fs/xfs/xfs_log.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   4  * All Rights Reserved.
   5  */
   6 #include "xfs.h"
   7 #include "xfs_fs.h"
   8 #include "xfs_shared.h"
   9 #include "xfs_format.h"
  10 #include "xfs_log_format.h"
  11 #include "xfs_trans_resv.h"
  12 #include "xfs_mount.h"
  13 #include "xfs_errortag.h"
  14 #include "xfs_error.h"
  15 #include "xfs_trans.h"
  16 #include "xfs_trans_priv.h"
  17 #include "xfs_log.h"
  18 #include "xfs_log_priv.h"
  19 #include "xfs_trace.h"
  20 #include "xfs_sysfs.h"
  21 #include "xfs_sb.h"
  22 #include "xfs_health.h"
  23
  24 kmem_zone_t     *xfs_log_ticket_zone;
  25
  26 /* Local miscellaneous function prototypes */
  27 STATIC int
  28 xlog_commit_record(
  29         struct xlog             *log,
  30         struct xlog_ticket      *ticket,
  31         struct xlog_in_core     **iclog,
  32         xfs_lsn_t               *commitlsnp);
  33
  34 STATIC struct xlog *
  35 xlog_alloc_log(
  36         struct xfs_mount        *mp,
  37         struct xfs_buftarg      *log_target,
  38         xfs_daddr_t             blk_offset,
  39         int                     num_bblks);
  40 STATIC int
  41 xlog_space_left(
  42         struct xlog             *log,
  43         atomic64_t              *head);
  44 STATIC void
  45 xlog_dealloc_log(
  46         struct xlog             *log);
  47
  48 /* local state machine functions */
  49 STATIC void xlog_state_done_syncing(
  50         struct xlog_in_core     *iclog,
  51         bool                    aborted);
  52 STATIC int
  53 xlog_state_get_iclog_space(
  54         struct xlog             *log,
  55         int                     len,
  56         struct xlog_in_core     **iclog,
  57         struct xlog_ticket      *ticket,
  58         int                     *continued_write,
  59         int                     *logoffsetp);
  60 STATIC void
  61 xlog_state_switch_iclogs(
  62         struct xlog             *log,
  63         struct xlog_in_core     *iclog,
  64         int                     eventual_size);
  65 STATIC void
  66 xlog_state_want_sync(
  67         struct xlog             *log,
  68         struct xlog_in_core     *iclog);
  69
  70 STATIC void
  71 xlog_grant_push_ail(
  72         struct xlog             *log,
  73         int                     need_bytes);
  74 STATIC void
  75 xlog_regrant_reserve_log_space(
  76         struct xlog             *log,
  77         struct xlog_ticket      *ticket);
  78 STATIC void
  79 xlog_ungrant_log_space(
  80         struct xlog             *log,
  81         struct xlog_ticket      *ticket);
  82 STATIC void
  83 xlog_sync(
  84         struct xlog             *log,
  85         struct xlog_in_core     *iclog);
  86 #if defined(DEBUG)
  87 STATIC void
  88 xlog_verify_dest_ptr(
  89         struct xlog             *log,
  90         void                    *ptr);
  91 STATIC void
  92 xlog_verify_grant_tail(
  93         struct xlog *log);
  94 STATIC void
  95 xlog_verify_iclog(
  96         struct xlog             *log,
  97         struct xlog_in_core     *iclog,
  98         int                     count);
  99 STATIC void
 100 xlog_verify_tail_lsn(
 101         struct xlog             *log,
 102         struct xlog_in_core     *iclog,
 103         xfs_lsn_t               tail_lsn);
 104 #else
 105 #define xlog_verify_dest_ptr(a,b)
 106 #define xlog_verify_grant_tail(a)
 107 #define xlog_verify_iclog(a,b,c)
 108 #define xlog_verify_tail_lsn(a,b,c)
 109 #endif
 110
 111 STATIC int
 112 xlog_iclogs_empty(
 113         struct xlog             *log);
 114
 115 static void
 116 xlog_grant_sub_space(
 117         struct xlog             *log,
 118         atomic64_t              *head,
 119         int                     bytes)
 120 {
 121         int64_t head_val = atomic64_read(head);
 122         int64_t new, old;
 123
 124         do {
 125                 int     cycle, space;
 126
 127                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 128
 129                 space -= bytes;
 130                 if (space < 0) {
 131                         space += log->l_logsize;
 132                         cycle--;
 133                 }
 134
 135                 old = head_val;
 136                 new = xlog_assign_grant_head_val(cycle, space);
 137                 head_val = atomic64_cmpxchg(head, old, new);
 138         } while (head_val != old);
 139 }
 140
 141 static void
 142 xlog_grant_add_space(
 143         struct xlog             *log,
 144         atomic64_t              *head,
 145         int                     bytes)
 146 {
 147         int64_t head_val = atomic64_read(head);
 148         int64_t new, old;
 149
 150         do {
 151                 int             tmp;
 152                 int             cycle, space;
 153
 154                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 155
 156                 tmp = log->l_logsize - space;
 157                 if (tmp > bytes)
 158                         space += bytes;
 159                 else {
 160                         space = bytes - tmp;
 161                         cycle++;
 162                 }
 163
 164                 old = head_val;
 165                 new = xlog_assign_grant_head_val(cycle, space);
 166                 head_val = atomic64_cmpxchg(head, old, new);
 167         } while (head_val != old);
 168 }
 169
 170 STATIC void
 171 xlog_grant_head_init(
 172         struct xlog_grant_head  *head)
 173 {
 174         xlog_assign_grant_head(&head->grant, 1, 0);
 175         INIT_LIST_HEAD(&head->waiters);
 176         spin_lock_init(&head->lock);
 177 }
 178
 179 STATIC void
 180 xlog_grant_head_wake_all(
 181         struct xlog_grant_head  *head)
 182 {
 183         struct xlog_ticket      *tic;
 184
 185         spin_lock(&head->lock);
 186         list_for_each_entry(tic, &head->waiters, t_queue)
 187                 wake_up_process(tic->t_task);
 188         spin_unlock(&head->lock);
 189 }
 190
 191 static inline int
 192 xlog_ticket_reservation(
 193         struct xlog             *log,
 194         struct xlog_grant_head  *head,
 195         struct xlog_ticket      *tic)
 196 {
 197         if (head == &log->l_write_head) {
 198                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 199                 return tic->t_unit_res;
 200         } else {
 201                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 202                         return tic->t_unit_res * tic->t_cnt;
 203                 else
 204                         return tic->t_unit_res;
 205         }
 206 }
 207
 208 STATIC bool
 209 xlog_grant_head_wake(
 210         struct xlog             *log,
 211         struct xlog_grant_head  *head,
 212         int                     *free_bytes)
 213 {
 214         struct xlog_ticket      *tic;
 215         int                     need_bytes;
 216         bool                    woken_task = false;
 217
 218         list_for_each_entry(tic, &head->waiters, t_queue) {
 219
 220                 /*
 221                  * There is a chance that the size of the CIL checkpoints in
 222                  * progress at the last AIL push target calculation resulted in
 223                  * limiting the target to the log head (l_last_sync_lsn) at the
 224                  * time. This may not reflect where the log head is now as the
 225                  * CIL checkpoints may have completed.
 226                  *
 227                  * Hence when we are woken here, it may be that the head of the
 228                  * log that has moved rather than the tail. As the tail didn't
 229                  * move, there still won't be space available for the
 230                  * reservation we require.  However, if the AIL has already
 231                  * pushed to the target defined by the old log head location, we
 232                  * will hang here waiting for something else to update the AIL
 233                  * push target.
 234                  *
 235                  * Therefore, if there isn't space to wake the first waiter on
 236                  * the grant head, we need to push the AIL again to ensure the
 237                  * target reflects both the current log tail and log head
 238                  * position before we wait for the tail to move again.
 239                  */
 240
 241                 need_bytes = xlog_ticket_reservation(log, head, tic);
 242                 if (*free_bytes < need_bytes) {
 243                         if (!woken_task)
 244                                 xlog_grant_push_ail(log, need_bytes);
 245                         return false;
 246                 }
 247
 248                 *free_bytes -= need_bytes;
 249                 trace_xfs_log_grant_wake_up(log, tic);
 250                 wake_up_process(tic->t_task);
 251                 woken_task = true;
 252         }
 253
 254         return true;
 255 }
 256
 257 STATIC int
 258 xlog_grant_head_wait(
 259         struct xlog             *log,
 260         struct xlog_grant_head  *head,
 261         struct xlog_ticket      *tic,
 262         int                     need_bytes) __releases(&head->lock)
 263                                             __acquires(&head->lock)
 264 {
 265         list_add_tail(&tic->t_queue, &head->waiters);
 266
 267         do {
 268                 if (XLOG_FORCED_SHUTDOWN(log))
 269                         goto shutdown;
 270                 xlog_grant_push_ail(log, need_bytes);
 271
 272                 __set_current_state(TASK_UNINTERRUPTIBLE);
 273                 spin_unlock(&head->lock);
 274
 275                 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
 276
 277                 trace_xfs_log_grant_sleep(log, tic);
 278                 schedule();
 279                 trace_xfs_log_grant_wake(log, tic);
 280
 281                 spin_lock(&head->lock);
 282                 if (XLOG_FORCED_SHUTDOWN(log))
 283                         goto shutdown;
 284         } while (xlog_space_left(log, &head->grant) < need_bytes);
 285
 286         list_del_init(&tic->t_queue);
 287         return 0;
 288 shutdown:
 289         list_del_init(&tic->t_queue);
 290         return -EIO;
 291 }
 292
 293 /*
 294  * Atomically get the log space required for a log ticket.
 295  *
 296  * Once a ticket gets put onto head->waiters, it will only return after the
 297  * needed reservation is satisfied.
 298  *
 299  * This function is structured so that it has a lock free fast path. This is
 300  * necessary because every new transaction reservation will come through this
 301  * path. Hence any lock will be globally hot if we take it unconditionally on
 302  * every pass.
 303  *
 304  * As tickets are only ever moved on and off head->waiters under head->lock, we
 305  * only need to take that lock if we are going to add the ticket to the queue
 306  * and sleep. We can avoid taking the lock if the ticket was never added to
 307  * head->waiters because the t_queue list head will be empty and we hold the
 308  * only reference to it so it can safely be checked unlocked.
 309  */
 310 STATIC int
 311 xlog_grant_head_check(
 312         struct xlog             *log,
 313         struct xlog_grant_head  *head,
 314         struct xlog_ticket      *tic,
 315         int                     *need_bytes)
 316 {
 317         int                     free_bytes;
 318         int                     error = 0;
 319
 320         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 321
 322         /*
 323          * If there are other waiters on the queue then give them a chance at
 324          * logspace before us.  Wake up the first waiters, if we do not wake
 325          * up all the waiters then go to sleep waiting for more free space,
 326          * otherwise try to get some space for this transaction.
 327          */
 328         *need_bytes = xlog_ticket_reservation(log, head, tic);
 329         free_bytes = xlog_space_left(log, &head->grant);
 330         if (!list_empty_careful(&head->waiters)) {
 331                 spin_lock(&head->lock);
 332                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 333                     free_bytes < *need_bytes) {
 334                         error = xlog_grant_head_wait(log, head, tic,
 335                                                      *need_bytes);
 336                 }
 337                 spin_unlock(&head->lock);
 338         } else if (free_bytes < *need_bytes) {
 339                 spin_lock(&head->lock);
 340                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 341                 spin_unlock(&head->lock);
 342         }
 343
 344         return error;
 345 }
 346
 347 static void
 348 xlog_tic_reset_res(xlog_ticket_t *tic)
 349 {
 350         tic->t_res_num = 0;
 351         tic->t_res_arr_sum = 0;
 352         tic->t_res_num_ophdrs = 0;
 353 }
 354
 355 static void
 356 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 357 {
 358         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 359                 /* add to overflow and start again */
 360                 tic->t_res_o_flow += tic->t_res_arr_sum;
 361                 tic->t_res_num = 0;
 362                 tic->t_res_arr_sum = 0;
 363         }
 364
 365         tic->t_res_arr[tic->t_res_num].r_len = len;
 366         tic->t_res_arr[tic->t_res_num].r_type = type;
 367         tic->t_res_arr_sum += len;
 368         tic->t_res_num++;
 369 }
 370
 371 /*
 372  * Replenish the byte reservation required by moving the grant write head.
 373  */
 374 int
 375 xfs_log_regrant(
 376         struct xfs_mount        *mp,
 377         struct xlog_ticket      *tic)
 378 {
 379         struct xlog             *log = mp->m_log;
 380         int                     need_bytes;
 381         int                     error = 0;
 382
 383         if (XLOG_FORCED_SHUTDOWN(log))
 384                 return -EIO;
 385
 386         XFS_STATS_INC(mp, xs_try_logspace);
 387
 388         /*
 389          * This is a new transaction on the ticket, so we need to change the
 390          * transaction ID so that the next transaction has a different TID in
 391          * the log. Just add one to the existing tid so that we can see chains
 392          * of rolling transactions in the log easily.
 393          */
 394         tic->t_tid++;
 395
 396         xlog_grant_push_ail(log, tic->t_unit_res);
 397
 398         tic->t_curr_res = tic->t_unit_res;
 399         xlog_tic_reset_res(tic);
 400
 401         if (tic->t_cnt > 0)
 402                 return 0;
 403
 404         trace_xfs_log_regrant(log, tic);
 405
 406         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 407                                       &need_bytes);
 408         if (error)
 409                 goto out_error;
 410
 411         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 412         trace_xfs_log_regrant_exit(log, tic);
 413         xlog_verify_grant_tail(log);
 414         return 0;
 415
 416 out_error:
 417         /*
 418          * If we are failing, make sure the ticket doesn't have any current
 419          * reservations.  We don't want to add this back when the ticket/
 420          * transaction gets cancelled.
 421          */
 422         tic->t_curr_res = 0;
 423         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 424         return error;
 425 }
 426
 427 /*
 428  * Reserve log space and return a ticket corresponding to the reservation.
 429  *
 430  * Each reservation is going to reserve extra space for a log record header.
 431  * When writes happen to the on-disk log, we don't subtract the length of the
 432  * log record header from any reservation.  By wasting space in each
 433  * reservation, we prevent over allocation problems.
 434  */
 435 int
 436 xfs_log_reserve(
 437         struct xfs_mount        *mp,
 438         int                     unit_bytes,
 439         int                     cnt,
 440         struct xlog_ticket      **ticp,
 441         uint8_t                 client,
 442         bool                    permanent)
 443 {
 444         struct xlog             *log = mp->m_log;
 445         struct xlog_ticket      *tic;
 446         int                     need_bytes;
 447         int                     error = 0;
 448
 449         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 450
 451         if (XLOG_FORCED_SHUTDOWN(log))
 452                 return -EIO;
 453
 454         XFS_STATS_INC(mp, xs_try_logspace);
 455
 456         ASSERT(*ticp == NULL);
 457         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent, 0);
 458         *ticp = tic;
 459
 460         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 461                                             : tic->t_unit_res);
 462
 463         trace_xfs_log_reserve(log, tic);
 464
 465         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 466                                       &need_bytes);
 467         if (error)
 468                 goto out_error;
 469
 470         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 471         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 472         trace_xfs_log_reserve_exit(log, tic);
 473         xlog_verify_grant_tail(log);
 474         return 0;
 475
 476 out_error:
 477         /*
 478          * If we are failing, make sure the ticket doesn't have any current
 479          * reservations.  We don't want to add this back when the ticket/
 480          * transaction gets cancelled.
 481          */
 482         tic->t_curr_res = 0;
 483         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 484         return error;
 485 }
 486
 487
 488 /*
 489  * NOTES:
 490  *
 491  *      1. currblock field gets updated at startup and after in-core logs
 492  *              marked as with WANT_SYNC.
 493  */
 494
 495 /*
 496  * This routine is called when a user of a log manager ticket is done with
 497  * the reservation.  If the ticket was ever used, then a commit record for
 498  * the associated transaction is written out as a log operation header with
 499  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 500  * a given ticket.  If the ticket was one with a permanent reservation, then
 501  * a few operations are done differently.  Permanent reservation tickets by
 502  * default don't release the reservation.  They just commit the current
 503  * transaction with the belief that the reservation is still needed.  A flag
 504  * must be passed in before permanent reservations are actually released.
 505  * When these type of tickets are not released, they need to be set into
 506  * the inited state again.  By doing this, a start record will be written
 507  * out when the next write occurs.
 508  */
 509 xfs_lsn_t
 510 xfs_log_done(
 511         struct xfs_mount        *mp,
 512         struct xlog_ticket      *ticket,
 513         struct xlog_in_core     **iclog,
 514         bool                    regrant)
 515 {
 516         struct xlog             *log = mp->m_log;
 517         xfs_lsn_t               lsn = 0;
 518
 519         if (XLOG_FORCED_SHUTDOWN(log) ||
 520             /*
 521              * If nothing was ever written, don't write out commit record.
 522              * If we get an error, just continue and give back the log ticket.
 523              */
 524             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 525              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 526                 lsn = (xfs_lsn_t) -1;
 527                 regrant = false;
 528         }
 529
 530
 531         if (!regrant) {
 532                 trace_xfs_log_done_nonperm(log, ticket);
 533
 534                 /*
 535                  * Release ticket if not permanent reservation or a specific
 536                  * request has been made to release a permanent reservation.
 537                  */
 538                 xlog_ungrant_log_space(log, ticket);
 539         } else {
 540                 trace_xfs_log_done_perm(log, ticket);
 541
 542                 xlog_regrant_reserve_log_space(log, ticket);
 543                 /* If this ticket was a permanent reservation and we aren't
 544                  * trying to release it, reset the inited flags; so next time
 545                  * we write, a start record will be written out.
 546                  */
 547                 ticket->t_flags |= XLOG_TIC_INITED;
 548         }
 549
 550         xfs_log_ticket_put(ticket);
 551         return lsn;
 552 }
 553
 554 static bool
 555 __xlog_state_release_iclog(
 556         struct xlog             *log,
 557         struct xlog_in_core     *iclog)
 558 {
 559         lockdep_assert_held(&log->l_icloglock);
 560
 561         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
 562                 /* update tail before writing to iclog */
 563                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
 564
 565                 iclog->ic_state = XLOG_STATE_SYNCING;
 566                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
 567                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
 568                 /* cycle incremented when incrementing curr_block */
 569                 return true;
 570         }
 571
 572         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
 573         return false;
 574 }
 575
 576 /*
 577  * Flush iclog to disk if this is the last reference to the given iclog and the
 578  * it is in the WANT_SYNC state.
 579  */
 580 static int
 581 xlog_state_release_iclog(
 582         struct xlog             *log,
 583         struct xlog_in_core     *iclog)
 584 {
 585         lockdep_assert_held(&log->l_icloglock);
 586
 587         if (iclog->ic_state == XLOG_STATE_IOERROR)
 588                 return -EIO;
 589
 590         if (atomic_dec_and_test(&iclog->ic_refcnt) &&
 591             __xlog_state_release_iclog(log, iclog)) {
 592                 spin_unlock(&log->l_icloglock);
 593                 xlog_sync(log, iclog);
 594                 spin_lock(&log->l_icloglock);
 595         }
 596
 597         return 0;
 598 }
 599
 600 int
 601 xfs_log_release_iclog(
 602         struct xfs_mount        *mp,
 603         struct xlog_in_core     *iclog)
 604 {
 605         struct xlog             *log = mp->m_log;
 606         bool                    sync;
 607
 608         if (iclog->ic_state == XLOG_STATE_IOERROR) {
 609                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 610                 return -EIO;
 611         }
 612
 613         if (atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock)) {
 614                 sync = __xlog_state_release_iclog(log, iclog);
 615                 spin_unlock(&log->l_icloglock);
 616                 if (sync)
 617                         xlog_sync(log, iclog);
 618         }
 619         return 0;
 620 }
 621
 622 /*
 623  * Mount a log filesystem
 624  *
 625  * mp           - ubiquitous xfs mount point structure
 626  * log_target   - buftarg of on-disk log device
 627  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 628  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 629  *
 630  * Return error or zero.
 631  */
 632 int
 633 xfs_log_mount(
 634         xfs_mount_t     *mp,
 635         xfs_buftarg_t   *log_target,
 636         xfs_daddr_t     blk_offset,
 637         int             num_bblks)
 638 {
 639         bool            fatal = xfs_sb_version_hascrc(&mp->m_sb);
 640         int             error = 0;
 641         int             min_logfsbs;
 642
 643         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 644                 xfs_notice(mp, "Mounting V%d Filesystem",
 645                            XFS_SB_VERSION_NUM(&mp->m_sb));
 646         } else {
 647                 xfs_notice(mp,
 648 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
 649                            XFS_SB_VERSION_NUM(&mp->m_sb));
 650                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 651         }
 652
 653         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 654         if (IS_ERR(mp->m_log)) {
 655                 error = PTR_ERR(mp->m_log);
 656                 goto out;
 657         }
 658
 659         /*
 660          * Validate the given log space and drop a critical message via syslog
 661          * if the log size is too small that would lead to some unexpected
 662          * situations in transaction log space reservation stage.
 663          *
 664          * Note: we can't just reject the mount if the validation fails.  This
 665          * would mean that people would have to downgrade their kernel just to
 666          * remedy the situation as there is no way to grow the log (short of
 667          * black magic surgery with xfs_db).
 668          *
 669          * We can, however, reject mounts for CRC format filesystems, as the
 670          * mkfs binary being used to make the filesystem should never create a
 671          * filesystem with a log that is too small.
 672          */
 673         min_logfsbs = xfs_log_calc_minimum_size(mp);
 674
 675         if (mp->m_sb.sb_logblocks < min_logfsbs) {
 676                 xfs_warn(mp,
 677                 "Log size %d blocks too small, minimum size is %d blocks",
 678                          mp->m_sb.sb_logblocks, min_logfsbs);
 679                 error = -EINVAL;
 680         } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
 681                 xfs_warn(mp,
 682                 "Log size %d blocks too large, maximum size is %lld blocks",
 683                          mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
 684                 error = -EINVAL;
 685         } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
 686                 xfs_warn(mp,
 687                 "log size %lld bytes too large, maximum size is %lld bytes",
 688                          XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
 689                          XFS_MAX_LOG_BYTES);
 690                 error = -EINVAL;
 691         } else if (mp->m_sb.sb_logsunit > 1 &&
 692                    mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
 693                 xfs_warn(mp,
 694                 "log stripe unit %u bytes must be a multiple of block size",
 695                          mp->m_sb.sb_logsunit);
 696                 error = -EINVAL;
 697                 fatal = true;
 698         }
 699         if (error) {
 700                 /*
 701                  * Log check errors are always fatal on v5; or whenever bad
 702                  * metadata leads to a crash.
 703                  */
 704                 if (fatal) {
 705                         xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 706                         ASSERT(0);
 707                         goto out_free_log;
 708                 }
 709                 xfs_crit(mp, "Log size out of supported range.");
 710                 xfs_crit(mp,
 711 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
 712         }
 713
 714         /*
 715          * Initialize the AIL now we have a log.
 716          */
 717         error = xfs_trans_ail_init(mp);
 718         if (error) {
 719                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 720                 goto out_free_log;
 721         }
 722         mp->m_log->l_ailp = mp->m_ail;
 723
 724         /*
 725          * skip log recovery on a norecovery mount.  pretend it all
 726          * just worked.
 727          */
 728         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 729                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 730
 731                 if (readonly)
 732                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 733
 734                 error = xlog_recover(mp->m_log);
 735
 736                 if (readonly)
 737                         mp->m_flags |= XFS_MOUNT_RDONLY;
 738                 if (error) {
 739                         xfs_warn(mp, "log mount/recovery failed: error %d",
 740                                 error);
 741                         xlog_recover_cancel(mp->m_log);
 742                         goto out_destroy_ail;
 743                 }
 744         }
 745
 746         error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 747                                "log");
 748         if (error)
 749                 goto out_destroy_ail;
 750
 751         /* Normal transactions can now occur */
 752         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 753
 754         /*
 755          * Now the log has been fully initialised and we know were our
 756          * space grant counters are, we can initialise the permanent ticket
 757          * needed for delayed logging to work.
 758          */
 759         xlog_cil_init_post_recovery(mp->m_log);
 760
 761         return 0;
 762
 763 out_destroy_ail:
 764         xfs_trans_ail_destroy(mp);
 765 out_free_log:
 766         xlog_dealloc_log(mp->m_log);
 767 out:
 768         return error;
 769 }
 770
 771 /*
 772  * Finish the recovery of the file system.  This is separate from the
 773  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 774  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 775  * here.
 776  *
 777  * If we finish recovery successfully, start the background log work. If we are
 778  * not doing recovery, then we have a RO filesystem and we don't need to start
 779  * it.
 780  */
 781 int
 782 xfs_log_mount_finish(
 783         struct xfs_mount        *mp)
 784 {
 785         int     error = 0;
 786         bool    readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 787         bool    recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
 788
 789         if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
 790                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 791                 return 0;
 792         } else if (readonly) {
 793                 /* Allow unlinked processing to proceed */
 794                 mp->m_flags &= ~XFS_MOUNT_RDONLY;
 795         }
 796
 797         /*
 798          * During the second phase of log recovery, we need iget and
 799          * iput to behave like they do for an active filesystem.
 800          * xfs_fs_drop_inode needs to be able to prevent the deletion
 801          * of inodes before we're done replaying log items on those
 802          * inodes.  Turn it off immediately after recovery finishes
 803          * so that we don't leak the quota inodes if subsequent mount
 804          * activities fail.
 805          *
 806          * We let all inodes involved in redo item processing end up on
 807          * the LRU instead of being evicted immediately so that if we do
 808          * something to an unlinked inode, the irele won't cause
 809          * premature truncation and freeing of the inode, which results
 810          * in log recovery failure.  We have to evict the unreferenced
 811          * lru inodes after clearing SB_ACTIVE because we don't
 812          * otherwise clean up the lru if there's a subsequent failure in
 813          * xfs_mountfs, which leads to us leaking the inodes if nothing
 814          * else (e.g. quotacheck) references the inodes before the
 815          * mount failure occurs.
 816          */
 817         mp->m_super->s_flags |= SB_ACTIVE;
 818         error = xlog_recover_finish(mp->m_log);
 819         if (!error)
 820                 xfs_log_work_queue(mp);
 821         mp->m_super->s_flags &= ~SB_ACTIVE;
 822         evict_inodes(mp->m_super);
 823
 824         /*
 825          * Drain the buffer LRU after log recovery. This is required for v4
 826          * filesystems to avoid leaving around buffers with NULL verifier ops,
 827          * but we do it unconditionally to make sure we're always in a clean
 828          * cache state after mount.
 829          *
 830          * Don't push in the error case because the AIL may have pending intents
 831          * that aren't removed until recovery is cancelled.
 832          */
 833         if (!error && recovered) {
 834                 xfs_log_force(mp, XFS_LOG_SYNC);
 835                 xfs_ail_push_all_sync(mp->m_ail);
 836         }
 837         xfs_wait_buftarg(mp->m_ddev_targp);
 838
 839         if (readonly)
 840                 mp->m_flags |= XFS_MOUNT_RDONLY;
 841
 842         return error;
 843 }
 844
 845 /*
 846  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
 847  * the log.
 848  */
 849 void
 850 xfs_log_mount_cancel(
 851         struct xfs_mount        *mp)
 852 {
 853         xlog_recover_cancel(mp->m_log);
 854         xfs_log_unmount(mp);
 855 }
 856
 857 /*
 858  * Final log writes as part of unmount.
 859  *
 860  * Mark the filesystem clean as unmount happens.  Note that during relocation
 861  * this routine needs to be executed as part of source-bag while the
 862  * deallocation must not be done until source-end.
 863  */
 864
 865 /* Actually write the unmount record to disk. */
 866 static void
 867 xfs_log_write_unmount_record(
 868         struct xfs_mount        *mp)
 869 {
 870         /* the data section must be 32 bit size aligned */
 871         struct xfs_unmount_log_format magic = {
 872                 .magic = XLOG_UNMOUNT_TYPE,
 873         };
 874         struct xfs_log_iovec reg = {
 875                 .i_addr = &magic,
 876                 .i_len = sizeof(magic),
 877                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 878         };
 879         struct xfs_log_vec vec = {
 880                 .lv_niovecs = 1,
 881                 .lv_iovecp = &reg,
 882         };
 883         struct xlog             *log = mp->m_log;
 884         struct xlog_in_core     *iclog;
 885         struct xlog_ticket      *tic = NULL;
 886         xfs_lsn_t               lsn;
 887         uint                    flags = XLOG_UNMOUNT_TRANS;
 888         int                     error;
 889
 890         error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
 891         if (error)
 892                 goto out_err;
 893
 894         /*
 895          * If we think the summary counters are bad, clear the unmount header
 896          * flag in the unmount record so that the summary counters will be
 897          * recalculated during log recovery at next mount.  Refer to
 898          * xlog_check_unmount_rec for more details.
 899          */
 900         if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
 901                         XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
 902                 xfs_alert(mp, "%s: will fix summary counters at next mount",
 903                                 __func__);
 904                 flags &= ~XLOG_UNMOUNT_TRANS;
 905         }
 906
 907         /* remove inited flag, and account for space used */
 908         tic->t_flags = 0;
 909         tic->t_curr_res -= sizeof(magic);
 910         error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
 911         /*
 912          * At this point, we're umounting anyway, so there's no point in
 913          * transitioning log state to IOERROR. Just continue...
 914          */
 915 out_err:
 916         if (error)
 917                 xfs_alert(mp, "%s: unmount record failed", __func__);
 918
 919         spin_lock(&log->l_icloglock);
 920         iclog = log->l_iclog;
 921         atomic_inc(&iclog->ic_refcnt);
 922         xlog_state_want_sync(log, iclog);
 923         error = xlog_state_release_iclog(log, iclog);
 924         switch (iclog->ic_state) {
 925         default:
 926                 if (!XLOG_FORCED_SHUTDOWN(log)) {
 927                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
 928                         break;
 929                 }
 930                 /* fall through */
 931         case XLOG_STATE_ACTIVE:
 932         case XLOG_STATE_DIRTY:
 933                 spin_unlock(&log->l_icloglock);
 934                 break;
 935         }
 936
 937         if (tic) {
 938                 trace_xfs_log_umount_write(log, tic);
 939                 xlog_ungrant_log_space(log, tic);
 940                 xfs_log_ticket_put(tic);
 941         }
 942 }
 943
 944 /*
 945  * Unmount record used to have a string "Unmount filesystem--" in the
 946  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 947  * We just write the magic number now since that particular field isn't
 948  * currently architecture converted and "Unmount" is a bit foo.
 949  * As far as I know, there weren't any dependencies on the old behaviour.
 950  */
 951
 952 static int
 953 xfs_log_unmount_write(xfs_mount_t *mp)
 954 {
 955         struct xlog      *log = mp->m_log;
 956         xlog_in_core_t   *iclog;
 957 #ifdef DEBUG
 958         xlog_in_core_t   *first_iclog;
 959 #endif
 960         int              error;
 961
 962         /*
 963          * Don't write out unmount record on norecovery mounts or ro devices.
 964          * Or, if we are doing a forced umount (typically because of IO errors).
 965          */
 966         if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
 967             xfs_readonly_buftarg(log->l_targ)) {
 968                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 969                 return 0;
 970         }
 971
 972         error = xfs_log_force(mp, XFS_LOG_SYNC);
 973         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 974
 975 #ifdef DEBUG
 976         first_iclog = iclog = log->l_iclog;
 977         do {
 978                 if (iclog->ic_state != XLOG_STATE_IOERROR) {
 979                         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
 980                         ASSERT(iclog->ic_offset == 0);
 981                 }
 982                 iclog = iclog->ic_next;
 983         } while (iclog != first_iclog);
 984 #endif
 985         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 986                 xfs_log_write_unmount_record(mp);
 987         } else {
 988                 /*
 989                  * We're already in forced_shutdown mode, couldn't
 990                  * even attempt to write out the unmount transaction.
 991                  *
 992                  * Go through the motions of sync'ing and releasing
 993                  * the iclog, even though no I/O will actually happen,
 994                  * we need to wait for other log I/Os that may already
 995                  * be in progress.  Do this as a separate section of
 996                  * code so we'll know if we ever get stuck here that
 997                  * we're in this odd situation of trying to unmount
 998                  * a file system that went into forced_shutdown as
 999                  * the result of an unmount..
1000                  */
1001                 spin_lock(&log->l_icloglock);
1002                 iclog = log->l_iclog;
1003                 atomic_inc(&iclog->ic_refcnt);
1004                 xlog_state_want_sync(log, iclog);
1005                 error =  xlog_state_release_iclog(log, iclog);
1006                 switch (iclog->ic_state) {
1007                 case XLOG_STATE_ACTIVE:
1008                 case XLOG_STATE_DIRTY:
1009                 case XLOG_STATE_IOERROR:
1010                         spin_unlock(&log->l_icloglock);
1011                         break;
1012                 default:
1013                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
1014                         break;
1015                 }
1016         }
1017
1018         return error;
1019 }       /* xfs_log_unmount_write */
1020
1021 /*
1022  * Empty the log for unmount/freeze.
1023  *
1024  * To do this, we first need to shut down the background log work so it is not
1025  * trying to cover the log as we clean up. We then need to unpin all objects in
1026  * the log so we can then flush them out. Once they have completed their IO and
1027  * run the callbacks removing themselves from the AIL, we can write the unmount
1028  * record.
1029  */
1030 void
1031 xfs_log_quiesce(
1032         struct xfs_mount        *mp)
1033 {
1034         cancel_delayed_work_sync(&mp->m_log->l_work);
1035         xfs_log_force(mp, XFS_LOG_SYNC);
1036
1037         /*
1038          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1039          * will push it, xfs_wait_buftarg() will not wait for it. Further,
1040          * xfs_buf_iowait() cannot be used because it was pushed with the
1041          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1042          * the IO to complete.
1043          */
1044         xfs_ail_push_all_sync(mp->m_ail);
1045         xfs_wait_buftarg(mp->m_ddev_targp);
1046         xfs_buf_lock(mp->m_sb_bp);
1047         xfs_buf_unlock(mp->m_sb_bp);
1048
1049         xfs_log_unmount_write(mp);
1050 }
1051
1052 /*
1053  * Shut down and release the AIL and Log.
1054  *
1055  * During unmount, we need to ensure we flush all the dirty metadata objects
1056  * from the AIL so that the log is empty before we write the unmount record to
1057  * the log. Once this is done, we can tear down the AIL and the log.
1058  */
1059 void
1060 xfs_log_unmount(
1061         struct xfs_mount        *mp)
1062 {
1063         xfs_log_quiesce(mp);
1064
1065         xfs_trans_ail_destroy(mp);
1066
1067         xfs_sysfs_del(&mp->m_log->l_kobj);
1068
1069         xlog_dealloc_log(mp->m_log);
1070 }
1071
1072 void
1073 xfs_log_item_init(
1074         struct xfs_mount        *mp,
1075         struct xfs_log_item     *item,
1076         int                     type,
1077         const struct xfs_item_ops *ops)
1078 {
1079         item->li_mountp = mp;
1080         item->li_ailp = mp->m_ail;
1081         item->li_type = type;
1082         item->li_ops = ops;
1083         item->li_lv = NULL;
1084
1085         INIT_LIST_HEAD(&item->li_ail);
1086         INIT_LIST_HEAD(&item->li_cil);
1087         INIT_LIST_HEAD(&item->li_bio_list);
1088         INIT_LIST_HEAD(&item->li_trans);
1089 }
1090
1091 /*
1092  * Wake up processes waiting for log space after we have moved the log tail.
1093  */
1094 void
1095 xfs_log_space_wake(
1096         struct xfs_mount        *mp)
1097 {
1098         struct xlog             *log = mp->m_log;
1099         int                     free_bytes;
1100
1101         if (XLOG_FORCED_SHUTDOWN(log))
1102                 return;
1103
1104         if (!list_empty_careful(&log->l_write_head.waiters)) {
1105                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1106
1107                 spin_lock(&log->l_write_head.lock);
1108                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1109                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1110                 spin_unlock(&log->l_write_head.lock);
1111         }
1112
1113         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1114                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1115
1116                 spin_lock(&log->l_reserve_head.lock);
1117                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1118                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1119                 spin_unlock(&log->l_reserve_head.lock);
1120         }
1121 }
1122
1123 /*
1124  * Determine if we have a transaction that has gone to disk that needs to be
1125  * covered. To begin the transition to the idle state firstly the log needs to
1126  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1127  * we start attempting to cover the log.
1128  *
1129  * Only if we are then in a state where covering is needed, the caller is
1130  * informed that dummy transactions are required to move the log into the idle
1131  * state.
1132  *
1133  * If there are any items in the AIl or CIL, then we do not want to attempt to
1134  * cover the log as we may be in a situation where there isn't log space
1135  * available to run a dummy transaction and this can lead to deadlocks when the
1136  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1137  * there's no point in running a dummy transaction at this point because we
1138  * can't start trying to idle the log until both the CIL and AIL are empty.
1139  */
1140 static int
1141 xfs_log_need_covered(xfs_mount_t *mp)
1142 {
1143         struct xlog     *log = mp->m_log;
1144         int             needed = 0;
1145
1146         if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1147                 return 0;
1148
1149         if (!xlog_cil_empty(log))
1150                 return 0;
1151
1152         spin_lock(&log->l_icloglock);
1153         switch (log->l_covered_state) {
1154         case XLOG_STATE_COVER_DONE:
1155         case XLOG_STATE_COVER_DONE2:
1156         case XLOG_STATE_COVER_IDLE:
1157                 break;
1158         case XLOG_STATE_COVER_NEED:
1159         case XLOG_STATE_COVER_NEED2:
1160                 if (xfs_ail_min_lsn(log->l_ailp))
1161                         break;
1162                 if (!xlog_iclogs_empty(log))
1163                         break;
1164
1165                 needed = 1;
1166                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1167                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1168                 else
1169                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1170                 break;
1171         default:
1172                 needed = 1;
1173                 break;
1174         }
1175         spin_unlock(&log->l_icloglock);
1176         return needed;
1177 }
1178
1179 /*
1180  * We may be holding the log iclog lock upon entering this routine.
1181  */
1182 xfs_lsn_t
1183 xlog_assign_tail_lsn_locked(
1184         struct xfs_mount        *mp)
1185 {
1186         struct xlog             *log = mp->m_log;
1187         struct xfs_log_item     *lip;
1188         xfs_lsn_t               tail_lsn;
1189
1190         assert_spin_locked(&mp->m_ail->ail_lock);
1191
1192         /*
1193          * To make sure we always have a valid LSN for the log tail we keep
1194          * track of the last LSN which was committed in log->l_last_sync_lsn,
1195          * and use that when the AIL was empty.
1196          */
1197         lip = xfs_ail_min(mp->m_ail);
1198         if (lip)
1199                 tail_lsn = lip->li_lsn;
1200         else
1201                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1202         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1203         atomic64_set(&log->l_tail_lsn, tail_lsn);
1204         return tail_lsn;
1205 }
1206
1207 xfs_lsn_t
1208 xlog_assign_tail_lsn(
1209         struct xfs_mount        *mp)
1210 {
1211         xfs_lsn_t               tail_lsn;
1212
1213         spin_lock(&mp->m_ail->ail_lock);
1214         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1215         spin_unlock(&mp->m_ail->ail_lock);
1216
1217         return tail_lsn;
1218 }
1219
1220 /*
1221  * Return the space in the log between the tail and the head.  The head
1222  * is passed in the cycle/bytes formal parms.  In the special case where
1223  * the reserve head has wrapped passed the tail, this calculation is no
1224  * longer valid.  In this case, just return 0 which means there is no space
1225  * in the log.  This works for all places where this function is called
1226  * with the reserve head.  Of course, if the write head were to ever
1227  * wrap the tail, we should blow up.  Rather than catch this case here,
1228  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1229  *
1230  * This code also handles the case where the reservation head is behind
1231  * the tail.  The details of this case are described below, but the end
1232  * result is that we return the size of the log as the amount of space left.
1233  */
1234 STATIC int
1235 xlog_space_left(
1236         struct xlog     *log,
1237         atomic64_t      *head)
1238 {
1239         int             free_bytes;
1240         int             tail_bytes;
1241         int             tail_cycle;
1242         int             head_cycle;
1243         int             head_bytes;
1244
1245         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1246         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1247         tail_bytes = BBTOB(tail_bytes);
1248         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1249                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1250         else if (tail_cycle + 1 < head_cycle)
1251                 return 0;
1252         else if (tail_cycle < head_cycle) {
1253                 ASSERT(tail_cycle == (head_cycle - 1));
1254                 free_bytes = tail_bytes - head_bytes;
1255         } else {
1256                 /*
1257                  * The reservation head is behind the tail.
1258                  * In this case we just want to return the size of the
1259                  * log as the amount of space left.
1260                  */
1261                 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1262                 xfs_alert(log->l_mp,
1263                           "  tail_cycle = %d, tail_bytes = %d",
1264                           tail_cycle, tail_bytes);
1265                 xfs_alert(log->l_mp,
1266                           "  GH   cycle = %d, GH   bytes = %d",
1267                           head_cycle, head_bytes);
1268                 ASSERT(0);
1269                 free_bytes = log->l_logsize;
1270         }
1271         return free_bytes;
1272 }
1273
1274
1275 static void
1276 xlog_ioend_work(
1277         struct work_struct      *work)
1278 {
1279         struct xlog_in_core     *iclog =
1280                 container_of(work, struct xlog_in_core, ic_end_io_work);
1281         struct xlog             *log = iclog->ic_log;
1282         bool                    aborted = false;
1283         int                     error;
1284
1285         error = blk_status_to_errno(iclog->ic_bio.bi_status);
1286 #ifdef DEBUG
1287         /* treat writes with injected CRC errors as failed */
1288         if (iclog->ic_fail_crc)
1289                 error = -EIO;
1290 #endif
1291
1292         /*
1293          * Race to shutdown the filesystem if we see an error.
1294          */
1295         if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1296                 xfs_alert(log->l_mp, "log I/O error %d", error);
1297                 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1298                 /*
1299                  * This flag will be propagated to the trans-committed
1300                  * callback routines to let them know that the log-commit
1301                  * didn't succeed.
1302                  */
1303                 aborted = true;
1304         } else if (iclog->ic_state == XLOG_STATE_IOERROR) {
1305                 aborted = true;
1306         }
1307
1308         xlog_state_done_syncing(iclog, aborted);
1309         bio_uninit(&iclog->ic_bio);
1310
1311         /*
1312          * Drop the lock to signal that we are done. Nothing references the
1313          * iclog after this, so an unmount waiting on this lock can now tear it
1314          * down safely. As such, it is unsafe to reference the iclog after the
1315          * unlock as we could race with it being freed.
1316          */
1317         up(&iclog->ic_sema);
1318 }
1319
1320 /*
1321  * Return size of each in-core log record buffer.
1322  *
1323  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1324  *
1325  * If the filesystem blocksize is too large, we may need to choose a
1326  * larger size since the directory code currently logs entire blocks.
1327  */
1328 STATIC void
1329 xlog_get_iclog_buffer_size(
1330         struct xfs_mount        *mp,
1331         struct xlog             *log)
1332 {
1333         if (mp->m_logbufs <= 0)
1334                 mp->m_logbufs = XLOG_MAX_ICLOGS;
1335         if (mp->m_logbsize <= 0)
1336                 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1337
1338         log->l_iclog_bufs = mp->m_logbufs;
1339         log->l_iclog_size = mp->m_logbsize;
1340
1341         /*
1342          * # headers = size / 32k - one header holds cycles from 32k of data.
1343          */
1344         log->l_iclog_heads =
1345                 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1346         log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1347 }
1348
1349 void
1350 xfs_log_work_queue(
1351         struct xfs_mount        *mp)
1352 {
1353         queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1354                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1355 }
1356
1357 /*
1358  * Every sync period we need to unpin all items in the AIL and push them to
1359  * disk. If there is nothing dirty, then we might need to cover the log to
1360  * indicate that the filesystem is idle.
1361  */
1362 static void
1363 xfs_log_worker(
1364         struct work_struct      *work)
1365 {
1366         struct xlog             *log = container_of(to_delayed_work(work),
1367                                                 struct xlog, l_work);
1368         struct xfs_mount        *mp = log->l_mp;
1369
1370         /* dgc: errors ignored - not fatal and nowhere to report them */
1371         if (xfs_log_need_covered(mp)) {
1372                 /*
1373                  * Dump a transaction into the log that contains no real change.
1374                  * This is needed to stamp the current tail LSN into the log
1375                  * during the covering operation.
1376                  *
1377                  * We cannot use an inode here for this - that will push dirty
1378                  * state back up into the VFS and then periodic inode flushing
1379                  * will prevent log covering from making progress. Hence we
1380                  * synchronously log the superblock instead to ensure the
1381                  * superblock is immediately unpinned and can be written back.
1382                  */
1383                 xfs_sync_sb(mp, true);
1384         } else
1385                 xfs_log_force(mp, 0);
1386
1387         /* start pushing all the metadata that is currently dirty */
1388         xfs_ail_push_all(mp->m_ail);
1389
1390         /* queue us up again */
1391         xfs_log_work_queue(mp);
1392 }
1393
1394 /*
1395  * This routine initializes some of the log structure for a given mount point.
1396  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1397  * some other stuff may be filled in too.
1398  */
1399 STATIC struct xlog *
1400 xlog_alloc_log(
1401         struct xfs_mount        *mp,
1402         struct xfs_buftarg      *log_target,
1403         xfs_daddr_t             blk_offset,
1404         int                     num_bblks)
1405 {
1406         struct xlog             *log;
1407         xlog_rec_header_t       *head;
1408         xlog_in_core_t          **iclogp;
1409         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1410         int                     i;
1411         int                     error = -ENOMEM;
1412         uint                    log2_size = 0;
1413
1414         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1415         if (!log) {
1416                 xfs_warn(mp, "Log allocation failed: No memory!");
1417                 goto out;
1418         }
1419
1420         log->l_mp          = mp;
1421         log->l_targ        = log_target;
1422         log->l_logsize     = BBTOB(num_bblks);
1423         log->l_logBBstart  = blk_offset;
1424         log->l_logBBsize   = num_bblks;
1425         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1426         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1427         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1428
1429         log->l_prev_block  = -1;
1430         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1431         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1432         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1433         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1434
1435         xlog_grant_head_init(&log->l_reserve_head);
1436         xlog_grant_head_init(&log->l_write_head);
1437
1438         error = -EFSCORRUPTED;
1439         if (xfs_sb_version_hassector(&mp->m_sb)) {
1440                 log2_size = mp->m_sb.sb_logsectlog;
1441                 if (log2_size < BBSHIFT) {
1442                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1443                                 log2_size, BBSHIFT);
1444                         goto out_free_log;
1445                 }
1446
1447                 log2_size -= BBSHIFT;
1448                 if (log2_size > mp->m_sectbb_log) {
1449                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1450                                 log2_size, mp->m_sectbb_log);
1451                         goto out_free_log;
1452                 }
1453
1454                 /* for larger sector sizes, must have v2 or external log */
1455                 if (log2_size && log->l_logBBstart > 0 &&
1456                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1457                         xfs_warn(mp,
1458                 "log sector size (0x%x) invalid for configuration.",
1459                                 log2_size);
1460                         goto out_free_log;
1461                 }
1462         }
1463         log->l_sectBBsize = 1 << log2_size;
1464
1465         xlog_get_iclog_buffer_size(mp, log);
1466
1467         spin_lock_init(&log->l_icloglock);
1468         init_waitqueue_head(&log->l_flush_wait);
1469
1470         iclogp = &log->l_iclog;
1471         /*
1472          * The amount of memory to allocate for the iclog structure is
1473          * rather funky due to the way the structure is defined.  It is
1474          * done this way so that we can use different sizes for machines
1475          * with different amounts of memory.  See the definition of
1476          * xlog_in_core_t in xfs_log_priv.h for details.
1477          */
1478         ASSERT(log->l_iclog_size >= 4096);
1479         for (i = 0; i < log->l_iclog_bufs; i++) {
1480                 int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1481                 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1482                                 sizeof(struct bio_vec);
1483
1484                 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1485                 if (!iclog)
1486                         goto out_free_iclog;
1487
1488                 *iclogp = iclog;
1489                 iclog->ic_prev = prev_iclog;
1490                 prev_iclog = iclog;
1491
1492                 iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1493                                                 KM_MAYFAIL | KM_ZERO);
1494                 if (!iclog->ic_data)
1495                         goto out_free_iclog;
1496 #ifdef DEBUG
1497                 log->l_iclog_bak[i] = &iclog->ic_header;
1498 #endif
1499                 head = &iclog->ic_header;
1500                 memset(head, 0, sizeof(xlog_rec_header_t));
1501                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1502                 head->h_version = cpu_to_be32(
1503                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1504                 head->h_size = cpu_to_be32(log->l_iclog_size);
1505                 /* new fields */
1506                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1507                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1508
1509                 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1510                 iclog->ic_state = XLOG_STATE_ACTIVE;
1511                 iclog->ic_log = log;
1512                 atomic_set(&iclog->ic_refcnt, 0);
1513                 spin_lock_init(&iclog->ic_callback_lock);
1514                 INIT_LIST_HEAD(&iclog->ic_callbacks);
1515                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1516
1517                 init_waitqueue_head(&iclog->ic_force_wait);
1518                 init_waitqueue_head(&iclog->ic_write_wait);
1519                 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1520                 sema_init(&iclog->ic_sema, 1);
1521
1522                 iclogp = &iclog->ic_next;
1523         }
1524         *iclogp = log->l_iclog;                 /* complete ring */
1525         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1526
1527         log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1528                         WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI, 0,
1529                         mp->m_super->s_id);
1530         if (!log->l_ioend_workqueue)
1531                 goto out_free_iclog;
1532
1533         error = xlog_cil_init(log);
1534         if (error)
1535                 goto out_destroy_workqueue;
1536         return log;
1537
1538 out_destroy_workqueue:
1539         destroy_workqueue(log->l_ioend_workqueue);
1540 out_free_iclog:
1541         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1542                 prev_iclog = iclog->ic_next;
1543                 kmem_free(iclog->ic_data);
1544                 kmem_free(iclog);
1545                 if (prev_iclog == log->l_iclog)
1546                         break;
1547         }
1548 out_free_log:
1549         kmem_free(log);
1550 out:
1551         return ERR_PTR(error);
1552 }       /* xlog_alloc_log */
1553
1554
1555 /*
1556  * Write out the commit record of a transaction associated with the given
1557  * ticket.  Return the lsn of the commit record.
1558  */
1559 STATIC int
1560 xlog_commit_record(
1561         struct xlog             *log,
1562         struct xlog_ticket      *ticket,
1563         struct xlog_in_core     **iclog,
1564         xfs_lsn_t               *commitlsnp)
1565 {
1566         struct xfs_mount *mp = log->l_mp;
1567         int     error;
1568         struct xfs_log_iovec reg = {
1569                 .i_addr = NULL,
1570                 .i_len = 0,
1571                 .i_type = XLOG_REG_TYPE_COMMIT,
1572         };
1573         struct xfs_log_vec vec = {
1574                 .lv_niovecs = 1,
1575                 .lv_iovecp = &reg,
1576         };
1577
1578         ASSERT_ALWAYS(iclog);
1579         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1580                                         XLOG_COMMIT_TRANS);
1581         if (error)
1582                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1583         return error;
1584 }
1585
1586 /*
1587  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1588  * log space.  This code pushes on the lsn which would supposedly free up
1589  * the 25% which we want to leave free.  We may need to adopt a policy which
1590  * pushes on an lsn which is further along in the log once we reach the high
1591  * water mark.  In this manner, we would be creating a low water mark.
1592  */
1593 STATIC void
1594 xlog_grant_push_ail(
1595         struct xlog     *log,
1596         int             need_bytes)
1597 {
1598         xfs_lsn_t       threshold_lsn = 0;
1599         xfs_lsn_t       last_sync_lsn;
1600         int             free_blocks;
1601         int             free_bytes;
1602         int             threshold_block;
1603         int             threshold_cycle;
1604         int             free_threshold;
1605
1606         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1607
1608         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1609         free_blocks = BTOBBT(free_bytes);
1610
1611         /*
1612          * Set the threshold for the minimum number of free blocks in the
1613          * log to the maximum of what the caller needs, one quarter of the
1614          * log, and 256 blocks.
1615          */
1616         free_threshold = BTOBB(need_bytes);
1617         free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1618         free_threshold = max(free_threshold, 256);
1619         if (free_blocks >= free_threshold)
1620                 return;
1621
1622         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1623                                                 &threshold_block);
1624         threshold_block += free_threshold;
1625         if (threshold_block >= log->l_logBBsize) {
1626                 threshold_block -= log->l_logBBsize;
1627                 threshold_cycle += 1;
1628         }
1629         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1630                                         threshold_block);
1631         /*
1632          * Don't pass in an lsn greater than the lsn of the last
1633          * log record known to be on disk. Use a snapshot of the last sync lsn
1634          * so that it doesn't change between the compare and the set.
1635          */
1636         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1637         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1638                 threshold_lsn = last_sync_lsn;
1639
1640         /*
1641          * Get the transaction layer to kick the dirty buffers out to
1642          * disk asynchronously. No point in trying to do this if
1643          * the filesystem is shutting down.
1644          */
1645         if (!XLOG_FORCED_SHUTDOWN(log))
1646                 xfs_ail_push(log->l_ailp, threshold_lsn);
1647 }
1648
1649 /*
1650  * Stamp cycle number in every block
1651  */
1652 STATIC void
1653 xlog_pack_data(
1654         struct xlog             *log,
1655         struct xlog_in_core     *iclog,
1656         int                     roundoff)
1657 {
1658         int                     i, j, k;
1659         int                     size = iclog->ic_offset + roundoff;
1660         __be32                  cycle_lsn;
1661         char                    *dp;
1662
1663         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1664
1665         dp = iclog->ic_datap;
1666         for (i = 0; i < BTOBB(size); i++) {
1667                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1668                         break;
1669                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1670                 *(__be32 *)dp = cycle_lsn;
1671                 dp += BBSIZE;
1672         }
1673
1674         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1675                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1676
1677                 for ( ; i < BTOBB(size); i++) {
1678                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1679                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1680                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1681                         *(__be32 *)dp = cycle_lsn;
1682                         dp += BBSIZE;
1683                 }
1684
1685                 for (i = 1; i < log->l_iclog_heads; i++)
1686                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1687         }
1688 }
1689
1690 /*
1691  * Calculate the checksum for a log buffer.
1692  *
1693  * This is a little more complicated than it should be because the various
1694  * headers and the actual data are non-contiguous.
1695  */
1696 __le32
1697 xlog_cksum(
1698         struct xlog             *log,
1699         struct xlog_rec_header  *rhead,
1700         char                    *dp,
1701         int                     size)
1702 {
1703         uint32_t                crc;
1704
1705         /* first generate the crc for the record header ... */
1706         crc = xfs_start_cksum_update((char *)rhead,
1707                               sizeof(struct xlog_rec_header),
1708                               offsetof(struct xlog_rec_header, h_crc));
1709
1710         /* ... then for additional cycle data for v2 logs ... */
1711         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1712                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1713                 int             i;
1714                 int             xheads;
1715
1716                 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1717                 if (size % XLOG_HEADER_CYCLE_SIZE)
1718                         xheads++;
1719
1720                 for (i = 1; i < xheads; i++) {
1721                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1722                                      sizeof(struct xlog_rec_ext_header));
1723                 }
1724         }
1725
1726         /* ... and finally for the payload */
1727         crc = crc32c(crc, dp, size);
1728
1729         return xfs_end_cksum(crc);
1730 }
1731
1732 static void
1733 xlog_bio_end_io(
1734         struct bio              *bio)
1735 {
1736         struct xlog_in_core     *iclog = bio->bi_private;
1737
1738         queue_work(iclog->ic_log->l_ioend_workqueue,
1739                    &iclog->ic_end_io_work);
1740 }
1741
1742 static void
1743 xlog_map_iclog_data(
1744         struct bio              *bio,
1745         void                    *data,
1746         size_t                  count)
1747 {
1748         do {
1749                 struct page     *page = kmem_to_page(data);
1750                 unsigned int    off = offset_in_page(data);
1751                 size_t          len = min_t(size_t, count, PAGE_SIZE - off);
1752
1753                 WARN_ON_ONCE(bio_add_page(bio, page, len, off) != len);
1754
1755                 data += len;
1756                 count -= len;
1757         } while (count);
1758 }
1759
1760 STATIC void
1761 xlog_write_iclog(
1762         struct xlog             *log,
1763         struct xlog_in_core     *iclog,
1764         uint64_t                bno,
1765         unsigned int            count,
1766         bool                    need_flush)
1767 {
1768         ASSERT(bno < log->l_logBBsize);
1769
1770         /*
1771          * We lock the iclogbufs here so that we can serialise against I/O
1772          * completion during unmount.  We might be processing a shutdown
1773          * triggered during unmount, and that can occur asynchronously to the
1774          * unmount thread, and hence we need to ensure that completes before
1775          * tearing down the iclogbufs.  Hence we need to hold the buffer lock
1776          * across the log IO to archieve that.
1777          */
1778         down(&iclog->ic_sema);
1779         if (unlikely(iclog->ic_state == XLOG_STATE_IOERROR)) {
1780                 /*
1781                  * It would seem logical to return EIO here, but we rely on
1782                  * the log state machine to propagate I/O errors instead of
1783                  * doing it here.  We kick of the state machine and unlock
1784                  * the buffer manually, the code needs to be kept in sync
1785                  * with the I/O completion path.
1786                  */
1787                 xlog_state_done_syncing(iclog, true);
1788                 up(&iclog->ic_sema);
1789                 return;
1790         }
1791
1792         bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1793         bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1794         iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1795         iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1796         iclog->ic_bio.bi_private = iclog;
1797         iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_FUA;
1798         if (need_flush)
1799                 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1800
1801         xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count);
1802         if (is_vmalloc_addr(iclog->ic_data))
1803                 flush_kernel_vmap_range(iclog->ic_data, count);
1804
1805         /*
1806          * If this log buffer would straddle the end of the log we will have
1807          * to split it up into two bios, so that we can continue at the start.
1808          */
1809         if (bno + BTOBB(count) > log->l_logBBsize) {
1810                 struct bio *split;
1811
1812                 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1813                                   GFP_NOIO, &fs_bio_set);
1814                 bio_chain(split, &iclog->ic_bio);
1815                 submit_bio(split);
1816
1817                 /* restart at logical offset zero for the remainder */
1818                 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1819         }
1820
1821         submit_bio(&iclog->ic_bio);
1822 }
1823
1824 /*
1825  * We need to bump cycle number for the part of the iclog that is
1826  * written to the start of the log. Watch out for the header magic
1827  * number case, though.
1828  */
1829 static void
1830 xlog_split_iclog(
1831         struct xlog             *log,
1832         void                    *data,
1833         uint64_t                bno,
1834         unsigned int            count)
1835 {
1836         unsigned int            split_offset = BBTOB(log->l_logBBsize - bno);
1837         unsigned int            i;
1838
1839         for (i = split_offset; i < count; i += BBSIZE) {
1840                 uint32_t cycle = get_unaligned_be32(data + i);
1841
1842                 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1843                         cycle++;
1844                 put_unaligned_be32(cycle, data + i);
1845         }
1846 }
1847
1848 static int
1849 xlog_calc_iclog_size(
1850         struct xlog             *log,
1851         struct xlog_in_core     *iclog,
1852         uint32_t                *roundoff)
1853 {
1854         uint32_t                count_init, count;
1855         bool                    use_lsunit;
1856
1857         use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1858                         log->l_mp->m_sb.sb_logsunit > 1;
1859
1860         /* Add for LR header */
1861         count_init = log->l_iclog_hsize + iclog->ic_offset;
1862
1863         /* Round out the log write size */
1864         if (use_lsunit) {
1865                 /* we have a v2 stripe unit to use */
1866                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1867         } else {
1868                 count = BBTOB(BTOBB(count_init));
1869         }
1870
1871         ASSERT(count >= count_init);
1872         *roundoff = count - count_init;
1873
1874         if (use_lsunit)
1875                 ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1876         else
1877                 ASSERT(*roundoff < BBTOB(1));
1878         return count;
1879 }
1880
1881 /*
1882  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1883  * fashion.  Previously, we should have moved the current iclog
1884  * ptr in the log to point to the next available iclog.  This allows further
1885  * write to continue while this code syncs out an iclog ready to go.
1886  * Before an in-core log can be written out, the data section must be scanned
1887  * to save away the 1st word of each BBSIZE block into the header.  We replace
1888  * it with the current cycle count.  Each BBSIZE block is tagged with the
1889  * cycle count because there in an implicit assumption that drives will
1890  * guarantee that entire 512 byte blocks get written at once.  In other words,
1891  * we can't have part of a 512 byte block written and part not written.  By
1892  * tagging each block, we will know which blocks are valid when recovering
1893  * after an unclean shutdown.
1894  *
1895  * This routine is single threaded on the iclog.  No other thread can be in
1896  * this routine with the same iclog.  Changing contents of iclog can there-
1897  * fore be done without grabbing the state machine lock.  Updating the global
1898  * log will require grabbing the lock though.
1899  *
1900  * The entire log manager uses a logical block numbering scheme.  Only
1901  * xlog_write_iclog knows about the fact that the log may not start with
1902  * block zero on a given device.
1903  */
1904 STATIC void
1905 xlog_sync(
1906         struct xlog             *log,
1907         struct xlog_in_core     *iclog)
1908 {
1909         unsigned int            count;          /* byte count of bwrite */
1910         unsigned int            roundoff;       /* roundoff to BB or stripe */
1911         uint64_t                bno;
1912         unsigned int            size;
1913         bool                    need_flush = true, split = false;
1914
1915         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1916
1917         count = xlog_calc_iclog_size(log, iclog, &roundoff);
1918
1919         /* move grant heads by roundoff in sync */
1920         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1921         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1922
1923         /* put cycle number in every block */
1924         xlog_pack_data(log, iclog, roundoff);
1925
1926         /* real byte length */
1927         size = iclog->ic_offset;
1928         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1929                 size += roundoff;
1930         iclog->ic_header.h_len = cpu_to_be32(size);
1931
1932         XFS_STATS_INC(log->l_mp, xs_log_writes);
1933         XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1934
1935         bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1936
1937         /* Do we need to split this write into 2 parts? */
1938         if (bno + BTOBB(count) > log->l_logBBsize) {
1939                 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1940                 split = true;
1941         }
1942
1943         /* calculcate the checksum */
1944         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1945                                             iclog->ic_datap, size);
1946         /*
1947          * Intentionally corrupt the log record CRC based on the error injection
1948          * frequency, if defined. This facilitates testing log recovery in the
1949          * event of torn writes. Hence, set the IOABORT state to abort the log
1950          * write on I/O completion and shutdown the fs. The subsequent mount
1951          * detects the bad CRC and attempts to recover.
1952          */
1953 #ifdef DEBUG
1954         if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1955                 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1956                 iclog->ic_fail_crc = true;
1957                 xfs_warn(log->l_mp,
1958         "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1959                          be64_to_cpu(iclog->ic_header.h_lsn));
1960         }
1961 #endif
1962
1963         /*
1964          * Flush the data device before flushing the log to make sure all meta
1965          * data written back from the AIL actually made it to disk before
1966          * stamping the new log tail LSN into the log buffer.  For an external
1967          * log we need to issue the flush explicitly, and unfortunately
1968          * synchronously here; for an internal log we can simply use the block
1969          * layer state machine for preflushes.
1970          */
1971         if (log->l_targ != log->l_mp->m_ddev_targp || split) {
1972                 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1973                 need_flush = false;
1974         }
1975
1976         xlog_verify_iclog(log, iclog, count);
1977         xlog_write_iclog(log, iclog, bno, count, need_flush);
1978 }
1979
1980 /*
1981  * Deallocate a log structure
1982  */
1983 STATIC void
1984 xlog_dealloc_log(
1985         struct xlog     *log)
1986 {
1987         xlog_in_core_t  *iclog, *next_iclog;
1988         int             i;
1989
1990         xlog_cil_destroy(log);
1991
1992         /*
1993          * Cycle all the iclogbuf locks to make sure all log IO completion
1994          * is done before we tear down these buffers.
1995          */
1996         iclog = log->l_iclog;
1997         for (i = 0; i < log->l_iclog_bufs; i++) {
1998                 down(&iclog->ic_sema);
1999                 up(&iclog->ic_sema);
2000                 iclog = iclog->ic_next;
2001         }
2002
2003         iclog = log->l_iclog;
2004         for (i = 0; i < log->l_iclog_bufs; i++) {
2005                 next_iclog = iclog->ic_next;
2006                 kmem_free(iclog->ic_data);
2007                 kmem_free(iclog);
2008                 iclog = next_iclog;
2009         }
2010
2011         log->l_mp->m_log = NULL;
2012         destroy_workqueue(log->l_ioend_workqueue);
2013         kmem_free(log);
2014 }       /* xlog_dealloc_log */
2015
2016 /*
2017  * Update counters atomically now that memcpy is done.
2018  */
2019 static inline void
2020 xlog_state_finish_copy(
2021         struct xlog             *log,
2022         struct xlog_in_core     *iclog,
2023         int                     record_cnt,
2024         int                     copy_bytes)
2025 {
2026         lockdep_assert_held(&log->l_icloglock);
2027
2028         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2029         iclog->ic_offset += copy_bytes;
2030 }
2031
2032 /*
2033  * print out info relating to regions written which consume
2034  * the reservation
2035  */
2036 void
2037 xlog_print_tic_res(
2038         struct xfs_mount        *mp,
2039         struct xlog_ticket      *ticket)
2040 {
2041         uint i;
2042         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2043
2044         /* match with XLOG_REG_TYPE_* in xfs_log.h */
2045 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2046         static char *res_type_str[] = {
2047             REG_TYPE_STR(BFORMAT, "bformat"),
2048             REG_TYPE_STR(BCHUNK, "bchunk"),
2049             REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2050             REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2051             REG_TYPE_STR(IFORMAT, "iformat"),
2052             REG_TYPE_STR(ICORE, "icore"),
2053             REG_TYPE_STR(IEXT, "iext"),
2054             REG_TYPE_STR(IBROOT, "ibroot"),
2055             REG_TYPE_STR(ILOCAL, "ilocal"),
2056             REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2057             REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2058             REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2059             REG_TYPE_STR(QFORMAT, "qformat"),
2060             REG_TYPE_STR(DQUOT, "dquot"),
2061             REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2062             REG_TYPE_STR(LRHEADER, "LR header"),
2063             REG_TYPE_STR(UNMOUNT, "unmount"),
2064             REG_TYPE_STR(COMMIT, "commit"),
2065             REG_TYPE_STR(TRANSHDR, "trans header"),
2066             REG_TYPE_STR(ICREATE, "inode create"),
2067             REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2068             REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2069             REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2070             REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2071             REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2072             REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2073         };
2074         BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2075 #undef REG_TYPE_STR
2076
2077         xfs_warn(mp, "ticket reservation summary:");
2078         xfs_warn(mp, "  unit res    = %d bytes",
2079                  ticket->t_unit_res);
2080         xfs_warn(mp, "  current res = %d bytes",
2081                  ticket->t_curr_res);
2082         xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2083                  ticket->t_res_arr_sum, ticket->t_res_o_flow);
2084         xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2085                  ticket->t_res_num_ophdrs, ophdr_spc);
2086         xfs_warn(mp, "  ophdr + reg = %u bytes",
2087                  ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2088         xfs_warn(mp, "  num regions = %u",
2089                  ticket->t_res_num);
2090
2091         for (i = 0; i < ticket->t_res_num; i++) {
2092                 uint r_type = ticket->t_res_arr[i].r_type;
2093                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2094                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2095                             "bad-rtype" : res_type_str[r_type]),
2096                             ticket->t_res_arr[i].r_len);
2097         }
2098 }
2099
2100 /*
2101  * Print a summary of the transaction.
2102  */
2103 void
2104 xlog_print_trans(
2105         struct xfs_trans        *tp)
2106 {
2107         struct xfs_mount        *mp = tp->t_mountp;
2108         struct xfs_log_item     *lip;
2109
2110         /* dump core transaction and ticket info */
2111         xfs_warn(mp, "transaction summary:");
2112         xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2113         xfs_warn(mp, "  log count = %d", tp->t_log_count);
2114         xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2115
2116         xlog_print_tic_res(mp, tp->t_ticket);
2117
2118         /* dump each log item */
2119         list_for_each_entry(lip, &tp->t_items, li_trans) {
2120                 struct xfs_log_vec      *lv = lip->li_lv;
2121                 struct xfs_log_iovec    *vec;
2122                 int                     i;
2123
2124                 xfs_warn(mp, "log item: ");
2125                 xfs_warn(mp, "  type    = 0x%x", lip->li_type);
2126                 xfs_warn(mp, "  flags   = 0x%lx", lip->li_flags);
2127                 if (!lv)
2128                         continue;
2129                 xfs_warn(mp, "  niovecs = %d", lv->lv_niovecs);
2130                 xfs_warn(mp, "  size    = %d", lv->lv_size);
2131                 xfs_warn(mp, "  bytes   = %d", lv->lv_bytes);
2132                 xfs_warn(mp, "  buf len = %d", lv->lv_buf_len);
2133
2134                 /* dump each iovec for the log item */
2135                 vec = lv->lv_iovecp;
2136                 for (i = 0; i < lv->lv_niovecs; i++) {
2137                         int dumplen = min(vec->i_len, 32);
2138
2139                         xfs_warn(mp, "  iovec[%d]", i);
2140                         xfs_warn(mp, "    type  = 0x%x", vec->i_type);
2141                         xfs_warn(mp, "    len   = %d", vec->i_len);
2142                         xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2143                         xfs_hex_dump(vec->i_addr, dumplen);
2144
2145                         vec++;
2146                 }
2147         }
2148 }
2149
2150 /*
2151  * Calculate the potential space needed by the log vector.  Each region gets
2152  * its own xlog_op_header_t and may need to be double word aligned.
2153  */
2154 static int
2155 xlog_write_calc_vec_length(
2156         struct xlog_ticket      *ticket,
2157         struct xfs_log_vec      *log_vector)
2158 {
2159         struct xfs_log_vec      *lv;
2160         int                     headers = 0;
2161         int                     len = 0;
2162         int                     i;
2163
2164         /* acct for start rec of xact */
2165         if (ticket->t_flags & XLOG_TIC_INITED)
2166                 headers++;
2167
2168         for (lv = log_vector; lv; lv = lv->lv_next) {
2169                 /* we don't write ordered log vectors */
2170                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2171                         continue;
2172
2173                 headers += lv->lv_niovecs;
2174
2175                 for (i = 0; i < lv->lv_niovecs; i++) {
2176                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2177
2178                         len += vecp->i_len;
2179                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2180                 }
2181         }
2182
2183         ticket->t_res_num_ophdrs += headers;
2184         len += headers * sizeof(struct xlog_op_header);
2185
2186         return len;
2187 }
2188
2189 /*
2190  * If first write for transaction, insert start record  We can't be trying to
2191  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2192  */
2193 static int
2194 xlog_write_start_rec(
2195         struct xlog_op_header   *ophdr,
2196         struct xlog_ticket      *ticket)
2197 {
2198         if (!(ticket->t_flags & XLOG_TIC_INITED))
2199                 return 0;
2200
2201         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2202         ophdr->oh_clientid = ticket->t_clientid;
2203         ophdr->oh_len = 0;
2204         ophdr->oh_flags = XLOG_START_TRANS;
2205         ophdr->oh_res2 = 0;
2206
2207         ticket->t_flags &= ~XLOG_TIC_INITED;
2208
2209         return sizeof(struct xlog_op_header);
2210 }
2211
2212 static xlog_op_header_t *
2213 xlog_write_setup_ophdr(
2214         struct xlog             *log,
2215         struct xlog_op_header   *ophdr,
2216         struct xlog_ticket      *ticket,
2217         uint                    flags)
2218 {
2219         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2220         ophdr->oh_clientid = ticket->t_clientid;
2221         ophdr->oh_res2 = 0;
2222
2223         /* are we copying a commit or unmount record? */
2224         ophdr->oh_flags = flags;
2225
2226         /*
2227          * We've seen logs corrupted with bad transaction client ids.  This
2228          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2229          * and shut down the filesystem.
2230          */
2231         switch (ophdr->oh_clientid)  {
2232         case XFS_TRANSACTION:
2233         case XFS_VOLUME:
2234         case XFS_LOG:
2235                 break;
2236         default:
2237                 xfs_warn(log->l_mp,
2238                         "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2239                         ophdr->oh_clientid, ticket);
2240                 return NULL;
2241         }
2242
2243         return ophdr;
2244 }
2245
2246 /*
2247  * Set up the parameters of the region copy into the log. This has
2248  * to handle region write split across multiple log buffers - this
2249  * state is kept external to this function so that this code can
2250  * be written in an obvious, self documenting manner.
2251  */
2252 static int
2253 xlog_write_setup_copy(
2254         struct xlog_ticket      *ticket,
2255         struct xlog_op_header   *ophdr,
2256         int                     space_available,
2257         int                     space_required,
2258         int                     *copy_off,
2259         int                     *copy_len,
2260         int                     *last_was_partial_copy,
2261         int                     *bytes_consumed)
2262 {
2263         int                     still_to_copy;
2264
2265         still_to_copy = space_required - *bytes_consumed;
2266         *copy_off = *bytes_consumed;
2267
2268         if (still_to_copy <= space_available) {
2269                 /* write of region completes here */
2270                 *copy_len = still_to_copy;
2271                 ophdr->oh_len = cpu_to_be32(*copy_len);
2272                 if (*last_was_partial_copy)
2273                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2274                 *last_was_partial_copy = 0;
2275                 *bytes_consumed = 0;
2276                 return 0;
2277         }
2278
2279         /* partial write of region, needs extra log op header reservation */
2280         *copy_len = space_available;
2281         ophdr->oh_len = cpu_to_be32(*copy_len);
2282         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2283         if (*last_was_partial_copy)
2284                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2285         *bytes_consumed += *copy_len;
2286         (*last_was_partial_copy)++;
2287
2288         /* account for new log op header */
2289         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2290         ticket->t_res_num_ophdrs++;
2291
2292         return sizeof(struct xlog_op_header);
2293 }
2294
2295 static int
2296 xlog_write_copy_finish(
2297         struct xlog             *log,
2298         struct xlog_in_core     *iclog,
2299         uint                    flags,
2300         int                     *record_cnt,
2301         int                     *data_cnt,
2302         int                     *partial_copy,
2303         int                     *partial_copy_len,
2304         int                     log_offset,
2305         struct xlog_in_core     **commit_iclog)
2306 {
2307         int                     error;
2308
2309         if (*partial_copy) {
2310                 /*
2311                  * This iclog has already been marked WANT_SYNC by
2312                  * xlog_state_get_iclog_space.
2313                  */
2314                 spin_lock(&log->l_icloglock);
2315                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2316                 *record_cnt = 0;
2317                 *data_cnt = 0;
2318                 goto release_iclog;
2319         }
2320
2321         *partial_copy = 0;
2322         *partial_copy_len = 0;
2323
2324         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2325                 /* no more space in this iclog - push it. */
2326                 spin_lock(&log->l_icloglock);
2327                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2328                 *record_cnt = 0;
2329                 *data_cnt = 0;
2330
2331                 xlog_state_want_sync(log, iclog);
2332                 if (!commit_iclog)
2333                         goto release_iclog;
2334                 spin_unlock(&log->l_icloglock);
2335                 ASSERT(flags & XLOG_COMMIT_TRANS);
2336                 *commit_iclog = iclog;
2337         }
2338
2339         return 0;
2340
2341 release_iclog:
2342         error = xlog_state_release_iclog(log, iclog);
2343         spin_unlock(&log->l_icloglock);
2344         return error;
2345 }
2346
2347 /*
2348  * Write some region out to in-core log
2349  *
2350  * This will be called when writing externally provided regions or when
2351  * writing out a commit record for a given transaction.
2352  *
2353  * General algorithm:
2354  *      1. Find total length of this write.  This may include adding to the
2355  *              lengths passed in.
2356  *      2. Check whether we violate the tickets reservation.
2357  *      3. While writing to this iclog
2358  *          A. Reserve as much space in this iclog as can get
2359  *          B. If this is first write, save away start lsn
2360  *          C. While writing this region:
2361  *              1. If first write of transaction, write start record
2362  *              2. Write log operation header (header per region)
2363  *              3. Find out if we can fit entire region into this iclog
2364  *              4. Potentially, verify destination memcpy ptr
2365  *              5. Memcpy (partial) region
2366  *              6. If partial copy, release iclog; otherwise, continue
2367  *                      copying more regions into current iclog
2368  *      4. Mark want sync bit (in simulation mode)
2369  *      5. Release iclog for potential flush to on-disk log.
2370  *
2371  * ERRORS:
2372  * 1.   Panic if reservation is overrun.  This should never happen since
2373  *      reservation amounts are generated internal to the filesystem.
2374  * NOTES:
2375  * 1. Tickets are single threaded data structures.
2376  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2377  *      syncing routine.  When a single log_write region needs to span
2378  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2379  *      on all log operation writes which don't contain the end of the
2380  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2381  *      operation which contains the end of the continued log_write region.
2382  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2383  *      we don't really know exactly how much space will be used.  As a result,
2384  *      we don't update ic_offset until the end when we know exactly how many
2385  *      bytes have been written out.
2386  */
2387 int
2388 xlog_write(
2389         struct xlog             *log,
2390         struct xfs_log_vec      *log_vector,
2391         struct xlog_ticket      *ticket,
2392         xfs_lsn_t               *start_lsn,
2393         struct xlog_in_core     **commit_iclog,
2394         uint                    flags)
2395 {
2396         struct xlog_in_core     *iclog = NULL;
2397         struct xfs_log_iovec    *vecp;
2398         struct xfs_log_vec      *lv;
2399         int                     len;
2400         int                     index;
2401         int                     partial_copy = 0;
2402         int                     partial_copy_len = 0;
2403         int                     contwr = 0;
2404         int                     record_cnt = 0;
2405         int                     data_cnt = 0;
2406         int                     error = 0;
2407
2408         *start_lsn = 0;
2409
2410         len = xlog_write_calc_vec_length(ticket, log_vector);
2411
2412         /*
2413          * Region headers and bytes are already accounted for.
2414          * We only need to take into account start records and
2415          * split regions in this function.
2416          */
2417         if (ticket->t_flags & XLOG_TIC_INITED)
2418                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2419
2420         /*
2421          * Commit record headers need to be accounted for. These
2422          * come in as separate writes so are easy to detect.
2423          */
2424         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2425                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2426
2427         if (ticket->t_curr_res < 0) {
2428                 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2429                      "ctx ticket reservation ran out. Need to up reservation");
2430                 xlog_print_tic_res(log->l_mp, ticket);
2431                 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2432         }
2433
2434         index = 0;
2435         lv = log_vector;
2436         vecp = lv->lv_iovecp;
2437         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2438                 void            *ptr;
2439                 int             log_offset;
2440
2441                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2442                                                    &contwr, &log_offset);
2443                 if (error)
2444                         return error;
2445
2446                 ASSERT(log_offset <= iclog->ic_size - 1);
2447                 ptr = iclog->ic_datap + log_offset;
2448
2449                 /* start_lsn is the first lsn written to. That's all we need. */
2450                 if (!*start_lsn)
2451                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2452
2453                 /*
2454                  * This loop writes out as many regions as can fit in the amount
2455                  * of space which was allocated by xlog_state_get_iclog_space().
2456                  */
2457                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2458                         struct xfs_log_iovec    *reg;
2459                         struct xlog_op_header   *ophdr;
2460                         int                     start_rec_copy;
2461                         int                     copy_len;
2462                         int                     copy_off;
2463                         bool                    ordered = false;
2464
2465                         /* ordered log vectors have no regions to write */
2466                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2467                                 ASSERT(lv->lv_niovecs == 0);
2468                                 ordered = true;
2469                                 goto next_lv;
2470                         }
2471
2472                         reg = &vecp[index];
2473                         ASSERT(reg->i_len % sizeof(int32_t) == 0);
2474                         ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2475
2476                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2477                         if (start_rec_copy) {
2478                                 record_cnt++;
2479                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2480                                                    start_rec_copy);
2481                         }
2482
2483                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2484                         if (!ophdr)
2485                                 return -EIO;
2486
2487                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2488                                            sizeof(struct xlog_op_header));
2489
2490                         len += xlog_write_setup_copy(ticket, ophdr,
2491                                                      iclog->ic_size-log_offset,
2492                                                      reg->i_len,
2493                                                      &copy_off, &copy_len,
2494                                                      &partial_copy,
2495                                                      &partial_copy_len);
2496                         xlog_verify_dest_ptr(log, ptr);
2497
2498                         /*
2499                          * Copy region.
2500                          *
2501                          * Unmount records just log an opheader, so can have
2502                          * empty payloads with no data region to copy. Hence we
2503                          * only copy the payload if the vector says it has data
2504                          * to copy.
2505                          */
2506                         ASSERT(copy_len >= 0);
2507                         if (copy_len > 0) {
2508                                 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2509                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2510                                                    copy_len);
2511                         }
2512                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2513                         record_cnt++;
2514                         data_cnt += contwr ? copy_len : 0;
2515
2516                         error = xlog_write_copy_finish(log, iclog, flags,
2517                                                        &record_cnt, &data_cnt,
2518                                                        &partial_copy,
2519                                                        &partial_copy_len,
2520                                                        log_offset,
2521                                                        commit_iclog);
2522                         if (error)
2523                                 return error;
2524
2525                         /*
2526                          * if we had a partial copy, we need to get more iclog
2527                          * space but we don't want to increment the region
2528                          * index because there is still more is this region to
2529                          * write.
2530                          *
2531                          * If we completed writing this region, and we flushed
2532                          * the iclog (indicated by resetting of the record
2533                          * count), then we also need to get more log space. If
2534                          * this was the last record, though, we are done and
2535                          * can just return.
2536                          */
2537                         if (partial_copy)
2538                                 break;
2539
2540                         if (++index == lv->lv_niovecs) {
2541 next_lv:
2542                                 lv = lv->lv_next;
2543                                 index = 0;
2544                                 if (lv)
2545                                         vecp = lv->lv_iovecp;
2546                         }
2547                         if (record_cnt == 0 && !ordered) {
2548                                 if (!lv)
2549                                         return 0;
2550                                 break;
2551                         }
2552                 }
2553         }
2554
2555         ASSERT(len == 0);
2556
2557         spin_lock(&log->l_icloglock);
2558         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2559         if (commit_iclog) {
2560                 ASSERT(flags & XLOG_COMMIT_TRANS);
2561                 *commit_iclog = iclog;
2562         } else {
2563                 error = xlog_state_release_iclog(log, iclog);
2564         }
2565         spin_unlock(&log->l_icloglock);
2566
2567         return error;
2568 }
2569
2570
2571 /*****************************************************************************
2572  *
2573  *              State Machine functions
2574  *
2575  *****************************************************************************
2576  */
2577
2578 /*
2579  * An iclog has just finished IO completion processing, so we need to update
2580  * the iclog state and propagate that up into the overall log state. Hence we
2581  * prepare the iclog for cleaning, and then clean all the pending dirty iclogs
2582  * starting from the head, and then wake up any threads that are waiting for the
2583  * iclog to be marked clean.
2584  *
2585  * The ordering of marking iclogs ACTIVE must be maintained, so an iclog
2586  * doesn't become ACTIVE beyond one that is SYNCING.  This is also required to
2587  * maintain the notion that we use a ordered wait queue to hold off would be
2588  * writers to the log when every iclog is trying to sync to disk.
2589  *
2590  * Caller must hold the icloglock before calling us.
2591  *
2592  * State Change: !IOERROR -> DIRTY -> ACTIVE
2593  */
2594 STATIC void
2595 xlog_state_clean_iclog(
2596         struct xlog             *log,
2597         struct xlog_in_core     *dirty_iclog)
2598 {
2599         struct xlog_in_core     *iclog;
2600         int                     changed = 0;
2601
2602         /* Prepare the completed iclog. */
2603         if (dirty_iclog->ic_state != XLOG_STATE_IOERROR)
2604                 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2605
2606         /* Walk all the iclogs to update the ordered active state. */
2607         iclog = log->l_iclog;
2608         do {
2609                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2610                         iclog->ic_state = XLOG_STATE_ACTIVE;
2611                         iclog->ic_offset       = 0;
2612                         ASSERT(list_empty_careful(&iclog->ic_callbacks));
2613                         /*
2614                          * If the number of ops in this iclog indicate it just
2615                          * contains the dummy transaction, we can
2616                          * change state into IDLE (the second time around).
2617                          * Otherwise we should change the state into
2618                          * NEED a dummy.
2619                          * We don't need to cover the dummy.
2620                          */
2621                         if (!changed &&
2622                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2623                                         XLOG_COVER_OPS)) {
2624                                 changed = 1;
2625                         } else {
2626                                 /*
2627                                  * We have two dirty iclogs so start over
2628                                  * This could also be num of ops indicates
2629                                  * this is not the dummy going out.
2630                                  */
2631                                 changed = 2;
2632                         }
2633                         iclog->ic_header.h_num_logops = 0;
2634                         memset(iclog->ic_header.h_cycle_data, 0,
2635                               sizeof(iclog->ic_header.h_cycle_data));
2636                         iclog->ic_header.h_lsn = 0;
2637                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2638                         /* do nothing */;
2639                 else
2640                         break;  /* stop cleaning */
2641                 iclog = iclog->ic_next;
2642         } while (iclog != log->l_iclog);
2643
2644
2645         /*
2646          * Wake up threads waiting in xfs_log_force() for the dirty iclog
2647          * to be cleaned.
2648          */
2649         wake_up_all(&dirty_iclog->ic_force_wait);
2650
2651         /*
2652          * Change state for the dummy log recording.
2653          * We usually go to NEED. But we go to NEED2 if the changed indicates
2654          * we are done writing the dummy record.
2655          * If we are done with the second dummy recored (DONE2), then
2656          * we go to IDLE.
2657          */
2658         if (changed) {
2659                 switch (log->l_covered_state) {
2660                 case XLOG_STATE_COVER_IDLE:
2661                 case XLOG_STATE_COVER_NEED:
2662                 case XLOG_STATE_COVER_NEED2:
2663                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2664                         break;
2665
2666                 case XLOG_STATE_COVER_DONE:
2667                         if (changed == 1)
2668                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2669                         else
2670                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2671                         break;
2672
2673                 case XLOG_STATE_COVER_DONE2:
2674                         if (changed == 1)
2675                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2676                         else
2677                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2678                         break;
2679
2680                 default:
2681                         ASSERT(0);
2682                 }
2683         }
2684 }
2685
2686 STATIC xfs_lsn_t
2687 xlog_get_lowest_lsn(
2688         struct xlog             *log)
2689 {
2690         struct xlog_in_core     *iclog = log->l_iclog;
2691         xfs_lsn_t               lowest_lsn = 0, lsn;
2692
2693         do {
2694                 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2695                     iclog->ic_state == XLOG_STATE_DIRTY)
2696                         continue;
2697
2698                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2699                 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2700                         lowest_lsn = lsn;
2701         } while ((iclog = iclog->ic_next) != log->l_iclog);
2702
2703         return lowest_lsn;
2704 }
2705
2706 /*
2707  * Completion of a iclog IO does not imply that a transaction has completed, as
2708  * transactions can be large enough to span many iclogs. We cannot change the
2709  * tail of the log half way through a transaction as this may be the only
2710  * transaction in the log and moving the tail to point to the middle of it
2711  * will prevent recovery from finding the start of the transaction. Hence we
2712  * should only update the last_sync_lsn if this iclog contains transaction
2713  * completion callbacks on it.
2714  *
2715  * We have to do this before we drop the icloglock to ensure we are the only one
2716  * that can update it.
2717  *
2718  * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2719  * the reservation grant head pushing. This is due to the fact that the push
2720  * target is bound by the current last_sync_lsn value. Hence if we have a large
2721  * amount of log space bound up in this committing transaction then the
2722  * last_sync_lsn value may be the limiting factor preventing tail pushing from
2723  * freeing space in the log. Hence once we've updated the last_sync_lsn we
2724  * should push the AIL to ensure the push target (and hence the grant head) is
2725  * no longer bound by the old log head location and can move forwards and make
2726  * progress again.
2727  */
2728 static void
2729 xlog_state_set_callback(
2730         struct xlog             *log,
2731         struct xlog_in_core     *iclog,
2732         xfs_lsn_t               header_lsn)
2733 {
2734         iclog->ic_state = XLOG_STATE_CALLBACK;
2735
2736         ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2737                            header_lsn) <= 0);
2738
2739         if (list_empty_careful(&iclog->ic_callbacks))
2740                 return;
2741
2742         atomic64_set(&log->l_last_sync_lsn, header_lsn);
2743         xlog_grant_push_ail(log, 0);
2744 }
2745
2746 /*
2747  * Return true if we need to stop processing, false to continue to the next
2748  * iclog. The caller will need to run callbacks if the iclog is returned in the
2749  * XLOG_STATE_CALLBACK state.
2750  */
2751 static bool
2752 xlog_state_iodone_process_iclog(
2753         struct xlog             *log,
2754         struct xlog_in_core     *iclog,
2755         bool                    *ioerror)
2756 {
2757         xfs_lsn_t               lowest_lsn;
2758         xfs_lsn_t               header_lsn;
2759
2760         switch (iclog->ic_state) {
2761         case XLOG_STATE_ACTIVE:
2762         case XLOG_STATE_DIRTY:
2763                 /*
2764                  * Skip all iclogs in the ACTIVE & DIRTY states:
2765                  */
2766                 return false;
2767         case XLOG_STATE_IOERROR:
2768                 /*
2769                  * Between marking a filesystem SHUTDOWN and stopping the log,
2770                  * we do flush all iclogs to disk (if there wasn't a log I/O
2771                  * error). So, we do want things to go smoothly in case of just
2772                  * a SHUTDOWN w/o a LOG_IO_ERROR.
2773                  */
2774                 *ioerror = true;
2775                 return false;
2776         case XLOG_STATE_DONE_SYNC:
2777                 /*
2778                  * Now that we have an iclog that is in the DONE_SYNC state, do
2779                  * one more check here to see if we have chased our tail around.
2780                  * If this is not the lowest lsn iclog, then we will leave it
2781                  * for another completion to process.
2782                  */
2783                 header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2784                 lowest_lsn = xlog_get_lowest_lsn(log);
2785                 if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2786                         return false;
2787                 xlog_state_set_callback(log, iclog, header_lsn);
2788                 return false;
2789         default:
2790                 /*
2791                  * Can only perform callbacks in order.  Since this iclog is not
2792                  * in the DONE_SYNC state, we skip the rest and just try to
2793                  * clean up.
2794                  */
2795                 return true;
2796         }
2797 }
2798
2799 /*
2800  * Keep processing entries in the iclog callback list until we come around and
2801  * it is empty.  We need to atomically see that the list is empty and change the
2802  * state to DIRTY so that we don't miss any more callbacks being added.
2803  *
2804  * This function is called with the icloglock held and returns with it held. We
2805  * drop it while running callbacks, however, as holding it over thousands of
2806  * callbacks is unnecessary and causes excessive contention if we do.
2807  */
2808 static void
2809 xlog_state_do_iclog_callbacks(
2810         struct xlog             *log,
2811         struct xlog_in_core     *iclog,
2812         bool                    aborted)
2813                 __releases(&log->l_icloglock)
2814                 __acquires(&log->l_icloglock)
2815 {
2816         spin_unlock(&log->l_icloglock);
2817         spin_lock(&iclog->ic_callback_lock);
2818         while (!list_empty(&iclog->ic_callbacks)) {
2819                 LIST_HEAD(tmp);
2820
2821                 list_splice_init(&iclog->ic_callbacks, &tmp);
2822
2823                 spin_unlock(&iclog->ic_callback_lock);
2824                 xlog_cil_process_committed(&tmp, aborted);
2825                 spin_lock(&iclog->ic_callback_lock);
2826         }
2827
2828         /*
2829          * Pick up the icloglock while still holding the callback lock so we
2830          * serialise against anyone trying to add more callbacks to this iclog
2831          * now we've finished processing.
2832          */
2833         spin_lock(&log->l_icloglock);
2834         spin_unlock(&iclog->ic_callback_lock);
2835 }
2836
2837 STATIC void
2838 xlog_state_do_callback(
2839         struct xlog             *log,
2840         bool                    aborted)
2841 {
2842         struct xlog_in_core     *iclog;
2843         struct xlog_in_core     *first_iclog;
2844         bool                    cycled_icloglock;
2845         bool                    ioerror;
2846         int                     flushcnt = 0;
2847         int                     repeats = 0;
2848
2849         spin_lock(&log->l_icloglock);
2850         do {
2851                 /*
2852                  * Scan all iclogs starting with the one pointed to by the
2853                  * log.  Reset this starting point each time the log is
2854                  * unlocked (during callbacks).
2855                  *
2856                  * Keep looping through iclogs until one full pass is made
2857                  * without running any callbacks.
2858                  */
2859                 first_iclog = log->l_iclog;
2860                 iclog = log->l_iclog;
2861                 cycled_icloglock = false;
2862                 ioerror = false;
2863                 repeats++;
2864
2865                 do {
2866                         if (xlog_state_iodone_process_iclog(log, iclog,
2867                                                         &ioerror))
2868                                 break;
2869
2870                         if (iclog->ic_state != XLOG_STATE_CALLBACK &&
2871                             iclog->ic_state != XLOG_STATE_IOERROR) {
2872                                 iclog = iclog->ic_next;
2873                                 continue;
2874                         }
2875
2876                         /*
2877                          * Running callbacks will drop the icloglock which means
2878                          * we'll have to run at least one more complete loop.
2879                          */
2880                         cycled_icloglock = true;
2881                         xlog_state_do_iclog_callbacks(log, iclog, aborted);
2882
2883                         xlog_state_clean_iclog(log, iclog);
2884                         iclog = iclog->ic_next;
2885                 } while (first_iclog != iclog);
2886
2887                 if (repeats > 5000) {
2888                         flushcnt += repeats;
2889                         repeats = 0;
2890                         xfs_warn(log->l_mp,
2891                                 "%s: possible infinite loop (%d iterations)",
2892                                 __func__, flushcnt);
2893                 }
2894         } while (!ioerror && cycled_icloglock);
2895
2896         if (log->l_iclog->ic_state == XLOG_STATE_ACTIVE ||
2897             log->l_iclog->ic_state == XLOG_STATE_IOERROR)
2898                 wake_up_all(&log->l_flush_wait);
2899
2900         spin_unlock(&log->l_icloglock);
2901 }
2902
2903
2904 /*
2905  * Finish transitioning this iclog to the dirty state.
2906  *
2907  * Make sure that we completely execute this routine only when this is
2908  * the last call to the iclog.  There is a good chance that iclog flushes,
2909  * when we reach the end of the physical log, get turned into 2 separate
2910  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2911  * routine.  By using the reference count bwritecnt, we guarantee that only
2912  * the second completion goes through.
2913  *
2914  * Callbacks could take time, so they are done outside the scope of the
2915  * global state machine log lock.
2916  */
2917 STATIC void
2918 xlog_state_done_syncing(
2919         struct xlog_in_core     *iclog,
2920         bool                    aborted)
2921 {
2922         struct xlog             *log = iclog->ic_log;
2923
2924         spin_lock(&log->l_icloglock);
2925
2926         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2927
2928         /*
2929          * If we got an error, either on the first buffer, or in the case of
2930          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2931          * and none should ever be attempted to be written to disk
2932          * again.
2933          */
2934         if (iclog->ic_state == XLOG_STATE_SYNCING)
2935                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2936         else
2937                 ASSERT(iclog->ic_state == XLOG_STATE_IOERROR);
2938
2939         /*
2940          * Someone could be sleeping prior to writing out the next
2941          * iclog buffer, we wake them all, one will get to do the
2942          * I/O, the others get to wait for the result.
2943          */
2944         wake_up_all(&iclog->ic_write_wait);
2945         spin_unlock(&log->l_icloglock);
2946         xlog_state_do_callback(log, aborted);   /* also cleans log */
2947 }       /* xlog_state_done_syncing */
2948
2949
2950 /*
2951  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2952  * sleep.  We wait on the flush queue on the head iclog as that should be
2953  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2954  * we will wait here and all new writes will sleep until a sync completes.
2955  *
2956  * The in-core logs are used in a circular fashion. They are not used
2957  * out-of-order even when an iclog past the head is free.
2958  *
2959  * return:
2960  *      * log_offset where xlog_write() can start writing into the in-core
2961  *              log's data space.
2962  *      * in-core log pointer to which xlog_write() should write.
2963  *      * boolean indicating this is a continued write to an in-core log.
2964  *              If this is the last write, then the in-core log's offset field
2965  *              needs to be incremented, depending on the amount of data which
2966  *              is copied.
2967  */
2968 STATIC int
2969 xlog_state_get_iclog_space(
2970         struct xlog             *log,
2971         int                     len,
2972         struct xlog_in_core     **iclogp,
2973         struct xlog_ticket      *ticket,
2974         int                     *continued_write,
2975         int                     *logoffsetp)
2976 {
2977         int               log_offset;
2978         xlog_rec_header_t *head;
2979         xlog_in_core_t    *iclog;
2980
2981 restart:
2982         spin_lock(&log->l_icloglock);
2983         if (XLOG_FORCED_SHUTDOWN(log)) {
2984                 spin_unlock(&log->l_icloglock);
2985                 return -EIO;
2986         }
2987
2988         iclog = log->l_iclog;
2989         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2990                 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2991
2992                 /* Wait for log writes to have flushed */
2993                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2994                 goto restart;
2995         }
2996
2997         head = &iclog->ic_header;
2998
2999         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
3000         log_offset = iclog->ic_offset;
3001
3002         /* On the 1st write to an iclog, figure out lsn.  This works
3003          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3004          * committing to.  If the offset is set, that's how many blocks
3005          * must be written.
3006          */
3007         if (log_offset == 0) {
3008                 ticket->t_curr_res -= log->l_iclog_hsize;
3009                 xlog_tic_add_region(ticket,
3010                                     log->l_iclog_hsize,
3011                                     XLOG_REG_TYPE_LRHEADER);
3012                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3013                 head->h_lsn = cpu_to_be64(
3014                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3015                 ASSERT(log->l_curr_block >= 0);
3016         }
3017
3018         /* If there is enough room to write everything, then do it.  Otherwise,
3019          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3020          * bit is on, so this will get flushed out.  Don't update ic_offset
3021          * until you know exactly how many bytes get copied.  Therefore, wait
3022          * until later to update ic_offset.
3023          *
3024          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3025          * can fit into remaining data section.
3026          */
3027         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3028                 int             error = 0;
3029
3030                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3031
3032                 /*
3033                  * If we are the only one writing to this iclog, sync it to
3034                  * disk.  We need to do an atomic compare and decrement here to
3035                  * avoid racing with concurrent atomic_dec_and_lock() calls in
3036                  * xlog_state_release_iclog() when there is more than one
3037                  * reference to the iclog.
3038                  */
3039                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
3040                         error = xlog_state_release_iclog(log, iclog);
3041                 spin_unlock(&log->l_icloglock);
3042                 if (error)
3043                         return error;
3044                 goto restart;
3045         }
3046
3047         /* Do we have enough room to write the full amount in the remainder
3048          * of this iclog?  Or must we continue a write on the next iclog and
3049          * mark this iclog as completely taken?  In the case where we switch
3050          * iclogs (to mark it taken), this particular iclog will release/sync
3051          * to disk in xlog_write().
3052          */
3053         if (len <= iclog->ic_size - iclog->ic_offset) {
3054                 *continued_write = 0;
3055                 iclog->ic_offset += len;
3056         } else {
3057                 *continued_write = 1;
3058                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3059         }
3060         *iclogp = iclog;
3061
3062         ASSERT(iclog->ic_offset <= iclog->ic_size);
3063         spin_unlock(&log->l_icloglock);
3064
3065         *logoffsetp = log_offset;
3066         return 0;
3067 }       /* xlog_state_get_iclog_space */
3068
3069 /* The first cnt-1 times through here we don't need to
3070  * move the grant write head because the permanent
3071  * reservation has reserved cnt times the unit amount.
3072  * Release part of current permanent unit reservation and
3073  * reset current reservation to be one units worth.  Also
3074  * move grant reservation head forward.
3075  */
3076 STATIC void
3077 xlog_regrant_reserve_log_space(
3078         struct xlog             *log,
3079         struct xlog_ticket      *ticket)
3080 {
3081         trace_xfs_log_regrant_reserve_enter(log, ticket);
3082
3083         if (ticket->t_cnt > 0)
3084                 ticket->t_cnt--;
3085
3086         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3087                                         ticket->t_curr_res);
3088         xlog_grant_sub_space(log, &log->l_write_head.grant,
3089                                         ticket->t_curr_res);
3090         ticket->t_curr_res = ticket->t_unit_res;
3091         xlog_tic_reset_res(ticket);
3092
3093         trace_xfs_log_regrant_reserve_sub(log, ticket);
3094
3095         /* just return if we still have some of the pre-reserved space */
3096         if (ticket->t_cnt > 0)
3097                 return;
3098
3099         xlog_grant_add_space(log, &log->l_reserve_head.grant,
3100                                         ticket->t_unit_res);
3101
3102         trace_xfs_log_regrant_reserve_exit(log, ticket);
3103
3104         ticket->t_curr_res = ticket->t_unit_res;
3105         xlog_tic_reset_res(ticket);
3106 }       /* xlog_regrant_reserve_log_space */
3107
3108
3109 /*
3110  * Give back the space left from a reservation.
3111  *
3112  * All the information we need to make a correct determination of space left
3113  * is present.  For non-permanent reservations, things are quite easy.  The
3114  * count should have been decremented to zero.  We only need to deal with the
3115  * space remaining in the current reservation part of the ticket.  If the
3116  * ticket contains a permanent reservation, there may be left over space which
3117  * needs to be released.  A count of N means that N-1 refills of the current
3118  * reservation can be done before we need to ask for more space.  The first
3119  * one goes to fill up the first current reservation.  Once we run out of
3120  * space, the count will stay at zero and the only space remaining will be
3121  * in the current reservation field.
3122  */
3123 STATIC void
3124 xlog_ungrant_log_space(
3125         struct xlog             *log,
3126         struct xlog_ticket      *ticket)
3127 {
3128         int     bytes;
3129
3130         if (ticket->t_cnt > 0)
3131                 ticket->t_cnt--;
3132
3133         trace_xfs_log_ungrant_enter(log, ticket);
3134         trace_xfs_log_ungrant_sub(log, ticket);
3135
3136         /*
3137          * If this is a permanent reservation ticket, we may be able to free
3138          * up more space based on the remaining count.
3139          */
3140         bytes = ticket->t_curr_res;
3141         if (ticket->t_cnt > 0) {
3142                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3143                 bytes += ticket->t_unit_res*ticket->t_cnt;
3144         }
3145
3146         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3147         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3148
3149         trace_xfs_log_ungrant_exit(log, ticket);
3150
3151         xfs_log_space_wake(log->l_mp);
3152 }
3153
3154 /*
3155  * This routine will mark the current iclog in the ring as WANT_SYNC
3156  * and move the current iclog pointer to the next iclog in the ring.
3157  * When this routine is called from xlog_state_get_iclog_space(), the
3158  * exact size of the iclog has not yet been determined.  All we know is
3159  * that every data block.  We have run out of space in this log record.
3160  */
3161 STATIC void
3162 xlog_state_switch_iclogs(
3163         struct xlog             *log,
3164         struct xlog_in_core     *iclog,
3165         int                     eventual_size)
3166 {
3167         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3168         if (!eventual_size)
3169                 eventual_size = iclog->ic_offset;
3170         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3171         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3172         log->l_prev_block = log->l_curr_block;
3173         log->l_prev_cycle = log->l_curr_cycle;
3174
3175         /* roll log?: ic_offset changed later */
3176         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3177
3178         /* Round up to next log-sunit */
3179         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3180             log->l_mp->m_sb.sb_logsunit > 1) {
3181                 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3182                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3183         }
3184
3185         if (log->l_curr_block >= log->l_logBBsize) {
3186                 /*
3187                  * Rewind the current block before the cycle is bumped to make
3188                  * sure that the combined LSN never transiently moves forward
3189                  * when the log wraps to the next cycle. This is to support the
3190                  * unlocked sample of these fields from xlog_valid_lsn(). Most
3191                  * other cases should acquire l_icloglock.
3192                  */
3193                 log->l_curr_block -= log->l_logBBsize;
3194                 ASSERT(log->l_curr_block >= 0);
3195                 smp_wmb();
3196                 log->l_curr_cycle++;
3197                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3198                         log->l_curr_cycle++;
3199         }
3200         ASSERT(iclog == log->l_iclog);
3201         log->l_iclog = iclog->ic_next;
3202 }       /* xlog_state_switch_iclogs */
3203
3204 /*
3205  * Write out all data in the in-core log as of this exact moment in time.
3206  *
3207  * Data may be written to the in-core log during this call.  However,
3208  * we don't guarantee this data will be written out.  A change from past
3209  * implementation means this routine will *not* write out zero length LRs.
3210  *
3211  * Basically, we try and perform an intelligent scan of the in-core logs.
3212  * If we determine there is no flushable data, we just return.  There is no
3213  * flushable data if:
3214  *
3215  *      1. the current iclog is active and has no data; the previous iclog
3216  *              is in the active or dirty state.
3217  *      2. the current iclog is drity, and the previous iclog is in the
3218  *              active or dirty state.
3219  *
3220  * We may sleep if:
3221  *
3222  *      1. the current iclog is not in the active nor dirty state.
3223  *      2. the current iclog dirty, and the previous iclog is not in the
3224  *              active nor dirty state.
3225  *      3. the current iclog is active, and there is another thread writing
3226  *              to this particular iclog.
3227  *      4. a) the current iclog is active and has no other writers
3228  *         b) when we return from flushing out this iclog, it is still
3229  *              not in the active nor dirty state.
3230  */
3231 int
3232 xfs_log_force(
3233         struct xfs_mount        *mp,
3234         uint                    flags)
3235 {
3236         struct xlog             *log = mp->m_log;
3237         struct xlog_in_core     *iclog;
3238         xfs_lsn_t               lsn;
3239
3240         XFS_STATS_INC(mp, xs_log_force);
3241         trace_xfs_log_force(mp, 0, _RET_IP_);
3242
3243         xlog_cil_force(log);
3244
3245         spin_lock(&log->l_icloglock);
3246         iclog = log->l_iclog;
3247         if (iclog->ic_state == XLOG_STATE_IOERROR)
3248                 goto out_error;
3249
3250         if (iclog->ic_state == XLOG_STATE_DIRTY ||
3251             (iclog->ic_state == XLOG_STATE_ACTIVE &&
3252              atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3253                 /*
3254                  * If the head is dirty or (active and empty), then we need to
3255                  * look at the previous iclog.
3256                  *
3257                  * If the previous iclog is active or dirty we are done.  There
3258                  * is nothing to sync out. Otherwise, we attach ourselves to the
3259                  * previous iclog and go to sleep.
3260                  */
3261                 iclog = iclog->ic_prev;
3262                 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3263                     iclog->ic_state == XLOG_STATE_DIRTY)
3264                         goto out_unlock;
3265         } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3266                 if (atomic_read(&iclog->ic_refcnt) == 0) {
3267                         /*
3268                          * We are the only one with access to this iclog.
3269                          *
3270                          * Flush it out now.  There should be a roundoff of zero
3271                          * to show that someone has already taken care of the
3272                          * roundoff from the previous sync.
3273                          */
3274                         atomic_inc(&iclog->ic_refcnt);
3275                         lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3276                         xlog_state_switch_iclogs(log, iclog, 0);
3277                         if (xlog_state_release_iclog(log, iclog))
3278                                 goto out_error;
3279
3280                         if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3281                             iclog->ic_state == XLOG_STATE_DIRTY)
3282                                 goto out_unlock;
3283                 } else {
3284                         /*
3285                          * Someone else is writing to this iclog.
3286                          *
3287                          * Use its call to flush out the data.  However, the
3288                          * other thread may not force out this LR, so we mark
3289                          * it WANT_SYNC.
3290                          */
3291                         xlog_state_switch_iclogs(log, iclog, 0);
3292                 }
3293         } else {
3294                 /*
3295                  * If the head iclog is not active nor dirty, we just attach
3296                  * ourselves to the head and go to sleep if necessary.
3297                  */
3298                 ;
3299         }
3300
3301         if (!(flags & XFS_LOG_SYNC))
3302                 goto out_unlock;
3303
3304         if (iclog->ic_state == XLOG_STATE_IOERROR)
3305                 goto out_error;
3306         XFS_STATS_INC(mp, xs_log_force_sleep);
3307         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3308         if (iclog->ic_state == XLOG_STATE_IOERROR)
3309                 return -EIO;
3310         return 0;
3311
3312 out_unlock:
3313         spin_unlock(&log->l_icloglock);
3314         return 0;
3315 out_error:
3316         spin_unlock(&log->l_icloglock);
3317         return -EIO;
3318 }
3319
3320 static int
3321 __xfs_log_force_lsn(
3322         struct xfs_mount        *mp,
3323         xfs_lsn_t               lsn,
3324         uint                    flags,
3325         int                     *log_flushed,
3326         bool                    already_slept)
3327 {
3328         struct xlog             *log = mp->m_log;
3329         struct xlog_in_core     *iclog;
3330
3331         spin_lock(&log->l_icloglock);
3332         iclog = log->l_iclog;
3333         if (iclog->ic_state == XLOG_STATE_IOERROR)
3334                 goto out_error;
3335
3336         while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3337                 iclog = iclog->ic_next;
3338                 if (iclog == log->l_iclog)
3339                         goto out_unlock;
3340         }
3341
3342         if (iclog->ic_state == XLOG_STATE_DIRTY)
3343                 goto out_unlock;
3344
3345         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3346                 /*
3347                  * We sleep here if we haven't already slept (e.g. this is the
3348                  * first time we've looked at the correct iclog buf) and the
3349                  * buffer before us is going to be sync'ed.  The reason for this
3350                  * is that if we are doing sync transactions here, by waiting
3351                  * for the previous I/O to complete, we can allow a few more
3352                  * transactions into this iclog before we close it down.
3353                  *
3354                  * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3355                  * refcnt so we can release the log (which drops the ref count).
3356                  * The state switch keeps new transaction commits from using
3357                  * this buffer.  When the current commits finish writing into
3358                  * the buffer, the refcount will drop to zero and the buffer
3359                  * will go out then.
3360                  */
3361                 if (!already_slept &&
3362                     (iclog->ic_prev->ic_state == XLOG_STATE_WANT_SYNC ||
3363                      iclog->ic_prev->ic_state == XLOG_STATE_SYNCING)) {
3364                         XFS_STATS_INC(mp, xs_log_force_sleep);
3365
3366                         xlog_wait(&iclog->ic_prev->ic_write_wait,
3367                                         &log->l_icloglock);
3368                         return -EAGAIN;
3369                 }
3370                 atomic_inc(&iclog->ic_refcnt);
3371                 xlog_state_switch_iclogs(log, iclog, 0);
3372                 if (xlog_state_release_iclog(log, iclog))
3373                         goto out_error;
3374                 if (log_flushed)
3375                         *log_flushed = 1;
3376         }
3377
3378         if (!(flags & XFS_LOG_SYNC) ||
3379             (iclog->ic_state == XLOG_STATE_ACTIVE ||
3380              iclog->ic_state == XLOG_STATE_DIRTY))
3381                 goto out_unlock;
3382
3383         if (iclog->ic_state == XLOG_STATE_IOERROR)
3384                 goto out_error;
3385
3386         XFS_STATS_INC(mp, xs_log_force_sleep);
3387         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3388         if (iclog->ic_state == XLOG_STATE_IOERROR)
3389                 return -EIO;
3390         return 0;
3391
3392 out_unlock:
3393         spin_unlock(&log->l_icloglock);
3394         return 0;
3395 out_error:
3396         spin_unlock(&log->l_icloglock);
3397         return -EIO;
3398 }
3399
3400 /*
3401  * Force the in-core log to disk for a specific LSN.
3402  *
3403  * Find in-core log with lsn.
3404  *      If it is in the DIRTY state, just return.
3405  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3406  *              state and go to sleep or return.
3407  *      If it is in any other state, go to sleep or return.
3408  *
3409  * Synchronous forces are implemented with a wait queue.  All callers trying
3410  * to force a given lsn to disk must wait on the queue attached to the
3411  * specific in-core log.  When given in-core log finally completes its write
3412  * to disk, that thread will wake up all threads waiting on the queue.
3413  */
3414 int
3415 xfs_log_force_lsn(
3416         struct xfs_mount        *mp,
3417         xfs_lsn_t               lsn,
3418         uint                    flags,
3419         int                     *log_flushed)
3420 {
3421         int                     ret;
3422         ASSERT(lsn != 0);
3423
3424         XFS_STATS_INC(mp, xs_log_force);
3425         trace_xfs_log_force(mp, lsn, _RET_IP_);
3426
3427         lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3428         if (lsn == NULLCOMMITLSN)
3429                 return 0;
3430
3431         ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3432         if (ret == -EAGAIN)
3433                 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3434         return ret;
3435 }
3436
3437 /*
3438  * Called when we want to mark the current iclog as being ready to sync to
3439  * disk.
3440  */
3441 STATIC void
3442 xlog_state_want_sync(
3443         struct xlog             *log,
3444         struct xlog_in_core     *iclog)
3445 {
3446         assert_spin_locked(&log->l_icloglock);
3447
3448         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3449                 xlog_state_switch_iclogs(log, iclog, 0);
3450         } else {
3451                 ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
3452                        iclog->ic_state == XLOG_STATE_IOERROR);
3453         }
3454 }
3455
3456
3457 /*****************************************************************************
3458  *
3459  *              TICKET functions
3460  *
3461  *****************************************************************************
3462  */
3463
3464 /*
3465  * Free a used ticket when its refcount falls to zero.
3466  */
3467 void
3468 xfs_log_ticket_put(
3469         xlog_ticket_t   *ticket)
3470 {
3471         ASSERT(atomic_read(&ticket->t_ref) > 0);
3472         if (atomic_dec_and_test(&ticket->t_ref))
3473                 kmem_cache_free(xfs_log_ticket_zone, ticket);
3474 }
3475
3476 xlog_ticket_t *
3477 xfs_log_ticket_get(
3478         xlog_ticket_t   *ticket)
3479 {
3480         ASSERT(atomic_read(&ticket->t_ref) > 0);
3481         atomic_inc(&ticket->t_ref);
3482         return ticket;
3483 }
3484
3485 /*
3486  * Figure out the total log space unit (in bytes) that would be
3487  * required for a log ticket.
3488  */
3489 int
3490 xfs_log_calc_unit_res(
3491         struct xfs_mount        *mp,
3492         int                     unit_bytes)
3493 {
3494         struct xlog             *log = mp->m_log;
3495         int                     iclog_space;
3496         uint                    num_headers;
3497
3498         /*
3499          * Permanent reservations have up to 'cnt'-1 active log operations
3500          * in the log.  A unit in this case is the amount of space for one
3501          * of these log operations.  Normal reservations have a cnt of 1
3502          * and their unit amount is the total amount of space required.
3503          *
3504          * The following lines of code account for non-transaction data
3505          * which occupy space in the on-disk log.
3506          *
3507          * Normal form of a transaction is:
3508          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3509          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3510          *
3511          * We need to account for all the leadup data and trailer data
3512          * around the transaction data.
3513          * And then we need to account for the worst case in terms of using
3514          * more space.
3515          * The worst case will happen if:
3516          * - the placement of the transaction happens to be such that the
3517          *   roundoff is at its maximum
3518          * - the transaction data is synced before the commit record is synced
3519          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3520          *   Therefore the commit record is in its own Log Record.
3521          *   This can happen as the commit record is called with its
3522          *   own region to xlog_write().
3523          *   This then means that in the worst case, roundoff can happen for
3524          *   the commit-rec as well.
3525          *   The commit-rec is smaller than padding in this scenario and so it is
3526          *   not added separately.
3527          */
3528
3529         /* for trans header */
3530         unit_bytes += sizeof(xlog_op_header_t);
3531         unit_bytes += sizeof(xfs_trans_header_t);
3532
3533         /* for start-rec */
3534         unit_bytes += sizeof(xlog_op_header_t);
3535
3536         /*
3537          * for LR headers - the space for data in an iclog is the size minus
3538          * the space used for the headers. If we use the iclog size, then we
3539          * undercalculate the number of headers required.
3540          *
3541          * Furthermore - the addition of op headers for split-recs might
3542          * increase the space required enough to require more log and op
3543          * headers, so take that into account too.
3544          *
3545          * IMPORTANT: This reservation makes the assumption that if this
3546          * transaction is the first in an iclog and hence has the LR headers
3547          * accounted to it, then the remaining space in the iclog is
3548          * exclusively for this transaction.  i.e. if the transaction is larger
3549          * than the iclog, it will be the only thing in that iclog.
3550          * Fundamentally, this means we must pass the entire log vector to
3551          * xlog_write to guarantee this.
3552          */
3553         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3554         num_headers = howmany(unit_bytes, iclog_space);
3555
3556         /* for split-recs - ophdrs added when data split over LRs */
3557         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3558
3559         /* add extra header reservations if we overrun */
3560         while (!num_headers ||
3561                howmany(unit_bytes, iclog_space) > num_headers) {
3562                 unit_bytes += sizeof(xlog_op_header_t);
3563                 num_headers++;
3564         }
3565         unit_bytes += log->l_iclog_hsize * num_headers;
3566
3567         /* for commit-rec LR header - note: padding will subsume the ophdr */
3568         unit_bytes += log->l_iclog_hsize;
3569
3570         /* for roundoff padding for transaction data and one for commit record */
3571         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3572                 /* log su roundoff */
3573                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3574         } else {
3575                 /* BB roundoff */
3576                 unit_bytes += 2 * BBSIZE;
3577         }
3578
3579         return unit_bytes;
3580 }
3581
3582 /*
3583  * Allocate and initialise a new log ticket.
3584  */
3585 struct xlog_ticket *
3586 xlog_ticket_alloc(
3587         struct xlog             *log,
3588         int                     unit_bytes,
3589         int                     cnt,
3590         char                    client,
3591         bool                    permanent,
3592         xfs_km_flags_t          alloc_flags)
3593 {
3594         struct xlog_ticket      *tic;
3595         int                     unit_res;
3596
3597         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3598         if (!tic)
3599                 return NULL;
3600
3601         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3602
3603         atomic_set(&tic->t_ref, 1);
3604         tic->t_task             = current;
3605         INIT_LIST_HEAD(&tic->t_queue);
3606         tic->t_unit_res         = unit_res;
3607         tic->t_curr_res         = unit_res;
3608         tic->t_cnt              = cnt;
3609         tic->t_ocnt             = cnt;
3610         tic->t_tid              = prandom_u32();
3611         tic->t_clientid         = client;
3612         tic->t_flags            = XLOG_TIC_INITED;
3613         if (permanent)
3614                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3615
3616         xlog_tic_reset_res(tic);
3617
3618         return tic;
3619 }
3620
3621
3622 /******************************************************************************
3623  *
3624  *              Log debug routines
3625  *
3626  ******************************************************************************
3627  */
3628 #if defined(DEBUG)
3629 /*
3630  * Make sure that the destination ptr is within the valid data region of
3631  * one of the iclogs.  This uses backup pointers stored in a different
3632  * part of the log in case we trash the log structure.
3633  */
3634 STATIC void
3635 xlog_verify_dest_ptr(
3636         struct xlog     *log,
3637         void            *ptr)
3638 {
3639         int i;
3640         int good_ptr = 0;
3641
3642         for (i = 0; i < log->l_iclog_bufs; i++) {
3643                 if (ptr >= log->l_iclog_bak[i] &&
3644                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3645                         good_ptr++;
3646         }
3647
3648         if (!good_ptr)
3649                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3650 }
3651
3652 /*
3653  * Check to make sure the grant write head didn't just over lap the tail.  If
3654  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3655  * the cycles differ by exactly one and check the byte count.
3656  *
3657  * This check is run unlocked, so can give false positives. Rather than assert
3658  * on failures, use a warn-once flag and a panic tag to allow the admin to
3659  * determine if they want to panic the machine when such an error occurs. For
3660  * debug kernels this will have the same effect as using an assert but, unlinke
3661  * an assert, it can be turned off at runtime.
3662  */
3663 STATIC void
3664 xlog_verify_grant_tail(
3665         struct xlog     *log)
3666 {
3667         int             tail_cycle, tail_blocks;
3668         int             cycle, space;
3669
3670         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3671         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3672         if (tail_cycle != cycle) {
3673                 if (cycle - 1 != tail_cycle &&
3674                     !(log->l_flags & XLOG_TAIL_WARN)) {
3675                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3676                                 "%s: cycle - 1 != tail_cycle", __func__);
3677                         log->l_flags |= XLOG_TAIL_WARN;
3678                 }
3679
3680                 if (space > BBTOB(tail_blocks) &&
3681                     !(log->l_flags & XLOG_TAIL_WARN)) {
3682                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3683                                 "%s: space > BBTOB(tail_blocks)", __func__);
3684                         log->l_flags |= XLOG_TAIL_WARN;
3685                 }
3686         }
3687 }
3688
3689 /* check if it will fit */
3690 STATIC void
3691 xlog_verify_tail_lsn(
3692         struct xlog             *log,
3693         struct xlog_in_core     *iclog,
3694         xfs_lsn_t               tail_lsn)
3695 {
3696     int blocks;
3697
3698     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3699         blocks =
3700             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3701         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3702                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3703     } else {
3704         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3705
3706         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3707                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3708
3709         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3710         if (blocks < BTOBB(iclog->ic_offset) + 1)
3711                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3712     }
3713 }       /* xlog_verify_tail_lsn */
3714
3715 /*
3716  * Perform a number of checks on the iclog before writing to disk.
3717  *
3718  * 1. Make sure the iclogs are still circular
3719  * 2. Make sure we have a good magic number
3720  * 3. Make sure we don't have magic numbers in the data
3721  * 4. Check fields of each log operation header for:
3722  *      A. Valid client identifier
3723  *      B. tid ptr value falls in valid ptr space (user space code)
3724  *      C. Length in log record header is correct according to the
3725  *              individual operation headers within record.
3726  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3727  *      log, check the preceding blocks of the physical log to make sure all
3728  *      the cycle numbers agree with the current cycle number.
3729  */
3730 STATIC void
3731 xlog_verify_iclog(
3732         struct xlog             *log,
3733         struct xlog_in_core     *iclog,
3734         int                     count)
3735 {
3736         xlog_op_header_t        *ophead;
3737         xlog_in_core_t          *icptr;
3738         xlog_in_core_2_t        *xhdr;
3739         void                    *base_ptr, *ptr, *p;
3740         ptrdiff_t               field_offset;
3741         uint8_t                 clientid;
3742         int                     len, i, j, k, op_len;
3743         int                     idx;
3744
3745         /* check validity of iclog pointers */
3746         spin_lock(&log->l_icloglock);
3747         icptr = log->l_iclog;
3748         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3749                 ASSERT(icptr);
3750
3751         if (icptr != log->l_iclog)
3752                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3753         spin_unlock(&log->l_icloglock);
3754
3755         /* check log magic numbers */
3756         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3757                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3758
3759         base_ptr = ptr = &iclog->ic_header;
3760         p = &iclog->ic_header;
3761         for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3762                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3763                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3764                                 __func__);
3765         }
3766
3767         /* check fields */
3768         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3769         base_ptr = ptr = iclog->ic_datap;
3770         ophead = ptr;
3771         xhdr = iclog->ic_data;
3772         for (i = 0; i < len; i++) {
3773                 ophead = ptr;
3774
3775                 /* clientid is only 1 byte */
3776                 p = &ophead->oh_clientid;
3777                 field_offset = p - base_ptr;
3778                 if (field_offset & 0x1ff) {
3779                         clientid = ophead->oh_clientid;
3780                 } else {
3781                         idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3782                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3783                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3784                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3785                                 clientid = xlog_get_client_id(
3786                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3787                         } else {
3788                                 clientid = xlog_get_client_id(
3789                                         iclog->ic_header.h_cycle_data[idx]);
3790                         }
3791                 }
3792                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3793                         xfs_warn(log->l_mp,
3794                                 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3795                                 __func__, clientid, ophead,
3796                                 (unsigned long)field_offset);
3797
3798                 /* check length */
3799                 p = &ophead->oh_len;
3800                 field_offset = p - base_ptr;
3801                 if (field_offset & 0x1ff) {
3802                         op_len = be32_to_cpu(ophead->oh_len);
3803                 } else {
3804                         idx = BTOBBT((uintptr_t)&ophead->oh_len -
3805                                     (uintptr_t)iclog->ic_datap);
3806                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3807                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3808                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3809                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3810                         } else {
3811                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3812                         }
3813                 }
3814                 ptr += sizeof(xlog_op_header_t) + op_len;
3815         }
3816 }       /* xlog_verify_iclog */
3817 #endif
3818
3819 /*
3820  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3821  */
3822 STATIC int
3823 xlog_state_ioerror(
3824         struct xlog     *log)
3825 {
3826         xlog_in_core_t  *iclog, *ic;
3827
3828         iclog = log->l_iclog;
3829         if (iclog->ic_state != XLOG_STATE_IOERROR) {
3830                 /*
3831                  * Mark all the incore logs IOERROR.
3832                  * From now on, no log flushes will result.
3833                  */
3834                 ic = iclog;
3835                 do {
3836                         ic->ic_state = XLOG_STATE_IOERROR;
3837                         ic = ic->ic_next;
3838                 } while (ic != iclog);
3839                 return 0;
3840         }
3841         /*
3842          * Return non-zero, if state transition has already happened.
3843          */
3844         return 1;
3845 }
3846
3847 /*
3848  * This is called from xfs_force_shutdown, when we're forcibly
3849  * shutting down the filesystem, typically because of an IO error.
3850  * Our main objectives here are to make sure that:
3851  *      a. if !logerror, flush the logs to disk. Anything modified
3852  *         after this is ignored.
3853  *      b. the filesystem gets marked 'SHUTDOWN' for all interested
3854  *         parties to find out, 'atomically'.
3855  *      c. those who're sleeping on log reservations, pinned objects and
3856  *          other resources get woken up, and be told the bad news.
3857  *      d. nothing new gets queued up after (b) and (c) are done.
3858  *
3859  * Note: for the !logerror case we need to flush the regions held in memory out
3860  * to disk first. This needs to be done before the log is marked as shutdown,
3861  * otherwise the iclog writes will fail.
3862  */
3863 int
3864 xfs_log_force_umount(
3865         struct xfs_mount        *mp,
3866         int                     logerror)
3867 {
3868         struct xlog     *log;
3869         int             retval;
3870
3871         log = mp->m_log;
3872
3873         /*
3874          * If this happens during log recovery, don't worry about
3875          * locking; the log isn't open for business yet.
3876          */
3877         if (!log ||
3878             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3879                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3880                 if (mp->m_sb_bp)
3881                         mp->m_sb_bp->b_flags |= XBF_DONE;
3882                 return 0;
3883         }
3884
3885         /*
3886          * Somebody could've already done the hard work for us.
3887          * No need to get locks for this.
3888          */
3889         if (logerror && log->l_iclog->ic_state == XLOG_STATE_IOERROR) {
3890                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3891                 return 1;
3892         }
3893
3894         /*
3895          * Flush all the completed transactions to disk before marking the log
3896          * being shut down. We need to do it in this order to ensure that
3897          * completed operations are safely on disk before we shut down, and that
3898          * we don't have to issue any buffer IO after the shutdown flags are set
3899          * to guarantee this.
3900          */
3901         if (!logerror)
3902                 xfs_log_force(mp, XFS_LOG_SYNC);
3903
3904         /*
3905          * mark the filesystem and the as in a shutdown state and wake
3906          * everybody up to tell them the bad news.
3907          */
3908         spin_lock(&log->l_icloglock);
3909         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3910         if (mp->m_sb_bp)
3911                 mp->m_sb_bp->b_flags |= XBF_DONE;
3912
3913         /*
3914          * Mark the log and the iclogs with IO error flags to prevent any
3915          * further log IO from being issued or completed.
3916          */
3917         log->l_flags |= XLOG_IO_ERROR;
3918         retval = xlog_state_ioerror(log);
3919         spin_unlock(&log->l_icloglock);
3920
3921         /*
3922          * We don't want anybody waiting for log reservations after this. That
3923          * means we have to wake up everybody queued up on reserveq as well as
3924          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3925          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3926          * action is protected by the grant locks.
3927          */
3928         xlog_grant_head_wake_all(&log->l_reserve_head);
3929         xlog_grant_head_wake_all(&log->l_write_head);
3930
3931         /*
3932          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3933          * as if the log writes were completed. The abort handling in the log
3934          * item committed callback functions will do this again under lock to
3935          * avoid races.
3936          */
3937         spin_lock(&log->l_cilp->xc_push_lock);
3938         wake_up_all(&log->l_cilp->xc_commit_wait);
3939         spin_unlock(&log->l_cilp->xc_push_lock);
3940         xlog_state_do_callback(log, true);
3941
3942         /* return non-zero if log IOERROR transition had already happened */
3943         return retval;
3944 }
3945
3946 STATIC int
3947 xlog_iclogs_empty(
3948         struct xlog     *log)
3949 {
3950         xlog_in_core_t  *iclog;
3951
3952         iclog = log->l_iclog;
3953         do {
3954                 /* endianness does not matter here, zero is zero in
3955                  * any language.
3956                  */
3957                 if (iclog->ic_header.h_num_logops)
3958                         return 0;
3959                 iclog = iclog->ic_next;
3960         } while (iclog != log->l_iclog);
3961         return 1;
3962 }
3963
3964 /*
3965  * Verify that an LSN stamped into a piece of metadata is valid. This is
3966  * intended for use in read verifiers on v5 superblocks.
3967  */
3968 bool
3969 xfs_log_check_lsn(
3970         struct xfs_mount        *mp,
3971         xfs_lsn_t               lsn)
3972 {
3973         struct xlog             *log = mp->m_log;
3974         bool                    valid;
3975
3976         /*
3977          * norecovery mode skips mount-time log processing and unconditionally
3978          * resets the in-core LSN. We can't validate in this mode, but
3979          * modifications are not allowed anyways so just return true.
3980          */
3981         if (mp->m_flags & XFS_MOUNT_NORECOVERY)
3982                 return true;
3983
3984         /*
3985          * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
3986          * handled by recovery and thus safe to ignore here.
3987          */
3988         if (lsn == NULLCOMMITLSN)
3989                 return true;
3990
3991         valid = xlog_valid_lsn(mp->m_log, lsn);
3992
3993         /* warn the user about what's gone wrong before verifier failure */
3994         if (!valid) {
3995                 spin_lock(&log->l_icloglock);
3996                 xfs_warn(mp,
3997 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
3998 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
3999                          CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4000                          log->l_curr_cycle, log->l_curr_block);
4001                 spin_unlock(&log->l_icloglock);
4002         }
4003
4004         return valid;
4005 }
4006
4007 bool
4008 xfs_log_in_recovery(
4009         struct xfs_mount        *mp)
4010 {
4011         struct xlog             *log = mp->m_log;
4012
4013         return log->l_flags & XLOG_ACTIVE_RECOVERY;
4014 }