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