1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2000-2005 Silicon Graphics, Inc. 4 * All Rights Reserved. 5 */ 6 #include "xfs.h" 7 #include "xfs_fs.h" 8 #include "xfs_shared.h" 9 #include "xfs_format.h" 10 #include "xfs_log_format.h" 11 #include "xfs_trans_resv.h" 12 #include "xfs_sb.h" 13 #include "xfs_mount.h" 14 #include "xfs_trans.h" 15 #include "xfs_error.h" 16 #include "xfs_alloc.h" 17 #include "xfs_fsops.h" 18 #include "xfs_trans_space.h" 19 #include "xfs_log.h" 20 #include "xfs_log_priv.h" 21 #include "xfs_ag.h" 22 #include "xfs_ag_resv.h" 23 #include "xfs_trace.h" 24 25 /* 26 * Write new AG headers to disk. Non-transactional, but need to be 27 * written and completed prior to the growfs transaction being logged. 28 * To do this, we use a delayed write buffer list and wait for 29 * submission and IO completion of the list as a whole. This allows the 30 * IO subsystem to merge all the AG headers in a single AG into a single 31 * IO and hide most of the latency of the IO from us. 32 * 33 * This also means that if we get an error whilst building the buffer 34 * list to write, we can cancel the entire list without having written 35 * anything. 36 */ 37 static int 38 xfs_resizefs_init_new_ags( 39 struct xfs_trans *tp, 40 struct aghdr_init_data *id, 41 xfs_agnumber_t oagcount, 42 xfs_agnumber_t nagcount, 43 xfs_rfsblock_t delta, 44 struct xfs_perag *last_pag, 45 bool *lastag_extended) 46 { 47 struct xfs_mount *mp = tp->t_mountp; 48 xfs_rfsblock_t nb = mp->m_sb.sb_dblocks + delta; 49 int error; 50 51 *lastag_extended = false; 52 53 INIT_LIST_HEAD(&id->buffer_list); 54 for (id->agno = nagcount - 1; 55 id->agno >= oagcount; 56 id->agno--, delta -= id->agsize) { 57 58 if (id->agno == nagcount - 1) 59 id->agsize = nb - (id->agno * 60 (xfs_rfsblock_t)mp->m_sb.sb_agblocks); 61 else 62 id->agsize = mp->m_sb.sb_agblocks; 63 64 error = xfs_ag_init_headers(mp, id); 65 if (error) { 66 xfs_buf_delwri_cancel(&id->buffer_list); 67 return error; 68 } 69 } 70 71 error = xfs_buf_delwri_submit(&id->buffer_list); 72 if (error) 73 return error; 74 75 if (delta) { 76 *lastag_extended = true; 77 error = xfs_ag_extend_space(last_pag, tp, delta); 78 } 79 return error; 80 } 81 82 /* 83 * growfs operations 84 */ 85 static int 86 xfs_growfs_data_private( 87 struct xfs_mount *mp, /* mount point for filesystem */ 88 struct xfs_growfs_data *in) /* growfs data input struct */ 89 { 90 struct xfs_buf *bp; 91 int error; 92 xfs_agnumber_t nagcount; 93 xfs_agnumber_t nagimax = 0; 94 xfs_rfsblock_t nb, nb_div, nb_mod; 95 int64_t delta; 96 bool lastag_extended = false; 97 xfs_agnumber_t oagcount; 98 struct xfs_trans *tp; 99 struct aghdr_init_data id = {}; 100 struct xfs_perag *last_pag; 101 102 nb = in->newblocks; 103 error = xfs_sb_validate_fsb_count(&mp->m_sb, nb); 104 if (error) 105 return error; 106 107 if (nb > mp->m_sb.sb_dblocks) { 108 error = xfs_buf_read_uncached(mp->m_ddev_targp, 109 XFS_FSB_TO_BB(mp, nb) - XFS_FSS_TO_BB(mp, 1), 110 XFS_FSS_TO_BB(mp, 1), 0, &bp, NULL); 111 if (error) 112 return error; 113 xfs_buf_relse(bp); 114 } 115 116 nb_div = nb; 117 nb_mod = do_div(nb_div, mp->m_sb.sb_agblocks); 118 if (nb_mod && nb_mod >= XFS_MIN_AG_BLOCKS) 119 nb_div++; 120 else if (nb_mod) 121 nb = nb_div * mp->m_sb.sb_agblocks; 122 123 if (nb_div > XFS_MAX_AGNUMBER + 1) { 124 nb_div = XFS_MAX_AGNUMBER + 1; 125 nb = nb_div * mp->m_sb.sb_agblocks; 126 } 127 nagcount = nb_div; 128 delta = nb - mp->m_sb.sb_dblocks; 129 /* 130 * Reject filesystems with a single AG because they are not 131 * supported, and reject a shrink operation that would cause a 132 * filesystem to become unsupported. 133 */ 134 if (delta < 0 && nagcount < 2) 135 return -EINVAL; 136 137 /* No work to do */ 138 if (delta == 0) 139 return 0; 140 141 oagcount = mp->m_sb.sb_agcount; 142 /* allocate the new per-ag structures */ 143 if (nagcount > oagcount) { 144 error = xfs_initialize_perag(mp, nagcount, nb, &nagimax); 145 if (error) 146 return error; 147 } else if (nagcount < oagcount) { 148 /* TODO: shrinking the entire AGs hasn't yet completed */ 149 return -EINVAL; 150 } 151 152 if (delta > 0) 153 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_growdata, 154 XFS_GROWFS_SPACE_RES(mp), 0, XFS_TRANS_RESERVE, 155 &tp); 156 else 157 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_growdata, -delta, 0, 158 0, &tp); 159 if (error) 160 return error; 161 162 last_pag = xfs_perag_get(mp, oagcount - 1); 163 if (delta > 0) { 164 error = xfs_resizefs_init_new_ags(tp, &id, oagcount, nagcount, 165 delta, last_pag, &lastag_extended); 166 } else { 167 xfs_warn_mount(mp, XFS_OPSTATE_WARNED_SHRINK, 168 "EXPERIMENTAL online shrink feature in use. Use at your own risk!"); 169 170 error = xfs_ag_shrink_space(last_pag, &tp, -delta); 171 } 172 xfs_perag_put(last_pag); 173 if (error) 174 goto out_trans_cancel; 175 176 /* 177 * Update changed superblock fields transactionally. These are not 178 * seen by the rest of the world until the transaction commit applies 179 * them atomically to the superblock. 180 */ 181 if (nagcount > oagcount) 182 xfs_trans_mod_sb(tp, XFS_TRANS_SB_AGCOUNT, nagcount - oagcount); 183 if (delta) 184 xfs_trans_mod_sb(tp, XFS_TRANS_SB_DBLOCKS, delta); 185 if (id.nfree) 186 xfs_trans_mod_sb(tp, XFS_TRANS_SB_FDBLOCKS, id.nfree); 187 188 /* 189 * Sync sb counters now to reflect the updated values. This is 190 * particularly important for shrink because the write verifier 191 * will fail if sb_fdblocks is ever larger than sb_dblocks. 192 */ 193 if (xfs_has_lazysbcount(mp)) 194 xfs_log_sb(tp); 195 196 xfs_trans_set_sync(tp); 197 error = xfs_trans_commit(tp); 198 if (error) 199 return error; 200 201 /* New allocation groups fully initialized, so update mount struct */ 202 if (nagimax) 203 mp->m_maxagi = nagimax; 204 xfs_set_low_space_thresholds(mp); 205 mp->m_alloc_set_aside = xfs_alloc_set_aside(mp); 206 207 if (delta > 0) { 208 /* 209 * If we expanded the last AG, free the per-AG reservation 210 * so we can reinitialize it with the new size. 211 */ 212 if (lastag_extended) { 213 struct xfs_perag *pag; 214 215 pag = xfs_perag_get(mp, id.agno); 216 error = xfs_ag_resv_free(pag); 217 xfs_perag_put(pag); 218 if (error) 219 return error; 220 } 221 /* 222 * Reserve AG metadata blocks. ENOSPC here does not mean there 223 * was a growfs failure, just that there still isn't space for 224 * new user data after the grow has been run. 225 */ 226 error = xfs_fs_reserve_ag_blocks(mp); 227 if (error == -ENOSPC) 228 error = 0; 229 } 230 return error; 231 232 out_trans_cancel: 233 xfs_trans_cancel(tp); 234 return error; 235 } 236 237 static int 238 xfs_growfs_log_private( 239 struct xfs_mount *mp, /* mount point for filesystem */ 240 struct xfs_growfs_log *in) /* growfs log input struct */ 241 { 242 xfs_extlen_t nb; 243 244 nb = in->newblocks; 245 if (nb < XFS_MIN_LOG_BLOCKS || nb < XFS_B_TO_FSB(mp, XFS_MIN_LOG_BYTES)) 246 return -EINVAL; 247 if (nb == mp->m_sb.sb_logblocks && 248 in->isint == (mp->m_sb.sb_logstart != 0)) 249 return -EINVAL; 250 /* 251 * Moving the log is hard, need new interfaces to sync 252 * the log first, hold off all activity while moving it. 253 * Can have shorter or longer log in the same space, 254 * or transform internal to external log or vice versa. 255 */ 256 return -ENOSYS; 257 } 258 259 static int 260 xfs_growfs_imaxpct( 261 struct xfs_mount *mp, 262 __u32 imaxpct) 263 { 264 struct xfs_trans *tp; 265 int dpct; 266 int error; 267 268 if (imaxpct > 100) 269 return -EINVAL; 270 271 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_growdata, 272 XFS_GROWFS_SPACE_RES(mp), 0, XFS_TRANS_RESERVE, &tp); 273 if (error) 274 return error; 275 276 dpct = imaxpct - mp->m_sb.sb_imax_pct; 277 xfs_trans_mod_sb(tp, XFS_TRANS_SB_IMAXPCT, dpct); 278 xfs_trans_set_sync(tp); 279 return xfs_trans_commit(tp); 280 } 281 282 /* 283 * protected versions of growfs function acquire and release locks on the mount 284 * point - exported through ioctls: XFS_IOC_FSGROWFSDATA, XFS_IOC_FSGROWFSLOG, 285 * XFS_IOC_FSGROWFSRT 286 */ 287 int 288 xfs_growfs_data( 289 struct xfs_mount *mp, 290 struct xfs_growfs_data *in) 291 { 292 int error = 0; 293 294 if (!capable(CAP_SYS_ADMIN)) 295 return -EPERM; 296 if (!mutex_trylock(&mp->m_growlock)) 297 return -EWOULDBLOCK; 298 299 /* update imaxpct separately to the physical grow of the filesystem */ 300 if (in->imaxpct != mp->m_sb.sb_imax_pct) { 301 error = xfs_growfs_imaxpct(mp, in->imaxpct); 302 if (error) 303 goto out_error; 304 } 305 306 if (in->newblocks != mp->m_sb.sb_dblocks) { 307 error = xfs_growfs_data_private(mp, in); 308 if (error) 309 goto out_error; 310 } 311 312 /* Post growfs calculations needed to reflect new state in operations */ 313 if (mp->m_sb.sb_imax_pct) { 314 uint64_t icount = mp->m_sb.sb_dblocks * mp->m_sb.sb_imax_pct; 315 do_div(icount, 100); 316 M_IGEO(mp)->maxicount = XFS_FSB_TO_INO(mp, icount); 317 } else 318 M_IGEO(mp)->maxicount = 0; 319 320 /* Update secondary superblocks now the physical grow has completed */ 321 error = xfs_update_secondary_sbs(mp); 322 323 out_error: 324 /* 325 * Increment the generation unconditionally, the error could be from 326 * updating the secondary superblocks, in which case the new size 327 * is live already. 328 */ 329 mp->m_generation++; 330 mutex_unlock(&mp->m_growlock); 331 return error; 332 } 333 334 int 335 xfs_growfs_log( 336 xfs_mount_t *mp, 337 struct xfs_growfs_log *in) 338 { 339 int error; 340 341 if (!capable(CAP_SYS_ADMIN)) 342 return -EPERM; 343 if (!mutex_trylock(&mp->m_growlock)) 344 return -EWOULDBLOCK; 345 error = xfs_growfs_log_private(mp, in); 346 mutex_unlock(&mp->m_growlock); 347 return error; 348 } 349 350 /* 351 * exported through ioctl XFS_IOC_FSCOUNTS 352 */ 353 354 void 355 xfs_fs_counts( 356 xfs_mount_t *mp, 357 xfs_fsop_counts_t *cnt) 358 { 359 cnt->allocino = percpu_counter_read_positive(&mp->m_icount); 360 cnt->freeino = percpu_counter_read_positive(&mp->m_ifree); 361 cnt->freedata = percpu_counter_read_positive(&mp->m_fdblocks) - 362 xfs_fdblocks_unavailable(mp); 363 cnt->freertx = percpu_counter_read_positive(&mp->m_frextents); 364 } 365 366 /* 367 * exported through ioctl XFS_IOC_SET_RESBLKS & XFS_IOC_GET_RESBLKS 368 * 369 * xfs_reserve_blocks is called to set m_resblks 370 * in the in-core mount table. The number of unused reserved blocks 371 * is kept in m_resblks_avail. 372 * 373 * Reserve the requested number of blocks if available. Otherwise return 374 * as many as possible to satisfy the request. The actual number 375 * reserved are returned in outval 376 * 377 * A null inval pointer indicates that only the current reserved blocks 378 * available should be returned no settings are changed. 379 */ 380 381 int 382 xfs_reserve_blocks( 383 xfs_mount_t *mp, 384 uint64_t *inval, 385 xfs_fsop_resblks_t *outval) 386 { 387 int64_t lcounter, delta; 388 int64_t fdblks_delta = 0; 389 uint64_t request; 390 int64_t free; 391 int error = 0; 392 393 /* If inval is null, report current values and return */ 394 if (inval == (uint64_t *)NULL) { 395 if (!outval) 396 return -EINVAL; 397 outval->resblks = mp->m_resblks; 398 outval->resblks_avail = mp->m_resblks_avail; 399 return 0; 400 } 401 402 request = *inval; 403 404 /* 405 * With per-cpu counters, this becomes an interesting problem. we need 406 * to work out if we are freeing or allocation blocks first, then we can 407 * do the modification as necessary. 408 * 409 * We do this under the m_sb_lock so that if we are near ENOSPC, we will 410 * hold out any changes while we work out what to do. This means that 411 * the amount of free space can change while we do this, so we need to 412 * retry if we end up trying to reserve more space than is available. 413 */ 414 spin_lock(&mp->m_sb_lock); 415 416 /* 417 * If our previous reservation was larger than the current value, 418 * then move any unused blocks back to the free pool. Modify the resblks 419 * counters directly since we shouldn't have any problems unreserving 420 * space. 421 */ 422 if (mp->m_resblks > request) { 423 lcounter = mp->m_resblks_avail - request; 424 if (lcounter > 0) { /* release unused blocks */ 425 fdblks_delta = lcounter; 426 mp->m_resblks_avail -= lcounter; 427 } 428 mp->m_resblks = request; 429 if (fdblks_delta) { 430 spin_unlock(&mp->m_sb_lock); 431 error = xfs_mod_fdblocks(mp, fdblks_delta, 0); 432 spin_lock(&mp->m_sb_lock); 433 } 434 435 goto out; 436 } 437 438 /* 439 * If the request is larger than the current reservation, reserve the 440 * blocks before we update the reserve counters. Sample m_fdblocks and 441 * perform a partial reservation if the request exceeds free space. 442 * 443 * The code below estimates how many blocks it can request from 444 * fdblocks to stash in the reserve pool. This is a classic TOCTOU 445 * race since fdblocks updates are not always coordinated via 446 * m_sb_lock. Set the reserve size even if there's not enough free 447 * space to fill it because mod_fdblocks will refill an undersized 448 * reserve when it can. 449 */ 450 free = percpu_counter_sum(&mp->m_fdblocks) - 451 xfs_fdblocks_unavailable(mp); 452 delta = request - mp->m_resblks; 453 mp->m_resblks = request; 454 if (delta > 0 && free > 0) { 455 /* 456 * We'll either succeed in getting space from the free block 457 * count or we'll get an ENOSPC. Don't set the reserved flag 458 * here - we don't want to reserve the extra reserve blocks 459 * from the reserve. 460 * 461 * The desired reserve size can change after we drop the lock. 462 * Use mod_fdblocks to put the space into the reserve or into 463 * fdblocks as appropriate. 464 */ 465 fdblks_delta = min(free, delta); 466 spin_unlock(&mp->m_sb_lock); 467 error = xfs_mod_fdblocks(mp, -fdblks_delta, 0); 468 if (!error) 469 xfs_mod_fdblocks(mp, fdblks_delta, 0); 470 spin_lock(&mp->m_sb_lock); 471 } 472 out: 473 if (outval) { 474 outval->resblks = mp->m_resblks; 475 outval->resblks_avail = mp->m_resblks_avail; 476 } 477 478 spin_unlock(&mp->m_sb_lock); 479 return error; 480 } 481 482 int 483 xfs_fs_goingdown( 484 xfs_mount_t *mp, 485 uint32_t inflags) 486 { 487 switch (inflags) { 488 case XFS_FSOP_GOING_FLAGS_DEFAULT: { 489 if (!freeze_bdev(mp->m_super->s_bdev)) { 490 xfs_force_shutdown(mp, SHUTDOWN_FORCE_UMOUNT); 491 thaw_bdev(mp->m_super->s_bdev); 492 } 493 break; 494 } 495 case XFS_FSOP_GOING_FLAGS_LOGFLUSH: 496 xfs_force_shutdown(mp, SHUTDOWN_FORCE_UMOUNT); 497 break; 498 case XFS_FSOP_GOING_FLAGS_NOLOGFLUSH: 499 xfs_force_shutdown(mp, 500 SHUTDOWN_FORCE_UMOUNT | SHUTDOWN_LOG_IO_ERROR); 501 break; 502 default: 503 return -EINVAL; 504 } 505 506 return 0; 507 } 508 509 /* 510 * Force a shutdown of the filesystem instantly while keeping the filesystem 511 * consistent. We don't do an unmount here; just shutdown the shop, make sure 512 * that absolutely nothing persistent happens to this filesystem after this 513 * point. 514 * 515 * The shutdown state change is atomic, resulting in the first and only the 516 * first shutdown call processing the shutdown. This means we only shutdown the 517 * log once as it requires, and we don't spam the logs when multiple concurrent 518 * shutdowns race to set the shutdown flags. 519 */ 520 void 521 xfs_do_force_shutdown( 522 struct xfs_mount *mp, 523 uint32_t flags, 524 char *fname, 525 int lnnum) 526 { 527 int tag; 528 const char *why; 529 530 531 if (test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &mp->m_opstate)) { 532 xlog_shutdown_wait(mp->m_log); 533 return; 534 } 535 if (mp->m_sb_bp) 536 mp->m_sb_bp->b_flags |= XBF_DONE; 537 538 if (flags & SHUTDOWN_FORCE_UMOUNT) 539 xfs_alert(mp, "User initiated shutdown received."); 540 541 if (xlog_force_shutdown(mp->m_log, flags)) { 542 tag = XFS_PTAG_SHUTDOWN_LOGERROR; 543 why = "Log I/O Error"; 544 } else if (flags & SHUTDOWN_CORRUPT_INCORE) { 545 tag = XFS_PTAG_SHUTDOWN_CORRUPT; 546 why = "Corruption of in-memory data"; 547 } else if (flags & SHUTDOWN_CORRUPT_ONDISK) { 548 tag = XFS_PTAG_SHUTDOWN_CORRUPT; 549 why = "Corruption of on-disk metadata"; 550 } else if (flags & SHUTDOWN_DEVICE_REMOVED) { 551 tag = XFS_PTAG_SHUTDOWN_IOERROR; 552 why = "Block device removal"; 553 } else { 554 tag = XFS_PTAG_SHUTDOWN_IOERROR; 555 why = "Metadata I/O Error"; 556 } 557 558 trace_xfs_force_shutdown(mp, tag, flags, fname, lnnum); 559 560 xfs_alert_tag(mp, tag, 561 "%s (0x%x) detected at %pS (%s:%d). Shutting down filesystem.", 562 why, flags, __return_address, fname, lnnum); 563 xfs_alert(mp, 564 "Please unmount the filesystem and rectify the problem(s)"); 565 if (xfs_error_level >= XFS_ERRLEVEL_HIGH) 566 xfs_stack_trace(); 567 } 568 569 /* 570 * Reserve free space for per-AG metadata. 571 */ 572 int 573 xfs_fs_reserve_ag_blocks( 574 struct xfs_mount *mp) 575 { 576 xfs_agnumber_t agno; 577 struct xfs_perag *pag; 578 int error = 0; 579 int err2; 580 581 mp->m_finobt_nores = false; 582 for_each_perag(mp, agno, pag) { 583 err2 = xfs_ag_resv_init(pag, NULL); 584 if (err2 && !error) 585 error = err2; 586 } 587 588 if (error && error != -ENOSPC) { 589 xfs_warn(mp, 590 "Error %d reserving per-AG metadata reserve pool.", error); 591 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE); 592 } 593 594 return error; 595 } 596 597 /* 598 * Free space reserved for per-AG metadata. 599 */ 600 int 601 xfs_fs_unreserve_ag_blocks( 602 struct xfs_mount *mp) 603 { 604 xfs_agnumber_t agno; 605 struct xfs_perag *pag; 606 int error = 0; 607 int err2; 608 609 for_each_perag(mp, agno, pag) { 610 err2 = xfs_ag_resv_free(pag); 611 if (err2 && !error) 612 error = err2; 613 } 614 615 if (error) 616 xfs_warn(mp, 617 "Error %d freeing per-AG metadata reserve pool.", error); 618 619 return error; 620 } 621