->ki_pos; 4242 struct address_space *mapping = file->f_mapping; 4243 const struct address_space_operations *a_ops = mapping->a_ops; 4244 size_t chunk = mapping_max_folio_size(mapping); 4245 long status = 0; 4246 ssize_t written = 0; 4247 4248 do { 4249 struct folio *folio; 4250 size_t offset; /* Offset into folio */ 4251 size_t bytes; /* Bytes to write to folio */ 4252 size_t copied; /* Bytes copied from user */ 4253 void *fsdata = NULL; 4254 4255 bytes = iov_iter_count(i); 4256 retry: 4257 offset = pos & (chunk - 1); 4258 bytes = min(chunk - offset, bytes); 4259 balance_dirty_pages_ratelimited(mapping); 4260 4261 if (fatal_signal_pending(current)) { 4262 status = -EINTR; 4263 break; 4264 } 4265 4266 status = a_ops->write_begin(iocb, mapping, pos, bytes, 4267 &folio, &fsdata); 4268 if (unlikely(status < 0)) 4269 break; 4270 4271 offset = offset_in_folio(folio, pos); 4272 if (bytes > folio_size(folio) - offset) 4273 bytes = folio_size(folio) - offset; 4274 4275 if (mapping_writably_mapped(mapping)) 4276 flush_dcache_folio(folio); 4277 4278 /* 4279 * Faults here on mmap()s can recurse into arbitrary 4280 * filesystem code. Lots of locks are held that can 4281 * deadlock. Use an atomic copy to avoid deadlocking 4282 * in page fault handling. 4283 */ 4284 copied = copy_folio_from_iter_atomic(folio, offset, bytes, i); 4285 flush_dcache_folio(folio); 4286 4287 status = a_ops->write_end(iocb, mapping, pos, bytes, copied, 4288 folio, fsdata); 4289 if (unlikely(status != copied)) { 4290 iov_iter_revert(i, copied - max(status, 0L)); 4291 if (unlikely(status < 0)) 4292 break; 4293 } 4294 cond_resched(); 4295 4296 if (unlikely(status == 0)) { 4297 /* 4298 * A short copy made ->write_end() reject the 4299 * thing entirely. Might be memory poisoning 4300 * halfway through, might be a race with munmap, 4301 * might be severe memory pressure. 4302 */ 4303 if (chunk > PAGE_SIZE) 4304 chunk /= 2; 4305 if (copied) { 4306 bytes = copied; 4307 goto retry; 4308 } 4309 4310 /* 4311 * 'folio' is now unlocked and faults on it can be 4312 * handled. Ensure forward progress by trying to 4313 * fault it in now. 4314 */ 4315 if (fault_in_iov_iter_readable(i, bytes) == bytes) { 4316 status = -EFAULT; 4317 break; 4318 } 4319 } else { 4320 pos += status; 4321 written += status; 4322 } 4323 } while (iov_iter_count(i)); 4324 4325 if (!written) 4326 return status; 4327 iocb->ki_pos += written; 4328 return written; 4329 } 4330 EXPORT_SYMBOL(generic_perform_write); 4331 4332 /** 4333 * __generic_file_write_iter - write data to a file 4334 * @iocb: IO state structure (file, offset, etc.) 4335 * @from: iov_iter with data to write 4336 * 4337 * This function does all the work needed for actually writing data to a 4338 * file. It does all basic checks, removes SUID from the file, updates 4339 * modification times and calls proper subroutines depending on whether we 4340 * do direct IO or a standard buffered write. 4341 * 4342 * It expects i_rwsem to be grabbed unless we work on a block device or similar 4343 * object which does not need locking at all. 4344 * 4345 * This function does *not* take care of syncing data in case of O_SYNC write. 4346 * A caller has to handle it. This is mainly due to the fact that we want to 4347 * avoid syncing under i_rwsem. 4348 * 4349 * Return: 4350 * * number of bytes written, even for truncated writes 4351 * * negative error code if no data has been written at all 4352 */ 4353 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from) 4354 { 4355 struct file *file = iocb->ki_filp; 4356 struct address_space *mapping = file->f_mapping; 4357 struct inode *inode = mapping->host; 4358 ssize_t ret; 4359 4360 ret = file_remove_privs(file); 4361 if (ret) 4362 return ret; 4363 4364 ret = file_update_time(file); 4365 if (ret) 4366 return ret; 4367 4368 if (iocb->ki_flags & IOCB_DIRECT) { 4369 ret = generic_file_direct_write(iocb, from); 4370 /* 4371 * If the write stopped short of completing, fall back to 4372 * buffered writes. Some filesystems do this for writes to 4373 * holes, for example. For DAX files, a buffered write will 4374 * not succeed (even if it did, DAX does not handle dirty 4375 * page-cache pages correctly). 4376 */ 4377 if (ret < 0 || !iov_iter_count(from) || IS_DAX(inode)) 4378 return ret; 4379 return direct_write_fallback(iocb, from, ret, 4380 generic_perform_write(iocb, from)); 4381 } 4382 4383 return generic_perform_write(iocb, from); 4384 } 4385 EXPORT_SYMBOL(__generic_file_write_iter); 4386 4387 /** 4388 * generic_file_write_iter - write data to a file 4389 * @iocb: IO state structure 4390 * @from: iov_iter with data to write 4391 * 4392 * This is a wrapper around __generic_file_write_iter() to be used by most 4393 * filesystems. It takes care of syncing the file in case of O_SYNC file 4394 * and acquires i_rwsem as needed. 4395 * Return: 4396 * * negative error code if no data has been written at all of 4397 * vfs_fsync_range() failed for a synchronous write 4398 * * number of bytes written, even for truncated writes 4399 */ 4400 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from) 4401 { 4402 struct file *file = iocb->ki_filp; 4403 struct inode *inode = file->f_mapping->host; 4404 ssize_t ret; 4405 4406 inode_lock(inode); 4407 ret = generic_write_checks(iocb, from); 4408 if (ret > 0) 4409 ret = __generic_file_write_iter(iocb, from); 4410 inode_unlock(inode); 4411 4412 if (ret > 0) 4413 ret = generic_write_sync(iocb, ret); 4414 return ret; 4415 } 4416 EXPORT_SYMBOL(generic_file_write_iter); 4417 4418 /** 4419 * filemap_release_folio() - Release fs-specific metadata on a folio. 4420 * @folio: The folio which the kernel is trying to free. 4421 * @gfp: Memory allocation flags (and I/O mode). 4422 * 4423 * The address_space is trying to release any data attached to a folio 4424 * (presumably at folio->private). 4425 * 4426 * This will also be called if the private_2 flag is set on a page, 4427 * indicating that the folio has other metadata associated with it. 4428 * 4429 * The @gfp argument specifies whether I/O may be performed to release 4430 * this page (__GFP_IO), and whether the call may block 4431 * (__GFP_RECLAIM & __GFP_FS). 4432 * 4433 * Return: %true if the release was successful, otherwise %false. 4434 */ 4435 bool filemap_release_folio(struct folio *folio, gfp_t gfp) 4436 { 4437 struct address_space * const mapping = folio->mapping; 4438 4439 BUG_ON(!folio_test_locked(folio)); 4440 if (!folio_needs_release(folio)) 4441 return true; 4442 if (folio_test_writeback(folio)) 4443 return false; 4444 4445 if (mapping && mapping->a_ops->release_folio) 4446 return mapping->a_ops->release_folio(folio, gfp); 4447 return try_to_free_buffers(folio); 4448 } 4449 EXPORT_SYMBOL(filemap_release_folio); 4450 4451 /** 4452 * filemap_invalidate_inode - Invalidate/forcibly write back a range of an inode's pagecache 4453 * @inode: The inode to flush 4454 * @flush: Set to write back rather than simply invalidate. 4455 * @start: First byte to in range. 4456 * @end: Last byte in range (inclusive), or LLONG_MAX for everything from start 4457 * onwards. 4458 * 4459 * Invalidate all the folios on an inode that contribute to the specified 4460 * range, possibly writing them back first. Whilst the operation is 4461 * undertaken, the invalidate lock is held to prevent new folios from being 4462 * installed. 4463 */ 4464 int filemap_invalidate_inode(struct inode *inode, bool flush, 4465 loff_t start, loff_t end) 4466 { 4467 struct address_space *mapping = inode->i_mapping; 4468 pgoff_t first = start >> PAGE_SHIFT; 4469 pgoff_t last = end >> PAGE_SHIFT; 4470 pgoff_t nr = end == LLONG_MAX ? ULONG_MAX : last - first + 1; 4471 4472 if (!mapping || !mapping->nrpages || end < start) 4473 goto out; 4474 4475 /* Prevent new folios from being added to the inode. */ 4476 filemap_invalidate_lock(mapping); 4477 4478 if (!mapping->nrpages) 4479 goto unlock; 4480 4481 unmap_mapping_pages(mapping, first, nr, false); 4482 4483 /* Write back the data if we're asked to. */ 4484 if (flush) { 4485 struct writeback_control wbc = { 4486 .sync_mode = WB_SYNC_ALL, 4487 .nr_to_write = LONG_MAX, 4488 .range_start = start, 4489 .range_end = end, 4490 }; 4491 4492 filemap_fdatawrite_wbc(mapping, &wbc); 4493 } 4494 4495 /* Wait for writeback to complete on all folios and discard. */ 4496 invalidate_inode_pages2_range(mapping, start / PAGE_SIZE, end / PAGE_SIZE); 4497 4498 unlock: 4499 filemap_invalidate_unlock(mapping); 4500 out: 4501 return filemap_check_errors(mapping); 4502 } 4503 EXPORT_SYMBOL_GPL(filemap_invalidate_inode); 4504 4505 #ifdef CONFIG_CACHESTAT_SYSCALL 4506 /** 4507 * filemap_cachestat() - compute the page cache statistics of a mapping 4508 * @mapping: The mapping to compute the statistics for. 4509 * @first_index: The starting page cache index. 4510 * @last_index: The final page index (inclusive). 4511 * @cs: the cachestat struct to write the result to. 4512 * 4513 * This will query the page cache statistics of a mapping in the 4514 * page range of [first_index, last_index] (inclusive). The statistics 4515 * queried include: number of dirty pages, number of pages marked for 4516 * writeback, and the number of (recently) evicted pages. 4517 */ 4518 static void filemap_cachestat(struct address_space *mapping, 4519 pgoff_t first_index, pgoff_t last_index, struct cachestat *cs) 4520 { 4521 XA_STATE(xas, &mapping->i_pages, first_index); 4522 struct folio *folio; 4523 4524 /* Flush stats (and potentially sleep) outside the RCU read section. */ 4525 mem_cgroup_flush_stats_ratelimited(NULL); 4526 4527 rcu_read_lock(); 4528 xas_for_each(&xas, folio, last_index) { 4529 int order; 4530 unsigned long nr_pages; 4531 pgoff_t folio_first_index, folio_last_index; 4532 4533 /* 4534 * Don't deref the folio. It is not pinned, and might 4535 * get freed (and reused) underneath us. 4536 * 4537 * We *could* pin it, but that would be expensive for 4538 * what should be a fast and lightweight syscall. 4539 * 4540 * Instead, derive all information of interest from 4541 * the rcu-protected xarray. 4542 */ 4543 4544 if (xas_retry(&xas, folio)) 4545 continue; 4546 4547 order = xas_get_order(&xas); 4548 nr_pages = 1 << order; 4549 folio_first_index = round_down(xas.xa_index, 1 << order); 4550 folio_last_index = folio_first_index + nr_pages - 1; 4551 4552 /* Folios might straddle the range boundaries, only count covered pages */ 4553 if (folio_first_index < first_index) 4554 nr_pages -= first_index - folio_first_index; 4555 4556 if (folio_last_index > last_index) 4557 nr_pages -= folio_last_index - last_index; 4558 4559 if (xa_is_value(folio)) { 4560 /* page is evicted */ 4561 void *shadow = (void *)folio; 4562 bool workingset; /* not used */ 4563 4564 cs->nr_evicted += nr_pages; 4565 4566 #ifdef CONFIG_SWAP /* implies CONFIG_MMU */ 4567 if (shmem_mapping(mapping)) { 4568 /* shmem file - in swap cache */ 4569 swp_entry_t swp = radix_to_swp_entry(folio); 4570 4571 /* swapin error results in poisoned entry */ 4572 if (non_swap_entry(swp)) 4573 goto resched; 4574 4575 /* 4576 * Getting a swap entry from the shmem 4577 * inode means we beat 4578 * shmem_unuse(). rcu_read_lock() 4579 * ensures swapoff waits for us before 4580 * freeing the swapper space. However, 4581 * we can race with swapping and 4582 * invalidation, so there might not be 4583 * a shadow in the swapcache (yet). 4584 */ 4585 shadow = swap_cache_get_shadow(swp); 4586 if (!shadow) 4587 goto resched; 4588 } 4589 #endif 4590 if (workingset_test_recent(shadow, true, &workingset, false)) 4591 cs->nr_recently_evicted += nr_pages; 4592 4593 goto resched; 4594 } 4595 4596 /* page is in cache */ 4597 cs->nr_cache += nr_pages; 4598 4599 if (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY)) 4600 cs->nr_dirty += nr_pages; 4601 4602 if (xas_get_mark(&xas, PAGECACHE_TAG_WRITEBACK)) 4603 cs->nr_writeback += nr_pages; 4604 4605 resched: 4606 if (need_resched()) { 4607 xas_pause(&xas); 4608 cond_resched_rcu(); 4609 } 4610 } 4611 rcu_read_unlock(); 4612 } 4613 4614 /* 4615 * See mincore: reveal pagecache information only for files 4616 * that the calling process has write access to, or could (if 4617 * tried) open for writing. 4618 */ 4619 static inline bool can_do_cachestat(struct file *f) 4620 { 4621 if (f->f_mode & FMODE_WRITE) 4622 return true; 4623 if (inode_owner_or_capable(file_mnt_idmap(f), file_inode(f))) 4624 return true; 4625 return file_permission(f, MAY_WRITE) == 0; 4626 } 4627 4628 /* 4629 * The cachestat(2) system call. 4630 * 4631 * cachestat() returns the page cache statistics of a file in the 4632 * bytes range specified by `off` and `len`: number of cached pages, 4633 * number of dirty pages, number of pages marked for writeback, 4634 * number of evicted pages, and number of recently evicted pages. 4635 * 4636 * An evicted page is a page that is previously in the page cache 4637 * but has been evicted since. A page is recently evicted if its last 4638 * eviction was recent enough that its reentry to the cache would 4639 * indicate that it is actively being used by the system, and that 4640 * there is memory pressure on the system. 4641 * 4642 * `off` and `len` must be non-negative integers. If `len` > 0, 4643 * the queried range is [`off`, `off` + `len`]. If `len` == 0, 4644 * we will query in the range from `off` to the end of the file. 4645 * 4646 * The `flags` argument is unused for now, but is included for future 4647 * extensibility. User should pass 0 (i.e no flag specified). 4648 * 4649 * Currently, hugetlbfs is not supported. 4650 * 4651 * Because the status of a page can change after cachestat() checks it 4652 * but before it returns to the application, the returned values may 4653 * contain stale information. 4654 * 4655 * return values: 4656 * zero - success 4657 * -EFAULT - cstat or cstat_range points to an illegal address 4658 * -EINVAL - invalid flags 4659 * -EBADF - invalid file descriptor 4660 * -EOPNOTSUPP - file descriptor is of a hugetlbfs file 4661 */ 4662 SYSCALL_DEFINE4(cachestat, unsigned int, fd, 4663 struct cachestat_range __user *, cstat_range, 4664 struct cachestat __user *, cstat, unsigned int, flags) 4665 { 4666 CLASS(fd, f)(fd); 4667 struct address_space *mapping; 4668 struct cachestat_range csr; 4669 struct cachestat cs; 4670 pgoff_t first_index, last_index; 4671 4672 if (fd_empty(f)) 4673 return -EBADF; 4674 4675 if (copy_from_user(&csr, cstat_range, 4676 sizeof(struct cachestat_range))) 4677 return -EFAULT; 4678 4679 /* hugetlbfs is not supported */ 4680 if (is_file_hugepages(fd_file(f))) 4681 return -EOPNOTSUPP; 4682 4683 if (!can_do_cachestat(fd_file(f))) 4684 return -EPERM; 4685 4686 if (flags != 0) 4687 return -EINVAL; 4688 4689 first_index = csr.off >> PAGE_SHIFT; 4690 last_index = 4691 csr.len == 0 ? ULONG_MAX : (csr.off + csr.len - 1) >> PAGE_SHIFT; 4692 memset(&cs, 0, sizeof(struct cachestat)); 4693 mapping = fd_file(f)->f_mapping; 4694 filemap_cachestat(mapping, first_index, last_index, &cs); 4695 4696 if (copy_to_user(cstat, &cs, sizeof(struct cachestat))) 4697 return -EFAULT; 4698 4699 return 0; 4700 } 4701 #endif /* CONFIG_CACHESTAT_SYSCALL */