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 static inline bool dax_page_is_shared(struct page *page) 338 { 339 return page->mapping == PAGE_MAPPING_DAX_SHARED; 340 } 341 342 /* 343 * Set the page->mapping with PAGE_MAPPING_DAX_SHARED flag, increase the 344 * refcount. 345 */ 346 static inline void dax_page_share_get(struct page *page) 347 { 348 if (page->mapping != PAGE_MAPPING_DAX_SHARED) { 349 /* 350 * Reset the index if the page was already mapped 351 * regularly before. 352 */ 353 if (page->mapping) 354 page->share = 1; 355 page->mapping = PAGE_MAPPING_DAX_SHARED; 356 } 357 page->share++; 358 } 359 360 static inline unsigned long dax_page_share_put(struct page *page) 361 { 362 return --page->share; 363 } 364 365 /* 366 * When it is called in dax_insert_entry(), the shared flag will indicate that 367 * whether this entry is shared by multiple files. If so, set the page->mapping 368 * PAGE_MAPPING_DAX_SHARED, and use page->share as refcount. 369 */ 370 static void dax_associate_entry(void *entry, struct address_space *mapping, 371 struct vm_area_struct *vma, unsigned long address, bool shared) 372 { 373 unsigned long size = dax_entry_size(entry), pfn, index; 374 int i = 0; 375 376 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 377 return; 378 379 index = linear_page_index(vma, address & ~(size - 1)); 380 for_each_mapped_pfn(entry, pfn) { 381 struct page *page = pfn_to_page(pfn); 382 383 if (shared) { 384 dax_page_share_get(page); 385 } else { 386 WARN_ON_ONCE(page->mapping); 387 page->mapping = mapping; 388 page->index = index + i++; 389 } 390 } 391 } 392 393 static void dax_disassociate_entry(void *entry, struct address_space *mapping, 394 bool trunc) 395 { 396 unsigned long pfn; 397 398 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 399 return; 400 401 for_each_mapped_pfn(entry, pfn) { 402 struct page *page = pfn_to_page(pfn); 403 404 WARN_ON_ONCE(trunc && page_ref_count(page) > 1); 405 if (dax_page_is_shared(page)) { 406 /* keep the shared flag if this page is still shared */ 407 if (dax_page_share_put(page) > 0) 408 continue; 409 } else 410 WARN_ON_ONCE(page->mapping && page->mapping != mapping); 411 page->mapping = NULL; 412 page->index = 0; 413 } 414 } 415 416 static struct page *dax_busy_page(void *entry) 417 { 418 unsigned long pfn; 419 420 for_each_mapped_pfn(entry, pfn) { 421 struct page *page = pfn_to_page(pfn); 422 423 if (page_ref_count(page) > 1) 424 return page; 425 } 426 return NULL; 427 } 428 429 /* 430 * dax_lock_page - Lock the DAX entry corresponding to a page 431 * @page: The page whose entry we want to lock 432 * 433 * Context: Process context. 434 * Return: A cookie to pass to dax_unlock_page() or 0 if the entry could 435 * not be locked. 436 */ 437 dax_entry_t dax_lock_page(struct page *page) 438 { 439 XA_STATE(xas, NULL, 0); 440 void *entry; 441 442 /* Ensure page->mapping isn't freed while we look at it */ 443 rcu_read_lock(); 444 for (;;) { 445 struct address_space *mapping = READ_ONCE(page->mapping); 446 447 entry = NULL; 448 if (!mapping || !dax_mapping(mapping)) 449 break; 450 451 /* 452 * In the device-dax case there's no need to lock, a 453 * struct dev_pagemap pin is sufficient to keep the 454 * inode alive, and we assume we have dev_pagemap pin 455 * otherwise we would not have a valid pfn_to_page() 456 * translation. 457 */ 458 entry = (void *)~0UL; 459 if (S_ISCHR(mapping->host->i_mode)) 460 break; 461 462 xas.xa = &mapping->i_pages; 463 xas_lock_irq(&xas); 464 if (mapping != page->mapping) { 465 xas_unlock_irq(&xas); 466 continue; 467 } 468 xas_set(&xas, page->index); 469 entry = xas_load(&xas); 470 if (dax_is_locked(entry)) { 471 rcu_read_unlock(); 472 wait_entry_unlocked(&xas, entry); 473 rcu_read_lock(); 474 continue; 475 } 476 dax_lock_entry(&xas, entry); 477 xas_unlock_irq(&xas); 478 break; 479 } 480 rcu_read_unlock(); 481 return (dax_entry_t)entry; 482 } 483 484 void dax_unlock_page(struct page *page, dax_entry_t cookie) 485 { 486 struct address_space *mapping = page->mapping; 487 XA_STATE(xas, &mapping->i_pages, page->index); 488 489 if (S_ISCHR(mapping->host->i_mode)) 490 return; 491 492 dax_unlock_entry(&xas, (void *)cookie); 493 } 494 495 /* 496 * dax_lock_mapping_entry - Lock the DAX entry corresponding to a mapping 497 * @mapping: the file's mapping whose entry we want to lock 498 * @index: the offset within this file 499 * @page: output the dax page corresponding to this dax entry 500 * 501 * Return: A cookie to pass to dax_unlock_mapping_entry() or 0 if the entry 502 * could not be locked. 503 */ 504 dax_entry_t dax_lock_mapping_entry(struct address_space *mapping, pgoff_t index, 505 struct page **page) 506 { 507 XA_STATE(xas, NULL, 0); 508 void *entry; 509 510 rcu_read_lock(); 511 for (;;) { 512 entry = NULL; 513 if (!dax_mapping(mapping)) 514 break; 515 516 xas.xa = &mapping->i_pages; 517 xas_lock_irq(&xas); 518 xas_set(&xas, index); 519 entry = xas_load(&xas); 520 if (dax_is_locked(entry)) { 521 rcu_read_unlock(); 522 wait_entry_unlocked(&xas, entry); 523 rcu_read_lock(); 524 continue; 525 } 526 if (!entry || 527 dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { 528 /* 529 * Because we are looking for entry from file's mapping 530 * and index, so the entry may not be inserted for now, 531 * or even a zero/empty entry. We don't think this is 532 * an error case. So, return a special value and do 533 * not output @page. 534 */ 535 entry = (void *)~0UL; 536 } else { 537 *page = pfn_to_page(dax_to_pfn(entry)); 538 dax_lock_entry(&xas, entry); 539 } 540 xas_unlock_irq(&xas); 541 break; 542 } 543 rcu_read_unlock(); 544 return (dax_entry_t)entry; 545 } 546 547 void dax_unlock_mapping_entry(struct address_space *mapping, pgoff_t index, 548 dax_entry_t cookie) 549 { 550 XA_STATE(xas, &mapping->i_pages, index); 551 552 if (cookie == ~0UL) 553 return; 554 555 dax_unlock_entry(&xas, (void *)cookie); 556 } 557 558 /* 559 * Find page cache entry at given index. If it is a DAX entry, return it 560 * with the entry locked. If the page cache doesn't contain an entry at 561 * that index, add a locked empty entry. 562 * 563 * When requesting an entry with size DAX_PMD, grab_mapping_entry() will 564 * either return that locked entry or will return VM_FAULT_FALLBACK. 565 * This will happen if there are any PTE entries within the PMD range 566 * that we are requesting. 567 * 568 * We always favor PTE entries over PMD entries. There isn't a flow where we 569 * evict PTE entries in order to 'upgrade' them to a PMD entry. A PMD 570 * insertion will fail if it finds any PTE entries already in the tree, and a 571 * PTE insertion will cause an existing PMD entry to be unmapped and 572 * downgraded to PTE entries. This happens for both PMD zero pages as 573 * well as PMD empty entries. 574 * 575 * The exception to this downgrade path is for PMD entries that have 576 * real storage backing them. We will leave these real PMD entries in 577 * the tree, and PTE writes will simply dirty the entire PMD entry. 578 * 579 * Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For 580 * persistent memory the benefit is doubtful. We can add that later if we can 581 * show it helps. 582 * 583 * On error, this function does not return an ERR_PTR. Instead it returns 584 * a VM_FAULT code, encoded as an xarray internal entry. The ERR_PTR values 585 * overlap with xarray value entries. 586 */ 587 static void *grab_mapping_entry(struct xa_state *xas, 588 struct address_space *mapping, unsigned int order) 589 { 590 unsigned long index = xas->xa_index; 591 bool pmd_downgrade; /* splitting PMD entry into PTE entries? */ 592 void *entry; 593 594 retry: 595 pmd_downgrade = false; 596 xas_lock_irq(xas); 597 entry = get_unlocked_entry(xas, order); 598 599 if (entry) { 600 if (dax_is_conflict(entry)) 601 goto fallback; 602 if (!xa_is_value(entry)) { 603 xas_set_err(xas, -EIO); 604 goto out_unlock; 605 } 606 607 if (order == 0) { 608 if (dax_is_pmd_entry(entry) && 609 (dax_is_zero_entry(entry) || 610 dax_is_empty_entry(entry))) { 611 pmd_downgrade = true; 612 } 613 } 614 } 615 616 if (pmd_downgrade) { 617 /* 618 * Make sure 'entry' remains valid while we drop 619 * the i_pages lock. 620 */ 621 dax_lock_entry(xas, entry); 622 623 /* 624 * Besides huge zero pages the only other thing that gets 625 * downgraded are empty entries which don't need to be 626 * unmapped. 627 */ 628 if (dax_is_zero_entry(entry)) { 629 xas_unlock_irq(xas); 630 unmap_mapping_pages(mapping, 631 xas->xa_index & ~PG_PMD_COLOUR, 632 PG_PMD_NR, false); 633 xas_reset(xas); 634 xas_lock_irq(xas); 635 } 636 637 dax_disassociate_entry(entry, mapping, false); 638 xas_store(xas, NULL); /* undo the PMD join */ 639 dax_wake_entry(xas, entry, WAKE_ALL); 640 mapping->nrpages -= PG_PMD_NR; 641 entry = NULL; 642 xas_set(xas, index); 643 } 644 645 if (entry) { 646 dax_lock_entry(xas, entry); 647 } else { 648 unsigned long flags = DAX_EMPTY; 649 650 if (order > 0) 651 flags |= DAX_PMD; 652 entry = dax_make_entry(pfn_to_pfn_t(0), flags); 653 dax_lock_entry(xas, entry); 654 if (xas_error(xas)) 655 goto out_unlock; 656 mapping->nrpages += 1UL << order; 657 } 658 659 out_unlock: 660 xas_unlock_irq(xas); 661 if (xas_nomem(xas, mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM)) 662 goto retry; 663 if (xas->xa_node == XA_ERROR(-ENOMEM)) 664 return xa_mk_internal(VM_FAULT_OOM); 665 if (xas_error(xas)) 666 return xa_mk_internal(VM_FAULT_SIGBUS); 667 return entry; 668 fallback: 669 xas_unlock_irq(xas); 670 return xa_mk_internal(VM_FAULT_FALLBACK); 671 } 672 673 /** 674 * dax_layout_busy_page_range - find first pinned page in @mapping 675 * @mapping: address space to scan for a page with ref count > 1 676 * @start: Starting offset. Page containing 'start' is included. 677 * @end: End offset. Page containing 'end' is included. If 'end' is LLONG_MAX, 678 * pages from 'start' till the end of file are included. 679 * 680 * DAX requires ZONE_DEVICE mapped pages. These pages are never 681 * 'onlined' to the page allocator so they are considered idle when 682 * page->count == 1. A filesystem uses this interface to determine if 683 * any page in the mapping is busy, i.e. for DMA, or other 684 * get_user_pages() usages. 685 * 686 * It is expected that the filesystem is holding locks to block the 687 * establishment of new mappings in this address_space. I.e. it expects 688 * to be able to run unmap_mapping_range() and subsequently not race 689 * mapping_mapped() becoming true. 690 */ 691 struct page *dax_layout_busy_page_range(struct address_space *mapping, 692 loff_t start, loff_t end) 693 { 694 void *entry; 695 unsigned int scanned = 0; 696 struct page *page = NULL; 697 pgoff_t start_idx = start >> PAGE_SHIFT; 698 pgoff_t end_idx; 699 XA_STATE(xas, &mapping->i_pages, start_idx); 700 701 /* 702 * In the 'limited' case get_user_pages() for dax is disabled. 703 */ 704 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) 705 return NULL; 706 707 if (!dax_mapping(mapping) || !mapping_mapped(mapping)) 708 return NULL; 709 710 /* If end == LLONG_MAX, all pages from start to till end of file */ 711 if (end == LLONG_MAX) 712 end_idx = ULONG_MAX; 713 else 714 end_idx = end >> PAGE_SHIFT; 715 /* 716 * If we race get_user_pages_fast() here either we'll see the 717 * elevated page count in the iteration and wait, or 718 * get_user_pages_fast() will see that the page it took a reference 719 * against is no longer mapped in the page tables and bail to the 720 * get_user_pages() slow path. The slow path is protected by 721 * pte_lock() and pmd_lock(). New references are not taken without 722 * holding those locks, and unmap_mapping_pages() will not zero the 723 * pte or pmd without holding the respective lock, so we are 724 * guaranteed to either see new references or prevent new 725 * references from being established. 726 */ 727 unmap_mapping_pages(mapping, start_idx, end_idx - start_idx + 1, 0); 728 729 xas_lock_irq(&xas); 730 xas_for_each(&xas, entry, end_idx) { 731 if (WARN_ON_ONCE(!xa_is_value(entry))) 732 continue; 733 if (unlikely(dax_is_locked(entry))) 734 entry = get_unlocked_entry(&xas, 0); 735 if (entry) 736 page = dax_busy_page(entry); 737 put_unlocked_entry(&xas, entry, WAKE_NEXT); 738 if (page) 739 break; 740 if (++scanned % XA_CHECK_SCHED) 741 continue; 742 743 xas_pause(&xas); 744 xas_unlock_irq(&xas); 745 cond_resched(); 746 xas_lock_irq(&xas); 747 } 748 xas_unlock_irq(&xas); 749 return page; 750 } 751 EXPORT_SYMBOL_GPL(dax_layout_busy_page_range); 752 753 struct page *dax_layout_busy_page(struct address_space *mapping) 754 { 755 return dax_layout_busy_page_range(mapping, 0, LLONG_MAX); 756 } 757 EXPORT_SYMBOL_GPL(dax_layout_busy_page); 758 759 static int __dax_invalidate_entry(struct address_space *mapping, 760 pgoff_t index, bool trunc) 761 { 762 XA_STATE(xas, &mapping->i_pages, index); 763 int ret = 0; 764 void *entry; 765 766 xas_lock_irq(&xas); 767 entry = get_unlocked_entry(&xas, 0); 768 if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) 769 goto out; 770 if (!trunc && 771 (xas_get_mark(&xas, PAGECACHE_TAG_DIRTY) || 772 xas_get_mark(&xas, PAGECACHE_TAG_TOWRITE))) 773 goto out; 774 dax_disassociate_entry(entry, mapping, trunc); 775 xas_store(&xas, NULL); 776 mapping->nrpages -= 1UL << dax_entry_order(entry); 777 ret = 1; 778 out: 779 put_unlocked_entry(&xas, entry, WAKE_ALL); 780 xas_unlock_irq(&xas); 781 return ret; 782 } 783 784 /* 785 * Delete DAX entry at @index from @mapping. Wait for it 786 * to be unlocked before deleting it. 787 */ 788 int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) 789 { 790 int ret = __dax_invalidate_entry(mapping, index, true); 791 792 /* 793 * This gets called from truncate / punch_hole path. As such, the caller 794 * must hold locks protecting against concurrent modifications of the 795 * page cache (usually fs-private i_mmap_sem for writing). Since the 796 * caller has seen a DAX entry for this index, we better find it 797 * at that index as well... 798 */ 799 WARN_ON_ONCE(!ret); 800 return ret; 801 } 802 803 /* 804 * Invalidate DAX entry if it is clean. 805 */ 806 int dax_invalidate_mapping_entry_sync(struct address_space *mapping, 807 pgoff_t index) 808 { 809 return __dax_invalidate_entry(mapping, index, false); 810 } 811 812 static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos) 813 { 814 return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset); 815 } 816 817 static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter) 818 { 819 pgoff_t pgoff = dax_iomap_pgoff(&iter->iomap, iter->pos); 820 void *vto, *kaddr; 821 long rc; 822 int id; 823 824 id = dax_read_lock(); 825 rc = dax_direct_access(iter->iomap.dax_dev, pgoff, 1, DAX_ACCESS, 826 &kaddr, NULL); 827 if (rc < 0) { 828 dax_read_unlock(id); 829 return rc; 830 } 831 vto = kmap_atomic(vmf->cow_page); 832 copy_user_page(vto, kaddr, vmf->address, vmf->cow_page); 833 kunmap_atomic(vto); 834 dax_read_unlock(id); 835 return 0; 836 } 837 838 /* 839 * MAP_SYNC on a dax mapping guarantees dirty metadata is 840 * flushed on write-faults (non-cow), but not read-faults. 841 */ 842 static bool dax_fault_is_synchronous(const struct iomap_iter *iter, 843 struct vm_area_struct *vma) 844 { 845 return (iter->flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) && 846 (iter->iomap.flags & IOMAP_F_DIRTY); 847 } 848 849 /* 850 * By this point grab_mapping_entry() has ensured that we have a locked entry 851 * of the appropriate size so we don't have to worry about downgrading PMDs to 852 * PTEs. If we happen to be trying to insert a PTE and there is a PMD 853 * already in the tree, we will skip the insertion and just dirty the PMD as 854 * appropriate. 855 */ 856 static void *dax_insert_entry(struct xa_state *xas, struct vm_fault *vmf, 857 const struct iomap_iter *iter, void *entry, pfn_t pfn, 858 unsigned long flags) 859 { 860 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 861 void *new_entry = dax_make_entry(pfn, flags); 862 bool write = iter->flags & IOMAP_WRITE; 863 bool dirty = write && !dax_fault_is_synchronous(iter, vmf->vma); 864 bool shared = iter->iomap.flags & IOMAP_F_SHARED; 865 866 if (dirty) 867 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); 868 869 if (shared || (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE))) { 870 unsigned long index = xas->xa_index; 871 /* we are replacing a zero page with block mapping */ 872 if (dax_is_pmd_entry(entry)) 873 unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR, 874 PG_PMD_NR, false); 875 else /* pte entry */ 876 unmap_mapping_pages(mapping, index, 1, false); 877 } 878 879 xas_reset(xas); 880 xas_lock_irq(xas); 881 if (shared || dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { 882 void *old; 883 884 dax_disassociate_entry(entry, mapping, false); 885 dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address, 886 shared); 887 /* 888 * Only swap our new entry into the page cache if the current 889 * entry is a zero page or an empty entry. If a normal PTE or 890 * PMD entry is already in the cache, we leave it alone. This 891 * means that if we are trying to insert a PTE and the 892 * existing entry is a PMD, we will just leave the PMD in the 893 * tree and dirty it if necessary. 894 */ 895 old = dax_lock_entry(xas, new_entry); 896 WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) | 897 DAX_LOCKED)); 898 entry = new_entry; 899 } else { 900 xas_load(xas); /* Walk the xa_state */ 901 } 902 903 if (dirty) 904 xas_set_mark(xas, PAGECACHE_TAG_DIRTY); 905 906 if (write && shared) 907 xas_set_mark(xas, PAGECACHE_TAG_TOWRITE); 908 909 xas_unlock_irq(xas); 910 return entry; 911 } 912 913 static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev, 914 struct address_space *mapping, void *entry) 915 { 916 unsigned long pfn, index, count, end; 917 long ret = 0; 918 struct vm_area_struct *vma; 919 920 /* 921 * A page got tagged dirty in DAX mapping? Something is seriously 922 * wrong. 923 */ 924 if (WARN_ON(!xa_is_value(entry))) 925 return -EIO; 926 927 if (unlikely(dax_is_locked(entry))) { 928 void *old_entry = entry; 929 930 entry = get_unlocked_entry(xas, 0); 931 932 /* Entry got punched out / reallocated? */ 933 if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) 934 goto put_unlocked; 935 /* 936 * Entry got reallocated elsewhere? No need to writeback. 937 * We have to compare pfns as we must not bail out due to 938 * difference in lockbit or entry type. 939 */ 940 if (dax_to_pfn(old_entry) != dax_to_pfn(entry)) 941 goto put_unlocked; 942 if (WARN_ON_ONCE(dax_is_empty_entry(entry) || 943 dax_is_zero_entry(entry))) { 944 ret = -EIO; 945 goto put_unlocked; 946 } 947 948 /* Another fsync thread may have already done this entry */ 949 if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE)) 950 goto put_unlocked; 951 } 952 953 /* Lock the entry to serialize with page faults */ 954 dax_lock_entry(xas, entry); 955 956 /* 957 * We can clear the tag now but we have to be careful so that concurrent 958 * dax_writeback_one() calls for the same index cannot finish before we 959 * actually flush the caches. This is achieved as the calls will look 960 * at the entry only under the i_pages lock and once they do that 961 * they will see the entry locked and wait for it to unlock. 962 */ 963 xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE); 964 xas_unlock_irq(xas); 965 966 /* 967 * If dax_writeback_mapping_range() was given a wbc->range_start 968 * in the middle of a PMD, the 'index' we use needs to be 969 * aligned to the start of the PMD. 970 * This allows us to flush for PMD_SIZE and not have to worry about 971 * partial PMD writebacks. 972 */ 973 pfn = dax_to_pfn(entry); 974 count = 1UL << dax_entry_order(entry); 975 index = xas->xa_index & ~(count - 1); 976 end = index + count - 1; 977 978 /* Walk all mappings of a given index of a file and writeprotect them */ 979 i_mmap_lock_read(mapping); 980 vma_interval_tree_foreach(vma, &mapping->i_mmap, index, end) { 981 pfn_mkclean_range(pfn, count, index, vma); 982 cond_resched(); 983 } 984 i_mmap_unlock_read(mapping); 985 986 dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE); 987 /* 988 * After we have flushed the cache, we can clear the dirty tag. There 989 * cannot be new dirty data in the pfn after the flush has completed as 990 * the pfn mappings are writeprotected and fault waits for mapping 991 * entry lock. 992 */ 993 xas_reset(xas); 994 xas_lock_irq(xas); 995 xas_store(xas, entry); 996 xas_clear_mark(xas, PAGECACHE_TAG_DIRTY); 997 dax_wake_entry(xas, entry, WAKE_NEXT); 998 999 trace_dax_writeback_one(mapping->host, index, count); 1000 return ret; 1001 1002 put_unlocked: 1003 put_unlocked_entry(xas, entry, WAKE_NEXT); 1004 return ret; 1005 } 1006 1007 /* 1008 * Flush the mapping to the persistent domain within the byte range of [start, 1009 * end]. This is required by data integrity operations to ensure file data is 1010 * on persistent storage prior to completion of the operation. 1011 */ 1012 int dax_writeback_mapping_range(struct address_space *mapping, 1013 struct dax_device *dax_dev, struct writeback_control *wbc) 1014 { 1015 XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT); 1016 struct inode *inode = mapping->host; 1017 pgoff_t end_index = wbc->range_end >> PAGE_SHIFT; 1018 void *entry; 1019 int ret = 0; 1020 unsigned int scanned = 0; 1021 1022 if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) 1023 return -EIO; 1024 1025 if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL) 1026 return 0; 1027 1028 trace_dax_writeback_range(inode, xas.xa_index, end_index); 1029 1030 tag_pages_for_writeback(mapping, xas.xa_index, end_index); 1031 1032 xas_lock_irq(&xas); 1033 xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) { 1034 ret = dax_writeback_one(&xas, dax_dev, mapping, entry); 1035 if (ret < 0) { 1036 mapping_set_error(mapping, ret); 1037 break; 1038 } 1039 if (++scanned % XA_CHECK_SCHED) 1040 continue; 1041 1042 xas_pause(&xas); 1043 xas_unlock_irq(&xas); 1044 cond_resched(); 1045 xas_lock_irq(&xas); 1046 } 1047 xas_unlock_irq(&xas); 1048 trace_dax_writeback_range_done(inode, xas.xa_index, end_index); 1049 return ret; 1050 } 1051 EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); 1052 1053 static int dax_iomap_direct_access(const struct iomap *iomap, loff_t pos, 1054 size_t size, void **kaddr, pfn_t *pfnp) 1055 { 1056 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); 1057 int id, rc = 0; 1058 long length; 1059 1060 id = dax_read_lock(); 1061 length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size), 1062 DAX_ACCESS, kaddr, pfnp); 1063 if (length < 0) { 1064 rc = length; 1065 goto out; 1066 } 1067 if (!pfnp) 1068 goto out_check_addr; 1069 rc = -EINVAL; 1070 if (PFN_PHYS(length) < size) 1071 goto out; 1072 if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1)) 1073 goto out; 1074 /* For larger pages we need devmap */ 1075 if (length > 1 && !pfn_t_devmap(*pfnp)) 1076 goto out; 1077 rc = 0; 1078 1079 out_check_addr: 1080 if (!kaddr) 1081 goto out; 1082 if (!*kaddr) 1083 rc = -EFAULT; 1084 out: 1085 dax_read_unlock(id); 1086 return rc; 1087 } 1088 1089 /** 1090 * dax_iomap_copy_around - Prepare for an unaligned write to a shared/cow page 1091 * by copying the data before and after the range to be written. 1092 * @pos: address to do copy from. 1093 * @length: size of copy operation. 1094 * @align_size: aligned w.r.t align_size (either PMD_SIZE or PAGE_SIZE) 1095 * @srcmap: iomap srcmap 1096 * @daddr: destination address to copy to. 1097 * 1098 * This can be called from two places. Either during DAX write fault (page 1099 * aligned), to copy the length size data to daddr. Or, while doing normal DAX 1100 * write operation, dax_iomap_iter() might call this to do the copy of either 1101 * start or end unaligned address. In the latter case the rest of the copy of 1102 * aligned ranges is taken care by dax_iomap_iter() itself. 1103 * If the srcmap contains invalid data, such as HOLE and UNWRITTEN, zero the 1104 * area to make sure no old data remains. 1105 */ 1106 static int dax_iomap_copy_around(loff_t pos, uint64_t length, size_t align_size, 1107 const struct iomap *srcmap, void *daddr) 1108 { 1109 loff_t head_off = pos & (align_size - 1); 1110 size_t size = ALIGN(head_off + length, align_size); 1111 loff_t end = pos + length; 1112 loff_t pg_end = round_up(end, align_size); 1113 /* copy_all is usually in page fault case */ 1114 bool copy_all = head_off == 0 && end == pg_end; 1115 /* zero the edges if srcmap is a HOLE or IOMAP_UNWRITTEN */ 1116 bool zero_edge = srcmap->flags & IOMAP_F_SHARED || 1117 srcmap->type == IOMAP_UNWRITTEN; 1118 void *saddr = 0; 1119 int ret = 0; 1120 1121 if (!zero_edge) { 1122 ret = dax_iomap_direct_access(srcmap, pos, size, &saddr, NULL); 1123 if (ret) 1124 return ret; 1125 } 1126 1127 if (copy_all) { 1128 if (zero_edge) 1129 memset(daddr, 0, size); 1130 else 1131 ret = copy_mc_to_kernel(daddr, saddr, length); 1132 goto out; 1133 } 1134 1135 /* Copy the head part of the range */ 1136 if (head_off) { 1137 if (zero_edge) 1138 memset(daddr, 0, head_off); 1139 else { 1140 ret = copy_mc_to_kernel(daddr, saddr, head_off); 1141 if (ret) 1142 return -EIO; 1143 } 1144 } 1145 1146 /* Copy the tail part of the range */ 1147 if (end < pg_end) { 1148 loff_t tail_off = head_off + length; 1149 loff_t tail_len = pg_end - end; 1150 1151 if (zero_edge) 1152 memset(daddr + tail_off, 0, tail_len); 1153 else { 1154 ret = copy_mc_to_kernel(daddr + tail_off, 1155 saddr + tail_off, tail_len); 1156 if (ret) 1157 return -EIO; 1158 } 1159 } 1160 out: 1161 if (zero_edge) 1162 dax_flush(srcmap->dax_dev, daddr, size); 1163 return ret ? -EIO : 0; 1164 } 1165 1166 /* 1167 * The user has performed a load from a hole in the file. Allocating a new 1168 * page in the file would cause excessive storage usage for workloads with 1169 * sparse files. Instead we insert a read-only mapping of the 4k zero page. 1170 * If this page is ever written to we will re-fault and change the mapping to 1171 * point to real DAX storage instead. 1172 */ 1173 static vm_fault_t dax_load_hole(struct xa_state *xas, struct vm_fault *vmf, 1174 const struct iomap_iter *iter, void **entry) 1175 { 1176 struct inode *inode = iter->inode; 1177 unsigned long vaddr = vmf->address; 1178 pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr)); 1179 vm_fault_t ret; 1180 1181 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, DAX_ZERO_PAGE); 1182 1183 ret = vmf_insert_mixed(vmf->vma, vaddr, pfn); 1184 trace_dax_load_hole(inode, vmf, ret); 1185 return ret; 1186 } 1187 1188 #ifdef CONFIG_FS_DAX_PMD 1189 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, 1190 const struct iomap_iter *iter, void **entry) 1191 { 1192 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1193 unsigned long pmd_addr = vmf->address & PMD_MASK; 1194 struct vm_area_struct *vma = vmf->vma; 1195 struct inode *inode = mapping->host; 1196 pgtable_t pgtable = NULL; 1197 struct page *zero_page; 1198 spinlock_t *ptl; 1199 pmd_t pmd_entry; 1200 pfn_t pfn; 1201 1202 zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); 1203 1204 if (unlikely(!zero_page)) 1205 goto fallback; 1206 1207 pfn = page_to_pfn_t(zero_page); 1208 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, 1209 DAX_PMD | DAX_ZERO_PAGE); 1210 1211 if (arch_needs_pgtable_deposit()) { 1212 pgtable = pte_alloc_one(vma->vm_mm); 1213 if (!pgtable) 1214 return VM_FAULT_OOM; 1215 } 1216 1217 ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); 1218 if (!pmd_none(*(vmf->pmd))) { 1219 spin_unlock(ptl); 1220 goto fallback; 1221 } 1222 1223 if (pgtable) { 1224 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); 1225 mm_inc_nr_ptes(vma->vm_mm); 1226 } 1227 pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); 1228 pmd_entry = pmd_mkhuge(pmd_entry); 1229 set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); 1230 spin_unlock(ptl); 1231 trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry); 1232 return VM_FAULT_NOPAGE; 1233 1234 fallback: 1235 if (pgtable) 1236 pte_free(vma->vm_mm, pgtable); 1237 trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry); 1238 return VM_FAULT_FALLBACK; 1239 } 1240 #else 1241 static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, 1242 const struct iomap_iter *iter, void **entry) 1243 { 1244 return VM_FAULT_FALLBACK; 1245 } 1246 #endif /* CONFIG_FS_DAX_PMD */ 1247 1248 static s64 dax_unshare_iter(struct iomap_iter *iter) 1249 { 1250 struct iomap *iomap = &iter->iomap; 1251 const struct iomap *srcmap = iomap_iter_srcmap(iter); 1252 loff_t pos = iter->pos; 1253 loff_t length = iomap_length(iter); 1254 int id = 0; 1255 s64 ret = 0; 1256 void *daddr = NULL, *saddr = NULL; 1257 1258 /* don't bother with blocks that are not shared to start with */ 1259 if (!(iomap->flags & IOMAP_F_SHARED)) 1260 return length; 1261 1262 id = dax_read_lock(); 1263 ret = dax_iomap_direct_access(iomap, pos, length, &daddr, NULL); 1264 if (ret < 0) 1265 goto out_unlock; 1266 1267 /* zero the distance if srcmap is HOLE or UNWRITTEN */ 1268 if (srcmap->flags & IOMAP_F_SHARED || srcmap->type == IOMAP_UNWRITTEN) { 1269 memset(daddr, 0, length); 1270 dax_flush(iomap->dax_dev, daddr, length); 1271 ret = length; 1272 goto out_unlock; 1273 } 1274 1275 ret = dax_iomap_direct_access(srcmap, pos, length, &saddr, NULL); 1276 if (ret < 0) 1277 goto out_unlock; 1278 1279 if (copy_mc_to_kernel(daddr, saddr, length) == 0) 1280 ret = length; 1281 else 1282 ret = -EIO; 1283 1284 out_unlock: 1285 dax_read_unlock(id); 1286 return ret; 1287 } 1288 1289 int dax_file_unshare(struct inode *inode, loff_t pos, loff_t len, 1290 const struct iomap_ops *ops) 1291 { 1292 struct iomap_iter iter = { 1293 .inode = inode, 1294 .pos = pos, 1295 .len = len, 1296 .flags = IOMAP_WRITE | IOMAP_UNSHARE | IOMAP_DAX, 1297 }; 1298 int ret; 1299 1300 while ((ret = iomap_iter(&iter, ops)) > 0) 1301 iter.processed = dax_unshare_iter(&iter); 1302 return ret; 1303 } 1304 EXPORT_SYMBOL_GPL(dax_file_unshare); 1305 1306 static int dax_memzero(struct iomap_iter *iter, loff_t pos, size_t size) 1307 { 1308 const struct iomap *iomap = &iter->iomap; 1309 const struct iomap *srcmap = iomap_iter_srcmap(iter); 1310 unsigned offset = offset_in_page(pos); 1311 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); 1312 void *kaddr; 1313 long ret; 1314 1315 ret = dax_direct_access(iomap->dax_dev, pgoff, 1, DAX_ACCESS, &kaddr, 1316 NULL); 1317 if (ret < 0) 1318 return ret; 1319 memset(kaddr + offset, 0, size); 1320 if (iomap->flags & IOMAP_F_SHARED) 1321 ret = dax_iomap_copy_around(pos, size, PAGE_SIZE, srcmap, 1322 kaddr); 1323 else 1324 dax_flush(iomap->dax_dev, kaddr + offset, size); 1325 return ret; 1326 } 1327 1328 static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero) 1329 { 1330 const struct iomap *iomap = &iter->iomap; 1331 const struct iomap *srcmap = iomap_iter_srcmap(iter); 1332 loff_t pos = iter->pos; 1333 u64 length = iomap_length(iter); 1334 s64 written = 0; 1335 1336 /* already zeroed? we're done. */ 1337 if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN) 1338 return length; 1339 1340 /* 1341 * invalidate the pages whose sharing state is to be changed 1342 * because of CoW. 1343 */ 1344 if (iomap->flags & IOMAP_F_SHARED) 1345 invalidate_inode_pages2_range(iter->inode->i_mapping, 1346 pos >> PAGE_SHIFT, 1347 (pos + length - 1) >> PAGE_SHIFT); 1348 1349 do { 1350 unsigned offset = offset_in_page(pos); 1351 unsigned size = min_t(u64, PAGE_SIZE - offset, length); 1352 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); 1353 long rc; 1354 int id; 1355 1356 id = dax_read_lock(); 1357 if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE) 1358 rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1); 1359 else 1360 rc = dax_memzero(iter, pos, size); 1361 dax_read_unlock(id); 1362 1363 if (rc < 0) 1364 return rc; 1365 pos += size; 1366 length -= size; 1367 written += size; 1368 } while (length > 0); 1369 1370 if (did_zero) 1371 *did_zero = true; 1372 return written; 1373 } 1374 1375 int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero, 1376 const struct iomap_ops *ops) 1377 { 1378 struct iomap_iter iter = { 1379 .inode = inode, 1380 .pos = pos, 1381 .len = len, 1382 .flags = IOMAP_DAX | IOMAP_ZERO, 1383 }; 1384 int ret; 1385 1386 while ((ret = iomap_iter(&iter, ops)) > 0) 1387 iter.processed = dax_zero_iter(&iter, did_zero); 1388 return ret; 1389 } 1390 EXPORT_SYMBOL_GPL(dax_zero_range); 1391 1392 int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero, 1393 const struct iomap_ops *ops) 1394 { 1395 unsigned int blocksize = i_blocksize(inode); 1396 unsigned int off = pos & (blocksize - 1); 1397 1398 /* Block boundary? Nothing to do */ 1399 if (!off) 1400 return 0; 1401 return dax_zero_range(inode, pos, blocksize - off, did_zero, ops); 1402 } 1403 EXPORT_SYMBOL_GPL(dax_truncate_page); 1404 1405 static loff_t dax_iomap_iter(const struct iomap_iter *iomi, 1406 struct iov_iter *iter) 1407 { 1408 const struct iomap *iomap = &iomi->iomap; 1409 const struct iomap *srcmap = iomap_iter_srcmap(iomi); 1410 loff_t length = iomap_length(iomi); 1411 loff_t pos = iomi->pos; 1412 struct dax_device *dax_dev = iomap->dax_dev; 1413 loff_t end = pos + length, done = 0; 1414 bool write = iov_iter_rw(iter) == WRITE; 1415 bool cow = write && iomap->flags & IOMAP_F_SHARED; 1416 ssize_t ret = 0; 1417 size_t xfer; 1418 int id; 1419 1420 if (!write) { 1421 end = min(end, i_size_read(iomi->inode)); 1422 if (pos >= end) 1423 return 0; 1424 1425 if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) 1426 return iov_iter_zero(min(length, end - pos), iter); 1427 } 1428 1429 /* 1430 * In DAX mode, enforce either pure overwrites of written extents, or 1431 * writes to unwritten extents as part of a copy-on-write operation. 1432 */ 1433 if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED && 1434 !(iomap->flags & IOMAP_F_SHARED))) 1435 return -EIO; 1436 1437 /* 1438 * Write can allocate block for an area which has a hole page mapped 1439 * into page tables. We have to tear down these mappings so that data 1440 * written by write(2) is visible in mmap. 1441 */ 1442 if (iomap->flags & IOMAP_F_NEW || cow) { 1443 invalidate_inode_pages2_range(iomi->inode->i_mapping, 1444 pos >> PAGE_SHIFT, 1445 (end - 1) >> PAGE_SHIFT); 1446 } 1447 1448 id = dax_read_lock(); 1449 while (pos < end) { 1450 unsigned offset = pos & (PAGE_SIZE - 1); 1451 const size_t size = ALIGN(length + offset, PAGE_SIZE); 1452 pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); 1453 ssize_t map_len; 1454 bool recovery = false; 1455 void *kaddr; 1456 1457 if (fatal_signal_pending(current)) { 1458 ret = -EINTR; 1459 break; 1460 } 1461 1462 map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), 1463 DAX_ACCESS, &kaddr, NULL); 1464 if (map_len == -EIO && iov_iter_rw(iter) == WRITE) { 1465 map_len = dax_direct_access(dax_dev, pgoff, 1466 PHYS_PFN(size), DAX_RECOVERY_WRITE, 1467 &kaddr, NULL); 1468 if (map_len > 0) 1469 recovery = true; 1470 } 1471 if (map_len < 0) { 1472 ret = map_len; 1473 break; 1474 } 1475 1476 if (cow) { 1477 ret = dax_iomap_copy_around(pos, length, PAGE_SIZE, 1478 srcmap, kaddr); 1479 if (ret) 1480 break; 1481 } 1482 1483 map_len = PFN_PHYS(map_len); 1484 kaddr += offset; 1485 map_len -= offset; 1486 if (map_len > end - pos) 1487 map_len = end - pos; 1488 1489 if (recovery) 1490 xfer = dax_recovery_write(dax_dev, pgoff, kaddr, 1491 map_len, iter); 1492 else if (write) 1493 xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr, 1494 map_len, iter); 1495 else 1496 xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr, 1497 map_len, iter); 1498 1499 pos += xfer; 1500 length -= xfer; 1501 done += xfer; 1502 1503 if (xfer == 0) 1504 ret = -EFAULT; 1505 if (xfer < map_len) 1506 break; 1507 } 1508 dax_read_unlock(id); 1509 1510 return done ? done : ret; 1511 } 1512 1513 /** 1514 * dax_iomap_rw - Perform I/O to a DAX file 1515 * @iocb: The control block for this I/O 1516 * @iter: The addresses to do I/O from or to 1517 * @ops: iomap ops passed from the file system 1518 * 1519 * This function performs read and write operations to directly mapped 1520 * persistent memory. The callers needs to take care of read/write exclusion 1521 * and evicting any page cache pages in the region under I/O. 1522 */ 1523 ssize_t 1524 dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, 1525 const struct iomap_ops *ops) 1526 { 1527 struct iomap_iter iomi = { 1528 .inode = iocb->ki_filp->f_mapping->host, 1529 .pos = iocb->ki_pos, 1530 .len = iov_iter_count(iter), 1531 .flags = IOMAP_DAX, 1532 }; 1533 loff_t done = 0; 1534 int ret; 1535 1536 if (!iomi.len) 1537 return 0; 1538 1539 if (iov_iter_rw(iter) == WRITE) { 1540 lockdep_assert_held_write(&iomi.inode->i_rwsem); 1541 iomi.flags |= IOMAP_WRITE; 1542 } else { 1543 lockdep_assert_held(&iomi.inode->i_rwsem); 1544 } 1545 1546 if (iocb->ki_flags & IOCB_NOWAIT) 1547 iomi.flags |= IOMAP_NOWAIT; 1548 1549 while ((ret = iomap_iter(&iomi, ops)) > 0) 1550 iomi.processed = dax_iomap_iter(&iomi, iter); 1551 1552 done = iomi.pos - iocb->ki_pos; 1553 iocb->ki_pos = iomi.pos; 1554 return done ? done : ret; 1555 } 1556 EXPORT_SYMBOL_GPL(dax_iomap_rw); 1557 1558 static vm_fault_t dax_fault_return(int error) 1559 { 1560 if (error == 0) 1561 return VM_FAULT_NOPAGE; 1562 return vmf_error(error); 1563 } 1564 1565 /* 1566 * When handling a synchronous page fault and the inode need a fsync, we can 1567 * insert the PTE/PMD into page tables only after that fsync happened. Skip 1568 * insertion for now and return the pfn so that caller can insert it after the 1569 * fsync is done. 1570 */ 1571 static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn) 1572 { 1573 if (WARN_ON_ONCE(!pfnp)) 1574 return VM_FAULT_SIGBUS; 1575 *pfnp = pfn; 1576 return VM_FAULT_NEEDDSYNC; 1577 } 1578 1579 static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf, 1580 const struct iomap_iter *iter) 1581 { 1582 vm_fault_t ret; 1583 int error = 0; 1584 1585 switch (iter->iomap.type) { 1586 case IOMAP_HOLE: 1587 case IOMAP_UNWRITTEN: 1588 clear_user_highpage(vmf->cow_page, vmf->address); 1589 break; 1590 case IOMAP_MAPPED: 1591 error = copy_cow_page_dax(vmf, iter); 1592 break; 1593 default: 1594 WARN_ON_ONCE(1); 1595 error = -EIO; 1596 break; 1597 } 1598 1599 if (error) 1600 return dax_fault_return(error); 1601 1602 __SetPageUptodate(vmf->cow_page); 1603 ret = finish_fault(vmf); 1604 if (!ret) 1605 return VM_FAULT_DONE_COW; 1606 return ret; 1607 } 1608 1609 /** 1610 * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault. 1611 * @vmf: vm fault instance 1612 * @iter: iomap iter 1613 * @pfnp: pfn to be returned 1614 * @xas: the dax mapping tree of a file 1615 * @entry: an unlocked dax entry to be inserted 1616 * @pmd: distinguish whether it is a pmd fault 1617 */ 1618 static vm_fault_t dax_fault_iter(struct vm_fault *vmf, 1619 const struct iomap_iter *iter, pfn_t *pfnp, 1620 struct xa_state *xas, void **entry, bool pmd) 1621 { 1622 const struct iomap *iomap = &iter->iomap; 1623 const struct iomap *srcmap = iomap_iter_srcmap(iter); 1624 size_t size = pmd ? PMD_SIZE : PAGE_SIZE; 1625 loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT; 1626 bool write = iter->flags & IOMAP_WRITE; 1627 unsigned long entry_flags = pmd ? DAX_PMD : 0; 1628 int err = 0; 1629 pfn_t pfn; 1630 void *kaddr; 1631 1632 if (!pmd && vmf->cow_page) 1633 return dax_fault_cow_page(vmf, iter); 1634 1635 /* if we are reading UNWRITTEN and HOLE, return a hole. */ 1636 if (!write && 1637 (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) { 1638 if (!pmd) 1639 return dax_load_hole(xas, vmf, iter, entry); 1640 return dax_pmd_load_hole(xas, vmf, iter, entry); 1641 } 1642 1643 if (iomap->type != IOMAP_MAPPED && !(iomap->flags & IOMAP_F_SHARED)) { 1644 WARN_ON_ONCE(1); 1645 return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS; 1646 } 1647 1648 err = dax_iomap_direct_access(iomap, pos, size, &kaddr, &pfn); 1649 if (err) 1650 return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err); 1651 1652 *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, entry_flags); 1653 1654 if (write && iomap->flags & IOMAP_F_SHARED) { 1655 err = dax_iomap_copy_around(pos, size, size, srcmap, kaddr); 1656 if (err) 1657 return dax_fault_return(err); 1658 } 1659 1660 if (dax_fault_is_synchronous(iter, vmf->vma)) 1661 return dax_fault_synchronous_pfnp(pfnp, pfn); 1662 1663 /* insert PMD pfn */ 1664 if (pmd) 1665 return vmf_insert_pfn_pmd(vmf, pfn, write); 1666 1667 /* insert PTE pfn */ 1668 if (write) 1669 return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); 1670 return vmf_insert_mixed(vmf->vma, vmf->address, pfn); 1671 } 1672 1673 static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp, 1674 int *iomap_errp, const struct iomap_ops *ops) 1675 { 1676 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1677 XA_STATE(xas, &mapping->i_pages, vmf->pgoff); 1678 struct iomap_iter iter = { 1679 .inode = mapping->host, 1680 .pos = (loff_t)vmf->pgoff << PAGE_SHIFT, 1681 .len = PAGE_SIZE, 1682 .flags = IOMAP_DAX | IOMAP_FAULT, 1683 }; 1684 vm_fault_t ret = 0; 1685 void *entry; 1686 int error; 1687 1688 trace_dax_pte_fault(iter.inode, vmf, ret); 1689 /* 1690 * Check whether offset isn't beyond end of file now. Caller is supposed 1691 * to hold locks serializing us with truncate / punch hole so this is 1692 * a reliable test. 1693 */ 1694 if (iter.pos >= i_size_read(iter.inode)) { 1695 ret = VM_FAULT_SIGBUS; 1696 goto out; 1697 } 1698 1699 if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) 1700 iter.flags |= IOMAP_WRITE; 1701 1702 entry = grab_mapping_entry(&xas, mapping, 0); 1703 if (xa_is_internal(entry)) { 1704 ret = xa_to_internal(entry); 1705 goto out; 1706 } 1707 1708 /* 1709 * It is possible, particularly with mixed reads & writes to private 1710 * mappings, that we have raced with a PMD fault that overlaps with 1711 * the PTE we need to set up. If so just return and the fault will be 1712 * retried. 1713 */ 1714 if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) { 1715 ret = VM_FAULT_NOPAGE; 1716 goto unlock_entry; 1717 } 1718 1719 while ((error = iomap_iter(&iter, ops)) > 0) { 1720 if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) { 1721 iter.processed = -EIO; /* fs corruption? */ 1722 continue; 1723 } 1724 1725 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false); 1726 if (ret != VM_FAULT_SIGBUS && 1727 (iter.iomap.flags & IOMAP_F_NEW)) { 1728 count_vm_event(PGMAJFAULT); 1729 count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT); 1730 ret |= VM_FAULT_MAJOR; 1731 } 1732 1733 if (!(ret & VM_FAULT_ERROR)) 1734 iter.processed = PAGE_SIZE; 1735 } 1736 1737 if (iomap_errp) 1738 *iomap_errp = error; 1739 if (!ret && error) 1740 ret = dax_fault_return(error); 1741 1742 unlock_entry: 1743 dax_unlock_entry(&xas, entry); 1744 out: 1745 trace_dax_pte_fault_done(iter.inode, vmf, ret); 1746 return ret; 1747 } 1748 1749 #ifdef CONFIG_FS_DAX_PMD 1750 static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas, 1751 pgoff_t max_pgoff) 1752 { 1753 unsigned long pmd_addr = vmf->address & PMD_MASK; 1754 bool write = vmf->flags & FAULT_FLAG_WRITE; 1755 1756 /* 1757 * Make sure that the faulting address's PMD offset (color) matches 1758 * the PMD offset from the start of the file. This is necessary so 1759 * that a PMD range in the page table overlaps exactly with a PMD 1760 * range in the page cache. 1761 */ 1762 if ((vmf->pgoff & PG_PMD_COLOUR) != 1763 ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR)) 1764 return true; 1765 1766 /* Fall back to PTEs if we're going to COW */ 1767 if (write && !(vmf->vma->vm_flags & VM_SHARED)) 1768 return true; 1769 1770 /* If the PMD would extend outside the VMA */ 1771 if (pmd_addr < vmf->vma->vm_start) 1772 return true; 1773 if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end) 1774 return true; 1775 1776 /* If the PMD would extend beyond the file size */ 1777 if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff) 1778 return true; 1779 1780 return false; 1781 } 1782 1783 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, 1784 const struct iomap_ops *ops) 1785 { 1786 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1787 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER); 1788 struct iomap_iter iter = { 1789 .inode = mapping->host, 1790 .len = PMD_SIZE, 1791 .flags = IOMAP_DAX | IOMAP_FAULT, 1792 }; 1793 vm_fault_t ret = VM_FAULT_FALLBACK; 1794 pgoff_t max_pgoff; 1795 void *entry; 1796 int error; 1797 1798 if (vmf->flags & FAULT_FLAG_WRITE) 1799 iter.flags |= IOMAP_WRITE; 1800 1801 /* 1802 * Check whether offset isn't beyond end of file now. Caller is 1803 * supposed to hold locks serializing us with truncate / punch hole so 1804 * this is a reliable test. 1805 */ 1806 max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE); 1807 1808 trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0); 1809 1810 if (xas.xa_index >= max_pgoff) { 1811 ret = VM_FAULT_SIGBUS; 1812 goto out; 1813 } 1814 1815 if (dax_fault_check_fallback(vmf, &xas, max_pgoff)) 1816 goto fallback; 1817 1818 /* 1819 * grab_mapping_entry() will make sure we get an empty PMD entry, 1820 * a zero PMD entry or a DAX PMD. If it can't (because a PTE 1821 * entry is already in the array, for instance), it will return 1822 * VM_FAULT_FALLBACK. 1823 */ 1824 entry = grab_mapping_entry(&xas, mapping, PMD_ORDER); 1825 if (xa_is_internal(entry)) { 1826 ret = xa_to_internal(entry); 1827 goto fallback; 1828 } 1829 1830 /* 1831 * It is possible, particularly with mixed reads & writes to private 1832 * mappings, that we have raced with a PTE fault that overlaps with 1833 * the PMD we need to set up. If so just return and the fault will be 1834 * retried. 1835 */ 1836 if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) && 1837 !pmd_devmap(*vmf->pmd)) { 1838 ret = 0; 1839 goto unlock_entry; 1840 } 1841 1842 iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT; 1843 while ((error = iomap_iter(&iter, ops)) > 0) { 1844 if (iomap_length(&iter) < PMD_SIZE) 1845 continue; /* actually breaks out of the loop */ 1846 1847 ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true); 1848 if (ret != VM_FAULT_FALLBACK) 1849 iter.processed = PMD_SIZE; 1850 } 1851 1852 unlock_entry: 1853 dax_unlock_entry(&xas, entry); 1854 fallback: 1855 if (ret == VM_FAULT_FALLBACK) { 1856 split_huge_pmd(vmf->vma, vmf->pmd, vmf->address); 1857 count_vm_event(THP_FAULT_FALLBACK); 1858 } 1859 out: 1860 trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret); 1861 return ret; 1862 } 1863 #else 1864 static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, 1865 const struct iomap_ops *ops) 1866 { 1867 return VM_FAULT_FALLBACK; 1868 } 1869 #endif /* CONFIG_FS_DAX_PMD */ 1870 1871 /** 1872 * dax_iomap_fault - handle a page fault on a DAX file 1873 * @vmf: The description of the fault 1874 * @pe_size: Size of the page to fault in 1875 * @pfnp: PFN to insert for synchronous faults if fsync is required 1876 * @iomap_errp: Storage for detailed error code in case of error 1877 * @ops: Iomap ops passed from the file system 1878 * 1879 * When a page fault occurs, filesystems may call this helper in 1880 * their fault handler for DAX files. dax_iomap_fault() assumes the caller 1881 * has done all the necessary locking for page fault to proceed 1882 * successfully. 1883 */ 1884 vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, 1885 pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops) 1886 { 1887 switch (pe_size) { 1888 case PE_SIZE_PTE: 1889 return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops); 1890 case PE_SIZE_PMD: 1891 return dax_iomap_pmd_fault(vmf, pfnp, ops); 1892 default: 1893 return VM_FAULT_FALLBACK; 1894 } 1895 } 1896 EXPORT_SYMBOL_GPL(dax_iomap_fault); 1897 1898 /* 1899 * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables 1900 * @vmf: The description of the fault 1901 * @pfn: PFN to insert 1902 * @order: Order of entry to insert. 1903 * 1904 * This function inserts a writeable PTE or PMD entry into the page tables 1905 * for an mmaped DAX file. It also marks the page cache entry as dirty. 1906 */ 1907 static vm_fault_t 1908 dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order) 1909 { 1910 struct address_space *mapping = vmf->vma->vm_file->f_mapping; 1911 XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order); 1912 void *entry; 1913 vm_fault_t ret; 1914 1915 xas_lock_irq(&xas); 1916 entry = get_unlocked_entry(&xas, order); 1917 /* Did we race with someone splitting entry or so? */ 1918 if (!entry || dax_is_conflict(entry) || 1919 (order == 0 && !dax_is_pte_entry(entry))) { 1920 put_unlocked_entry(&xas, entry, WAKE_NEXT); 1921 xas_unlock_irq(&xas); 1922 trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf, 1923 VM_FAULT_NOPAGE); 1924 return VM_FAULT_NOPAGE; 1925 } 1926 xas_set_mark(&xas, PAGECACHE_TAG_DIRTY); 1927 dax_lock_entry(&xas, entry); 1928 xas_unlock_irq(&xas); 1929 if (order == 0) 1930 ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); 1931 #ifdef CONFIG_FS_DAX_PMD 1932 else if (order == PMD_ORDER) 1933 ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE); 1934 #endif 1935 else 1936 ret = VM_FAULT_FALLBACK; 1937 dax_unlock_entry(&xas, entry); 1938 trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret); 1939 return ret; 1940 } 1941 1942 /** 1943 * dax_finish_sync_fault - finish synchronous page fault 1944 * @vmf: The description of the fault 1945 * @pe_size: Size of entry to be inserted 1946 * @pfn: PFN to insert 1947 * 1948 * This function ensures that the file range touched by the page fault is 1949 * stored persistently on the media and handles inserting of appropriate page 1950 * table entry. 1951 */ 1952 vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, 1953 enum page_entry_size pe_size, pfn_t pfn) 1954 { 1955 int err; 1956 loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT; 1957 unsigned int order = pe_order(pe_size); 1958 size_t len = PAGE_SIZE << order; 1959 1960 err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1); 1961 if (err) 1962 return VM_FAULT_SIGBUS; 1963 return dax_insert_pfn_mkwrite(vmf, pfn, order); 1964 } 1965 EXPORT_SYMBOL_GPL(dax_finish_sync_fault); 1966 1967 static loff_t dax_range_compare_iter(struct iomap_iter *it_src, 1968 struct iomap_iter *it_dest, u64 len, bool *same) 1969 { 1970 const struct iomap *smap = &it_src->iomap; 1971 const struct iomap *dmap = &it_dest->iomap; 1972 loff_t pos1 = it_src->pos, pos2 = it_dest->pos; 1973 void *saddr, *daddr; 1974 int id, ret; 1975 1976 len = min(len, min(smap->length, dmap->length)); 1977 1978 if (smap->type == IOMAP_HOLE && dmap->type == IOMAP_HOLE) { 1979 *same = true; 1980 return len; 1981 } 1982 1983 if (smap->type == IOMAP_HOLE || dmap->type == IOMAP_HOLE) { 1984 *same = false; 1985 return 0; 1986 } 1987 1988 id = dax_read_lock(); 1989 ret = dax_iomap_direct_access(smap, pos1, ALIGN(pos1 + len, PAGE_SIZE), 1990 &saddr, NULL); 1991 if (ret < 0) 1992 goto out_unlock; 1993 1994 ret = dax_iomap_direct_access(dmap, pos2, ALIGN(pos2 + len, PAGE_SIZE), 1995 &daddr, NULL); 1996 if (ret < 0) 1997 goto out_unlock; 1998 1999 *same = !memcmp(saddr, daddr, len); 2000 if (!*same) 2001 len = 0; 2002 dax_read_unlock(id); 2003 return len; 2004 2005 out_unlock: 2006 dax_read_unlock(id); 2007 return -EIO; 2008 } 2009 2010 int dax_dedupe_file_range_compare(struct inode *src, loff_t srcoff, 2011 struct inode *dst, loff_t dstoff, loff_t len, bool *same, 2012 const struct iomap_ops *ops) 2013 { 2014 struct iomap_iter src_iter = { 2015 .inode = src, 2016 .pos = srcoff, 2017 .len = len, 2018 .flags = IOMAP_DAX, 2019 }; 2020 struct iomap_iter dst_iter = { 2021 .inode = dst, 2022 .pos = dstoff, 2023 .len = len, 2024 .flags = IOMAP_DAX, 2025 }; 2026 int ret, compared = 0; 2027 2028 while ((ret = iomap_iter(&src_iter, ops)) > 0 && 2029 (ret = iomap_iter(&dst_iter, ops)) > 0) { 2030 compared = dax_range_compare_iter(&src_iter, &dst_iter, 2031 min(src_iter.len, dst_iter.len), same); 2032 if (compared < 0) 2033 return ret; 2034 src_iter.processed = dst_iter.processed = compared; 2035 } 2036 return ret; 2037 } 2038 2039 int dax_remap_file_range_prep(struct file *file_in, loff_t pos_in, 2040 struct file *file_out, loff_t pos_out, 2041 loff_t *len, unsigned int remap_flags, 2042 const struct iomap_ops *ops) 2043 { 2044 return __generic_remap_file_range_prep(file_in, pos_in, file_out, 2045 pos_out, len, remap_flags, ops); 2046 } 2047 EXPORT_SYMBOL_GPL(dax_remap_file_range_prep); 2048