1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * fs/dax.c - Direct Access filesystem code 4 * Copyright (c) 2013-2014 Intel Corporation 5 * Author: Matthew Wilcox <matthew.r.wilcox@intel.com> 6 * Author: Ross Zwisler <ross.zwisler@linux.intel.com> 7 */ 8 9 #include <linux/atomic.h> 10 #include <linux/blkdev.h> 11 #include <linux/buffer_head.h> 12 #include <linux/dax.h> 13 #include <linux/fs.h> 14 #include <linux/highmem.h> 15 #include <linux/memcontrol.h> 16 #include <linux/mm.h> 17 #include <linux/mutex.h> 18 #include <linux/pagevec.h> 19 #include <linux/sched.h> 20 #include <linux/sched/signal.h> 21 #include <linux/uio.h> 22 #include <linux/vmstat.h> 23 #include <linux/pfn_t.h> 24 #include <linux/sizes.h> 25 #include <linux/mmu_notifier.h> 26 #include <linux/iomap.h> 27 #include <linux/rmap.h> 28 #include <asm/pgalloc.h> 29 30 #define CREATE_TRACE_POINTS 31 #include <trace/events/fs_dax.h> 32 33 static inline unsigned int pe_order(enum page_entry_size pe_size) 34 { 35 if (pe_size == PE_SIZE_PTE) 36 return PAGE_SHIFT - PAGE_SHIFT; 37 if (pe_size == PE_SIZE_PMD) 38 return PMD_SHIFT - PAGE_SHIFT; 39 if (pe_size == PE_SIZE_PUD) 40 return PUD_SHIFT - PAGE_SHIFT; 41 return ~0; 42 } 43 44 /* We choose 4096 entries - same as per-zone page wait tables */ 45 #define DAX_WAIT_TABLE_BITS 12 46 #define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS) 47 48 /* The 'colour' (ie low bits) within a PMD of a page offset. */ 49 #define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1) 50 #define PG_PMD_NR (PMD_SIZE >> PAGE_SHIFT) 51 52 /* The order of a PMD entry */ 53 #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT) 54 55 static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES]; 56 57 static int __init init_dax_wait_table(void) 58 { 59 int i; 60 61 for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++) 62 init_waitqueue_head(wait_table + i); 63 return 0; 64 } 65 fs_initcall(init_dax_wait_table); 66 67 /* 68 * DAX pagecache entries use XArray value entries so they can't be mistaken 69 * for pages. We use one bit for locking, one bit for the entry size (PMD) 70 * and two more to tell us if the entry is a zero page or an empty entry that 71 * is just used for locking. In total four special bits. 72 * 73 * If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE 74 * and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem 75 * block allocation. 76 */ 77 #define DAX_SHIFT (4) 78 #define DAX_LOCKED (1UL << 0) 79 #define DAX_PMD (1UL << 1) 80 #define DAX_ZERO_PAGE (1UL << 2) 81 #define DAX_EMPTY (1UL << 3) 82 83 static unsigned long dax_to_pfn(void *entry) 84 { 85 return xa_to_value(entry) >> DAX_SHIFT; 86 } 87 88 static void *dax_make_entry(pfn_t pfn, unsigned long flags) 89 { 90 return xa_mk_value(flags | (pfn_t_to_pfn(pfn) << DAX_SHIFT)); 91 } 92 93 static bool dax_is_locked(void *entry) 94 { 95 return xa_to_value(entry) & DAX_LOCKED; 96 } 97 98 static unsigned int dax_entry_order(void *entry) 99 { 100 if (xa_to_value(entry) & DAX_PMD) 101 return PMD_ORDER; 102 return 0; 103 } 104 105 static unsigned long dax_is_pmd_entry(void *entry) 106 { 107 return xa_to_value(entry) & DAX_PMD; 108 } 109 110 static bool dax_is_pte_entry(void *entry) 111 { 112 return !(xa_to_value(entry) & DAX_PMD); 113 } 114 115 static int dax_is_zero_entry(void *entry) 116 { 117 return xa_to_value(entry) & DAX_ZERO_PAGE; 118 } 119 120 static int dax_is_empty_entry(void *entry) 121 { 122 return xa_to_value(entry) & DAX_EMPTY; 123 } 124 125 /* 126 * true if the entry that was found is of a smaller order than the entry 127 * we were looking for 128 */ 129 static bool dax_is_conflict(void *entry) 130 { 131 return entry == XA_RETRY_ENTRY; 132 } 133 134 /* 135 * DAX page cache entry locking 136 */ 137 struct exceptional_entry_key { 138 struct xarray *xa; 139 pgoff_t entry_start; 140 }; 141 142 struct wait_exceptional_entry_queue { 143 wait_queue_entry_t wait; 144 struct exceptional_entry_key key; 145 }; 146 147 /** 148 * enum dax_wake_mode: waitqueue wakeup behaviour 149 * @WAKE_ALL: wake all waiters in the waitqueue 150 * @WAKE_NEXT: wake only the first waiter in the waitqueue 151 */ 152 enum dax_wake_mode { 153 WAKE_ALL, 154 WAKE_NEXT, 155 }; 156 157 static wait_queue_head_t *dax_entry_waitqueue(struct xa_state *xas, 158 void *entry, struct exceptional_entry_key *key) 159 { 160 unsigned long hash; 161 unsigned long index = xas->xa_index; 162 163 /* 164 * If 'entry' is a PMD, align the 'index' that we use for the wait 165 * queue to the start of that PMD. This ensures that all offsets in 166 * the range covered by the PMD map to the same bit lock. 167 */ 168 if (dax_is_pmd_entry(entry)) 169 index &= ~PG_PMD_COLOUR; 170 key->xa = xas->xa; 171 key->entry_start = index; 172 173 hash = hash_long((unsigned long)xas->xa ^ index, DAX_WAIT_TABLE_BITS); 174 return wait_table + hash; 175 } 176 177 static int wake_exceptional_entry_func(wait_queue_entry_t *wait, 178 unsigned int mode, int sync, void *keyp) 179 { 180 struct exceptional_entry_key *key = keyp; 181 struct wait_exceptional_entry_queue *ewait = 182 container_of(wait, struct wait_exceptional_entry_queue, wait); 183 184 if (key->xa != ewait->key.xa || 185 key->entry_start != ewait->key.entry_start) 186 return 0; 187 return autoremove_wake_function(wait, mode, sync, NULL); 188 } 189 190 /* 191 * @entry may no longer be the entry at the index in the mapping. 192 * The important information it's conveying is whether the entry at 193 * this index used to be a PMD entry. 194 */ 195 static void dax_wake_entry(struct xa_state *xas, void *entry, 196 enum dax_wake_mode mode) 197 { 198 struct exceptional_entry_key key; 199 wait_queue_head_t *wq; 200 201 wq = dax_entry_waitqueue(xas, entry, &key); 202 203 /* 204 * Checking for locked entry and prepare_to_wait_exclusive() happens 205 * under the i_pages lock, ditto for entry handling in our callers. 206 * So at this point all tasks that could have seen our entry locked 207 * must be in the waitqueue and the following check will see them. 208 */ 209 if (waitqueue_active(wq)) 210 __wake_up(wq, TASK_NORMAL, mode == WAKE_ALL ? 0 : 1, &key); 211 } 212 213 /* 214 * Look up entry in page cache, wait for it to become unlocked if it 215 * is a DAX entry and return it. The caller must subsequently call 216 * put_unlocked_entry() if it did not lock the entry or dax_unlock_entry() 217 * if it did. The entry returned may have a larger order than @order. 218 * If @order is larger than the order of the entry found in i_pages, this 219 * function returns a dax_is_conflict entry. 220 * 221 * Must be called with the i_pages lock held. 222 */ 223 static void *get_unlocked_entry(struct xa_state *xas, unsigned int order) 224 { 225 void *entry; 226 struct wait_exceptional_entry_queue ewait; 227 wait_queue_head_t *wq; 228 229 init_wait(&ewait.wait); 230 ewait.wait.func = wake_exceptional_entry_func; 231 232 for (;;) { 233 entry = xas_find_conflict(xas); 234 if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) 235 return entry; 236 if (dax_entry_order(entry) < order) 237 return XA_RETRY_ENTRY; 238 if (!dax_is_locked(entry)) 239 return entry; 240 241 wq = dax_entry_waitqueue(xas, entry, &ewait.key); 242 prepare_to_wait_exclusive(wq, &ewait.wait, 243 TASK_UNINTERRUPTIBLE); 244 xas_unlock_irq(xas); 245 xas_reset(xas); 246 schedule(); 247 finish_wait(wq, &ewait.wait); 248 xas_lock_irq(xas); 249 } 250 } 251 252 /* 253 * The only thing keeping the address space around is the i_pages lock 254 * (it's cycled in clear_inode() after removing the entries from i_pages) 255 * After we call xas_unlock_irq(), we cannot touch xas->xa. 256 */ 257 static void wait_entry_unlocked(struct xa_state *xas, void *entry) 258 { 259 struct wait_exceptional_entry_queue ewait; 260 wait_queue_head_t *wq; 261 262 init_wait(&ewait.wait); 263 ewait.wait.func = wake_exceptional_entry_func; 264 265 wq = dax_entry_waitqueue(xas, entry, &ewait.key); 266 /* 267 * Unlike get_unlocked_entry() there is no guarantee that this 268 * path ever successfully retrieves an unlocked entry before an 269 * inode dies. Perform a non-exclusive wait in case this path 270 * never successfully performs its own wake up. 271 */ 272 prepare_to_wait(wq, &ewait.wait, TASK_UNINTERRUPTIBLE); 273 xas_unlock_irq(xas); 274 schedule(); 275 finish_wait(wq, &ewait.wait); 276 } 277 278 static void put_unlocked_entry(struct xa_state *xas, void *entry, 279 enum dax_wake_mode mode) 280 { 281 if (entry && !dax_is_conflict(entry)) 282 dax_wake_entry(xas, entry, mode); 283 } 284 285 /* 286 * We used the xa_state to get the entry, but then we locked the entry and 287 * dropped the xa_lock, so we know the xa_state is stale and must be reset 288 * before use. 289 */ 290 static void dax_unlock_entry(struct xa_state *xas, void *entry) 291 { 292 void *old; 293 294 BUG_ON(dax_is_locked(entry)); 295 xas_reset(xas); 296 xas_lock_irq(xas); 297 old = xas_store(xas, entry); 298 xas_unlock_irq(xas); 299 BUG_ON(!dax_is_locked(old)); 300 dax_wake_entry(xas, entry, WAKE_NEXT); 301 } 302 303 /* 304 * Return: The entry stored at this location before it was locked. 305 */ 306 static void *dax_lock_entry(struct xa_state *xas, void *entry) 307 { 308 unsigned long v = xa_to_value(entry); 309 return xas_store(xas, xa_mk_value(v | DAX_LOCKED)); 310 } 311 312 static unsigned long dax_entry_size(void *entry) 313 { 314 if (dax_is_zero_entry(entry)) 315 return 0; 316 else if (dax_is_empty_entry(entry)) 317 return 0; 318 else if (dax_is_pmd_entry(entry)) 319 return PMD_SIZE; 320 else 321 return PAGE_SIZE; 322 } 323 324 static unsigned long dax_end_pfn(void *entry) 325 { 326 return dax_to_pfn(entry) + dax_entry_size(entry) / PAGE_SIZE; 327 } 328 329 /* 330 * Iterate through all mapped pfns represented by an entry, i.e. skip 331 * 'empty' and 'zero' entries. 332 */ 333 #define for_each_mapped_pfn(entry, pfn) \ 334 for (pfn = dax_to_pfn(entry); \ 335 pfn < dax_end_pfn(entry); pfn++) 336 337 /* 338 * TODO: for reflink+dax we need a way to associate a single page with 339 * multiple address_space instances at different linear_page_index() 340 * offsets. 341 */ 342 static void dax_associate_entry(void *entry, struct address_space *mapping, 343 struct vm_area_struct *vma, unsigned long address) 344 { 345 unsigned long size = dax_entry_size(entry), pfn, index; 346 int i = 0; 347 348 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 349 return; 350 351 index = linear_page_index(vma, address & ~(size - 1)); 352 for_each_mapped_pfn(entry, pfn) { 353 struct page *page = pfn_to_page(pfn); 354 355 WARN_ON_ONCE(page->mapping); 356 page->mapping = mapping; 357 page->index = index + i++; 358 } 359 } 360 361 static void dax_disassociate_entry(void *entry, struct address_space *mapping, 362 bool trunc) 363 { 364 unsigned long pfn; 365 366 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 367 return; 368 369 for_each_mapped_pfn(entry, pfn) { 370 struct page *page = pfn_to_page(pfn); 371 372 WARN_ON_ONCE(trunc && page_ref_count(page) > 1); 373 WARN_ON_ONCE(page->mapping && page->mapping != mapping); 374 page->mapping = NULL; 375 page->index = 0; 376 } 377 } 378 379 static struct page *dax_busy_page(void *entry) 380 { 381 unsigned long pfn; 382 383 for_each_mapped_pfn(entry, pfn) { 384 struct page *page = pfn_to_page(pfn); 385 386 if (page_ref_count(page) > 1) 387 return page; 388 } 389 return NULL; 390 } 391 392 /* 393 * dax_lock_page - Lock the DAX entry corresponding to a page 394 * @page: The page whose entry we want to lock 395 * 396 * Context: Process context. 397 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could 398 * not be locked. 399 */ 400 dax_entry_t dax_lock_page(struct page *page) 401 { 402 XA_STATE(xas, NULL, 0); 403 void *entry; 404 405 /* Ensure page->mapping isn't freed while we look at it */ 406 rcu_read_lock(); 407 for (;;) { 408 struct address_space *mapping = READ_ONCE(page->mapping); 409 410 entry = NULL; 411 if (!mapping || !dax_mapping(mapping)) 412 break; 413 414 /* 415 * In the device-dax case there's no need to lock, a 416 * struct dev_pagemap pin is sufficient to keep the 417 * inode alive, and we assume we have dev_pagemap pin 418 * otherwise we would not have a valid pfn_to_page() 419 * translation. 420 */ 421 entry = (void *)~0UL; 422 if (S_ISCHR(mapping->host->i_mode)) 423 break; 424 425 xas.xa = &mapping->i_pages; 426 xas_lock_irq(&xas); 427 if (mapping != page->mapping) { 428 xas_unlock_irq(&xas); 429 continue; 430 } 431 xas_set(&xas, page->index); 432 entry = xas_load(&xas); 433 if (dax_is_locked(entry)) { 434 rcu_read_unlock(); 435 wait_entry_unlocked(&xas, entry); 436 rcu_read_lock(); 437 continue; 438 } 439 dax_lock_entry(&xas, entry); 440 xas_unlock_irq(&xas); 441 break; 442 } 443 rcu_read_unlock(); 444 return (dax_entry_t)entry; 445 } 446 447 void dax_unlock_page(struct page *page, dax_entry_t cookie) 448 { 449 struct address_space *mapping = page->mapping; 450 XA_STATE(xas, &mapping->i_pages, page->index); 451 452 if (S_ISCHR(mapping->host->i_mode)) 453 return; 454 455 dax_unlock_entry(&xas, (void *)cookie); 456 } 457 458 /* 459 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping 460 * @mapping: the file's mapping whose entry we want to lock 461 * @index: the offset within this file 462 * @page: output the dax page corresponding to this dax entry 463 * 464 * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry 465 * could not be locked. 466 */ 467 dax_entry_t dax_lock_mapping_entry(struct address_space *mapping, pgoff_t index, 468 struct page **page) 469 { 470 XA_STATE(xas, NULL, 0); 471 void *entry; 472 473 rcu_read_lock(); 474 for (;;) { 475 entry = NULL; 476 if (!dax_mapping(mapping)) 477 break; 478 479 xas.xa = &mapping->i_pages; 480 xas_lock_irq(&xas); 481 xas_set(&xas, index); 482 entry = xas_load(&xas); 483 if (dax_is_locked(entry)) { 484 rcu_read_unlock(); 485 wait_entry_unlocked(&xas, entry); 486 rcu_read_lock(); 487 continue; 488 } 489 if (!entry || 490 dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { 491 /* 492 * Because we are looking for entry from file's mapping 493 * and index, so the entry may not be inserted for now, 494 * or even a zero/empty entry. We don't think this is 495 * an error case. So, return a special value and do 496 * not output @page. 497 */ 498 entry = (void *)~0UL; 499 } else { 500 *page = pfn_to_page(dax_to_pfn(entry)); 501 dax_lock_entry(&xas, entry); 502 } 503 xas_unlock_irq(&xas); 504 break; 505 } 506 rcu_read_unlock(); 507 return (dax_entry_t)entry; 508 } 509 510 void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index, 511 dax_entry_t cookie) 512 { 513 XA_STATE(xas, &mapping->i_pages, index); 514 515 if (cookie == ~0UL) 516 return; 517 518 dax_unlock_entry(&xas, (void *)cookie); 519 } 520 521 /* 522 * Find page cache entry at given index. If it is a DAX entry, return it 523 * with the entry locked. If the page cache doesn't contain an entry at 524 * that index, add a locked empty entry. 525 * 526 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will 527 * either return that locked entry or will return VM_FAULT_FALLBACK. 528 * This will happen if there are any PTE entries within the PMD range 529 * that we are requesting. 530 * 531 * We always favor PTE entries over PMD entries. There isn't a flow where we 532 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD 533 * insertion will fail if it finds any PTE entries already in the tree, and a 534 * PTE insertion will cause an existing PMD entry to be unmapped and 535 * downgraded to PTE entries. This happens for both PMD zero pages as 536 * well as PMD empty entries. 537 * 538 * The exception to this downgrade path is for PMD entries that have 539 * real storage backing them. We will leave these real PMD entries in 540 * the tree, and PTE writes will simply dirty the entire PMD entry. 541 * 542 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For 543 * persistent memory the benefit is doubtful. We can add that later if we can 544 * show it helps. 545 * 546 * On error, this function does not return an ERR_PTR. Instead it returns 547 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values 548 * overlap with xarray value entries. 549 */ 550 static void *grab_mapping_entry(struct xa_state *xas, 551 struct address_space *mapping, unsigned int order) 552 { 553 unsigned long index = xas->xa_index; 554 bool pmd_downgrade; /* splitting PMD entry into PTE entries? */ 555 void *entry; 556 557 retry: 558 pmd_downgrade = false; 559 xas_lock_irq(xas); 560 entry = get_unlocked_entry(xas, order); 561 562 if (entry) { 563 if (dax_is_conflict(entry)) 564 goto fallback; 565 if (!xa_is_value(entry)) { 566 xas_set_err(xas, -EIO); 567 goto out_unlock; 568 } 569 570 if (order == 0) { 571 if (dax_is_pmd_entry(entry) && 572 (dax_is_zero_entry(entry) || 573 dax_is_empty_entry(entry))) { 574 pmd_downgrade = true; 575 } 576 } 577 } 578 579 if (pmd_downgrade) { 580 /* 581 * Make sure 'entry' remains valid while we drop 582 * the i_pages lock. 583 */ 584 dax_lock_entry(xas, entry); 585 586 /* 587 * Besides huge zero pages the only other thing that gets 588 * downgraded are empty entries which don't need to be 589 * unmapped. 590 */ 591 if (dax_is_zero_entry(entry)) { 592 xas_unlock_irq(xas); 593 unmap_mapping_pages(mapping, 594 xas->xa_index & ~PG_PMD_COLOUR, 595 PG_PMD_NR, false); 596 xas_reset(xas); 597 xas_lock_irq(xas); 598 } 599 600 dax_disassociate_entry(entry, mapping, false); 601 xas_store(xas, NULL); /* undo the PMD join */ 602 dax_wake_entry(xas, entry, WAKE_ALL); 603 mapping->nrpages -= PG_PMD_NR; 604 entry = NULL; 605 xas_set(xas, index); 606 } 607 608 if (entry) { 609 dax_lock_entry(xas, entry); 610 } else { 611 unsigned long flags = DAX_EMPTY; 612 613 if (order > 0) 614 flags |= DAX_PMD; 615 entry = dax_make_entry(pfn_to_pfn_t(0), flags); 616 dax_lock_entry(xas, entry); 617 if (xas_error(xas)) 618 goto out_unlock; 619 mapping->nrpages += 1UL << order; 620 } 621 622 out_unlock: 623 xas_unlock_irq(xas); 624 if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM)) 625 goto retry; 626 if (xas->xa_node == XA_ERROR(-ENOMEM)) 627 return xa_mk_internal(VM_FAULT_OOM); 628 if (xas_error(xas)) 629 return xa_mk_internal(VM_FAULT_SIGBUS); 630 return entry; 631 fallback: 632 xas_unlock_irq(xas); 633 return xa_mk_internal(VM_FAULT_FALLBACK); 634 } 635 636 /** 637 * dax_layout_busy_page_range - find first pinned page in @mapping 638 * @mapping: address space to scan for a page with ref count > 1 639 * @start: Starting offset. Page containing 'start' is included. 640 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX, 641 * pages from 'start' till the end of file are included. 642 * 643 * DAX requires ZONE_DEVICE mapped pages. These pages are never 644 * 'onlined' to the page allocator so they are considered idle when 645 * page->count == 1. A filesystem uses this interface to determine if 646 * any page in the mapping is busy, i.e. for DMA, or other 647 * get_user_pages() usages. 648 * 649 * It is expected that the filesystem is holding locks to block the 650 * establishment of new mappings in this address_space. I.e. it expects 651 * to be able to run unmap_mapping_range() and subsequently not race 652 * mapping_mapped() becoming true. 653 */ 654 struct page *dax_layout_busy_page_range(struct address_space *mapping, 655 loff_t start, loff_t end) 656 { 657 void *entry; 658 unsigned int scanned = 0; 659 struct page *page = NULL; 660 pgoff_t start_idx = start >> PAGE_SHIFT; 661 pgoff_t end_idx; 662 XA_STATE(xas, &mapping->i_pages, start_idx); 663 664 /* 665 * In the 'limited' case get_user_pages() for dax is disabled. 666 */ 667 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 668 return NULL; 669 670 if (!dax_mapping(mapping) || !mapping_mapped(mapping)) 671 return NULL; 672 673 /* If end == LLONG_MAX, all pages from start to till end of file */ 674 if (end == LLONG_MAX) 675 end_idx = ULONG_MAX; 676 else 677 end_idx = end >> PAGE_SHIFT; 678 /* 679 * If we race get_user_pages_fast() here either we'll see the 680 * elevated page count in the iteration and wait, or 681 * get_user_pages_fast() will see that the page it took a reference 682 * against is no longer mapped in the page tables and bail to the 683 * get_user_pages() slow path. The slow path is protected by 684 * pte_lock() and pmd_lock(). New references are not taken without 685 * holding those locks, and unmap_mapping_pages() will not zero the 686 * pte or pmd without holding the respective lock, so we are 687 * guaranteed to either see new references or prevent new 688 * references from being established. 689 */ 690 unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0); 691 692 xas_lock_irq(&xas); 693 xas_for_each(&xas, entry, end_idx) { 694 if (WARN_ON_ONCE(!xa_is_value(entry))) 695 continue; 696 if (unlikely(dax_is_locked(entry))) 697 entry = get_unlocked_entry(&xas, 0); 698 if (entry) 699 page = dax_busy_page(entry); 700 put_unlocked_entry(&xas, entry, WAKE_NEXT); 701 if (page) 702 break; 703 if (++scanned % XA_CHECK_SCHED) 704 continue; 705 706 xas_pause(&xas); 707 xas_unlock_irq(&xas); 708 cond_resched(); 709 xas_lock_irq(&xas); 710 } 711 xas_unlock_irq(&xas); 712 return page; 713 } 714 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range); 715 716 struct page *dax_layout_busy_page(struct address_space *mapping) 717 { 718 return dax_layout_busy_page_range(mapping, 0, LLONG_MAX); 719 } 720 EXPORT_SYMBOL_GPL(dax_layout_busy_page); 721 722 static int __dax_invalidate_entry(struct address_space *mapping, 723 pgoff_t index, bool trunc) 724 { 725 XA_STATE(xas, &mapping->i_pages, index); 726 int ret = 0; 727 void *entry; 728 729 xas_lock_irq(&xas); 730 entry = get_unlocked_entry(&xas, 0); 731 if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) 732 goto out; 733 if (!trunc && 734 (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) || 735 xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE))) 736 goto out; 737 dax_disassociate_entry(entry, mapping, trunc); 738 xas_store(&xas, NULL); 739 mapping->nrpages -= 1UL << dax_entry_order(entry); 740 ret = 1; 741 out: 742 put_unlocked_entry(&xas, entry, WAKE_ALL); 743 xas_unlock_irq(&xas); 744 return ret; 745 } 746 747 /* 748 * Delete DAX entry at @index from @mapping. Wait for it 749 * to be unlocked before deleting it. 750 */ 751 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) 752 { 753 int ret = __dax_invalidate_entry(mapping, index, true); 754 755 /* 756 * This gets called from truncate / punch_hole path. As such, the caller 757 * must hold locks protecting against concurrent modifications of the 758 * page cache (usually fs-private i_mmap_sem for writing). Since the 759 * caller has seen a DAX entry for this index, we better find it 760 * at that index as well... 761 */ 762 WARN_ON_ONCE(!ret); 763 return ret; 764 } 765 766 /* 767 * Invalidate DAX entry if it is clean. 768 */ 769 int dax_invalidate_mapping_entry_sync(struct address_space *mapping, 770 pgoff_t index) 771 { 772 return __dax_invalidate_entry(mapping, index, false); 773 } 774 775 static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos) 776 { 777 return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset); 778 } 779 780 static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter) 781 { 782 pgoff_t pgoff = dax_iomap_pgoff(&iter->iomap, iter->pos); 783 void *vto, *kaddr; 784 long rc; 785 int id; 786 787 id = dax_read_lock(); 788 rc = dax_direct_access(iter->iomap.dax_dev, pgoff, 1, DAX_ACCESS, 789 &kaddr, NULL); 790 if (rc < 0) { 791 dax_read_unlock(id); 792 return rc; 793 } 794 vto = kmap_atomic(vmf->cow_page); 795 copy_user_page(vto, kaddr, vmf->address, vmf->cow_page); 796 kunmap_atomic(vto); 797 dax_read_unlock(id); 798 return 0; 799 } 800 801 /* 802 * By this point grab_mapping_entry() has ensured that we have a locked entry 803 * of the appropriate size so we don't have to worry about downgrading PMDs to 804 * PTEs. If we happen to be trying to insert a PTE and there is a PMD 805 * already in the tree, we will skip the insertion and just dirty the PMD as 806 * appropriate. 807 */ 808 static void *dax_insert_entry(struct xa_state *xas, 809 struct address_space *mapping, struct vm_fault *vmf, 810 void *entry, pfn_t pfn, unsigned long flags, bool dirty) 811 { 812 void *new_entry = dax_make_entry(pfn, flags); 813 814 if (dirty) 815 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 816 817 if (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE)) { 818 unsigned long index = xas->xa_index; 819 /* we are replacing a zero page with block mapping */ 820 if (dax_is_pmd_entry(entry)) 821 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR, 822 PG_PMD_NR, false); 823 else /* pte entry */ 824 unmap_mapping_pages(mapping, index, 1, false); 825 } 826 827 xas_reset(xas); 828 xas_lock_irq(xas); 829 if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { 830 void *old; 831 832 dax_disassociate_entry(entry, mapping, false); 833 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address); 834 /* 835 * Only swap our new entry into the page cache if the current 836 * entry is a zero page or an empty entry. If a normal PTE or 837 * PMD entry is already in the cache, we leave it alone. This 838 * means that if we are trying to insert a PTE and the 839 * existing entry is a PMD, we will just leave the PMD in the 840 * tree and dirty it if necessary. 841 */ 842 old = dax_lock_entry(xas, new_entry); 843 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) | 844 DAX_LOCKED)); 845 entry = new_entry; 846 } else { 847 xas_load(xas); /* Walk the xa_state */ 848 } 849 850 if (dirty) 851 xas_set_mark(xas, PAGECACHE_TAG_DIRTY); 852 853 xas_unlock_irq(xas); 854 return entry; 855 } 856 857 static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev, 858 struct address_space *mapping, void *entry) 859 { 860 unsigned long pfn, index, count, end; 861 long ret = 0; 862 struct vm_area_struct *vma; 863 864 /* 865 * A page got tagged dirty in DAX mapping? Something is seriously 866 * wrong. 867 */ 868 if (WARN_ON(!xa_is_value(entry))) 869 return -EIO; 870 871 if (unlikely(dax_is_locked(entry))) { 872 void *old_entry = entry; 873 874 entry = get_unlocked_entry(xas, 0); 875 876 /* Entry got punched out / reallocated? */ 877 if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) 878 goto put_unlocked; 879 /* 880 * Entry got reallocated elsewhere? No need to writeback. 881 * We have to compare pfns as we must not bail out due to 882 * difference in lockbit or entry type. 883 */ 884 if (dax_to_pfn(old_entry) != dax_to_pfn(entry)) 885 goto put_unlocked; 886 if (WARN_ON_ONCE(dax_is_empty_entry(entry) || 887 dax_is_zero_entry(entry))) { 888 ret = -EIO; 889 goto put_unlocked; 890 } 891 892 /* Another fsync thread may have already done this entry */ 893 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE)) 894 goto put_unlocked; 895 } 896 897 /* Lock the entry to serialize with page faults */ 898 dax_lock_entry(xas, entry); 899 900 /* 901 * We can clear the tag now but we have to be careful so that concurrent 902 * dax_writeback_one() calls for the same index cannot finish before we 903 * actually flush the caches. This is achieved as the calls will look 904 * at the entry only under the i_pages lock and once they do that 905 * they will see the entry locked and wait for it to unlock. 906 */ 907 xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE); 908 xas_unlock_irq(xas); 909 910 /* 911 * If dax_writeback_mapping_range() was given a wbc->range_start 912 * in the middle of a PMD, the 'index' we use needs to be 913 * aligned to the start of the PMD. 914 * This allows us to flush for PMD_SIZE and not have to worry about 915 * partial PMD writebacks. 916 */ 917 pfn = dax_to_pfn(entry); 918 count = 1UL << dax_entry_order(entry); 919 index = xas->xa_index & ~(count - 1); 920 end = index + count - 1; 921 922 /* Walk all mappings of a given index of a file and writeprotect them */ 923 i_mmap_lock_read(mapping); 924 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, end) { 925 pfn_mkclean_range(pfn, count, index, vma); 926 cond_resched(); 927 } 928 i_mmap_unlock_read(mapping); 929 930 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE); 931 /* 932 * After we have flushed the cache, we can clear the dirty tag. There 933 * cannot be new dirty data in the pfn after the flush has completed as 934 * the pfn mappings are writeprotected and fault waits for mapping 935 * entry lock. 936 */ 937 xas_reset(xas); 938 xas_lock_irq(xas); 939 xas_store(xas, entry); 940 xas_clear_mark(xas, PAGECACHE_TAG_DIRTY); 941 dax_wake_entry(xas, entry, WAKE_NEXT); 942 943 trace_dax_writeback_one(mapping->host, index, count); 944 return ret; 945 946 put_unlocked: 947 put_unlocked_entry(xas, entry, WAKE_NEXT); 948 return ret; 949 } 950 951 /* 952 * Flush the mapping to the persistent domain within the byte range of [start, 953 * end]. This is required by data integrity operations to ensure file data is 954 * on persistent storage prior to completion of the operation. 955 */ 956 int dax_writeback_mapping_range(struct address_space *mapping, 957 struct dax_device *dax_dev, struct writeback_control *wbc) 958 { 959 XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT); 960 struct inode *inode = mapping->host; 961 pgoff_t end_index = wbc->range_end >> PAGE_SHIFT; 962 void *entry; 963 int ret = 0; 964 unsigned int scanned = 0; 965 966 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) 967 return -EIO; 968 969 if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL) 970 return 0; 971 972 trace_dax_writeback_range(inode, xas.xa_index, end_index); 973 974 tag_pages_for_writeback(mapping, xas.xa_index, end_index); 975 976 xas_lock_irq(&xas); 977 xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) { 978 ret = dax_writeback_one(&xas, dax_dev, mapping, entry); 979 if (ret < 0) { 980 mapping_set_error(mapping, ret); 981 break; 982 } 983 if (++scanned % XA_CHECK_SCHED) 984 continue; 985 986 xas_pause(&xas); 987 xas_unlock_irq(&xas); 988 cond_resched(); 989 xas_lock_irq(&xas); 990 } 991 xas_unlock_irq(&xas); 992 trace_dax_writeback_range_done(inode, xas.xa_index, end_index); 993 return ret; 994 } 995 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); 996 997 static int dax_iomap_pfn(const struct iomap *iomap, loff_t pos, size_t size, 998 pfn_t *pfnp) 999 { 1000 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); 1001 int id, rc; 1002 long length; 1003 1004 id = dax_read_lock(); 1005 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size), 1006 DAX_ACCESS, NULL, pfnp); 1007 if (length < 0) { 1008 rc = length; 1009 goto out; 1010 } 1011 rc = -EINVAL; 1012 if (PFN_PHYS(length) < size) 1013 goto out; 1014 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1)) 1015 goto out; 1016 /* For larger pages we need devmap */ 1017 if (length > 1 && !pfn_t_devmap(*pfnp)) 1018 goto out; 1019 rc = 0; 1020 out: 1021 dax_read_unlock(id); 1022 return rc; 1023 } 1024 1025 /* 1026 * The user has performed a load from a hole in the file. Allocating a new 1027 * page in the file would cause excessive storage usage for workloads with 1028 * sparse files. Instead we insert a read-only mapping of the 4k zero page. 1029 * If this page is ever written to we will re-fault and change the mapping to 1030 * point to real DAX storage instead. 1031 */ 1032 static vm_fault_t dax_load_hole(struct xa_state *xas, 1033 struct address_space *mapping, void **entry, 1034 struct vm_fault *vmf) 1035 { 1036 struct inode *inode = mapping->host; 1037 unsigned long vaddr = vmf->address; 1038 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr)); 1039 vm_fault_t ret; 1040 1041 *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, 1042 DAX_ZERO_PAGE, false); 1043 1044 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn); 1045 trace_dax_load_hole(inode, vmf, ret); 1046 return ret; 1047 } 1048 1049 #ifdef CONFIG_FS_DAX_PMD 1050 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, 1051 const struct iomap *iomap, void **entry) 1052 { 1053 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1054 unsigned long pmd_addr = vmf->address & PMD_MASK; 1055 struct vm_area_struct *vma = vmf->vma; 1056 struct inode *inode = mapping->host; 1057 pgtable_t pgtable = NULL; 1058 struct page *zero_page; 1059 spinlock_t *ptl; 1060 pmd_t pmd_entry; 1061 pfn_t pfn; 1062 1063 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); 1064 1065 if (unlikely(!zero_page)) 1066 goto fallback; 1067 1068 pfn = page_to_pfn_t(zero_page); 1069 *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, 1070 DAX_PMD | DAX_ZERO_PAGE, false); 1071 1072 if (arch_needs_pgtable_deposit()) { 1073 pgtable = pte_alloc_one(vma->vm_mm); 1074 if (!pgtable) 1075 return VM_FAULT_OOM; 1076 } 1077 1078 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1079 if (!pmd_none(*(vmf->pmd))) { 1080 spin_unlock(ptl); 1081 goto fallback; 1082 } 1083 1084 if (pgtable) { 1085 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 1086 mm_inc_nr_ptes(vma->vm_mm); 1087 } 1088 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); 1089 pmd_entry = pmd_mkhuge(pmd_entry); 1090 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); 1091 spin_unlock(ptl); 1092 trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry); 1093 return VM_FAULT_NOPAGE; 1094 1095 fallback: 1096 if (pgtable) 1097 pte_free(vma->vm_mm, pgtable); 1098 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry); 1099 return VM_FAULT_FALLBACK; 1100 } 1101 #else 1102 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, 1103 const struct iomap *iomap, void **entry) 1104 { 1105 return VM_FAULT_FALLBACK; 1106 } 1107 #endif /* CONFIG_FS_DAX_PMD */ 1108 1109 static int dax_memzero(struct dax_device *dax_dev, pgoff_t pgoff, 1110 unsigned int offset, size_t size) 1111 { 1112 void *kaddr; 1113 long ret; 1114 1115 ret = dax_direct_access(dax_dev, pgoff, 1, DAX_ACCESS, &kaddr, NULL); 1116 if (ret > 0) { 1117 memset(kaddr + offset, 0, size); 1118 dax_flush(dax_dev, kaddr + offset, size); 1119 } 1120 return ret; 1121 } 1122 1123 static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero) 1124 { 1125 const struct iomap *iomap = &iter->iomap; 1126 const struct iomap *srcmap = iomap_iter_srcmap(iter); 1127 loff_t pos = iter->pos; 1128 u64 length = iomap_length(iter); 1129 s64 written = 0; 1130 1131 /* already zeroed? we're done. */ 1132 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN) 1133 return length; 1134 1135 do { 1136 unsigned offset = offset_in_page(pos); 1137 unsigned size = min_t(u64, PAGE_SIZE - offset, length); 1138 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); 1139 long rc; 1140 int id; 1141 1142 id = dax_read_lock(); 1143 if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE) 1144 rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1); 1145 else 1146 rc = dax_memzero(iomap->dax_dev, pgoff, offset, size); 1147 dax_read_unlock(id); 1148 1149 if (rc < 0) 1150 return rc; 1151 pos += size; 1152 length -= size; 1153 written += size; 1154 if (did_zero) 1155 *did_zero = true; 1156 } while (length > 0); 1157 1158 return written; 1159 } 1160 1161 int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero, 1162 const struct iomap_ops *ops) 1163 { 1164 struct iomap_iter iter = { 1165 .inode = inode, 1166 .pos = pos, 1167 .len = len, 1168 .flags = IOMAP_DAX | IOMAP_ZERO, 1169 }; 1170 int ret; 1171 1172 while ((ret = iomap_iter(&iter, ops)) > 0) 1173 iter.processed = dax_zero_iter(&iter, did_zero); 1174 return ret; 1175 } 1176 EXPORT_SYMBOL_GPL(dax_zero_range); 1177 1178 int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero, 1179 const struct iomap_ops *ops) 1180 { 1181 unsigned int blocksize = i_blocksize(inode); 1182 unsigned int off = pos & (blocksize - 1); 1183 1184 /* Block boundary? Nothing to do */ 1185 if (!off) 1186 return 0; 1187 return dax_zero_range(inode, pos, blocksize - off, did_zero, ops); 1188 } 1189 EXPORT_SYMBOL_GPL(dax_truncate_page); 1190 1191 static loff_t dax_iomap_iter(const struct iomap_iter *iomi, 1192 struct iov_iter *iter) 1193 { 1194 const struct iomap *iomap = &iomi->iomap; 1195 loff_t length = iomap_length(iomi); 1196 loff_t pos = iomi->pos; 1197 struct dax_device *dax_dev = iomap->dax_dev; 1198 loff_t end = pos + length, done = 0; 1199 ssize_t ret = 0; 1200 size_t xfer; 1201 int id; 1202 1203 if (iov_iter_rw(iter) == READ) { 1204 end = min(end, i_size_read(iomi->inode)); 1205 if (pos >= end) 1206 return 0; 1207 1208 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) 1209 return iov_iter_zero(min(length, end - pos), iter); 1210 } 1211 1212 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED)) 1213 return -EIO; 1214 1215 /* 1216 * Write can allocate block for an area which has a hole page mapped 1217 * into page tables. We have to tear down these mappings so that data 1218 * written by write(2) is visible in mmap. 1219 */ 1220 if (iomap->flags & IOMAP_F_NEW) { 1221 invalidate_inode_pages2_range(iomi->inode->i_mapping, 1222 pos >> PAGE_SHIFT, 1223 (end - 1) >> PAGE_SHIFT); 1224 } 1225 1226 id = dax_read_lock(); 1227 while (pos < end) { 1228 unsigned offset = pos & (PAGE_SIZE - 1); 1229 const size_t size = ALIGN(length + offset, PAGE_SIZE); 1230 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); 1231 ssize_t map_len; 1232 bool recovery = false; 1233 void *kaddr; 1234 1235 if (fatal_signal_pending(current)) { 1236 ret = -EINTR; 1237 break; 1238 } 1239 1240 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), 1241 DAX_ACCESS, &kaddr, NULL); 1242 if (map_len == -EIO && iov_iter_rw(iter) == WRITE) { 1243 map_len = dax_direct_access(dax_dev, pgoff, 1244 PHYS_PFN(size), DAX_RECOVERY_WRITE, 1245 &kaddr, NULL); 1246 if (map_len > 0) 1247 recovery = true; 1248 } 1249 if (map_len < 0) { 1250 ret = map_len; 1251 break; 1252 } 1253 1254 map_len = PFN_PHYS(map_len); 1255 kaddr += offset; 1256 map_len -= offset; 1257 if (map_len > end - pos) 1258 map_len = end - pos; 1259 1260 if (recovery) 1261 xfer = dax_recovery_write(dax_dev, pgoff, kaddr, 1262 map_len, iter); 1263 else if (iov_iter_rw(iter) == WRITE) 1264 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr, 1265 map_len, iter); 1266 else 1267 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr, 1268 map_len, iter); 1269 1270 pos += xfer; 1271 length -= xfer; 1272 done += xfer; 1273 1274 if (xfer == 0) 1275 ret = -EFAULT; 1276 if (xfer < map_len) 1277 break; 1278 } 1279 dax_read_unlock(id); 1280 1281 return done ? done : ret; 1282 } 1283 1284 /** 1285 * dax_iomap_rw - Perform I/O to a DAX file 1286 * @iocb: The control block for this I/O 1287 * @iter: The addresses to do I/O from or to 1288 * @ops: iomap ops passed from the file system 1289 * 1290 * This function performs read and write operations to directly mapped 1291 * persistent memory. The callers needs to take care of read/write exclusion 1292 * and evicting any page cache pages in the region under I/O. 1293 */ 1294 ssize_t 1295 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, 1296 const struct iomap_ops *ops) 1297 { 1298 struct iomap_iter iomi = { 1299 .inode = iocb->ki_filp->f_mapping->host, 1300 .pos = iocb->ki_pos, 1301 .len = iov_iter_count(iter), 1302 .flags = IOMAP_DAX, 1303 }; 1304 loff_t done = 0; 1305 int ret; 1306 1307 if (iov_iter_rw(iter) == WRITE) { 1308 lockdep_assert_held_write(&iomi.inode->i_rwsem); 1309 iomi.flags |= IOMAP_WRITE; 1310 } else { 1311 lockdep_assert_held(&iomi.inode->i_rwsem); 1312 } 1313 1314 if (iocb->ki_flags & IOCB_NOWAIT) 1315 iomi.flags |= IOMAP_NOWAIT; 1316 1317 while ((ret = iomap_iter(&iomi, ops)) > 0) 1318 iomi.processed = dax_iomap_iter(&iomi, iter); 1319 1320 done = iomi.pos - iocb->ki_pos; 1321 iocb->ki_pos = iomi.pos; 1322 return done ? done : ret; 1323 } 1324 EXPORT_SYMBOL_GPL(dax_iomap_rw); 1325 1326 static vm_fault_t dax_fault_return(int error) 1327 { 1328 if (error == 0) 1329 return VM_FAULT_NOPAGE; 1330 return vmf_error(error); 1331 } 1332 1333 /* 1334 * MAP_SYNC on a dax mapping guarantees dirty metadata is 1335 * flushed on write-faults (non-cow), but not read-faults. 1336 */ 1337 static bool dax_fault_is_synchronous(unsigned long flags, 1338 struct vm_area_struct *vma, const struct iomap *iomap) 1339 { 1340 return (flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) 1341 && (iomap->flags & IOMAP_F_DIRTY); 1342 } 1343 1344 /* 1345 * When handling a synchronous page fault and the inode need a fsync, we can 1346 * insert the PTE/PMD into page tables only after that fsync happened. Skip 1347 * insertion for now and return the pfn so that caller can insert it after the 1348 * fsync is done. 1349 */ 1350 static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn) 1351 { 1352 if (WARN_ON_ONCE(!pfnp)) 1353 return VM_FAULT_SIGBUS; 1354 *pfnp = pfn; 1355 return VM_FAULT_NEEDDSYNC; 1356 } 1357 1358 static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf, 1359 const struct iomap_iter *iter) 1360 { 1361 vm_fault_t ret; 1362 int error = 0; 1363 1364 switch (iter->iomap.type) { 1365 case IOMAP_HOLE: 1366 case IOMAP_UNWRITTEN: 1367 clear_user_highpage(vmf->cow_page, vmf->address); 1368 break; 1369 case IOMAP_MAPPED: 1370 error = copy_cow_page_dax(vmf, iter); 1371 break; 1372 default: 1373 WARN_ON_ONCE(1); 1374 error = -EIO; 1375 break; 1376 } 1377 1378 if (error) 1379 return dax_fault_return(error); 1380 1381 __SetPageUptodate(vmf->cow_page); 1382 ret = finish_fault(vmf); 1383 if (!ret) 1384 return VM_FAULT_DONE_COW; 1385 return ret; 1386 } 1387 1388 /** 1389 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault. 1390 * @vmf: vm fault instance 1391 * @iter: iomap iter 1392 * @pfnp: pfn to be returned 1393 * @xas: the dax mapping tree of a file 1394 * @entry: an unlocked dax entry to be inserted 1395 * @pmd: distinguish whether it is a pmd fault 1396 */ 1397 static vm_fault_t dax_fault_iter(struct vm_fault *vmf, 1398 const struct iomap_iter *iter, pfn_t *pfnp, 1399 struct xa_state *xas, void **entry, bool pmd) 1400 { 1401 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1402 const struct iomap *iomap = &iter->iomap; 1403 size_t size = pmd ? PMD_SIZE : PAGE_SIZE; 1404 loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT; 1405 bool write = vmf->flags & FAULT_FLAG_WRITE; 1406 bool sync = dax_fault_is_synchronous(iter->flags, vmf->vma, iomap); 1407 unsigned long entry_flags = pmd ? DAX_PMD : 0; 1408 int err = 0; 1409 pfn_t pfn; 1410 1411 if (!pmd && vmf->cow_page) 1412 return dax_fault_cow_page(vmf, iter); 1413 1414 /* if we are reading UNWRITTEN and HOLE, return a hole. */ 1415 if (!write && 1416 (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) { 1417 if (!pmd) 1418 return dax_load_hole(xas, mapping, entry, vmf); 1419 return dax_pmd_load_hole(xas, vmf, iomap, entry); 1420 } 1421 1422 if (iomap->type != IOMAP_MAPPED) { 1423 WARN_ON_ONCE(1); 1424 return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS; 1425 } 1426 1427 err = dax_iomap_pfn(&iter->iomap, pos, size, &pfn); 1428 if (err) 1429 return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err); 1430 1431 *entry = dax_insert_entry(xas, mapping, vmf, *entry, pfn, entry_flags, 1432 write && !sync); 1433 1434 if (sync) 1435 return dax_fault_synchronous_pfnp(pfnp, pfn); 1436 1437 /* insert PMD pfn */ 1438 if (pmd) 1439 return vmf_insert_pfn_pmd(vmf, pfn, write); 1440 1441 /* insert PTE pfn */ 1442 if (write) 1443 return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); 1444 return vmf_insert_mixed(vmf->vma, vmf->address, pfn); 1445 } 1446 1447 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp, 1448 int *iomap_errp, const struct iomap_ops *ops) 1449 { 1450 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1451 XA_STATE(xas, &mapping->i_pages, vmf->pgoff); 1452 struct iomap_iter iter = { 1453 .inode = mapping->host, 1454 .pos = (loff_t)vmf->pgoff << PAGE_SHIFT, 1455 .len = PAGE_SIZE, 1456 .flags = IOMAP_DAX | IOMAP_FAULT, 1457 }; 1458 vm_fault_t ret = 0; 1459 void *entry; 1460 int error; 1461 1462 trace_dax_pte_fault(iter.inode, vmf, ret); 1463 /* 1464 * Check whether offset isn't beyond end of file now. Caller is supposed 1465 * to hold locks serializing us with truncate / punch hole so this is 1466 * a reliable test. 1467 */ 1468 if (iter.pos >= i_size_read(iter.inode)) { 1469 ret = VM_FAULT_SIGBUS; 1470 goto out; 1471 } 1472 1473 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) 1474 iter.flags |= IOMAP_WRITE; 1475 1476 entry = grab_mapping_entry(&xas, mapping, 0); 1477 if (xa_is_internal(entry)) { 1478 ret = xa_to_internal(entry); 1479 goto out; 1480 } 1481 1482 /* 1483 * It is possible, particularly with mixed reads & writes to private 1484 * mappings, that we have raced with a PMD fault that overlaps with 1485 * the PTE we need to set up. If so just return and the fault will be 1486 * retried. 1487 */ 1488 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) { 1489 ret = VM_FAULT_NOPAGE; 1490 goto unlock_entry; 1491 } 1492 1493 while ((error = iomap_iter(&iter, ops)) > 0) { 1494 if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) { 1495 iter.processed = -EIO; /* fs corruption? */ 1496 continue; 1497 } 1498 1499 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false); 1500 if (ret != VM_FAULT_SIGBUS && 1501 (iter.iomap.flags & IOMAP_F_NEW)) { 1502 count_vm_event(PGMAJFAULT); 1503 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT); 1504 ret |= VM_FAULT_MAJOR; 1505 } 1506 1507 if (!(ret & VM_FAULT_ERROR)) 1508 iter.processed = PAGE_SIZE; 1509 } 1510 1511 if (iomap_errp) 1512 *iomap_errp = error; 1513 if (!ret && error) 1514 ret = dax_fault_return(error); 1515 1516 unlock_entry: 1517 dax_unlock_entry(&xas, entry); 1518 out: 1519 trace_dax_pte_fault_done(iter.inode, vmf, ret); 1520 return ret; 1521 } 1522 1523 #ifdef CONFIG_FS_DAX_PMD 1524 static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas, 1525 pgoff_t max_pgoff) 1526 { 1527 unsigned long pmd_addr = vmf->address & PMD_MASK; 1528 bool write = vmf->flags & FAULT_FLAG_WRITE; 1529 1530 /* 1531 * Make sure that the faulting address's PMD offset (color) matches 1532 * the PMD offset from the start of the file. This is necessary so 1533 * that a PMD range in the page table overlaps exactly with a PMD 1534 * range in the page cache. 1535 */ 1536 if ((vmf->pgoff & PG_PMD_COLOUR) != 1537 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR)) 1538 return true; 1539 1540 /* Fall back to PTEs if we're going to COW */ 1541 if (write && !(vmf->vma->vm_flags & VM_SHARED)) 1542 return true; 1543 1544 /* If the PMD would extend outside the VMA */ 1545 if (pmd_addr < vmf->vma->vm_start) 1546 return true; 1547 if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end) 1548 return true; 1549 1550 /* If the PMD would extend beyond the file size */ 1551 if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff) 1552 return true; 1553 1554 return false; 1555 } 1556 1557 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, 1558 const struct iomap_ops *ops) 1559 { 1560 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1561 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER); 1562 struct iomap_iter iter = { 1563 .inode = mapping->host, 1564 .len = PMD_SIZE, 1565 .flags = IOMAP_DAX | IOMAP_FAULT, 1566 }; 1567 vm_fault_t ret = VM_FAULT_FALLBACK; 1568 pgoff_t max_pgoff; 1569 void *entry; 1570 int error; 1571 1572 if (vmf->flags & FAULT_FLAG_WRITE) 1573 iter.flags |= IOMAP_WRITE; 1574 1575 /* 1576 * Check whether offset isn't beyond end of file now. Caller is 1577 * supposed to hold locks serializing us with truncate / punch hole so 1578 * this is a reliable test. 1579 */ 1580 max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE); 1581 1582 trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0); 1583 1584 if (xas.xa_index >= max_pgoff) { 1585 ret = VM_FAULT_SIGBUS; 1586 goto out; 1587 } 1588 1589 if (dax_fault_check_fallback(vmf, &xas, max_pgoff)) 1590 goto fallback; 1591 1592 /* 1593 * grab_mapping_entry() will make sure we get an empty PMD entry, 1594 * a zero PMD entry or a DAX PMD. If it can't (because a PTE 1595 * entry is already in the array, for instance), it will return 1596 * VM_FAULT_FALLBACK. 1597 */ 1598 entry = grab_mapping_entry(&xas, mapping, PMD_ORDER); 1599 if (xa_is_internal(entry)) { 1600 ret = xa_to_internal(entry); 1601 goto fallback; 1602 } 1603 1604 /* 1605 * It is possible, particularly with mixed reads & writes to private 1606 * mappings, that we have raced with a PTE fault that overlaps with 1607 * the PMD we need to set up. If so just return and the fault will be 1608 * retried. 1609 */ 1610 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) && 1611 !pmd_devmap(*vmf->pmd)) { 1612 ret = 0; 1613 goto unlock_entry; 1614 } 1615 1616 iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT; 1617 while ((error = iomap_iter(&iter, ops)) > 0) { 1618 if (iomap_length(&iter) < PMD_SIZE) 1619 continue; /* actually breaks out of the loop */ 1620 1621 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true); 1622 if (ret != VM_FAULT_FALLBACK) 1623 iter.processed = PMD_SIZE; 1624 } 1625 1626 unlock_entry: 1627 dax_unlock_entry(&xas, entry); 1628 fallback: 1629 if (ret == VM_FAULT_FALLBACK) { 1630 split_huge_pmd(vmf->vma, vmf->pmd, vmf->address); 1631 count_vm_event(THP_FAULT_FALLBACK); 1632 } 1633 out: 1634 trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret); 1635 return ret; 1636 } 1637 #else 1638 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, 1639 const struct iomap_ops *ops) 1640 { 1641 return VM_FAULT_FALLBACK; 1642 } 1643 #endif /* CONFIG_FS_DAX_PMD */ 1644 1645 /** 1646 * dax_iomap_fault - handle a page fault on a DAX file 1647 * @vmf: The description of the fault 1648 * @pe_size: Size of the page to fault in 1649 * @pfnp: PFN to insert for synchronous faults if fsync is required 1650 * @iomap_errp: Storage for detailed error code in case of error 1651 * @ops: Iomap ops passed from the file system 1652 * 1653 * When a page fault occurs, filesystems may call this helper in 1654 * their fault handler for DAX files. dax_iomap_fault() assumes the caller 1655 * has done all the necessary locking for page fault to proceed 1656 * successfully. 1657 */ 1658 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, 1659 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops) 1660 { 1661 switch (pe_size) { 1662 case PE_SIZE_PTE: 1663 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops); 1664 case PE_SIZE_PMD: 1665 return dax_iomap_pmd_fault(vmf, pfnp, ops); 1666 default: 1667 return VM_FAULT_FALLBACK; 1668 } 1669 } 1670 EXPORT_SYMBOL_GPL(dax_iomap_fault); 1671 1672 /* 1673 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables 1674 * @vmf: The description of the fault 1675 * @pfn: PFN to insert 1676 * @order: Order of entry to insert. 1677 * 1678 * This function inserts a writeable PTE or PMD entry into the page tables 1679 * for an mmaped DAX file. It also marks the page cache entry as dirty. 1680 */ 1681 static vm_fault_t 1682 dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order) 1683 { 1684 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1685 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order); 1686 void *entry; 1687 vm_fault_t ret; 1688 1689 xas_lock_irq(&xas); 1690 entry = get_unlocked_entry(&xas, order); 1691 /* Did we race with someone splitting entry or so? */ 1692 if (!entry || dax_is_conflict(entry) || 1693 (order == 0 && !dax_is_pte_entry(entry))) { 1694 put_unlocked_entry(&xas, entry, WAKE_NEXT); 1695 xas_unlock_irq(&xas); 1696 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf, 1697 VM_FAULT_NOPAGE); 1698 return VM_FAULT_NOPAGE; 1699 } 1700 xas_set_mark(&xas, PAGECACHE_TAG_DIRTY); 1701 dax_lock_entry(&xas, entry); 1702 xas_unlock_irq(&xas); 1703 if (order == 0) 1704 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); 1705 #ifdef CONFIG_FS_DAX_PMD 1706 else if (order == PMD_ORDER) 1707 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE); 1708 #endif 1709 else 1710 ret = VM_FAULT_FALLBACK; 1711 dax_unlock_entry(&xas, entry); 1712 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret); 1713 return ret; 1714 } 1715 1716 /** 1717 * dax_finish_sync_fault - finish synchronous page fault 1718 * @vmf: The description of the fault 1719 * @pe_size: Size of entry to be inserted 1720 * @pfn: PFN to insert 1721 * 1722 * This function ensures that the file range touched by the page fault is 1723 * stored persistently on the media and handles inserting of appropriate page 1724 * table entry. 1725 */ 1726 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, 1727 enum page_entry_size pe_size, pfn_t pfn) 1728 { 1729 int err; 1730 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT; 1731 unsigned int order = pe_order(pe_size); 1732 size_t len = PAGE_SIZE << order; 1733 1734 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1); 1735 if (err) 1736 return VM_FAULT_SIGBUS; 1737 return dax_insert_pfn_mkwrite(vmf, pfn, order); 1738 } 1739 EXPORT_SYMBOL_GPL(dax_finish_sync_fault); 1740