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/mman.h> 25 #include <linux/pagemap.h> 26 #include <linux/file.h> 27 #include <linux/uio.h> 28 #include <linux/error-injection.h> 29 #include <linux/hash.h> 30 #include <linux/writeback.h> 31 #include <linux/backing-dev.h> 32 #include <linux/pagevec.h> 33 #include <linux/security.h> 34 #include <linux/cpuset.h> 35 #include <linux/hugetlb.h> 36 #include <linux/memcontrol.h> 37 #include <linux/cleancache.h> 38 #include <linux/shmem_fs.h> 39 #include <linux/rmap.h> 40 #include <linux/delayacct.h> 41 #include <linux/psi.h> 42 #include <linux/ramfs.h> 43 #include <linux/page_idle.h> 44 #include <asm/pgalloc.h> 45 #include <asm/tlbflush.h> 46 #include "internal.h" 47 48 #define CREATE_TRACE_POINTS 49 #include <trace/events/filemap.h> 50 51 /* 52 * FIXME: remove all knowledge of the buffer layer from the core VM 53 */ 54 #include <linux/buffer_head.h> /* for try_to_free_buffers */ 55 56 #include <asm/mman.h> 57 58 /* 59 * Shared mappings implemented 30.11.1994. It's not fully working yet, 60 * though. 61 * 62 * Shared mappings now work. 15.8.1995 Bruno. 63 * 64 * finished 'unifying' the page and buffer cache and SMP-threaded the 65 * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com> 66 * 67 * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de> 68 */ 69 70 /* 71 * Lock ordering: 72 * 73 * ->i_mmap_rwsem (truncate_pagecache) 74 * ->private_lock (__free_pte->__set_page_dirty_buffers) 75 * ->swap_lock (exclusive_swap_page, others) 76 * ->i_pages lock 77 * 78 * ->i_rwsem 79 * ->invalidate_lock (acquired by fs in truncate path) 80 * ->i_mmap_rwsem (truncate->unmap_mapping_range) 81 * 82 * ->mmap_lock 83 * ->i_mmap_rwsem 84 * ->page_table_lock or pte_lock (various, mainly in memory.c) 85 * ->i_pages lock (arch-dependent flush_dcache_mmap_lock) 86 * 87 * ->mmap_lock 88 * ->invalidate_lock (filemap_fault) 89 * ->lock_page (filemap_fault, access_process_vm) 90 * 91 * ->i_rwsem (generic_perform_write) 92 * ->mmap_lock (fault_in_readable->do_page_fault) 93 * 94 * bdi->wb.list_lock 95 * sb_lock (fs/fs-writeback.c) 96 * ->i_pages lock (__sync_single_inode) 97 * 98 * ->i_mmap_rwsem 99 * ->anon_vma.lock (vma_adjust) 100 * 101 * ->anon_vma.lock 102 * ->page_table_lock or pte_lock (anon_vma_prepare and various) 103 * 104 * ->page_table_lock or pte_lock 105 * ->swap_lock (try_to_unmap_one) 106 * ->private_lock (try_to_unmap_one) 107 * ->i_pages lock (try_to_unmap_one) 108 * ->lruvec->lru_lock (follow_page->mark_page_accessed) 109 * ->lruvec->lru_lock (check_pte_range->isolate_lru_page) 110 * ->private_lock (page_remove_rmap->set_page_dirty) 111 * ->i_pages lock (page_remove_rmap->set_page_dirty) 112 * bdi.wb->list_lock (page_remove_rmap->set_page_dirty) 113 * ->inode->i_lock (page_remove_rmap->set_page_dirty) 114 * ->memcg->move_lock (page_remove_rmap->lock_page_memcg) 115 * bdi.wb->list_lock (zap_pte_range->set_page_dirty) 116 * ->inode->i_lock (zap_pte_range->set_page_dirty) 117 * ->private_lock (zap_pte_range->__set_page_dirty_buffers) 118 * 119 * ->i_mmap_rwsem 120 * ->tasklist_lock (memory_failure, collect_procs_ao) 121 */ 122 123 static void page_cache_delete(struct address_space *mapping, 124 struct folio *folio, void *shadow) 125 { 126 XA_STATE(xas, &mapping->i_pages, folio->index); 127 long nr = 1; 128 129 mapping_set_update(&xas, mapping); 130 131 /* hugetlb pages are represented by a single entry in the xarray */ 132 if (!folio_test_hugetlb(folio)) { 133 xas_set_order(&xas, folio->index, folio_order(folio)); 134 nr = folio_nr_pages(folio); 135 } 136 137 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 138 139 xas_store(&xas, shadow); 140 xas_init_marks(&xas); 141 142 folio->mapping = NULL; 143 /* Leave page->index set: truncation lookup relies upon it */ 144 mapping->nrpages -= nr; 145 } 146 147 static void filemap_unaccount_folio(struct address_space *mapping, 148 struct folio *folio) 149 { 150 long nr; 151 152 /* 153 * if we're uptodate, flush out into the cleancache, otherwise 154 * invalidate any existing cleancache entries. We can't leave 155 * stale data around in the cleancache once our page is gone 156 */ 157 if (folio_test_uptodate(folio) && folio_test_mappedtodisk(folio)) 158 cleancache_put_page(&folio->page); 159 else 160 cleancache_invalidate_page(mapping, &folio->page); 161 162 VM_BUG_ON_FOLIO(folio_mapped(folio), folio); 163 if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(folio_mapped(folio))) { 164 int mapcount; 165 166 pr_alert("BUG: Bad page cache in process %s pfn:%05lx\n", 167 current->comm, folio_pfn(folio)); 168 dump_page(&folio->page, "still mapped when deleted"); 169 dump_stack(); 170 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 171 172 mapcount = page_mapcount(&folio->page); 173 if (mapping_exiting(mapping) && 174 folio_ref_count(folio) >= mapcount + 2) { 175 /* 176 * All vmas have already been torn down, so it's 177 * a good bet that actually the folio is unmapped, 178 * and we'd prefer not to leak it: if we're wrong, 179 * some other bad page check should catch it later. 180 */ 181 page_mapcount_reset(&folio->page); 182 folio_ref_sub(folio, mapcount); 183 } 184 } 185 186 /* hugetlb folios do not participate in page cache accounting. */ 187 if (folio_test_hugetlb(folio)) 188 return; 189 190 nr = folio_nr_pages(folio); 191 192 __lruvec_stat_mod_folio(folio, NR_FILE_PAGES, -nr); 193 if (folio_test_swapbacked(folio)) { 194 __lruvec_stat_mod_folio(folio, NR_SHMEM, -nr); 195 if (folio_test_pmd_mappable(folio)) 196 __lruvec_stat_mod_folio(folio, NR_SHMEM_THPS, -nr); 197 } else if (folio_test_pmd_mappable(folio)) { 198 __lruvec_stat_mod_folio(folio, NR_FILE_THPS, -nr); 199 filemap_nr_thps_dec(mapping); 200 } 201 202 /* 203 * At this point folio must be either written or cleaned by 204 * truncate. Dirty folio here signals a bug and loss of 205 * unwritten data. 206 * 207 * This fixes dirty accounting after removing the folio entirely 208 * but leaves the dirty flag set: it has no effect for truncated 209 * folio and anyway will be cleared before returning folio to 210 * buddy allocator. 211 */ 212 if (WARN_ON_ONCE(folio_test_dirty(folio))) 213 folio_account_cleaned(folio, mapping, 214 inode_to_wb(mapping->host)); 215 } 216 217 /* 218 * Delete a page from the page cache and free it. Caller has to make 219 * sure the page is locked and that nobody else uses it - or that usage 220 * is safe. The caller must hold the i_pages lock. 221 */ 222 void __filemap_remove_folio(struct folio *folio, void *shadow) 223 { 224 struct address_space *mapping = folio->mapping; 225 226 trace_mm_filemap_delete_from_page_cache(folio); 227 filemap_unaccount_folio(mapping, folio); 228 page_cache_delete(mapping, folio, shadow); 229 } 230 231 void filemap_free_folio(struct address_space *mapping, struct folio *folio) 232 { 233 void (*freepage)(struct page *); 234 235 freepage = mapping->a_ops->freepage; 236 if (freepage) 237 freepage(&folio->page); 238 239 if (folio_test_large(folio) && !folio_test_hugetlb(folio)) { 240 folio_ref_sub(folio, folio_nr_pages(folio)); 241 VM_BUG_ON_FOLIO(folio_ref_count(folio) <= 0, folio); 242 } else { 243 folio_put(folio); 244 } 245 } 246 247 /** 248 * filemap_remove_folio - Remove folio from page cache. 249 * @folio: The folio. 250 * 251 * This must be called only on folios that are locked and have been 252 * verified to be in the page cache. It will never put the folio into 253 * the free list because the caller has a reference on the page. 254 */ 255 void filemap_remove_folio(struct folio *folio) 256 { 257 struct address_space *mapping = folio->mapping; 258 259 BUG_ON(!folio_test_locked(folio)); 260 spin_lock(&mapping->host->i_lock); 261 xa_lock_irq(&mapping->i_pages); 262 __filemap_remove_folio(folio, NULL); 263 xa_unlock_irq(&mapping->i_pages); 264 if (mapping_shrinkable(mapping)) 265 inode_add_lru(mapping->host); 266 spin_unlock(&mapping->host->i_lock); 267 268 filemap_free_folio(mapping, folio); 269 } 270 271 /* 272 * page_cache_delete_batch - delete several folios from page cache 273 * @mapping: the mapping to which folios belong 274 * @fbatch: batch of folios to delete 275 * 276 * The function walks over mapping->i_pages and removes folios passed in 277 * @fbatch from the mapping. The function expects @fbatch to be sorted 278 * by page index and is optimised for it to be dense. 279 * It tolerates holes in @fbatch (mapping entries at those indices are not 280 * modified). 281 * 282 * The function expects the i_pages lock to be held. 283 */ 284 static void page_cache_delete_batch(struct address_space *mapping, 285 struct folio_batch *fbatch) 286 { 287 XA_STATE(xas, &mapping->i_pages, fbatch->folios[0]->index); 288 long total_pages = 0; 289 int i = 0; 290 struct folio *folio; 291 292 mapping_set_update(&xas, mapping); 293 xas_for_each(&xas, folio, ULONG_MAX) { 294 if (i >= folio_batch_count(fbatch)) 295 break; 296 297 /* A swap/dax/shadow entry got inserted? Skip it. */ 298 if (xa_is_value(folio)) 299 continue; 300 /* 301 * A page got inserted in our range? Skip it. We have our 302 * pages locked so they are protected from being removed. 303 * If we see a page whose index is higher than ours, it 304 * means our page has been removed, which shouldn't be 305 * possible because we're holding the PageLock. 306 */ 307 if (folio != fbatch->folios[i]) { 308 VM_BUG_ON_FOLIO(folio->index > 309 fbatch->folios[i]->index, folio); 310 continue; 311 } 312 313 WARN_ON_ONCE(!folio_test_locked(folio)); 314 315 folio->mapping = NULL; 316 /* Leave folio->index set: truncation lookup relies on it */ 317 318 i++; 319 xas_store(&xas, NULL); 320 total_pages += folio_nr_pages(folio); 321 } 322 mapping->nrpages -= total_pages; 323 } 324 325 void delete_from_page_cache_batch(struct address_space *mapping, 326 struct folio_batch *fbatch) 327 { 328 int i; 329 330 if (!folio_batch_count(fbatch)) 331 return; 332 333 spin_lock(&mapping->host->i_lock); 334 xa_lock_irq(&mapping->i_pages); 335 for (i = 0; i < folio_batch_count(fbatch); i++) { 336 struct folio *folio = fbatch->folios[i]; 337 338 trace_mm_filemap_delete_from_page_cache(folio); 339 filemap_unaccount_folio(mapping, folio); 340 } 341 page_cache_delete_batch(mapping, fbatch); 342 xa_unlock_irq(&mapping->i_pages); 343 if (mapping_shrinkable(mapping)) 344 inode_add_lru(mapping->host); 345 spin_unlock(&mapping->host->i_lock); 346 347 for (i = 0; i < folio_batch_count(fbatch); i++) 348 filemap_free_folio(mapping, fbatch->folios[i]); 349 } 350 351 int filemap_check_errors(struct address_space *mapping) 352 { 353 int ret = 0; 354 /* Check for outstanding write errors */ 355 if (test_bit(AS_ENOSPC, &mapping->flags) && 356 test_and_clear_bit(AS_ENOSPC, &mapping->flags)) 357 ret = -ENOSPC; 358 if (test_bit(AS_EIO, &mapping->flags) && 359 test_and_clear_bit(AS_EIO, &mapping->flags)) 360 ret = -EIO; 361 return ret; 362 } 363 EXPORT_SYMBOL(filemap_check_errors); 364 365 static int filemap_check_and_keep_errors(struct address_space *mapping) 366 { 367 /* Check for outstanding write errors */ 368 if (test_bit(AS_EIO, &mapping->flags)) 369 return -EIO; 370 if (test_bit(AS_ENOSPC, &mapping->flags)) 371 return -ENOSPC; 372 return 0; 373 } 374 375 /** 376 * filemap_fdatawrite_wbc - start writeback on mapping dirty pages in range 377 * @mapping: address space structure to write 378 * @wbc: the writeback_control controlling the writeout 379 * 380 * Call writepages on the mapping using the provided wbc to control the 381 * writeout. 382 * 383 * Return: %0 on success, negative error code otherwise. 384 */ 385 int filemap_fdatawrite_wbc(struct address_space *mapping, 386 struct writeback_control *wbc) 387 { 388 int ret; 389 390 if (!mapping_can_writeback(mapping) || 391 !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) 392 return 0; 393 394 wbc_attach_fdatawrite_inode(wbc, mapping->host); 395 ret = do_writepages(mapping, wbc); 396 wbc_detach_inode(wbc); 397 return ret; 398 } 399 EXPORT_SYMBOL(filemap_fdatawrite_wbc); 400 401 /** 402 * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range 403 * @mapping: address space structure to write 404 * @start: offset in bytes where the range starts 405 * @end: offset in bytes where the range ends (inclusive) 406 * @sync_mode: enable synchronous operation 407 * 408 * Start writeback against all of a mapping's dirty pages that lie 409 * within the byte offsets <start, end> inclusive. 410 * 411 * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as 412 * opposed to a regular memory cleansing writeback. The difference between 413 * these two operations is that if a dirty page/buffer is encountered, it must 414 * be waited upon, and not just skipped over. 415 * 416 * Return: %0 on success, negative error code otherwise. 417 */ 418 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start, 419 loff_t end, int sync_mode) 420 { 421 struct writeback_control wbc = { 422 .sync_mode = sync_mode, 423 .nr_to_write = LONG_MAX, 424 .range_start = start, 425 .range_end = end, 426 }; 427 428 return filemap_fdatawrite_wbc(mapping, &wbc); 429 } 430 431 static inline int __filemap_fdatawrite(struct address_space *mapping, 432 int sync_mode) 433 { 434 return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode); 435 } 436 437 int filemap_fdatawrite(struct address_space *mapping) 438 { 439 return __filemap_fdatawrite(mapping, WB_SYNC_ALL); 440 } 441 EXPORT_SYMBOL(filemap_fdatawrite); 442 443 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start, 444 loff_t end) 445 { 446 return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL); 447 } 448 EXPORT_SYMBOL(filemap_fdatawrite_range); 449 450 /** 451 * filemap_flush - mostly a non-blocking flush 452 * @mapping: target address_space 453 * 454 * This is a mostly non-blocking flush. Not suitable for data-integrity 455 * purposes - I/O may not be started against all dirty pages. 456 * 457 * Return: %0 on success, negative error code otherwise. 458 */ 459 int filemap_flush(struct address_space *mapping) 460 { 461 return __filemap_fdatawrite(mapping, WB_SYNC_NONE); 462 } 463 EXPORT_SYMBOL(filemap_flush); 464 465 /** 466 * filemap_range_has_page - check if a page exists in range. 467 * @mapping: address space within which to check 468 * @start_byte: offset in bytes where the range starts 469 * @end_byte: offset in bytes where the range ends (inclusive) 470 * 471 * Find at least one page in the range supplied, usually used to check if 472 * direct writing in this range will trigger a writeback. 473 * 474 * Return: %true if at least one page exists in the specified range, 475 * %false otherwise. 476 */ 477 bool filemap_range_has_page(struct address_space *mapping, 478 loff_t start_byte, loff_t end_byte) 479 { 480 struct page *page; 481 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT); 482 pgoff_t max = end_byte >> PAGE_SHIFT; 483 484 if (end_byte < start_byte) 485 return false; 486 487 rcu_read_lock(); 488 for (;;) { 489 page = xas_find(&xas, max); 490 if (xas_retry(&xas, page)) 491 continue; 492 /* Shadow entries don't count */ 493 if (xa_is_value(page)) 494 continue; 495 /* 496 * We don't need to try to pin this page; we're about to 497 * release the RCU lock anyway. It is enough to know that 498 * there was a page here recently. 499 */ 500 break; 501 } 502 rcu_read_unlock(); 503 504 return page != NULL; 505 } 506 EXPORT_SYMBOL(filemap_range_has_page); 507 508 static void __filemap_fdatawait_range(struct address_space *mapping, 509 loff_t start_byte, loff_t end_byte) 510 { 511 pgoff_t index = start_byte >> PAGE_SHIFT; 512 pgoff_t end = end_byte >> PAGE_SHIFT; 513 struct pagevec pvec; 514 int nr_pages; 515 516 if (end_byte < start_byte) 517 return; 518 519 pagevec_init(&pvec); 520 while (index <= end) { 521 unsigned i; 522 523 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, 524 end, PAGECACHE_TAG_WRITEBACK); 525 if (!nr_pages) 526 break; 527 528 for (i = 0; i < nr_pages; i++) { 529 struct page *page = pvec.pages[i]; 530 531 wait_on_page_writeback(page); 532 ClearPageError(page); 533 } 534 pagevec_release(&pvec); 535 cond_resched(); 536 } 537 } 538 539 /** 540 * filemap_fdatawait_range - wait for writeback to complete 541 * @mapping: address space structure to wait for 542 * @start_byte: offset in bytes where the range starts 543 * @end_byte: offset in bytes where the range ends (inclusive) 544 * 545 * Walk the list of under-writeback pages of the given address space 546 * in the given range and wait for all of them. Check error status of 547 * the address space and return it. 548 * 549 * Since the error status of the address space is cleared by this function, 550 * callers are responsible for checking the return value and handling and/or 551 * reporting the error. 552 * 553 * Return: error status of the address space. 554 */ 555 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte, 556 loff_t end_byte) 557 { 558 __filemap_fdatawait_range(mapping, start_byte, end_byte); 559 return filemap_check_errors(mapping); 560 } 561 EXPORT_SYMBOL(filemap_fdatawait_range); 562 563 /** 564 * filemap_fdatawait_range_keep_errors - wait for writeback to complete 565 * @mapping: address space structure to wait for 566 * @start_byte: offset in bytes where the range starts 567 * @end_byte: offset in bytes where the range ends (inclusive) 568 * 569 * Walk the list of under-writeback pages of the given address space in the 570 * given range and wait for all of them. Unlike filemap_fdatawait_range(), 571 * this function does not clear error status of the address space. 572 * 573 * Use this function if callers don't handle errors themselves. Expected 574 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2), 575 * fsfreeze(8) 576 */ 577 int filemap_fdatawait_range_keep_errors(struct address_space *mapping, 578 loff_t start_byte, loff_t end_byte) 579 { 580 __filemap_fdatawait_range(mapping, start_byte, end_byte); 581 return filemap_check_and_keep_errors(mapping); 582 } 583 EXPORT_SYMBOL(filemap_fdatawait_range_keep_errors); 584 585 /** 586 * file_fdatawait_range - wait for writeback to complete 587 * @file: file pointing to address space structure to wait for 588 * @start_byte: offset in bytes where the range starts 589 * @end_byte: offset in bytes where the range ends (inclusive) 590 * 591 * Walk the list of under-writeback pages of the address space that file 592 * refers to, in the given range and wait for all of them. Check error 593 * status of the address space vs. the file->f_wb_err cursor and return it. 594 * 595 * Since the error status of the file is advanced by this function, 596 * callers are responsible for checking the return value and handling and/or 597 * reporting the error. 598 * 599 * Return: error status of the address space vs. the file->f_wb_err cursor. 600 */ 601 int file_fdatawait_range(struct file *file, loff_t start_byte, loff_t end_byte) 602 { 603 struct address_space *mapping = file->f_mapping; 604 605 __filemap_fdatawait_range(mapping, start_byte, end_byte); 606 return file_check_and_advance_wb_err(file); 607 } 608 EXPORT_SYMBOL(file_fdatawait_range); 609 610 /** 611 * filemap_fdatawait_keep_errors - wait for writeback without clearing errors 612 * @mapping: address space structure to wait for 613 * 614 * Walk the list of under-writeback pages of the given address space 615 * and wait for all of them. Unlike filemap_fdatawait(), this function 616 * does not clear error status of the address space. 617 * 618 * Use this function if callers don't handle errors themselves. Expected 619 * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2), 620 * fsfreeze(8) 621 * 622 * Return: error status of the address space. 623 */ 624 int filemap_fdatawait_keep_errors(struct address_space *mapping) 625 { 626 __filemap_fdatawait_range(mapping, 0, LLONG_MAX); 627 return filemap_check_and_keep_errors(mapping); 628 } 629 EXPORT_SYMBOL(filemap_fdatawait_keep_errors); 630 631 /* Returns true if writeback might be needed or already in progress. */ 632 static bool mapping_needs_writeback(struct address_space *mapping) 633 { 634 return mapping->nrpages; 635 } 636 637 bool filemap_range_has_writeback(struct address_space *mapping, 638 loff_t start_byte, loff_t end_byte) 639 { 640 XA_STATE(xas, &mapping->i_pages, start_byte >> PAGE_SHIFT); 641 pgoff_t max = end_byte >> PAGE_SHIFT; 642 struct page *page; 643 644 if (end_byte < start_byte) 645 return false; 646 647 rcu_read_lock(); 648 xas_for_each(&xas, page, max) { 649 if (xas_retry(&xas, page)) 650 continue; 651 if (xa_is_value(page)) 652 continue; 653 if (PageDirty(page) || PageLocked(page) || PageWriteback(page)) 654 break; 655 } 656 rcu_read_unlock(); 657 return page != NULL; 658 } 659 EXPORT_SYMBOL_GPL(filemap_range_has_writeback); 660 661 /** 662 * filemap_write_and_wait_range - write out & wait on a file range 663 * @mapping: the address_space for the pages 664 * @lstart: offset in bytes where the range starts 665 * @lend: offset in bytes where the range ends (inclusive) 666 * 667 * Write out and wait upon file offsets lstart->lend, inclusive. 668 * 669 * Note that @lend is inclusive (describes the last byte to be written) so 670 * that this function can be used to write to the very end-of-file (end = -1). 671 * 672 * Return: error status of the address space. 673 */ 674 int filemap_write_and_wait_range(struct address_space *mapping, 675 loff_t lstart, loff_t lend) 676 { 677 int err = 0; 678 679 if (mapping_needs_writeback(mapping)) { 680 err = __filemap_fdatawrite_range(mapping, lstart, lend, 681 WB_SYNC_ALL); 682 /* 683 * Even if the above returned error, the pages may be 684 * written partially (e.g. -ENOSPC), so we wait for it. 685 * But the -EIO is special case, it may indicate the worst 686 * thing (e.g. bug) happened, so we avoid waiting for it. 687 */ 688 if (err != -EIO) { 689 int err2 = filemap_fdatawait_range(mapping, 690 lstart, lend); 691 if (!err) 692 err = err2; 693 } else { 694 /* Clear any previously stored errors */ 695 filemap_check_errors(mapping); 696 } 697 } else { 698 err = filemap_check_errors(mapping); 699 } 700 return err; 701 } 702 EXPORT_SYMBOL(filemap_write_and_wait_range); 703 704 void __filemap_set_wb_err(struct address_space *mapping, int err) 705 { 706 errseq_t eseq = errseq_set(&mapping->wb_err, err); 707 708 trace_filemap_set_wb_err(mapping, eseq); 709 } 710 EXPORT_SYMBOL(__filemap_set_wb_err); 711 712 /** 713 * file_check_and_advance_wb_err - report wb error (if any) that was previously 714 * and advance wb_err to current one 715 * @file: struct file on which the error is being reported 716 * 717 * When userland calls fsync (or something like nfsd does the equivalent), we 718 * want to report any writeback errors that occurred since the last fsync (or 719 * since the file was opened if there haven't been any). 720 * 721 * Grab the wb_err from the mapping. If it matches what we have in the file, 722 * then just quickly return 0. The file is all caught up. 723 * 724 * If it doesn't match, then take the mapping value, set the "seen" flag in 725 * it and try to swap it into place. If it works, or another task beat us 726 * to it with the new value, then update the f_wb_err and return the error 727 * portion. The error at this point must be reported via proper channels 728 * (a'la fsync, or NFS COMMIT operation, etc.). 729 * 730 * While we handle mapping->wb_err with atomic operations, the f_wb_err 731 * value is protected by the f_lock since we must ensure that it reflects 732 * the latest value swapped in for this file descriptor. 733 * 734 * Return: %0 on success, negative error code otherwise. 735 */ 736 int file_check_and_advance_wb_err(struct file *file) 737 { 738 int err = 0; 739 errseq_t old = READ_ONCE(file->f_wb_err); 740 struct address_space *mapping = file->f_mapping; 741 742 /* Locklessly handle the common case where nothing has changed */ 743 if (errseq_check(&mapping->wb_err, old)) { 744 /* Something changed, must use slow path */ 745 spin_lock(&file->f_lock); 746 old = file->f_wb_err; 747 err = errseq_check_and_advance(&mapping->wb_err, 748 &file->f_wb_err); 749 trace_file_check_and_advance_wb_err(file, old); 750 spin_unlock(&file->f_lock); 751 } 752 753 /* 754 * We're mostly using this function as a drop in replacement for 755 * filemap_check_errors. Clear AS_EIO/AS_ENOSPC to emulate the effect 756 * that the legacy code would have had on these flags. 757 */ 758 clear_bit(AS_EIO, &mapping->flags); 759 clear_bit(AS_ENOSPC, &mapping->flags); 760 return err; 761 } 762 EXPORT_SYMBOL(file_check_and_advance_wb_err); 763 764 /** 765 * file_write_and_wait_range - write out & wait on a file range 766 * @file: file pointing to address_space with pages 767 * @lstart: offset in bytes where the range starts 768 * @lend: offset in bytes where the range ends (inclusive) 769 * 770 * Write out and wait upon file offsets lstart->lend, inclusive. 771 * 772 * Note that @lend is inclusive (describes the last byte to be written) so 773 * that this function can be used to write to the very end-of-file (end = -1). 774 * 775 * After writing out and waiting on the data, we check and advance the 776 * f_wb_err cursor to the latest value, and return any errors detected there. 777 * 778 * Return: %0 on success, negative error code otherwise. 779 */ 780 int file_write_and_wait_range(struct file *file, loff_t lstart, loff_t lend) 781 { 782 int err = 0, err2; 783 struct address_space *mapping = file->f_mapping; 784 785 if (mapping_needs_writeback(mapping)) { 786 err = __filemap_fdatawrite_range(mapping, lstart, lend, 787 WB_SYNC_ALL); 788 /* See comment of filemap_write_and_wait() */ 789 if (err != -EIO) 790 __filemap_fdatawait_range(mapping, lstart, lend); 791 } 792 err2 = file_check_and_advance_wb_err(file); 793 if (!err) 794 err = err2; 795 return err; 796 } 797 EXPORT_SYMBOL(file_write_and_wait_range); 798 799 /** 800 * replace_page_cache_page - replace a pagecache page with a new one 801 * @old: page to be replaced 802 * @new: page to replace with 803 * 804 * This function replaces a page in the pagecache with a new one. On 805 * success it acquires the pagecache reference for the new page and 806 * drops it for the old page. Both the old and new pages must be 807 * locked. This function does not add the new page to the LRU, the 808 * caller must do that. 809 * 810 * The remove + add is atomic. This function cannot fail. 811 */ 812 void replace_page_cache_page(struct page *old, struct page *new) 813 { 814 struct folio *fold = page_folio(old); 815 struct folio *fnew = page_folio(new); 816 struct address_space *mapping = old->mapping; 817 void (*freepage)(struct page *) = mapping->a_ops->freepage; 818 pgoff_t offset = old->index; 819 XA_STATE(xas, &mapping->i_pages, offset); 820 821 VM_BUG_ON_PAGE(!PageLocked(old), old); 822 VM_BUG_ON_PAGE(!PageLocked(new), new); 823 VM_BUG_ON_PAGE(new->mapping, new); 824 825 get_page(new); 826 new->mapping = mapping; 827 new->index = offset; 828 829 mem_cgroup_migrate(fold, fnew); 830 831 xas_lock_irq(&xas); 832 xas_store(&xas, new); 833 834 old->mapping = NULL; 835 /* hugetlb pages do not participate in page cache accounting. */ 836 if (!PageHuge(old)) 837 __dec_lruvec_page_state(old, NR_FILE_PAGES); 838 if (!PageHuge(new)) 839 __inc_lruvec_page_state(new, NR_FILE_PAGES); 840 if (PageSwapBacked(old)) 841 __dec_lruvec_page_state(old, NR_SHMEM); 842 if (PageSwapBacked(new)) 843 __inc_lruvec_page_state(new, NR_SHMEM); 844 xas_unlock_irq(&xas); 845 if (freepage) 846 freepage(old); 847 put_page(old); 848 } 849 EXPORT_SYMBOL_GPL(replace_page_cache_page); 850 851 noinline int __filemap_add_folio(struct address_space *mapping, 852 struct folio *folio, pgoff_t index, gfp_t gfp, void **shadowp) 853 { 854 XA_STATE(xas, &mapping->i_pages, index); 855 int huge = folio_test_hugetlb(folio); 856 int error; 857 bool charged = false; 858 859 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 860 VM_BUG_ON_FOLIO(folio_test_swapbacked(folio), folio); 861 mapping_set_update(&xas, mapping); 862 863 folio_get(folio); 864 folio->mapping = mapping; 865 folio->index = index; 866 867 if (!huge) { 868 error = mem_cgroup_charge(folio, NULL, gfp); 869 VM_BUG_ON_FOLIO(index & (folio_nr_pages(folio) - 1), folio); 870 if (error) 871 goto error; 872 charged = true; 873 } 874 875 gfp &= GFP_RECLAIM_MASK; 876 877 do { 878 unsigned int order = xa_get_order(xas.xa, xas.xa_index); 879 void *entry, *old = NULL; 880 881 if (order > folio_order(folio)) 882 xas_split_alloc(&xas, xa_load(xas.xa, xas.xa_index), 883 order, gfp); 884 xas_lock_irq(&xas); 885 xas_for_each_conflict(&xas, entry) { 886 old = entry; 887 if (!xa_is_value(entry)) { 888 xas_set_err(&xas, -EEXIST); 889 goto unlock; 890 } 891 } 892 893 if (old) { 894 if (shadowp) 895 *shadowp = old; 896 /* entry may have been split before we acquired lock */ 897 order = xa_get_order(xas.xa, xas.xa_index); 898 if (order > folio_order(folio)) { 899 xas_split(&xas, old, order); 900 xas_reset(&xas); 901 } 902 } 903 904 xas_store(&xas, folio); 905 if (xas_error(&xas)) 906 goto unlock; 907 908 mapping->nrpages++; 909 910 /* hugetlb pages do not participate in page cache accounting */ 911 if (!huge) 912 __lruvec_stat_add_folio(folio, NR_FILE_PAGES); 913 unlock: 914 xas_unlock_irq(&xas); 915 } while (xas_nomem(&xas, gfp)); 916 917 if (xas_error(&xas)) { 918 error = xas_error(&xas); 919 if (charged) 920 mem_cgroup_uncharge(folio); 921 goto error; 922 } 923 924 trace_mm_filemap_add_to_page_cache(folio); 925 return 0; 926 error: 927 folio->mapping = NULL; 928 /* Leave page->index set: truncation relies upon it */ 929 folio_put(folio); 930 return error; 931 } 932 ALLOW_ERROR_INJECTION(__filemap_add_folio, ERRNO); 933 934 /** 935 * add_to_page_cache_locked - add a locked page to the pagecache 936 * @page: page to add 937 * @mapping: the page's address_space 938 * @offset: page index 939 * @gfp_mask: page allocation mode 940 * 941 * This function is used to add a page to the pagecache. It must be locked. 942 * This function does not add the page to the LRU. The caller must do that. 943 * 944 * Return: %0 on success, negative error code otherwise. 945 */ 946 int add_to_page_cache_locked(struct page *page, struct address_space *mapping, 947 pgoff_t offset, gfp_t gfp_mask) 948 { 949 return __filemap_add_folio(mapping, page_folio(page), offset, 950 gfp_mask, NULL); 951 } 952 EXPORT_SYMBOL(add_to_page_cache_locked); 953 954 int filemap_add_folio(struct address_space *mapping, struct folio *folio, 955 pgoff_t index, gfp_t gfp) 956 { 957 void *shadow = NULL; 958 int ret; 959 960 __folio_set_locked(folio); 961 ret = __filemap_add_folio(mapping, folio, index, gfp, &shadow); 962 if (unlikely(ret)) 963 __folio_clear_locked(folio); 964 else { 965 /* 966 * The folio might have been evicted from cache only 967 * recently, in which case it should be activated like 968 * any other repeatedly accessed folio. 969 * The exception is folios getting rewritten; evicting other 970 * data from the working set, only to cache data that will 971 * get overwritten with something else, is a waste of memory. 972 */ 973 WARN_ON_ONCE(folio_test_active(folio)); 974 if (!(gfp & __GFP_WRITE) && shadow) 975 workingset_refault(folio, shadow); 976 folio_add_lru(folio); 977 } 978 return ret; 979 } 980 EXPORT_SYMBOL_GPL(filemap_add_folio); 981 982 #ifdef CONFIG_NUMA 983 struct folio *filemap_alloc_folio(gfp_t gfp, unsigned int order) 984 { 985 int n; 986 struct folio *folio; 987 988 if (cpuset_do_page_mem_spread()) { 989 unsigned int cpuset_mems_cookie; 990 do { 991 cpuset_mems_cookie = read_mems_allowed_begin(); 992 n = cpuset_mem_spread_node(); 993 folio = __folio_alloc_node(gfp, order, n); 994 } while (!folio && read_mems_allowed_retry(cpuset_mems_cookie)); 995 996 return folio; 997 } 998 return folio_alloc(gfp, order); 999 } 1000 EXPORT_SYMBOL(filemap_alloc_folio); 1001 #endif 1002 1003 /* 1004 * filemap_invalidate_lock_two - lock invalidate_lock for two mappings 1005 * 1006 * Lock exclusively invalidate_lock of any passed mapping that is not NULL. 1007 * 1008 * @mapping1: the first mapping to lock 1009 * @mapping2: the second mapping to lock 1010 */ 1011 void filemap_invalidate_lock_two(struct address_space *mapping1, 1012 struct address_space *mapping2) 1013 { 1014 if (mapping1 > mapping2) 1015 swap(mapping1, mapping2); 1016 if (mapping1) 1017 down_write(&mapping1->invalidate_lock); 1018 if (mapping2 && mapping1 != mapping2) 1019 down_write_nested(&mapping2->invalidate_lock, 1); 1020 } 1021 EXPORT_SYMBOL(filemap_invalidate_lock_two); 1022 1023 /* 1024 * filemap_invalidate_unlock_two - unlock invalidate_lock for two mappings 1025 * 1026 * Unlock exclusive invalidate_lock of any passed mapping that is not NULL. 1027 * 1028 * @mapping1: the first mapping to unlock 1029 * @mapping2: the second mapping to unlock 1030 */ 1031 void filemap_invalidate_unlock_two(struct address_space *mapping1, 1032 struct address_space *mapping2) 1033 { 1034 if (mapping1) 1035 up_write(&mapping1->invalidate_lock); 1036 if (mapping2 && mapping1 != mapping2) 1037 up_write(&mapping2->invalidate_lock); 1038 } 1039 EXPORT_SYMBOL(filemap_invalidate_unlock_two); 1040 1041 /* 1042 * In order to wait for pages to become available there must be 1043 * waitqueues associated with pages. By using a hash table of 1044 * waitqueues where the bucket discipline is to maintain all 1045 * waiters on the same queue and wake all when any of the pages 1046 * become available, and for the woken contexts to check to be 1047 * sure the appropriate page became available, this saves space 1048 * at a cost of "thundering herd" phenomena during rare hash 1049 * collisions. 1050 */ 1051 #define PAGE_WAIT_TABLE_BITS 8 1052 #define PAGE_WAIT_TABLE_SIZE (1 << PAGE_WAIT_TABLE_BITS) 1053 static wait_queue_head_t folio_wait_table[PAGE_WAIT_TABLE_SIZE] __cacheline_aligned; 1054 1055 static wait_queue_head_t *folio_waitqueue(struct folio *folio) 1056 { 1057 return &folio_wait_table[hash_ptr(folio, PAGE_WAIT_TABLE_BITS)]; 1058 } 1059 1060 void __init pagecache_init(void) 1061 { 1062 int i; 1063 1064 for (i = 0; i < PAGE_WAIT_TABLE_SIZE; i++) 1065 init_waitqueue_head(&folio_wait_table[i]); 1066 1067 page_writeback_init(); 1068 } 1069 1070 /* 1071 * The page wait code treats the "wait->flags" somewhat unusually, because 1072 * we have multiple different kinds of waits, not just the usual "exclusive" 1073 * one. 1074 * 1075 * We have: 1076 * 1077 * (a) no special bits set: 1078 * 1079 * We're just waiting for the bit to be released, and when a waker 1080 * calls the wakeup function, we set WQ_FLAG_WOKEN and wake it up, 1081 * and remove it from the wait queue. 1082 * 1083 * Simple and straightforward. 1084 * 1085 * (b) WQ_FLAG_EXCLUSIVE: 1086 * 1087 * The waiter is waiting to get the lock, and only one waiter should 1088 * be woken up to avoid any thundering herd behavior. We'll set the 1089 * WQ_FLAG_WOKEN bit, wake it up, and remove it from the wait queue. 1090 * 1091 * This is the traditional exclusive wait. 1092 * 1093 * (c) WQ_FLAG_EXCLUSIVE | WQ_FLAG_CUSTOM: 1094 * 1095 * The waiter is waiting to get the bit, and additionally wants the 1096 * lock to be transferred to it for fair lock behavior. If the lock 1097 * cannot be taken, we stop walking the wait queue without waking 1098 * the waiter. 1099 * 1100 * This is the "fair lock handoff" case, and in addition to setting 1101 * WQ_FLAG_WOKEN, we set WQ_FLAG_DONE to let the waiter easily see 1102 * that it now has the lock. 1103 */ 1104 static int wake_page_function(wait_queue_entry_t *wait, unsigned mode, int sync, void *arg) 1105 { 1106 unsigned int flags; 1107 struct wait_page_key *key = arg; 1108 struct wait_page_queue *wait_page 1109 = container_of(wait, struct wait_page_queue, wait); 1110 1111 if (!wake_page_match(wait_page, key)) 1112 return 0; 1113 1114 /* 1115 * If it's a lock handoff wait, we get the bit for it, and 1116 * stop walking (and do not wake it up) if we can't. 1117 */ 1118 flags = wait->flags; 1119 if (flags & WQ_FLAG_EXCLUSIVE) { 1120 if (test_bit(key->bit_nr, &key->folio->flags)) 1121 return -1; 1122 if (flags & WQ_FLAG_CUSTOM) { 1123 if (test_and_set_bit(key->bit_nr, &key->folio->flags)) 1124 return -1; 1125 flags |= WQ_FLAG_DONE; 1126 } 1127 } 1128 1129 /* 1130 * We are holding the wait-queue lock, but the waiter that 1131 * is waiting for this will be checking the flags without 1132 * any locking. 1133 * 1134 * So update the flags atomically, and wake up the waiter 1135 * afterwards to avoid any races. This store-release pairs 1136 * with the load-acquire in folio_wait_bit_common(). 1137 */ 1138 smp_store_release(&wait->flags, flags | WQ_FLAG_WOKEN); 1139 wake_up_state(wait->private, mode); 1140 1141 /* 1142 * Ok, we have successfully done what we're waiting for, 1143 * and we can unconditionally remove the wait entry. 1144 * 1145 * Note that this pairs with the "finish_wait()" in the 1146 * waiter, and has to be the absolute last thing we do. 1147 * After this list_del_init(&wait->entry) the wait entry 1148 * might be de-allocated and the process might even have 1149 * exited. 1150 */ 1151 list_del_init_careful(&wait->entry); 1152 return (flags & WQ_FLAG_EXCLUSIVE) != 0; 1153 } 1154 1155 static void folio_wake_bit(struct folio *folio, int bit_nr) 1156 { 1157 wait_queue_head_t *q = folio_waitqueue(folio); 1158 struct wait_page_key key; 1159 unsigned long flags; 1160 wait_queue_entry_t bookmark; 1161 1162 key.folio = folio; 1163 key.bit_nr = bit_nr; 1164 key.page_match = 0; 1165 1166 bookmark.flags = 0; 1167 bookmark.private = NULL; 1168 bookmark.func = NULL; 1169 INIT_LIST_HEAD(&bookmark.entry); 1170 1171 spin_lock_irqsave(&q->lock, flags); 1172 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark); 1173 1174 while (bookmark.flags & WQ_FLAG_BOOKMARK) { 1175 /* 1176 * Take a breather from holding the lock, 1177 * allow pages that finish wake up asynchronously 1178 * to acquire the lock and remove themselves 1179 * from wait queue 1180 */ 1181 spin_unlock_irqrestore(&q->lock, flags); 1182 cpu_relax(); 1183 spin_lock_irqsave(&q->lock, flags); 1184 __wake_up_locked_key_bookmark(q, TASK_NORMAL, &key, &bookmark); 1185 } 1186 1187 /* 1188 * It is possible for other pages to have collided on the waitqueue 1189 * hash, so in that case check for a page match. That prevents a long- 1190 * term waiter 1191 * 1192 * It is still possible to miss a case here, when we woke page waiters 1193 * and removed them from the waitqueue, but there are still other 1194 * page waiters. 1195 */ 1196 if (!waitqueue_active(q) || !key.page_match) { 1197 folio_clear_waiters(folio); 1198 /* 1199 * It's possible to miss clearing Waiters here, when we woke 1200 * our page waiters, but the hashed waitqueue has waiters for 1201 * other pages on it. 1202 * 1203 * That's okay, it's a rare case. The next waker will clear it. 1204 */ 1205 } 1206 spin_unlock_irqrestore(&q->lock, flags); 1207 } 1208 1209 static void folio_wake(struct folio *folio, int bit) 1210 { 1211 if (!folio_test_waiters(folio)) 1212 return; 1213 folio_wake_bit(folio, bit); 1214 } 1215 1216 /* 1217 * A choice of three behaviors for folio_wait_bit_common(): 1218 */ 1219 enum behavior { 1220 EXCLUSIVE, /* Hold ref to page and take the bit when woken, like 1221 * __folio_lock() waiting on then setting PG_locked. 1222 */ 1223 SHARED, /* Hold ref to page and check the bit when woken, like 1224 * folio_wait_writeback() waiting on PG_writeback. 1225 */ 1226 DROP, /* Drop ref to page before wait, no check when woken, 1227 * like folio_put_wait_locked() on PG_locked. 1228 */ 1229 }; 1230 1231 /* 1232 * Attempt to check (or get) the folio flag, and mark us done 1233 * if successful. 1234 */ 1235 static inline bool folio_trylock_flag(struct folio *folio, int bit_nr, 1236 struct wait_queue_entry *wait) 1237 { 1238 if (wait->flags & WQ_FLAG_EXCLUSIVE) { 1239 if (test_and_set_bit(bit_nr, &folio->flags)) 1240 return false; 1241 } else if (test_bit(bit_nr, &folio->flags)) 1242 return false; 1243 1244 wait->flags |= WQ_FLAG_WOKEN | WQ_FLAG_DONE; 1245 return true; 1246 } 1247 1248 /* How many times do we accept lock stealing from under a waiter? */ 1249 int sysctl_page_lock_unfairness = 5; 1250 1251 static inline int folio_wait_bit_common(struct folio *folio, int bit_nr, 1252 int state, enum behavior behavior) 1253 { 1254 wait_queue_head_t *q = folio_waitqueue(folio); 1255 int unfairness = sysctl_page_lock_unfairness; 1256 struct wait_page_queue wait_page; 1257 wait_queue_entry_t *wait = &wait_page.wait; 1258 bool thrashing = false; 1259 bool delayacct = false; 1260 unsigned long pflags; 1261 1262 if (bit_nr == PG_locked && 1263 !folio_test_uptodate(folio) && folio_test_workingset(folio)) { 1264 if (!folio_test_swapbacked(folio)) { 1265 delayacct_thrashing_start(); 1266 delayacct = true; 1267 } 1268 psi_memstall_enter(&pflags); 1269 thrashing = true; 1270 } 1271 1272 init_wait(wait); 1273 wait->func = wake_page_function; 1274 wait_page.folio = folio; 1275 wait_page.bit_nr = bit_nr; 1276 1277 repeat: 1278 wait->flags = 0; 1279 if (behavior == EXCLUSIVE) { 1280 wait->flags = WQ_FLAG_EXCLUSIVE; 1281 if (--unfairness < 0) 1282 wait->flags |= WQ_FLAG_CUSTOM; 1283 } 1284 1285 /* 1286 * Do one last check whether we can get the 1287 * page bit synchronously. 1288 * 1289 * Do the folio_set_waiters() marking before that 1290 * to let any waker we _just_ missed know they 1291 * need to wake us up (otherwise they'll never 1292 * even go to the slow case that looks at the 1293 * page queue), and add ourselves to the wait 1294 * queue if we need to sleep. 1295 * 1296 * This part needs to be done under the queue 1297 * lock to avoid races. 1298 */ 1299 spin_lock_irq(&q->lock); 1300 folio_set_waiters(folio); 1301 if (!folio_trylock_flag(folio, bit_nr, wait)) 1302 __add_wait_queue_entry_tail(q, wait); 1303 spin_unlock_irq(&q->lock); 1304 1305 /* 1306 * From now on, all the logic will be based on 1307 * the WQ_FLAG_WOKEN and WQ_FLAG_DONE flag, to 1308 * see whether the page bit testing has already 1309 * been done by the wake function. 1310 * 1311 * We can drop our reference to the folio. 1312 */ 1313 if (behavior == DROP) 1314 folio_put(folio); 1315 1316 /* 1317 * Note that until the "finish_wait()", or until 1318 * we see the WQ_FLAG_WOKEN flag, we need to 1319 * be very careful with the 'wait->flags', because 1320 * we may race with a waker that sets them. 1321 */ 1322 for (;;) { 1323 unsigned int flags; 1324 1325 set_current_state(state); 1326 1327 /* Loop until we've been woken or interrupted */ 1328 flags = smp_load_acquire(&wait->flags); 1329 if (!(flags & WQ_FLAG_WOKEN)) { 1330 if (signal_pending_state(state, current)) 1331 break; 1332 1333 io_schedule(); 1334 continue; 1335 } 1336 1337 /* If we were non-exclusive, we're done */ 1338 if (behavior != EXCLUSIVE) 1339 break; 1340 1341 /* If the waker got the lock for us, we're done */ 1342 if (flags & WQ_FLAG_DONE) 1343 break; 1344 1345 /* 1346 * Otherwise, if we're getting the lock, we need to 1347 * try to get it ourselves. 1348 * 1349 * And if that fails, we'll have to retry this all. 1350 */ 1351 if (unlikely(test_and_set_bit(bit_nr, folio_flags(folio, 0)))) 1352 goto repeat; 1353 1354 wait->flags |= WQ_FLAG_DONE; 1355 break; 1356 } 1357 1358 /* 1359 * If a signal happened, this 'finish_wait()' may remove the last 1360 * waiter from the wait-queues, but the folio waiters bit will remain 1361 * set. That's ok. The next wakeup will take care of it, and trying 1362 * to do it here would be difficult and prone to races. 1363 */ 1364 finish_wait(q, wait); 1365 1366 if (thrashing) { 1367 if (delayacct) 1368 delayacct_thrashing_end(); 1369 psi_memstall_leave(&pflags); 1370 } 1371 1372 /* 1373 * NOTE! The wait->flags weren't stable until we've done the 1374 * 'finish_wait()', and we could have exited the loop above due 1375 * to a signal, and had a wakeup event happen after the signal 1376 * test but before the 'finish_wait()'. 1377 * 1378 * So only after the finish_wait() can we reliably determine 1379 * if we got woken up or not, so we can now figure out the final 1380 * return value based on that state without races. 1381 * 1382 * Also note that WQ_FLAG_WOKEN is sufficient for a non-exclusive 1383 * waiter, but an exclusive one requires WQ_FLAG_DONE. 1384 */ 1385 if (behavior == EXCLUSIVE) 1386 return wait->flags & WQ_FLAG_DONE ? 0 : -EINTR; 1387 1388 return wait->flags & WQ_FLAG_WOKEN ? 0 : -EINTR; 1389 } 1390 1391 void folio_wait_bit(struct folio *folio, int bit_nr) 1392 { 1393 folio_wait_bit_common(folio, bit_nr, TASK_UNINTERRUPTIBLE, SHARED); 1394 } 1395 EXPORT_SYMBOL(folio_wait_bit); 1396 1397 int folio_wait_bit_killable(struct folio *folio, int bit_nr) 1398 { 1399 return folio_wait_bit_common(folio, bit_nr, TASK_KILLABLE, SHARED); 1400 } 1401 EXPORT_SYMBOL(folio_wait_bit_killable); 1402 1403 /** 1404 * folio_put_wait_locked - Drop a reference and wait for it to be unlocked 1405 * @folio: The folio to wait for. 1406 * @state: The sleep state (TASK_KILLABLE, TASK_UNINTERRUPTIBLE, etc). 1407 * 1408 * The caller should hold a reference on @folio. They expect the page to 1409 * become unlocked relatively soon, but do not wish to hold up migration 1410 * (for example) by holding the reference while waiting for the folio to 1411 * come unlocked. After this function returns, the caller should not 1412 * dereference @folio. 1413 * 1414 * Return: 0 if the folio was unlocked or -EINTR if interrupted by a signal. 1415 */ 1416 int folio_put_wait_locked(struct folio *folio, int state) 1417 { 1418 return folio_wait_bit_common(folio, PG_locked, state, DROP); 1419 } 1420 1421 /** 1422 * folio_add_wait_queue - Add an arbitrary waiter to a folio's wait queue 1423 * @folio: Folio defining the wait queue of interest 1424 * @waiter: Waiter to add to the queue 1425 * 1426 * Add an arbitrary @waiter to the wait queue for the nominated @folio. 1427 */ 1428 void folio_add_wait_queue(struct folio *folio, wait_queue_entry_t *waiter) 1429 { 1430 wait_queue_head_t *q = folio_waitqueue(folio); 1431 unsigned long flags; 1432 1433 spin_lock_irqsave(&q->lock, flags); 1434 __add_wait_queue_entry_tail(q, waiter); 1435 folio_set_waiters(folio); 1436 spin_unlock_irqrestore(&q->lock, flags); 1437 } 1438 EXPORT_SYMBOL_GPL(folio_add_wait_queue); 1439 1440 #ifndef clear_bit_unlock_is_negative_byte 1441 1442 /* 1443 * PG_waiters is the high bit in the same byte as PG_lock. 1444 * 1445 * On x86 (and on many other architectures), we can clear PG_lock and 1446 * test the sign bit at the same time. But if the architecture does 1447 * not support that special operation, we just do this all by hand 1448 * instead. 1449 * 1450 * The read of PG_waiters has to be after (or concurrently with) PG_locked 1451 * being cleared, but a memory barrier should be unnecessary since it is 1452 * in the same byte as PG_locked. 1453 */ 1454 static inline bool clear_bit_unlock_is_negative_byte(long nr, volatile void *mem) 1455 { 1456 clear_bit_unlock(nr, mem); 1457 /* smp_mb__after_atomic(); */ 1458 return test_bit(PG_waiters, mem); 1459 } 1460 1461 #endif 1462 1463 /** 1464 * folio_unlock - Unlock a locked folio. 1465 * @folio: The folio. 1466 * 1467 * Unlocks the folio and wakes up any thread sleeping on the page lock. 1468 * 1469 * Context: May be called from interrupt or process context. May not be 1470 * called from NMI context. 1471 */ 1472 void folio_unlock(struct folio *folio) 1473 { 1474 /* Bit 7 allows x86 to check the byte's sign bit */ 1475 BUILD_BUG_ON(PG_waiters != 7); 1476 BUILD_BUG_ON(PG_locked > 7); 1477 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); 1478 if (clear_bit_unlock_is_negative_byte(PG_locked, folio_flags(folio, 0))) 1479 folio_wake_bit(folio, PG_locked); 1480 } 1481 EXPORT_SYMBOL(folio_unlock); 1482 1483 /** 1484 * folio_end_private_2 - Clear PG_private_2 and wake any waiters. 1485 * @folio: The folio. 1486 * 1487 * Clear the PG_private_2 bit on a folio and wake up any sleepers waiting for 1488 * it. The folio reference held for PG_private_2 being set is released. 1489 * 1490 * This is, for example, used when a netfs folio is being written to a local 1491 * disk cache, thereby allowing writes to the cache for the same folio to be 1492 * serialised. 1493 */ 1494 void folio_end_private_2(struct folio *folio) 1495 { 1496 VM_BUG_ON_FOLIO(!folio_test_private_2(folio), folio); 1497 clear_bit_unlock(PG_private_2, folio_flags(folio, 0)); 1498 folio_wake_bit(folio, PG_private_2); 1499 folio_put(folio); 1500 } 1501 EXPORT_SYMBOL(folio_end_private_2); 1502 1503 /** 1504 * folio_wait_private_2 - Wait for PG_private_2 to be cleared on a folio. 1505 * @folio: The folio to wait on. 1506 * 1507 * Wait for PG_private_2 (aka PG_fscache) to be cleared on a folio. 1508 */ 1509 void folio_wait_private_2(struct folio *folio) 1510 { 1511 while (folio_test_private_2(folio)) 1512 folio_wait_bit(folio, PG_private_2); 1513 } 1514 EXPORT_SYMBOL(folio_wait_private_2); 1515 1516 /** 1517 * folio_wait_private_2_killable - Wait for PG_private_2 to be cleared on a folio. 1518 * @folio: The folio to wait on. 1519 * 1520 * Wait for PG_private_2 (aka PG_fscache) to be cleared on a folio or until a 1521 * fatal signal is received by the calling task. 1522 * 1523 * Return: 1524 * - 0 if successful. 1525 * - -EINTR if a fatal signal was encountered. 1526 */ 1527 int folio_wait_private_2_killable(struct folio *folio) 1528 { 1529 int ret = 0; 1530 1531 while (folio_test_private_2(folio)) { 1532 ret = folio_wait_bit_killable(folio, PG_private_2); 1533 if (ret < 0) 1534 break; 1535 } 1536 1537 return ret; 1538 } 1539 EXPORT_SYMBOL(folio_wait_private_2_killable); 1540 1541 /** 1542 * folio_end_writeback - End writeback against a folio. 1543 * @folio: The folio. 1544 */ 1545 void folio_end_writeback(struct folio *folio) 1546 { 1547 /* 1548 * folio_test_clear_reclaim() could be used here but it is an 1549 * atomic operation and overkill in this particular case. Failing 1550 * to shuffle a folio marked for immediate reclaim is too mild 1551 * a gain to justify taking an atomic operation penalty at the 1552 * end of every folio writeback. 1553 */ 1554 if (folio_test_reclaim(folio)) { 1555 folio_clear_reclaim(folio); 1556 folio_rotate_reclaimable(folio); 1557 } 1558 1559 /* 1560 * Writeback does not hold a folio reference of its own, relying 1561 * on truncation to wait for the clearing of PG_writeback. 1562 * But here we must make sure that the folio is not freed and 1563 * reused before the folio_wake(). 1564 */ 1565 folio_get(folio); 1566 if (!__folio_end_writeback(folio)) 1567 BUG(); 1568 1569 smp_mb__after_atomic(); 1570 folio_wake(folio, PG_writeback); 1571 acct_reclaim_writeback(folio); 1572 folio_put(folio); 1573 } 1574 EXPORT_SYMBOL(folio_end_writeback); 1575 1576 /* 1577 * After completing I/O on a page, call this routine to update the page 1578 * flags appropriately 1579 */ 1580 void page_endio(struct page *page, bool is_write, int err) 1581 { 1582 if (!is_write) { 1583 if (!err) { 1584 SetPageUptodate(page); 1585 } else { 1586 ClearPageUptodate(page); 1587 SetPageError(page); 1588 } 1589 unlock_page(page); 1590 } else { 1591 if (err) { 1592 struct address_space *mapping; 1593 1594 SetPageError(page); 1595 mapping = page_mapping(page); 1596 if (mapping) 1597 mapping_set_error(mapping, err); 1598 } 1599 end_page_writeback(page); 1600 } 1601 } 1602 EXPORT_SYMBOL_GPL(page_endio); 1603 1604 /** 1605 * __folio_lock - Get a lock on the folio, assuming we need to sleep to get it. 1606 * @folio: The folio to lock 1607 */ 1608 void __folio_lock(struct folio *folio) 1609 { 1610 folio_wait_bit_common(folio, PG_locked, TASK_UNINTERRUPTIBLE, 1611 EXCLUSIVE); 1612 } 1613 EXPORT_SYMBOL(__folio_lock); 1614 1615 int __folio_lock_killable(struct folio *folio) 1616 { 1617 return folio_wait_bit_common(folio, PG_locked, TASK_KILLABLE, 1618 EXCLUSIVE); 1619 } 1620 EXPORT_SYMBOL_GPL(__folio_lock_killable); 1621 1622 static int __folio_lock_async(struct folio *folio, struct wait_page_queue *wait) 1623 { 1624 struct wait_queue_head *q = folio_waitqueue(folio); 1625 int ret = 0; 1626 1627 wait->folio = folio; 1628 wait->bit_nr = PG_locked; 1629 1630 spin_lock_irq(&q->lock); 1631 __add_wait_queue_entry_tail(q, &wait->wait); 1632 folio_set_waiters(folio); 1633 ret = !folio_trylock(folio); 1634 /* 1635 * If we were successful now, we know we're still on the 1636 * waitqueue as we're still under the lock. This means it's 1637 * safe to remove and return success, we know the callback 1638 * isn't going to trigger. 1639 */ 1640 if (!ret) 1641 __remove_wait_queue(q, &wait->wait); 1642 else 1643 ret = -EIOCBQUEUED; 1644 spin_unlock_irq(&q->lock); 1645 return ret; 1646 } 1647 1648 /* 1649 * Return values: 1650 * true - folio is locked; mmap_lock is still held. 1651 * false - folio is not locked. 1652 * mmap_lock has been released (mmap_read_unlock(), unless flags had both 1653 * FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in 1654 * which case mmap_lock is still held. 1655 * 1656 * If neither ALLOW_RETRY nor KILLABLE are set, will always return true 1657 * with the folio locked and the mmap_lock unperturbed. 1658 */ 1659 bool __folio_lock_or_retry(struct folio *folio, struct mm_struct *mm, 1660 unsigned int flags) 1661 { 1662 if (fault_flag_allow_retry_first(flags)) { 1663 /* 1664 * CAUTION! In this case, mmap_lock is not released 1665 * even though return 0. 1666 */ 1667 if (flags & FAULT_FLAG_RETRY_NOWAIT) 1668 return false; 1669 1670 mmap_read_unlock(mm); 1671 if (flags & FAULT_FLAG_KILLABLE) 1672 folio_wait_locked_killable(folio); 1673 else 1674 folio_wait_locked(folio); 1675 return false; 1676 } 1677 if (flags & FAULT_FLAG_KILLABLE) { 1678 bool ret; 1679 1680 ret = __folio_lock_killable(folio); 1681 if (ret) { 1682 mmap_read_unlock(mm); 1683 return false; 1684 } 1685 } else { 1686 __folio_lock(folio); 1687 } 1688 1689 return true; 1690 } 1691 1692 /** 1693 * page_cache_next_miss() - Find the next gap in the page cache. 1694 * @mapping: Mapping. 1695 * @index: Index. 1696 * @max_scan: Maximum range to search. 1697 * 1698 * Search the range [index, min(index + max_scan - 1, ULONG_MAX)] for the 1699 * gap with the lowest index. 1700 * 1701 * This function may be called under the rcu_read_lock. However, this will 1702 * not atomically search a snapshot of the cache at a single point in time. 1703 * For example, if a gap is created at index 5, then subsequently a gap is 1704 * created at index 10, page_cache_next_miss covering both indices may 1705 * return 10 if called under the rcu_read_lock. 1706 * 1707 * Return: The index of the gap if found, otherwise an index outside the 1708 * range specified (in which case 'return - index >= max_scan' will be true). 1709 * In the rare case of index wrap-around, 0 will be returned. 1710 */ 1711 pgoff_t page_cache_next_miss(struct address_space *mapping, 1712 pgoff_t index, unsigned long max_scan) 1713 { 1714 XA_STATE(xas, &mapping->i_pages, index); 1715 1716 while (max_scan--) { 1717 void *entry = xas_next(&xas); 1718 if (!entry || xa_is_value(entry)) 1719 break; 1720 if (xas.xa_index == 0) 1721 break; 1722 } 1723 1724 return xas.xa_index; 1725 } 1726 EXPORT_SYMBOL(page_cache_next_miss); 1727 1728 /** 1729 * page_cache_prev_miss() - Find the previous gap in the page cache. 1730 * @mapping: Mapping. 1731 * @index: Index. 1732 * @max_scan: Maximum range to search. 1733 * 1734 * Search the range [max(index - max_scan + 1, 0), index] for the 1735 * gap with the highest index. 1736 * 1737 * This function may be called under the rcu_read_lock. However, this will 1738 * not atomically search a snapshot of the cache at a single point in time. 1739 * For example, if a gap is created at index 10, then subsequently a gap is 1740 * created at index 5, page_cache_prev_miss() covering both indices may 1741 * return 5 if called under the rcu_read_lock. 1742 * 1743 * Return: The index of the gap if found, otherwise an index outside the 1744 * range specified (in which case 'index - return >= max_scan' will be true). 1745 * In the rare case of wrap-around, ULONG_MAX will be returned. 1746 */ 1747 pgoff_t page_cache_prev_miss(struct address_space *mapping, 1748 pgoff_t index, unsigned long max_scan) 1749 { 1750 XA_STATE(xas, &mapping->i_pages, index); 1751 1752 while (max_scan--) { 1753 void *entry = xas_prev(&xas); 1754 if (!entry || xa_is_value(entry)) 1755 break; 1756 if (xas.xa_index == ULONG_MAX) 1757 break; 1758 } 1759 1760 return xas.xa_index; 1761 } 1762 EXPORT_SYMBOL(page_cache_prev_miss); 1763 1764 /* 1765 * Lockless page cache protocol: 1766 * On the lookup side: 1767 * 1. Load the folio from i_pages 1768 * 2. Increment the refcount if it's not zero 1769 * 3. If the folio is not found by xas_reload(), put the refcount and retry 1770 * 1771 * On the removal side: 1772 * A. Freeze the page (by zeroing the refcount if nobody else has a reference) 1773 * B. Remove the page from i_pages 1774 * C. Return the page to the page allocator 1775 * 1776 * This means that any page may have its reference count temporarily 1777 * increased by a speculative page cache (or fast GUP) lookup as it can 1778 * be allocated by another user before the RCU grace period expires. 1779 * Because the refcount temporarily acquired here may end up being the 1780 * last refcount on the page, any page allocation must be freeable by 1781 * folio_put(). 1782 */ 1783 1784 /* 1785 * mapping_get_entry - Get a page cache entry. 1786 * @mapping: the address_space to search 1787 * @index: The page cache index. 1788 * 1789 * Looks up the page cache entry at @mapping & @index. If it is a folio, 1790 * it is returned with an increased refcount. If it is a shadow entry 1791 * of a previously evicted folio, or a swap entry from shmem/tmpfs, 1792 * it is returned without further action. 1793 * 1794 * Return: The folio, swap or shadow entry, %NULL if nothing is found. 1795 */ 1796 static void *mapping_get_entry(struct address_space *mapping, pgoff_t index) 1797 { 1798 XA_STATE(xas, &mapping->i_pages, index); 1799 struct folio *folio; 1800 1801 rcu_read_lock(); 1802 repeat: 1803 xas_reset(&xas); 1804 folio = xas_load(&xas); 1805 if (xas_retry(&xas, folio)) 1806 goto repeat; 1807 /* 1808 * A shadow entry of a recently evicted page, or a swap entry from 1809 * shmem/tmpfs. Return it without attempting to raise page count. 1810 */ 1811 if (!folio || xa_is_value(folio)) 1812 goto out; 1813 1814 if (!folio_try_get_rcu(folio)) 1815 goto repeat; 1816 1817 if (unlikely(folio != xas_reload(&xas))) { 1818 folio_put(folio); 1819 goto repeat; 1820 } 1821 out: 1822 rcu_read_unlock(); 1823 1824 return folio; 1825 } 1826 1827 /** 1828 * __filemap_get_folio - Find and get a reference to a folio. 1829 * @mapping: The address_space to search. 1830 * @index: The page index. 1831 * @fgp_flags: %FGP flags modify how the folio is returned. 1832 * @gfp: Memory allocation flags to use if %FGP_CREAT is specified. 1833 * 1834 * Looks up the page cache entry at @mapping & @index. 1835 * 1836 * @fgp_flags can be zero or more of these flags: 1837 * 1838 * * %FGP_ACCESSED - The folio will be marked accessed. 1839 * * %FGP_LOCK - The folio is returned locked. 1840 * * %FGP_ENTRY - If there is a shadow / swap / DAX entry, return it 1841 * instead of allocating a new folio to replace it. 1842 * * %FGP_CREAT - If no page is present then a new page is allocated using 1843 * @gfp and added to the page cache and the VM's LRU list. 1844 * The page is returned locked and with an increased refcount. 1845 * * %FGP_FOR_MMAP - The caller wants to do its own locking dance if the 1846 * page is already in cache. If the page was allocated, unlock it before 1847 * returning so the caller can do the same dance. 1848 * * %FGP_WRITE - The page will be written to by the caller. 1849 * * %FGP_NOFS - __GFP_FS will get cleared in gfp. 1850 * * %FGP_NOWAIT - Don't get blocked by page lock. 1851 * * %FGP_STABLE - Wait for the folio to be stable (finished writeback) 1852 * 1853 * If %FGP_LOCK or %FGP_CREAT are specified then the function may sleep even 1854 * if the %GFP flags specified for %FGP_CREAT are atomic. 1855 * 1856 * If there is a page cache page, it is returned with an increased refcount. 1857 * 1858 * Return: The found folio or %NULL otherwise. 1859 */ 1860 struct folio *__filemap_get_folio(struct address_space *mapping, pgoff_t index, 1861 int fgp_flags, gfp_t gfp) 1862 { 1863 struct folio *folio; 1864 1865 repeat: 1866 folio = mapping_get_entry(mapping, index); 1867 if (xa_is_value(folio)) { 1868 if (fgp_flags & FGP_ENTRY) 1869 return folio; 1870 folio = NULL; 1871 } 1872 if (!folio) 1873 goto no_page; 1874 1875 if (fgp_flags & FGP_LOCK) { 1876 if (fgp_flags & FGP_NOWAIT) { 1877 if (!folio_trylock(folio)) { 1878 folio_put(folio); 1879 return NULL; 1880 } 1881 } else { 1882 folio_lock(folio); 1883 } 1884 1885 /* Has the page been truncated? */ 1886 if (unlikely(folio->mapping != mapping)) { 1887 folio_unlock(folio); 1888 folio_put(folio); 1889 goto repeat; 1890 } 1891 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio); 1892 } 1893 1894 if (fgp_flags & FGP_ACCESSED) 1895 folio_mark_accessed(folio); 1896 else if (fgp_flags & FGP_WRITE) { 1897 /* Clear idle flag for buffer write */ 1898 if (folio_test_idle(folio)) 1899 folio_clear_idle(folio); 1900 } 1901 1902 if (fgp_flags & FGP_STABLE) 1903 folio_wait_stable(folio); 1904 no_page: 1905 if (!folio && (fgp_flags & FGP_CREAT)) { 1906 int err; 1907 if ((fgp_flags & FGP_WRITE) && mapping_can_writeback(mapping)) 1908 gfp |= __GFP_WRITE; 1909 if (fgp_flags & FGP_NOFS) 1910 gfp &= ~__GFP_FS; 1911 1912 folio = filemap_alloc_folio(gfp, 0); 1913 if (!folio) 1914 return NULL; 1915 1916 if (WARN_ON_ONCE(!(fgp_flags & (FGP_LOCK | FGP_FOR_MMAP)))) 1917 fgp_flags |= FGP_LOCK; 1918 1919 /* Init accessed so avoid atomic mark_page_accessed later */ 1920 if (fgp_flags & FGP_ACCESSED) 1921 __folio_set_referenced(folio); 1922 1923 err = filemap_add_folio(mapping, folio, index, gfp); 1924 if (unlikely(err)) { 1925 folio_put(folio); 1926 folio = NULL; 1927 if (err == -EEXIST) 1928 goto repeat; 1929 } 1930 1931 /* 1932 * filemap_add_folio locks the page, and for mmap 1933 * we expect an unlocked page. 1934 */ 1935 if (folio && (fgp_flags & FGP_FOR_MMAP)) 1936 folio_unlock(folio); 1937 } 1938 1939 return folio; 1940 } 1941 EXPORT_SYMBOL(__filemap_get_folio); 1942 1943 static inline struct folio *find_get_entry(struct xa_state *xas, pgoff_t max, 1944 xa_mark_t mark) 1945 { 1946 struct folio *folio; 1947 1948 retry: 1949 if (mark == XA_PRESENT) 1950 folio = xas_find(xas, max); 1951 else 1952 folio = xas_find_marked(xas, max, mark); 1953 1954 if (xas_retry(xas, folio)) 1955 goto retry; 1956 /* 1957 * A shadow entry of a recently evicted page, a swap 1958 * entry from shmem/tmpfs or a DAX entry. Return it 1959 * without attempting to raise page count. 1960 */ 1961 if (!folio || xa_is_value(folio)) 1962 return folio; 1963 1964 if (!folio_try_get_rcu(folio)) 1965 goto reset; 1966 1967 if (unlikely(folio != xas_reload(xas))) { 1968 folio_put(folio); 1969 goto reset; 1970 } 1971 1972 return folio; 1973 reset: 1974 xas_reset(xas); 1975 goto retry; 1976 } 1977 1978 /** 1979 * find_get_entries - gang pagecache lookup 1980 * @mapping: The address_space to search 1981 * @start: The starting page cache index 1982 * @end: The final page index (inclusive). 1983 * @fbatch: Where the resulting entries are placed. 1984 * @indices: The cache indices corresponding to the entries in @entries 1985 * 1986 * find_get_entries() will search for and return a batch of entries in 1987 * the mapping. The entries are placed in @fbatch. find_get_entries() 1988 * takes a reference on any actual folios it returns. 1989 * 1990 * The entries have ascending indexes. The indices may not be consecutive 1991 * due to not-present entries or large folios. 1992 * 1993 * Any shadow entries of evicted folios, or swap entries from 1994 * shmem/tmpfs, are included in the returned array. 1995 * 1996 * Return: The number of entries which were found. 1997 */ 1998 unsigned find_get_entries(struct address_space *mapping, pgoff_t start, 1999 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices) 2000 { 2001 XA_STATE(xas, &mapping->i_pages, start); 2002 struct folio *folio; 2003 2004 rcu_read_lock(); 2005 while ((folio = find_get_entry(&xas, end, XA_PRESENT)) != NULL) { 2006 indices[fbatch->nr] = xas.xa_index; 2007 if (!folio_batch_add(fbatch, folio)) 2008 break; 2009 } 2010 rcu_read_unlock(); 2011 2012 return folio_batch_count(fbatch); 2013 } 2014 2015 /** 2016 * find_lock_entries - Find a batch of pagecache entries. 2017 * @mapping: The address_space to search. 2018 * @start: The starting page cache index. 2019 * @end: The final page index (inclusive). 2020 * @fbatch: Where the resulting entries are placed. 2021 * @indices: The cache indices of the entries in @fbatch. 2022 * 2023 * find_lock_entries() will return a batch of entries from @mapping. 2024 * Swap, shadow and DAX entries are included. Folios are returned 2025 * locked and with an incremented refcount. Folios which are locked 2026 * by somebody else or under writeback are skipped. Folios which are 2027 * partially outside the range are not returned. 2028 * 2029 * The entries have ascending indexes. The indices may not be consecutive 2030 * due to not-present entries, large folios, folios which could not be 2031 * locked or folios under writeback. 2032 * 2033 * Return: The number of entries which were found. 2034 */ 2035 unsigned find_lock_entries(struct address_space *mapping, pgoff_t start, 2036 pgoff_t end, struct folio_batch *fbatch, pgoff_t *indices) 2037 { 2038 XA_STATE(xas, &mapping->i_pages, start); 2039 struct folio *folio; 2040 2041 rcu_read_lock(); 2042 while ((folio = find_get_entry(&xas, end, XA_PRESENT))) { 2043 if (!xa_is_value(folio)) { 2044 if (folio->index < start) 2045 goto put; 2046 if (folio->index + folio_nr_pages(folio) - 1 > end) 2047 goto put; 2048 if (!folio_trylock(folio)) 2049 goto put; 2050 if (folio->mapping != mapping || 2051 folio_test_writeback(folio)) 2052 goto unlock; 2053 VM_BUG_ON_FOLIO(!folio_contains(folio, xas.xa_index), 2054 folio); 2055 } 2056 indices[fbatch->nr] = xas.xa_index; 2057 if (!folio_batch_add(fbatch, folio)) 2058 break; 2059 continue; 2060 unlock: 2061 folio_unlock(folio); 2062 put: 2063 folio_put(folio); 2064 } 2065 rcu_read_unlock(); 2066 2067 return folio_batch_count(fbatch); 2068 } 2069 2070 static inline 2071 bool folio_more_pages(struct folio *folio, pgoff_t index, pgoff_t max) 2072 { 2073 if (!folio_test_large(folio) || folio_test_hugetlb(folio)) 2074 return false; 2075 if (index >= max) 2076 return false; 2077 return index < folio->index + folio_nr_pages(folio) - 1; 2078 } 2079 2080 /** 2081 * find_get_pages_range - gang pagecache lookup 2082 * @mapping: The address_space to search 2083 * @start: The starting page index 2084 * @end: The final page index (inclusive) 2085 * @nr_pages: The maximum number of pages 2086 * @pages: Where the resulting pages are placed 2087 * 2088 * find_get_pages_range() will search for and return a group of up to @nr_pages 2089 * pages in the mapping starting at index @start and up to index @end 2090 * (inclusive). The pages are placed at @pages. find_get_pages_range() takes 2091 * a reference against the returned pages. 2092 * 2093 * The search returns a group of mapping-contiguous pages with ascending 2094 * indexes. There may be holes in the indices due to not-present pages. 2095 * We also update @start to index the next page for the traversal. 2096 * 2097 * Return: the number of pages which were found. If this number is 2098 * smaller than @nr_pages, the end of specified range has been 2099 * reached. 2100 */ 2101 unsigned find_get_pages_range(struct address_space *mapping, pgoff_t *start, 2102 pgoff_t end, unsigned int nr_pages, 2103 struct page **pages) 2104 { 2105 XA_STATE(xas, &mapping->i_pages, *start); 2106 struct folio *folio; 2107 unsigned ret = 0; 2108 2109 if (unlikely(!nr_pages)) 2110 return 0; 2111 2112 rcu_read_lock(); 2113 while ((folio = find_get_entry(&xas, end, XA_PRESENT))) { 2114 /* Skip over shadow, swap and DAX entries */ 2115 if (xa_is_value(folio)) 2116 continue; 2117 2118 again: 2119 pages[ret] = folio_file_page(folio, xas.xa_index); 2120 if (++ret == nr_pages) { 2121 *start = xas.xa_index + 1; 2122 goto out; 2123 } 2124 if (folio_more_pages(folio, xas.xa_index, end)) { 2125 xas.xa_index++; 2126 folio_ref_inc(folio); 2127 goto again; 2128 } 2129 } 2130 2131 /* 2132 * We come here when there is no page beyond @end. We take care to not 2133 * overflow the index @start as it confuses some of the callers. This 2134 * breaks the iteration when there is a page at index -1 but that is 2135 * already broken anyway. 2136 */ 2137 if (end == (pgoff_t)-1) 2138 *start = (pgoff_t)-1; 2139 else 2140 *start = end + 1; 2141 out: 2142 rcu_read_unlock(); 2143 2144 return ret; 2145 } 2146 2147 /** 2148 * find_get_pages_contig - gang contiguous pagecache lookup 2149 * @mapping: The address_space to search 2150 * @index: The starting page index 2151 * @nr_pages: The maximum number of pages 2152 * @pages: Where the resulting pages are placed 2153 * 2154 * find_get_pages_contig() works exactly like find_get_pages(), except 2155 * that the returned number of pages are guaranteed to be contiguous. 2156 * 2157 * Return: the number of pages which were found. 2158 */ 2159 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index, 2160 unsigned int nr_pages, struct page **pages) 2161 { 2162 XA_STATE(xas, &mapping->i_pages, index); 2163 struct folio *folio; 2164 unsigned int ret = 0; 2165 2166 if (unlikely(!nr_pages)) 2167 return 0; 2168 2169 rcu_read_lock(); 2170 for (folio = xas_load(&xas); folio; folio = xas_next(&xas)) { 2171 if (xas_retry(&xas, folio)) 2172 continue; 2173 /* 2174 * If the entry has been swapped out, we can stop looking. 2175 * No current caller is looking for DAX entries. 2176 */ 2177 if (xa_is_value(folio)) 2178 break; 2179 2180 if (!folio_try_get_rcu(folio)) 2181 goto retry; 2182 2183 if (unlikely(folio != xas_reload(&xas))) 2184 goto put_page; 2185 2186 again: 2187 pages[ret] = folio_file_page(folio, xas.xa_index); 2188 if (++ret == nr_pages) 2189 break; 2190 if (folio_more_pages(folio, xas.xa_index, ULONG_MAX)) { 2191 xas.xa_index++; 2192 folio_ref_inc(folio); 2193 goto again; 2194 } 2195 continue; 2196 put_page: 2197 folio_put(folio); 2198 retry: 2199 xas_reset(&xas); 2200 } 2201 rcu_read_unlock(); 2202 return ret; 2203 } 2204 EXPORT_SYMBOL(find_get_pages_contig); 2205 2206 /** 2207 * find_get_pages_range_tag - Find and return head pages matching @tag. 2208 * @mapping: the address_space to search 2209 * @index: the starting page index 2210 * @end: The final page index (inclusive) 2211 * @tag: the tag index 2212 * @nr_pages: the maximum number of pages 2213 * @pages: where the resulting pages are placed 2214 * 2215 * Like find_get_pages(), except we only return head pages which are tagged 2216 * with @tag. @index is updated to the index immediately after the last 2217 * page we return, ready for the next iteration. 2218 * 2219 * Return: the number of pages which were found. 2220 */ 2221 unsigned find_get_pages_range_tag(struct address_space *mapping, pgoff_t *index, 2222 pgoff_t end, xa_mark_t tag, unsigned int nr_pages, 2223 struct page **pages) 2224 { 2225 XA_STATE(xas, &mapping->i_pages, *index); 2226 struct folio *folio; 2227 unsigned ret = 0; 2228 2229 if (unlikely(!nr_pages)) 2230 return 0; 2231 2232 rcu_read_lock(); 2233 while ((folio = find_get_entry(&xas, end, tag))) { 2234 /* 2235 * Shadow entries should never be tagged, but this iteration 2236 * is lockless so there is a window for page reclaim to evict 2237 * a page we saw tagged. Skip over it. 2238 */ 2239 if (xa_is_value(folio)) 2240 continue; 2241 2242 pages[ret] = &folio->page; 2243 if (++ret == nr_pages) { 2244 *index = folio->index + folio_nr_pages(folio); 2245 goto out; 2246 } 2247 } 2248 2249 /* 2250 * We come here when we got to @end. We take care to not overflow the 2251 * index @index as it confuses some of the callers. This breaks the 2252 * iteration when there is a page at index -1 but that is already 2253 * broken anyway. 2254 */ 2255 if (end == (pgoff_t)-1) 2256 *index = (pgoff_t)-1; 2257 else 2258 *index = end + 1; 2259 out: 2260 rcu_read_unlock(); 2261 2262 return ret; 2263 } 2264 EXPORT_SYMBOL(find_get_pages_range_tag); 2265 2266 /* 2267 * CD/DVDs are error prone. When a medium error occurs, the driver may fail 2268 * a _large_ part of the i/o request. Imagine the worst scenario: 2269 * 2270 * ---R__________________________________________B__________ 2271 * ^ reading here ^ bad block(assume 4k) 2272 * 2273 * read(R) => miss => readahead(R...B) => media error => frustrating retries 2274 * => failing the whole request => read(R) => read(R+1) => 2275 * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) => 2276 * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) => 2277 * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ...... 2278 * 2279 * It is going insane. Fix it by quickly scaling down the readahead size. 2280 */ 2281 static void shrink_readahead_size_eio(struct file_ra_state *ra) 2282 { 2283 ra->ra_pages /= 4; 2284 } 2285 2286 /* 2287 * filemap_get_read_batch - Get a batch of folios for read 2288 * 2289 * Get a batch of folios which represent a contiguous range of bytes in 2290 * the file. No exceptional entries will be returned. If @index is in 2291 * the middle of a folio, the entire folio will be returned. The last 2292 * folio in the batch may have the readahead flag set or the uptodate flag 2293 * clear so that the caller can take the appropriate action. 2294 */ 2295 static void filemap_get_read_batch(struct address_space *mapping, 2296 pgoff_t index, pgoff_t max, struct folio_batch *fbatch) 2297 { 2298 XA_STATE(xas, &mapping->i_pages, index); 2299 struct folio *folio; 2300 2301 rcu_read_lock(); 2302 for (folio = xas_load(&xas); folio; folio = xas_next(&xas)) { 2303 if (xas_retry(&xas, folio)) 2304 continue; 2305 if (xas.xa_index > max || xa_is_value(folio)) 2306 break; 2307 if (!folio_try_get_rcu(folio)) 2308 goto retry; 2309 2310 if (unlikely(folio != xas_reload(&xas))) 2311 goto put_folio; 2312 2313 if (!folio_batch_add(fbatch, folio)) 2314 break; 2315 if (!folio_test_uptodate(folio)) 2316 break; 2317 if (folio_test_readahead(folio)) 2318 break; 2319 xas_advance(&xas, folio->index + folio_nr_pages(folio) - 1); 2320 continue; 2321 put_folio: 2322 folio_put(folio); 2323 retry: 2324 xas_reset(&xas); 2325 } 2326 rcu_read_unlock(); 2327 } 2328 2329 static int filemap_read_folio(struct file *file, struct address_space *mapping, 2330 struct folio *folio) 2331 { 2332 int error; 2333 2334 /* 2335 * A previous I/O error may have been due to temporary failures, 2336 * eg. multipath errors. PG_error will be set again if readpage 2337 * fails. 2338 */ 2339 folio_clear_error(folio); 2340 /* Start the actual read. The read will unlock the page. */ 2341 error = mapping->a_ops->readpage(file, &folio->page); 2342 if (error) 2343 return error; 2344 2345 error = folio_wait_locked_killable(folio); 2346 if (error) 2347 return error; 2348 if (folio_test_uptodate(folio)) 2349 return 0; 2350 shrink_readahead_size_eio(&file->f_ra); 2351 return -EIO; 2352 } 2353 2354 static bool filemap_range_uptodate(struct address_space *mapping, 2355 loff_t pos, struct iov_iter *iter, struct folio *folio) 2356 { 2357 int count; 2358 2359 if (folio_test_uptodate(folio)) 2360 return true; 2361 /* pipes can't handle partially uptodate pages */ 2362 if (iov_iter_is_pipe(iter)) 2363 return false; 2364 if (!mapping->a_ops->is_partially_uptodate) 2365 return false; 2366 if (mapping->host->i_blkbits >= folio_shift(folio)) 2367 return false; 2368 2369 count = iter->count; 2370 if (folio_pos(folio) > pos) { 2371 count -= folio_pos(folio) - pos; 2372 pos = 0; 2373 } else { 2374 pos -= folio_pos(folio); 2375 } 2376 2377 return mapping->a_ops->is_partially_uptodate(&folio->page, pos, count); 2378 } 2379 2380 static int filemap_update_page(struct kiocb *iocb, 2381 struct address_space *mapping, struct iov_iter *iter, 2382 struct folio *folio) 2383 { 2384 int error; 2385 2386 if (iocb->ki_flags & IOCB_NOWAIT) { 2387 if (!filemap_invalidate_trylock_shared(mapping)) 2388 return -EAGAIN; 2389 } else { 2390 filemap_invalidate_lock_shared(mapping); 2391 } 2392 2393 if (!folio_trylock(folio)) { 2394 error = -EAGAIN; 2395 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_NOIO)) 2396 goto unlock_mapping; 2397 if (!(iocb->ki_flags & IOCB_WAITQ)) { 2398 filemap_invalidate_unlock_shared(mapping); 2399 /* 2400 * This is where we usually end up waiting for a 2401 * previously submitted readahead to finish. 2402 */ 2403 folio_put_wait_locked(folio, TASK_KILLABLE); 2404 return AOP_TRUNCATED_PAGE; 2405 } 2406 error = __folio_lock_async(folio, iocb->ki_waitq); 2407 if (error) 2408 goto unlock_mapping; 2409 } 2410 2411 error = AOP_TRUNCATED_PAGE; 2412 if (!folio->mapping) 2413 goto unlock; 2414 2415 error = 0; 2416 if (filemap_range_uptodate(mapping, iocb->ki_pos, iter, folio)) 2417 goto unlock; 2418 2419 error = -EAGAIN; 2420 if (iocb->ki_flags & (IOCB_NOIO | IOCB_NOWAIT | IOCB_WAITQ)) 2421 goto unlock; 2422 2423 error = filemap_read_folio(iocb->ki_filp, mapping, folio); 2424 goto unlock_mapping; 2425 unlock: 2426 folio_unlock(folio); 2427 unlock_mapping: 2428 filemap_invalidate_unlock_shared(mapping); 2429 if (error == AOP_TRUNCATED_PAGE) 2430 folio_put(folio); 2431 return error; 2432 } 2433 2434 static int filemap_create_folio(struct file *file, 2435 struct address_space *mapping, pgoff_t index, 2436 struct folio_batch *fbatch) 2437 { 2438 struct folio *folio; 2439 int error; 2440 2441 folio = filemap_alloc_folio(mapping_gfp_mask(mapping), 0); 2442 if (!folio) 2443 return -ENOMEM; 2444 2445 /* 2446 * Protect against truncate / hole punch. Grabbing invalidate_lock 2447 * here assures we cannot instantiate and bring uptodate new 2448 * pagecache folios after evicting page cache during truncate 2449 * and before actually freeing blocks. Note that we could 2450 * release invalidate_lock after inserting the folio into 2451 * the page cache as the locked folio would then be enough to 2452 * synchronize with hole punching. But there are code paths 2453 * such as filemap_update_page() filling in partially uptodate 2454 * pages or ->readpages() that need to hold invalidate_lock 2455 * while mapping blocks for IO so let's hold the lock here as 2456 * well to keep locking rules simple. 2457 */ 2458 filemap_invalidate_lock_shared(mapping); 2459 error = filemap_add_folio(mapping, folio, index, 2460 mapping_gfp_constraint(mapping, GFP_KERNEL)); 2461 if (error == -EEXIST) 2462 error = AOP_TRUNCATED_PAGE; 2463 if (error) 2464 goto error; 2465 2466 error = filemap_read_folio(file, mapping, folio); 2467 if (error) 2468 goto error; 2469 2470 filemap_invalidate_unlock_shared(mapping); 2471 folio_batch_add(fbatch, folio); 2472 return 0; 2473 error: 2474 filemap_invalidate_unlock_shared(mapping); 2475 folio_put(folio); 2476 return error; 2477 } 2478 2479 static int filemap_readahead(struct kiocb *iocb, struct file *file, 2480 struct address_space *mapping, struct folio *folio, 2481 pgoff_t last_index) 2482 { 2483 DEFINE_READAHEAD(ractl, file, &file->f_ra, mapping, folio->index); 2484 2485 if (iocb->ki_flags & IOCB_NOIO) 2486 return -EAGAIN; 2487 page_cache_async_ra(&ractl, folio, last_index - folio->index); 2488 return 0; 2489 } 2490 2491 static int filemap_get_pages(struct kiocb *iocb, struct iov_iter *iter, 2492 struct folio_batch *fbatch) 2493 { 2494 struct file *filp = iocb->ki_filp; 2495 struct address_space *mapping = filp->f_mapping; 2496 struct file_ra_state *ra = &filp->f_ra; 2497 pgoff_t index = iocb->ki_pos >> PAGE_SHIFT; 2498 pgoff_t last_index; 2499 struct folio *folio; 2500 int err = 0; 2501 2502 last_index = DIV_ROUND_UP(iocb->ki_pos + iter->count, PAGE_SIZE); 2503 retry: 2504 if (fatal_signal_pending(current)) 2505 return -EINTR; 2506 2507 filemap_get_read_batch(mapping, index, last_index, fbatch); 2508 if (!folio_batch_count(fbatch)) { 2509 if (iocb->ki_flags & IOCB_NOIO) 2510 return -EAGAIN; 2511 page_cache_sync_readahead(mapping, ra, filp, index, 2512 last_index - index); 2513 filemap_get_read_batch(mapping, index, last_index, fbatch); 2514 } 2515 if (!folio_batch_count(fbatch)) { 2516 if (iocb->ki_flags & (IOCB_NOWAIT | IOCB_WAITQ)) 2517 return -EAGAIN; 2518 err = filemap_create_folio(filp, mapping, 2519 iocb->ki_pos >> PAGE_SHIFT, fbatch); 2520 if (err == AOP_TRUNCATED_PAGE) 2521 goto retry; 2522 return err; 2523 } 2524 2525 folio = fbatch->folios[folio_batch_count(fbatch) - 1]; 2526 if (folio_test_readahead(folio)) { 2527 err = filemap_readahead(iocb, filp, mapping, folio, last_index); 2528 if (err) 2529 goto err; 2530 } 2531 if (!folio_test_uptodate(folio)) { 2532 if ((iocb->ki_flags & IOCB_WAITQ) && 2533 folio_batch_count(fbatch) > 1) 2534 iocb->ki_flags |= IOCB_NOWAIT; 2535 err = filemap_update_page(iocb, mapping, iter, folio); 2536 if (err) 2537 goto err; 2538 } 2539 2540 return 0; 2541 err: 2542 if (err < 0) 2543 folio_put(folio); 2544 if (likely(--fbatch->nr)) 2545 return 0; 2546 if (err == AOP_TRUNCATED_PAGE) 2547 goto retry; 2548 return err; 2549 } 2550 2551 /** 2552 * filemap_read - Read data from the page cache. 2553 * @iocb: The iocb to read. 2554 * @iter: Destination for the data. 2555 * @already_read: Number of bytes already read by the caller. 2556 * 2557 * Copies data from the page cache. If the data is not currently present, 2558 * uses the readahead and readpage address_space operations to fetch it. 2559 * 2560 * Return: Total number of bytes copied, including those already read by 2561 * the caller. If an error happens before any bytes are copied, returns 2562 * a negative error number. 2563 */ 2564 ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter, 2565 ssize_t already_read) 2566 { 2567 struct file *filp = iocb->ki_filp; 2568 struct file_ra_state *ra = &filp->f_ra; 2569 struct address_space *mapping = filp->f_mapping; 2570 struct inode *inode = mapping->host; 2571 struct folio_batch fbatch; 2572 int i, error = 0; 2573 bool writably_mapped; 2574 loff_t isize, end_offset; 2575 2576 if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes)) 2577 return 0; 2578 if (unlikely(!iov_iter_count(iter))) 2579 return 0; 2580 2581 iov_iter_truncate(iter, inode->i_sb->s_maxbytes); 2582 folio_batch_init(&fbatch); 2583 2584 do { 2585 cond_resched(); 2586 2587 /* 2588 * If we've already successfully copied some data, then we 2589 * can no longer safely return -EIOCBQUEUED. Hence mark 2590 * an async read NOWAIT at that point. 2591 */ 2592 if ((iocb->ki_flags & IOCB_WAITQ) && already_read) 2593 iocb->ki_flags |= IOCB_NOWAIT; 2594 2595 if (unlikely(iocb->ki_pos >= i_size_read(inode))) 2596 break; 2597 2598 error = filemap_get_pages(iocb, iter, &fbatch); 2599 if (error < 0) 2600 break; 2601 2602 /* 2603 * i_size must be checked after we know the pages are Uptodate. 2604 * 2605 * Checking i_size after the check allows us to calculate 2606 * the correct value for "nr", which means the zero-filled 2607 * part of the page is not copied back to userspace (unless 2608 * another truncate extends the file - this is desired though). 2609 */ 2610 isize = i_size_read(inode); 2611 if (unlikely(iocb->ki_pos >= isize)) 2612 goto put_folios; 2613 end_offset = min_t(loff_t, isize, iocb->ki_pos + iter->count); 2614 2615 /* 2616 * Once we start copying data, we don't want to be touching any 2617 * cachelines that might be contended: 2618 */ 2619 writably_mapped = mapping_writably_mapped(mapping); 2620 2621 /* 2622 * When a sequential read accesses a page several times, only 2623 * mark it as accessed the first time. 2624 */ 2625 if (iocb->ki_pos >> PAGE_SHIFT != 2626 ra->prev_pos >> PAGE_SHIFT) 2627 folio_mark_accessed(fbatch.folios[0]); 2628 2629 for (i = 0; i < folio_batch_count(&fbatch); i++) { 2630 struct folio *folio = fbatch.folios[i]; 2631 size_t fsize = folio_size(folio); 2632 size_t offset = iocb->ki_pos & (fsize - 1); 2633 size_t bytes = min_t(loff_t, end_offset - iocb->ki_pos, 2634 fsize - offset); 2635 size_t copied; 2636 2637 if (end_offset < folio_pos(folio)) 2638 break; 2639 if (i > 0) 2640 folio_mark_accessed(folio); 2641 /* 2642 * If users can be writing to this folio using arbitrary 2643 * virtual addresses, take care of potential aliasing 2644 * before reading the folio on the kernel side. 2645 */ 2646 if (writably_mapped) 2647 flush_dcache_folio(folio); 2648 2649 copied = copy_folio_to_iter(folio, offset, bytes, iter); 2650 2651 already_read += copied; 2652 iocb->ki_pos += copied; 2653 ra->prev_pos = iocb->ki_pos; 2654 2655 if (copied < bytes) { 2656 error = -EFAULT; 2657 break; 2658 } 2659 } 2660 put_folios: 2661 for (i = 0; i < folio_batch_count(&fbatch); i++) 2662 folio_put(fbatch.folios[i]); 2663 folio_batch_init(&fbatch); 2664 } while (iov_iter_count(iter) && iocb->ki_pos < isize && !error); 2665 2666 file_accessed(filp); 2667 2668 return already_read ? already_read : error; 2669 } 2670 EXPORT_SYMBOL_GPL(filemap_read); 2671 2672 /** 2673 * generic_file_read_iter - generic filesystem read routine 2674 * @iocb: kernel I/O control block 2675 * @iter: destination for the data read 2676 * 2677 * This is the "read_iter()" routine for all filesystems 2678 * that can use the page cache directly. 2679 * 2680 * The IOCB_NOWAIT flag in iocb->ki_flags indicates that -EAGAIN shall 2681 * be returned when no data can be read without waiting for I/O requests 2682 * to complete; it doesn't prevent readahead. 2683 * 2684 * The IOCB_NOIO flag in iocb->ki_flags indicates that no new I/O 2685 * requests shall be made for the read or for readahead. When no data 2686 * can be read, -EAGAIN shall be returned. When readahead would be 2687 * triggered, a partial, possibly empty read shall be returned. 2688 * 2689 * Return: 2690 * * number of bytes copied, even for partial reads 2691 * * negative error code (or 0 if IOCB_NOIO) if nothing was read 2692 */ 2693 ssize_t 2694 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter) 2695 { 2696 size_t count = iov_iter_count(iter); 2697 ssize_t retval = 0; 2698 2699 if (!count) 2700 return 0; /* skip atime */ 2701 2702 if (iocb->ki_flags & IOCB_DIRECT) { 2703 struct file *file = iocb->ki_filp; 2704 struct address_space *mapping = file->f_mapping; 2705 struct inode *inode = mapping->host; 2706 2707 if (iocb->ki_flags & IOCB_NOWAIT) { 2708 if (filemap_range_needs_writeback(mapping, iocb->ki_pos, 2709 iocb->ki_pos + count - 1)) 2710 return -EAGAIN; 2711 } else { 2712 retval = filemap_write_and_wait_range(mapping, 2713 iocb->ki_pos, 2714 iocb->ki_pos + count - 1); 2715 if (retval < 0) 2716 return retval; 2717 } 2718 2719 file_accessed(file); 2720 2721 retval = mapping->a_ops->direct_IO(iocb, iter); 2722 if (retval >= 0) { 2723 iocb->ki_pos += retval; 2724 count -= retval; 2725 } 2726 if (retval != -EIOCBQUEUED) 2727 iov_iter_revert(iter, count - iov_iter_count(iter)); 2728 2729 /* 2730 * Btrfs can have a short DIO read if we encounter 2731 * compressed extents, so if there was an error, or if 2732 * we've already read everything we wanted to, or if 2733 * there was a short read because we hit EOF, go ahead 2734 * and return. Otherwise fallthrough to buffered io for 2735 * the rest of the read. Buffered reads will not work for 2736 * DAX files, so don't bother trying. 2737 */ 2738 if (retval < 0 || !count || IS_DAX(inode)) 2739 return retval; 2740 if (iocb->ki_pos >= i_size_read(inode)) 2741 return retval; 2742 } 2743 2744 return filemap_read(iocb, iter, retval); 2745 } 2746 EXPORT_SYMBOL(generic_file_read_iter); 2747 2748 static inline loff_t folio_seek_hole_data(struct xa_state *xas, 2749 struct address_space *mapping, struct folio *folio, 2750 loff_t start, loff_t end, bool seek_data) 2751 { 2752 const struct address_space_operations *ops = mapping->a_ops; 2753 size_t offset, bsz = i_blocksize(mapping->host); 2754 2755 if (xa_is_value(folio) || folio_test_uptodate(folio)) 2756 return seek_data ? start : end; 2757 if (!ops->is_partially_uptodate) 2758 return seek_data ? end : start; 2759 2760 xas_pause(xas); 2761 rcu_read_unlock(); 2762 folio_lock(folio); 2763 if (unlikely(folio->mapping != mapping)) 2764 goto unlock; 2765 2766 offset = offset_in_folio(folio, start) & ~(bsz - 1); 2767 2768 do { 2769 if (ops->is_partially_uptodate(&folio->page, offset, bsz) == 2770 seek_data) 2771 break; 2772 start = (start + bsz) & ~(bsz - 1); 2773 offset += bsz; 2774 } while (offset < folio_size(folio)); 2775 unlock: 2776 folio_unlock(folio); 2777 rcu_read_lock(); 2778 return start; 2779 } 2780 2781 static inline size_t seek_folio_size(struct xa_state *xas, struct folio *folio) 2782 { 2783 if (xa_is_value(folio)) 2784 return PAGE_SIZE << xa_get_order(xas->xa, xas->xa_index); 2785 return folio_size(folio); 2786 } 2787 2788 /** 2789 * mapping_seek_hole_data - Seek for SEEK_DATA / SEEK_HOLE in the page cache. 2790 * @mapping: Address space to search. 2791 * @start: First byte to consider. 2792 * @end: Limit of search (exclusive). 2793 * @whence: Either SEEK_HOLE or SEEK_DATA. 2794 * 2795 * If the page cache knows which blocks contain holes and which blocks 2796 * contain data, your filesystem can use this function to implement 2797 * SEEK_HOLE and SEEK_DATA. This is useful for filesystems which are 2798 * entirely memory-based such as tmpfs, and filesystems which support 2799 * unwritten extents. 2800 * 2801 * Return: The requested offset on success, or -ENXIO if @whence specifies 2802 * SEEK_DATA and there is no data after @start. There is an implicit hole 2803 * after @end - 1, so SEEK_HOLE returns @end if all the bytes between @start 2804 * and @end contain data. 2805 */ 2806 loff_t mapping_seek_hole_data(struct address_space *mapping, loff_t start, 2807 loff_t end, int whence) 2808 { 2809 XA_STATE(xas, &mapping->i_pages, start >> PAGE_SHIFT); 2810 pgoff_t max = (end - 1) >> PAGE_SHIFT; 2811 bool seek_data = (whence == SEEK_DATA); 2812 struct folio *folio; 2813 2814 if (end <= start) 2815 return -ENXIO; 2816 2817 rcu_read_lock(); 2818 while ((folio = find_get_entry(&xas, max, XA_PRESENT))) { 2819 loff_t pos = (u64)xas.xa_index << PAGE_SHIFT; 2820 size_t seek_size; 2821 2822 if (start < pos) { 2823 if (!seek_data) 2824 goto unlock; 2825 start = pos; 2826 } 2827 2828 seek_size = seek_folio_size(&xas, folio); 2829 pos = round_up((u64)pos + 1, seek_size); 2830 start = folio_seek_hole_data(&xas, mapping, folio, start, pos, 2831 seek_data); 2832 if (start < pos) 2833 goto unlock; 2834 if (start >= end) 2835 break; 2836 if (seek_size > PAGE_SIZE) 2837 xas_set(&xas, pos >> PAGE_SHIFT); 2838 if (!xa_is_value(folio)) 2839 folio_put(folio); 2840 } 2841 if (seek_data) 2842 start = -ENXIO; 2843 unlock: 2844 rcu_read_unlock(); 2845 if (folio && !xa_is_value(folio)) 2846 folio_put(folio); 2847 if (start > end) 2848 return end; 2849 return start; 2850 } 2851 2852 #ifdef CONFIG_MMU 2853 #define MMAP_LOTSAMISS (100) 2854 /* 2855 * lock_folio_maybe_drop_mmap - lock the page, possibly dropping the mmap_lock 2856 * @vmf - the vm_fault for this fault. 2857 * @folio - the folio to lock. 2858 * @fpin - the pointer to the file we may pin (or is already pinned). 2859 * 2860 * This works similar to lock_folio_or_retry in that it can drop the 2861 * mmap_lock. It differs in that it actually returns the folio locked 2862 * if it returns 1 and 0 if it couldn't lock the folio. If we did have 2863 * to drop the mmap_lock then fpin will point to the pinned file and 2864 * needs to be fput()'ed at a later point. 2865 */ 2866 static int lock_folio_maybe_drop_mmap(struct vm_fault *vmf, struct folio *folio, 2867 struct file **fpin) 2868 { 2869 if (folio_trylock(folio)) 2870 return 1; 2871 2872 /* 2873 * NOTE! This will make us return with VM_FAULT_RETRY, but with 2874 * the mmap_lock still held. That's how FAULT_FLAG_RETRY_NOWAIT 2875 * is supposed to work. We have way too many special cases.. 2876 */ 2877 if (vmf->flags & FAULT_FLAG_RETRY_NOWAIT) 2878 return 0; 2879 2880 *fpin = maybe_unlock_mmap_for_io(vmf, *fpin); 2881 if (vmf->flags & FAULT_FLAG_KILLABLE) { 2882 if (__folio_lock_killable(folio)) { 2883 /* 2884 * We didn't have the right flags to drop the mmap_lock, 2885 * but all fault_handlers only check for fatal signals 2886 * if we return VM_FAULT_RETRY, so we need to drop the 2887 * mmap_lock here and return 0 if we don't have a fpin. 2888 */ 2889 if (*fpin == NULL) 2890 mmap_read_unlock(vmf->vma->vm_mm); 2891 return 0; 2892 } 2893 } else 2894 __folio_lock(folio); 2895 2896 return 1; 2897 } 2898 2899 /* 2900 * Synchronous readahead happens when we don't even find a page in the page 2901 * cache at all. We don't want to perform IO under the mmap sem, so if we have 2902 * to drop the mmap sem we return the file that was pinned in order for us to do 2903 * that. If we didn't pin a file then we return NULL. The file that is 2904 * returned needs to be fput()'ed when we're done with it. 2905 */ 2906 static struct file *do_sync_mmap_readahead(struct vm_fault *vmf) 2907 { 2908 struct file *file = vmf->vma->vm_file; 2909 struct file_ra_state *ra = &file->f_ra; 2910 struct address_space *mapping = file->f_mapping; 2911 DEFINE_READAHEAD(ractl, file, ra, mapping, vmf->pgoff); 2912 struct file *fpin = NULL; 2913 unsigned int mmap_miss; 2914 2915 /* If we don't want any read-ahead, don't bother */ 2916 if (vmf->vma->vm_flags & VM_RAND_READ) 2917 return fpin; 2918 if (!ra->ra_pages) 2919 return fpin; 2920 2921 if (vmf->vma->vm_flags & VM_SEQ_READ) { 2922 fpin = maybe_unlock_mmap_for_io(vmf, fpin); 2923 page_cache_sync_ra(&ractl, ra->ra_pages); 2924 return fpin; 2925 } 2926 2927 /* Avoid banging the cache line if not needed */ 2928 mmap_miss = READ_ONCE(ra->mmap_miss); 2929 if (mmap_miss < MMAP_LOTSAMISS * 10) 2930 WRITE_ONCE(ra->mmap_miss, ++mmap_miss); 2931 2932 /* 2933 * Do we miss much more than hit in this file? If so, 2934 * stop bothering with read-ahead. It will only hurt. 2935 */ 2936 if (mmap_miss > MMAP_LOTSAMISS) 2937 return fpin; 2938 2939 /* 2940 * mmap read-around 2941 */ 2942 fpin = maybe_unlock_mmap_for_io(vmf, fpin); 2943 ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2); 2944 ra->size = ra->ra_pages; 2945 ra->async_size = ra->ra_pages / 4; 2946 ractl._index = ra->start; 2947 do_page_cache_ra(&ractl, ra->size, ra->async_size); 2948 return fpin; 2949 } 2950 2951 /* 2952 * Asynchronous readahead happens when we find the page and PG_readahead, 2953 * so we want to possibly extend the readahead further. We return the file that 2954 * was pinned if we have to drop the mmap_lock in order to do IO. 2955 */ 2956 static struct file *do_async_mmap_readahead(struct vm_fault *vmf, 2957 struct folio *folio) 2958 { 2959 struct file *file = vmf->vma->vm_file; 2960 struct file_ra_state *ra = &file->f_ra; 2961 DEFINE_READAHEAD(ractl, file, ra, file->f_mapping, vmf->pgoff); 2962 struct file *fpin = NULL; 2963 unsigned int mmap_miss; 2964 2965 /* If we don't want any read-ahead, don't bother */ 2966 if (vmf->vma->vm_flags & VM_RAND_READ || !ra->ra_pages) 2967 return fpin; 2968 2969 mmap_miss = READ_ONCE(ra->mmap_miss); 2970 if (mmap_miss) 2971 WRITE_ONCE(ra->mmap_miss, --mmap_miss); 2972 2973 if (folio_test_readahead(folio)) { 2974 fpin = maybe_unlock_mmap_for_io(vmf, fpin); 2975 page_cache_async_ra(&ractl, folio, ra->ra_pages); 2976 } 2977 return fpin; 2978 } 2979 2980 /** 2981 * filemap_fault - read in file data for page fault handling 2982 * @vmf: struct vm_fault containing details of the fault 2983 * 2984 * filemap_fault() is invoked via the vma operations vector for a 2985 * mapped memory region to read in file data during a page fault. 2986 * 2987 * The goto's are kind of ugly, but this streamlines the normal case of having 2988 * it in the page cache, and handles the special cases reasonably without 2989 * having a lot of duplicated code. 2990 * 2991 * vma->vm_mm->mmap_lock must be held on entry. 2992 * 2993 * If our return value has VM_FAULT_RETRY set, it's because the mmap_lock 2994 * may be dropped before doing I/O or by lock_folio_maybe_drop_mmap(). 2995 * 2996 * If our return value does not have VM_FAULT_RETRY set, the mmap_lock 2997 * has not been released. 2998 * 2999 * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set. 3000 * 3001 * Return: bitwise-OR of %VM_FAULT_ codes. 3002 */ 3003 vm_fault_t filemap_fault(struct vm_fault *vmf) 3004 { 3005 int error; 3006 struct file *file = vmf->vma->vm_file; 3007 struct file *fpin = NULL; 3008 struct address_space *mapping = file->f_mapping; 3009 struct inode *inode = mapping->host; 3010 pgoff_t max_idx, index = vmf->pgoff; 3011 struct folio *folio; 3012 vm_fault_t ret = 0; 3013 bool mapping_locked = false; 3014 3015 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE); 3016 if (unlikely(index >= max_idx)) 3017 return VM_FAULT_SIGBUS; 3018 3019 /* 3020 * Do we have something in the page cache already? 3021 */ 3022 folio = filemap_get_folio(mapping, index); 3023 if (likely(folio)) { 3024 /* 3025 * We found the page, so try async readahead before waiting for 3026 * the lock. 3027 */ 3028 if (!(vmf->flags & FAULT_FLAG_TRIED)) 3029 fpin = do_async_mmap_readahead(vmf, folio); 3030 if (unlikely(!folio_test_uptodate(folio))) { 3031 filemap_invalidate_lock_shared(mapping); 3032 mapping_locked = true; 3033 } 3034 } else { 3035 /* No page in the page cache at all */ 3036 count_vm_event(PGMAJFAULT); 3037 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT); 3038 ret = VM_FAULT_MAJOR; 3039 fpin = do_sync_mmap_readahead(vmf); 3040 retry_find: 3041 /* 3042 * See comment in filemap_create_folio() why we need 3043 * invalidate_lock 3044 */ 3045 if (!mapping_locked) { 3046 filemap_invalidate_lock_shared(mapping); 3047 mapping_locked = true; 3048 } 3049 folio = __filemap_get_folio(mapping, index, 3050 FGP_CREAT|FGP_FOR_MMAP, 3051 vmf->gfp_mask); 3052 if (!folio) { 3053 if (fpin) 3054 goto out_retry; 3055 filemap_invalidate_unlock_shared(mapping); 3056 return VM_FAULT_OOM; 3057 } 3058 } 3059 3060 if (!lock_folio_maybe_drop_mmap(vmf, folio, &fpin)) 3061 goto out_retry; 3062 3063 /* Did it get truncated? */ 3064 if (unlikely(folio->mapping != mapping)) { 3065 folio_unlock(folio); 3066 folio_put(folio); 3067 goto retry_find; 3068 } 3069 VM_BUG_ON_FOLIO(!folio_contains(folio, index), folio); 3070 3071 /* 3072 * We have a locked page in the page cache, now we need to check 3073 * that it's up-to-date. If not, it is going to be due to an error. 3074 */ 3075 if (unlikely(!folio_test_uptodate(folio))) { 3076 /* 3077 * The page was in cache and uptodate and now it is not. 3078 * Strange but possible since we didn't hold the page lock all 3079 * the time. Let's drop everything get the invalidate lock and 3080 * try again. 3081 */ 3082 if (!mapping_locked) { 3083 folio_unlock(folio); 3084 folio_put(folio); 3085 goto retry_find; 3086 } 3087 goto page_not_uptodate; 3088 } 3089 3090 /* 3091 * We've made it this far and we had to drop our mmap_lock, now is the 3092 * time to return to the upper layer and have it re-find the vma and 3093 * redo the fault. 3094 */ 3095 if (fpin) { 3096 folio_unlock(folio); 3097 goto out_retry; 3098 } 3099 if (mapping_locked) 3100 filemap_invalidate_unlock_shared(mapping); 3101 3102 /* 3103 * Found the page and have a reference on it. 3104 * We must recheck i_size under page lock. 3105 */ 3106 max_idx = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE); 3107 if (unlikely(index >= max_idx)) { 3108 folio_unlock(folio); 3109 folio_put(folio); 3110 return VM_FAULT_SIGBUS; 3111 } 3112 3113 vmf->page = folio_file_page(folio, index); 3114 return ret | VM_FAULT_LOCKED; 3115 3116 page_not_uptodate: 3117 /* 3118 * Umm, take care of errors if the page isn't up-to-date. 3119 * Try to re-read it _once_. We do this synchronously, 3120 * because there really aren't any performance issues here 3121 * and we need to check for errors. 3122 */ 3123 fpin = maybe_unlock_mmap_for_io(vmf, fpin); 3124 error = filemap_read_folio(file, mapping, folio); 3125 if (fpin) 3126 goto out_retry; 3127 folio_put(folio); 3128 3129 if (!error || error == AOP_TRUNCATED_PAGE) 3130 goto retry_find; 3131 filemap_invalidate_unlock_shared(mapping); 3132 3133 return VM_FAULT_SIGBUS; 3134 3135 out_retry: 3136 /* 3137 * We dropped the mmap_lock, we need to return to the fault handler to 3138 * re-find the vma and come back and find our hopefully still populated 3139 * page. 3140 */ 3141 if (folio) 3142 folio_put(folio); 3143 if (mapping_locked) 3144 filemap_invalidate_unlock_shared(mapping); 3145 if (fpin) 3146 fput(fpin); 3147 return ret | VM_FAULT_RETRY; 3148 } 3149 EXPORT_SYMBOL(filemap_fault); 3150 3151 static bool filemap_map_pmd(struct vm_fault *vmf, struct page *page) 3152 { 3153 struct mm_struct *mm = vmf->vma->vm_mm; 3154 3155 /* Huge page is mapped? No need to proceed. */ 3156 if (pmd_trans_huge(*vmf->pmd)) { 3157 unlock_page(page); 3158 put_page(page); 3159 return true; 3160 } 3161 3162 if (pmd_none(*vmf->pmd) && PageTransHuge(page)) { 3163 vm_fault_t ret = do_set_pmd(vmf, page); 3164 if (!ret) { 3165 /* The page is mapped successfully, reference consumed. */ 3166 unlock_page(page); 3167 return true; 3168 } 3169 } 3170 3171 if (pmd_none(*vmf->pmd)) 3172 pmd_install(mm, vmf->pmd, &vmf->prealloc_pte); 3173 3174 /* See comment in handle_pte_fault() */ 3175 if (pmd_devmap_trans_unstable(vmf->pmd)) { 3176 unlock_page(page); 3177 put_page(page); 3178 return true; 3179 } 3180 3181 return false; 3182 } 3183 3184 static struct folio *next_uptodate_page(struct folio *folio, 3185 struct address_space *mapping, 3186 struct xa_state *xas, pgoff_t end_pgoff) 3187 { 3188 unsigned long max_idx; 3189 3190 do { 3191 if (!folio) 3192 return NULL; 3193 if (xas_retry(xas, folio)) 3194 continue; 3195 if (xa_is_value(folio)) 3196 continue; 3197 if (folio_test_locked(folio)) 3198 continue; 3199 if (!folio_try_get_rcu(folio)) 3200 continue; 3201 /* Has the page moved or been split? */ 3202 if (unlikely(folio != xas_reload(xas))) 3203 goto skip; 3204 if (!folio_test_uptodate(folio) || folio_test_readahead(folio)) 3205 goto skip; 3206 if (!folio_trylock(folio)) 3207 goto skip; 3208 if (folio->mapping != mapping) 3209 goto unlock; 3210 if (!folio_test_uptodate(folio)) 3211 goto unlock; 3212 max_idx = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE); 3213 if (xas->xa_index >= max_idx) 3214 goto unlock; 3215 return folio; 3216 unlock: 3217 folio_unlock(folio); 3218 skip: 3219 folio_put(folio); 3220 } while ((folio = xas_next_entry(xas, end_pgoff)) != NULL); 3221 3222 return NULL; 3223 } 3224 3225 static inline struct folio *first_map_page(struct address_space *mapping, 3226 struct xa_state *xas, 3227 pgoff_t end_pgoff) 3228 { 3229 return next_uptodate_page(xas_find(xas, end_pgoff), 3230 mapping, xas, end_pgoff); 3231 } 3232 3233 static inline struct folio *next_map_page(struct address_space *mapping, 3234 struct xa_state *xas, 3235 pgoff_t end_pgoff) 3236 { 3237 return next_uptodate_page(xas_next_entry(xas, end_pgoff), 3238 mapping, xas, end_pgoff); 3239 } 3240 3241 vm_fault_t filemap_map_pages(struct vm_fault *vmf, 3242 pgoff_t start_pgoff, pgoff_t end_pgoff) 3243 { 3244 struct vm_area_struct *vma = vmf->vma; 3245 struct file *file = vma->vm_file; 3246 struct address_space *mapping = file->f_mapping; 3247 pgoff_t last_pgoff = start_pgoff; 3248 unsigned long addr; 3249 XA_STATE(xas, &mapping->i_pages, start_pgoff); 3250 struct folio *folio; 3251 struct page *page; 3252 unsigned int mmap_miss = READ_ONCE(file->f_ra.mmap_miss); 3253 vm_fault_t ret = 0; 3254 3255 rcu_read_lock(); 3256 folio = first_map_page(mapping, &xas, end_pgoff); 3257 if (!folio) 3258 goto out; 3259 3260 if (filemap_map_pmd(vmf, &folio->page)) { 3261 ret = VM_FAULT_NOPAGE; 3262 goto out; 3263 } 3264 3265 addr = vma->vm_start + ((start_pgoff - vma->vm_pgoff) << PAGE_SHIFT); 3266 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl); 3267 do { 3268 again: 3269 page = folio_file_page(folio, xas.xa_index); 3270 if (PageHWPoison(page)) 3271 goto unlock; 3272 3273 if (mmap_miss > 0) 3274 mmap_miss--; 3275 3276 addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT; 3277 vmf->pte += xas.xa_index - last_pgoff; 3278 last_pgoff = xas.xa_index; 3279 3280 if (!pte_none(*vmf->pte)) 3281 goto unlock; 3282 3283 /* We're about to handle the fault */ 3284 if (vmf->address == addr) 3285 ret = VM_FAULT_NOPAGE; 3286 3287 do_set_pte(vmf, page, addr); 3288 /* no need to invalidate: a not-present page won't be cached */ 3289 update_mmu_cache(vma, addr, vmf->pte); 3290 if (folio_more_pages(folio, xas.xa_index, end_pgoff)) { 3291 xas.xa_index++; 3292 folio_ref_inc(folio); 3293 goto again; 3294 } 3295 folio_unlock(folio); 3296 continue; 3297 unlock: 3298 if (folio_more_pages(folio, xas.xa_index, end_pgoff)) { 3299 xas.xa_index++; 3300 goto again; 3301 } 3302 folio_unlock(folio); 3303 folio_put(folio); 3304 } while ((folio = next_map_page(mapping, &xas, end_pgoff)) != NULL); 3305 pte_unmap_unlock(vmf->pte, vmf->ptl); 3306 out: 3307 rcu_read_unlock(); 3308 WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss); 3309 return ret; 3310 } 3311 EXPORT_SYMBOL(filemap_map_pages); 3312 3313 vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf) 3314 { 3315 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 3316 struct folio *folio = page_folio(vmf->page); 3317 vm_fault_t ret = VM_FAULT_LOCKED; 3318 3319 sb_start_pagefault(mapping->host->i_sb); 3320 file_update_time(vmf->vma->vm_file); 3321 folio_lock(folio); 3322 if (folio->mapping != mapping) { 3323 folio_unlock(folio); 3324 ret = VM_FAULT_NOPAGE; 3325 goto out; 3326 } 3327 /* 3328 * We mark the folio dirty already here so that when freeze is in 3329 * progress, we are guaranteed that writeback during freezing will 3330 * see the dirty folio and writeprotect it again. 3331 */ 3332 folio_mark_dirty(folio); 3333 folio_wait_stable(folio); 3334 out: 3335 sb_end_pagefault(mapping->host->i_sb); 3336 return ret; 3337 } 3338 3339 const struct vm_operations_struct generic_file_vm_ops = { 3340 .fault = filemap_fault, 3341 .map_pages = filemap_map_pages, 3342 .page_mkwrite = filemap_page_mkwrite, 3343 }; 3344 3345 /* This is used for a general mmap of a disk file */ 3346 3347 int generic_file_mmap(struct file *file, struct vm_area_struct *vma) 3348 { 3349 struct address_space *mapping = file->f_mapping; 3350 3351 if (!mapping->a_ops->readpage) 3352 return -ENOEXEC; 3353 file_accessed(file); 3354 vma->vm_ops = &generic_file_vm_ops; 3355 return 0; 3356 } 3357 3358 /* 3359 * This is for filesystems which do not implement ->writepage. 3360 */ 3361 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma) 3362 { 3363 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE)) 3364 return -EINVAL; 3365 return generic_file_mmap(file, vma); 3366 } 3367 #else 3368 vm_fault_t filemap_page_mkwrite(struct vm_fault *vmf) 3369 { 3370 return VM_FAULT_SIGBUS; 3371 } 3372 int generic_file_mmap(struct file *file, struct vm_area_struct *vma) 3373 { 3374 return -ENOSYS; 3375 } 3376 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma) 3377 { 3378 return -ENOSYS; 3379 } 3380 #endif /* CONFIG_MMU */ 3381 3382 EXPORT_SYMBOL(filemap_page_mkwrite); 3383 EXPORT_SYMBOL(generic_file_mmap); 3384 EXPORT_SYMBOL(generic_file_readonly_mmap); 3385 3386 static struct folio *do_read_cache_folio(struct address_space *mapping, 3387 pgoff_t index, filler_t filler, void *data, gfp_t gfp) 3388 { 3389 struct folio *folio; 3390 int err; 3391 repeat: 3392 folio = filemap_get_folio(mapping, index); 3393 if (!folio) { 3394 folio = filemap_alloc_folio(gfp, 0); 3395 if (!folio) 3396 return ERR_PTR(-ENOMEM); 3397 err = filemap_add_folio(mapping, folio, index, gfp); 3398 if (unlikely(err)) { 3399 folio_put(folio); 3400 if (err == -EEXIST) 3401 goto repeat; 3402 /* Presumably ENOMEM for xarray node */ 3403 return ERR_PTR(err); 3404 } 3405 3406 filler: 3407 if (filler) 3408 err = filler(data, &folio->page); 3409 else 3410 err = mapping->a_ops->readpage(data, &folio->page); 3411 3412 if (err < 0) { 3413 folio_put(folio); 3414 return ERR_PTR(err); 3415 } 3416 3417 folio_wait_locked(folio); 3418 if (!folio_test_uptodate(folio)) { 3419 folio_put(folio); 3420 return ERR_PTR(-EIO); 3421 } 3422 3423 goto out; 3424 } 3425 if (folio_test_uptodate(folio)) 3426 goto out; 3427 3428 if (!folio_trylock(folio)) { 3429 folio_put_wait_locked(folio, TASK_UNINTERRUPTIBLE); 3430 goto repeat; 3431 } 3432 3433 /* Folio was truncated from mapping */ 3434 if (!folio->mapping) { 3435 folio_unlock(folio); 3436 folio_put(folio); 3437 goto repeat; 3438 } 3439 3440 /* Someone else locked and filled the page in a very small window */ 3441 if (folio_test_uptodate(folio)) { 3442 folio_unlock(folio); 3443 goto out; 3444 } 3445 3446 /* 3447 * A previous I/O error may have been due to temporary 3448 * failures. 3449 * Clear page error before actual read, PG_error will be 3450 * set again if read page fails. 3451 */ 3452 folio_clear_error(folio); 3453 goto filler; 3454 3455 out: 3456 folio_mark_accessed(folio); 3457 return folio; 3458 } 3459 3460 /** 3461 * read_cache_folio - read into page cache, fill it if needed 3462 * @mapping: the page's address_space 3463 * @index: the page index 3464 * @filler: function to perform the read 3465 * @data: first arg to filler(data, page) function, often left as NULL 3466 * 3467 * Read into the page cache. If a page already exists, and PageUptodate() is 3468 * not set, try to fill the page and wait for it to become unlocked. 3469 * 3470 * If the page does not get brought uptodate, return -EIO. 3471 * 3472 * The function expects mapping->invalidate_lock to be already held. 3473 * 3474 * Return: up to date page on success, ERR_PTR() on failure. 3475 */ 3476 struct folio *read_cache_folio(struct address_space *mapping, pgoff_t index, 3477 filler_t filler, void *data) 3478 { 3479 return do_read_cache_folio(mapping, index, filler, data, 3480 mapping_gfp_mask(mapping)); 3481 } 3482 EXPORT_SYMBOL(read_cache_folio); 3483 3484 static struct page *do_read_cache_page(struct address_space *mapping, 3485 pgoff_t index, filler_t *filler, void *data, gfp_t gfp) 3486 { 3487 struct folio *folio; 3488 3489 folio = do_read_cache_folio(mapping, index, filler, data, gfp); 3490 if (IS_ERR(folio)) 3491 return &folio->page; 3492 return folio_file_page(folio, index); 3493 } 3494 3495 struct page *read_cache_page(struct address_space *mapping, 3496 pgoff_t index, filler_t *filler, void *data) 3497 { 3498 return do_read_cache_page(mapping, index, filler, data, 3499 mapping_gfp_mask(mapping)); 3500 } 3501 EXPORT_SYMBOL(read_cache_page); 3502 3503 /** 3504 * read_cache_page_gfp - read into page cache, using specified page allocation flags. 3505 * @mapping: the page's address_space 3506 * @index: the page index 3507 * @gfp: the page allocator flags to use if allocating 3508 * 3509 * This is the same as "read_mapping_page(mapping, index, NULL)", but with 3510 * any new page allocations done using the specified allocation flags. 3511 * 3512 * If the page does not get brought uptodate, return -EIO. 3513 * 3514 * The function expects mapping->invalidate_lock to be already held. 3515 * 3516 * Return: up to date page on success, ERR_PTR() on failure. 3517 */ 3518 struct page *read_cache_page_gfp(struct address_space *mapping, 3519 pgoff_t index, 3520 gfp_t gfp) 3521 { 3522 return do_read_cache_page(mapping, index, NULL, NULL, gfp); 3523 } 3524 EXPORT_SYMBOL(read_cache_page_gfp); 3525 3526 int pagecache_write_begin(struct file *file, struct address_space *mapping, 3527 loff_t pos, unsigned len, unsigned flags, 3528 struct page **pagep, void **fsdata) 3529 { 3530 const struct address_space_operations *aops = mapping->a_ops; 3531 3532 return aops->write_begin(file, mapping, pos, len, flags, 3533 pagep, fsdata); 3534 } 3535 EXPORT_SYMBOL(pagecache_write_begin); 3536 3537 int pagecache_write_end(struct file *file, struct address_space *mapping, 3538 loff_t pos, unsigned len, unsigned copied, 3539 struct page *page, void *fsdata) 3540 { 3541 const struct address_space_operations *aops = mapping->a_ops; 3542 3543 return aops->write_end(file, mapping, pos, len, copied, page, fsdata); 3544 } 3545 EXPORT_SYMBOL(pagecache_write_end); 3546 3547 /* 3548 * Warn about a page cache invalidation failure during a direct I/O write. 3549 */ 3550 void dio_warn_stale_pagecache(struct file *filp) 3551 { 3552 static DEFINE_RATELIMIT_STATE(_rs, 86400 * HZ, DEFAULT_RATELIMIT_BURST); 3553 char pathname[128]; 3554 char *path; 3555 3556 errseq_set(&filp->f_mapping->wb_err, -EIO); 3557 if (__ratelimit(&_rs)) { 3558 path = file_path(filp, pathname, sizeof(pathname)); 3559 if (IS_ERR(path)) 3560 path = "(unknown)"; 3561 pr_crit("Page cache invalidation failure on direct I/O. Possible data corruption due to collision with buffered I/O!\n"); 3562 pr_crit("File: %s PID: %d Comm: %.20s\n", path, current->pid, 3563 current->comm); 3564 } 3565 } 3566 3567 ssize_t 3568 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from) 3569 { 3570 struct file *file = iocb->ki_filp; 3571 struct address_space *mapping = file->f_mapping; 3572 struct inode *inode = mapping->host; 3573 loff_t pos = iocb->ki_pos; 3574 ssize_t written; 3575 size_t write_len; 3576 pgoff_t end; 3577 3578 write_len = iov_iter_count(from); 3579 end = (pos + write_len - 1) >> PAGE_SHIFT; 3580 3581 if (iocb->ki_flags & IOCB_NOWAIT) { 3582 /* If there are pages to writeback, return */ 3583 if (filemap_range_has_page(file->f_mapping, pos, 3584 pos + write_len - 1)) 3585 return -EAGAIN; 3586 } else { 3587 written = filemap_write_and_wait_range(mapping, pos, 3588 pos + write_len - 1); 3589 if (written) 3590 goto out; 3591 } 3592 3593 /* 3594 * After a write we want buffered reads to be sure to go to disk to get 3595 * the new data. We invalidate clean cached page from the region we're 3596 * about to write. We do this *before* the write so that we can return 3597 * without clobbering -EIOCBQUEUED from ->direct_IO(). 3598 */ 3599 written = invalidate_inode_pages2_range(mapping, 3600 pos >> PAGE_SHIFT, end); 3601 /* 3602 * If a page can not be invalidated, return 0 to fall back 3603 * to buffered write. 3604 */ 3605 if (written) { 3606 if (written == -EBUSY) 3607 return 0; 3608 goto out; 3609 } 3610 3611 written = mapping->a_ops->direct_IO(iocb, from); 3612 3613 /* 3614 * Finally, try again to invalidate clean pages which might have been 3615 * cached by non-direct readahead, or faulted in by get_user_pages() 3616 * if the source of the write was an mmap'ed region of the file 3617 * we're writing. Either one is a pretty crazy thing to do, 3618 * so we don't support it 100%. If this invalidation 3619 * fails, tough, the write still worked... 3620 * 3621 * Most of the time we do not need this since dio_complete() will do 3622 * the invalidation for us. However there are some file systems that 3623 * do not end up with dio_complete() being called, so let's not break 3624 * them by removing it completely. 3625 * 3626 * Noticeable example is a blkdev_direct_IO(). 3627 * 3628 * Skip invalidation for async writes or if mapping has no pages. 3629 */ 3630 if (written > 0 && mapping->nrpages && 3631 invalidate_inode_pages2_range(mapping, pos >> PAGE_SHIFT, end)) 3632 dio_warn_stale_pagecache(file); 3633 3634 if (written > 0) { 3635 pos += written; 3636 write_len -= written; 3637 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) { 3638 i_size_write(inode, pos); 3639 mark_inode_dirty(inode); 3640 } 3641 iocb->ki_pos = pos; 3642 } 3643 if (written != -EIOCBQUEUED) 3644 iov_iter_revert(from, write_len - iov_iter_count(from)); 3645 out: 3646 return written; 3647 } 3648 EXPORT_SYMBOL(generic_file_direct_write); 3649 3650 ssize_t generic_perform_write(struct file *file, 3651 struct iov_iter *i, loff_t pos) 3652 { 3653 struct address_space *mapping = file->f_mapping; 3654 const struct address_space_operations *a_ops = mapping->a_ops; 3655 long status = 0; 3656 ssize_t written = 0; 3657 unsigned int flags = 0; 3658 3659 do { 3660 struct page *page; 3661 unsigned long offset; /* Offset into pagecache page */ 3662 unsigned long bytes; /* Bytes to write to page */ 3663 size_t copied; /* Bytes copied from user */ 3664 void *fsdata; 3665 3666 offset = (pos & (PAGE_SIZE - 1)); 3667 bytes = min_t(unsigned long, PAGE_SIZE - offset, 3668 iov_iter_count(i)); 3669 3670 again: 3671 /* 3672 * Bring in the user page that we will copy from _first_. 3673 * Otherwise there's a nasty deadlock on copying from the 3674 * same page as we're writing to, without it being marked 3675 * up-to-date. 3676 */ 3677 if (unlikely(fault_in_iov_iter_readable(i, bytes))) { 3678 status = -EFAULT; 3679 break; 3680 } 3681 3682 if (fatal_signal_pending(current)) { 3683 status = -EINTR; 3684 break; 3685 } 3686 3687 status = a_ops->write_begin(file, mapping, pos, bytes, flags, 3688 &page, &fsdata); 3689 if (unlikely(status < 0)) 3690 break; 3691 3692 if (mapping_writably_mapped(mapping)) 3693 flush_dcache_page(page); 3694 3695 copied = copy_page_from_iter_atomic(page, offset, bytes, i); 3696 flush_dcache_page(page); 3697 3698 status = a_ops->write_end(file, mapping, pos, bytes, copied, 3699 page, fsdata); 3700 if (unlikely(status != copied)) { 3701 iov_iter_revert(i, copied - max(status, 0L)); 3702 if (unlikely(status < 0)) 3703 break; 3704 } 3705 cond_resched(); 3706 3707 if (unlikely(status == 0)) { 3708 /* 3709 * A short copy made ->write_end() reject the 3710 * thing entirely. Might be memory poisoning 3711 * halfway through, might be a race with munmap, 3712 * might be severe memory pressure. 3713 */ 3714 if (copied) 3715 bytes = copied; 3716 goto again; 3717 } 3718 pos += status; 3719 written += status; 3720 3721 balance_dirty_pages_ratelimited(mapping); 3722 } while (iov_iter_count(i)); 3723 3724 return written ? written : status; 3725 } 3726 EXPORT_SYMBOL(generic_perform_write); 3727 3728 /** 3729 * __generic_file_write_iter - write data to a file 3730 * @iocb: IO state structure (file, offset, etc.) 3731 * @from: iov_iter with data to write 3732 * 3733 * This function does all the work needed for actually writing data to a 3734 * file. It does all basic checks, removes SUID from the file, updates 3735 * modification times and calls proper subroutines depending on whether we 3736 * do direct IO or a standard buffered write. 3737 * 3738 * It expects i_rwsem to be grabbed unless we work on a block device or similar 3739 * object which does not need locking at all. 3740 * 3741 * This function does *not* take care of syncing data in case of O_SYNC write. 3742 * A caller has to handle it. This is mainly due to the fact that we want to 3743 * avoid syncing under i_rwsem. 3744 * 3745 * Return: 3746 * * number of bytes written, even for truncated writes 3747 * * negative error code if no data has been written at all 3748 */ 3749 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from) 3750 { 3751 struct file *file = iocb->ki_filp; 3752 struct address_space *mapping = file->f_mapping; 3753 struct inode *inode = mapping->host; 3754 ssize_t written = 0; 3755 ssize_t err; 3756 ssize_t status; 3757 3758 /* We can write back this queue in page reclaim */ 3759 current->backing_dev_info = inode_to_bdi(inode); 3760 err = file_remove_privs(file); 3761 if (err) 3762 goto out; 3763 3764 err = file_update_time(file); 3765 if (err) 3766 goto out; 3767 3768 if (iocb->ki_flags & IOCB_DIRECT) { 3769 loff_t pos, endbyte; 3770 3771 written = generic_file_direct_write(iocb, from); 3772 /* 3773 * If the write stopped short of completing, fall back to 3774 * buffered writes. Some filesystems do this for writes to 3775 * holes, for example. For DAX files, a buffered write will 3776 * not succeed (even if it did, DAX does not handle dirty 3777 * page-cache pages correctly). 3778 */ 3779 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode)) 3780 goto out; 3781 3782 status = generic_perform_write(file, from, pos = iocb->ki_pos); 3783 /* 3784 * If generic_perform_write() returned a synchronous error 3785 * then we want to return the number of bytes which were 3786 * direct-written, or the error code if that was zero. Note 3787 * that this differs from normal direct-io semantics, which 3788 * will return -EFOO even if some bytes were written. 3789 */ 3790 if (unlikely(status < 0)) { 3791 err = status; 3792 goto out; 3793 } 3794 /* 3795 * We need to ensure that the page cache pages are written to 3796 * disk and invalidated to preserve the expected O_DIRECT 3797 * semantics. 3798 */ 3799 endbyte = pos + status - 1; 3800 err = filemap_write_and_wait_range(mapping, pos, endbyte); 3801 if (err == 0) { 3802 iocb->ki_pos = endbyte + 1; 3803 written += status; 3804 invalidate_mapping_pages(mapping, 3805 pos >> PAGE_SHIFT, 3806 endbyte >> PAGE_SHIFT); 3807 } else { 3808 /* 3809 * We don't know how much we wrote, so just return 3810 * the number of bytes which were direct-written 3811 */ 3812 } 3813 } else { 3814 written = generic_perform_write(file, from, iocb->ki_pos); 3815 if (likely(written > 0)) 3816 iocb->ki_pos += written; 3817 } 3818 out: 3819 current->backing_dev_info = NULL; 3820 return written ? written : err; 3821 } 3822 EXPORT_SYMBOL(__generic_file_write_iter); 3823 3824 /** 3825 * generic_file_write_iter - write data to a file 3826 * @iocb: IO state structure 3827 * @from: iov_iter with data to write 3828 * 3829 * This is a wrapper around __generic_file_write_iter() to be used by most 3830 * filesystems. It takes care of syncing the file in case of O_SYNC file 3831 * and acquires i_rwsem as needed. 3832 * Return: 3833 * * negative error code if no data has been written at all of 3834 * vfs_fsync_range() failed for a synchronous write 3835 * * number of bytes written, even for truncated writes 3836 */ 3837 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from) 3838 { 3839 struct file *file = iocb->ki_filp; 3840 struct inode *inode = file->f_mapping->host; 3841 ssize_t ret; 3842 3843 inode_lock(inode); 3844 ret = generic_write_checks(iocb, from); 3845 if (ret > 0) 3846 ret = __generic_file_write_iter(iocb, from); 3847 inode_unlock(inode); 3848 3849 if (ret > 0) 3850 ret = generic_write_sync(iocb, ret); 3851 return ret; 3852 } 3853 EXPORT_SYMBOL(generic_file_write_iter); 3854 3855 /** 3856 * filemap_release_folio() - Release fs-specific metadata on a folio. 3857 * @folio: The folio which the kernel is trying to free. 3858 * @gfp: Memory allocation flags (and I/O mode). 3859 * 3860 * The address_space is trying to release any data attached to a folio 3861 * (presumably at folio->private). 3862 * 3863 * This will also be called if the private_2 flag is set on a page, 3864 * indicating that the folio has other metadata associated with it. 3865 * 3866 * The @gfp argument specifies whether I/O may be performed to release 3867 * this page (__GFP_IO), and whether the call may block 3868 * (__GFP_RECLAIM & __GFP_FS). 3869 * 3870 * Return: %true if the release was successful, otherwise %false. 3871 */ 3872 bool filemap_release_folio(struct folio *folio, gfp_t gfp) 3873 { 3874 struct address_space * const mapping = folio->mapping; 3875 3876 BUG_ON(!folio_test_locked(folio)); 3877 if (folio_test_writeback(folio)) 3878 return false; 3879 3880 if (mapping && mapping->a_ops->releasepage) 3881 return mapping->a_ops->releasepage(&folio->page, gfp); 3882 return try_to_free_buffers(&folio->page); 3883 } 3884 EXPORT_SYMBOL(filemap_release_folio); 3885