1 /* 2 * Copyright (c) 2000-2005 Silicon Graphics, Inc. 3 * All Rights Reserved. 4 * 5 * This program is free software; you can redistribute it and/or 6 * modify it under the terms of the GNU General Public License as 7 * published by the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it would be useful, 10 * but WITHOUT ANY WARRANTY; without even the implied warranty of 11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 12 * GNU General Public License for more details. 13 * 14 * You should have received a copy of the GNU General Public License 15 * along with this program; if not, write the Free Software Foundation, 16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA 17 */ 18 #include "xfs.h" 19 #include "xfs_fs.h" 20 #include "xfs_types.h" 21 #include "xfs_bit.h" 22 #include "xfs_log.h" 23 #include "xfs_inum.h" 24 #include "xfs_trans.h" 25 #include "xfs_sb.h" 26 #include "xfs_ag.h" 27 #include "xfs_mount.h" 28 #include "xfs_error.h" 29 #include "xfs_log_priv.h" 30 #include "xfs_buf_item.h" 31 #include "xfs_bmap_btree.h" 32 #include "xfs_alloc_btree.h" 33 #include "xfs_ialloc_btree.h" 34 #include "xfs_log_recover.h" 35 #include "xfs_trans_priv.h" 36 #include "xfs_dinode.h" 37 #include "xfs_inode.h" 38 #include "xfs_rw.h" 39 #include "xfs_trace.h" 40 41 kmem_zone_t *xfs_log_ticket_zone; 42 43 /* Local miscellaneous function prototypes */ 44 STATIC int xlog_commit_record(struct log *log, struct xlog_ticket *ticket, 45 xlog_in_core_t **, xfs_lsn_t *); 46 STATIC xlog_t * xlog_alloc_log(xfs_mount_t *mp, 47 xfs_buftarg_t *log_target, 48 xfs_daddr_t blk_offset, 49 int num_bblks); 50 STATIC int xlog_space_left(struct log *log, atomic64_t *head); 51 STATIC int xlog_sync(xlog_t *log, xlog_in_core_t *iclog); 52 STATIC void xlog_dealloc_log(xlog_t *log); 53 54 /* local state machine functions */ 55 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int); 56 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog); 57 STATIC int xlog_state_get_iclog_space(xlog_t *log, 58 int len, 59 xlog_in_core_t **iclog, 60 xlog_ticket_t *ticket, 61 int *continued_write, 62 int *logoffsetp); 63 STATIC int xlog_state_release_iclog(xlog_t *log, 64 xlog_in_core_t *iclog); 65 STATIC void xlog_state_switch_iclogs(xlog_t *log, 66 xlog_in_core_t *iclog, 67 int eventual_size); 68 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog); 69 70 /* local functions to manipulate grant head */ 71 STATIC int xlog_grant_log_space(xlog_t *log, 72 xlog_ticket_t *xtic); 73 STATIC void xlog_grant_push_ail(struct log *log, 74 int need_bytes); 75 STATIC void xlog_regrant_reserve_log_space(xlog_t *log, 76 xlog_ticket_t *ticket); 77 STATIC int xlog_regrant_write_log_space(xlog_t *log, 78 xlog_ticket_t *ticket); 79 STATIC void xlog_ungrant_log_space(xlog_t *log, 80 xlog_ticket_t *ticket); 81 82 #if defined(DEBUG) 83 STATIC void xlog_verify_dest_ptr(xlog_t *log, char *ptr); 84 STATIC void xlog_verify_grant_tail(struct log *log); 85 STATIC void xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog, 86 int count, boolean_t syncing); 87 STATIC void xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog, 88 xfs_lsn_t tail_lsn); 89 #else 90 #define xlog_verify_dest_ptr(a,b) 91 #define xlog_verify_grant_tail(a) 92 #define xlog_verify_iclog(a,b,c,d) 93 #define xlog_verify_tail_lsn(a,b,c) 94 #endif 95 96 STATIC int xlog_iclogs_empty(xlog_t *log); 97 98 static void 99 xlog_grant_sub_space( 100 struct log *log, 101 atomic64_t *head, 102 int bytes) 103 { 104 int64_t head_val = atomic64_read(head); 105 int64_t new, old; 106 107 do { 108 int cycle, space; 109 110 xlog_crack_grant_head_val(head_val, &cycle, &space); 111 112 space -= bytes; 113 if (space < 0) { 114 space += log->l_logsize; 115 cycle--; 116 } 117 118 old = head_val; 119 new = xlog_assign_grant_head_val(cycle, space); 120 head_val = atomic64_cmpxchg(head, old, new); 121 } while (head_val != old); 122 } 123 124 static void 125 xlog_grant_add_space( 126 struct log *log, 127 atomic64_t *head, 128 int bytes) 129 { 130 int64_t head_val = atomic64_read(head); 131 int64_t new, old; 132 133 do { 134 int tmp; 135 int cycle, space; 136 137 xlog_crack_grant_head_val(head_val, &cycle, &space); 138 139 tmp = log->l_logsize - space; 140 if (tmp > bytes) 141 space += bytes; 142 else { 143 space = bytes - tmp; 144 cycle++; 145 } 146 147 old = head_val; 148 new = xlog_assign_grant_head_val(cycle, space); 149 head_val = atomic64_cmpxchg(head, old, new); 150 } while (head_val != old); 151 } 152 153 static void 154 xlog_tic_reset_res(xlog_ticket_t *tic) 155 { 156 tic->t_res_num = 0; 157 tic->t_res_arr_sum = 0; 158 tic->t_res_num_ophdrs = 0; 159 } 160 161 static void 162 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type) 163 { 164 if (tic->t_res_num == XLOG_TIC_LEN_MAX) { 165 /* add to overflow and start again */ 166 tic->t_res_o_flow += tic->t_res_arr_sum; 167 tic->t_res_num = 0; 168 tic->t_res_arr_sum = 0; 169 } 170 171 tic->t_res_arr[tic->t_res_num].r_len = len; 172 tic->t_res_arr[tic->t_res_num].r_type = type; 173 tic->t_res_arr_sum += len; 174 tic->t_res_num++; 175 } 176 177 /* 178 * NOTES: 179 * 180 * 1. currblock field gets updated at startup and after in-core logs 181 * marked as with WANT_SYNC. 182 */ 183 184 /* 185 * This routine is called when a user of a log manager ticket is done with 186 * the reservation. If the ticket was ever used, then a commit record for 187 * the associated transaction is written out as a log operation header with 188 * no data. The flag XLOG_TIC_INITED is set when the first write occurs with 189 * a given ticket. If the ticket was one with a permanent reservation, then 190 * a few operations are done differently. Permanent reservation tickets by 191 * default don't release the reservation. They just commit the current 192 * transaction with the belief that the reservation is still needed. A flag 193 * must be passed in before permanent reservations are actually released. 194 * When these type of tickets are not released, they need to be set into 195 * the inited state again. By doing this, a start record will be written 196 * out when the next write occurs. 197 */ 198 xfs_lsn_t 199 xfs_log_done( 200 struct xfs_mount *mp, 201 struct xlog_ticket *ticket, 202 struct xlog_in_core **iclog, 203 uint flags) 204 { 205 struct log *log = mp->m_log; 206 xfs_lsn_t lsn = 0; 207 208 if (XLOG_FORCED_SHUTDOWN(log) || 209 /* 210 * If nothing was ever written, don't write out commit record. 211 * If we get an error, just continue and give back the log ticket. 212 */ 213 (((ticket->t_flags & XLOG_TIC_INITED) == 0) && 214 (xlog_commit_record(log, ticket, iclog, &lsn)))) { 215 lsn = (xfs_lsn_t) -1; 216 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) { 217 flags |= XFS_LOG_REL_PERM_RESERV; 218 } 219 } 220 221 222 if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 || 223 (flags & XFS_LOG_REL_PERM_RESERV)) { 224 trace_xfs_log_done_nonperm(log, ticket); 225 226 /* 227 * Release ticket if not permanent reservation or a specific 228 * request has been made to release a permanent reservation. 229 */ 230 xlog_ungrant_log_space(log, ticket); 231 xfs_log_ticket_put(ticket); 232 } else { 233 trace_xfs_log_done_perm(log, ticket); 234 235 xlog_regrant_reserve_log_space(log, ticket); 236 /* If this ticket was a permanent reservation and we aren't 237 * trying to release it, reset the inited flags; so next time 238 * we write, a start record will be written out. 239 */ 240 ticket->t_flags |= XLOG_TIC_INITED; 241 } 242 243 return lsn; 244 } 245 246 /* 247 * Attaches a new iclog I/O completion callback routine during 248 * transaction commit. If the log is in error state, a non-zero 249 * return code is handed back and the caller is responsible for 250 * executing the callback at an appropriate time. 251 */ 252 int 253 xfs_log_notify( 254 struct xfs_mount *mp, 255 struct xlog_in_core *iclog, 256 xfs_log_callback_t *cb) 257 { 258 int abortflg; 259 260 spin_lock(&iclog->ic_callback_lock); 261 abortflg = (iclog->ic_state & XLOG_STATE_IOERROR); 262 if (!abortflg) { 263 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) || 264 (iclog->ic_state == XLOG_STATE_WANT_SYNC)); 265 cb->cb_next = NULL; 266 *(iclog->ic_callback_tail) = cb; 267 iclog->ic_callback_tail = &(cb->cb_next); 268 } 269 spin_unlock(&iclog->ic_callback_lock); 270 return abortflg; 271 } 272 273 int 274 xfs_log_release_iclog( 275 struct xfs_mount *mp, 276 struct xlog_in_core *iclog) 277 { 278 if (xlog_state_release_iclog(mp->m_log, iclog)) { 279 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR); 280 return EIO; 281 } 282 283 return 0; 284 } 285 286 /* 287 * 1. Reserve an amount of on-disk log space and return a ticket corresponding 288 * to the reservation. 289 * 2. Potentially, push buffers at tail of log to disk. 290 * 291 * Each reservation is going to reserve extra space for a log record header. 292 * When writes happen to the on-disk log, we don't subtract the length of the 293 * log record header from any reservation. By wasting space in each 294 * reservation, we prevent over allocation problems. 295 */ 296 int 297 xfs_log_reserve( 298 struct xfs_mount *mp, 299 int unit_bytes, 300 int cnt, 301 struct xlog_ticket **ticket, 302 __uint8_t client, 303 uint flags, 304 uint t_type) 305 { 306 struct log *log = mp->m_log; 307 struct xlog_ticket *internal_ticket; 308 int retval = 0; 309 310 ASSERT(client == XFS_TRANSACTION || client == XFS_LOG); 311 312 if (XLOG_FORCED_SHUTDOWN(log)) 313 return XFS_ERROR(EIO); 314 315 XFS_STATS_INC(xs_try_logspace); 316 317 318 if (*ticket != NULL) { 319 ASSERT(flags & XFS_LOG_PERM_RESERV); 320 internal_ticket = *ticket; 321 322 /* 323 * this is a new transaction on the ticket, so we need to 324 * change the transaction ID so that the next transaction has a 325 * different TID in the log. Just add one to the existing tid 326 * so that we can see chains of rolling transactions in the log 327 * easily. 328 */ 329 internal_ticket->t_tid++; 330 331 trace_xfs_log_reserve(log, internal_ticket); 332 333 xlog_grant_push_ail(log, internal_ticket->t_unit_res); 334 retval = xlog_regrant_write_log_space(log, internal_ticket); 335 } else { 336 /* may sleep if need to allocate more tickets */ 337 internal_ticket = xlog_ticket_alloc(log, unit_bytes, cnt, 338 client, flags, 339 KM_SLEEP|KM_MAYFAIL); 340 if (!internal_ticket) 341 return XFS_ERROR(ENOMEM); 342 internal_ticket->t_trans_type = t_type; 343 *ticket = internal_ticket; 344 345 trace_xfs_log_reserve(log, internal_ticket); 346 347 xlog_grant_push_ail(log, 348 (internal_ticket->t_unit_res * 349 internal_ticket->t_cnt)); 350 retval = xlog_grant_log_space(log, internal_ticket); 351 } 352 353 return retval; 354 } /* xfs_log_reserve */ 355 356 357 /* 358 * Mount a log filesystem 359 * 360 * mp - ubiquitous xfs mount point structure 361 * log_target - buftarg of on-disk log device 362 * blk_offset - Start block # where block size is 512 bytes (BBSIZE) 363 * num_bblocks - Number of BBSIZE blocks in on-disk log 364 * 365 * Return error or zero. 366 */ 367 int 368 xfs_log_mount( 369 xfs_mount_t *mp, 370 xfs_buftarg_t *log_target, 371 xfs_daddr_t blk_offset, 372 int num_bblks) 373 { 374 int error; 375 376 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) 377 xfs_notice(mp, "Mounting Filesystem"); 378 else { 379 xfs_notice(mp, 380 "Mounting filesystem in no-recovery mode. Filesystem will be inconsistent."); 381 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY); 382 } 383 384 mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks); 385 if (IS_ERR(mp->m_log)) { 386 error = -PTR_ERR(mp->m_log); 387 goto out; 388 } 389 390 /* 391 * Initialize the AIL now we have a log. 392 */ 393 error = xfs_trans_ail_init(mp); 394 if (error) { 395 xfs_warn(mp, "AIL initialisation failed: error %d", error); 396 goto out_free_log; 397 } 398 mp->m_log->l_ailp = mp->m_ail; 399 400 /* 401 * skip log recovery on a norecovery mount. pretend it all 402 * just worked. 403 */ 404 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) { 405 int readonly = (mp->m_flags & XFS_MOUNT_RDONLY); 406 407 if (readonly) 408 mp->m_flags &= ~XFS_MOUNT_RDONLY; 409 410 error = xlog_recover(mp->m_log); 411 412 if (readonly) 413 mp->m_flags |= XFS_MOUNT_RDONLY; 414 if (error) { 415 xfs_warn(mp, "log mount/recovery failed: error %d", 416 error); 417 goto out_destroy_ail; 418 } 419 } 420 421 /* Normal transactions can now occur */ 422 mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY; 423 424 /* 425 * Now the log has been fully initialised and we know were our 426 * space grant counters are, we can initialise the permanent ticket 427 * needed for delayed logging to work. 428 */ 429 xlog_cil_init_post_recovery(mp->m_log); 430 431 return 0; 432 433 out_destroy_ail: 434 xfs_trans_ail_destroy(mp); 435 out_free_log: 436 xlog_dealloc_log(mp->m_log); 437 out: 438 return error; 439 } 440 441 /* 442 * Finish the recovery of the file system. This is separate from 443 * the xfs_log_mount() call, because it depends on the code in 444 * xfs_mountfs() to read in the root and real-time bitmap inodes 445 * between calling xfs_log_mount() and here. 446 * 447 * mp - ubiquitous xfs mount point structure 448 */ 449 int 450 xfs_log_mount_finish(xfs_mount_t *mp) 451 { 452 int error; 453 454 if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) 455 error = xlog_recover_finish(mp->m_log); 456 else { 457 error = 0; 458 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY); 459 } 460 461 return error; 462 } 463 464 /* 465 * Final log writes as part of unmount. 466 * 467 * Mark the filesystem clean as unmount happens. Note that during relocation 468 * this routine needs to be executed as part of source-bag while the 469 * deallocation must not be done until source-end. 470 */ 471 472 /* 473 * Unmount record used to have a string "Unmount filesystem--" in the 474 * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE). 475 * We just write the magic number now since that particular field isn't 476 * currently architecture converted and "nUmount" is a bit foo. 477 * As far as I know, there weren't any dependencies on the old behaviour. 478 */ 479 480 int 481 xfs_log_unmount_write(xfs_mount_t *mp) 482 { 483 xlog_t *log = mp->m_log; 484 xlog_in_core_t *iclog; 485 #ifdef DEBUG 486 xlog_in_core_t *first_iclog; 487 #endif 488 xlog_ticket_t *tic = NULL; 489 xfs_lsn_t lsn; 490 int error; 491 492 /* 493 * Don't write out unmount record on read-only mounts. 494 * Or, if we are doing a forced umount (typically because of IO errors). 495 */ 496 if (mp->m_flags & XFS_MOUNT_RDONLY) 497 return 0; 498 499 error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL); 500 ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log))); 501 502 #ifdef DEBUG 503 first_iclog = iclog = log->l_iclog; 504 do { 505 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) { 506 ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE); 507 ASSERT(iclog->ic_offset == 0); 508 } 509 iclog = iclog->ic_next; 510 } while (iclog != first_iclog); 511 #endif 512 if (! (XLOG_FORCED_SHUTDOWN(log))) { 513 error = xfs_log_reserve(mp, 600, 1, &tic, 514 XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE); 515 if (!error) { 516 /* the data section must be 32 bit size aligned */ 517 struct { 518 __uint16_t magic; 519 __uint16_t pad1; 520 __uint32_t pad2; /* may as well make it 64 bits */ 521 } magic = { 522 .magic = XLOG_UNMOUNT_TYPE, 523 }; 524 struct xfs_log_iovec reg = { 525 .i_addr = &magic, 526 .i_len = sizeof(magic), 527 .i_type = XLOG_REG_TYPE_UNMOUNT, 528 }; 529 struct xfs_log_vec vec = { 530 .lv_niovecs = 1, 531 .lv_iovecp = ®, 532 }; 533 534 /* remove inited flag */ 535 tic->t_flags = 0; 536 error = xlog_write(log, &vec, tic, &lsn, 537 NULL, XLOG_UNMOUNT_TRANS); 538 /* 539 * At this point, we're umounting anyway, 540 * so there's no point in transitioning log state 541 * to IOERROR. Just continue... 542 */ 543 } 544 545 if (error) 546 xfs_alert(mp, "%s: unmount record failed", __func__); 547 548 549 spin_lock(&log->l_icloglock); 550 iclog = log->l_iclog; 551 atomic_inc(&iclog->ic_refcnt); 552 xlog_state_want_sync(log, iclog); 553 spin_unlock(&log->l_icloglock); 554 error = xlog_state_release_iclog(log, iclog); 555 556 spin_lock(&log->l_icloglock); 557 if (!(iclog->ic_state == XLOG_STATE_ACTIVE || 558 iclog->ic_state == XLOG_STATE_DIRTY)) { 559 if (!XLOG_FORCED_SHUTDOWN(log)) { 560 xlog_wait(&iclog->ic_force_wait, 561 &log->l_icloglock); 562 } else { 563 spin_unlock(&log->l_icloglock); 564 } 565 } else { 566 spin_unlock(&log->l_icloglock); 567 } 568 if (tic) { 569 trace_xfs_log_umount_write(log, tic); 570 xlog_ungrant_log_space(log, tic); 571 xfs_log_ticket_put(tic); 572 } 573 } else { 574 /* 575 * We're already in forced_shutdown mode, couldn't 576 * even attempt to write out the unmount transaction. 577 * 578 * Go through the motions of sync'ing and releasing 579 * the iclog, even though no I/O will actually happen, 580 * we need to wait for other log I/Os that may already 581 * be in progress. Do this as a separate section of 582 * code so we'll know if we ever get stuck here that 583 * we're in this odd situation of trying to unmount 584 * a file system that went into forced_shutdown as 585 * the result of an unmount.. 586 */ 587 spin_lock(&log->l_icloglock); 588 iclog = log->l_iclog; 589 atomic_inc(&iclog->ic_refcnt); 590 591 xlog_state_want_sync(log, iclog); 592 spin_unlock(&log->l_icloglock); 593 error = xlog_state_release_iclog(log, iclog); 594 595 spin_lock(&log->l_icloglock); 596 597 if ( ! ( iclog->ic_state == XLOG_STATE_ACTIVE 598 || iclog->ic_state == XLOG_STATE_DIRTY 599 || iclog->ic_state == XLOG_STATE_IOERROR) ) { 600 601 xlog_wait(&iclog->ic_force_wait, 602 &log->l_icloglock); 603 } else { 604 spin_unlock(&log->l_icloglock); 605 } 606 } 607 608 return error; 609 } /* xfs_log_unmount_write */ 610 611 /* 612 * Deallocate log structures for unmount/relocation. 613 * 614 * We need to stop the aild from running before we destroy 615 * and deallocate the log as the aild references the log. 616 */ 617 void 618 xfs_log_unmount(xfs_mount_t *mp) 619 { 620 xfs_trans_ail_destroy(mp); 621 xlog_dealloc_log(mp->m_log); 622 } 623 624 void 625 xfs_log_item_init( 626 struct xfs_mount *mp, 627 struct xfs_log_item *item, 628 int type, 629 struct xfs_item_ops *ops) 630 { 631 item->li_mountp = mp; 632 item->li_ailp = mp->m_ail; 633 item->li_type = type; 634 item->li_ops = ops; 635 item->li_lv = NULL; 636 637 INIT_LIST_HEAD(&item->li_ail); 638 INIT_LIST_HEAD(&item->li_cil); 639 } 640 641 /* 642 * Write region vectors to log. The write happens using the space reservation 643 * of the ticket (tic). It is not a requirement that all writes for a given 644 * transaction occur with one call to xfs_log_write(). However, it is important 645 * to note that the transaction reservation code makes an assumption about the 646 * number of log headers a transaction requires that may be violated if you 647 * don't pass all the transaction vectors in one call.... 648 */ 649 int 650 xfs_log_write( 651 struct xfs_mount *mp, 652 struct xfs_log_iovec reg[], 653 int nentries, 654 struct xlog_ticket *tic, 655 xfs_lsn_t *start_lsn) 656 { 657 struct log *log = mp->m_log; 658 int error; 659 struct xfs_log_vec vec = { 660 .lv_niovecs = nentries, 661 .lv_iovecp = reg, 662 }; 663 664 if (XLOG_FORCED_SHUTDOWN(log)) 665 return XFS_ERROR(EIO); 666 667 error = xlog_write(log, &vec, tic, start_lsn, NULL, 0); 668 if (error) 669 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR); 670 return error; 671 } 672 673 void 674 xfs_log_move_tail(xfs_mount_t *mp, 675 xfs_lsn_t tail_lsn) 676 { 677 xlog_ticket_t *tic; 678 xlog_t *log = mp->m_log; 679 int need_bytes, free_bytes; 680 681 if (XLOG_FORCED_SHUTDOWN(log)) 682 return; 683 684 if (tail_lsn == 0) 685 tail_lsn = atomic64_read(&log->l_last_sync_lsn); 686 687 /* tail_lsn == 1 implies that we weren't passed a valid value. */ 688 if (tail_lsn != 1) 689 atomic64_set(&log->l_tail_lsn, tail_lsn); 690 691 if (!list_empty_careful(&log->l_writeq)) { 692 #ifdef DEBUG 693 if (log->l_flags & XLOG_ACTIVE_RECOVERY) 694 panic("Recovery problem"); 695 #endif 696 spin_lock(&log->l_grant_write_lock); 697 free_bytes = xlog_space_left(log, &log->l_grant_write_head); 698 list_for_each_entry(tic, &log->l_writeq, t_queue) { 699 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV); 700 701 if (free_bytes < tic->t_unit_res && tail_lsn != 1) 702 break; 703 tail_lsn = 0; 704 free_bytes -= tic->t_unit_res; 705 trace_xfs_log_regrant_write_wake_up(log, tic); 706 wake_up(&tic->t_wait); 707 } 708 spin_unlock(&log->l_grant_write_lock); 709 } 710 711 if (!list_empty_careful(&log->l_reserveq)) { 712 #ifdef DEBUG 713 if (log->l_flags & XLOG_ACTIVE_RECOVERY) 714 panic("Recovery problem"); 715 #endif 716 spin_lock(&log->l_grant_reserve_lock); 717 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head); 718 list_for_each_entry(tic, &log->l_reserveq, t_queue) { 719 if (tic->t_flags & XLOG_TIC_PERM_RESERV) 720 need_bytes = tic->t_unit_res*tic->t_cnt; 721 else 722 need_bytes = tic->t_unit_res; 723 if (free_bytes < need_bytes && tail_lsn != 1) 724 break; 725 tail_lsn = 0; 726 free_bytes -= need_bytes; 727 trace_xfs_log_grant_wake_up(log, tic); 728 wake_up(&tic->t_wait); 729 } 730 spin_unlock(&log->l_grant_reserve_lock); 731 } 732 } 733 734 /* 735 * Determine if we have a transaction that has gone to disk 736 * that needs to be covered. To begin the transition to the idle state 737 * firstly the log needs to be idle (no AIL and nothing in the iclogs). 738 * If we are then in a state where covering is needed, the caller is informed 739 * that dummy transactions are required to move the log into the idle state. 740 * 741 * Because this is called as part of the sync process, we should also indicate 742 * that dummy transactions should be issued in anything but the covered or 743 * idle states. This ensures that the log tail is accurately reflected in 744 * the log at the end of the sync, hence if a crash occurrs avoids replay 745 * of transactions where the metadata is already on disk. 746 */ 747 int 748 xfs_log_need_covered(xfs_mount_t *mp) 749 { 750 int needed = 0; 751 xlog_t *log = mp->m_log; 752 753 if (!xfs_fs_writable(mp)) 754 return 0; 755 756 spin_lock(&log->l_icloglock); 757 switch (log->l_covered_state) { 758 case XLOG_STATE_COVER_DONE: 759 case XLOG_STATE_COVER_DONE2: 760 case XLOG_STATE_COVER_IDLE: 761 break; 762 case XLOG_STATE_COVER_NEED: 763 case XLOG_STATE_COVER_NEED2: 764 if (!xfs_ail_min_lsn(log->l_ailp) && 765 xlog_iclogs_empty(log)) { 766 if (log->l_covered_state == XLOG_STATE_COVER_NEED) 767 log->l_covered_state = XLOG_STATE_COVER_DONE; 768 else 769 log->l_covered_state = XLOG_STATE_COVER_DONE2; 770 } 771 /* FALLTHRU */ 772 default: 773 needed = 1; 774 break; 775 } 776 spin_unlock(&log->l_icloglock); 777 return needed; 778 } 779 780 /****************************************************************************** 781 * 782 * local routines 783 * 784 ****************************************************************************** 785 */ 786 787 /* xfs_trans_tail_ail returns 0 when there is nothing in the list. 788 * The log manager must keep track of the last LR which was committed 789 * to disk. The lsn of this LR will become the new tail_lsn whenever 790 * xfs_trans_tail_ail returns 0. If we don't do this, we run into 791 * the situation where stuff could be written into the log but nothing 792 * was ever in the AIL when asked. Eventually, we panic since the 793 * tail hits the head. 794 * 795 * We may be holding the log iclog lock upon entering this routine. 796 */ 797 xfs_lsn_t 798 xlog_assign_tail_lsn( 799 struct xfs_mount *mp) 800 { 801 xfs_lsn_t tail_lsn; 802 struct log *log = mp->m_log; 803 804 tail_lsn = xfs_ail_min_lsn(mp->m_ail); 805 if (!tail_lsn) 806 tail_lsn = atomic64_read(&log->l_last_sync_lsn); 807 808 atomic64_set(&log->l_tail_lsn, tail_lsn); 809 return tail_lsn; 810 } 811 812 /* 813 * Return the space in the log between the tail and the head. The head 814 * is passed in the cycle/bytes formal parms. In the special case where 815 * the reserve head has wrapped passed the tail, this calculation is no 816 * longer valid. In this case, just return 0 which means there is no space 817 * in the log. This works for all places where this function is called 818 * with the reserve head. Of course, if the write head were to ever 819 * wrap the tail, we should blow up. Rather than catch this case here, 820 * we depend on other ASSERTions in other parts of the code. XXXmiken 821 * 822 * This code also handles the case where the reservation head is behind 823 * the tail. The details of this case are described below, but the end 824 * result is that we return the size of the log as the amount of space left. 825 */ 826 STATIC int 827 xlog_space_left( 828 struct log *log, 829 atomic64_t *head) 830 { 831 int free_bytes; 832 int tail_bytes; 833 int tail_cycle; 834 int head_cycle; 835 int head_bytes; 836 837 xlog_crack_grant_head(head, &head_cycle, &head_bytes); 838 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes); 839 tail_bytes = BBTOB(tail_bytes); 840 if (tail_cycle == head_cycle && head_bytes >= tail_bytes) 841 free_bytes = log->l_logsize - (head_bytes - tail_bytes); 842 else if (tail_cycle + 1 < head_cycle) 843 return 0; 844 else if (tail_cycle < head_cycle) { 845 ASSERT(tail_cycle == (head_cycle - 1)); 846 free_bytes = tail_bytes - head_bytes; 847 } else { 848 /* 849 * The reservation head is behind the tail. 850 * In this case we just want to return the size of the 851 * log as the amount of space left. 852 */ 853 xfs_alert(log->l_mp, 854 "xlog_space_left: head behind tail\n" 855 " tail_cycle = %d, tail_bytes = %d\n" 856 " GH cycle = %d, GH bytes = %d", 857 tail_cycle, tail_bytes, head_cycle, head_bytes); 858 ASSERT(0); 859 free_bytes = log->l_logsize; 860 } 861 return free_bytes; 862 } 863 864 865 /* 866 * Log function which is called when an io completes. 867 * 868 * The log manager needs its own routine, in order to control what 869 * happens with the buffer after the write completes. 870 */ 871 void 872 xlog_iodone(xfs_buf_t *bp) 873 { 874 xlog_in_core_t *iclog; 875 xlog_t *l; 876 int aborted; 877 878 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *); 879 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long) 2); 880 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1); 881 aborted = 0; 882 l = iclog->ic_log; 883 884 /* 885 * Race to shutdown the filesystem if we see an error. 886 */ 887 if (XFS_TEST_ERROR((XFS_BUF_GETERROR(bp)), l->l_mp, 888 XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) { 889 xfs_ioerror_alert("xlog_iodone", l->l_mp, bp, XFS_BUF_ADDR(bp)); 890 XFS_BUF_STALE(bp); 891 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR); 892 /* 893 * This flag will be propagated to the trans-committed 894 * callback routines to let them know that the log-commit 895 * didn't succeed. 896 */ 897 aborted = XFS_LI_ABORTED; 898 } else if (iclog->ic_state & XLOG_STATE_IOERROR) { 899 aborted = XFS_LI_ABORTED; 900 } 901 902 /* log I/O is always issued ASYNC */ 903 ASSERT(XFS_BUF_ISASYNC(bp)); 904 xlog_state_done_syncing(iclog, aborted); 905 /* 906 * do not reference the buffer (bp) here as we could race 907 * with it being freed after writing the unmount record to the 908 * log. 909 */ 910 911 } /* xlog_iodone */ 912 913 /* 914 * Return size of each in-core log record buffer. 915 * 916 * All machines get 8 x 32kB buffers by default, unless tuned otherwise. 917 * 918 * If the filesystem blocksize is too large, we may need to choose a 919 * larger size since the directory code currently logs entire blocks. 920 */ 921 922 STATIC void 923 xlog_get_iclog_buffer_size(xfs_mount_t *mp, 924 xlog_t *log) 925 { 926 int size; 927 int xhdrs; 928 929 if (mp->m_logbufs <= 0) 930 log->l_iclog_bufs = XLOG_MAX_ICLOGS; 931 else 932 log->l_iclog_bufs = mp->m_logbufs; 933 934 /* 935 * Buffer size passed in from mount system call. 936 */ 937 if (mp->m_logbsize > 0) { 938 size = log->l_iclog_size = mp->m_logbsize; 939 log->l_iclog_size_log = 0; 940 while (size != 1) { 941 log->l_iclog_size_log++; 942 size >>= 1; 943 } 944 945 if (xfs_sb_version_haslogv2(&mp->m_sb)) { 946 /* # headers = size / 32k 947 * one header holds cycles from 32k of data 948 */ 949 950 xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE; 951 if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE) 952 xhdrs++; 953 log->l_iclog_hsize = xhdrs << BBSHIFT; 954 log->l_iclog_heads = xhdrs; 955 } else { 956 ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE); 957 log->l_iclog_hsize = BBSIZE; 958 log->l_iclog_heads = 1; 959 } 960 goto done; 961 } 962 963 /* All machines use 32kB buffers by default. */ 964 log->l_iclog_size = XLOG_BIG_RECORD_BSIZE; 965 log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT; 966 967 /* the default log size is 16k or 32k which is one header sector */ 968 log->l_iclog_hsize = BBSIZE; 969 log->l_iclog_heads = 1; 970 971 done: 972 /* are we being asked to make the sizes selected above visible? */ 973 if (mp->m_logbufs == 0) 974 mp->m_logbufs = log->l_iclog_bufs; 975 if (mp->m_logbsize == 0) 976 mp->m_logbsize = log->l_iclog_size; 977 } /* xlog_get_iclog_buffer_size */ 978 979 980 /* 981 * This routine initializes some of the log structure for a given mount point. 982 * Its primary purpose is to fill in enough, so recovery can occur. However, 983 * some other stuff may be filled in too. 984 */ 985 STATIC xlog_t * 986 xlog_alloc_log(xfs_mount_t *mp, 987 xfs_buftarg_t *log_target, 988 xfs_daddr_t blk_offset, 989 int num_bblks) 990 { 991 xlog_t *log; 992 xlog_rec_header_t *head; 993 xlog_in_core_t **iclogp; 994 xlog_in_core_t *iclog, *prev_iclog=NULL; 995 xfs_buf_t *bp; 996 int i; 997 int error = ENOMEM; 998 uint log2_size = 0; 999 1000 log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL); 1001 if (!log) { 1002 xfs_warn(mp, "Log allocation failed: No memory!"); 1003 goto out; 1004 } 1005 1006 log->l_mp = mp; 1007 log->l_targ = log_target; 1008 log->l_logsize = BBTOB(num_bblks); 1009 log->l_logBBstart = blk_offset; 1010 log->l_logBBsize = num_bblks; 1011 log->l_covered_state = XLOG_STATE_COVER_IDLE; 1012 log->l_flags |= XLOG_ACTIVE_RECOVERY; 1013 1014 log->l_prev_block = -1; 1015 /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */ 1016 xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0); 1017 xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0); 1018 log->l_curr_cycle = 1; /* 0 is bad since this is initial value */ 1019 xlog_assign_grant_head(&log->l_grant_reserve_head, 1, 0); 1020 xlog_assign_grant_head(&log->l_grant_write_head, 1, 0); 1021 INIT_LIST_HEAD(&log->l_reserveq); 1022 INIT_LIST_HEAD(&log->l_writeq); 1023 spin_lock_init(&log->l_grant_reserve_lock); 1024 spin_lock_init(&log->l_grant_write_lock); 1025 1026 error = EFSCORRUPTED; 1027 if (xfs_sb_version_hassector(&mp->m_sb)) { 1028 log2_size = mp->m_sb.sb_logsectlog; 1029 if (log2_size < BBSHIFT) { 1030 xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)", 1031 log2_size, BBSHIFT); 1032 goto out_free_log; 1033 } 1034 1035 log2_size -= BBSHIFT; 1036 if (log2_size > mp->m_sectbb_log) { 1037 xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)", 1038 log2_size, mp->m_sectbb_log); 1039 goto out_free_log; 1040 } 1041 1042 /* for larger sector sizes, must have v2 or external log */ 1043 if (log2_size && log->l_logBBstart > 0 && 1044 !xfs_sb_version_haslogv2(&mp->m_sb)) { 1045 xfs_warn(mp, 1046 "log sector size (0x%x) invalid for configuration.", 1047 log2_size); 1048 goto out_free_log; 1049 } 1050 } 1051 log->l_sectBBsize = 1 << log2_size; 1052 1053 xlog_get_iclog_buffer_size(mp, log); 1054 1055 error = ENOMEM; 1056 bp = xfs_buf_get_empty(log->l_iclog_size, mp->m_logdev_targp); 1057 if (!bp) 1058 goto out_free_log; 1059 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone); 1060 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1); 1061 ASSERT(XFS_BUF_ISBUSY(bp)); 1062 ASSERT(XFS_BUF_VALUSEMA(bp) <= 0); 1063 log->l_xbuf = bp; 1064 1065 spin_lock_init(&log->l_icloglock); 1066 init_waitqueue_head(&log->l_flush_wait); 1067 1068 /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */ 1069 ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0); 1070 1071 iclogp = &log->l_iclog; 1072 /* 1073 * The amount of memory to allocate for the iclog structure is 1074 * rather funky due to the way the structure is defined. It is 1075 * done this way so that we can use different sizes for machines 1076 * with different amounts of memory. See the definition of 1077 * xlog_in_core_t in xfs_log_priv.h for details. 1078 */ 1079 ASSERT(log->l_iclog_size >= 4096); 1080 for (i=0; i < log->l_iclog_bufs; i++) { 1081 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL); 1082 if (!*iclogp) 1083 goto out_free_iclog; 1084 1085 iclog = *iclogp; 1086 iclog->ic_prev = prev_iclog; 1087 prev_iclog = iclog; 1088 1089 bp = xfs_buf_get_uncached(mp->m_logdev_targp, 1090 log->l_iclog_size, 0); 1091 if (!bp) 1092 goto out_free_iclog; 1093 if (!XFS_BUF_CPSEMA(bp)) 1094 ASSERT(0); 1095 XFS_BUF_SET_IODONE_FUNC(bp, xlog_iodone); 1096 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)1); 1097 iclog->ic_bp = bp; 1098 iclog->ic_data = bp->b_addr; 1099 #ifdef DEBUG 1100 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header); 1101 #endif 1102 head = &iclog->ic_header; 1103 memset(head, 0, sizeof(xlog_rec_header_t)); 1104 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM); 1105 head->h_version = cpu_to_be32( 1106 xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1); 1107 head->h_size = cpu_to_be32(log->l_iclog_size); 1108 /* new fields */ 1109 head->h_fmt = cpu_to_be32(XLOG_FMT); 1110 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t)); 1111 1112 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize; 1113 iclog->ic_state = XLOG_STATE_ACTIVE; 1114 iclog->ic_log = log; 1115 atomic_set(&iclog->ic_refcnt, 0); 1116 spin_lock_init(&iclog->ic_callback_lock); 1117 iclog->ic_callback_tail = &(iclog->ic_callback); 1118 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize; 1119 1120 ASSERT(XFS_BUF_ISBUSY(iclog->ic_bp)); 1121 ASSERT(XFS_BUF_VALUSEMA(iclog->ic_bp) <= 0); 1122 init_waitqueue_head(&iclog->ic_force_wait); 1123 init_waitqueue_head(&iclog->ic_write_wait); 1124 1125 iclogp = &iclog->ic_next; 1126 } 1127 *iclogp = log->l_iclog; /* complete ring */ 1128 log->l_iclog->ic_prev = prev_iclog; /* re-write 1st prev ptr */ 1129 1130 error = xlog_cil_init(log); 1131 if (error) 1132 goto out_free_iclog; 1133 return log; 1134 1135 out_free_iclog: 1136 for (iclog = log->l_iclog; iclog; iclog = prev_iclog) { 1137 prev_iclog = iclog->ic_next; 1138 if (iclog->ic_bp) 1139 xfs_buf_free(iclog->ic_bp); 1140 kmem_free(iclog); 1141 } 1142 spinlock_destroy(&log->l_icloglock); 1143 xfs_buf_free(log->l_xbuf); 1144 out_free_log: 1145 kmem_free(log); 1146 out: 1147 return ERR_PTR(-error); 1148 } /* xlog_alloc_log */ 1149 1150 1151 /* 1152 * Write out the commit record of a transaction associated with the given 1153 * ticket. Return the lsn of the commit record. 1154 */ 1155 STATIC int 1156 xlog_commit_record( 1157 struct log *log, 1158 struct xlog_ticket *ticket, 1159 struct xlog_in_core **iclog, 1160 xfs_lsn_t *commitlsnp) 1161 { 1162 struct xfs_mount *mp = log->l_mp; 1163 int error; 1164 struct xfs_log_iovec reg = { 1165 .i_addr = NULL, 1166 .i_len = 0, 1167 .i_type = XLOG_REG_TYPE_COMMIT, 1168 }; 1169 struct xfs_log_vec vec = { 1170 .lv_niovecs = 1, 1171 .lv_iovecp = ®, 1172 }; 1173 1174 ASSERT_ALWAYS(iclog); 1175 error = xlog_write(log, &vec, ticket, commitlsnp, iclog, 1176 XLOG_COMMIT_TRANS); 1177 if (error) 1178 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR); 1179 return error; 1180 } 1181 1182 /* 1183 * Push on the buffer cache code if we ever use more than 75% of the on-disk 1184 * log space. This code pushes on the lsn which would supposedly free up 1185 * the 25% which we want to leave free. We may need to adopt a policy which 1186 * pushes on an lsn which is further along in the log once we reach the high 1187 * water mark. In this manner, we would be creating a low water mark. 1188 */ 1189 STATIC void 1190 xlog_grant_push_ail( 1191 struct log *log, 1192 int need_bytes) 1193 { 1194 xfs_lsn_t threshold_lsn = 0; 1195 xfs_lsn_t last_sync_lsn; 1196 int free_blocks; 1197 int free_bytes; 1198 int threshold_block; 1199 int threshold_cycle; 1200 int free_threshold; 1201 1202 ASSERT(BTOBB(need_bytes) < log->l_logBBsize); 1203 1204 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head); 1205 free_blocks = BTOBBT(free_bytes); 1206 1207 /* 1208 * Set the threshold for the minimum number of free blocks in the 1209 * log to the maximum of what the caller needs, one quarter of the 1210 * log, and 256 blocks. 1211 */ 1212 free_threshold = BTOBB(need_bytes); 1213 free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2)); 1214 free_threshold = MAX(free_threshold, 256); 1215 if (free_blocks >= free_threshold) 1216 return; 1217 1218 xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle, 1219 &threshold_block); 1220 threshold_block += free_threshold; 1221 if (threshold_block >= log->l_logBBsize) { 1222 threshold_block -= log->l_logBBsize; 1223 threshold_cycle += 1; 1224 } 1225 threshold_lsn = xlog_assign_lsn(threshold_cycle, 1226 threshold_block); 1227 /* 1228 * Don't pass in an lsn greater than the lsn of the last 1229 * log record known to be on disk. Use a snapshot of the last sync lsn 1230 * so that it doesn't change between the compare and the set. 1231 */ 1232 last_sync_lsn = atomic64_read(&log->l_last_sync_lsn); 1233 if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0) 1234 threshold_lsn = last_sync_lsn; 1235 1236 /* 1237 * Get the transaction layer to kick the dirty buffers out to 1238 * disk asynchronously. No point in trying to do this if 1239 * the filesystem is shutting down. 1240 */ 1241 if (!XLOG_FORCED_SHUTDOWN(log)) 1242 xfs_ail_push(log->l_ailp, threshold_lsn); 1243 } 1244 1245 /* 1246 * The bdstrat callback function for log bufs. This gives us a central 1247 * place to trap bufs in case we get hit by a log I/O error and need to 1248 * shutdown. Actually, in practice, even when we didn't get a log error, 1249 * we transition the iclogs to IOERROR state *after* flushing all existing 1250 * iclogs to disk. This is because we don't want anymore new transactions to be 1251 * started or completed afterwards. 1252 */ 1253 STATIC int 1254 xlog_bdstrat( 1255 struct xfs_buf *bp) 1256 { 1257 struct xlog_in_core *iclog; 1258 1259 iclog = XFS_BUF_FSPRIVATE(bp, xlog_in_core_t *); 1260 if (iclog->ic_state & XLOG_STATE_IOERROR) { 1261 XFS_BUF_ERROR(bp, EIO); 1262 XFS_BUF_STALE(bp); 1263 xfs_buf_ioend(bp, 0); 1264 /* 1265 * It would seem logical to return EIO here, but we rely on 1266 * the log state machine to propagate I/O errors instead of 1267 * doing it here. 1268 */ 1269 return 0; 1270 } 1271 1272 bp->b_flags |= _XBF_RUN_QUEUES; 1273 xfs_buf_iorequest(bp); 1274 return 0; 1275 } 1276 1277 /* 1278 * Flush out the in-core log (iclog) to the on-disk log in an asynchronous 1279 * fashion. Previously, we should have moved the current iclog 1280 * ptr in the log to point to the next available iclog. This allows further 1281 * write to continue while this code syncs out an iclog ready to go. 1282 * Before an in-core log can be written out, the data section must be scanned 1283 * to save away the 1st word of each BBSIZE block into the header. We replace 1284 * it with the current cycle count. Each BBSIZE block is tagged with the 1285 * cycle count because there in an implicit assumption that drives will 1286 * guarantee that entire 512 byte blocks get written at once. In other words, 1287 * we can't have part of a 512 byte block written and part not written. By 1288 * tagging each block, we will know which blocks are valid when recovering 1289 * after an unclean shutdown. 1290 * 1291 * This routine is single threaded on the iclog. No other thread can be in 1292 * this routine with the same iclog. Changing contents of iclog can there- 1293 * fore be done without grabbing the state machine lock. Updating the global 1294 * log will require grabbing the lock though. 1295 * 1296 * The entire log manager uses a logical block numbering scheme. Only 1297 * log_sync (and then only bwrite()) know about the fact that the log may 1298 * not start with block zero on a given device. The log block start offset 1299 * is added immediately before calling bwrite(). 1300 */ 1301 1302 STATIC int 1303 xlog_sync(xlog_t *log, 1304 xlog_in_core_t *iclog) 1305 { 1306 xfs_caddr_t dptr; /* pointer to byte sized element */ 1307 xfs_buf_t *bp; 1308 int i; 1309 uint count; /* byte count of bwrite */ 1310 uint count_init; /* initial count before roundup */ 1311 int roundoff; /* roundoff to BB or stripe */ 1312 int split = 0; /* split write into two regions */ 1313 int error; 1314 int v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb); 1315 1316 XFS_STATS_INC(xs_log_writes); 1317 ASSERT(atomic_read(&iclog->ic_refcnt) == 0); 1318 1319 /* Add for LR header */ 1320 count_init = log->l_iclog_hsize + iclog->ic_offset; 1321 1322 /* Round out the log write size */ 1323 if (v2 && log->l_mp->m_sb.sb_logsunit > 1) { 1324 /* we have a v2 stripe unit to use */ 1325 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init)); 1326 } else { 1327 count = BBTOB(BTOBB(count_init)); 1328 } 1329 roundoff = count - count_init; 1330 ASSERT(roundoff >= 0); 1331 ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 && 1332 roundoff < log->l_mp->m_sb.sb_logsunit) 1333 || 1334 (log->l_mp->m_sb.sb_logsunit <= 1 && 1335 roundoff < BBTOB(1))); 1336 1337 /* move grant heads by roundoff in sync */ 1338 xlog_grant_add_space(log, &log->l_grant_reserve_head, roundoff); 1339 xlog_grant_add_space(log, &log->l_grant_write_head, roundoff); 1340 1341 /* put cycle number in every block */ 1342 xlog_pack_data(log, iclog, roundoff); 1343 1344 /* real byte length */ 1345 if (v2) { 1346 iclog->ic_header.h_len = 1347 cpu_to_be32(iclog->ic_offset + roundoff); 1348 } else { 1349 iclog->ic_header.h_len = 1350 cpu_to_be32(iclog->ic_offset); 1351 } 1352 1353 bp = iclog->ic_bp; 1354 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == (unsigned long)1); 1355 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2); 1356 XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn))); 1357 1358 XFS_STATS_ADD(xs_log_blocks, BTOBB(count)); 1359 1360 /* Do we need to split this write into 2 parts? */ 1361 if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) { 1362 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp))); 1363 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)); 1364 iclog->ic_bwritecnt = 2; /* split into 2 writes */ 1365 } else { 1366 iclog->ic_bwritecnt = 1; 1367 } 1368 XFS_BUF_SET_COUNT(bp, count); 1369 XFS_BUF_SET_FSPRIVATE(bp, iclog); /* save for later */ 1370 XFS_BUF_ZEROFLAGS(bp); 1371 XFS_BUF_BUSY(bp); 1372 XFS_BUF_ASYNC(bp); 1373 bp->b_flags |= XBF_LOG_BUFFER; 1374 1375 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) { 1376 /* 1377 * If we have an external log device, flush the data device 1378 * before flushing the log to make sure all meta data 1379 * written back from the AIL actually made it to disk 1380 * before writing out the new log tail LSN in the log buffer. 1381 */ 1382 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp) 1383 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp); 1384 XFS_BUF_ORDERED(bp); 1385 } 1386 1387 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1); 1388 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize); 1389 1390 xlog_verify_iclog(log, iclog, count, B_TRUE); 1391 1392 /* account for log which doesn't start at block #0 */ 1393 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart); 1394 /* 1395 * Don't call xfs_bwrite here. We do log-syncs even when the filesystem 1396 * is shutting down. 1397 */ 1398 XFS_BUF_WRITE(bp); 1399 1400 if ((error = xlog_bdstrat(bp))) { 1401 xfs_ioerror_alert("xlog_sync", log->l_mp, bp, 1402 XFS_BUF_ADDR(bp)); 1403 return error; 1404 } 1405 if (split) { 1406 bp = iclog->ic_log->l_xbuf; 1407 ASSERT(XFS_BUF_FSPRIVATE2(bp, unsigned long) == 1408 (unsigned long)1); 1409 XFS_BUF_SET_FSPRIVATE2(bp, (unsigned long)2); 1410 XFS_BUF_SET_ADDR(bp, 0); /* logical 0 */ 1411 XFS_BUF_SET_PTR(bp, (xfs_caddr_t)((__psint_t)&(iclog->ic_header)+ 1412 (__psint_t)count), split); 1413 XFS_BUF_SET_FSPRIVATE(bp, iclog); 1414 XFS_BUF_ZEROFLAGS(bp); 1415 XFS_BUF_BUSY(bp); 1416 XFS_BUF_ASYNC(bp); 1417 bp->b_flags |= XBF_LOG_BUFFER; 1418 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) 1419 XFS_BUF_ORDERED(bp); 1420 dptr = XFS_BUF_PTR(bp); 1421 /* 1422 * Bump the cycle numbers at the start of each block 1423 * since this part of the buffer is at the start of 1424 * a new cycle. Watch out for the header magic number 1425 * case, though. 1426 */ 1427 for (i = 0; i < split; i += BBSIZE) { 1428 be32_add_cpu((__be32 *)dptr, 1); 1429 if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM) 1430 be32_add_cpu((__be32 *)dptr, 1); 1431 dptr += BBSIZE; 1432 } 1433 1434 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1); 1435 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize); 1436 1437 /* account for internal log which doesn't start at block #0 */ 1438 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart); 1439 XFS_BUF_WRITE(bp); 1440 if ((error = xlog_bdstrat(bp))) { 1441 xfs_ioerror_alert("xlog_sync (split)", log->l_mp, 1442 bp, XFS_BUF_ADDR(bp)); 1443 return error; 1444 } 1445 } 1446 return 0; 1447 } /* xlog_sync */ 1448 1449 1450 /* 1451 * Deallocate a log structure 1452 */ 1453 STATIC void 1454 xlog_dealloc_log(xlog_t *log) 1455 { 1456 xlog_in_core_t *iclog, *next_iclog; 1457 int i; 1458 1459 xlog_cil_destroy(log); 1460 1461 /* 1462 * always need to ensure that the extra buffer does not point to memory 1463 * owned by another log buffer before we free it. 1464 */ 1465 xfs_buf_set_empty(log->l_xbuf, log->l_iclog_size); 1466 xfs_buf_free(log->l_xbuf); 1467 1468 iclog = log->l_iclog; 1469 for (i=0; i<log->l_iclog_bufs; i++) { 1470 xfs_buf_free(iclog->ic_bp); 1471 next_iclog = iclog->ic_next; 1472 kmem_free(iclog); 1473 iclog = next_iclog; 1474 } 1475 spinlock_destroy(&log->l_icloglock); 1476 1477 log->l_mp->m_log = NULL; 1478 kmem_free(log); 1479 } /* xlog_dealloc_log */ 1480 1481 /* 1482 * Update counters atomically now that memcpy is done. 1483 */ 1484 /* ARGSUSED */ 1485 static inline void 1486 xlog_state_finish_copy(xlog_t *log, 1487 xlog_in_core_t *iclog, 1488 int record_cnt, 1489 int copy_bytes) 1490 { 1491 spin_lock(&log->l_icloglock); 1492 1493 be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt); 1494 iclog->ic_offset += copy_bytes; 1495 1496 spin_unlock(&log->l_icloglock); 1497 } /* xlog_state_finish_copy */ 1498 1499 1500 1501 1502 /* 1503 * print out info relating to regions written which consume 1504 * the reservation 1505 */ 1506 void 1507 xlog_print_tic_res( 1508 struct xfs_mount *mp, 1509 struct xlog_ticket *ticket) 1510 { 1511 uint i; 1512 uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t); 1513 1514 /* match with XLOG_REG_TYPE_* in xfs_log.h */ 1515 static char *res_type_str[XLOG_REG_TYPE_MAX] = { 1516 "bformat", 1517 "bchunk", 1518 "efi_format", 1519 "efd_format", 1520 "iformat", 1521 "icore", 1522 "iext", 1523 "ibroot", 1524 "ilocal", 1525 "iattr_ext", 1526 "iattr_broot", 1527 "iattr_local", 1528 "qformat", 1529 "dquot", 1530 "quotaoff", 1531 "LR header", 1532 "unmount", 1533 "commit", 1534 "trans header" 1535 }; 1536 static char *trans_type_str[XFS_TRANS_TYPE_MAX] = { 1537 "SETATTR_NOT_SIZE", 1538 "SETATTR_SIZE", 1539 "INACTIVE", 1540 "CREATE", 1541 "CREATE_TRUNC", 1542 "TRUNCATE_FILE", 1543 "REMOVE", 1544 "LINK", 1545 "RENAME", 1546 "MKDIR", 1547 "RMDIR", 1548 "SYMLINK", 1549 "SET_DMATTRS", 1550 "GROWFS", 1551 "STRAT_WRITE", 1552 "DIOSTRAT", 1553 "WRITE_SYNC", 1554 "WRITEID", 1555 "ADDAFORK", 1556 "ATTRINVAL", 1557 "ATRUNCATE", 1558 "ATTR_SET", 1559 "ATTR_RM", 1560 "ATTR_FLAG", 1561 "CLEAR_AGI_BUCKET", 1562 "QM_SBCHANGE", 1563 "DUMMY1", 1564 "DUMMY2", 1565 "QM_QUOTAOFF", 1566 "QM_DQALLOC", 1567 "QM_SETQLIM", 1568 "QM_DQCLUSTER", 1569 "QM_QINOCREATE", 1570 "QM_QUOTAOFF_END", 1571 "SB_UNIT", 1572 "FSYNC_TS", 1573 "GROWFSRT_ALLOC", 1574 "GROWFSRT_ZERO", 1575 "GROWFSRT_FREE", 1576 "SWAPEXT" 1577 }; 1578 1579 xfs_warn(mp, 1580 "xfs_log_write: reservation summary:\n" 1581 " trans type = %s (%u)\n" 1582 " unit res = %d bytes\n" 1583 " current res = %d bytes\n" 1584 " total reg = %u bytes (o/flow = %u bytes)\n" 1585 " ophdrs = %u (ophdr space = %u bytes)\n" 1586 " ophdr + reg = %u bytes\n" 1587 " num regions = %u\n", 1588 ((ticket->t_trans_type <= 0 || 1589 ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ? 1590 "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]), 1591 ticket->t_trans_type, 1592 ticket->t_unit_res, 1593 ticket->t_curr_res, 1594 ticket->t_res_arr_sum, ticket->t_res_o_flow, 1595 ticket->t_res_num_ophdrs, ophdr_spc, 1596 ticket->t_res_arr_sum + 1597 ticket->t_res_o_flow + ophdr_spc, 1598 ticket->t_res_num); 1599 1600 for (i = 0; i < ticket->t_res_num; i++) { 1601 uint r_type = ticket->t_res_arr[i].r_type; 1602 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i, 1603 ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ? 1604 "bad-rtype" : res_type_str[r_type-1]), 1605 ticket->t_res_arr[i].r_len); 1606 } 1607 1608 xfs_alert_tag(mp, XFS_PTAG_LOGRES, 1609 "xfs_log_write: reservation ran out. Need to up reservation"); 1610 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE); 1611 } 1612 1613 /* 1614 * Calculate the potential space needed by the log vector. Each region gets 1615 * its own xlog_op_header_t and may need to be double word aligned. 1616 */ 1617 static int 1618 xlog_write_calc_vec_length( 1619 struct xlog_ticket *ticket, 1620 struct xfs_log_vec *log_vector) 1621 { 1622 struct xfs_log_vec *lv; 1623 int headers = 0; 1624 int len = 0; 1625 int i; 1626 1627 /* acct for start rec of xact */ 1628 if (ticket->t_flags & XLOG_TIC_INITED) 1629 headers++; 1630 1631 for (lv = log_vector; lv; lv = lv->lv_next) { 1632 headers += lv->lv_niovecs; 1633 1634 for (i = 0; i < lv->lv_niovecs; i++) { 1635 struct xfs_log_iovec *vecp = &lv->lv_iovecp[i]; 1636 1637 len += vecp->i_len; 1638 xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type); 1639 } 1640 } 1641 1642 ticket->t_res_num_ophdrs += headers; 1643 len += headers * sizeof(struct xlog_op_header); 1644 1645 return len; 1646 } 1647 1648 /* 1649 * If first write for transaction, insert start record We can't be trying to 1650 * commit if we are inited. We can't have any "partial_copy" if we are inited. 1651 */ 1652 static int 1653 xlog_write_start_rec( 1654 struct xlog_op_header *ophdr, 1655 struct xlog_ticket *ticket) 1656 { 1657 if (!(ticket->t_flags & XLOG_TIC_INITED)) 1658 return 0; 1659 1660 ophdr->oh_tid = cpu_to_be32(ticket->t_tid); 1661 ophdr->oh_clientid = ticket->t_clientid; 1662 ophdr->oh_len = 0; 1663 ophdr->oh_flags = XLOG_START_TRANS; 1664 ophdr->oh_res2 = 0; 1665 1666 ticket->t_flags &= ~XLOG_TIC_INITED; 1667 1668 return sizeof(struct xlog_op_header); 1669 } 1670 1671 static xlog_op_header_t * 1672 xlog_write_setup_ophdr( 1673 struct log *log, 1674 struct xlog_op_header *ophdr, 1675 struct xlog_ticket *ticket, 1676 uint flags) 1677 { 1678 ophdr->oh_tid = cpu_to_be32(ticket->t_tid); 1679 ophdr->oh_clientid = ticket->t_clientid; 1680 ophdr->oh_res2 = 0; 1681 1682 /* are we copying a commit or unmount record? */ 1683 ophdr->oh_flags = flags; 1684 1685 /* 1686 * We've seen logs corrupted with bad transaction client ids. This 1687 * makes sure that XFS doesn't generate them on. Turn this into an EIO 1688 * and shut down the filesystem. 1689 */ 1690 switch (ophdr->oh_clientid) { 1691 case XFS_TRANSACTION: 1692 case XFS_VOLUME: 1693 case XFS_LOG: 1694 break; 1695 default: 1696 xfs_warn(log->l_mp, 1697 "Bad XFS transaction clientid 0x%x in ticket 0x%p", 1698 ophdr->oh_clientid, ticket); 1699 return NULL; 1700 } 1701 1702 return ophdr; 1703 } 1704 1705 /* 1706 * Set up the parameters of the region copy into the log. This has 1707 * to handle region write split across multiple log buffers - this 1708 * state is kept external to this function so that this code can 1709 * can be written in an obvious, self documenting manner. 1710 */ 1711 static int 1712 xlog_write_setup_copy( 1713 struct xlog_ticket *ticket, 1714 struct xlog_op_header *ophdr, 1715 int space_available, 1716 int space_required, 1717 int *copy_off, 1718 int *copy_len, 1719 int *last_was_partial_copy, 1720 int *bytes_consumed) 1721 { 1722 int still_to_copy; 1723 1724 still_to_copy = space_required - *bytes_consumed; 1725 *copy_off = *bytes_consumed; 1726 1727 if (still_to_copy <= space_available) { 1728 /* write of region completes here */ 1729 *copy_len = still_to_copy; 1730 ophdr->oh_len = cpu_to_be32(*copy_len); 1731 if (*last_was_partial_copy) 1732 ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS); 1733 *last_was_partial_copy = 0; 1734 *bytes_consumed = 0; 1735 return 0; 1736 } 1737 1738 /* partial write of region, needs extra log op header reservation */ 1739 *copy_len = space_available; 1740 ophdr->oh_len = cpu_to_be32(*copy_len); 1741 ophdr->oh_flags |= XLOG_CONTINUE_TRANS; 1742 if (*last_was_partial_copy) 1743 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS; 1744 *bytes_consumed += *copy_len; 1745 (*last_was_partial_copy)++; 1746 1747 /* account for new log op header */ 1748 ticket->t_curr_res -= sizeof(struct xlog_op_header); 1749 ticket->t_res_num_ophdrs++; 1750 1751 return sizeof(struct xlog_op_header); 1752 } 1753 1754 static int 1755 xlog_write_copy_finish( 1756 struct log *log, 1757 struct xlog_in_core *iclog, 1758 uint flags, 1759 int *record_cnt, 1760 int *data_cnt, 1761 int *partial_copy, 1762 int *partial_copy_len, 1763 int log_offset, 1764 struct xlog_in_core **commit_iclog) 1765 { 1766 if (*partial_copy) { 1767 /* 1768 * This iclog has already been marked WANT_SYNC by 1769 * xlog_state_get_iclog_space. 1770 */ 1771 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt); 1772 *record_cnt = 0; 1773 *data_cnt = 0; 1774 return xlog_state_release_iclog(log, iclog); 1775 } 1776 1777 *partial_copy = 0; 1778 *partial_copy_len = 0; 1779 1780 if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) { 1781 /* no more space in this iclog - push it. */ 1782 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt); 1783 *record_cnt = 0; 1784 *data_cnt = 0; 1785 1786 spin_lock(&log->l_icloglock); 1787 xlog_state_want_sync(log, iclog); 1788 spin_unlock(&log->l_icloglock); 1789 1790 if (!commit_iclog) 1791 return xlog_state_release_iclog(log, iclog); 1792 ASSERT(flags & XLOG_COMMIT_TRANS); 1793 *commit_iclog = iclog; 1794 } 1795 1796 return 0; 1797 } 1798 1799 /* 1800 * Write some region out to in-core log 1801 * 1802 * This will be called when writing externally provided regions or when 1803 * writing out a commit record for a given transaction. 1804 * 1805 * General algorithm: 1806 * 1. Find total length of this write. This may include adding to the 1807 * lengths passed in. 1808 * 2. Check whether we violate the tickets reservation. 1809 * 3. While writing to this iclog 1810 * A. Reserve as much space in this iclog as can get 1811 * B. If this is first write, save away start lsn 1812 * C. While writing this region: 1813 * 1. If first write of transaction, write start record 1814 * 2. Write log operation header (header per region) 1815 * 3. Find out if we can fit entire region into this iclog 1816 * 4. Potentially, verify destination memcpy ptr 1817 * 5. Memcpy (partial) region 1818 * 6. If partial copy, release iclog; otherwise, continue 1819 * copying more regions into current iclog 1820 * 4. Mark want sync bit (in simulation mode) 1821 * 5. Release iclog for potential flush to on-disk log. 1822 * 1823 * ERRORS: 1824 * 1. Panic if reservation is overrun. This should never happen since 1825 * reservation amounts are generated internal to the filesystem. 1826 * NOTES: 1827 * 1. Tickets are single threaded data structures. 1828 * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the 1829 * syncing routine. When a single log_write region needs to span 1830 * multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set 1831 * on all log operation writes which don't contain the end of the 1832 * region. The XLOG_END_TRANS bit is used for the in-core log 1833 * operation which contains the end of the continued log_write region. 1834 * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog, 1835 * we don't really know exactly how much space will be used. As a result, 1836 * we don't update ic_offset until the end when we know exactly how many 1837 * bytes have been written out. 1838 */ 1839 int 1840 xlog_write( 1841 struct log *log, 1842 struct xfs_log_vec *log_vector, 1843 struct xlog_ticket *ticket, 1844 xfs_lsn_t *start_lsn, 1845 struct xlog_in_core **commit_iclog, 1846 uint flags) 1847 { 1848 struct xlog_in_core *iclog = NULL; 1849 struct xfs_log_iovec *vecp; 1850 struct xfs_log_vec *lv; 1851 int len; 1852 int index; 1853 int partial_copy = 0; 1854 int partial_copy_len = 0; 1855 int contwr = 0; 1856 int record_cnt = 0; 1857 int data_cnt = 0; 1858 int error; 1859 1860 *start_lsn = 0; 1861 1862 len = xlog_write_calc_vec_length(ticket, log_vector); 1863 if (log->l_cilp) { 1864 /* 1865 * Region headers and bytes are already accounted for. 1866 * We only need to take into account start records and 1867 * split regions in this function. 1868 */ 1869 if (ticket->t_flags & XLOG_TIC_INITED) 1870 ticket->t_curr_res -= sizeof(xlog_op_header_t); 1871 1872 /* 1873 * Commit record headers need to be accounted for. These 1874 * come in as separate writes so are easy to detect. 1875 */ 1876 if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS)) 1877 ticket->t_curr_res -= sizeof(xlog_op_header_t); 1878 } else 1879 ticket->t_curr_res -= len; 1880 1881 if (ticket->t_curr_res < 0) 1882 xlog_print_tic_res(log->l_mp, ticket); 1883 1884 index = 0; 1885 lv = log_vector; 1886 vecp = lv->lv_iovecp; 1887 while (lv && index < lv->lv_niovecs) { 1888 void *ptr; 1889 int log_offset; 1890 1891 error = xlog_state_get_iclog_space(log, len, &iclog, ticket, 1892 &contwr, &log_offset); 1893 if (error) 1894 return error; 1895 1896 ASSERT(log_offset <= iclog->ic_size - 1); 1897 ptr = iclog->ic_datap + log_offset; 1898 1899 /* start_lsn is the first lsn written to. That's all we need. */ 1900 if (!*start_lsn) 1901 *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn); 1902 1903 /* 1904 * This loop writes out as many regions as can fit in the amount 1905 * of space which was allocated by xlog_state_get_iclog_space(). 1906 */ 1907 while (lv && index < lv->lv_niovecs) { 1908 struct xfs_log_iovec *reg = &vecp[index]; 1909 struct xlog_op_header *ophdr; 1910 int start_rec_copy; 1911 int copy_len; 1912 int copy_off; 1913 1914 ASSERT(reg->i_len % sizeof(__int32_t) == 0); 1915 ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0); 1916 1917 start_rec_copy = xlog_write_start_rec(ptr, ticket); 1918 if (start_rec_copy) { 1919 record_cnt++; 1920 xlog_write_adv_cnt(&ptr, &len, &log_offset, 1921 start_rec_copy); 1922 } 1923 1924 ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags); 1925 if (!ophdr) 1926 return XFS_ERROR(EIO); 1927 1928 xlog_write_adv_cnt(&ptr, &len, &log_offset, 1929 sizeof(struct xlog_op_header)); 1930 1931 len += xlog_write_setup_copy(ticket, ophdr, 1932 iclog->ic_size-log_offset, 1933 reg->i_len, 1934 ©_off, ©_len, 1935 &partial_copy, 1936 &partial_copy_len); 1937 xlog_verify_dest_ptr(log, ptr); 1938 1939 /* copy region */ 1940 ASSERT(copy_len >= 0); 1941 memcpy(ptr, reg->i_addr + copy_off, copy_len); 1942 xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len); 1943 1944 copy_len += start_rec_copy + sizeof(xlog_op_header_t); 1945 record_cnt++; 1946 data_cnt += contwr ? copy_len : 0; 1947 1948 error = xlog_write_copy_finish(log, iclog, flags, 1949 &record_cnt, &data_cnt, 1950 &partial_copy, 1951 &partial_copy_len, 1952 log_offset, 1953 commit_iclog); 1954 if (error) 1955 return error; 1956 1957 /* 1958 * if we had a partial copy, we need to get more iclog 1959 * space but we don't want to increment the region 1960 * index because there is still more is this region to 1961 * write. 1962 * 1963 * If we completed writing this region, and we flushed 1964 * the iclog (indicated by resetting of the record 1965 * count), then we also need to get more log space. If 1966 * this was the last record, though, we are done and 1967 * can just return. 1968 */ 1969 if (partial_copy) 1970 break; 1971 1972 if (++index == lv->lv_niovecs) { 1973 lv = lv->lv_next; 1974 index = 0; 1975 if (lv) 1976 vecp = lv->lv_iovecp; 1977 } 1978 if (record_cnt == 0) { 1979 if (!lv) 1980 return 0; 1981 break; 1982 } 1983 } 1984 } 1985 1986 ASSERT(len == 0); 1987 1988 xlog_state_finish_copy(log, iclog, record_cnt, data_cnt); 1989 if (!commit_iclog) 1990 return xlog_state_release_iclog(log, iclog); 1991 1992 ASSERT(flags & XLOG_COMMIT_TRANS); 1993 *commit_iclog = iclog; 1994 return 0; 1995 } 1996 1997 1998 /***************************************************************************** 1999 * 2000 * State Machine functions 2001 * 2002 ***************************************************************************** 2003 */ 2004 2005 /* Clean iclogs starting from the head. This ordering must be 2006 * maintained, so an iclog doesn't become ACTIVE beyond one that 2007 * is SYNCING. This is also required to maintain the notion that we use 2008 * a ordered wait queue to hold off would be writers to the log when every 2009 * iclog is trying to sync to disk. 2010 * 2011 * State Change: DIRTY -> ACTIVE 2012 */ 2013 STATIC void 2014 xlog_state_clean_log(xlog_t *log) 2015 { 2016 xlog_in_core_t *iclog; 2017 int changed = 0; 2018 2019 iclog = log->l_iclog; 2020 do { 2021 if (iclog->ic_state == XLOG_STATE_DIRTY) { 2022 iclog->ic_state = XLOG_STATE_ACTIVE; 2023 iclog->ic_offset = 0; 2024 ASSERT(iclog->ic_callback == NULL); 2025 /* 2026 * If the number of ops in this iclog indicate it just 2027 * contains the dummy transaction, we can 2028 * change state into IDLE (the second time around). 2029 * Otherwise we should change the state into 2030 * NEED a dummy. 2031 * We don't need to cover the dummy. 2032 */ 2033 if (!changed && 2034 (be32_to_cpu(iclog->ic_header.h_num_logops) == 2035 XLOG_COVER_OPS)) { 2036 changed = 1; 2037 } else { 2038 /* 2039 * We have two dirty iclogs so start over 2040 * This could also be num of ops indicates 2041 * this is not the dummy going out. 2042 */ 2043 changed = 2; 2044 } 2045 iclog->ic_header.h_num_logops = 0; 2046 memset(iclog->ic_header.h_cycle_data, 0, 2047 sizeof(iclog->ic_header.h_cycle_data)); 2048 iclog->ic_header.h_lsn = 0; 2049 } else if (iclog->ic_state == XLOG_STATE_ACTIVE) 2050 /* do nothing */; 2051 else 2052 break; /* stop cleaning */ 2053 iclog = iclog->ic_next; 2054 } while (iclog != log->l_iclog); 2055 2056 /* log is locked when we are called */ 2057 /* 2058 * Change state for the dummy log recording. 2059 * We usually go to NEED. But we go to NEED2 if the changed indicates 2060 * we are done writing the dummy record. 2061 * If we are done with the second dummy recored (DONE2), then 2062 * we go to IDLE. 2063 */ 2064 if (changed) { 2065 switch (log->l_covered_state) { 2066 case XLOG_STATE_COVER_IDLE: 2067 case XLOG_STATE_COVER_NEED: 2068 case XLOG_STATE_COVER_NEED2: 2069 log->l_covered_state = XLOG_STATE_COVER_NEED; 2070 break; 2071 2072 case XLOG_STATE_COVER_DONE: 2073 if (changed == 1) 2074 log->l_covered_state = XLOG_STATE_COVER_NEED2; 2075 else 2076 log->l_covered_state = XLOG_STATE_COVER_NEED; 2077 break; 2078 2079 case XLOG_STATE_COVER_DONE2: 2080 if (changed == 1) 2081 log->l_covered_state = XLOG_STATE_COVER_IDLE; 2082 else 2083 log->l_covered_state = XLOG_STATE_COVER_NEED; 2084 break; 2085 2086 default: 2087 ASSERT(0); 2088 } 2089 } 2090 } /* xlog_state_clean_log */ 2091 2092 STATIC xfs_lsn_t 2093 xlog_get_lowest_lsn( 2094 xlog_t *log) 2095 { 2096 xlog_in_core_t *lsn_log; 2097 xfs_lsn_t lowest_lsn, lsn; 2098 2099 lsn_log = log->l_iclog; 2100 lowest_lsn = 0; 2101 do { 2102 if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) { 2103 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn); 2104 if ((lsn && !lowest_lsn) || 2105 (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) { 2106 lowest_lsn = lsn; 2107 } 2108 } 2109 lsn_log = lsn_log->ic_next; 2110 } while (lsn_log != log->l_iclog); 2111 return lowest_lsn; 2112 } 2113 2114 2115 STATIC void 2116 xlog_state_do_callback( 2117 xlog_t *log, 2118 int aborted, 2119 xlog_in_core_t *ciclog) 2120 { 2121 xlog_in_core_t *iclog; 2122 xlog_in_core_t *first_iclog; /* used to know when we've 2123 * processed all iclogs once */ 2124 xfs_log_callback_t *cb, *cb_next; 2125 int flushcnt = 0; 2126 xfs_lsn_t lowest_lsn; 2127 int ioerrors; /* counter: iclogs with errors */ 2128 int loopdidcallbacks; /* flag: inner loop did callbacks*/ 2129 int funcdidcallbacks; /* flag: function did callbacks */ 2130 int repeats; /* for issuing console warnings if 2131 * looping too many times */ 2132 int wake = 0; 2133 2134 spin_lock(&log->l_icloglock); 2135 first_iclog = iclog = log->l_iclog; 2136 ioerrors = 0; 2137 funcdidcallbacks = 0; 2138 repeats = 0; 2139 2140 do { 2141 /* 2142 * Scan all iclogs starting with the one pointed to by the 2143 * log. Reset this starting point each time the log is 2144 * unlocked (during callbacks). 2145 * 2146 * Keep looping through iclogs until one full pass is made 2147 * without running any callbacks. 2148 */ 2149 first_iclog = log->l_iclog; 2150 iclog = log->l_iclog; 2151 loopdidcallbacks = 0; 2152 repeats++; 2153 2154 do { 2155 2156 /* skip all iclogs in the ACTIVE & DIRTY states */ 2157 if (iclog->ic_state & 2158 (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) { 2159 iclog = iclog->ic_next; 2160 continue; 2161 } 2162 2163 /* 2164 * Between marking a filesystem SHUTDOWN and stopping 2165 * the log, we do flush all iclogs to disk (if there 2166 * wasn't a log I/O error). So, we do want things to 2167 * go smoothly in case of just a SHUTDOWN w/o a 2168 * LOG_IO_ERROR. 2169 */ 2170 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) { 2171 /* 2172 * Can only perform callbacks in order. Since 2173 * this iclog is not in the DONE_SYNC/ 2174 * DO_CALLBACK state, we skip the rest and 2175 * just try to clean up. If we set our iclog 2176 * to DO_CALLBACK, we will not process it when 2177 * we retry since a previous iclog is in the 2178 * CALLBACK and the state cannot change since 2179 * we are holding the l_icloglock. 2180 */ 2181 if (!(iclog->ic_state & 2182 (XLOG_STATE_DONE_SYNC | 2183 XLOG_STATE_DO_CALLBACK))) { 2184 if (ciclog && (ciclog->ic_state == 2185 XLOG_STATE_DONE_SYNC)) { 2186 ciclog->ic_state = XLOG_STATE_DO_CALLBACK; 2187 } 2188 break; 2189 } 2190 /* 2191 * We now have an iclog that is in either the 2192 * DO_CALLBACK or DONE_SYNC states. The other 2193 * states (WANT_SYNC, SYNCING, or CALLBACK were 2194 * caught by the above if and are going to 2195 * clean (i.e. we aren't doing their callbacks) 2196 * see the above if. 2197 */ 2198 2199 /* 2200 * We will do one more check here to see if we 2201 * have chased our tail around. 2202 */ 2203 2204 lowest_lsn = xlog_get_lowest_lsn(log); 2205 if (lowest_lsn && 2206 XFS_LSN_CMP(lowest_lsn, 2207 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) { 2208 iclog = iclog->ic_next; 2209 continue; /* Leave this iclog for 2210 * another thread */ 2211 } 2212 2213 iclog->ic_state = XLOG_STATE_CALLBACK; 2214 2215 2216 /* 2217 * update the last_sync_lsn before we drop the 2218 * icloglock to ensure we are the only one that 2219 * can update it. 2220 */ 2221 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn), 2222 be64_to_cpu(iclog->ic_header.h_lsn)) <= 0); 2223 atomic64_set(&log->l_last_sync_lsn, 2224 be64_to_cpu(iclog->ic_header.h_lsn)); 2225 2226 } else 2227 ioerrors++; 2228 2229 spin_unlock(&log->l_icloglock); 2230 2231 /* 2232 * Keep processing entries in the callback list until 2233 * we come around and it is empty. We need to 2234 * atomically see that the list is empty and change the 2235 * state to DIRTY so that we don't miss any more 2236 * callbacks being added. 2237 */ 2238 spin_lock(&iclog->ic_callback_lock); 2239 cb = iclog->ic_callback; 2240 while (cb) { 2241 iclog->ic_callback_tail = &(iclog->ic_callback); 2242 iclog->ic_callback = NULL; 2243 spin_unlock(&iclog->ic_callback_lock); 2244 2245 /* perform callbacks in the order given */ 2246 for (; cb; cb = cb_next) { 2247 cb_next = cb->cb_next; 2248 cb->cb_func(cb->cb_arg, aborted); 2249 } 2250 spin_lock(&iclog->ic_callback_lock); 2251 cb = iclog->ic_callback; 2252 } 2253 2254 loopdidcallbacks++; 2255 funcdidcallbacks++; 2256 2257 spin_lock(&log->l_icloglock); 2258 ASSERT(iclog->ic_callback == NULL); 2259 spin_unlock(&iclog->ic_callback_lock); 2260 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) 2261 iclog->ic_state = XLOG_STATE_DIRTY; 2262 2263 /* 2264 * Transition from DIRTY to ACTIVE if applicable. 2265 * NOP if STATE_IOERROR. 2266 */ 2267 xlog_state_clean_log(log); 2268 2269 /* wake up threads waiting in xfs_log_force() */ 2270 wake_up_all(&iclog->ic_force_wait); 2271 2272 iclog = iclog->ic_next; 2273 } while (first_iclog != iclog); 2274 2275 if (repeats > 5000) { 2276 flushcnt += repeats; 2277 repeats = 0; 2278 xfs_warn(log->l_mp, 2279 "%s: possible infinite loop (%d iterations)", 2280 __func__, flushcnt); 2281 } 2282 } while (!ioerrors && loopdidcallbacks); 2283 2284 /* 2285 * make one last gasp attempt to see if iclogs are being left in 2286 * limbo.. 2287 */ 2288 #ifdef DEBUG 2289 if (funcdidcallbacks) { 2290 first_iclog = iclog = log->l_iclog; 2291 do { 2292 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK); 2293 /* 2294 * Terminate the loop if iclogs are found in states 2295 * which will cause other threads to clean up iclogs. 2296 * 2297 * SYNCING - i/o completion will go through logs 2298 * DONE_SYNC - interrupt thread should be waiting for 2299 * l_icloglock 2300 * IOERROR - give up hope all ye who enter here 2301 */ 2302 if (iclog->ic_state == XLOG_STATE_WANT_SYNC || 2303 iclog->ic_state == XLOG_STATE_SYNCING || 2304 iclog->ic_state == XLOG_STATE_DONE_SYNC || 2305 iclog->ic_state == XLOG_STATE_IOERROR ) 2306 break; 2307 iclog = iclog->ic_next; 2308 } while (first_iclog != iclog); 2309 } 2310 #endif 2311 2312 if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR)) 2313 wake = 1; 2314 spin_unlock(&log->l_icloglock); 2315 2316 if (wake) 2317 wake_up_all(&log->l_flush_wait); 2318 } 2319 2320 2321 /* 2322 * Finish transitioning this iclog to the dirty state. 2323 * 2324 * Make sure that we completely execute this routine only when this is 2325 * the last call to the iclog. There is a good chance that iclog flushes, 2326 * when we reach the end of the physical log, get turned into 2 separate 2327 * calls to bwrite. Hence, one iclog flush could generate two calls to this 2328 * routine. By using the reference count bwritecnt, we guarantee that only 2329 * the second completion goes through. 2330 * 2331 * Callbacks could take time, so they are done outside the scope of the 2332 * global state machine log lock. 2333 */ 2334 STATIC void 2335 xlog_state_done_syncing( 2336 xlog_in_core_t *iclog, 2337 int aborted) 2338 { 2339 xlog_t *log = iclog->ic_log; 2340 2341 spin_lock(&log->l_icloglock); 2342 2343 ASSERT(iclog->ic_state == XLOG_STATE_SYNCING || 2344 iclog->ic_state == XLOG_STATE_IOERROR); 2345 ASSERT(atomic_read(&iclog->ic_refcnt) == 0); 2346 ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2); 2347 2348 2349 /* 2350 * If we got an error, either on the first buffer, or in the case of 2351 * split log writes, on the second, we mark ALL iclogs STATE_IOERROR, 2352 * and none should ever be attempted to be written to disk 2353 * again. 2354 */ 2355 if (iclog->ic_state != XLOG_STATE_IOERROR) { 2356 if (--iclog->ic_bwritecnt == 1) { 2357 spin_unlock(&log->l_icloglock); 2358 return; 2359 } 2360 iclog->ic_state = XLOG_STATE_DONE_SYNC; 2361 } 2362 2363 /* 2364 * Someone could be sleeping prior to writing out the next 2365 * iclog buffer, we wake them all, one will get to do the 2366 * I/O, the others get to wait for the result. 2367 */ 2368 wake_up_all(&iclog->ic_write_wait); 2369 spin_unlock(&log->l_icloglock); 2370 xlog_state_do_callback(log, aborted, iclog); /* also cleans log */ 2371 } /* xlog_state_done_syncing */ 2372 2373 2374 /* 2375 * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must 2376 * sleep. We wait on the flush queue on the head iclog as that should be 2377 * the first iclog to complete flushing. Hence if all iclogs are syncing, 2378 * we will wait here and all new writes will sleep until a sync completes. 2379 * 2380 * The in-core logs are used in a circular fashion. They are not used 2381 * out-of-order even when an iclog past the head is free. 2382 * 2383 * return: 2384 * * log_offset where xlog_write() can start writing into the in-core 2385 * log's data space. 2386 * * in-core log pointer to which xlog_write() should write. 2387 * * boolean indicating this is a continued write to an in-core log. 2388 * If this is the last write, then the in-core log's offset field 2389 * needs to be incremented, depending on the amount of data which 2390 * is copied. 2391 */ 2392 STATIC int 2393 xlog_state_get_iclog_space(xlog_t *log, 2394 int len, 2395 xlog_in_core_t **iclogp, 2396 xlog_ticket_t *ticket, 2397 int *continued_write, 2398 int *logoffsetp) 2399 { 2400 int log_offset; 2401 xlog_rec_header_t *head; 2402 xlog_in_core_t *iclog; 2403 int error; 2404 2405 restart: 2406 spin_lock(&log->l_icloglock); 2407 if (XLOG_FORCED_SHUTDOWN(log)) { 2408 spin_unlock(&log->l_icloglock); 2409 return XFS_ERROR(EIO); 2410 } 2411 2412 iclog = log->l_iclog; 2413 if (iclog->ic_state != XLOG_STATE_ACTIVE) { 2414 XFS_STATS_INC(xs_log_noiclogs); 2415 2416 /* Wait for log writes to have flushed */ 2417 xlog_wait(&log->l_flush_wait, &log->l_icloglock); 2418 goto restart; 2419 } 2420 2421 head = &iclog->ic_header; 2422 2423 atomic_inc(&iclog->ic_refcnt); /* prevents sync */ 2424 log_offset = iclog->ic_offset; 2425 2426 /* On the 1st write to an iclog, figure out lsn. This works 2427 * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are 2428 * committing to. If the offset is set, that's how many blocks 2429 * must be written. 2430 */ 2431 if (log_offset == 0) { 2432 ticket->t_curr_res -= log->l_iclog_hsize; 2433 xlog_tic_add_region(ticket, 2434 log->l_iclog_hsize, 2435 XLOG_REG_TYPE_LRHEADER); 2436 head->h_cycle = cpu_to_be32(log->l_curr_cycle); 2437 head->h_lsn = cpu_to_be64( 2438 xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block)); 2439 ASSERT(log->l_curr_block >= 0); 2440 } 2441 2442 /* If there is enough room to write everything, then do it. Otherwise, 2443 * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC 2444 * bit is on, so this will get flushed out. Don't update ic_offset 2445 * until you know exactly how many bytes get copied. Therefore, wait 2446 * until later to update ic_offset. 2447 * 2448 * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's 2449 * can fit into remaining data section. 2450 */ 2451 if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) { 2452 xlog_state_switch_iclogs(log, iclog, iclog->ic_size); 2453 2454 /* 2455 * If I'm the only one writing to this iclog, sync it to disk. 2456 * We need to do an atomic compare and decrement here to avoid 2457 * racing with concurrent atomic_dec_and_lock() calls in 2458 * xlog_state_release_iclog() when there is more than one 2459 * reference to the iclog. 2460 */ 2461 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) { 2462 /* we are the only one */ 2463 spin_unlock(&log->l_icloglock); 2464 error = xlog_state_release_iclog(log, iclog); 2465 if (error) 2466 return error; 2467 } else { 2468 spin_unlock(&log->l_icloglock); 2469 } 2470 goto restart; 2471 } 2472 2473 /* Do we have enough room to write the full amount in the remainder 2474 * of this iclog? Or must we continue a write on the next iclog and 2475 * mark this iclog as completely taken? In the case where we switch 2476 * iclogs (to mark it taken), this particular iclog will release/sync 2477 * to disk in xlog_write(). 2478 */ 2479 if (len <= iclog->ic_size - iclog->ic_offset) { 2480 *continued_write = 0; 2481 iclog->ic_offset += len; 2482 } else { 2483 *continued_write = 1; 2484 xlog_state_switch_iclogs(log, iclog, iclog->ic_size); 2485 } 2486 *iclogp = iclog; 2487 2488 ASSERT(iclog->ic_offset <= iclog->ic_size); 2489 spin_unlock(&log->l_icloglock); 2490 2491 *logoffsetp = log_offset; 2492 return 0; 2493 } /* xlog_state_get_iclog_space */ 2494 2495 /* 2496 * Atomically get the log space required for a log ticket. 2497 * 2498 * Once a ticket gets put onto the reserveq, it will only return after 2499 * the needed reservation is satisfied. 2500 * 2501 * This function is structured so that it has a lock free fast path. This is 2502 * necessary because every new transaction reservation will come through this 2503 * path. Hence any lock will be globally hot if we take it unconditionally on 2504 * every pass. 2505 * 2506 * As tickets are only ever moved on and off the reserveq under the 2507 * l_grant_reserve_lock, we only need to take that lock if we are going 2508 * to add the ticket to the queue and sleep. We can avoid taking the lock if the 2509 * ticket was never added to the reserveq because the t_queue list head will be 2510 * empty and we hold the only reference to it so it can safely be checked 2511 * unlocked. 2512 */ 2513 STATIC int 2514 xlog_grant_log_space(xlog_t *log, 2515 xlog_ticket_t *tic) 2516 { 2517 int free_bytes; 2518 int need_bytes; 2519 2520 #ifdef DEBUG 2521 if (log->l_flags & XLOG_ACTIVE_RECOVERY) 2522 panic("grant Recovery problem"); 2523 #endif 2524 2525 trace_xfs_log_grant_enter(log, tic); 2526 2527 need_bytes = tic->t_unit_res; 2528 if (tic->t_flags & XFS_LOG_PERM_RESERV) 2529 need_bytes *= tic->t_ocnt; 2530 2531 /* something is already sleeping; insert new transaction at end */ 2532 if (!list_empty_careful(&log->l_reserveq)) { 2533 spin_lock(&log->l_grant_reserve_lock); 2534 /* recheck the queue now we are locked */ 2535 if (list_empty(&log->l_reserveq)) { 2536 spin_unlock(&log->l_grant_reserve_lock); 2537 goto redo; 2538 } 2539 list_add_tail(&tic->t_queue, &log->l_reserveq); 2540 2541 trace_xfs_log_grant_sleep1(log, tic); 2542 2543 /* 2544 * Gotta check this before going to sleep, while we're 2545 * holding the grant lock. 2546 */ 2547 if (XLOG_FORCED_SHUTDOWN(log)) 2548 goto error_return; 2549 2550 XFS_STATS_INC(xs_sleep_logspace); 2551 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock); 2552 2553 /* 2554 * If we got an error, and the filesystem is shutting down, 2555 * we'll catch it down below. So just continue... 2556 */ 2557 trace_xfs_log_grant_wake1(log, tic); 2558 } 2559 2560 redo: 2561 if (XLOG_FORCED_SHUTDOWN(log)) 2562 goto error_return_unlocked; 2563 2564 free_bytes = xlog_space_left(log, &log->l_grant_reserve_head); 2565 if (free_bytes < need_bytes) { 2566 spin_lock(&log->l_grant_reserve_lock); 2567 if (list_empty(&tic->t_queue)) 2568 list_add_tail(&tic->t_queue, &log->l_reserveq); 2569 2570 trace_xfs_log_grant_sleep2(log, tic); 2571 2572 if (XLOG_FORCED_SHUTDOWN(log)) 2573 goto error_return; 2574 2575 xlog_grant_push_ail(log, need_bytes); 2576 2577 XFS_STATS_INC(xs_sleep_logspace); 2578 xlog_wait(&tic->t_wait, &log->l_grant_reserve_lock); 2579 2580 trace_xfs_log_grant_wake2(log, tic); 2581 goto redo; 2582 } 2583 2584 if (!list_empty(&tic->t_queue)) { 2585 spin_lock(&log->l_grant_reserve_lock); 2586 list_del_init(&tic->t_queue); 2587 spin_unlock(&log->l_grant_reserve_lock); 2588 } 2589 2590 /* we've got enough space */ 2591 xlog_grant_add_space(log, &log->l_grant_reserve_head, need_bytes); 2592 xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes); 2593 trace_xfs_log_grant_exit(log, tic); 2594 xlog_verify_grant_tail(log); 2595 return 0; 2596 2597 error_return_unlocked: 2598 spin_lock(&log->l_grant_reserve_lock); 2599 error_return: 2600 list_del_init(&tic->t_queue); 2601 spin_unlock(&log->l_grant_reserve_lock); 2602 trace_xfs_log_grant_error(log, tic); 2603 2604 /* 2605 * If we are failing, make sure the ticket doesn't have any 2606 * current reservations. We don't want to add this back when 2607 * the ticket/transaction gets cancelled. 2608 */ 2609 tic->t_curr_res = 0; 2610 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */ 2611 return XFS_ERROR(EIO); 2612 } /* xlog_grant_log_space */ 2613 2614 2615 /* 2616 * Replenish the byte reservation required by moving the grant write head. 2617 * 2618 * Similar to xlog_grant_log_space, the function is structured to have a lock 2619 * free fast path. 2620 */ 2621 STATIC int 2622 xlog_regrant_write_log_space(xlog_t *log, 2623 xlog_ticket_t *tic) 2624 { 2625 int free_bytes, need_bytes; 2626 2627 tic->t_curr_res = tic->t_unit_res; 2628 xlog_tic_reset_res(tic); 2629 2630 if (tic->t_cnt > 0) 2631 return 0; 2632 2633 #ifdef DEBUG 2634 if (log->l_flags & XLOG_ACTIVE_RECOVERY) 2635 panic("regrant Recovery problem"); 2636 #endif 2637 2638 trace_xfs_log_regrant_write_enter(log, tic); 2639 if (XLOG_FORCED_SHUTDOWN(log)) 2640 goto error_return_unlocked; 2641 2642 /* If there are other waiters on the queue then give them a 2643 * chance at logspace before us. Wake up the first waiters, 2644 * if we do not wake up all the waiters then go to sleep waiting 2645 * for more free space, otherwise try to get some space for 2646 * this transaction. 2647 */ 2648 need_bytes = tic->t_unit_res; 2649 if (!list_empty_careful(&log->l_writeq)) { 2650 struct xlog_ticket *ntic; 2651 2652 spin_lock(&log->l_grant_write_lock); 2653 free_bytes = xlog_space_left(log, &log->l_grant_write_head); 2654 list_for_each_entry(ntic, &log->l_writeq, t_queue) { 2655 ASSERT(ntic->t_flags & XLOG_TIC_PERM_RESERV); 2656 2657 if (free_bytes < ntic->t_unit_res) 2658 break; 2659 free_bytes -= ntic->t_unit_res; 2660 wake_up(&ntic->t_wait); 2661 } 2662 2663 if (ntic != list_first_entry(&log->l_writeq, 2664 struct xlog_ticket, t_queue)) { 2665 if (list_empty(&tic->t_queue)) 2666 list_add_tail(&tic->t_queue, &log->l_writeq); 2667 trace_xfs_log_regrant_write_sleep1(log, tic); 2668 2669 xlog_grant_push_ail(log, need_bytes); 2670 2671 XFS_STATS_INC(xs_sleep_logspace); 2672 xlog_wait(&tic->t_wait, &log->l_grant_write_lock); 2673 trace_xfs_log_regrant_write_wake1(log, tic); 2674 } else 2675 spin_unlock(&log->l_grant_write_lock); 2676 } 2677 2678 redo: 2679 if (XLOG_FORCED_SHUTDOWN(log)) 2680 goto error_return_unlocked; 2681 2682 free_bytes = xlog_space_left(log, &log->l_grant_write_head); 2683 if (free_bytes < need_bytes) { 2684 spin_lock(&log->l_grant_write_lock); 2685 if (list_empty(&tic->t_queue)) 2686 list_add_tail(&tic->t_queue, &log->l_writeq); 2687 2688 if (XLOG_FORCED_SHUTDOWN(log)) 2689 goto error_return; 2690 2691 xlog_grant_push_ail(log, need_bytes); 2692 2693 XFS_STATS_INC(xs_sleep_logspace); 2694 trace_xfs_log_regrant_write_sleep2(log, tic); 2695 xlog_wait(&tic->t_wait, &log->l_grant_write_lock); 2696 2697 trace_xfs_log_regrant_write_wake2(log, tic); 2698 goto redo; 2699 } 2700 2701 if (!list_empty(&tic->t_queue)) { 2702 spin_lock(&log->l_grant_write_lock); 2703 list_del_init(&tic->t_queue); 2704 spin_unlock(&log->l_grant_write_lock); 2705 } 2706 2707 /* we've got enough space */ 2708 xlog_grant_add_space(log, &log->l_grant_write_head, need_bytes); 2709 trace_xfs_log_regrant_write_exit(log, tic); 2710 xlog_verify_grant_tail(log); 2711 return 0; 2712 2713 2714 error_return_unlocked: 2715 spin_lock(&log->l_grant_write_lock); 2716 error_return: 2717 list_del_init(&tic->t_queue); 2718 spin_unlock(&log->l_grant_write_lock); 2719 trace_xfs_log_regrant_write_error(log, tic); 2720 2721 /* 2722 * If we are failing, make sure the ticket doesn't have any 2723 * current reservations. We don't want to add this back when 2724 * the ticket/transaction gets cancelled. 2725 */ 2726 tic->t_curr_res = 0; 2727 tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */ 2728 return XFS_ERROR(EIO); 2729 } /* xlog_regrant_write_log_space */ 2730 2731 2732 /* The first cnt-1 times through here we don't need to 2733 * move the grant write head because the permanent 2734 * reservation has reserved cnt times the unit amount. 2735 * Release part of current permanent unit reservation and 2736 * reset current reservation to be one units worth. Also 2737 * move grant reservation head forward. 2738 */ 2739 STATIC void 2740 xlog_regrant_reserve_log_space(xlog_t *log, 2741 xlog_ticket_t *ticket) 2742 { 2743 trace_xfs_log_regrant_reserve_enter(log, ticket); 2744 2745 if (ticket->t_cnt > 0) 2746 ticket->t_cnt--; 2747 2748 xlog_grant_sub_space(log, &log->l_grant_reserve_head, 2749 ticket->t_curr_res); 2750 xlog_grant_sub_space(log, &log->l_grant_write_head, 2751 ticket->t_curr_res); 2752 ticket->t_curr_res = ticket->t_unit_res; 2753 xlog_tic_reset_res(ticket); 2754 2755 trace_xfs_log_regrant_reserve_sub(log, ticket); 2756 2757 /* just return if we still have some of the pre-reserved space */ 2758 if (ticket->t_cnt > 0) 2759 return; 2760 2761 xlog_grant_add_space(log, &log->l_grant_reserve_head, 2762 ticket->t_unit_res); 2763 2764 trace_xfs_log_regrant_reserve_exit(log, ticket); 2765 2766 ticket->t_curr_res = ticket->t_unit_res; 2767 xlog_tic_reset_res(ticket); 2768 } /* xlog_regrant_reserve_log_space */ 2769 2770 2771 /* 2772 * Give back the space left from a reservation. 2773 * 2774 * All the information we need to make a correct determination of space left 2775 * is present. For non-permanent reservations, things are quite easy. The 2776 * count should have been decremented to zero. We only need to deal with the 2777 * space remaining in the current reservation part of the ticket. If the 2778 * ticket contains a permanent reservation, there may be left over space which 2779 * needs to be released. A count of N means that N-1 refills of the current 2780 * reservation can be done before we need to ask for more space. The first 2781 * one goes to fill up the first current reservation. Once we run out of 2782 * space, the count will stay at zero and the only space remaining will be 2783 * in the current reservation field. 2784 */ 2785 STATIC void 2786 xlog_ungrant_log_space(xlog_t *log, 2787 xlog_ticket_t *ticket) 2788 { 2789 int bytes; 2790 2791 if (ticket->t_cnt > 0) 2792 ticket->t_cnt--; 2793 2794 trace_xfs_log_ungrant_enter(log, ticket); 2795 trace_xfs_log_ungrant_sub(log, ticket); 2796 2797 /* 2798 * If this is a permanent reservation ticket, we may be able to free 2799 * up more space based on the remaining count. 2800 */ 2801 bytes = ticket->t_curr_res; 2802 if (ticket->t_cnt > 0) { 2803 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV); 2804 bytes += ticket->t_unit_res*ticket->t_cnt; 2805 } 2806 2807 xlog_grant_sub_space(log, &log->l_grant_reserve_head, bytes); 2808 xlog_grant_sub_space(log, &log->l_grant_write_head, bytes); 2809 2810 trace_xfs_log_ungrant_exit(log, ticket); 2811 2812 xfs_log_move_tail(log->l_mp, 1); 2813 } /* xlog_ungrant_log_space */ 2814 2815 2816 /* 2817 * Flush iclog to disk if this is the last reference to the given iclog and 2818 * the WANT_SYNC bit is set. 2819 * 2820 * When this function is entered, the iclog is not necessarily in the 2821 * WANT_SYNC state. It may be sitting around waiting to get filled. 2822 * 2823 * 2824 */ 2825 STATIC int 2826 xlog_state_release_iclog( 2827 xlog_t *log, 2828 xlog_in_core_t *iclog) 2829 { 2830 int sync = 0; /* do we sync? */ 2831 2832 if (iclog->ic_state & XLOG_STATE_IOERROR) 2833 return XFS_ERROR(EIO); 2834 2835 ASSERT(atomic_read(&iclog->ic_refcnt) > 0); 2836 if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock)) 2837 return 0; 2838 2839 if (iclog->ic_state & XLOG_STATE_IOERROR) { 2840 spin_unlock(&log->l_icloglock); 2841 return XFS_ERROR(EIO); 2842 } 2843 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE || 2844 iclog->ic_state == XLOG_STATE_WANT_SYNC); 2845 2846 if (iclog->ic_state == XLOG_STATE_WANT_SYNC) { 2847 /* update tail before writing to iclog */ 2848 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp); 2849 sync++; 2850 iclog->ic_state = XLOG_STATE_SYNCING; 2851 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn); 2852 xlog_verify_tail_lsn(log, iclog, tail_lsn); 2853 /* cycle incremented when incrementing curr_block */ 2854 } 2855 spin_unlock(&log->l_icloglock); 2856 2857 /* 2858 * We let the log lock go, so it's possible that we hit a log I/O 2859 * error or some other SHUTDOWN condition that marks the iclog 2860 * as XLOG_STATE_IOERROR before the bwrite. However, we know that 2861 * this iclog has consistent data, so we ignore IOERROR 2862 * flags after this point. 2863 */ 2864 if (sync) 2865 return xlog_sync(log, iclog); 2866 return 0; 2867 } /* xlog_state_release_iclog */ 2868 2869 2870 /* 2871 * This routine will mark the current iclog in the ring as WANT_SYNC 2872 * and move the current iclog pointer to the next iclog in the ring. 2873 * When this routine is called from xlog_state_get_iclog_space(), the 2874 * exact size of the iclog has not yet been determined. All we know is 2875 * that every data block. We have run out of space in this log record. 2876 */ 2877 STATIC void 2878 xlog_state_switch_iclogs(xlog_t *log, 2879 xlog_in_core_t *iclog, 2880 int eventual_size) 2881 { 2882 ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE); 2883 if (!eventual_size) 2884 eventual_size = iclog->ic_offset; 2885 iclog->ic_state = XLOG_STATE_WANT_SYNC; 2886 iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block); 2887 log->l_prev_block = log->l_curr_block; 2888 log->l_prev_cycle = log->l_curr_cycle; 2889 2890 /* roll log?: ic_offset changed later */ 2891 log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize); 2892 2893 /* Round up to next log-sunit */ 2894 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) && 2895 log->l_mp->m_sb.sb_logsunit > 1) { 2896 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit); 2897 log->l_curr_block = roundup(log->l_curr_block, sunit_bb); 2898 } 2899 2900 if (log->l_curr_block >= log->l_logBBsize) { 2901 log->l_curr_cycle++; 2902 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM) 2903 log->l_curr_cycle++; 2904 log->l_curr_block -= log->l_logBBsize; 2905 ASSERT(log->l_curr_block >= 0); 2906 } 2907 ASSERT(iclog == log->l_iclog); 2908 log->l_iclog = iclog->ic_next; 2909 } /* xlog_state_switch_iclogs */ 2910 2911 /* 2912 * Write out all data in the in-core log as of this exact moment in time. 2913 * 2914 * Data may be written to the in-core log during this call. However, 2915 * we don't guarantee this data will be written out. A change from past 2916 * implementation means this routine will *not* write out zero length LRs. 2917 * 2918 * Basically, we try and perform an intelligent scan of the in-core logs. 2919 * If we determine there is no flushable data, we just return. There is no 2920 * flushable data if: 2921 * 2922 * 1. the current iclog is active and has no data; the previous iclog 2923 * is in the active or dirty state. 2924 * 2. the current iclog is drity, and the previous iclog is in the 2925 * active or dirty state. 2926 * 2927 * We may sleep if: 2928 * 2929 * 1. the current iclog is not in the active nor dirty state. 2930 * 2. the current iclog dirty, and the previous iclog is not in the 2931 * active nor dirty state. 2932 * 3. the current iclog is active, and there is another thread writing 2933 * to this particular iclog. 2934 * 4. a) the current iclog is active and has no other writers 2935 * b) when we return from flushing out this iclog, it is still 2936 * not in the active nor dirty state. 2937 */ 2938 int 2939 _xfs_log_force( 2940 struct xfs_mount *mp, 2941 uint flags, 2942 int *log_flushed) 2943 { 2944 struct log *log = mp->m_log; 2945 struct xlog_in_core *iclog; 2946 xfs_lsn_t lsn; 2947 2948 XFS_STATS_INC(xs_log_force); 2949 2950 if (log->l_cilp) 2951 xlog_cil_force(log); 2952 2953 spin_lock(&log->l_icloglock); 2954 2955 iclog = log->l_iclog; 2956 if (iclog->ic_state & XLOG_STATE_IOERROR) { 2957 spin_unlock(&log->l_icloglock); 2958 return XFS_ERROR(EIO); 2959 } 2960 2961 /* If the head iclog is not active nor dirty, we just attach 2962 * ourselves to the head and go to sleep. 2963 */ 2964 if (iclog->ic_state == XLOG_STATE_ACTIVE || 2965 iclog->ic_state == XLOG_STATE_DIRTY) { 2966 /* 2967 * If the head is dirty or (active and empty), then 2968 * we need to look at the previous iclog. If the previous 2969 * iclog is active or dirty we are done. There is nothing 2970 * to sync out. Otherwise, we attach ourselves to the 2971 * previous iclog and go to sleep. 2972 */ 2973 if (iclog->ic_state == XLOG_STATE_DIRTY || 2974 (atomic_read(&iclog->ic_refcnt) == 0 2975 && iclog->ic_offset == 0)) { 2976 iclog = iclog->ic_prev; 2977 if (iclog->ic_state == XLOG_STATE_ACTIVE || 2978 iclog->ic_state == XLOG_STATE_DIRTY) 2979 goto no_sleep; 2980 else 2981 goto maybe_sleep; 2982 } else { 2983 if (atomic_read(&iclog->ic_refcnt) == 0) { 2984 /* We are the only one with access to this 2985 * iclog. Flush it out now. There should 2986 * be a roundoff of zero to show that someone 2987 * has already taken care of the roundoff from 2988 * the previous sync. 2989 */ 2990 atomic_inc(&iclog->ic_refcnt); 2991 lsn = be64_to_cpu(iclog->ic_header.h_lsn); 2992 xlog_state_switch_iclogs(log, iclog, 0); 2993 spin_unlock(&log->l_icloglock); 2994 2995 if (xlog_state_release_iclog(log, iclog)) 2996 return XFS_ERROR(EIO); 2997 2998 if (log_flushed) 2999 *log_flushed = 1; 3000 spin_lock(&log->l_icloglock); 3001 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn && 3002 iclog->ic_state != XLOG_STATE_DIRTY) 3003 goto maybe_sleep; 3004 else 3005 goto no_sleep; 3006 } else { 3007 /* Someone else is writing to this iclog. 3008 * Use its call to flush out the data. However, 3009 * the other thread may not force out this LR, 3010 * so we mark it WANT_SYNC. 3011 */ 3012 xlog_state_switch_iclogs(log, iclog, 0); 3013 goto maybe_sleep; 3014 } 3015 } 3016 } 3017 3018 /* By the time we come around again, the iclog could've been filled 3019 * which would give it another lsn. If we have a new lsn, just 3020 * return because the relevant data has been flushed. 3021 */ 3022 maybe_sleep: 3023 if (flags & XFS_LOG_SYNC) { 3024 /* 3025 * We must check if we're shutting down here, before 3026 * we wait, while we're holding the l_icloglock. 3027 * Then we check again after waking up, in case our 3028 * sleep was disturbed by a bad news. 3029 */ 3030 if (iclog->ic_state & XLOG_STATE_IOERROR) { 3031 spin_unlock(&log->l_icloglock); 3032 return XFS_ERROR(EIO); 3033 } 3034 XFS_STATS_INC(xs_log_force_sleep); 3035 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock); 3036 /* 3037 * No need to grab the log lock here since we're 3038 * only deciding whether or not to return EIO 3039 * and the memory read should be atomic. 3040 */ 3041 if (iclog->ic_state & XLOG_STATE_IOERROR) 3042 return XFS_ERROR(EIO); 3043 if (log_flushed) 3044 *log_flushed = 1; 3045 } else { 3046 3047 no_sleep: 3048 spin_unlock(&log->l_icloglock); 3049 } 3050 return 0; 3051 } 3052 3053 /* 3054 * Wrapper for _xfs_log_force(), to be used when caller doesn't care 3055 * about errors or whether the log was flushed or not. This is the normal 3056 * interface to use when trying to unpin items or move the log forward. 3057 */ 3058 void 3059 xfs_log_force( 3060 xfs_mount_t *mp, 3061 uint flags) 3062 { 3063 int error; 3064 3065 error = _xfs_log_force(mp, flags, NULL); 3066 if (error) 3067 xfs_warn(mp, "%s: error %d returned.", __func__, error); 3068 } 3069 3070 /* 3071 * Force the in-core log to disk for a specific LSN. 3072 * 3073 * Find in-core log with lsn. 3074 * If it is in the DIRTY state, just return. 3075 * If it is in the ACTIVE state, move the in-core log into the WANT_SYNC 3076 * state and go to sleep or return. 3077 * If it is in any other state, go to sleep or return. 3078 * 3079 * Synchronous forces are implemented with a signal variable. All callers 3080 * to force a given lsn to disk will wait on a the sv attached to the 3081 * specific in-core log. When given in-core log finally completes its 3082 * write to disk, that thread will wake up all threads waiting on the 3083 * sv. 3084 */ 3085 int 3086 _xfs_log_force_lsn( 3087 struct xfs_mount *mp, 3088 xfs_lsn_t lsn, 3089 uint flags, 3090 int *log_flushed) 3091 { 3092 struct log *log = mp->m_log; 3093 struct xlog_in_core *iclog; 3094 int already_slept = 0; 3095 3096 ASSERT(lsn != 0); 3097 3098 XFS_STATS_INC(xs_log_force); 3099 3100 if (log->l_cilp) { 3101 lsn = xlog_cil_force_lsn(log, lsn); 3102 if (lsn == NULLCOMMITLSN) 3103 return 0; 3104 } 3105 3106 try_again: 3107 spin_lock(&log->l_icloglock); 3108 iclog = log->l_iclog; 3109 if (iclog->ic_state & XLOG_STATE_IOERROR) { 3110 spin_unlock(&log->l_icloglock); 3111 return XFS_ERROR(EIO); 3112 } 3113 3114 do { 3115 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) { 3116 iclog = iclog->ic_next; 3117 continue; 3118 } 3119 3120 if (iclog->ic_state == XLOG_STATE_DIRTY) { 3121 spin_unlock(&log->l_icloglock); 3122 return 0; 3123 } 3124 3125 if (iclog->ic_state == XLOG_STATE_ACTIVE) { 3126 /* 3127 * We sleep here if we haven't already slept (e.g. 3128 * this is the first time we've looked at the correct 3129 * iclog buf) and the buffer before us is going to 3130 * be sync'ed. The reason for this is that if we 3131 * are doing sync transactions here, by waiting for 3132 * the previous I/O to complete, we can allow a few 3133 * more transactions into this iclog before we close 3134 * it down. 3135 * 3136 * Otherwise, we mark the buffer WANT_SYNC, and bump 3137 * up the refcnt so we can release the log (which 3138 * drops the ref count). The state switch keeps new 3139 * transaction commits from using this buffer. When 3140 * the current commits finish writing into the buffer, 3141 * the refcount will drop to zero and the buffer will 3142 * go out then. 3143 */ 3144 if (!already_slept && 3145 (iclog->ic_prev->ic_state & 3146 (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) { 3147 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR)); 3148 3149 XFS_STATS_INC(xs_log_force_sleep); 3150 3151 xlog_wait(&iclog->ic_prev->ic_write_wait, 3152 &log->l_icloglock); 3153 if (log_flushed) 3154 *log_flushed = 1; 3155 already_slept = 1; 3156 goto try_again; 3157 } 3158 atomic_inc(&iclog->ic_refcnt); 3159 xlog_state_switch_iclogs(log, iclog, 0); 3160 spin_unlock(&log->l_icloglock); 3161 if (xlog_state_release_iclog(log, iclog)) 3162 return XFS_ERROR(EIO); 3163 if (log_flushed) 3164 *log_flushed = 1; 3165 spin_lock(&log->l_icloglock); 3166 } 3167 3168 if ((flags & XFS_LOG_SYNC) && /* sleep */ 3169 !(iclog->ic_state & 3170 (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) { 3171 /* 3172 * Don't wait on completion if we know that we've 3173 * gotten a log write error. 3174 */ 3175 if (iclog->ic_state & XLOG_STATE_IOERROR) { 3176 spin_unlock(&log->l_icloglock); 3177 return XFS_ERROR(EIO); 3178 } 3179 XFS_STATS_INC(xs_log_force_sleep); 3180 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock); 3181 /* 3182 * No need to grab the log lock here since we're 3183 * only deciding whether or not to return EIO 3184 * and the memory read should be atomic. 3185 */ 3186 if (iclog->ic_state & XLOG_STATE_IOERROR) 3187 return XFS_ERROR(EIO); 3188 3189 if (log_flushed) 3190 *log_flushed = 1; 3191 } else { /* just return */ 3192 spin_unlock(&log->l_icloglock); 3193 } 3194 3195 return 0; 3196 } while (iclog != log->l_iclog); 3197 3198 spin_unlock(&log->l_icloglock); 3199 return 0; 3200 } 3201 3202 /* 3203 * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care 3204 * about errors or whether the log was flushed or not. This is the normal 3205 * interface to use when trying to unpin items or move the log forward. 3206 */ 3207 void 3208 xfs_log_force_lsn( 3209 xfs_mount_t *mp, 3210 xfs_lsn_t lsn, 3211 uint flags) 3212 { 3213 int error; 3214 3215 error = _xfs_log_force_lsn(mp, lsn, flags, NULL); 3216 if (error) 3217 xfs_warn(mp, "%s: error %d returned.", __func__, error); 3218 } 3219 3220 /* 3221 * Called when we want to mark the current iclog as being ready to sync to 3222 * disk. 3223 */ 3224 STATIC void 3225 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog) 3226 { 3227 assert_spin_locked(&log->l_icloglock); 3228 3229 if (iclog->ic_state == XLOG_STATE_ACTIVE) { 3230 xlog_state_switch_iclogs(log, iclog, 0); 3231 } else { 3232 ASSERT(iclog->ic_state & 3233 (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR)); 3234 } 3235 } 3236 3237 3238 /***************************************************************************** 3239 * 3240 * TICKET functions 3241 * 3242 ***************************************************************************** 3243 */ 3244 3245 /* 3246 * Free a used ticket when its refcount falls to zero. 3247 */ 3248 void 3249 xfs_log_ticket_put( 3250 xlog_ticket_t *ticket) 3251 { 3252 ASSERT(atomic_read(&ticket->t_ref) > 0); 3253 if (atomic_dec_and_test(&ticket->t_ref)) 3254 kmem_zone_free(xfs_log_ticket_zone, ticket); 3255 } 3256 3257 xlog_ticket_t * 3258 xfs_log_ticket_get( 3259 xlog_ticket_t *ticket) 3260 { 3261 ASSERT(atomic_read(&ticket->t_ref) > 0); 3262 atomic_inc(&ticket->t_ref); 3263 return ticket; 3264 } 3265 3266 /* 3267 * Allocate and initialise a new log ticket. 3268 */ 3269 xlog_ticket_t * 3270 xlog_ticket_alloc( 3271 struct log *log, 3272 int unit_bytes, 3273 int cnt, 3274 char client, 3275 uint xflags, 3276 int alloc_flags) 3277 { 3278 struct xlog_ticket *tic; 3279 uint num_headers; 3280 int iclog_space; 3281 3282 tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags); 3283 if (!tic) 3284 return NULL; 3285 3286 /* 3287 * Permanent reservations have up to 'cnt'-1 active log operations 3288 * in the log. A unit in this case is the amount of space for one 3289 * of these log operations. Normal reservations have a cnt of 1 3290 * and their unit amount is the total amount of space required. 3291 * 3292 * The following lines of code account for non-transaction data 3293 * which occupy space in the on-disk log. 3294 * 3295 * Normal form of a transaction is: 3296 * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph> 3297 * and then there are LR hdrs, split-recs and roundoff at end of syncs. 3298 * 3299 * We need to account for all the leadup data and trailer data 3300 * around the transaction data. 3301 * And then we need to account for the worst case in terms of using 3302 * more space. 3303 * The worst case will happen if: 3304 * - the placement of the transaction happens to be such that the 3305 * roundoff is at its maximum 3306 * - the transaction data is synced before the commit record is synced 3307 * i.e. <transaction-data><roundoff> | <commit-rec><roundoff> 3308 * Therefore the commit record is in its own Log Record. 3309 * This can happen as the commit record is called with its 3310 * own region to xlog_write(). 3311 * This then means that in the worst case, roundoff can happen for 3312 * the commit-rec as well. 3313 * The commit-rec is smaller than padding in this scenario and so it is 3314 * not added separately. 3315 */ 3316 3317 /* for trans header */ 3318 unit_bytes += sizeof(xlog_op_header_t); 3319 unit_bytes += sizeof(xfs_trans_header_t); 3320 3321 /* for start-rec */ 3322 unit_bytes += sizeof(xlog_op_header_t); 3323 3324 /* 3325 * for LR headers - the space for data in an iclog is the size minus 3326 * the space used for the headers. If we use the iclog size, then we 3327 * undercalculate the number of headers required. 3328 * 3329 * Furthermore - the addition of op headers for split-recs might 3330 * increase the space required enough to require more log and op 3331 * headers, so take that into account too. 3332 * 3333 * IMPORTANT: This reservation makes the assumption that if this 3334 * transaction is the first in an iclog and hence has the LR headers 3335 * accounted to it, then the remaining space in the iclog is 3336 * exclusively for this transaction. i.e. if the transaction is larger 3337 * than the iclog, it will be the only thing in that iclog. 3338 * Fundamentally, this means we must pass the entire log vector to 3339 * xlog_write to guarantee this. 3340 */ 3341 iclog_space = log->l_iclog_size - log->l_iclog_hsize; 3342 num_headers = howmany(unit_bytes, iclog_space); 3343 3344 /* for split-recs - ophdrs added when data split over LRs */ 3345 unit_bytes += sizeof(xlog_op_header_t) * num_headers; 3346 3347 /* add extra header reservations if we overrun */ 3348 while (!num_headers || 3349 howmany(unit_bytes, iclog_space) > num_headers) { 3350 unit_bytes += sizeof(xlog_op_header_t); 3351 num_headers++; 3352 } 3353 unit_bytes += log->l_iclog_hsize * num_headers; 3354 3355 /* for commit-rec LR header - note: padding will subsume the ophdr */ 3356 unit_bytes += log->l_iclog_hsize; 3357 3358 /* for roundoff padding for transaction data and one for commit record */ 3359 if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) && 3360 log->l_mp->m_sb.sb_logsunit > 1) { 3361 /* log su roundoff */ 3362 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit; 3363 } else { 3364 /* BB roundoff */ 3365 unit_bytes += 2*BBSIZE; 3366 } 3367 3368 atomic_set(&tic->t_ref, 1); 3369 INIT_LIST_HEAD(&tic->t_queue); 3370 tic->t_unit_res = unit_bytes; 3371 tic->t_curr_res = unit_bytes; 3372 tic->t_cnt = cnt; 3373 tic->t_ocnt = cnt; 3374 tic->t_tid = random32(); 3375 tic->t_clientid = client; 3376 tic->t_flags = XLOG_TIC_INITED; 3377 tic->t_trans_type = 0; 3378 if (xflags & XFS_LOG_PERM_RESERV) 3379 tic->t_flags |= XLOG_TIC_PERM_RESERV; 3380 init_waitqueue_head(&tic->t_wait); 3381 3382 xlog_tic_reset_res(tic); 3383 3384 return tic; 3385 } 3386 3387 3388 /****************************************************************************** 3389 * 3390 * Log debug routines 3391 * 3392 ****************************************************************************** 3393 */ 3394 #if defined(DEBUG) 3395 /* 3396 * Make sure that the destination ptr is within the valid data region of 3397 * one of the iclogs. This uses backup pointers stored in a different 3398 * part of the log in case we trash the log structure. 3399 */ 3400 void 3401 xlog_verify_dest_ptr( 3402 struct log *log, 3403 char *ptr) 3404 { 3405 int i; 3406 int good_ptr = 0; 3407 3408 for (i = 0; i < log->l_iclog_bufs; i++) { 3409 if (ptr >= log->l_iclog_bak[i] && 3410 ptr <= log->l_iclog_bak[i] + log->l_iclog_size) 3411 good_ptr++; 3412 } 3413 3414 if (!good_ptr) 3415 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__); 3416 } 3417 3418 /* 3419 * Check to make sure the grant write head didn't just over lap the tail. If 3420 * the cycles are the same, we can't be overlapping. Otherwise, make sure that 3421 * the cycles differ by exactly one and check the byte count. 3422 * 3423 * This check is run unlocked, so can give false positives. Rather than assert 3424 * on failures, use a warn-once flag and a panic tag to allow the admin to 3425 * determine if they want to panic the machine when such an error occurs. For 3426 * debug kernels this will have the same effect as using an assert but, unlinke 3427 * an assert, it can be turned off at runtime. 3428 */ 3429 STATIC void 3430 xlog_verify_grant_tail( 3431 struct log *log) 3432 { 3433 int tail_cycle, tail_blocks; 3434 int cycle, space; 3435 3436 xlog_crack_grant_head(&log->l_grant_write_head, &cycle, &space); 3437 xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks); 3438 if (tail_cycle != cycle) { 3439 if (cycle - 1 != tail_cycle && 3440 !(log->l_flags & XLOG_TAIL_WARN)) { 3441 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES, 3442 "%s: cycle - 1 != tail_cycle", __func__); 3443 log->l_flags |= XLOG_TAIL_WARN; 3444 } 3445 3446 if (space > BBTOB(tail_blocks) && 3447 !(log->l_flags & XLOG_TAIL_WARN)) { 3448 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES, 3449 "%s: space > BBTOB(tail_blocks)", __func__); 3450 log->l_flags |= XLOG_TAIL_WARN; 3451 } 3452 } 3453 } 3454 3455 /* check if it will fit */ 3456 STATIC void 3457 xlog_verify_tail_lsn(xlog_t *log, 3458 xlog_in_core_t *iclog, 3459 xfs_lsn_t tail_lsn) 3460 { 3461 int blocks; 3462 3463 if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) { 3464 blocks = 3465 log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn)); 3466 if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize)) 3467 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__); 3468 } else { 3469 ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle); 3470 3471 if (BLOCK_LSN(tail_lsn) == log->l_prev_block) 3472 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__); 3473 3474 blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block; 3475 if (blocks < BTOBB(iclog->ic_offset) + 1) 3476 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__); 3477 } 3478 } /* xlog_verify_tail_lsn */ 3479 3480 /* 3481 * Perform a number of checks on the iclog before writing to disk. 3482 * 3483 * 1. Make sure the iclogs are still circular 3484 * 2. Make sure we have a good magic number 3485 * 3. Make sure we don't have magic numbers in the data 3486 * 4. Check fields of each log operation header for: 3487 * A. Valid client identifier 3488 * B. tid ptr value falls in valid ptr space (user space code) 3489 * C. Length in log record header is correct according to the 3490 * individual operation headers within record. 3491 * 5. When a bwrite will occur within 5 blocks of the front of the physical 3492 * log, check the preceding blocks of the physical log to make sure all 3493 * the cycle numbers agree with the current cycle number. 3494 */ 3495 STATIC void 3496 xlog_verify_iclog(xlog_t *log, 3497 xlog_in_core_t *iclog, 3498 int count, 3499 boolean_t syncing) 3500 { 3501 xlog_op_header_t *ophead; 3502 xlog_in_core_t *icptr; 3503 xlog_in_core_2_t *xhdr; 3504 xfs_caddr_t ptr; 3505 xfs_caddr_t base_ptr; 3506 __psint_t field_offset; 3507 __uint8_t clientid; 3508 int len, i, j, k, op_len; 3509 int idx; 3510 3511 /* check validity of iclog pointers */ 3512 spin_lock(&log->l_icloglock); 3513 icptr = log->l_iclog; 3514 for (i=0; i < log->l_iclog_bufs; i++) { 3515 if (icptr == NULL) 3516 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__); 3517 icptr = icptr->ic_next; 3518 } 3519 if (icptr != log->l_iclog) 3520 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__); 3521 spin_unlock(&log->l_icloglock); 3522 3523 /* check log magic numbers */ 3524 if (be32_to_cpu(iclog->ic_header.h_magicno) != XLOG_HEADER_MAGIC_NUM) 3525 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__); 3526 3527 ptr = (xfs_caddr_t) &iclog->ic_header; 3528 for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count; 3529 ptr += BBSIZE) { 3530 if (be32_to_cpu(*(__be32 *)ptr) == XLOG_HEADER_MAGIC_NUM) 3531 xfs_emerg(log->l_mp, "%s: unexpected magic num", 3532 __func__); 3533 } 3534 3535 /* check fields */ 3536 len = be32_to_cpu(iclog->ic_header.h_num_logops); 3537 ptr = iclog->ic_datap; 3538 base_ptr = ptr; 3539 ophead = (xlog_op_header_t *)ptr; 3540 xhdr = iclog->ic_data; 3541 for (i = 0; i < len; i++) { 3542 ophead = (xlog_op_header_t *)ptr; 3543 3544 /* clientid is only 1 byte */ 3545 field_offset = (__psint_t) 3546 ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr); 3547 if (syncing == B_FALSE || (field_offset & 0x1ff)) { 3548 clientid = ophead->oh_clientid; 3549 } else { 3550 idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap); 3551 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) { 3552 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE); 3553 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE); 3554 clientid = xlog_get_client_id( 3555 xhdr[j].hic_xheader.xh_cycle_data[k]); 3556 } else { 3557 clientid = xlog_get_client_id( 3558 iclog->ic_header.h_cycle_data[idx]); 3559 } 3560 } 3561 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG) 3562 xfs_warn(log->l_mp, 3563 "%s: invalid clientid %d op 0x%p offset 0x%lx", 3564 __func__, clientid, ophead, 3565 (unsigned long)field_offset); 3566 3567 /* check length */ 3568 field_offset = (__psint_t) 3569 ((xfs_caddr_t)&(ophead->oh_len) - base_ptr); 3570 if (syncing == B_FALSE || (field_offset & 0x1ff)) { 3571 op_len = be32_to_cpu(ophead->oh_len); 3572 } else { 3573 idx = BTOBBT((__psint_t)&ophead->oh_len - 3574 (__psint_t)iclog->ic_datap); 3575 if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) { 3576 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE); 3577 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE); 3578 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]); 3579 } else { 3580 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]); 3581 } 3582 } 3583 ptr += sizeof(xlog_op_header_t) + op_len; 3584 } 3585 } /* xlog_verify_iclog */ 3586 #endif 3587 3588 /* 3589 * Mark all iclogs IOERROR. l_icloglock is held by the caller. 3590 */ 3591 STATIC int 3592 xlog_state_ioerror( 3593 xlog_t *log) 3594 { 3595 xlog_in_core_t *iclog, *ic; 3596 3597 iclog = log->l_iclog; 3598 if (! (iclog->ic_state & XLOG_STATE_IOERROR)) { 3599 /* 3600 * Mark all the incore logs IOERROR. 3601 * From now on, no log flushes will result. 3602 */ 3603 ic = iclog; 3604 do { 3605 ic->ic_state = XLOG_STATE_IOERROR; 3606 ic = ic->ic_next; 3607 } while (ic != iclog); 3608 return 0; 3609 } 3610 /* 3611 * Return non-zero, if state transition has already happened. 3612 */ 3613 return 1; 3614 } 3615 3616 /* 3617 * This is called from xfs_force_shutdown, when we're forcibly 3618 * shutting down the filesystem, typically because of an IO error. 3619 * Our main objectives here are to make sure that: 3620 * a. the filesystem gets marked 'SHUTDOWN' for all interested 3621 * parties to find out, 'atomically'. 3622 * b. those who're sleeping on log reservations, pinned objects and 3623 * other resources get woken up, and be told the bad news. 3624 * c. nothing new gets queued up after (a) and (b) are done. 3625 * d. if !logerror, flush the iclogs to disk, then seal them off 3626 * for business. 3627 * 3628 * Note: for delayed logging the !logerror case needs to flush the regions 3629 * held in memory out to the iclogs before flushing them to disk. This needs 3630 * to be done before the log is marked as shutdown, otherwise the flush to the 3631 * iclogs will fail. 3632 */ 3633 int 3634 xfs_log_force_umount( 3635 struct xfs_mount *mp, 3636 int logerror) 3637 { 3638 xlog_ticket_t *tic; 3639 xlog_t *log; 3640 int retval; 3641 3642 log = mp->m_log; 3643 3644 /* 3645 * If this happens during log recovery, don't worry about 3646 * locking; the log isn't open for business yet. 3647 */ 3648 if (!log || 3649 log->l_flags & XLOG_ACTIVE_RECOVERY) { 3650 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN; 3651 if (mp->m_sb_bp) 3652 XFS_BUF_DONE(mp->m_sb_bp); 3653 return 0; 3654 } 3655 3656 /* 3657 * Somebody could've already done the hard work for us. 3658 * No need to get locks for this. 3659 */ 3660 if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) { 3661 ASSERT(XLOG_FORCED_SHUTDOWN(log)); 3662 return 1; 3663 } 3664 retval = 0; 3665 3666 /* 3667 * Flush the in memory commit item list before marking the log as 3668 * being shut down. We need to do it in this order to ensure all the 3669 * completed transactions are flushed to disk with the xfs_log_force() 3670 * call below. 3671 */ 3672 if (!logerror && (mp->m_flags & XFS_MOUNT_DELAYLOG)) 3673 xlog_cil_force(log); 3674 3675 /* 3676 * mark the filesystem and the as in a shutdown state and wake 3677 * everybody up to tell them the bad news. 3678 */ 3679 spin_lock(&log->l_icloglock); 3680 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN; 3681 if (mp->m_sb_bp) 3682 XFS_BUF_DONE(mp->m_sb_bp); 3683 3684 /* 3685 * This flag is sort of redundant because of the mount flag, but 3686 * it's good to maintain the separation between the log and the rest 3687 * of XFS. 3688 */ 3689 log->l_flags |= XLOG_IO_ERROR; 3690 3691 /* 3692 * If we hit a log error, we want to mark all the iclogs IOERROR 3693 * while we're still holding the loglock. 3694 */ 3695 if (logerror) 3696 retval = xlog_state_ioerror(log); 3697 spin_unlock(&log->l_icloglock); 3698 3699 /* 3700 * We don't want anybody waiting for log reservations after this. That 3701 * means we have to wake up everybody queued up on reserveq as well as 3702 * writeq. In addition, we make sure in xlog_{re}grant_log_space that 3703 * we don't enqueue anything once the SHUTDOWN flag is set, and this 3704 * action is protected by the grant locks. 3705 */ 3706 spin_lock(&log->l_grant_reserve_lock); 3707 list_for_each_entry(tic, &log->l_reserveq, t_queue) 3708 wake_up(&tic->t_wait); 3709 spin_unlock(&log->l_grant_reserve_lock); 3710 3711 spin_lock(&log->l_grant_write_lock); 3712 list_for_each_entry(tic, &log->l_writeq, t_queue) 3713 wake_up(&tic->t_wait); 3714 spin_unlock(&log->l_grant_write_lock); 3715 3716 if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) { 3717 ASSERT(!logerror); 3718 /* 3719 * Force the incore logs to disk before shutting the 3720 * log down completely. 3721 */ 3722 _xfs_log_force(mp, XFS_LOG_SYNC, NULL); 3723 3724 spin_lock(&log->l_icloglock); 3725 retval = xlog_state_ioerror(log); 3726 spin_unlock(&log->l_icloglock); 3727 } 3728 /* 3729 * Wake up everybody waiting on xfs_log_force. 3730 * Callback all log item committed functions as if the 3731 * log writes were completed. 3732 */ 3733 xlog_state_do_callback(log, XFS_LI_ABORTED, NULL); 3734 3735 #ifdef XFSERRORDEBUG 3736 { 3737 xlog_in_core_t *iclog; 3738 3739 spin_lock(&log->l_icloglock); 3740 iclog = log->l_iclog; 3741 do { 3742 ASSERT(iclog->ic_callback == 0); 3743 iclog = iclog->ic_next; 3744 } while (iclog != log->l_iclog); 3745 spin_unlock(&log->l_icloglock); 3746 } 3747 #endif 3748 /* return non-zero if log IOERROR transition had already happened */ 3749 return retval; 3750 } 3751 3752 STATIC int 3753 xlog_iclogs_empty(xlog_t *log) 3754 { 3755 xlog_in_core_t *iclog; 3756 3757 iclog = log->l_iclog; 3758 do { 3759 /* endianness does not matter here, zero is zero in 3760 * any language. 3761 */ 3762 if (iclog->ic_header.h_num_logops) 3763 return 0; 3764 iclog = iclog->ic_next; 3765 } while (iclog != log->l_iclog); 3766 return 1; 3767 } 3768