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