| 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 | static int __dax_clear_dirty_range(struct address_space *mapping, |
| 785 | pgoff_t start, pgoff_t end) |
| 786 | { |
| 787 | XA_STATE(xas, &mapping->i_pages, start); |
| 788 | unsigned int scanned = 0; |
| 789 | void *entry; |
| 790 | |
| 791 | xas_lock_irq(&xas); |
| 792 | xas_for_each(&xas, entry, end) { |
| 793 | entry = get_unlocked_entry(&xas, 0); |
| 794 | xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY); |
| 795 | xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE); |
| 796 | put_unlocked_entry(&xas, entry, WAKE_NEXT); |
| 797 | |
| 798 | if (++scanned % XA_CHECK_SCHED) |
| 799 | continue; |
| 800 | |
| 801 | xas_pause(&xas); |
| 802 | xas_unlock_irq(&xas); |
| 803 | cond_resched(); |
| 804 | xas_lock_irq(&xas); |
| 805 | } |
| 806 | xas_unlock_irq(&xas); |
| 807 | |
| 808 | return 0; |
| 809 | } |
| 810 | |
| 811 | /* |
| 812 | * Delete DAX entry at @index from @mapping. Wait for it |
| 813 | * to be unlocked before deleting it. |
| 814 | */ |
| 815 | int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index) |
| 816 | { |
| 817 | int ret = __dax_invalidate_entry(mapping, index, true); |
| 818 | |
| 819 | /* |
| 820 | * This gets called from truncate / punch_hole path. As such, the caller |
| 821 | * must hold locks protecting against concurrent modifications of the |
| 822 | * page cache (usually fs-private i_mmap_sem for writing). Since the |
| 823 | * caller has seen a DAX entry for this index, we better find it |
| 824 | * at that index as well... |
| 825 | */ |
| 826 | WARN_ON_ONCE(!ret); |
| 827 | return ret; |
| 828 | } |
| 829 | |
| 830 | /* |
| 831 | * Invalidate DAX entry if it is clean. |
| 832 | */ |
| 833 | int dax_invalidate_mapping_entry_sync(struct address_space *mapping, |
| 834 | pgoff_t index) |
| 835 | { |
| 836 | return __dax_invalidate_entry(mapping, index, false); |
| 837 | } |
| 838 | |
| 839 | static pgoff_t dax_iomap_pgoff(const struct iomap *iomap, loff_t pos) |
| 840 | { |
| 841 | return PHYS_PFN(iomap->addr + (pos & PAGE_MASK) - iomap->offset); |
| 842 | } |
| 843 | |
| 844 | static int copy_cow_page_dax(struct vm_fault *vmf, const struct iomap_iter *iter) |
| 845 | { |
| 846 | pgoff_t pgoff = dax_iomap_pgoff(&iter->iomap, iter->pos); |
| 847 | void *vto, *kaddr; |
| 848 | long rc; |
| 849 | int id; |
| 850 | |
| 851 | id = dax_read_lock(); |
| 852 | rc = dax_direct_access(iter->iomap.dax_dev, pgoff, 1, DAX_ACCESS, |
| 853 | &kaddr, NULL); |
| 854 | if (rc < 0) { |
| 855 | dax_read_unlock(id); |
| 856 | return rc; |
| 857 | } |
| 858 | vto = kmap_atomic(vmf->cow_page); |
| 859 | copy_user_page(vto, kaddr, vmf->address, vmf->cow_page); |
| 860 | kunmap_atomic(vto); |
| 861 | dax_read_unlock(id); |
| 862 | return 0; |
| 863 | } |
| 864 | |
| 865 | /* |
| 866 | * MAP_SYNC on a dax mapping guarantees dirty metadata is |
| 867 | * flushed on write-faults (non-cow), but not read-faults. |
| 868 | */ |
| 869 | static bool dax_fault_is_synchronous(const struct iomap_iter *iter, |
| 870 | struct vm_area_struct *vma) |
| 871 | { |
| 872 | return (iter->flags & IOMAP_WRITE) && (vma->vm_flags & VM_SYNC) && |
| 873 | (iter->iomap.flags & IOMAP_F_DIRTY); |
| 874 | } |
| 875 | |
| 876 | /* |
| 877 | * By this point grab_mapping_entry() has ensured that we have a locked entry |
| 878 | * of the appropriate size so we don't have to worry about downgrading PMDs to |
| 879 | * PTEs. If we happen to be trying to insert a PTE and there is a PMD |
| 880 | * already in the tree, we will skip the insertion and just dirty the PMD as |
| 881 | * appropriate. |
| 882 | */ |
| 883 | static void *dax_insert_entry(struct xa_state *xas, struct vm_fault *vmf, |
| 884 | const struct iomap_iter *iter, void *entry, pfn_t pfn, |
| 885 | unsigned long flags) |
| 886 | { |
| 887 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
| 888 | void *new_entry = dax_make_entry(pfn, flags); |
| 889 | bool write = iter->flags & IOMAP_WRITE; |
| 890 | bool dirty = write && !dax_fault_is_synchronous(iter, vmf->vma); |
| 891 | bool shared = iter->iomap.flags & IOMAP_F_SHARED; |
| 892 | |
| 893 | if (dirty) |
| 894 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); |
| 895 | |
| 896 | if (shared || (dax_is_zero_entry(entry) && !(flags & DAX_ZERO_PAGE))) { |
| 897 | unsigned long index = xas->xa_index; |
| 898 | /* we are replacing a zero page with block mapping */ |
| 899 | if (dax_is_pmd_entry(entry)) |
| 900 | unmap_mapping_pages(mapping, index & ~PG_PMD_COLOUR, |
| 901 | PG_PMD_NR, false); |
| 902 | else /* pte entry */ |
| 903 | unmap_mapping_pages(mapping, index, 1, false); |
| 904 | } |
| 905 | |
| 906 | xas_reset(xas); |
| 907 | xas_lock_irq(xas); |
| 908 | if (shared || dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) { |
| 909 | void *old; |
| 910 | |
| 911 | dax_disassociate_entry(entry, mapping, false); |
| 912 | dax_associate_entry(new_entry, mapping, vmf->vma, vmf->address, |
| 913 | shared); |
| 914 | /* |
| 915 | * Only swap our new entry into the page cache if the current |
| 916 | * entry is a zero page or an empty entry. If a normal PTE or |
| 917 | * PMD entry is already in the cache, we leave it alone. This |
| 918 | * means that if we are trying to insert a PTE and the |
| 919 | * existing entry is a PMD, we will just leave the PMD in the |
| 920 | * tree and dirty it if necessary. |
| 921 | */ |
| 922 | old = dax_lock_entry(xas, new_entry); |
| 923 | WARN_ON_ONCE(old != xa_mk_value(xa_to_value(entry) | |
| 924 | DAX_LOCKED)); |
| 925 | entry = new_entry; |
| 926 | } else { |
| 927 | xas_load(xas); /* Walk the xa_state */ |
| 928 | } |
| 929 | |
| 930 | if (dirty) |
| 931 | xas_set_mark(xas, PAGECACHE_TAG_DIRTY); |
| 932 | |
| 933 | if (write && shared) |
| 934 | xas_set_mark(xas, PAGECACHE_TAG_TOWRITE); |
| 935 | |
| 936 | xas_unlock_irq(xas); |
| 937 | return entry; |
| 938 | } |
| 939 | |
| 940 | static int dax_writeback_one(struct xa_state *xas, struct dax_device *dax_dev, |
| 941 | struct address_space *mapping, void *entry) |
| 942 | { |
| 943 | unsigned long pfn, index, count, end; |
| 944 | long ret = 0; |
| 945 | struct vm_area_struct *vma; |
| 946 | |
| 947 | /* |
| 948 | * A page got tagged dirty in DAX mapping? Something is seriously |
| 949 | * wrong. |
| 950 | */ |
| 951 | if (WARN_ON(!xa_is_value(entry))) |
| 952 | return -EIO; |
| 953 | |
| 954 | if (unlikely(dax_is_locked(entry))) { |
| 955 | void *old_entry = entry; |
| 956 | |
| 957 | entry = get_unlocked_entry(xas, 0); |
| 958 | |
| 959 | /* Entry got punched out / reallocated? */ |
| 960 | if (!entry || WARN_ON_ONCE(!xa_is_value(entry))) |
| 961 | goto put_unlocked; |
| 962 | /* |
| 963 | * Entry got reallocated elsewhere? No need to writeback. |
| 964 | * We have to compare pfns as we must not bail out due to |
| 965 | * difference in lockbit or entry type. |
| 966 | */ |
| 967 | if (dax_to_pfn(old_entry) != dax_to_pfn(entry)) |
| 968 | goto put_unlocked; |
| 969 | if (WARN_ON_ONCE(dax_is_empty_entry(entry) || |
| 970 | dax_is_zero_entry(entry))) { |
| 971 | ret = -EIO; |
| 972 | goto put_unlocked; |
| 973 | } |
| 974 | |
| 975 | /* Another fsync thread may have already done this entry */ |
| 976 | if (!xas_get_mark(xas, PAGECACHE_TAG_TOWRITE)) |
| 977 | goto put_unlocked; |
| 978 | } |
| 979 | |
| 980 | /* Lock the entry to serialize with page faults */ |
| 981 | dax_lock_entry(xas, entry); |
| 982 | |
| 983 | /* |
| 984 | * We can clear the tag now but we have to be careful so that concurrent |
| 985 | * dax_writeback_one() calls for the same index cannot finish before we |
| 986 | * actually flush the caches. This is achieved as the calls will look |
| 987 | * at the entry only under the i_pages lock and once they do that |
| 988 | * they will see the entry locked and wait for it to unlock. |
| 989 | */ |
| 990 | xas_clear_mark(xas, PAGECACHE_TAG_TOWRITE); |
| 991 | xas_unlock_irq(xas); |
| 992 | |
| 993 | /* |
| 994 | * If dax_writeback_mapping_range() was given a wbc->range_start |
| 995 | * in the middle of a PMD, the 'index' we use needs to be |
| 996 | * aligned to the start of the PMD. |
| 997 | * This allows us to flush for PMD_SIZE and not have to worry about |
| 998 | * partial PMD writebacks. |
| 999 | */ |
| 1000 | pfn = dax_to_pfn(entry); |
| 1001 | count = 1UL << dax_entry_order(entry); |
| 1002 | index = xas->xa_index & ~(count - 1); |
| 1003 | end = index + count - 1; |
| 1004 | |
| 1005 | /* Walk all mappings of a given index of a file and writeprotect them */ |
| 1006 | i_mmap_lock_read(mapping); |
| 1007 | vma_interval_tree_foreach(vma, &mapping->i_mmap, index, end) { |
| 1008 | pfn_mkclean_range(pfn, count, index, vma); |
| 1009 | cond_resched(); |
| 1010 | } |
| 1011 | i_mmap_unlock_read(mapping); |
| 1012 | |
| 1013 | dax_flush(dax_dev, page_address(pfn_to_page(pfn)), count * PAGE_SIZE); |
| 1014 | /* |
| 1015 | * After we have flushed the cache, we can clear the dirty tag. There |
| 1016 | * cannot be new dirty data in the pfn after the flush has completed as |
| 1017 | * the pfn mappings are writeprotected and fault waits for mapping |
| 1018 | * entry lock. |
| 1019 | */ |
| 1020 | xas_reset(xas); |
| 1021 | xas_lock_irq(xas); |
| 1022 | xas_store(xas, entry); |
| 1023 | xas_clear_mark(xas, PAGECACHE_TAG_DIRTY); |
| 1024 | dax_wake_entry(xas, entry, WAKE_NEXT); |
| 1025 | |
| 1026 | trace_dax_writeback_one(mapping->host, index, count); |
| 1027 | return ret; |
| 1028 | |
| 1029 | put_unlocked: |
| 1030 | put_unlocked_entry(xas, entry, WAKE_NEXT); |
| 1031 | return ret; |
| 1032 | } |
| 1033 | |
| 1034 | /* |
| 1035 | * Flush the mapping to the persistent domain within the byte range of [start, |
| 1036 | * end]. This is required by data integrity operations to ensure file data is |
| 1037 | * on persistent storage prior to completion of the operation. |
| 1038 | */ |
| 1039 | int dax_writeback_mapping_range(struct address_space *mapping, |
| 1040 | struct dax_device *dax_dev, struct writeback_control *wbc) |
| 1041 | { |
| 1042 | XA_STATE(xas, &mapping->i_pages, wbc->range_start >> PAGE_SHIFT); |
| 1043 | struct inode *inode = mapping->host; |
| 1044 | pgoff_t end_index = wbc->range_end >> PAGE_SHIFT; |
| 1045 | void *entry; |
| 1046 | int ret = 0; |
| 1047 | unsigned int scanned = 0; |
| 1048 | |
| 1049 | if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT)) |
| 1050 | return -EIO; |
| 1051 | |
| 1052 | if (mapping_empty(mapping) || wbc->sync_mode != WB_SYNC_ALL) |
| 1053 | return 0; |
| 1054 | |
| 1055 | trace_dax_writeback_range(inode, xas.xa_index, end_index); |
| 1056 | |
| 1057 | tag_pages_for_writeback(mapping, xas.xa_index, end_index); |
| 1058 | |
| 1059 | xas_lock_irq(&xas); |
| 1060 | xas_for_each_marked(&xas, entry, end_index, PAGECACHE_TAG_TOWRITE) { |
| 1061 | ret = dax_writeback_one(&xas, dax_dev, mapping, entry); |
| 1062 | if (ret < 0) { |
| 1063 | mapping_set_error(mapping, ret); |
| 1064 | break; |
| 1065 | } |
| 1066 | if (++scanned % XA_CHECK_SCHED) |
| 1067 | continue; |
| 1068 | |
| 1069 | xas_pause(&xas); |
| 1070 | xas_unlock_irq(&xas); |
| 1071 | cond_resched(); |
| 1072 | xas_lock_irq(&xas); |
| 1073 | } |
| 1074 | xas_unlock_irq(&xas); |
| 1075 | trace_dax_writeback_range_done(inode, xas.xa_index, end_index); |
| 1076 | return ret; |
| 1077 | } |
| 1078 | EXPORT_SYMBOL_GPL(dax_writeback_mapping_range); |
| 1079 | |
| 1080 | static int dax_iomap_direct_access(const struct iomap *iomap, loff_t pos, |
| 1081 | size_t size, void **kaddr, pfn_t *pfnp) |
| 1082 | { |
| 1083 | pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); |
| 1084 | int id, rc = 0; |
| 1085 | long length; |
| 1086 | |
| 1087 | id = dax_read_lock(); |
| 1088 | length = dax_direct_access(iomap->dax_dev, pgoff, PHYS_PFN(size), |
| 1089 | DAX_ACCESS, kaddr, pfnp); |
| 1090 | if (length < 0) { |
| 1091 | rc = length; |
| 1092 | goto out; |
| 1093 | } |
| 1094 | if (!pfnp) |
| 1095 | goto out_check_addr; |
| 1096 | rc = -EINVAL; |
| 1097 | if (PFN_PHYS(length) < size) |
| 1098 | goto out; |
| 1099 | if (pfn_t_to_pfn(*pfnp) & (PHYS_PFN(size)-1)) |
| 1100 | goto out; |
| 1101 | /* For larger pages we need devmap */ |
| 1102 | if (length > 1 && !pfn_t_devmap(*pfnp)) |
| 1103 | goto out; |
| 1104 | rc = 0; |
| 1105 | |
| 1106 | out_check_addr: |
| 1107 | if (!kaddr) |
| 1108 | goto out; |
| 1109 | if (!*kaddr) |
| 1110 | rc = -EFAULT; |
| 1111 | out: |
| 1112 | dax_read_unlock(id); |
| 1113 | return rc; |
| 1114 | } |
| 1115 | |
| 1116 | /** |
| 1117 | * dax_iomap_copy_around - Prepare for an unaligned write to a shared/cow page |
| 1118 | * by copying the data before and after the range to be written. |
| 1119 | * @pos: address to do copy from. |
| 1120 | * @length: size of copy operation. |
| 1121 | * @align_size: aligned w.r.t align_size (either PMD_SIZE or PAGE_SIZE) |
| 1122 | * @srcmap: iomap srcmap |
| 1123 | * @daddr: destination address to copy to. |
| 1124 | * |
| 1125 | * This can be called from two places. Either during DAX write fault (page |
| 1126 | * aligned), to copy the length size data to daddr. Or, while doing normal DAX |
| 1127 | * write operation, dax_iomap_iter() might call this to do the copy of either |
| 1128 | * start or end unaligned address. In the latter case the rest of the copy of |
| 1129 | * aligned ranges is taken care by dax_iomap_iter() itself. |
| 1130 | * If the srcmap contains invalid data, such as HOLE and UNWRITTEN, zero the |
| 1131 | * area to make sure no old data remains. |
| 1132 | */ |
| 1133 | static int dax_iomap_copy_around(loff_t pos, uint64_t length, size_t align_size, |
| 1134 | const struct iomap *srcmap, void *daddr) |
| 1135 | { |
| 1136 | loff_t head_off = pos & (align_size - 1); |
| 1137 | size_t size = ALIGN(head_off + length, align_size); |
| 1138 | loff_t end = pos + length; |
| 1139 | loff_t pg_end = round_up(end, align_size); |
| 1140 | /* copy_all is usually in page fault case */ |
| 1141 | bool copy_all = head_off == 0 && end == pg_end; |
| 1142 | /* zero the edges if srcmap is a HOLE or IOMAP_UNWRITTEN */ |
| 1143 | bool zero_edge = srcmap->flags & IOMAP_F_SHARED || |
| 1144 | srcmap->type == IOMAP_UNWRITTEN; |
| 1145 | void *saddr = 0; |
| 1146 | int ret = 0; |
| 1147 | |
| 1148 | if (!zero_edge) { |
| 1149 | ret = dax_iomap_direct_access(srcmap, pos, size, &saddr, NULL); |
| 1150 | if (ret) |
| 1151 | return ret; |
| 1152 | } |
| 1153 | |
| 1154 | if (copy_all) { |
| 1155 | if (zero_edge) |
| 1156 | memset(daddr, 0, size); |
| 1157 | else |
| 1158 | ret = copy_mc_to_kernel(daddr, saddr, length); |
| 1159 | goto out; |
| 1160 | } |
| 1161 | |
| 1162 | /* Copy the head part of the range */ |
| 1163 | if (head_off) { |
| 1164 | if (zero_edge) |
| 1165 | memset(daddr, 0, head_off); |
| 1166 | else { |
| 1167 | ret = copy_mc_to_kernel(daddr, saddr, head_off); |
| 1168 | if (ret) |
| 1169 | return -EIO; |
| 1170 | } |
| 1171 | } |
| 1172 | |
| 1173 | /* Copy the tail part of the range */ |
| 1174 | if (end < pg_end) { |
| 1175 | loff_t tail_off = head_off + length; |
| 1176 | loff_t tail_len = pg_end - end; |
| 1177 | |
| 1178 | if (zero_edge) |
| 1179 | memset(daddr + tail_off, 0, tail_len); |
| 1180 | else { |
| 1181 | ret = copy_mc_to_kernel(daddr + tail_off, |
| 1182 | saddr + tail_off, tail_len); |
| 1183 | if (ret) |
| 1184 | return -EIO; |
| 1185 | } |
| 1186 | } |
| 1187 | out: |
| 1188 | if (zero_edge) |
| 1189 | dax_flush(srcmap->dax_dev, daddr, size); |
| 1190 | return ret ? -EIO : 0; |
| 1191 | } |
| 1192 | |
| 1193 | /* |
| 1194 | * The user has performed a load from a hole in the file. Allocating a new |
| 1195 | * page in the file would cause excessive storage usage for workloads with |
| 1196 | * sparse files. Instead we insert a read-only mapping of the 4k zero page. |
| 1197 | * If this page is ever written to we will re-fault and change the mapping to |
| 1198 | * point to real DAX storage instead. |
| 1199 | */ |
| 1200 | static vm_fault_t dax_load_hole(struct xa_state *xas, struct vm_fault *vmf, |
| 1201 | const struct iomap_iter *iter, void **entry) |
| 1202 | { |
| 1203 | struct inode *inode = iter->inode; |
| 1204 | unsigned long vaddr = vmf->address; |
| 1205 | pfn_t pfn = pfn_to_pfn_t(my_zero_pfn(vaddr)); |
| 1206 | vm_fault_t ret; |
| 1207 | |
| 1208 | *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, DAX_ZERO_PAGE); |
| 1209 | |
| 1210 | ret = vmf_insert_mixed(vmf->vma, vaddr, pfn); |
| 1211 | trace_dax_load_hole(inode, vmf, ret); |
| 1212 | return ret; |
| 1213 | } |
| 1214 | |
| 1215 | #ifdef CONFIG_FS_DAX_PMD |
| 1216 | static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, |
| 1217 | const struct iomap_iter *iter, void **entry) |
| 1218 | { |
| 1219 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
| 1220 | unsigned long pmd_addr = vmf->address & PMD_MASK; |
| 1221 | struct vm_area_struct *vma = vmf->vma; |
| 1222 | struct inode *inode = mapping->host; |
| 1223 | pgtable_t pgtable = NULL; |
| 1224 | struct page *zero_page; |
| 1225 | spinlock_t *ptl; |
| 1226 | pmd_t pmd_entry; |
| 1227 | pfn_t pfn; |
| 1228 | |
| 1229 | zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm); |
| 1230 | |
| 1231 | if (unlikely(!zero_page)) |
| 1232 | goto fallback; |
| 1233 | |
| 1234 | pfn = page_to_pfn_t(zero_page); |
| 1235 | *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, |
| 1236 | DAX_PMD | DAX_ZERO_PAGE); |
| 1237 | |
| 1238 | if (arch_needs_pgtable_deposit()) { |
| 1239 | pgtable = pte_alloc_one(vma->vm_mm); |
| 1240 | if (!pgtable) |
| 1241 | return VM_FAULT_OOM; |
| 1242 | } |
| 1243 | |
| 1244 | ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd); |
| 1245 | if (!pmd_none(*(vmf->pmd))) { |
| 1246 | spin_unlock(ptl); |
| 1247 | goto fallback; |
| 1248 | } |
| 1249 | |
| 1250 | if (pgtable) { |
| 1251 | pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable); |
| 1252 | mm_inc_nr_ptes(vma->vm_mm); |
| 1253 | } |
| 1254 | pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot); |
| 1255 | pmd_entry = pmd_mkhuge(pmd_entry); |
| 1256 | set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry); |
| 1257 | spin_unlock(ptl); |
| 1258 | trace_dax_pmd_load_hole(inode, vmf, zero_page, *entry); |
| 1259 | return VM_FAULT_NOPAGE; |
| 1260 | |
| 1261 | fallback: |
| 1262 | if (pgtable) |
| 1263 | pte_free(vma->vm_mm, pgtable); |
| 1264 | trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, *entry); |
| 1265 | return VM_FAULT_FALLBACK; |
| 1266 | } |
| 1267 | #else |
| 1268 | static vm_fault_t dax_pmd_load_hole(struct xa_state *xas, struct vm_fault *vmf, |
| 1269 | const struct iomap_iter *iter, void **entry) |
| 1270 | { |
| 1271 | return VM_FAULT_FALLBACK; |
| 1272 | } |
| 1273 | #endif /* CONFIG_FS_DAX_PMD */ |
| 1274 | |
| 1275 | static s64 dax_unshare_iter(struct iomap_iter *iter) |
| 1276 | { |
| 1277 | struct iomap *iomap = &iter->iomap; |
| 1278 | const struct iomap *srcmap = iomap_iter_srcmap(iter); |
| 1279 | loff_t pos = iter->pos; |
| 1280 | loff_t length = iomap_length(iter); |
| 1281 | int id = 0; |
| 1282 | s64 ret = 0; |
| 1283 | void *daddr = NULL, *saddr = NULL; |
| 1284 | |
| 1285 | /* don't bother with blocks that are not shared to start with */ |
| 1286 | if (!(iomap->flags & IOMAP_F_SHARED)) |
| 1287 | return length; |
| 1288 | |
| 1289 | id = dax_read_lock(); |
| 1290 | ret = dax_iomap_direct_access(iomap, pos, length, &daddr, NULL); |
| 1291 | if (ret < 0) |
| 1292 | goto out_unlock; |
| 1293 | |
| 1294 | /* zero the distance if srcmap is HOLE or UNWRITTEN */ |
| 1295 | if (srcmap->flags & IOMAP_F_SHARED || srcmap->type == IOMAP_UNWRITTEN) { |
| 1296 | memset(daddr, 0, length); |
| 1297 | dax_flush(iomap->dax_dev, daddr, length); |
| 1298 | ret = length; |
| 1299 | goto out_unlock; |
| 1300 | } |
| 1301 | |
| 1302 | ret = dax_iomap_direct_access(srcmap, pos, length, &saddr, NULL); |
| 1303 | if (ret < 0) |
| 1304 | goto out_unlock; |
| 1305 | |
| 1306 | if (copy_mc_to_kernel(daddr, saddr, length) == 0) |
| 1307 | ret = length; |
| 1308 | else |
| 1309 | ret = -EIO; |
| 1310 | |
| 1311 | out_unlock: |
| 1312 | dax_read_unlock(id); |
| 1313 | return ret; |
| 1314 | } |
| 1315 | |
| 1316 | int dax_file_unshare(struct inode *inode, loff_t pos, loff_t len, |
| 1317 | const struct iomap_ops *ops) |
| 1318 | { |
| 1319 | struct iomap_iter iter = { |
| 1320 | .inode = inode, |
| 1321 | .pos = pos, |
| 1322 | .len = len, |
| 1323 | .flags = IOMAP_WRITE | IOMAP_UNSHARE | IOMAP_DAX, |
| 1324 | }; |
| 1325 | int ret; |
| 1326 | |
| 1327 | while ((ret = iomap_iter(&iter, ops)) > 0) |
| 1328 | iter.processed = dax_unshare_iter(&iter); |
| 1329 | return ret; |
| 1330 | } |
| 1331 | EXPORT_SYMBOL_GPL(dax_file_unshare); |
| 1332 | |
| 1333 | static int dax_memzero(struct iomap_iter *iter, loff_t pos, size_t size) |
| 1334 | { |
| 1335 | const struct iomap *iomap = &iter->iomap; |
| 1336 | const struct iomap *srcmap = iomap_iter_srcmap(iter); |
| 1337 | unsigned offset = offset_in_page(pos); |
| 1338 | pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); |
| 1339 | void *kaddr; |
| 1340 | long ret; |
| 1341 | |
| 1342 | ret = dax_direct_access(iomap->dax_dev, pgoff, 1, DAX_ACCESS, &kaddr, |
| 1343 | NULL); |
| 1344 | if (ret < 0) |
| 1345 | return ret; |
| 1346 | memset(kaddr + offset, 0, size); |
| 1347 | if (iomap->flags & IOMAP_F_SHARED) |
| 1348 | ret = dax_iomap_copy_around(pos, size, PAGE_SIZE, srcmap, |
| 1349 | kaddr); |
| 1350 | else |
| 1351 | dax_flush(iomap->dax_dev, kaddr + offset, size); |
| 1352 | return ret; |
| 1353 | } |
| 1354 | |
| 1355 | static s64 dax_zero_iter(struct iomap_iter *iter, bool *did_zero) |
| 1356 | { |
| 1357 | const struct iomap *iomap = &iter->iomap; |
| 1358 | const struct iomap *srcmap = iomap_iter_srcmap(iter); |
| 1359 | loff_t pos = iter->pos; |
| 1360 | u64 length = iomap_length(iter); |
| 1361 | s64 written = 0; |
| 1362 | |
| 1363 | /* already zeroed? we're done. */ |
| 1364 | if (srcmap->type == IOMAP_HOLE || srcmap->type == IOMAP_UNWRITTEN) |
| 1365 | return length; |
| 1366 | |
| 1367 | /* |
| 1368 | * invalidate the pages whose sharing state is to be changed |
| 1369 | * because of CoW. |
| 1370 | */ |
| 1371 | if (iomap->flags & IOMAP_F_SHARED) |
| 1372 | invalidate_inode_pages2_range(iter->inode->i_mapping, |
| 1373 | pos >> PAGE_SHIFT, |
| 1374 | (pos + length - 1) >> PAGE_SHIFT); |
| 1375 | |
| 1376 | do { |
| 1377 | unsigned offset = offset_in_page(pos); |
| 1378 | unsigned size = min_t(u64, PAGE_SIZE - offset, length); |
| 1379 | pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); |
| 1380 | long rc; |
| 1381 | int id; |
| 1382 | |
| 1383 | id = dax_read_lock(); |
| 1384 | if (IS_ALIGNED(pos, PAGE_SIZE) && size == PAGE_SIZE) |
| 1385 | rc = dax_zero_page_range(iomap->dax_dev, pgoff, 1); |
| 1386 | else |
| 1387 | rc = dax_memzero(iter, pos, size); |
| 1388 | dax_read_unlock(id); |
| 1389 | |
| 1390 | if (rc < 0) |
| 1391 | return rc; |
| 1392 | pos += size; |
| 1393 | length -= size; |
| 1394 | written += size; |
| 1395 | } while (length > 0); |
| 1396 | |
| 1397 | if (did_zero) |
| 1398 | *did_zero = true; |
| 1399 | return written; |
| 1400 | } |
| 1401 | |
| 1402 | int dax_zero_range(struct inode *inode, loff_t pos, loff_t len, bool *did_zero, |
| 1403 | const struct iomap_ops *ops) |
| 1404 | { |
| 1405 | struct iomap_iter iter = { |
| 1406 | .inode = inode, |
| 1407 | .pos = pos, |
| 1408 | .len = len, |
| 1409 | .flags = IOMAP_DAX | IOMAP_ZERO, |
| 1410 | }; |
| 1411 | int ret; |
| 1412 | |
| 1413 | while ((ret = iomap_iter(&iter, ops)) > 0) |
| 1414 | iter.processed = dax_zero_iter(&iter, did_zero); |
| 1415 | return ret; |
| 1416 | } |
| 1417 | EXPORT_SYMBOL_GPL(dax_zero_range); |
| 1418 | |
| 1419 | int dax_truncate_page(struct inode *inode, loff_t pos, bool *did_zero, |
| 1420 | const struct iomap_ops *ops) |
| 1421 | { |
| 1422 | unsigned int blocksize = i_blocksize(inode); |
| 1423 | unsigned int off = pos & (blocksize - 1); |
| 1424 | |
| 1425 | /* Block boundary? Nothing to do */ |
| 1426 | if (!off) |
| 1427 | return 0; |
| 1428 | return dax_zero_range(inode, pos, blocksize - off, did_zero, ops); |
| 1429 | } |
| 1430 | EXPORT_SYMBOL_GPL(dax_truncate_page); |
| 1431 | |
| 1432 | static loff_t dax_iomap_iter(const struct iomap_iter *iomi, |
| 1433 | struct iov_iter *iter) |
| 1434 | { |
| 1435 | const struct iomap *iomap = &iomi->iomap; |
| 1436 | const struct iomap *srcmap = iomap_iter_srcmap(iomi); |
| 1437 | loff_t length = iomap_length(iomi); |
| 1438 | loff_t pos = iomi->pos; |
| 1439 | struct dax_device *dax_dev = iomap->dax_dev; |
| 1440 | loff_t end = pos + length, done = 0; |
| 1441 | bool write = iov_iter_rw(iter) == WRITE; |
| 1442 | bool cow = write && iomap->flags & IOMAP_F_SHARED; |
| 1443 | ssize_t ret = 0; |
| 1444 | size_t xfer; |
| 1445 | int id; |
| 1446 | |
| 1447 | if (!write) { |
| 1448 | end = min(end, i_size_read(iomi->inode)); |
| 1449 | if (pos >= end) |
| 1450 | return 0; |
| 1451 | |
| 1452 | if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN) |
| 1453 | return iov_iter_zero(min(length, end - pos), iter); |
| 1454 | } |
| 1455 | |
| 1456 | /* |
| 1457 | * In DAX mode, enforce either pure overwrites of written extents, or |
| 1458 | * writes to unwritten extents as part of a copy-on-write operation. |
| 1459 | */ |
| 1460 | if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED && |
| 1461 | !(iomap->flags & IOMAP_F_SHARED))) |
| 1462 | return -EIO; |
| 1463 | |
| 1464 | /* |
| 1465 | * Write can allocate block for an area which has a hole page mapped |
| 1466 | * into page tables. We have to tear down these mappings so that data |
| 1467 | * written by write(2) is visible in mmap. |
| 1468 | */ |
| 1469 | if (iomap->flags & IOMAP_F_NEW || cow) { |
| 1470 | /* |
| 1471 | * Filesystem allows CoW on non-shared extents. The src extents |
| 1472 | * may have been mmapped with dirty mark before. To be able to |
| 1473 | * invalidate its dax entries, we need to clear the dirty mark |
| 1474 | * in advance. |
| 1475 | */ |
| 1476 | if (cow) |
| 1477 | __dax_clear_dirty_range(iomi->inode->i_mapping, |
| 1478 | pos >> PAGE_SHIFT, |
| 1479 | (end - 1) >> PAGE_SHIFT); |
| 1480 | invalidate_inode_pages2_range(iomi->inode->i_mapping, |
| 1481 | pos >> PAGE_SHIFT, |
| 1482 | (end - 1) >> PAGE_SHIFT); |
| 1483 | } |
| 1484 | |
| 1485 | id = dax_read_lock(); |
| 1486 | while (pos < end) { |
| 1487 | unsigned offset = pos & (PAGE_SIZE - 1); |
| 1488 | const size_t size = ALIGN(length + offset, PAGE_SIZE); |
| 1489 | pgoff_t pgoff = dax_iomap_pgoff(iomap, pos); |
| 1490 | ssize_t map_len; |
| 1491 | bool recovery = false; |
| 1492 | void *kaddr; |
| 1493 | |
| 1494 | if (fatal_signal_pending(current)) { |
| 1495 | ret = -EINTR; |
| 1496 | break; |
| 1497 | } |
| 1498 | |
| 1499 | map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), |
| 1500 | DAX_ACCESS, &kaddr, NULL); |
| 1501 | if (map_len == -EIO && iov_iter_rw(iter) == WRITE) { |
| 1502 | map_len = dax_direct_access(dax_dev, pgoff, |
| 1503 | PHYS_PFN(size), DAX_RECOVERY_WRITE, |
| 1504 | &kaddr, NULL); |
| 1505 | if (map_len > 0) |
| 1506 | recovery = true; |
| 1507 | } |
| 1508 | if (map_len < 0) { |
| 1509 | ret = map_len; |
| 1510 | break; |
| 1511 | } |
| 1512 | |
| 1513 | if (cow) { |
| 1514 | ret = dax_iomap_copy_around(pos, length, PAGE_SIZE, |
| 1515 | srcmap, kaddr); |
| 1516 | if (ret) |
| 1517 | break; |
| 1518 | } |
| 1519 | |
| 1520 | map_len = PFN_PHYS(map_len); |
| 1521 | kaddr += offset; |
| 1522 | map_len -= offset; |
| 1523 | if (map_len > end - pos) |
| 1524 | map_len = end - pos; |
| 1525 | |
| 1526 | if (recovery) |
| 1527 | xfer = dax_recovery_write(dax_dev, pgoff, kaddr, |
| 1528 | map_len, iter); |
| 1529 | else if (write) |
| 1530 | xfer = dax_copy_from_iter(dax_dev, pgoff, kaddr, |
| 1531 | map_len, iter); |
| 1532 | else |
| 1533 | xfer = dax_copy_to_iter(dax_dev, pgoff, kaddr, |
| 1534 | map_len, iter); |
| 1535 | |
| 1536 | pos += xfer; |
| 1537 | length -= xfer; |
| 1538 | done += xfer; |
| 1539 | |
| 1540 | if (xfer == 0) |
| 1541 | ret = -EFAULT; |
| 1542 | if (xfer < map_len) |
| 1543 | break; |
| 1544 | } |
| 1545 | dax_read_unlock(id); |
| 1546 | |
| 1547 | return done ? done : ret; |
| 1548 | } |
| 1549 | |
| 1550 | /** |
| 1551 | * dax_iomap_rw - Perform I/O to a DAX file |
| 1552 | * @iocb: The control block for this I/O |
| 1553 | * @iter: The addresses to do I/O from or to |
| 1554 | * @ops: iomap ops passed from the file system |
| 1555 | * |
| 1556 | * This function performs read and write operations to directly mapped |
| 1557 | * persistent memory. The callers needs to take care of read/write exclusion |
| 1558 | * and evicting any page cache pages in the region under I/O. |
| 1559 | */ |
| 1560 | ssize_t |
| 1561 | dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter, |
| 1562 | const struct iomap_ops *ops) |
| 1563 | { |
| 1564 | struct iomap_iter iomi = { |
| 1565 | .inode = iocb->ki_filp->f_mapping->host, |
| 1566 | .pos = iocb->ki_pos, |
| 1567 | .len = iov_iter_count(iter), |
| 1568 | .flags = IOMAP_DAX, |
| 1569 | }; |
| 1570 | loff_t done = 0; |
| 1571 | int ret; |
| 1572 | |
| 1573 | if (!iomi.len) |
| 1574 | return 0; |
| 1575 | |
| 1576 | if (iov_iter_rw(iter) == WRITE) { |
| 1577 | lockdep_assert_held_write(&iomi.inode->i_rwsem); |
| 1578 | iomi.flags |= IOMAP_WRITE; |
| 1579 | } else { |
| 1580 | lockdep_assert_held(&iomi.inode->i_rwsem); |
| 1581 | } |
| 1582 | |
| 1583 | if (iocb->ki_flags & IOCB_NOWAIT) |
| 1584 | iomi.flags |= IOMAP_NOWAIT; |
| 1585 | |
| 1586 | while ((ret = iomap_iter(&iomi, ops)) > 0) |
| 1587 | iomi.processed = dax_iomap_iter(&iomi, iter); |
| 1588 | |
| 1589 | done = iomi.pos - iocb->ki_pos; |
| 1590 | iocb->ki_pos = iomi.pos; |
| 1591 | return done ? done : ret; |
| 1592 | } |
| 1593 | EXPORT_SYMBOL_GPL(dax_iomap_rw); |
| 1594 | |
| 1595 | static vm_fault_t dax_fault_return(int error) |
| 1596 | { |
| 1597 | if (error == 0) |
| 1598 | return VM_FAULT_NOPAGE; |
| 1599 | return vmf_error(error); |
| 1600 | } |
| 1601 | |
| 1602 | /* |
| 1603 | * When handling a synchronous page fault and the inode need a fsync, we can |
| 1604 | * insert the PTE/PMD into page tables only after that fsync happened. Skip |
| 1605 | * insertion for now and return the pfn so that caller can insert it after the |
| 1606 | * fsync is done. |
| 1607 | */ |
| 1608 | static vm_fault_t dax_fault_synchronous_pfnp(pfn_t *pfnp, pfn_t pfn) |
| 1609 | { |
| 1610 | if (WARN_ON_ONCE(!pfnp)) |
| 1611 | return VM_FAULT_SIGBUS; |
| 1612 | *pfnp = pfn; |
| 1613 | return VM_FAULT_NEEDDSYNC; |
| 1614 | } |
| 1615 | |
| 1616 | static vm_fault_t dax_fault_cow_page(struct vm_fault *vmf, |
| 1617 | const struct iomap_iter *iter) |
| 1618 | { |
| 1619 | vm_fault_t ret; |
| 1620 | int error = 0; |
| 1621 | |
| 1622 | switch (iter->iomap.type) { |
| 1623 | case IOMAP_HOLE: |
| 1624 | case IOMAP_UNWRITTEN: |
| 1625 | clear_user_highpage(vmf->cow_page, vmf->address); |
| 1626 | break; |
| 1627 | case IOMAP_MAPPED: |
| 1628 | error = copy_cow_page_dax(vmf, iter); |
| 1629 | break; |
| 1630 | default: |
| 1631 | WARN_ON_ONCE(1); |
| 1632 | error = -EIO; |
| 1633 | break; |
| 1634 | } |
| 1635 | |
| 1636 | if (error) |
| 1637 | return dax_fault_return(error); |
| 1638 | |
| 1639 | __SetPageUptodate(vmf->cow_page); |
| 1640 | ret = finish_fault(vmf); |
| 1641 | if (!ret) |
| 1642 | return VM_FAULT_DONE_COW; |
| 1643 | return ret; |
| 1644 | } |
| 1645 | |
| 1646 | /** |
| 1647 | * dax_fault_iter - Common actor to handle pfn insertion in PTE/PMD fault. |
| 1648 | * @vmf: vm fault instance |
| 1649 | * @iter: iomap iter |
| 1650 | * @pfnp: pfn to be returned |
| 1651 | * @xas: the dax mapping tree of a file |
| 1652 | * @entry: an unlocked dax entry to be inserted |
| 1653 | * @pmd: distinguish whether it is a pmd fault |
| 1654 | */ |
| 1655 | static vm_fault_t dax_fault_iter(struct vm_fault *vmf, |
| 1656 | const struct iomap_iter *iter, pfn_t *pfnp, |
| 1657 | struct xa_state *xas, void **entry, bool pmd) |
| 1658 | { |
| 1659 | const struct iomap *iomap = &iter->iomap; |
| 1660 | const struct iomap *srcmap = iomap_iter_srcmap(iter); |
| 1661 | size_t size = pmd ? PMD_SIZE : PAGE_SIZE; |
| 1662 | loff_t pos = (loff_t)xas->xa_index << PAGE_SHIFT; |
| 1663 | bool write = iter->flags & IOMAP_WRITE; |
| 1664 | unsigned long entry_flags = pmd ? DAX_PMD : 0; |
| 1665 | int err = 0; |
| 1666 | pfn_t pfn; |
| 1667 | void *kaddr; |
| 1668 | |
| 1669 | if (!pmd && vmf->cow_page) |
| 1670 | return dax_fault_cow_page(vmf, iter); |
| 1671 | |
| 1672 | /* if we are reading UNWRITTEN and HOLE, return a hole. */ |
| 1673 | if (!write && |
| 1674 | (iomap->type == IOMAP_UNWRITTEN || iomap->type == IOMAP_HOLE)) { |
| 1675 | if (!pmd) |
| 1676 | return dax_load_hole(xas, vmf, iter, entry); |
| 1677 | return dax_pmd_load_hole(xas, vmf, iter, entry); |
| 1678 | } |
| 1679 | |
| 1680 | if (iomap->type != IOMAP_MAPPED && !(iomap->flags & IOMAP_F_SHARED)) { |
| 1681 | WARN_ON_ONCE(1); |
| 1682 | return pmd ? VM_FAULT_FALLBACK : VM_FAULT_SIGBUS; |
| 1683 | } |
| 1684 | |
| 1685 | err = dax_iomap_direct_access(iomap, pos, size, &kaddr, &pfn); |
| 1686 | if (err) |
| 1687 | return pmd ? VM_FAULT_FALLBACK : dax_fault_return(err); |
| 1688 | |
| 1689 | *entry = dax_insert_entry(xas, vmf, iter, *entry, pfn, entry_flags); |
| 1690 | |
| 1691 | if (write && iomap->flags & IOMAP_F_SHARED) { |
| 1692 | err = dax_iomap_copy_around(pos, size, size, srcmap, kaddr); |
| 1693 | if (err) |
| 1694 | return dax_fault_return(err); |
| 1695 | } |
| 1696 | |
| 1697 | if (dax_fault_is_synchronous(iter, vmf->vma)) |
| 1698 | return dax_fault_synchronous_pfnp(pfnp, pfn); |
| 1699 | |
| 1700 | /* insert PMD pfn */ |
| 1701 | if (pmd) |
| 1702 | return vmf_insert_pfn_pmd(vmf, pfn, write); |
| 1703 | |
| 1704 | /* insert PTE pfn */ |
| 1705 | if (write) |
| 1706 | return vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); |
| 1707 | return vmf_insert_mixed(vmf->vma, vmf->address, pfn); |
| 1708 | } |
| 1709 | |
| 1710 | static vm_fault_t dax_iomap_pte_fault(struct vm_fault *vmf, pfn_t *pfnp, |
| 1711 | int *iomap_errp, const struct iomap_ops *ops) |
| 1712 | { |
| 1713 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
| 1714 | XA_STATE(xas, &mapping->i_pages, vmf->pgoff); |
| 1715 | struct iomap_iter iter = { |
| 1716 | .inode = mapping->host, |
| 1717 | .pos = (loff_t)vmf->pgoff << PAGE_SHIFT, |
| 1718 | .len = PAGE_SIZE, |
| 1719 | .flags = IOMAP_DAX | IOMAP_FAULT, |
| 1720 | }; |
| 1721 | vm_fault_t ret = 0; |
| 1722 | void *entry; |
| 1723 | int error; |
| 1724 | |
| 1725 | trace_dax_pte_fault(iter.inode, vmf, ret); |
| 1726 | /* |
| 1727 | * Check whether offset isn't beyond end of file now. Caller is supposed |
| 1728 | * to hold locks serializing us with truncate / punch hole so this is |
| 1729 | * a reliable test. |
| 1730 | */ |
| 1731 | if (iter.pos >= i_size_read(iter.inode)) { |
| 1732 | ret = VM_FAULT_SIGBUS; |
| 1733 | goto out; |
| 1734 | } |
| 1735 | |
| 1736 | if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page) |
| 1737 | iter.flags |= IOMAP_WRITE; |
| 1738 | |
| 1739 | entry = grab_mapping_entry(&xas, mapping, 0); |
| 1740 | if (xa_is_internal(entry)) { |
| 1741 | ret = xa_to_internal(entry); |
| 1742 | goto out; |
| 1743 | } |
| 1744 | |
| 1745 | /* |
| 1746 | * It is possible, particularly with mixed reads & writes to private |
| 1747 | * mappings, that we have raced with a PMD fault that overlaps with |
| 1748 | * the PTE we need to set up. If so just return and the fault will be |
| 1749 | * retried. |
| 1750 | */ |
| 1751 | if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) { |
| 1752 | ret = VM_FAULT_NOPAGE; |
| 1753 | goto unlock_entry; |
| 1754 | } |
| 1755 | |
| 1756 | while ((error = iomap_iter(&iter, ops)) > 0) { |
| 1757 | if (WARN_ON_ONCE(iomap_length(&iter) < PAGE_SIZE)) { |
| 1758 | iter.processed = -EIO; /* fs corruption? */ |
| 1759 | continue; |
| 1760 | } |
| 1761 | |
| 1762 | ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, false); |
| 1763 | if (ret != VM_FAULT_SIGBUS && |
| 1764 | (iter.iomap.flags & IOMAP_F_NEW)) { |
| 1765 | count_vm_event(PGMAJFAULT); |
| 1766 | count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT); |
| 1767 | ret |= VM_FAULT_MAJOR; |
| 1768 | } |
| 1769 | |
| 1770 | if (!(ret & VM_FAULT_ERROR)) |
| 1771 | iter.processed = PAGE_SIZE; |
| 1772 | } |
| 1773 | |
| 1774 | if (iomap_errp) |
| 1775 | *iomap_errp = error; |
| 1776 | if (!ret && error) |
| 1777 | ret = dax_fault_return(error); |
| 1778 | |
| 1779 | unlock_entry: |
| 1780 | dax_unlock_entry(&xas, entry); |
| 1781 | out: |
| 1782 | trace_dax_pte_fault_done(iter.inode, vmf, ret); |
| 1783 | return ret; |
| 1784 | } |
| 1785 | |
| 1786 | #ifdef CONFIG_FS_DAX_PMD |
| 1787 | static bool dax_fault_check_fallback(struct vm_fault *vmf, struct xa_state *xas, |
| 1788 | pgoff_t max_pgoff) |
| 1789 | { |
| 1790 | unsigned long pmd_addr = vmf->address & PMD_MASK; |
| 1791 | bool write = vmf->flags & FAULT_FLAG_WRITE; |
| 1792 | |
| 1793 | /* |
| 1794 | * Make sure that the faulting address's PMD offset (color) matches |
| 1795 | * the PMD offset from the start of the file. This is necessary so |
| 1796 | * that a PMD range in the page table overlaps exactly with a PMD |
| 1797 | * range in the page cache. |
| 1798 | */ |
| 1799 | if ((vmf->pgoff & PG_PMD_COLOUR) != |
| 1800 | ((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR)) |
| 1801 | return true; |
| 1802 | |
| 1803 | /* Fall back to PTEs if we're going to COW */ |
| 1804 | if (write && !(vmf->vma->vm_flags & VM_SHARED)) |
| 1805 | return true; |
| 1806 | |
| 1807 | /* If the PMD would extend outside the VMA */ |
| 1808 | if (pmd_addr < vmf->vma->vm_start) |
| 1809 | return true; |
| 1810 | if ((pmd_addr + PMD_SIZE) > vmf->vma->vm_end) |
| 1811 | return true; |
| 1812 | |
| 1813 | /* If the PMD would extend beyond the file size */ |
| 1814 | if ((xas->xa_index | PG_PMD_COLOUR) >= max_pgoff) |
| 1815 | return true; |
| 1816 | |
| 1817 | return false; |
| 1818 | } |
| 1819 | |
| 1820 | static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, |
| 1821 | const struct iomap_ops *ops) |
| 1822 | { |
| 1823 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
| 1824 | XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, PMD_ORDER); |
| 1825 | struct iomap_iter iter = { |
| 1826 | .inode = mapping->host, |
| 1827 | .len = PMD_SIZE, |
| 1828 | .flags = IOMAP_DAX | IOMAP_FAULT, |
| 1829 | }; |
| 1830 | vm_fault_t ret = VM_FAULT_FALLBACK; |
| 1831 | pgoff_t max_pgoff; |
| 1832 | void *entry; |
| 1833 | int error; |
| 1834 | |
| 1835 | if (vmf->flags & FAULT_FLAG_WRITE) |
| 1836 | iter.flags |= IOMAP_WRITE; |
| 1837 | |
| 1838 | /* |
| 1839 | * Check whether offset isn't beyond end of file now. Caller is |
| 1840 | * supposed to hold locks serializing us with truncate / punch hole so |
| 1841 | * this is a reliable test. |
| 1842 | */ |
| 1843 | max_pgoff = DIV_ROUND_UP(i_size_read(iter.inode), PAGE_SIZE); |
| 1844 | |
| 1845 | trace_dax_pmd_fault(iter.inode, vmf, max_pgoff, 0); |
| 1846 | |
| 1847 | if (xas.xa_index >= max_pgoff) { |
| 1848 | ret = VM_FAULT_SIGBUS; |
| 1849 | goto out; |
| 1850 | } |
| 1851 | |
| 1852 | if (dax_fault_check_fallback(vmf, &xas, max_pgoff)) |
| 1853 | goto fallback; |
| 1854 | |
| 1855 | /* |
| 1856 | * grab_mapping_entry() will make sure we get an empty PMD entry, |
| 1857 | * a zero PMD entry or a DAX PMD. If it can't (because a PTE |
| 1858 | * entry is already in the array, for instance), it will return |
| 1859 | * VM_FAULT_FALLBACK. |
| 1860 | */ |
| 1861 | entry = grab_mapping_entry(&xas, mapping, PMD_ORDER); |
| 1862 | if (xa_is_internal(entry)) { |
| 1863 | ret = xa_to_internal(entry); |
| 1864 | goto fallback; |
| 1865 | } |
| 1866 | |
| 1867 | /* |
| 1868 | * It is possible, particularly with mixed reads & writes to private |
| 1869 | * mappings, that we have raced with a PTE fault that overlaps with |
| 1870 | * the PMD we need to set up. If so just return and the fault will be |
| 1871 | * retried. |
| 1872 | */ |
| 1873 | if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) && |
| 1874 | !pmd_devmap(*vmf->pmd)) { |
| 1875 | ret = 0; |
| 1876 | goto unlock_entry; |
| 1877 | } |
| 1878 | |
| 1879 | iter.pos = (loff_t)xas.xa_index << PAGE_SHIFT; |
| 1880 | while ((error = iomap_iter(&iter, ops)) > 0) { |
| 1881 | if (iomap_length(&iter) < PMD_SIZE) |
| 1882 | continue; /* actually breaks out of the loop */ |
| 1883 | |
| 1884 | ret = dax_fault_iter(vmf, &iter, pfnp, &xas, &entry, true); |
| 1885 | if (ret != VM_FAULT_FALLBACK) |
| 1886 | iter.processed = PMD_SIZE; |
| 1887 | } |
| 1888 | |
| 1889 | unlock_entry: |
| 1890 | dax_unlock_entry(&xas, entry); |
| 1891 | fallback: |
| 1892 | if (ret == VM_FAULT_FALLBACK) { |
| 1893 | split_huge_pmd(vmf->vma, vmf->pmd, vmf->address); |
| 1894 | count_vm_event(THP_FAULT_FALLBACK); |
| 1895 | } |
| 1896 | out: |
| 1897 | trace_dax_pmd_fault_done(iter.inode, vmf, max_pgoff, ret); |
| 1898 | return ret; |
| 1899 | } |
| 1900 | #else |
| 1901 | static vm_fault_t dax_iomap_pmd_fault(struct vm_fault *vmf, pfn_t *pfnp, |
| 1902 | const struct iomap_ops *ops) |
| 1903 | { |
| 1904 | return VM_FAULT_FALLBACK; |
| 1905 | } |
| 1906 | #endif /* CONFIG_FS_DAX_PMD */ |
| 1907 | |
| 1908 | /** |
| 1909 | * dax_iomap_fault - handle a page fault on a DAX file |
| 1910 | * @vmf: The description of the fault |
| 1911 | * @pe_size: Size of the page to fault in |
| 1912 | * @pfnp: PFN to insert for synchronous faults if fsync is required |
| 1913 | * @iomap_errp: Storage for detailed error code in case of error |
| 1914 | * @ops: Iomap ops passed from the file system |
| 1915 | * |
| 1916 | * When a page fault occurs, filesystems may call this helper in |
| 1917 | * their fault handler for DAX files. dax_iomap_fault() assumes the caller |
| 1918 | * has done all the necessary locking for page fault to proceed |
| 1919 | * successfully. |
| 1920 | */ |
| 1921 | vm_fault_t dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size, |
| 1922 | pfn_t *pfnp, int *iomap_errp, const struct iomap_ops *ops) |
| 1923 | { |
| 1924 | switch (pe_size) { |
| 1925 | case PE_SIZE_PTE: |
| 1926 | return dax_iomap_pte_fault(vmf, pfnp, iomap_errp, ops); |
| 1927 | case PE_SIZE_PMD: |
| 1928 | return dax_iomap_pmd_fault(vmf, pfnp, ops); |
| 1929 | default: |
| 1930 | return VM_FAULT_FALLBACK; |
| 1931 | } |
| 1932 | } |
| 1933 | EXPORT_SYMBOL_GPL(dax_iomap_fault); |
| 1934 | |
| 1935 | /* |
| 1936 | * dax_insert_pfn_mkwrite - insert PTE or PMD entry into page tables |
| 1937 | * @vmf: The description of the fault |
| 1938 | * @pfn: PFN to insert |
| 1939 | * @order: Order of entry to insert. |
| 1940 | * |
| 1941 | * This function inserts a writeable PTE or PMD entry into the page tables |
| 1942 | * for an mmaped DAX file. It also marks the page cache entry as dirty. |
| 1943 | */ |
| 1944 | static vm_fault_t |
| 1945 | dax_insert_pfn_mkwrite(struct vm_fault *vmf, pfn_t pfn, unsigned int order) |
| 1946 | { |
| 1947 | struct address_space *mapping = vmf->vma->vm_file->f_mapping; |
| 1948 | XA_STATE_ORDER(xas, &mapping->i_pages, vmf->pgoff, order); |
| 1949 | void *entry; |
| 1950 | vm_fault_t ret; |
| 1951 | |
| 1952 | xas_lock_irq(&xas); |
| 1953 | entry = get_unlocked_entry(&xas, order); |
| 1954 | /* Did we race with someone splitting entry or so? */ |
| 1955 | if (!entry || dax_is_conflict(entry) || |
| 1956 | (order == 0 && !dax_is_pte_entry(entry))) { |
| 1957 | put_unlocked_entry(&xas, entry, WAKE_NEXT); |
| 1958 | xas_unlock_irq(&xas); |
| 1959 | trace_dax_insert_pfn_mkwrite_no_entry(mapping->host, vmf, |
| 1960 | VM_FAULT_NOPAGE); |
| 1961 | return VM_FAULT_NOPAGE; |
| 1962 | } |
| 1963 | xas_set_mark(&xas, PAGECACHE_TAG_DIRTY); |
| 1964 | dax_lock_entry(&xas, entry); |
| 1965 | xas_unlock_irq(&xas); |
| 1966 | if (order == 0) |
| 1967 | ret = vmf_insert_mixed_mkwrite(vmf->vma, vmf->address, pfn); |
| 1968 | #ifdef CONFIG_FS_DAX_PMD |
| 1969 | else if (order == PMD_ORDER) |
| 1970 | ret = vmf_insert_pfn_pmd(vmf, pfn, FAULT_FLAG_WRITE); |
| 1971 | #endif |
| 1972 | else |
| 1973 | ret = VM_FAULT_FALLBACK; |
| 1974 | dax_unlock_entry(&xas, entry); |
| 1975 | trace_dax_insert_pfn_mkwrite(mapping->host, vmf, ret); |
| 1976 | return ret; |
| 1977 | } |
| 1978 | |
| 1979 | /** |
| 1980 | * dax_finish_sync_fault - finish synchronous page fault |
| 1981 | * @vmf: The description of the fault |
| 1982 | * @pe_size: Size of entry to be inserted |
| 1983 | * @pfn: PFN to insert |
| 1984 | * |
| 1985 | * This function ensures that the file range touched by the page fault is |
| 1986 | * stored persistently on the media and handles inserting of appropriate page |
| 1987 | * table entry. |
| 1988 | */ |
| 1989 | vm_fault_t dax_finish_sync_fault(struct vm_fault *vmf, |
| 1990 | enum page_entry_size pe_size, pfn_t pfn) |
| 1991 | { |
| 1992 | int err; |
| 1993 | loff_t start = ((loff_t)vmf->pgoff) << PAGE_SHIFT; |
| 1994 | unsigned int order = pe_order(pe_size); |
| 1995 | size_t len = PAGE_SIZE << order; |
| 1996 | |
| 1997 | err = vfs_fsync_range(vmf->vma->vm_file, start, start + len - 1, 1); |
| 1998 | if (err) |
| 1999 | return VM_FAULT_SIGBUS; |
| 2000 | return dax_insert_pfn_mkwrite(vmf, pfn, order); |
| 2001 | } |
| 2002 | EXPORT_SYMBOL_GPL(dax_finish_sync_fault); |
| 2003 | |
| 2004 | static loff_t dax_range_compare_iter(struct iomap_iter *it_src, |
| 2005 | struct iomap_iter *it_dest, u64 len, bool *same) |
| 2006 | { |
| 2007 | const struct iomap *smap = &it_src->iomap; |
| 2008 | const struct iomap *dmap = &it_dest->iomap; |
| 2009 | loff_t pos1 = it_src->pos, pos2 = it_dest->pos; |
| 2010 | void *saddr, *daddr; |
| 2011 | int id, ret; |
| 2012 | |
| 2013 | len = min(len, min(smap->length, dmap->length)); |
| 2014 | |
| 2015 | if (smap->type == IOMAP_HOLE && dmap->type == IOMAP_HOLE) { |
| 2016 | *same = true; |
| 2017 | return len; |
| 2018 | } |
| 2019 | |
| 2020 | if (smap->type == IOMAP_HOLE || dmap->type == IOMAP_HOLE) { |
| 2021 | *same = false; |
| 2022 | return 0; |
| 2023 | } |
| 2024 | |
| 2025 | id = dax_read_lock(); |
| 2026 | ret = dax_iomap_direct_access(smap, pos1, ALIGN(pos1 + len, PAGE_SIZE), |
| 2027 | &saddr, NULL); |
| 2028 | if (ret < 0) |
| 2029 | goto out_unlock; |
| 2030 | |
| 2031 | ret = dax_iomap_direct_access(dmap, pos2, ALIGN(pos2 + len, PAGE_SIZE), |
| 2032 | &daddr, NULL); |
| 2033 | if (ret < 0) |
| 2034 | goto out_unlock; |
| 2035 | |
| 2036 | *same = !memcmp(saddr, daddr, len); |
| 2037 | if (!*same) |
| 2038 | len = 0; |
| 2039 | dax_read_unlock(id); |
| 2040 | return len; |
| 2041 | |
| 2042 | out_unlock: |
| 2043 | dax_read_unlock(id); |
| 2044 | return -EIO; |
| 2045 | } |
| 2046 | |
| 2047 | int dax_dedupe_file_range_compare(struct inode *src, loff_t srcoff, |
| 2048 | struct inode *dst, loff_t dstoff, loff_t len, bool *same, |
| 2049 | const struct iomap_ops *ops) |
| 2050 | { |
| 2051 | struct iomap_iter src_iter = { |
| 2052 | .inode = src, |
| 2053 | .pos = srcoff, |
| 2054 | .len = len, |
| 2055 | .flags = IOMAP_DAX, |
| 2056 | }; |
| 2057 | struct iomap_iter dst_iter = { |
| 2058 | .inode = dst, |
| 2059 | .pos = dstoff, |
| 2060 | .len = len, |
| 2061 | .flags = IOMAP_DAX, |
| 2062 | }; |
| 2063 | int ret, compared = 0; |
| 2064 | |
| 2065 | while ((ret = iomap_iter(&src_iter, ops)) > 0 && |
| 2066 | (ret = iomap_iter(&dst_iter, ops)) > 0) { |
| 2067 | compared = dax_range_compare_iter(&src_iter, &dst_iter, |
| 2068 | min(src_iter.len, dst_iter.len), same); |
| 2069 | if (compared < 0) |
| 2070 | return ret; |
| 2071 | src_iter.processed = dst_iter.processed = compared; |
| 2072 | } |
| 2073 | return ret; |
| 2074 | } |
| 2075 | |
| 2076 | int dax_remap_file_range_prep(struct file *file_in, loff_t pos_in, |
| 2077 | struct file *file_out, loff_t pos_out, |
| 2078 | loff_t *len, unsigned int remap_flags, |
| 2079 | const struct iomap_ops *ops) |
| 2080 | { |
| 2081 | return __generic_remap_file_range_prep(file_in, pos_in, file_out, |
| 2082 | pos_out, len, remap_flags, ops); |
| 2083 | } |
| 2084 | EXPORT_SYMBOL_GPL(dax_remap_file_range_prep); |