1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Copyright (C) 2017-2023 Oracle. All Rights Reserved. 4 * Author: Darrick J. Wong <djwong@kernel.org> 5 */ 6 #include "xfs.h" 7 #include "xfs_fs.h" 8 #include "xfs_shared.h" 9 #include "xfs_format.h" 10 #include "xfs_trans_resv.h" 11 #include "xfs_mount.h" 12 #include "xfs_log_format.h" 13 #include "xfs_trans.h" 14 #include "xfs_inode.h" 15 #include "xfs_quota.h" 16 #include "xfs_qm.h" 17 #include "xfs_errortag.h" 18 #include "xfs_error.h" 19 #include "xfs_scrub.h" 20 #include "scrub/scrub.h" 21 #include "scrub/common.h" 22 #include "scrub/trace.h" 23 #include "scrub/repair.h" 24 #include "scrub/health.h" 25 #include "scrub/stats.h" 26 #include "scrub/xfile.h" 27 28 /* 29 * Online Scrub and Repair 30 * 31 * Traditionally, XFS (the kernel driver) did not know how to check or 32 * repair on-disk data structures. That task was left to the xfs_check 33 * and xfs_repair tools, both of which require taking the filesystem 34 * offline for a thorough but time consuming examination. Online 35 * scrub & repair, on the other hand, enables us to check the metadata 36 * for obvious errors while carefully stepping around the filesystem's 37 * ongoing operations, locking rules, etc. 38 * 39 * Given that most XFS metadata consist of records stored in a btree, 40 * most of the checking functions iterate the btree blocks themselves 41 * looking for irregularities. When a record block is encountered, each 42 * record can be checked for obviously bad values. Record values can 43 * also be cross-referenced against other btrees to look for potential 44 * misunderstandings between pieces of metadata. 45 * 46 * It is expected that the checkers responsible for per-AG metadata 47 * structures will lock the AG headers (AGI, AGF, AGFL), iterate the 48 * metadata structure, and perform any relevant cross-referencing before 49 * unlocking the AG and returning the results to userspace. These 50 * scrubbers must not keep an AG locked for too long to avoid tying up 51 * the block and inode allocators. 52 * 53 * Block maps and b-trees rooted in an inode present a special challenge 54 * because they can involve extents from any AG. The general scrubber 55 * structure of lock -> check -> xref -> unlock still holds, but AG 56 * locking order rules /must/ be obeyed to avoid deadlocks. The 57 * ordering rule, of course, is that we must lock in increasing AG 58 * order. Helper functions are provided to track which AG headers we've 59 * already locked. If we detect an imminent locking order violation, we 60 * can signal a potential deadlock, in which case the scrubber can jump 61 * out to the top level, lock all the AGs in order, and retry the scrub. 62 * 63 * For file data (directories, extended attributes, symlinks) scrub, we 64 * can simply lock the inode and walk the data. For btree data 65 * (directories and attributes) we follow the same btree-scrubbing 66 * strategy outlined previously to check the records. 67 * 68 * We use a bit of trickery with transactions to avoid buffer deadlocks 69 * if there is a cycle in the metadata. The basic problem is that 70 * travelling down a btree involves locking the current buffer at each 71 * tree level. If a pointer should somehow point back to a buffer that 72 * we've already examined, we will deadlock due to the second buffer 73 * locking attempt. Note however that grabbing a buffer in transaction 74 * context links the locked buffer to the transaction. If we try to 75 * re-grab the buffer in the context of the same transaction, we avoid 76 * the second lock attempt and continue. Between the verifier and the 77 * scrubber, something will notice that something is amiss and report 78 * the corruption. Therefore, each scrubber will allocate an empty 79 * transaction, attach buffers to it, and cancel the transaction at the 80 * end of the scrub run. Cancelling a non-dirty transaction simply 81 * unlocks the buffers. 82 * 83 * There are four pieces of data that scrub can communicate to 84 * userspace. The first is the error code (errno), which can be used to 85 * communicate operational errors in performing the scrub. There are 86 * also three flags that can be set in the scrub context. If the data 87 * structure itself is corrupt, the CORRUPT flag will be set. If 88 * the metadata is correct but otherwise suboptimal, the PREEN flag 89 * will be set. 90 * 91 * We perform secondary validation of filesystem metadata by 92 * cross-referencing every record with all other available metadata. 93 * For example, for block mapping extents, we verify that there are no 94 * records in the free space and inode btrees corresponding to that 95 * space extent and that there is a corresponding entry in the reverse 96 * mapping btree. Inconsistent metadata is noted by setting the 97 * XCORRUPT flag; btree query function errors are noted by setting the 98 * XFAIL flag and deleting the cursor to prevent further attempts to 99 * cross-reference with a defective btree. 100 * 101 * If a piece of metadata proves corrupt or suboptimal, the userspace 102 * program can ask the kernel to apply some tender loving care (TLC) to 103 * the metadata object by setting the REPAIR flag and re-calling the 104 * scrub ioctl. "Corruption" is defined by metadata violating the 105 * on-disk specification; operations cannot continue if the violation is 106 * left untreated. It is possible for XFS to continue if an object is 107 * "suboptimal", however performance may be degraded. Repairs are 108 * usually performed by rebuilding the metadata entirely out of 109 * redundant metadata. Optimizing, on the other hand, can sometimes be 110 * done without rebuilding entire structures. 111 * 112 * Generally speaking, the repair code has the following code structure: 113 * Lock -> scrub -> repair -> commit -> re-lock -> re-scrub -> unlock. 114 * The first check helps us figure out if we need to rebuild or simply 115 * optimize the structure so that the rebuild knows what to do. The 116 * second check evaluates the completeness of the repair; that is what 117 * is reported to userspace. 118 * 119 * A quick note on symbol prefixes: 120 * - "xfs_" are general XFS symbols. 121 * - "xchk_" are symbols related to metadata checking. 122 * - "xrep_" are symbols related to metadata repair. 123 * - "xfs_scrub_" are symbols that tie online fsck to the rest of XFS. 124 */ 125 126 /* 127 * Scrub probe -- userspace uses this to probe if we're willing to scrub 128 * or repair a given mountpoint. This will be used by xfs_scrub to 129 * probe the kernel's abilities to scrub (and repair) the metadata. We 130 * do this by validating the ioctl inputs from userspace, preparing the 131 * filesystem for a scrub (or a repair) operation, and immediately 132 * returning to userspace. Userspace can use the returned errno and 133 * structure state to decide (in broad terms) if scrub/repair are 134 * supported by the running kernel. 135 */ 136 static int 137 xchk_probe( 138 struct xfs_scrub *sc) 139 { 140 int error = 0; 141 142 if (xchk_should_terminate(sc, &error)) 143 return error; 144 145 return 0; 146 } 147 148 /* Scrub setup and teardown */ 149 150 static inline void 151 xchk_fsgates_disable( 152 struct xfs_scrub *sc) 153 { 154 if (!(sc->flags & XCHK_FSGATES_ALL)) 155 return; 156 157 trace_xchk_fsgates_disable(sc, sc->flags & XCHK_FSGATES_ALL); 158 159 if (sc->flags & XCHK_FSGATES_DRAIN) 160 xfs_drain_wait_disable(); 161 162 sc->flags &= ~XCHK_FSGATES_ALL; 163 } 164 165 /* Free all the resources and finish the transactions. */ 166 STATIC int 167 xchk_teardown( 168 struct xfs_scrub *sc, 169 int error) 170 { 171 xchk_ag_free(sc, &sc->sa); 172 if (sc->tp) { 173 if (error == 0 && (sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR)) 174 error = xfs_trans_commit(sc->tp); 175 else 176 xfs_trans_cancel(sc->tp); 177 sc->tp = NULL; 178 } 179 if (sc->ip) { 180 if (sc->ilock_flags) 181 xchk_iunlock(sc, sc->ilock_flags); 182 xchk_irele(sc, sc->ip); 183 sc->ip = NULL; 184 } 185 if (sc->flags & XCHK_HAVE_FREEZE_PROT) { 186 sc->flags &= ~XCHK_HAVE_FREEZE_PROT; 187 mnt_drop_write_file(sc->file); 188 } 189 if (sc->xfile) { 190 xfile_destroy(sc->xfile); 191 sc->xfile = NULL; 192 } 193 if (sc->buf) { 194 if (sc->buf_cleanup) 195 sc->buf_cleanup(sc->buf); 196 kvfree(sc->buf); 197 sc->buf_cleanup = NULL; 198 sc->buf = NULL; 199 } 200 201 xchk_fsgates_disable(sc); 202 return error; 203 } 204 205 /* Scrubbing dispatch. */ 206 207 static const struct xchk_meta_ops meta_scrub_ops[] = { 208 [XFS_SCRUB_TYPE_PROBE] = { /* ioctl presence test */ 209 .type = ST_NONE, 210 .setup = xchk_setup_fs, 211 .scrub = xchk_probe, 212 .repair = xrep_probe, 213 }, 214 [XFS_SCRUB_TYPE_SB] = { /* superblock */ 215 .type = ST_PERAG, 216 .setup = xchk_setup_agheader, 217 .scrub = xchk_superblock, 218 .repair = xrep_superblock, 219 }, 220 [XFS_SCRUB_TYPE_AGF] = { /* agf */ 221 .type = ST_PERAG, 222 .setup = xchk_setup_agheader, 223 .scrub = xchk_agf, 224 .repair = xrep_agf, 225 }, 226 [XFS_SCRUB_TYPE_AGFL]= { /* agfl */ 227 .type = ST_PERAG, 228 .setup = xchk_setup_agheader, 229 .scrub = xchk_agfl, 230 .repair = xrep_agfl, 231 }, 232 [XFS_SCRUB_TYPE_AGI] = { /* agi */ 233 .type = ST_PERAG, 234 .setup = xchk_setup_agheader, 235 .scrub = xchk_agi, 236 .repair = xrep_agi, 237 }, 238 [XFS_SCRUB_TYPE_BNOBT] = { /* bnobt */ 239 .type = ST_PERAG, 240 .setup = xchk_setup_ag_allocbt, 241 .scrub = xchk_bnobt, 242 .repair = xrep_notsupported, 243 }, 244 [XFS_SCRUB_TYPE_CNTBT] = { /* cntbt */ 245 .type = ST_PERAG, 246 .setup = xchk_setup_ag_allocbt, 247 .scrub = xchk_cntbt, 248 .repair = xrep_notsupported, 249 }, 250 [XFS_SCRUB_TYPE_INOBT] = { /* inobt */ 251 .type = ST_PERAG, 252 .setup = xchk_setup_ag_iallocbt, 253 .scrub = xchk_inobt, 254 .repair = xrep_notsupported, 255 }, 256 [XFS_SCRUB_TYPE_FINOBT] = { /* finobt */ 257 .type = ST_PERAG, 258 .setup = xchk_setup_ag_iallocbt, 259 .scrub = xchk_finobt, 260 .has = xfs_has_finobt, 261 .repair = xrep_notsupported, 262 }, 263 [XFS_SCRUB_TYPE_RMAPBT] = { /* rmapbt */ 264 .type = ST_PERAG, 265 .setup = xchk_setup_ag_rmapbt, 266 .scrub = xchk_rmapbt, 267 .has = xfs_has_rmapbt, 268 .repair = xrep_notsupported, 269 }, 270 [XFS_SCRUB_TYPE_REFCNTBT] = { /* refcountbt */ 271 .type = ST_PERAG, 272 .setup = xchk_setup_ag_refcountbt, 273 .scrub = xchk_refcountbt, 274 .has = xfs_has_reflink, 275 .repair = xrep_notsupported, 276 }, 277 [XFS_SCRUB_TYPE_INODE] = { /* inode record */ 278 .type = ST_INODE, 279 .setup = xchk_setup_inode, 280 .scrub = xchk_inode, 281 .repair = xrep_notsupported, 282 }, 283 [XFS_SCRUB_TYPE_BMBTD] = { /* inode data fork */ 284 .type = ST_INODE, 285 .setup = xchk_setup_inode_bmap, 286 .scrub = xchk_bmap_data, 287 .repair = xrep_notsupported, 288 }, 289 [XFS_SCRUB_TYPE_BMBTA] = { /* inode attr fork */ 290 .type = ST_INODE, 291 .setup = xchk_setup_inode_bmap, 292 .scrub = xchk_bmap_attr, 293 .repair = xrep_notsupported, 294 }, 295 [XFS_SCRUB_TYPE_BMBTC] = { /* inode CoW fork */ 296 .type = ST_INODE, 297 .setup = xchk_setup_inode_bmap, 298 .scrub = xchk_bmap_cow, 299 .repair = xrep_notsupported, 300 }, 301 [XFS_SCRUB_TYPE_DIR] = { /* directory */ 302 .type = ST_INODE, 303 .setup = xchk_setup_directory, 304 .scrub = xchk_directory, 305 .repair = xrep_notsupported, 306 }, 307 [XFS_SCRUB_TYPE_XATTR] = { /* extended attributes */ 308 .type = ST_INODE, 309 .setup = xchk_setup_xattr, 310 .scrub = xchk_xattr, 311 .repair = xrep_notsupported, 312 }, 313 [XFS_SCRUB_TYPE_SYMLINK] = { /* symbolic link */ 314 .type = ST_INODE, 315 .setup = xchk_setup_symlink, 316 .scrub = xchk_symlink, 317 .repair = xrep_notsupported, 318 }, 319 [XFS_SCRUB_TYPE_PARENT] = { /* parent pointers */ 320 .type = ST_INODE, 321 .setup = xchk_setup_parent, 322 .scrub = xchk_parent, 323 .repair = xrep_notsupported, 324 }, 325 [XFS_SCRUB_TYPE_RTBITMAP] = { /* realtime bitmap */ 326 .type = ST_FS, 327 .setup = xchk_setup_rtbitmap, 328 .scrub = xchk_rtbitmap, 329 .has = xfs_has_realtime, 330 .repair = xrep_notsupported, 331 }, 332 [XFS_SCRUB_TYPE_RTSUM] = { /* realtime summary */ 333 .type = ST_FS, 334 .setup = xchk_setup_rtsummary, 335 .scrub = xchk_rtsummary, 336 .has = xfs_has_realtime, 337 .repair = xrep_notsupported, 338 }, 339 [XFS_SCRUB_TYPE_UQUOTA] = { /* user quota */ 340 .type = ST_FS, 341 .setup = xchk_setup_quota, 342 .scrub = xchk_quota, 343 .repair = xrep_notsupported, 344 }, 345 [XFS_SCRUB_TYPE_GQUOTA] = { /* group quota */ 346 .type = ST_FS, 347 .setup = xchk_setup_quota, 348 .scrub = xchk_quota, 349 .repair = xrep_notsupported, 350 }, 351 [XFS_SCRUB_TYPE_PQUOTA] = { /* project quota */ 352 .type = ST_FS, 353 .setup = xchk_setup_quota, 354 .scrub = xchk_quota, 355 .repair = xrep_notsupported, 356 }, 357 [XFS_SCRUB_TYPE_FSCOUNTERS] = { /* fs summary counters */ 358 .type = ST_FS, 359 .setup = xchk_setup_fscounters, 360 .scrub = xchk_fscounters, 361 .repair = xrep_notsupported, 362 }, 363 }; 364 365 static int 366 xchk_validate_inputs( 367 struct xfs_mount *mp, 368 struct xfs_scrub_metadata *sm) 369 { 370 int error; 371 const struct xchk_meta_ops *ops; 372 373 error = -EINVAL; 374 /* Check our inputs. */ 375 sm->sm_flags &= ~XFS_SCRUB_FLAGS_OUT; 376 if (sm->sm_flags & ~XFS_SCRUB_FLAGS_IN) 377 goto out; 378 /* sm_reserved[] must be zero */ 379 if (memchr_inv(sm->sm_reserved, 0, sizeof(sm->sm_reserved))) 380 goto out; 381 382 error = -ENOENT; 383 /* Do we know about this type of metadata? */ 384 if (sm->sm_type >= XFS_SCRUB_TYPE_NR) 385 goto out; 386 ops = &meta_scrub_ops[sm->sm_type]; 387 if (ops->setup == NULL || ops->scrub == NULL) 388 goto out; 389 /* Does this fs even support this type of metadata? */ 390 if (ops->has && !ops->has(mp)) 391 goto out; 392 393 error = -EINVAL; 394 /* restricting fields must be appropriate for type */ 395 switch (ops->type) { 396 case ST_NONE: 397 case ST_FS: 398 if (sm->sm_ino || sm->sm_gen || sm->sm_agno) 399 goto out; 400 break; 401 case ST_PERAG: 402 if (sm->sm_ino || sm->sm_gen || 403 sm->sm_agno >= mp->m_sb.sb_agcount) 404 goto out; 405 break; 406 case ST_INODE: 407 if (sm->sm_agno || (sm->sm_gen && !sm->sm_ino)) 408 goto out; 409 break; 410 default: 411 goto out; 412 } 413 414 /* No rebuild without repair. */ 415 if ((sm->sm_flags & XFS_SCRUB_IFLAG_FORCE_REBUILD) && 416 !(sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR)) 417 return -EINVAL; 418 419 /* 420 * We only want to repair read-write v5+ filesystems. Defer the check 421 * for ops->repair until after our scrub confirms that we need to 422 * perform repairs so that we avoid failing due to not supporting 423 * repairing an object that doesn't need repairs. 424 */ 425 if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) { 426 error = -EOPNOTSUPP; 427 if (!xfs_has_crc(mp)) 428 goto out; 429 430 error = -EROFS; 431 if (xfs_is_readonly(mp)) 432 goto out; 433 } 434 435 error = 0; 436 out: 437 return error; 438 } 439 440 #ifdef CONFIG_XFS_ONLINE_REPAIR 441 static inline void xchk_postmortem(struct xfs_scrub *sc) 442 { 443 /* 444 * Userspace asked us to repair something, we repaired it, rescanned 445 * it, and the rescan says it's still broken. Scream about this in 446 * the system logs. 447 */ 448 if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) && 449 (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT | 450 XFS_SCRUB_OFLAG_XCORRUPT))) 451 xrep_failure(sc->mp); 452 } 453 #else 454 static inline void xchk_postmortem(struct xfs_scrub *sc) 455 { 456 /* 457 * Userspace asked us to scrub something, it's broken, and we have no 458 * way of fixing it. Scream in the logs. 459 */ 460 if (sc->sm->sm_flags & (XFS_SCRUB_OFLAG_CORRUPT | 461 XFS_SCRUB_OFLAG_XCORRUPT)) 462 xfs_alert_ratelimited(sc->mp, 463 "Corruption detected during scrub."); 464 } 465 #endif /* CONFIG_XFS_ONLINE_REPAIR */ 466 467 /* Dispatch metadata scrubbing. */ 468 int 469 xfs_scrub_metadata( 470 struct file *file, 471 struct xfs_scrub_metadata *sm) 472 { 473 struct xchk_stats_run run = { }; 474 struct xfs_scrub *sc; 475 struct xfs_mount *mp = XFS_I(file_inode(file))->i_mount; 476 u64 check_start; 477 int error = 0; 478 479 BUILD_BUG_ON(sizeof(meta_scrub_ops) != 480 (sizeof(struct xchk_meta_ops) * XFS_SCRUB_TYPE_NR)); 481 482 trace_xchk_start(XFS_I(file_inode(file)), sm, error); 483 484 /* Forbidden if we are shut down or mounted norecovery. */ 485 error = -ESHUTDOWN; 486 if (xfs_is_shutdown(mp)) 487 goto out; 488 error = -ENOTRECOVERABLE; 489 if (xfs_has_norecovery(mp)) 490 goto out; 491 492 error = xchk_validate_inputs(mp, sm); 493 if (error) 494 goto out; 495 496 xfs_warn_mount(mp, XFS_OPSTATE_WARNED_SCRUB, 497 "EXPERIMENTAL online scrub feature in use. Use at your own risk!"); 498 499 sc = kzalloc(sizeof(struct xfs_scrub), XCHK_GFP_FLAGS); 500 if (!sc) { 501 error = -ENOMEM; 502 goto out; 503 } 504 505 sc->mp = mp; 506 sc->file = file; 507 sc->sm = sm; 508 sc->ops = &meta_scrub_ops[sm->sm_type]; 509 sc->sick_mask = xchk_health_mask_for_scrub_type(sm->sm_type); 510 retry_op: 511 /* 512 * When repairs are allowed, prevent freezing or readonly remount while 513 * scrub is running with a real transaction. 514 */ 515 if (sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) { 516 error = mnt_want_write_file(sc->file); 517 if (error) 518 goto out_sc; 519 520 sc->flags |= XCHK_HAVE_FREEZE_PROT; 521 } 522 523 /* Set up for the operation. */ 524 error = sc->ops->setup(sc); 525 if (error == -EDEADLOCK && !(sc->flags & XCHK_TRY_HARDER)) 526 goto try_harder; 527 if (error == -ECHRNG && !(sc->flags & XCHK_NEED_DRAIN)) 528 goto need_drain; 529 if (error) 530 goto out_teardown; 531 532 /* Scrub for errors. */ 533 check_start = xchk_stats_now(); 534 error = sc->ops->scrub(sc); 535 run.scrub_ns += xchk_stats_elapsed_ns(check_start); 536 if (error == -EDEADLOCK && !(sc->flags & XCHK_TRY_HARDER)) 537 goto try_harder; 538 if (error == -ECHRNG && !(sc->flags & XCHK_NEED_DRAIN)) 539 goto need_drain; 540 if (error || (sm->sm_flags & XFS_SCRUB_OFLAG_INCOMPLETE)) 541 goto out_teardown; 542 543 xchk_update_health(sc); 544 545 if ((sc->sm->sm_flags & XFS_SCRUB_IFLAG_REPAIR) && 546 !(sc->flags & XREP_ALREADY_FIXED)) { 547 bool needs_fix = xchk_needs_repair(sc->sm); 548 549 /* Userspace asked us to rebuild the structure regardless. */ 550 if (sc->sm->sm_flags & XFS_SCRUB_IFLAG_FORCE_REBUILD) 551 needs_fix = true; 552 553 /* Let debug users force us into the repair routines. */ 554 if (XFS_TEST_ERROR(needs_fix, mp, XFS_ERRTAG_FORCE_SCRUB_REPAIR)) 555 needs_fix = true; 556 557 /* 558 * If userspace asked for a repair but it wasn't necessary, 559 * report that back to userspace. 560 */ 561 if (!needs_fix) { 562 sc->sm->sm_flags |= XFS_SCRUB_OFLAG_NO_REPAIR_NEEDED; 563 goto out_nofix; 564 } 565 566 /* 567 * If it's broken, userspace wants us to fix it, and we haven't 568 * already tried to fix it, then attempt a repair. 569 */ 570 error = xrep_attempt(sc, &run); 571 if (error == -EAGAIN) { 572 /* 573 * Either the repair function succeeded or it couldn't 574 * get all the resources it needs; either way, we go 575 * back to the beginning and call the scrub function. 576 */ 577 error = xchk_teardown(sc, 0); 578 if (error) { 579 xrep_failure(mp); 580 goto out_sc; 581 } 582 goto retry_op; 583 } 584 } 585 586 out_nofix: 587 xchk_postmortem(sc); 588 out_teardown: 589 error = xchk_teardown(sc, error); 590 out_sc: 591 kfree(sc); 592 out: 593 trace_xchk_done(XFS_I(file_inode(file)), sm, error); 594 if (error == -EFSCORRUPTED || error == -EFSBADCRC) { 595 sm->sm_flags |= XFS_SCRUB_OFLAG_CORRUPT; 596 error = 0; 597 } 598 if (error != -ENOENT) 599 xchk_stats_merge(mp, sm, &run); 600 return error; 601 need_drain: 602 error = xchk_teardown(sc, 0); 603 if (error) 604 goto out_sc; 605 sc->flags |= XCHK_NEED_DRAIN; 606 run.retries++; 607 goto retry_op; 608 try_harder: 609 /* 610 * Scrubbers return -EDEADLOCK to mean 'try harder'. Tear down 611 * everything we hold, then set up again with preparation for 612 * worst-case scenarios. 613 */ 614 error = xchk_teardown(sc, 0); 615 if (error) 616 goto out_sc; 617 sc->flags |= XCHK_TRY_HARDER; 618 run.retries++; 619 goto retry_op; 620 } 621