1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc. 4 * Copyright (C) 2010 Red Hat, Inc. 5 * All Rights Reserved. 6 */ 7 #include "xfs.h" 8 #include "xfs_fs.h" 9 #include "xfs_shared.h" 10 #include "xfs_format.h" 11 #include "xfs_log_format.h" 12 #include "xfs_trans_resv.h" 13 #include "xfs_mount.h" 14 #include "xfs_extent_busy.h" 15 #include "xfs_quota.h" 16 #include "xfs_trans.h" 17 #include "xfs_trans_priv.h" 18 #include "xfs_log.h" 19 #include "xfs_log_priv.h" 20 #include "xfs_trace.h" 21 #include "xfs_error.h" 22 #include "xfs_defer.h" 23 #include "xfs_inode.h" 24 #include "xfs_dquot_item.h" 25 #include "xfs_dquot.h" 26 #include "xfs_icache.h" 27 #include "xfs_rtbitmap.h" 28 #include "xfs_rtgroup.h" 29 #include "xfs_sb.h" 30 31 struct kmem_cache *xfs_trans_cache; 32 33 #if defined(CONFIG_TRACEPOINTS) 34 static void 35 xfs_trans_trace_reservations( 36 struct xfs_mount *mp) 37 { 38 struct xfs_trans_res *res; 39 struct xfs_trans_res *end_res; 40 int i; 41 42 res = (struct xfs_trans_res *)M_RES(mp); 43 end_res = (struct xfs_trans_res *)(M_RES(mp) + 1); 44 for (i = 0; res < end_res; i++, res++) 45 trace_xfs_trans_resv_calc(mp, i, res); 46 } 47 #else 48 # define xfs_trans_trace_reservations(mp) 49 #endif 50 51 /* 52 * Initialize the precomputed transaction reservation values 53 * in the mount structure. 54 */ 55 void 56 xfs_trans_init( 57 struct xfs_mount *mp) 58 { 59 xfs_trans_resv_calc(mp, M_RES(mp)); 60 xfs_trans_trace_reservations(mp); 61 } 62 63 /* 64 * Free the transaction structure. If there is more clean up 65 * to do when the structure is freed, add it here. 66 */ 67 STATIC void 68 xfs_trans_free( 69 struct xfs_trans *tp) 70 { 71 xfs_extent_busy_sort(&tp->t_busy); 72 xfs_extent_busy_clear(&tp->t_busy, false); 73 74 trace_xfs_trans_free(tp, _RET_IP_); 75 xfs_trans_clear_context(tp); 76 if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT)) 77 sb_end_intwrite(tp->t_mountp->m_super); 78 xfs_trans_free_dqinfo(tp); 79 kmem_cache_free(xfs_trans_cache, tp); 80 } 81 82 /* 83 * This is called to create a new transaction which will share the 84 * permanent log reservation of the given transaction. The remaining 85 * unused block and rt extent reservations are also inherited. This 86 * implies that the original transaction is no longer allowed to allocate 87 * blocks. Locks and log items, however, are no inherited. They must 88 * be added to the new transaction explicitly. 89 */ 90 STATIC struct xfs_trans * 91 xfs_trans_dup( 92 struct xfs_trans *tp) 93 { 94 struct xfs_trans *ntp; 95 96 trace_xfs_trans_dup(tp, _RET_IP_); 97 98 ntp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL); 99 100 /* 101 * Initialize the new transaction structure. 102 */ 103 ntp->t_mountp = tp->t_mountp; 104 INIT_LIST_HEAD(&ntp->t_items); 105 INIT_LIST_HEAD(&ntp->t_busy); 106 INIT_LIST_HEAD(&ntp->t_dfops); 107 ntp->t_highest_agno = NULLAGNUMBER; 108 109 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES); 110 ASSERT(tp->t_ticket != NULL); 111 112 ntp->t_flags = XFS_TRANS_PERM_LOG_RES | 113 (tp->t_flags & XFS_TRANS_RESERVE) | 114 (tp->t_flags & XFS_TRANS_NO_WRITECOUNT) | 115 (tp->t_flags & XFS_TRANS_RES_FDBLKS); 116 /* We gave our writer reference to the new transaction */ 117 tp->t_flags |= XFS_TRANS_NO_WRITECOUNT; 118 ntp->t_ticket = xfs_log_ticket_get(tp->t_ticket); 119 120 ASSERT(tp->t_blk_res >= tp->t_blk_res_used); 121 ntp->t_blk_res = tp->t_blk_res - tp->t_blk_res_used; 122 tp->t_blk_res = tp->t_blk_res_used; 123 124 ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used; 125 tp->t_rtx_res = tp->t_rtx_res_used; 126 127 xfs_trans_switch_context(tp, ntp); 128 129 /* move deferred ops over to the new tp */ 130 xfs_defer_move(ntp, tp); 131 132 xfs_trans_dup_dqinfo(tp, ntp); 133 return ntp; 134 } 135 136 /* 137 * This is called to reserve free disk blocks and log space for the 138 * given transaction. This must be done before allocating any resources 139 * within the transaction. 140 * 141 * This will return ENOSPC if there are not enough blocks available. 142 * It will sleep waiting for available log space. 143 * The only valid value for the flags parameter is XFS_RES_LOG_PERM, which 144 * is used by long running transactions. If any one of the reservations 145 * fails then they will all be backed out. 146 * 147 * This does not do quota reservations. That typically is done by the 148 * caller afterwards. 149 */ 150 static int 151 xfs_trans_reserve( 152 struct xfs_trans *tp, 153 struct xfs_trans_res *resp, 154 uint blocks, 155 uint rtextents) 156 { 157 struct xfs_mount *mp = tp->t_mountp; 158 int error = 0; 159 bool rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0; 160 161 /* 162 * Attempt to reserve the needed disk blocks by decrementing 163 * the number needed from the number available. This will 164 * fail if the count would go below zero. 165 */ 166 if (blocks > 0) { 167 error = xfs_dec_fdblocks(mp, blocks, rsvd); 168 if (error != 0) 169 return -ENOSPC; 170 tp->t_blk_res += blocks; 171 } 172 173 /* 174 * Reserve the log space needed for this transaction. 175 */ 176 if (resp->tr_logres > 0) { 177 bool permanent = false; 178 179 ASSERT(tp->t_log_res == 0 || 180 tp->t_log_res == resp->tr_logres); 181 ASSERT(tp->t_log_count == 0 || 182 tp->t_log_count == resp->tr_logcount); 183 184 if (resp->tr_logflags & XFS_TRANS_PERM_LOG_RES) { 185 tp->t_flags |= XFS_TRANS_PERM_LOG_RES; 186 permanent = true; 187 } else { 188 ASSERT(tp->t_ticket == NULL); 189 ASSERT(!(tp->t_flags & XFS_TRANS_PERM_LOG_RES)); 190 } 191 192 if (tp->t_ticket != NULL) { 193 ASSERT(resp->tr_logflags & XFS_TRANS_PERM_LOG_RES); 194 error = xfs_log_regrant(mp, tp->t_ticket); 195 } else { 196 error = xfs_log_reserve(mp, resp->tr_logres, 197 resp->tr_logcount, 198 &tp->t_ticket, permanent); 199 } 200 201 if (error) 202 goto undo_blocks; 203 204 tp->t_log_res = resp->tr_logres; 205 tp->t_log_count = resp->tr_logcount; 206 } 207 208 /* 209 * Attempt to reserve the needed realtime extents by decrementing 210 * the number needed from the number available. This will 211 * fail if the count would go below zero. 212 */ 213 if (rtextents > 0) { 214 error = xfs_dec_frextents(mp, rtextents); 215 if (error) { 216 error = -ENOSPC; 217 goto undo_log; 218 } 219 tp->t_rtx_res += rtextents; 220 } 221 222 return 0; 223 224 /* 225 * Error cases jump to one of these labels to undo any 226 * reservations which have already been performed. 227 */ 228 undo_log: 229 if (resp->tr_logres > 0) { 230 xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket); 231 tp->t_ticket = NULL; 232 tp->t_log_res = 0; 233 tp->t_flags &= ~XFS_TRANS_PERM_LOG_RES; 234 } 235 236 undo_blocks: 237 if (blocks > 0) { 238 xfs_add_fdblocks(mp, blocks); 239 tp->t_blk_res = 0; 240 } 241 return error; 242 } 243 244 int 245 xfs_trans_alloc( 246 struct xfs_mount *mp, 247 struct xfs_trans_res *resp, 248 uint blocks, 249 uint rtextents, 250 uint flags, 251 struct xfs_trans **tpp) 252 { 253 struct xfs_trans *tp; 254 bool want_retry = true; 255 int error; 256 257 /* 258 * Allocate the handle before we do our freeze accounting and setting up 259 * GFP_NOFS allocation context so that we avoid lockdep false positives 260 * by doing GFP_KERNEL allocations inside sb_start_intwrite(). 261 */ 262 retry: 263 tp = kmem_cache_zalloc(xfs_trans_cache, GFP_KERNEL | __GFP_NOFAIL); 264 if (!(flags & XFS_TRANS_NO_WRITECOUNT)) 265 sb_start_intwrite(mp->m_super); 266 xfs_trans_set_context(tp); 267 268 /* 269 * Zero-reservation ("empty") transactions can't modify anything, so 270 * they're allowed to run while we're frozen. 271 */ 272 WARN_ON(resp->tr_logres > 0 && 273 mp->m_super->s_writers.frozen == SB_FREEZE_COMPLETE); 274 ASSERT(!(flags & XFS_TRANS_RES_FDBLKS) || 275 xfs_has_lazysbcount(mp)); 276 277 tp->t_flags = flags; 278 tp->t_mountp = mp; 279 INIT_LIST_HEAD(&tp->t_items); 280 INIT_LIST_HEAD(&tp->t_busy); 281 INIT_LIST_HEAD(&tp->t_dfops); 282 tp->t_highest_agno = NULLAGNUMBER; 283 284 error = xfs_trans_reserve(tp, resp, blocks, rtextents); 285 if (error == -ENOSPC && want_retry) { 286 xfs_trans_cancel(tp); 287 288 /* 289 * We weren't able to reserve enough space for the transaction. 290 * Flush the other speculative space allocations to free space. 291 * Do not perform a synchronous scan because callers can hold 292 * other locks. 293 */ 294 error = xfs_blockgc_flush_all(mp); 295 if (error) 296 return error; 297 want_retry = false; 298 goto retry; 299 } 300 if (error) { 301 xfs_trans_cancel(tp); 302 return error; 303 } 304 305 trace_xfs_trans_alloc(tp, _RET_IP_); 306 307 *tpp = tp; 308 return 0; 309 } 310 311 /* 312 * Create an empty transaction with no reservation. This is a defensive 313 * mechanism for routines that query metadata without actually modifying them -- 314 * if the metadata being queried is somehow cross-linked (think a btree block 315 * pointer that points higher in the tree), we risk deadlock. However, blocks 316 * grabbed as part of a transaction can be re-grabbed. The verifiers will 317 * notice the corrupt block and the operation will fail back to userspace 318 * without deadlocking. 319 * 320 * Note the zero-length reservation; this transaction MUST be cancelled without 321 * any dirty data. 322 * 323 * Callers should obtain freeze protection to avoid a conflict with fs freezing 324 * where we can be grabbing buffers at the same time that freeze is trying to 325 * drain the buffer LRU list. 326 */ 327 int 328 xfs_trans_alloc_empty( 329 struct xfs_mount *mp, 330 struct xfs_trans **tpp) 331 { 332 struct xfs_trans_res resv = {0}; 333 334 return xfs_trans_alloc(mp, &resv, 0, 0, XFS_TRANS_NO_WRITECOUNT, tpp); 335 } 336 337 /* 338 * Record the indicated change to the given field for application 339 * to the file system's superblock when the transaction commits. 340 * For now, just store the change in the transaction structure. 341 * 342 * Mark the transaction structure to indicate that the superblock 343 * needs to be updated before committing. 344 * 345 * Because we may not be keeping track of allocated/free inodes and 346 * used filesystem blocks in the superblock, we do not mark the 347 * superblock dirty in this transaction if we modify these fields. 348 * We still need to update the transaction deltas so that they get 349 * applied to the incore superblock, but we don't want them to 350 * cause the superblock to get locked and logged if these are the 351 * only fields in the superblock that the transaction modifies. 352 */ 353 void 354 xfs_trans_mod_sb( 355 xfs_trans_t *tp, 356 uint field, 357 int64_t delta) 358 { 359 uint32_t flags = (XFS_TRANS_DIRTY|XFS_TRANS_SB_DIRTY); 360 xfs_mount_t *mp = tp->t_mountp; 361 362 switch (field) { 363 case XFS_TRANS_SB_ICOUNT: 364 tp->t_icount_delta += delta; 365 if (xfs_has_lazysbcount(mp)) 366 flags &= ~XFS_TRANS_SB_DIRTY; 367 break; 368 case XFS_TRANS_SB_IFREE: 369 tp->t_ifree_delta += delta; 370 if (xfs_has_lazysbcount(mp)) 371 flags &= ~XFS_TRANS_SB_DIRTY; 372 break; 373 case XFS_TRANS_SB_FDBLOCKS: 374 /* 375 * Track the number of blocks allocated in the transaction. 376 * Make sure it does not exceed the number reserved. If so, 377 * shutdown as this can lead to accounting inconsistency. 378 */ 379 if (delta < 0) { 380 tp->t_blk_res_used += (uint)-delta; 381 if (tp->t_blk_res_used > tp->t_blk_res) 382 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE); 383 } else if (delta > 0 && (tp->t_flags & XFS_TRANS_RES_FDBLKS)) { 384 int64_t blkres_delta; 385 386 /* 387 * Return freed blocks directly to the reservation 388 * instead of the global pool, being careful not to 389 * overflow the trans counter. This is used to preserve 390 * reservation across chains of transaction rolls that 391 * repeatedly free and allocate blocks. 392 */ 393 blkres_delta = min_t(int64_t, delta, 394 UINT_MAX - tp->t_blk_res); 395 tp->t_blk_res += blkres_delta; 396 delta -= blkres_delta; 397 } 398 tp->t_fdblocks_delta += delta; 399 if (xfs_has_lazysbcount(mp)) 400 flags &= ~XFS_TRANS_SB_DIRTY; 401 break; 402 case XFS_TRANS_SB_RES_FDBLOCKS: 403 /* 404 * The allocation has already been applied to the 405 * in-core superblock's counter. This should only 406 * be applied to the on-disk superblock. 407 */ 408 tp->t_res_fdblocks_delta += delta; 409 if (xfs_has_lazysbcount(mp)) 410 flags &= ~XFS_TRANS_SB_DIRTY; 411 break; 412 case XFS_TRANS_SB_FREXTENTS: 413 /* 414 * Track the number of blocks allocated in the 415 * transaction. Make sure it does not exceed the 416 * number reserved. 417 */ 418 if (delta < 0) { 419 tp->t_rtx_res_used += (uint)-delta; 420 ASSERT(tp->t_rtx_res_used <= tp->t_rtx_res); 421 } 422 tp->t_frextents_delta += delta; 423 if (xfs_has_rtgroups(mp)) 424 flags &= ~XFS_TRANS_SB_DIRTY; 425 break; 426 case XFS_TRANS_SB_RES_FREXTENTS: 427 /* 428 * The allocation has already been applied to the 429 * in-core superblock's counter. This should only 430 * be applied to the on-disk superblock. 431 */ 432 ASSERT(delta < 0); 433 tp->t_res_frextents_delta += delta; 434 if (xfs_has_rtgroups(mp)) 435 flags &= ~XFS_TRANS_SB_DIRTY; 436 break; 437 case XFS_TRANS_SB_DBLOCKS: 438 tp->t_dblocks_delta += delta; 439 break; 440 case XFS_TRANS_SB_AGCOUNT: 441 ASSERT(delta > 0); 442 tp->t_agcount_delta += delta; 443 break; 444 case XFS_TRANS_SB_IMAXPCT: 445 tp->t_imaxpct_delta += delta; 446 break; 447 case XFS_TRANS_SB_REXTSIZE: 448 tp->t_rextsize_delta += delta; 449 break; 450 case XFS_TRANS_SB_RBMBLOCKS: 451 tp->t_rbmblocks_delta += delta; 452 break; 453 case XFS_TRANS_SB_RBLOCKS: 454 tp->t_rblocks_delta += delta; 455 break; 456 case XFS_TRANS_SB_REXTENTS: 457 tp->t_rextents_delta += delta; 458 break; 459 case XFS_TRANS_SB_REXTSLOG: 460 tp->t_rextslog_delta += delta; 461 break; 462 case XFS_TRANS_SB_RGCOUNT: 463 ASSERT(delta > 0); 464 tp->t_rgcount_delta += delta; 465 break; 466 default: 467 ASSERT(0); 468 return; 469 } 470 471 tp->t_flags |= flags; 472 } 473 474 /* 475 * xfs_trans_apply_sb_deltas() is called from the commit code 476 * to bring the superblock buffer into the current transaction 477 * and modify it as requested by earlier calls to xfs_trans_mod_sb(). 478 * 479 * For now we just look at each field allowed to change and change 480 * it if necessary. 481 */ 482 STATIC void 483 xfs_trans_apply_sb_deltas( 484 xfs_trans_t *tp) 485 { 486 struct xfs_dsb *sbp; 487 struct xfs_buf *bp; 488 int whole = 0; 489 490 bp = xfs_trans_getsb(tp); 491 sbp = bp->b_addr; 492 493 /* 494 * Only update the superblock counters if we are logging them 495 */ 496 if (!xfs_has_lazysbcount((tp->t_mountp))) { 497 if (tp->t_icount_delta) 498 be64_add_cpu(&sbp->sb_icount, tp->t_icount_delta); 499 if (tp->t_ifree_delta) 500 be64_add_cpu(&sbp->sb_ifree, tp->t_ifree_delta); 501 if (tp->t_fdblocks_delta) 502 be64_add_cpu(&sbp->sb_fdblocks, tp->t_fdblocks_delta); 503 if (tp->t_res_fdblocks_delta) 504 be64_add_cpu(&sbp->sb_fdblocks, tp->t_res_fdblocks_delta); 505 } 506 507 /* 508 * sb_frextents was added to the lazy sb counters when the rt groups 509 * feature was introduced. This is possible because we know that all 510 * kernels supporting rtgroups will also recompute frextents from the 511 * realtime bitmap. 512 * 513 * For older file systems, updating frextents requires careful handling 514 * because we cannot rely on log recovery in older kernels to recompute 515 * the value from the rtbitmap. This means that the ondisk frextents 516 * must be consistent with the rtbitmap. 517 * 518 * Therefore, log the frextents change to the ondisk superblock and 519 * update the incore superblock so that future calls to xfs_log_sb 520 * write the correct value ondisk. 521 */ 522 if ((tp->t_frextents_delta || tp->t_res_frextents_delta) && 523 !xfs_has_rtgroups(tp->t_mountp)) { 524 struct xfs_mount *mp = tp->t_mountp; 525 int64_t rtxdelta; 526 527 rtxdelta = tp->t_frextents_delta + tp->t_res_frextents_delta; 528 529 spin_lock(&mp->m_sb_lock); 530 be64_add_cpu(&sbp->sb_frextents, rtxdelta); 531 mp->m_sb.sb_frextents += rtxdelta; 532 spin_unlock(&mp->m_sb_lock); 533 } 534 535 if (tp->t_dblocks_delta) { 536 be64_add_cpu(&sbp->sb_dblocks, tp->t_dblocks_delta); 537 whole = 1; 538 } 539 if (tp->t_agcount_delta) { 540 be32_add_cpu(&sbp->sb_agcount, tp->t_agcount_delta); 541 whole = 1; 542 } 543 if (tp->t_imaxpct_delta) { 544 sbp->sb_imax_pct += tp->t_imaxpct_delta; 545 whole = 1; 546 } 547 if (tp->t_rextsize_delta) { 548 be32_add_cpu(&sbp->sb_rextsize, tp->t_rextsize_delta); 549 550 /* 551 * Because the ondisk sb records rtgroup size in units of rt 552 * extents, any time we update the rt extent size we have to 553 * recompute the ondisk rtgroup block log. The incore values 554 * will be recomputed in xfs_trans_unreserve_and_mod_sb. 555 */ 556 if (xfs_has_rtgroups(tp->t_mountp)) { 557 sbp->sb_rgblklog = xfs_compute_rgblklog( 558 be32_to_cpu(sbp->sb_rgextents), 559 be32_to_cpu(sbp->sb_rextsize)); 560 } 561 whole = 1; 562 } 563 if (tp->t_rbmblocks_delta) { 564 be32_add_cpu(&sbp->sb_rbmblocks, tp->t_rbmblocks_delta); 565 whole = 1; 566 } 567 if (tp->t_rblocks_delta) { 568 be64_add_cpu(&sbp->sb_rblocks, tp->t_rblocks_delta); 569 whole = 1; 570 } 571 if (tp->t_rextents_delta) { 572 be64_add_cpu(&sbp->sb_rextents, tp->t_rextents_delta); 573 whole = 1; 574 } 575 if (tp->t_rextslog_delta) { 576 sbp->sb_rextslog += tp->t_rextslog_delta; 577 whole = 1; 578 } 579 if (tp->t_rgcount_delta) { 580 be32_add_cpu(&sbp->sb_rgcount, tp->t_rgcount_delta); 581 whole = 1; 582 } 583 584 xfs_trans_buf_set_type(tp, bp, XFS_BLFT_SB_BUF); 585 if (whole) 586 /* 587 * Log the whole thing, the fields are noncontiguous. 588 */ 589 xfs_trans_log_buf(tp, bp, 0, sizeof(struct xfs_dsb) - 1); 590 else 591 /* 592 * Since all the modifiable fields are contiguous, we 593 * can get away with this. 594 */ 595 xfs_trans_log_buf(tp, bp, offsetof(struct xfs_dsb, sb_icount), 596 offsetof(struct xfs_dsb, sb_frextents) + 597 sizeof(sbp->sb_frextents) - 1); 598 } 599 600 /* 601 * xfs_trans_unreserve_and_mod_sb() is called to release unused reservations and 602 * apply superblock counter changes to the in-core superblock. The 603 * t_res_fdblocks_delta and t_res_frextents_delta fields are explicitly NOT 604 * applied to the in-core superblock. The idea is that that has already been 605 * done. 606 * 607 * If we are not logging superblock counters, then the inode allocated/free and 608 * used block counts are not updated in the on disk superblock. In this case, 609 * XFS_TRANS_SB_DIRTY will not be set when the transaction is updated but we 610 * still need to update the incore superblock with the changes. 611 * 612 * Deltas for the inode count are +/-64, hence we use a large batch size of 128 613 * so we don't need to take the counter lock on every update. 614 */ 615 #define XFS_ICOUNT_BATCH 128 616 617 void 618 xfs_trans_unreserve_and_mod_sb( 619 struct xfs_trans *tp) 620 { 621 struct xfs_mount *mp = tp->t_mountp; 622 int64_t blkdelta = tp->t_blk_res; 623 int64_t rtxdelta = tp->t_rtx_res; 624 int64_t idelta = 0; 625 int64_t ifreedelta = 0; 626 627 /* 628 * Calculate the deltas. 629 * 630 * t_fdblocks_delta and t_frextents_delta can be positive or negative: 631 * 632 * - positive values indicate blocks freed in the transaction. 633 * - negative values indicate blocks allocated in the transaction 634 * 635 * Negative values can only happen if the transaction has a block 636 * reservation that covers the allocated block. The end result is 637 * that the calculated delta values must always be positive and we 638 * can only put back previous allocated or reserved blocks here. 639 */ 640 ASSERT(tp->t_blk_res || tp->t_fdblocks_delta >= 0); 641 if (xfs_has_lazysbcount(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) { 642 blkdelta += tp->t_fdblocks_delta; 643 ASSERT(blkdelta >= 0); 644 } 645 646 ASSERT(tp->t_rtx_res || tp->t_frextents_delta >= 0); 647 if (xfs_has_rtgroups(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) { 648 rtxdelta += tp->t_frextents_delta; 649 ASSERT(rtxdelta >= 0); 650 } 651 652 if (xfs_has_lazysbcount(mp) || (tp->t_flags & XFS_TRANS_SB_DIRTY)) { 653 idelta = tp->t_icount_delta; 654 ifreedelta = tp->t_ifree_delta; 655 } 656 657 /* apply the per-cpu counters */ 658 if (blkdelta) 659 xfs_add_fdblocks(mp, blkdelta); 660 661 if (idelta) 662 percpu_counter_add_batch(&mp->m_icount, idelta, 663 XFS_ICOUNT_BATCH); 664 665 if (ifreedelta) 666 percpu_counter_add(&mp->m_ifree, ifreedelta); 667 668 if (rtxdelta) 669 xfs_add_frextents(mp, rtxdelta); 670 671 if (!(tp->t_flags & XFS_TRANS_SB_DIRTY)) 672 return; 673 674 /* apply remaining deltas */ 675 spin_lock(&mp->m_sb_lock); 676 mp->m_sb.sb_fdblocks += tp->t_fdblocks_delta + tp->t_res_fdblocks_delta; 677 mp->m_sb.sb_icount += idelta; 678 mp->m_sb.sb_ifree += ifreedelta; 679 /* 680 * Do not touch sb_frextents here because it is handled in 681 * xfs_trans_apply_sb_deltas for file systems where it isn't a lazy 682 * counter anyway. 683 */ 684 mp->m_sb.sb_dblocks += tp->t_dblocks_delta; 685 mp->m_sb.sb_agcount += tp->t_agcount_delta; 686 mp->m_sb.sb_imax_pct += tp->t_imaxpct_delta; 687 if (tp->t_rextsize_delta) 688 xfs_mount_sb_set_rextsize(mp, &mp->m_sb, 689 mp->m_sb.sb_rextsize + tp->t_rextsize_delta); 690 mp->m_sb.sb_rbmblocks += tp->t_rbmblocks_delta; 691 mp->m_sb.sb_rblocks += tp->t_rblocks_delta; 692 mp->m_sb.sb_rextents += tp->t_rextents_delta; 693 mp->m_sb.sb_rextslog += tp->t_rextslog_delta; 694 mp->m_sb.sb_rgcount += tp->t_rgcount_delta; 695 spin_unlock(&mp->m_sb_lock); 696 697 /* 698 * Debug checks outside of the spinlock so they don't lock up the 699 * machine if they fail. 700 */ 701 ASSERT(mp->m_sb.sb_imax_pct >= 0); 702 ASSERT(mp->m_sb.sb_rextslog >= 0); 703 } 704 705 /* Add the given log item to the transaction's list of log items. */ 706 void 707 xfs_trans_add_item( 708 struct xfs_trans *tp, 709 struct xfs_log_item *lip) 710 { 711 ASSERT(lip->li_log == tp->t_mountp->m_log); 712 ASSERT(lip->li_ailp == tp->t_mountp->m_ail); 713 ASSERT(list_empty(&lip->li_trans)); 714 ASSERT(!test_bit(XFS_LI_DIRTY, &lip->li_flags)); 715 716 list_add_tail(&lip->li_trans, &tp->t_items); 717 trace_xfs_trans_add_item(tp, _RET_IP_); 718 } 719 720 /* 721 * Unlink the log item from the transaction. the log item is no longer 722 * considered dirty in this transaction, as the linked transaction has 723 * finished, either by abort or commit completion. 724 */ 725 void 726 xfs_trans_del_item( 727 struct xfs_log_item *lip) 728 { 729 clear_bit(XFS_LI_DIRTY, &lip->li_flags); 730 list_del_init(&lip->li_trans); 731 } 732 733 /* Detach and unlock all of the items in a transaction */ 734 static void 735 xfs_trans_free_items( 736 struct xfs_trans *tp, 737 bool abort) 738 { 739 struct xfs_log_item *lip, *next; 740 741 trace_xfs_trans_free_items(tp, _RET_IP_); 742 743 list_for_each_entry_safe(lip, next, &tp->t_items, li_trans) { 744 xfs_trans_del_item(lip); 745 if (abort) { 746 trace_xfs_trans_free_abort(lip); 747 set_bit(XFS_LI_ABORTED, &lip->li_flags); 748 } 749 if (lip->li_ops->iop_release) 750 lip->li_ops->iop_release(lip); 751 } 752 } 753 754 /* 755 * Sort transaction items prior to running precommit operations. This will 756 * attempt to order the items such that they will always be locked in the same 757 * order. Items that have no sort function are moved to the end of the list 758 * and so are locked last. 759 * 760 * This may need refinement as different types of objects add sort functions. 761 * 762 * Function is more complex than it needs to be because we are comparing 64 bit 763 * values and the function only returns 32 bit values. 764 */ 765 static int 766 xfs_trans_precommit_sort( 767 void *unused_arg, 768 const struct list_head *a, 769 const struct list_head *b) 770 { 771 struct xfs_log_item *lia = container_of(a, 772 struct xfs_log_item, li_trans); 773 struct xfs_log_item *lib = container_of(b, 774 struct xfs_log_item, li_trans); 775 int64_t diff; 776 777 /* 778 * If both items are non-sortable, leave them alone. If only one is 779 * sortable, move the non-sortable item towards the end of the list. 780 */ 781 if (!lia->li_ops->iop_sort && !lib->li_ops->iop_sort) 782 return 0; 783 if (!lia->li_ops->iop_sort) 784 return 1; 785 if (!lib->li_ops->iop_sort) 786 return -1; 787 788 diff = lia->li_ops->iop_sort(lia) - lib->li_ops->iop_sort(lib); 789 if (diff < 0) 790 return -1; 791 if (diff > 0) 792 return 1; 793 return 0; 794 } 795 796 /* 797 * Run transaction precommit functions. 798 * 799 * If there is an error in any of the callouts, then stop immediately and 800 * trigger a shutdown to abort the transaction. There is no recovery possible 801 * from errors at this point as the transaction is dirty.... 802 */ 803 static int 804 xfs_trans_run_precommits( 805 struct xfs_trans *tp) 806 { 807 struct xfs_mount *mp = tp->t_mountp; 808 struct xfs_log_item *lip, *n; 809 int error = 0; 810 811 /* 812 * Sort the item list to avoid ABBA deadlocks with other transactions 813 * running precommit operations that lock multiple shared items such as 814 * inode cluster buffers. 815 */ 816 list_sort(NULL, &tp->t_items, xfs_trans_precommit_sort); 817 818 /* 819 * Precommit operations can remove the log item from the transaction 820 * if the log item exists purely to delay modifications until they 821 * can be ordered against other operations. Hence we have to use 822 * list_for_each_entry_safe() here. 823 */ 824 list_for_each_entry_safe(lip, n, &tp->t_items, li_trans) { 825 if (!test_bit(XFS_LI_DIRTY, &lip->li_flags)) 826 continue; 827 if (lip->li_ops->iop_precommit) { 828 error = lip->li_ops->iop_precommit(tp, lip); 829 if (error) 830 break; 831 } 832 } 833 if (error) 834 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE); 835 return error; 836 } 837 838 /* 839 * Commit the given transaction to the log. 840 * 841 * XFS disk error handling mechanism is not based on a typical 842 * transaction abort mechanism. Logically after the filesystem 843 * gets marked 'SHUTDOWN', we can't let any new transactions 844 * be durable - ie. committed to disk - because some metadata might 845 * be inconsistent. In such cases, this returns an error, and the 846 * caller may assume that all locked objects joined to the transaction 847 * have already been unlocked as if the commit had succeeded. 848 * Do not reference the transaction structure after this call. 849 */ 850 static int 851 __xfs_trans_commit( 852 struct xfs_trans *tp, 853 bool regrant) 854 { 855 struct xfs_mount *mp = tp->t_mountp; 856 struct xlog *log = mp->m_log; 857 xfs_csn_t commit_seq = 0; 858 int error = 0; 859 int sync = tp->t_flags & XFS_TRANS_SYNC; 860 861 trace_xfs_trans_commit(tp, _RET_IP_); 862 863 /* 864 * Commit per-transaction changes that are not already tracked through 865 * log items. This can add dirty log items to the transaction. 866 */ 867 if (tp->t_flags & XFS_TRANS_SB_DIRTY) 868 xfs_trans_apply_sb_deltas(tp); 869 xfs_trans_apply_dquot_deltas(tp); 870 871 error = xfs_trans_run_precommits(tp); 872 if (error) 873 goto out_unreserve; 874 875 /* 876 * If there is nothing to be logged by the transaction, 877 * then unlock all of the items associated with the 878 * transaction and free the transaction structure. 879 * Also make sure to return any reserved blocks to 880 * the free pool. 881 */ 882 if (!(tp->t_flags & XFS_TRANS_DIRTY)) 883 goto out_unreserve; 884 885 /* 886 * We must check against log shutdown here because we cannot abort log 887 * items and leave them dirty, inconsistent and unpinned in memory while 888 * the log is active. This leaves them open to being written back to 889 * disk, and that will lead to on-disk corruption. 890 */ 891 if (xlog_is_shutdown(log)) { 892 error = -EIO; 893 goto out_unreserve; 894 } 895 896 ASSERT(tp->t_ticket != NULL); 897 898 xlog_cil_commit(log, tp, &commit_seq, regrant); 899 900 xfs_trans_free(tp); 901 902 /* 903 * If the transaction needs to be synchronous, then force the 904 * log out now and wait for it. 905 */ 906 if (sync) { 907 error = xfs_log_force_seq(mp, commit_seq, XFS_LOG_SYNC, NULL); 908 XFS_STATS_INC(mp, xs_trans_sync); 909 } else { 910 XFS_STATS_INC(mp, xs_trans_async); 911 } 912 913 return error; 914 915 out_unreserve: 916 xfs_trans_unreserve_and_mod_sb(tp); 917 918 /* 919 * It is indeed possible for the transaction to be not dirty but 920 * the dqinfo portion to be. All that means is that we have some 921 * (non-persistent) quota reservations that need to be unreserved. 922 */ 923 xfs_trans_unreserve_and_mod_dquots(tp, true); 924 if (tp->t_ticket) { 925 if (regrant && !xlog_is_shutdown(log)) 926 xfs_log_ticket_regrant(log, tp->t_ticket); 927 else 928 xfs_log_ticket_ungrant(log, tp->t_ticket); 929 tp->t_ticket = NULL; 930 } 931 xfs_trans_free_items(tp, !!error); 932 xfs_trans_free(tp); 933 934 XFS_STATS_INC(mp, xs_trans_empty); 935 return error; 936 } 937 938 int 939 xfs_trans_commit( 940 struct xfs_trans *tp) 941 { 942 /* 943 * Finish deferred items on final commit. Only permanent transactions 944 * should ever have deferred ops. 945 */ 946 WARN_ON_ONCE(!list_empty(&tp->t_dfops) && 947 !(tp->t_flags & XFS_TRANS_PERM_LOG_RES)); 948 if (tp->t_flags & XFS_TRANS_PERM_LOG_RES) { 949 int error = xfs_defer_finish_noroll(&tp); 950 if (error) { 951 xfs_trans_cancel(tp); 952 return error; 953 } 954 } 955 956 return __xfs_trans_commit(tp, false); 957 } 958 959 /* 960 * Unlock all of the transaction's items and free the transaction. If the 961 * transaction is dirty, we must shut down the filesystem because there is no 962 * way to restore them to their previous state. 963 * 964 * If the transaction has made a log reservation, make sure to release it as 965 * well. 966 * 967 * This is a high level function (equivalent to xfs_trans_commit()) and so can 968 * be called after the transaction has effectively been aborted due to the mount 969 * being shut down. However, if the mount has not been shut down and the 970 * transaction is dirty we will shut the mount down and, in doing so, that 971 * guarantees that the log is shut down, too. Hence we don't need to be as 972 * careful with shutdown state and dirty items here as we need to be in 973 * xfs_trans_commit(). 974 */ 975 void 976 xfs_trans_cancel( 977 struct xfs_trans *tp) 978 { 979 struct xfs_mount *mp = tp->t_mountp; 980 struct xlog *log = mp->m_log; 981 bool dirty = (tp->t_flags & XFS_TRANS_DIRTY); 982 983 trace_xfs_trans_cancel(tp, _RET_IP_); 984 985 /* 986 * It's never valid to cancel a transaction with deferred ops attached, 987 * because the transaction is effectively dirty. Complain about this 988 * loudly before freeing the in-memory defer items and shutting down the 989 * filesystem. 990 */ 991 if (!list_empty(&tp->t_dfops)) { 992 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES); 993 dirty = true; 994 xfs_defer_cancel(tp); 995 } 996 997 /* 998 * See if the caller is relying on us to shut down the filesystem. We 999 * only want an error report if there isn't already a shutdown in 1000 * progress, so we only need to check against the mount shutdown state 1001 * here. 1002 */ 1003 if (dirty && !xfs_is_shutdown(mp)) { 1004 XFS_ERROR_REPORT("xfs_trans_cancel", XFS_ERRLEVEL_LOW, mp); 1005 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE); 1006 } 1007 #ifdef DEBUG 1008 /* Log items need to be consistent until the log is shut down. */ 1009 if (!dirty && !xlog_is_shutdown(log)) { 1010 struct xfs_log_item *lip; 1011 1012 list_for_each_entry(lip, &tp->t_items, li_trans) 1013 ASSERT(!xlog_item_is_intent_done(lip)); 1014 } 1015 #endif 1016 xfs_trans_unreserve_and_mod_sb(tp); 1017 xfs_trans_unreserve_and_mod_dquots(tp, false); 1018 1019 if (tp->t_ticket) { 1020 xfs_log_ticket_ungrant(log, tp->t_ticket); 1021 tp->t_ticket = NULL; 1022 } 1023 1024 xfs_trans_free_items(tp, dirty); 1025 xfs_trans_free(tp); 1026 } 1027 1028 /* 1029 * Roll from one trans in the sequence of PERMANENT transactions to 1030 * the next: permanent transactions are only flushed out when 1031 * committed with xfs_trans_commit(), but we still want as soon 1032 * as possible to let chunks of it go to the log. So we commit the 1033 * chunk we've been working on and get a new transaction to continue. 1034 */ 1035 int 1036 xfs_trans_roll( 1037 struct xfs_trans **tpp) 1038 { 1039 struct xfs_trans *trans = *tpp; 1040 struct xfs_trans_res tres; 1041 int error; 1042 1043 trace_xfs_trans_roll(trans, _RET_IP_); 1044 1045 /* 1046 * Copy the critical parameters from one trans to the next. 1047 */ 1048 tres.tr_logres = trans->t_log_res; 1049 tres.tr_logcount = trans->t_log_count; 1050 1051 *tpp = xfs_trans_dup(trans); 1052 1053 /* 1054 * Commit the current transaction. 1055 * If this commit failed, then it'd just unlock those items that 1056 * are not marked ihold. That also means that a filesystem shutdown 1057 * is in progress. The caller takes the responsibility to cancel 1058 * the duplicate transaction that gets returned. 1059 */ 1060 error = __xfs_trans_commit(trans, true); 1061 if (error) 1062 return error; 1063 1064 /* 1065 * Reserve space in the log for the next transaction. 1066 * This also pushes items in the "AIL", the list of logged items, 1067 * out to disk if they are taking up space at the tail of the log 1068 * that we want to use. This requires that either nothing be locked 1069 * across this call, or that anything that is locked be logged in 1070 * the prior and the next transactions. 1071 */ 1072 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES; 1073 return xfs_trans_reserve(*tpp, &tres, 0, 0); 1074 } 1075 1076 /* 1077 * Allocate an transaction, lock and join the inode to it, and reserve quota. 1078 * 1079 * The caller must ensure that the on-disk dquots attached to this inode have 1080 * already been allocated and initialized. The caller is responsible for 1081 * releasing ILOCK_EXCL if a new transaction is returned. 1082 */ 1083 int 1084 xfs_trans_alloc_inode( 1085 struct xfs_inode *ip, 1086 struct xfs_trans_res *resv, 1087 unsigned int dblocks, 1088 unsigned int rblocks, 1089 bool force, 1090 struct xfs_trans **tpp) 1091 { 1092 struct xfs_trans *tp; 1093 struct xfs_mount *mp = ip->i_mount; 1094 bool retried = false; 1095 int error; 1096 1097 retry: 1098 error = xfs_trans_alloc(mp, resv, dblocks, 1099 xfs_extlen_to_rtxlen(mp, rblocks), 1100 force ? XFS_TRANS_RESERVE : 0, &tp); 1101 if (error) 1102 return error; 1103 1104 xfs_ilock(ip, XFS_ILOCK_EXCL); 1105 xfs_trans_ijoin(tp, ip, 0); 1106 1107 error = xfs_qm_dqattach_locked(ip, false); 1108 if (error) { 1109 /* Caller should have allocated the dquots! */ 1110 ASSERT(error != -ENOENT); 1111 goto out_cancel; 1112 } 1113 1114 error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks, force); 1115 if ((error == -EDQUOT || error == -ENOSPC) && !retried) { 1116 xfs_trans_cancel(tp); 1117 xfs_iunlock(ip, XFS_ILOCK_EXCL); 1118 xfs_blockgc_free_quota(ip, 0); 1119 retried = true; 1120 goto retry; 1121 } 1122 if (error) 1123 goto out_cancel; 1124 1125 *tpp = tp; 1126 return 0; 1127 1128 out_cancel: 1129 xfs_trans_cancel(tp); 1130 xfs_iunlock(ip, XFS_ILOCK_EXCL); 1131 return error; 1132 } 1133 1134 /* 1135 * Try to reserve more blocks for a transaction. 1136 * 1137 * This is for callers that need to attach resources to a transaction, scan 1138 * those resources to determine the space reservation requirements, and then 1139 * modify the attached resources. In other words, online repair. This can 1140 * fail due to ENOSPC, so the caller must be able to cancel the transaction 1141 * without shutting down the fs. 1142 */ 1143 int 1144 xfs_trans_reserve_more( 1145 struct xfs_trans *tp, 1146 unsigned int blocks, 1147 unsigned int rtextents) 1148 { 1149 struct xfs_trans_res resv = { }; 1150 1151 return xfs_trans_reserve(tp, &resv, blocks, rtextents); 1152 } 1153 1154 /* 1155 * Try to reserve more blocks and file quota for a transaction. Same 1156 * conditions of usage as xfs_trans_reserve_more. 1157 */ 1158 int 1159 xfs_trans_reserve_more_inode( 1160 struct xfs_trans *tp, 1161 struct xfs_inode *ip, 1162 unsigned int dblocks, 1163 unsigned int rblocks, 1164 bool force_quota) 1165 { 1166 struct xfs_trans_res resv = { }; 1167 struct xfs_mount *mp = ip->i_mount; 1168 unsigned int rtx = xfs_extlen_to_rtxlen(mp, rblocks); 1169 int error; 1170 1171 xfs_assert_ilocked(ip, XFS_ILOCK_EXCL); 1172 1173 error = xfs_trans_reserve(tp, &resv, dblocks, rtx); 1174 if (error) 1175 return error; 1176 1177 if (!XFS_IS_QUOTA_ON(mp) || xfs_is_quota_inode(&mp->m_sb, ip->i_ino)) 1178 return 0; 1179 1180 if (tp->t_flags & XFS_TRANS_RESERVE) 1181 force_quota = true; 1182 1183 error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks, 1184 force_quota); 1185 if (!error) 1186 return 0; 1187 1188 /* Quota failed, give back the new reservation. */ 1189 xfs_add_fdblocks(mp, dblocks); 1190 tp->t_blk_res -= dblocks; 1191 xfs_add_frextents(mp, rtx); 1192 tp->t_rtx_res -= rtx; 1193 return error; 1194 } 1195 1196 /* 1197 * Allocate an transaction in preparation for inode creation by reserving quota 1198 * against the given dquots. Callers are not required to hold any inode locks. 1199 */ 1200 int 1201 xfs_trans_alloc_icreate( 1202 struct xfs_mount *mp, 1203 struct xfs_trans_res *resv, 1204 struct xfs_dquot *udqp, 1205 struct xfs_dquot *gdqp, 1206 struct xfs_dquot *pdqp, 1207 unsigned int dblocks, 1208 struct xfs_trans **tpp) 1209 { 1210 struct xfs_trans *tp; 1211 bool retried = false; 1212 int error; 1213 1214 retry: 1215 error = xfs_trans_alloc(mp, resv, dblocks, 0, 0, &tp); 1216 if (error) 1217 return error; 1218 1219 error = xfs_trans_reserve_quota_icreate(tp, udqp, gdqp, pdqp, dblocks); 1220 if ((error == -EDQUOT || error == -ENOSPC) && !retried) { 1221 xfs_trans_cancel(tp); 1222 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0); 1223 retried = true; 1224 goto retry; 1225 } 1226 if (error) { 1227 xfs_trans_cancel(tp); 1228 return error; 1229 } 1230 1231 *tpp = tp; 1232 return 0; 1233 } 1234 1235 /* 1236 * Allocate an transaction, lock and join the inode to it, and reserve quota 1237 * in preparation for inode attribute changes that include uid, gid, or prid 1238 * changes. 1239 * 1240 * The caller must ensure that the on-disk dquots attached to this inode have 1241 * already been allocated and initialized. The ILOCK will be dropped when the 1242 * transaction is committed or cancelled. 1243 */ 1244 int 1245 xfs_trans_alloc_ichange( 1246 struct xfs_inode *ip, 1247 struct xfs_dquot *new_udqp, 1248 struct xfs_dquot *new_gdqp, 1249 struct xfs_dquot *new_pdqp, 1250 bool force, 1251 struct xfs_trans **tpp) 1252 { 1253 struct xfs_trans *tp; 1254 struct xfs_mount *mp = ip->i_mount; 1255 struct xfs_dquot *udqp; 1256 struct xfs_dquot *gdqp; 1257 struct xfs_dquot *pdqp; 1258 bool retried = false; 1259 int error; 1260 1261 retry: 1262 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp); 1263 if (error) 1264 return error; 1265 1266 xfs_ilock(ip, XFS_ILOCK_EXCL); 1267 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL); 1268 1269 if (xfs_is_metadir_inode(ip)) 1270 goto out; 1271 1272 error = xfs_qm_dqattach_locked(ip, false); 1273 if (error) { 1274 /* Caller should have allocated the dquots! */ 1275 ASSERT(error != -ENOENT); 1276 goto out_cancel; 1277 } 1278 1279 /* 1280 * For each quota type, skip quota reservations if the inode's dquots 1281 * now match the ones that came from the caller, or the caller didn't 1282 * pass one in. The inode's dquots can change if we drop the ILOCK to 1283 * perform a blockgc scan, so we must preserve the caller's arguments. 1284 */ 1285 udqp = (new_udqp != ip->i_udquot) ? new_udqp : NULL; 1286 gdqp = (new_gdqp != ip->i_gdquot) ? new_gdqp : NULL; 1287 pdqp = (new_pdqp != ip->i_pdquot) ? new_pdqp : NULL; 1288 if (udqp || gdqp || pdqp) { 1289 xfs_filblks_t dblocks, rblocks; 1290 unsigned int qflags = XFS_QMOPT_RES_REGBLKS; 1291 bool isrt = XFS_IS_REALTIME_INODE(ip); 1292 1293 if (force) 1294 qflags |= XFS_QMOPT_FORCE_RES; 1295 1296 if (isrt) { 1297 error = xfs_iread_extents(tp, ip, XFS_DATA_FORK); 1298 if (error) 1299 goto out_cancel; 1300 } 1301 1302 xfs_inode_count_blocks(tp, ip, &dblocks, &rblocks); 1303 1304 if (isrt) 1305 rblocks += ip->i_delayed_blks; 1306 else 1307 dblocks += ip->i_delayed_blks; 1308 1309 /* 1310 * Reserve enough quota to handle blocks on disk and reserved 1311 * for a delayed allocation. We'll actually transfer the 1312 * delalloc reservation between dquots at chown time, even 1313 * though that part is only semi-transactional. 1314 */ 1315 error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp, 1316 pdqp, dblocks, 1, qflags); 1317 if ((error == -EDQUOT || error == -ENOSPC) && !retried) { 1318 xfs_trans_cancel(tp); 1319 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0); 1320 retried = true; 1321 goto retry; 1322 } 1323 if (error) 1324 goto out_cancel; 1325 1326 /* Do the same for realtime. */ 1327 qflags = XFS_QMOPT_RES_RTBLKS | (qflags & XFS_QMOPT_FORCE_RES); 1328 error = xfs_trans_reserve_quota_bydquots(tp, mp, udqp, gdqp, 1329 pdqp, rblocks, 0, qflags); 1330 if ((error == -EDQUOT || error == -ENOSPC) && !retried) { 1331 xfs_trans_cancel(tp); 1332 xfs_blockgc_free_dquots(mp, udqp, gdqp, pdqp, 0); 1333 retried = true; 1334 goto retry; 1335 } 1336 if (error) 1337 goto out_cancel; 1338 } 1339 1340 out: 1341 *tpp = tp; 1342 return 0; 1343 1344 out_cancel: 1345 xfs_trans_cancel(tp); 1346 return error; 1347 } 1348 1349 /* 1350 * Allocate an transaction, lock and join the directory and child inodes to it, 1351 * and reserve quota for a directory update. If there isn't sufficient space, 1352 * @dblocks will be set to zero for a reservationless directory update and 1353 * @nospace_error will be set to a negative errno describing the space 1354 * constraint we hit. 1355 * 1356 * The caller must ensure that the on-disk dquots attached to this inode have 1357 * already been allocated and initialized. The ILOCKs will be dropped when the 1358 * transaction is committed or cancelled. 1359 * 1360 * Caller is responsible for unlocking the inodes manually upon return 1361 */ 1362 int 1363 xfs_trans_alloc_dir( 1364 struct xfs_inode *dp, 1365 struct xfs_trans_res *resv, 1366 struct xfs_inode *ip, 1367 unsigned int *dblocks, 1368 struct xfs_trans **tpp, 1369 int *nospace_error) 1370 { 1371 struct xfs_trans *tp; 1372 struct xfs_mount *mp = ip->i_mount; 1373 unsigned int resblks; 1374 bool retried = false; 1375 int error; 1376 1377 retry: 1378 *nospace_error = 0; 1379 resblks = *dblocks; 1380 error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp); 1381 if (error == -ENOSPC) { 1382 *nospace_error = error; 1383 resblks = 0; 1384 error = xfs_trans_alloc(mp, resv, resblks, 0, 0, &tp); 1385 } 1386 if (error) 1387 return error; 1388 1389 xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL); 1390 1391 xfs_trans_ijoin(tp, dp, 0); 1392 xfs_trans_ijoin(tp, ip, 0); 1393 1394 error = xfs_qm_dqattach_locked(dp, false); 1395 if (error) { 1396 /* Caller should have allocated the dquots! */ 1397 ASSERT(error != -ENOENT); 1398 goto out_cancel; 1399 } 1400 1401 error = xfs_qm_dqattach_locked(ip, false); 1402 if (error) { 1403 /* Caller should have allocated the dquots! */ 1404 ASSERT(error != -ENOENT); 1405 goto out_cancel; 1406 } 1407 1408 if (resblks == 0) 1409 goto done; 1410 1411 error = xfs_trans_reserve_quota_nblks(tp, dp, resblks, 0, false); 1412 if (error == -EDQUOT || error == -ENOSPC) { 1413 if (!retried) { 1414 xfs_trans_cancel(tp); 1415 xfs_iunlock(dp, XFS_ILOCK_EXCL); 1416 if (dp != ip) 1417 xfs_iunlock(ip, XFS_ILOCK_EXCL); 1418 xfs_blockgc_free_quota(dp, 0); 1419 retried = true; 1420 goto retry; 1421 } 1422 1423 *nospace_error = error; 1424 resblks = 0; 1425 error = 0; 1426 } 1427 if (error) 1428 goto out_cancel; 1429 1430 done: 1431 *tpp = tp; 1432 *dblocks = resblks; 1433 return 0; 1434 1435 out_cancel: 1436 xfs_trans_cancel(tp); 1437 xfs_iunlock(dp, XFS_ILOCK_EXCL); 1438 if (dp != ip) 1439 xfs_iunlock(ip, XFS_ILOCK_EXCL); 1440 return error; 1441 } 1442