fs/xfs/xfs_log.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   4  * All Rights Reserved.
   5  */
   6 #include "xfs.h"
   7 #include "xfs_fs.h"
   8 #include "xfs_shared.h"
   9 #include "xfs_format.h"
  10 #include "xfs_log_format.h"
  11 #include "xfs_trans_resv.h"
  12 #include "xfs_mount.h"
  13 #include "xfs_errortag.h"
  14 #include "xfs_error.h"
  15 #include "xfs_trans.h"
  16 #include "xfs_trans_priv.h"
  17 #include "xfs_log.h"
  18 #include "xfs_log_priv.h"
  19 #include "xfs_trace.h"
  20 #include "xfs_sysfs.h"
  21 #include "xfs_sb.h"
  22 #include "xfs_health.h"
  23
  24 kmem_zone_t     *xfs_log_ticket_zone;
  25
  26 /* Local miscellaneous function prototypes */
  27 STATIC int
  28 xlog_commit_record(
  29         struct xlog             *log,
  30         struct xlog_ticket      *ticket,
  31         struct xlog_in_core     **iclog,
  32         xfs_lsn_t               *commitlsnp);
  33
  34 STATIC struct xlog *
  35 xlog_alloc_log(
  36         struct xfs_mount        *mp,
  37         struct xfs_buftarg      *log_target,
  38         xfs_daddr_t             blk_offset,
  39         int                     num_bblks);
  40 STATIC int
  41 xlog_space_left(
  42         struct xlog             *log,
  43         atomic64_t              *head);
  44 STATIC void
  45 xlog_dealloc_log(
  46         struct xlog             *log);
  47
  48 /* local state machine functions */
  49 STATIC void xlog_state_done_syncing(
  50         struct xlog_in_core     *iclog,
  51         bool                    aborted);
  52 STATIC int
  53 xlog_state_get_iclog_space(
  54         struct xlog             *log,
  55         int                     len,
  56         struct xlog_in_core     **iclog,
  57         struct xlog_ticket      *ticket,
  58         int                     *continued_write,
  59         int                     *logoffsetp);
  60 STATIC int
  61 xlog_state_release_iclog(
  62         struct xlog             *log,
  63         struct xlog_in_core     *iclog);
  64 STATIC void
  65 xlog_state_switch_iclogs(
  66         struct xlog             *log,
  67         struct xlog_in_core     *iclog,
  68         int                     eventual_size);
  69 STATIC void
  70 xlog_state_want_sync(
  71         struct xlog             *log,
  72         struct xlog_in_core     *iclog);
  73
  74 STATIC void
  75 xlog_grant_push_ail(
  76         struct xlog             *log,
  77         int                     need_bytes);
  78 STATIC void
  79 xlog_regrant_reserve_log_space(
  80         struct xlog             *log,
  81         struct xlog_ticket      *ticket);
  82 STATIC void
  83 xlog_ungrant_log_space(
  84         struct xlog             *log,
  85         struct xlog_ticket      *ticket);
  86
  87 #if defined(DEBUG)
  88 STATIC void
  89 xlog_verify_dest_ptr(
  90         struct xlog             *log,
  91         void                    *ptr);
  92 STATIC void
  93 xlog_verify_grant_tail(
  94         struct xlog *log);
  95 STATIC void
  96 xlog_verify_iclog(
  97         struct xlog             *log,
  98         struct xlog_in_core     *iclog,
  99         int                     count);
 100 STATIC void
 101 xlog_verify_tail_lsn(
 102         struct xlog             *log,
 103         struct xlog_in_core     *iclog,
 104         xfs_lsn_t               tail_lsn);
 105 #else
 106 #define xlog_verify_dest_ptr(a,b)
 107 #define xlog_verify_grant_tail(a)
 108 #define xlog_verify_iclog(a,b,c)
 109 #define xlog_verify_tail_lsn(a,b,c)
 110 #endif
 111
 112 STATIC int
 113 xlog_iclogs_empty(
 114         struct xlog             *log);
 115
 116 static void
 117 xlog_grant_sub_space(
 118         struct xlog             *log,
 119         atomic64_t              *head,
 120         int                     bytes)
 121 {
 122         int64_t head_val = atomic64_read(head);
 123         int64_t new, old;
 124
 125         do {
 126                 int     cycle, space;
 127
 128                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 129
 130                 space -= bytes;
 131                 if (space < 0) {
 132                         space += log->l_logsize;
 133                         cycle--;
 134                 }
 135
 136                 old = head_val;
 137                 new = xlog_assign_grant_head_val(cycle, space);
 138                 head_val = atomic64_cmpxchg(head, old, new);
 139         } while (head_val != old);
 140 }
 141
 142 static void
 143 xlog_grant_add_space(
 144         struct xlog             *log,
 145         atomic64_t              *head,
 146         int                     bytes)
 147 {
 148         int64_t head_val = atomic64_read(head);
 149         int64_t new, old;
 150
 151         do {
 152                 int             tmp;
 153                 int             cycle, space;
 154
 155                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 156
 157                 tmp = log->l_logsize - space;
 158                 if (tmp > bytes)
 159                         space += bytes;
 160                 else {
 161                         space = bytes - tmp;
 162                         cycle++;
 163                 }
 164
 165                 old = head_val;
 166                 new = xlog_assign_grant_head_val(cycle, space);
 167                 head_val = atomic64_cmpxchg(head, old, new);
 168         } while (head_val != old);
 169 }
 170
 171 STATIC void
 172 xlog_grant_head_init(
 173         struct xlog_grant_head  *head)
 174 {
 175         xlog_assign_grant_head(&head->grant, 1, 0);
 176         INIT_LIST_HEAD(&head->waiters);
 177         spin_lock_init(&head->lock);
 178 }
 179
 180 STATIC void
 181 xlog_grant_head_wake_all(
 182         struct xlog_grant_head  *head)
 183 {
 184         struct xlog_ticket      *tic;
 185
 186         spin_lock(&head->lock);
 187         list_for_each_entry(tic, &head->waiters, t_queue)
 188                 wake_up_process(tic->t_task);
 189         spin_unlock(&head->lock);
 190 }
 191
 192 static inline int
 193 xlog_ticket_reservation(
 194         struct xlog             *log,
 195         struct xlog_grant_head  *head,
 196         struct xlog_ticket      *tic)
 197 {
 198         if (head == &log->l_write_head) {
 199                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 200                 return tic->t_unit_res;
 201         } else {
 202                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 203                         return tic->t_unit_res * tic->t_cnt;
 204                 else
 205                         return tic->t_unit_res;
 206         }
 207 }
 208
 209 STATIC bool
 210 xlog_grant_head_wake(
 211         struct xlog             *log,
 212         struct xlog_grant_head  *head,
 213         int                     *free_bytes)
 214 {
 215         struct xlog_ticket      *tic;
 216         int                     need_bytes;
 217         bool                    woken_task = false;
 218
 219         list_for_each_entry(tic, &head->waiters, t_queue) {
 220
 221                 /*
 222                  * There is a chance that the size of the CIL checkpoints in
 223                  * progress at the last AIL push target calculation resulted in
 224                  * limiting the target to the log head (l_last_sync_lsn) at the
 225                  * time. This may not reflect where the log head is now as the
 226                  * CIL checkpoints may have completed.
 227                  *
 228                  * Hence when we are woken here, it may be that the head of the
 229                  * log that has moved rather than the tail. As the tail didn't
 230                  * move, there still won't be space available for the
 231                  * reservation we require.  However, if the AIL has already
 232                  * pushed to the target defined by the old log head location, we
 233                  * will hang here waiting for something else to update the AIL
 234                  * push target.
 235                  *
 236                  * Therefore, if there isn't space to wake the first waiter on
 237                  * the grant head, we need to push the AIL again to ensure the
 238                  * target reflects both the current log tail and log head
 239                  * position before we wait for the tail to move again.
 240                  */
 241
 242                 need_bytes = xlog_ticket_reservation(log, head, tic);
 243                 if (*free_bytes < need_bytes) {
 244                         if (!woken_task)
 245                                 xlog_grant_push_ail(log, need_bytes);
 246                         return false;
 247                 }
 248
 249                 *free_bytes -= need_bytes;
 250                 trace_xfs_log_grant_wake_up(log, tic);
 251                 wake_up_process(tic->t_task);
 252                 woken_task = true;
 253         }
 254
 255         return true;
 256 }
 257
 258 STATIC int
 259 xlog_grant_head_wait(
 260         struct xlog             *log,
 261         struct xlog_grant_head  *head,
 262         struct xlog_ticket      *tic,
 263         int                     need_bytes) __releases(&head->lock)
 264                                             __acquires(&head->lock)
 265 {
 266         list_add_tail(&tic->t_queue, &head->waiters);
 267
 268         do {
 269                 if (XLOG_FORCED_SHUTDOWN(log))
 270                         goto shutdown;
 271                 xlog_grant_push_ail(log, need_bytes);
 272
 273                 __set_current_state(TASK_UNINTERRUPTIBLE);
 274                 spin_unlock(&head->lock);
 275
 276                 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
 277
 278                 trace_xfs_log_grant_sleep(log, tic);
 279                 schedule();
 280                 trace_xfs_log_grant_wake(log, tic);
 281
 282                 spin_lock(&head->lock);
 283                 if (XLOG_FORCED_SHUTDOWN(log))
 284                         goto shutdown;
 285         } while (xlog_space_left(log, &head->grant) < need_bytes);
 286
 287         list_del_init(&tic->t_queue);
 288         return 0;
 289 shutdown:
 290         list_del_init(&tic->t_queue);
 291         return -EIO;
 292 }
 293
 294 /*
 295  * Atomically get the log space required for a log ticket.
 296  *
 297  * Once a ticket gets put onto head->waiters, it will only return after the
 298  * needed reservation is satisfied.
 299  *
 300  * This function is structured so that it has a lock free fast path. This is
 301  * necessary because every new transaction reservation will come through this
 302  * path. Hence any lock will be globally hot if we take it unconditionally on
 303  * every pass.
 304  *
 305  * As tickets are only ever moved on and off head->waiters under head->lock, we
 306  * only need to take that lock if we are going to add the ticket to the queue
 307  * and sleep. We can avoid taking the lock if the ticket was never added to
 308  * head->waiters because the t_queue list head will be empty and we hold the
 309  * only reference to it so it can safely be checked unlocked.
 310  */
 311 STATIC int
 312 xlog_grant_head_check(
 313         struct xlog             *log,
 314         struct xlog_grant_head  *head,
 315         struct xlog_ticket      *tic,
 316         int                     *need_bytes)
 317 {
 318         int                     free_bytes;
 319         int                     error = 0;
 320
 321         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 322
 323         /*
 324          * If there are other waiters on the queue then give them a chance at
 325          * logspace before us.  Wake up the first waiters, if we do not wake
 326          * up all the waiters then go to sleep waiting for more free space,
 327          * otherwise try to get some space for this transaction.
 328          */
 329         *need_bytes = xlog_ticket_reservation(log, head, tic);
 330         free_bytes = xlog_space_left(log, &head->grant);
 331         if (!list_empty_careful(&head->waiters)) {
 332                 spin_lock(&head->lock);
 333                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 334                     free_bytes < *need_bytes) {
 335                         error = xlog_grant_head_wait(log, head, tic,
 336                                                      *need_bytes);
 337                 }
 338                 spin_unlock(&head->lock);
 339         } else if (free_bytes < *need_bytes) {
 340                 spin_lock(&head->lock);
 341                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 342                 spin_unlock(&head->lock);
 343         }
 344
 345         return error;
 346 }
 347
 348 static void
 349 xlog_tic_reset_res(xlog_ticket_t *tic)
 350 {
 351         tic->t_res_num = 0;
 352         tic->t_res_arr_sum = 0;
 353         tic->t_res_num_ophdrs = 0;
 354 }
 355
 356 static void
 357 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 358 {
 359         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 360                 /* add to overflow and start again */
 361                 tic->t_res_o_flow += tic->t_res_arr_sum;
 362                 tic->t_res_num = 0;
 363                 tic->t_res_arr_sum = 0;
 364         }
 365
 366         tic->t_res_arr[tic->t_res_num].r_len = len;
 367         tic->t_res_arr[tic->t_res_num].r_type = type;
 368         tic->t_res_arr_sum += len;
 369         tic->t_res_num++;
 370 }
 371
 372 bool
 373 xfs_log_writable(
 374         struct xfs_mount        *mp)
 375 {
 376         /*
 377          * Never write to the log on norecovery mounts, if the block device is
 378          * read-only, or if the filesystem is shutdown. Read-only mounts still
 379          * allow internal writes for log recovery and unmount purposes, so don't
 380          * restrict that case here.
 381          */
 382         if (mp->m_flags & XFS_MOUNT_NORECOVERY)
 383                 return false;
 384         if (xfs_readonly_buftarg(mp->m_log->l_targ))
 385                 return false;
 386         if (XFS_FORCED_SHUTDOWN(mp))
 387                 return false;
 388         return true;
 389 }
 390
 391 /*
 392  * Replenish the byte reservation required by moving the grant write head.
 393  */
 394 int
 395 xfs_log_regrant(
 396         struct xfs_mount        *mp,
 397         struct xlog_ticket      *tic)
 398 {
 399         struct xlog             *log = mp->m_log;
 400         int                     need_bytes;
 401         int                     error = 0;
 402
 403         if (XLOG_FORCED_SHUTDOWN(log))
 404                 return -EIO;
 405
 406         XFS_STATS_INC(mp, xs_try_logspace);
 407
 408         /*
 409          * This is a new transaction on the ticket, so we need to change the
 410          * transaction ID so that the next transaction has a different TID in
 411          * the log. Just add one to the existing tid so that we can see chains
 412          * of rolling transactions in the log easily.
 413          */
 414         tic->t_tid++;
 415
 416         xlog_grant_push_ail(log, tic->t_unit_res);
 417
 418         tic->t_curr_res = tic->t_unit_res;
 419         xlog_tic_reset_res(tic);
 420
 421         if (tic->t_cnt > 0)
 422                 return 0;
 423
 424         trace_xfs_log_regrant(log, tic);
 425
 426         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 427                                       &need_bytes);
 428         if (error)
 429                 goto out_error;
 430
 431         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 432         trace_xfs_log_regrant_exit(log, tic);
 433         xlog_verify_grant_tail(log);
 434         return 0;
 435
 436 out_error:
 437         /*
 438          * If we are failing, make sure the ticket doesn't have any current
 439          * reservations.  We don't want to add this back when the ticket/
 440          * transaction gets cancelled.
 441          */
 442         tic->t_curr_res = 0;
 443         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 444         return error;
 445 }
 446
 447 /*
 448  * Reserve log space and return a ticket corresponding to the reservation.
 449  *
 450  * Each reservation is going to reserve extra space for a log record header.
 451  * When writes happen to the on-disk log, we don't subtract the length of the
 452  * log record header from any reservation.  By wasting space in each
 453  * reservation, we prevent over allocation problems.
 454  */
 455 int
 456 xfs_log_reserve(
 457         struct xfs_mount        *mp,
 458         int                     unit_bytes,
 459         int                     cnt,
 460         struct xlog_ticket      **ticp,
 461         uint8_t                 client,
 462         bool                    permanent)
 463 {
 464         struct xlog             *log = mp->m_log;
 465         struct xlog_ticket      *tic;
 466         int                     need_bytes;
 467         int                     error = 0;
 468
 469         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 470
 471         if (XLOG_FORCED_SHUTDOWN(log))
 472                 return -EIO;
 473
 474         XFS_STATS_INC(mp, xs_try_logspace);
 475
 476         ASSERT(*ticp == NULL);
 477         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent, 0);
 478         *ticp = tic;
 479
 480         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 481                                             : tic->t_unit_res);
 482
 483         trace_xfs_log_reserve(log, tic);
 484
 485         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 486                                       &need_bytes);
 487         if (error)
 488                 goto out_error;
 489
 490         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 491         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 492         trace_xfs_log_reserve_exit(log, tic);
 493         xlog_verify_grant_tail(log);
 494         return 0;
 495
 496 out_error:
 497         /*
 498          * If we are failing, make sure the ticket doesn't have any current
 499          * reservations.  We don't want to add this back when the ticket/
 500          * transaction gets cancelled.
 501          */
 502         tic->t_curr_res = 0;
 503         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 504         return error;
 505 }
 506
 507
 508 /*
 509  * NOTES:
 510  *
 511  *      1. currblock field gets updated at startup and after in-core logs
 512  *              marked as with WANT_SYNC.
 513  */
 514
 515 /*
 516  * This routine is called when a user of a log manager ticket is done with
 517  * the reservation.  If the ticket was ever used, then a commit record for
 518  * the associated transaction is written out as a log operation header with
 519  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 520  * a given ticket.  If the ticket was one with a permanent reservation, then
 521  * a few operations are done differently.  Permanent reservation tickets by
 522  * default don't release the reservation.  They just commit the current
 523  * transaction with the belief that the reservation is still needed.  A flag
 524  * must be passed in before permanent reservations are actually released.
 525  * When these type of tickets are not released, they need to be set into
 526  * the inited state again.  By doing this, a start record will be written
 527  * out when the next write occurs.
 528  */
 529 xfs_lsn_t
 530 xfs_log_done(
 531         struct xfs_mount        *mp,
 532         struct xlog_ticket      *ticket,
 533         struct xlog_in_core     **iclog,
 534         bool                    regrant)
 535 {
 536         struct xlog             *log = mp->m_log;
 537         xfs_lsn_t               lsn = 0;
 538
 539         if (XLOG_FORCED_SHUTDOWN(log) ||
 540             /*
 541              * If nothing was ever written, don't write out commit record.
 542              * If we get an error, just continue and give back the log ticket.
 543              */
 544             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 545              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 546                 lsn = (xfs_lsn_t) -1;
 547                 regrant = false;
 548         }
 549
 550
 551         if (!regrant) {
 552                 trace_xfs_log_done_nonperm(log, ticket);
 553
 554                 /*
 555                  * Release ticket if not permanent reservation or a specific
 556                  * request has been made to release a permanent reservation.
 557                  */
 558                 xlog_ungrant_log_space(log, ticket);
 559         } else {
 560                 trace_xfs_log_done_perm(log, ticket);
 561
 562                 xlog_regrant_reserve_log_space(log, ticket);
 563                 /* If this ticket was a permanent reservation and we aren't
 564                  * trying to release it, reset the inited flags; so next time
 565                  * we write, a start record will be written out.
 566                  */
 567                 ticket->t_flags |= XLOG_TIC_INITED;
 568         }
 569
 570         xfs_log_ticket_put(ticket);
 571         return lsn;
 572 }
 573
 574 int
 575 xfs_log_release_iclog(
 576         struct xfs_mount        *mp,
 577         struct xlog_in_core     *iclog)
 578 {
 579         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 580                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 581                 return -EIO;
 582         }
 583
 584         return 0;
 585 }
 586
 587 /*
 588  * Mount a log filesystem
 589  *
 590  * mp           - ubiquitous xfs mount point structure
 591  * log_target   - buftarg of on-disk log device
 592  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 593  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 594  *
 595  * Return error or zero.
 596  */
 597 int
 598 xfs_log_mount(
 599         xfs_mount_t     *mp,
 600         xfs_buftarg_t   *log_target,
 601         xfs_daddr_t     blk_offset,
 602         int             num_bblks)
 603 {
 604         bool            fatal = xfs_sb_version_hascrc(&mp->m_sb);
 605         int             error = 0;
 606         int             min_logfsbs;
 607
 608         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 609                 xfs_notice(mp, "Mounting V%d Filesystem",
 610                            XFS_SB_VERSION_NUM(&mp->m_sb));
 611         } else {
 612                 xfs_notice(mp,
 613 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
 614                            XFS_SB_VERSION_NUM(&mp->m_sb));
 615                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 616         }
 617
 618         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 619         if (IS_ERR(mp->m_log)) {
 620                 error = PTR_ERR(mp->m_log);
 621                 goto out;
 622         }
 623
 624         /*
 625          * Validate the given log space and drop a critical message via syslog
 626          * if the log size is too small that would lead to some unexpected
 627          * situations in transaction log space reservation stage.
 628          *
 629          * Note: we can't just reject the mount if the validation fails.  This
 630          * would mean that people would have to downgrade their kernel just to
 631          * remedy the situation as there is no way to grow the log (short of
 632          * black magic surgery with xfs_db).
 633          *
 634          * We can, however, reject mounts for CRC format filesystems, as the
 635          * mkfs binary being used to make the filesystem should never create a
 636          * filesystem with a log that is too small.
 637          */
 638         min_logfsbs = xfs_log_calc_minimum_size(mp);
 639
 640         if (mp->m_sb.sb_logblocks < min_logfsbs) {
 641                 xfs_warn(mp,
 642                 "Log size %d blocks too small, minimum size is %d blocks",
 643                          mp->m_sb.sb_logblocks, min_logfsbs);
 644                 error = -EINVAL;
 645         } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
 646                 xfs_warn(mp,
 647                 "Log size %d blocks too large, maximum size is %lld blocks",
 648                          mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
 649                 error = -EINVAL;
 650         } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
 651                 xfs_warn(mp,
 652                 "log size %lld bytes too large, maximum size is %lld bytes",
 653                          XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
 654                          XFS_MAX_LOG_BYTES);
 655                 error = -EINVAL;
 656         } else if (mp->m_sb.sb_logsunit > 1 &&
 657                    mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
 658                 xfs_warn(mp,
 659                 "log stripe unit %u bytes must be a multiple of block size",
 660                          mp->m_sb.sb_logsunit);
 661                 error = -EINVAL;
 662                 fatal = true;
 663         }
 664         if (error) {
 665                 /*
 666                  * Log check errors are always fatal on v5; or whenever bad
 667                  * metadata leads to a crash.
 668                  */
 669                 if (fatal) {
 670                         xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 671                         ASSERT(0);
 672                         goto out_free_log;
 673                 }
 674                 xfs_crit(mp, "Log size out of supported range.");
 675                 xfs_crit(mp,
 676 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
 677         }
 678
 679         /*
 680          * Initialize the AIL now we have a log.
 681          */
 682         error = xfs_trans_ail_init(mp);
 683         if (error) {
 684                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 685                 goto out_free_log;
 686         }
 687         mp->m_log->l_ailp = mp->m_ail;
 688
 689         /*
 690          * skip log recovery on a norecovery mount.  pretend it all
 691          * just worked.
 692          */
 693         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 694                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 695
 696                 if (readonly)
 697                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 698
 699                 error = xlog_recover(mp->m_log);
 700
 701                 if (readonly)
 702                         mp->m_flags |= XFS_MOUNT_RDONLY;
 703                 if (error) {
 704                         xfs_warn(mp, "log mount/recovery failed: error %d",
 705                                 error);
 706                         xlog_recover_cancel(mp->m_log);
 707                         goto out_destroy_ail;
 708                 }
 709         }
 710
 711         error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 712                                "log");
 713         if (error)
 714                 goto out_destroy_ail;
 715
 716         /* Normal transactions can now occur */
 717         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 718
 719         /*
 720          * Now the log has been fully initialised and we know were our
 721          * space grant counters are, we can initialise the permanent ticket
 722          * needed for delayed logging to work.
 723          */
 724         xlog_cil_init_post_recovery(mp->m_log);
 725
 726         return 0;
 727
 728 out_destroy_ail:
 729         xfs_trans_ail_destroy(mp);
 730 out_free_log:
 731         xlog_dealloc_log(mp->m_log);
 732 out:
 733         return error;
 734 }
 735
 736 /*
 737  * Finish the recovery of the file system.  This is separate from the
 738  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 739  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 740  * here.
 741  *
 742  * If we finish recovery successfully, start the background log work. If we are
 743  * not doing recovery, then we have a RO filesystem and we don't need to start
 744  * it.
 745  */
 746 int
 747 xfs_log_mount_finish(
 748         struct xfs_mount        *mp)
 749 {
 750         int     error = 0;
 751         bool    readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 752         bool    recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
 753
 754         if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
 755                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 756                 return 0;
 757         } else if (readonly) {
 758                 /* Allow unlinked processing to proceed */
 759                 mp->m_flags &= ~XFS_MOUNT_RDONLY;
 760         }
 761
 762         /*
 763          * During the second phase of log recovery, we need iget and
 764          * iput to behave like they do for an active filesystem.
 765          * xfs_fs_drop_inode needs to be able to prevent the deletion
 766          * of inodes before we're done replaying log items on those
 767          * inodes.  Turn it off immediately after recovery finishes
 768          * so that we don't leak the quota inodes if subsequent mount
 769          * activities fail.
 770          *
 771          * We let all inodes involved in redo item processing end up on
 772          * the LRU instead of being evicted immediately so that if we do
 773          * something to an unlinked inode, the irele won't cause
 774          * premature truncation and freeing of the inode, which results
 775          * in log recovery failure.  We have to evict the unreferenced
 776          * lru inodes after clearing SB_ACTIVE because we don't
 777          * otherwise clean up the lru if there's a subsequent failure in
 778          * xfs_mountfs, which leads to us leaking the inodes if nothing
 779          * else (e.g. quotacheck) references the inodes before the
 780          * mount failure occurs.
 781          */
 782         mp->m_super->s_flags |= SB_ACTIVE;
 783         error = xlog_recover_finish(mp->m_log);
 784         if (!error)
 785                 xfs_log_work_queue(mp);
 786         mp->m_super->s_flags &= ~SB_ACTIVE;
 787         evict_inodes(mp->m_super);
 788
 789         /*
 790          * Drain the buffer LRU after log recovery. This is required for v4
 791          * filesystems to avoid leaving around buffers with NULL verifier ops,
 792          * but we do it unconditionally to make sure we're always in a clean
 793          * cache state after mount.
 794          *
 795          * Don't push in the error case because the AIL may have pending intents
 796          * that aren't removed until recovery is cancelled.
 797          */
 798         if (!error && recovered) {
 799                 xfs_log_force(mp, XFS_LOG_SYNC);
 800                 xfs_ail_push_all_sync(mp->m_ail);
 801         }
 802         xfs_wait_buftarg(mp->m_ddev_targp);
 803
 804         if (readonly)
 805                 mp->m_flags |= XFS_MOUNT_RDONLY;
 806
 807         return error;
 808 }
 809
 810 /*
 811  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
 812  * the log.
 813  */
 814 void
 815 xfs_log_mount_cancel(
 816         struct xfs_mount        *mp)
 817 {
 818         xlog_recover_cancel(mp->m_log);
 819         xfs_log_unmount(mp);
 820 }
 821
 822 /*
 823  * Final log writes as part of unmount.
 824  *
 825  * Mark the filesystem clean as unmount happens.  Note that during relocation
 826  * this routine needs to be executed as part of source-bag while the
 827  * deallocation must not be done until source-end.
 828  */
 829
 830 /* Actually write the unmount record to disk. */
 831 static void
 832 xfs_log_write_unmount_record(
 833         struct xfs_mount        *mp)
 834 {
 835         /* the data section must be 32 bit size aligned */
 836         struct xfs_unmount_log_format magic = {
 837                 .magic = XLOG_UNMOUNT_TYPE,
 838         };
 839         struct xfs_log_iovec reg = {
 840                 .i_addr = &magic,
 841                 .i_len = sizeof(magic),
 842                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 843         };
 844         struct xfs_log_vec vec = {
 845                 .lv_niovecs = 1,
 846                 .lv_iovecp = &reg,
 847         };
 848         struct xlog             *log = mp->m_log;
 849         struct xlog_in_core     *iclog;
 850         struct xlog_ticket      *tic = NULL;
 851         xfs_lsn_t               lsn;
 852         uint                    flags = XLOG_UNMOUNT_TRANS;
 853         int                     error;
 854
 855         error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
 856         if (error)
 857                 goto out_err;
 858
 859         /* remove inited flag, and account for space used */
 860         tic->t_flags = 0;
 861         tic->t_curr_res -= sizeof(magic);
 862         error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
 863         /*
 864          * At this point, we're umounting anyway, so there's no point in
 865          * transitioning log state to IOERROR. Just continue...
 866          */
 867 out_err:
 868         if (error)
 869                 xfs_alert(mp, "%s: unmount record failed", __func__);
 870
 871         spin_lock(&log->l_icloglock);
 872         iclog = log->l_iclog;
 873         atomic_inc(&iclog->ic_refcnt);
 874         xlog_state_want_sync(log, iclog);
 875         spin_unlock(&log->l_icloglock);
 876         error = xlog_state_release_iclog(log, iclog);
 877
 878         spin_lock(&log->l_icloglock);
 879         switch (iclog->ic_state) {
 880         default:
 881                 if (!XLOG_FORCED_SHUTDOWN(log)) {
 882                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
 883                         break;
 884                 }
 885                 /* fall through */
 886         case XLOG_STATE_ACTIVE:
 887         case XLOG_STATE_DIRTY:
 888                 spin_unlock(&log->l_icloglock);
 889                 break;
 890         }
 891
 892         if (tic) {
 893                 trace_xfs_log_umount_write(log, tic);
 894                 xlog_ungrant_log_space(log, tic);
 895                 xfs_log_ticket_put(tic);
 896         }
 897 }
 898
 899 /*
 900  * Unmount record used to have a string "Unmount filesystem--" in the
 901  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 902  * We just write the magic number now since that particular field isn't
 903  * currently architecture converted and "Unmount" is a bit foo.
 904  * As far as I know, there weren't any dependencies on the old behaviour.
 905  */
 906
 907 static int
 908 xfs_log_unmount_write(xfs_mount_t *mp)
 909 {
 910         struct xlog      *log = mp->m_log;
 911         xlog_in_core_t   *iclog;
 912 #ifdef DEBUG
 913         xlog_in_core_t   *first_iclog;
 914 #endif
 915         int              error;
 916
 917         if (!xfs_log_writable(mp))
 918                 return 0;
 919
 920         error = xfs_log_force(mp, XFS_LOG_SYNC);
 921         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 922
 923 #ifdef DEBUG
 924         first_iclog = iclog = log->l_iclog;
 925         do {
 926                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 927                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 928                         ASSERT(iclog->ic_offset == 0);
 929                 }
 930                 iclog = iclog->ic_next;
 931         } while (iclog != first_iclog);
 932 #endif
 933         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 934                 /*
 935                  * If we think the summary counters are bad, avoid writing the
 936                  * unmount record to force log recovery at next mount, after
 937                  * which the summary counters will be recalculated.  Refer to
 938                  * xlog_check_unmount_rec for more details.
 939                  */
 940                 if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS),
 941                                 mp, XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
 942                         xfs_alert(mp,
 943                                 "%s: will fix summary counters at next mount",
 944                                 __func__);
 945                         return 0;
 946                 }
 947                 xfs_log_write_unmount_record(mp);
 948         } else {
 949                 /*
 950                  * We're already in forced_shutdown mode, couldn't
 951                  * even attempt to write out the unmount transaction.
 952                  *
 953                  * Go through the motions of sync'ing and releasing
 954                  * the iclog, even though no I/O will actually happen,
 955                  * we need to wait for other log I/Os that may already
 956                  * be in progress.  Do this as a separate section of
 957                  * code so we'll know if we ever get stuck here that
 958                  * we're in this odd situation of trying to unmount
 959                  * a file system that went into forced_shutdown as
 960                  * the result of an unmount..
 961                  */
 962                 spin_lock(&log->l_icloglock);
 963                 iclog = log->l_iclog;
 964                 atomic_inc(&iclog->ic_refcnt);
 965
 966                 xlog_state_want_sync(log, iclog);
 967                 spin_unlock(&log->l_icloglock);
 968                 error =  xlog_state_release_iclog(log, iclog);
 969
 970                 spin_lock(&log->l_icloglock);
 971
 972                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 973                         || iclog->ic_state == XLOG_STATE_DIRTY
 974                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 975
 976                                 xlog_wait(&iclog->ic_force_wait,
 977                                                         &log->l_icloglock);
 978                 } else {
 979                         spin_unlock(&log->l_icloglock);
 980                 }
 981         }
 982
 983         return error;
 984 }       /* xfs_log_unmount_write */
 985
 986 /*
 987  * Empty the log for unmount/freeze.
 988  *
 989  * To do this, we first need to shut down the background log work so it is not
 990  * trying to cover the log as we clean up. We then need to unpin all objects in
 991  * the log so we can then flush them out. Once they have completed their IO and
 992  * run the callbacks removing themselves from the AIL, we can write the unmount
 993  * record.
 994  */
 995 void
 996 xfs_log_quiesce(
 997         struct xfs_mount        *mp)
 998 {
 999         cancel_delayed_work_sync(&mp->m_log->l_work);
1000         xfs_log_force(mp, XFS_LOG_SYNC);
1001
1002         /*
1003          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1004          * will push it, xfs_wait_buftarg() will not wait for it. Further,
1005          * xfs_buf_iowait() cannot be used because it was pushed with the
1006          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1007          * the IO to complete.
1008          */
1009         xfs_ail_push_all_sync(mp->m_ail);
1010         xfs_wait_buftarg(mp->m_ddev_targp);
1011         xfs_buf_lock(mp->m_sb_bp);
1012         xfs_buf_unlock(mp->m_sb_bp);
1013
1014         xfs_log_unmount_write(mp);
1015 }
1016
1017 /*
1018  * Shut down and release the AIL and Log.
1019  *
1020  * During unmount, we need to ensure we flush all the dirty metadata objects
1021  * from the AIL so that the log is empty before we write the unmount record to
1022  * the log. Once this is done, we can tear down the AIL and the log.
1023  */
1024 void
1025 xfs_log_unmount(
1026         struct xfs_mount        *mp)
1027 {
1028         xfs_log_quiesce(mp);
1029
1030         xfs_trans_ail_destroy(mp);
1031
1032         xfs_sysfs_del(&mp->m_log->l_kobj);
1033
1034         xlog_dealloc_log(mp->m_log);
1035 }
1036
1037 void
1038 xfs_log_item_init(
1039         struct xfs_mount        *mp,
1040         struct xfs_log_item     *item,
1041         int                     type,
1042         const struct xfs_item_ops *ops)
1043 {
1044         item->li_mountp = mp;
1045         item->li_ailp = mp->m_ail;
1046         item->li_type = type;
1047         item->li_ops = ops;
1048         item->li_lv = NULL;
1049
1050         INIT_LIST_HEAD(&item->li_ail);
1051         INIT_LIST_HEAD(&item->li_cil);
1052         INIT_LIST_HEAD(&item->li_bio_list);
1053         INIT_LIST_HEAD(&item->li_trans);
1054 }
1055
1056 /*
1057  * Wake up processes waiting for log space after we have moved the log tail.
1058  */
1059 void
1060 xfs_log_space_wake(
1061         struct xfs_mount        *mp)
1062 {
1063         struct xlog             *log = mp->m_log;
1064         int                     free_bytes;
1065
1066         if (XLOG_FORCED_SHUTDOWN(log))
1067                 return;
1068
1069         if (!list_empty_careful(&log->l_write_head.waiters)) {
1070                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1071
1072                 spin_lock(&log->l_write_head.lock);
1073                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1074                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1075                 spin_unlock(&log->l_write_head.lock);
1076         }
1077
1078         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1079                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1080
1081                 spin_lock(&log->l_reserve_head.lock);
1082                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1083                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1084                 spin_unlock(&log->l_reserve_head.lock);
1085         }
1086 }
1087
1088 /*
1089  * Determine if we have a transaction that has gone to disk that needs to be
1090  * covered. To begin the transition to the idle state firstly the log needs to
1091  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1092  * we start attempting to cover the log.
1093  *
1094  * Only if we are then in a state where covering is needed, the caller is
1095  * informed that dummy transactions are required to move the log into the idle
1096  * state.
1097  *
1098  * If there are any items in the AIl or CIL, then we do not want to attempt to
1099  * cover the log as we may be in a situation where there isn't log space
1100  * available to run a dummy transaction and this can lead to deadlocks when the
1101  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1102  * there's no point in running a dummy transaction at this point because we
1103  * can't start trying to idle the log until both the CIL and AIL are empty.
1104  */
1105 static int
1106 xfs_log_need_covered(xfs_mount_t *mp)
1107 {
1108         struct xlog     *log = mp->m_log;
1109         int             needed = 0;
1110
1111         if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1112                 return 0;
1113
1114         if (!xlog_cil_empty(log))
1115                 return 0;
1116
1117         spin_lock(&log->l_icloglock);
1118         switch (log->l_covered_state) {
1119         case XLOG_STATE_COVER_DONE:
1120         case XLOG_STATE_COVER_DONE2:
1121         case XLOG_STATE_COVER_IDLE:
1122                 break;
1123         case XLOG_STATE_COVER_NEED:
1124         case XLOG_STATE_COVER_NEED2:
1125                 if (xfs_ail_min_lsn(log->l_ailp))
1126                         break;
1127                 if (!xlog_iclogs_empty(log))
1128                         break;
1129
1130                 needed = 1;
1131                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1132                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1133                 else
1134                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1135                 break;
1136         default:
1137                 needed = 1;
1138                 break;
1139         }
1140         spin_unlock(&log->l_icloglock);
1141         return needed;
1142 }
1143
1144 /*
1145  * We may be holding the log iclog lock upon entering this routine.
1146  */
1147 xfs_lsn_t
1148 xlog_assign_tail_lsn_locked(
1149         struct xfs_mount        *mp)
1150 {
1151         struct xlog             *log = mp->m_log;
1152         struct xfs_log_item     *lip;
1153         xfs_lsn_t               tail_lsn;
1154
1155         assert_spin_locked(&mp->m_ail->ail_lock);
1156
1157         /*
1158          * To make sure we always have a valid LSN for the log tail we keep
1159          * track of the last LSN which was committed in log->l_last_sync_lsn,
1160          * and use that when the AIL was empty.
1161          */
1162         lip = xfs_ail_min(mp->m_ail);
1163         if (lip)
1164                 tail_lsn = lip->li_lsn;
1165         else
1166                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1167         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1168         atomic64_set(&log->l_tail_lsn, tail_lsn);
1169         return tail_lsn;
1170 }
1171
1172 xfs_lsn_t
1173 xlog_assign_tail_lsn(
1174         struct xfs_mount        *mp)
1175 {
1176         xfs_lsn_t               tail_lsn;
1177
1178         spin_lock(&mp->m_ail->ail_lock);
1179         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1180         spin_unlock(&mp->m_ail->ail_lock);
1181
1182         return tail_lsn;
1183 }
1184
1185 /*
1186  * Return the space in the log between the tail and the head.  The head
1187  * is passed in the cycle/bytes formal parms.  In the special case where
1188  * the reserve head has wrapped passed the tail, this calculation is no
1189  * longer valid.  In this case, just return 0 which means there is no space
1190  * in the log.  This works for all places where this function is called
1191  * with the reserve head.  Of course, if the write head were to ever
1192  * wrap the tail, we should blow up.  Rather than catch this case here,
1193  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1194  *
1195  * This code also handles the case where the reservation head is behind
1196  * the tail.  The details of this case are described below, but the end
1197  * result is that we return the size of the log as the amount of space left.
1198  */
1199 STATIC int
1200 xlog_space_left(
1201         struct xlog     *log,
1202         atomic64_t      *head)
1203 {
1204         int             free_bytes;
1205         int             tail_bytes;
1206         int             tail_cycle;
1207         int             head_cycle;
1208         int             head_bytes;
1209
1210         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1211         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1212         tail_bytes = BBTOB(tail_bytes);
1213         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1214                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1215         else if (tail_cycle + 1 < head_cycle)
1216                 return 0;
1217         else if (tail_cycle < head_cycle) {
1218                 ASSERT(tail_cycle == (head_cycle - 1));
1219                 free_bytes = tail_bytes - head_bytes;
1220         } else {
1221                 /*
1222                  * The reservation head is behind the tail.
1223                  * In this case we just want to return the size of the
1224                  * log as the amount of space left.
1225                  */
1226                 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1227                 xfs_alert(log->l_mp,
1228                           "  tail_cycle = %d, tail_bytes = %d",
1229                           tail_cycle, tail_bytes);
1230                 xfs_alert(log->l_mp,
1231                           "  GH   cycle = %d, GH   bytes = %d",
1232                           head_cycle, head_bytes);
1233                 ASSERT(0);
1234                 free_bytes = log->l_logsize;
1235         }
1236         return free_bytes;
1237 }
1238
1239
1240 static void
1241 xlog_ioend_work(
1242         struct work_struct      *work)
1243 {
1244         struct xlog_in_core     *iclog =
1245                 container_of(work, struct xlog_in_core, ic_end_io_work);
1246         struct xlog             *log = iclog->ic_log;
1247         bool                    aborted = false;
1248         int                     error;
1249
1250         error = blk_status_to_errno(iclog->ic_bio.bi_status);
1251 #ifdef DEBUG
1252         /* treat writes with injected CRC errors as failed */
1253         if (iclog->ic_fail_crc)
1254                 error = -EIO;
1255 #endif
1256
1257         /*
1258          * Race to shutdown the filesystem if we see an error.
1259          */
1260         if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1261                 xfs_alert(log->l_mp, "log I/O error %d", error);
1262                 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
1263                 /*
1264                  * This flag will be propagated to the trans-committed
1265                  * callback routines to let them know that the log-commit
1266                  * didn't succeed.
1267                  */
1268                 aborted = true;
1269         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1270                 aborted = true;
1271         }
1272
1273         xlog_state_done_syncing(iclog, aborted);
1274         bio_uninit(&iclog->ic_bio);
1275
1276         /*
1277          * Drop the lock to signal that we are done. Nothing references the
1278          * iclog after this, so an unmount waiting on this lock can now tear it
1279          * down safely. As such, it is unsafe to reference the iclog after the
1280          * unlock as we could race with it being freed.
1281          */
1282         up(&iclog->ic_sema);
1283 }
1284
1285 /*
1286  * Return size of each in-core log record buffer.
1287  *
1288  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1289  *
1290  * If the filesystem blocksize is too large, we may need to choose a
1291  * larger size since the directory code currently logs entire blocks.
1292  */
1293 STATIC void
1294 xlog_get_iclog_buffer_size(
1295         struct xfs_mount        *mp,
1296         struct xlog             *log)
1297 {
1298         if (mp->m_logbufs <= 0)
1299                 mp->m_logbufs = XLOG_MAX_ICLOGS;
1300         if (mp->m_logbsize <= 0)
1301                 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1302
1303         log->l_iclog_bufs = mp->m_logbufs;
1304         log->l_iclog_size = mp->m_logbsize;
1305
1306         /*
1307          * # headers = size / 32k - one header holds cycles from 32k of data.
1308          */
1309         log->l_iclog_heads =
1310                 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1311         log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1312 }
1313
1314 void
1315 xfs_log_work_queue(
1316         struct xfs_mount        *mp)
1317 {
1318         queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1319                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1320 }
1321
1322 /*
1323  * Every sync period we need to unpin all items in the AIL and push them to
1324  * disk. If there is nothing dirty, then we might need to cover the log to
1325  * indicate that the filesystem is idle.
1326  */
1327 static void
1328 xfs_log_worker(
1329         struct work_struct      *work)
1330 {
1331         struct xlog             *log = container_of(to_delayed_work(work),
1332                                                 struct xlog, l_work);
1333         struct xfs_mount        *mp = log->l_mp;
1334
1335         /* dgc: errors ignored - not fatal and nowhere to report them */
1336         if (xfs_log_need_covered(mp)) {
1337                 /*
1338                  * Dump a transaction into the log that contains no real change.
1339                  * This is needed to stamp the current tail LSN into the log
1340                  * during the covering operation.
1341                  *
1342                  * We cannot use an inode here for this - that will push dirty
1343                  * state back up into the VFS and then periodic inode flushing
1344                  * will prevent log covering from making progress. Hence we
1345                  * synchronously log the superblock instead to ensure the
1346                  * superblock is immediately unpinned and can be written back.
1347                  */
1348                 xfs_sync_sb(mp, true);
1349         } else
1350                 xfs_log_force(mp, 0);
1351
1352         /* start pushing all the metadata that is currently dirty */
1353         xfs_ail_push_all(mp->m_ail);
1354
1355         /* queue us up again */
1356         xfs_log_work_queue(mp);
1357 }
1358
1359 /*
1360  * This routine initializes some of the log structure for a given mount point.
1361  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1362  * some other stuff may be filled in too.
1363  */
1364 STATIC struct xlog *
1365 xlog_alloc_log(
1366         struct xfs_mount        *mp,
1367         struct xfs_buftarg      *log_target,
1368         xfs_daddr_t             blk_offset,
1369         int                     num_bblks)
1370 {
1371         struct xlog             *log;
1372         xlog_rec_header_t       *head;
1373         xlog_in_core_t          **iclogp;
1374         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1375         int                     i;
1376         int                     error = -ENOMEM;
1377         uint                    log2_size = 0;
1378
1379         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1380         if (!log) {
1381                 xfs_warn(mp, "Log allocation failed: No memory!");
1382                 goto out;
1383         }
1384
1385         log->l_mp          = mp;
1386         log->l_targ        = log_target;
1387         log->l_logsize     = BBTOB(num_bblks);
1388         log->l_logBBstart  = blk_offset;
1389         log->l_logBBsize   = num_bblks;
1390         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1391         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1392         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1393
1394         log->l_prev_block  = -1;
1395         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1396         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1397         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1398         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1399
1400         xlog_grant_head_init(&log->l_reserve_head);
1401         xlog_grant_head_init(&log->l_write_head);
1402
1403         error = -EFSCORRUPTED;
1404         if (xfs_sb_version_hassector(&mp->m_sb)) {
1405                 log2_size = mp->m_sb.sb_logsectlog;
1406                 if (log2_size < BBSHIFT) {
1407                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1408                                 log2_size, BBSHIFT);
1409                         goto out_free_log;
1410                 }
1411
1412                 log2_size -= BBSHIFT;
1413                 if (log2_size > mp->m_sectbb_log) {
1414                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1415                                 log2_size, mp->m_sectbb_log);
1416                         goto out_free_log;
1417                 }
1418
1419                 /* for larger sector sizes, must have v2 or external log */
1420                 if (log2_size && log->l_logBBstart > 0 &&
1421                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1422                         xfs_warn(mp,
1423                 "log sector size (0x%x) invalid for configuration.",
1424                                 log2_size);
1425                         goto out_free_log;
1426                 }
1427         }
1428         log->l_sectBBsize = 1 << log2_size;
1429
1430         xlog_get_iclog_buffer_size(mp, log);
1431
1432         spin_lock_init(&log->l_icloglock);
1433         init_waitqueue_head(&log->l_flush_wait);
1434
1435         iclogp = &log->l_iclog;
1436         /*
1437          * The amount of memory to allocate for the iclog structure is
1438          * rather funky due to the way the structure is defined.  It is
1439          * done this way so that we can use different sizes for machines
1440          * with different amounts of memory.  See the definition of
1441          * xlog_in_core_t in xfs_log_priv.h for details.
1442          */
1443         ASSERT(log->l_iclog_size >= 4096);
1444         for (i = 0; i < log->l_iclog_bufs; i++) {
1445                 int align_mask = xfs_buftarg_dma_alignment(mp->m_logdev_targp);
1446                 size_t bvec_size = howmany(log->l_iclog_size, PAGE_SIZE) *
1447                                 sizeof(struct bio_vec);
1448
1449                 iclog = kmem_zalloc(sizeof(*iclog) + bvec_size, KM_MAYFAIL);
1450                 if (!iclog)
1451                         goto out_free_iclog;
1452
1453                 *iclogp = iclog;
1454                 iclog->ic_prev = prev_iclog;
1455                 prev_iclog = iclog;
1456
1457                 iclog->ic_data = kmem_alloc_io(log->l_iclog_size, align_mask,
1458                                                 KM_MAYFAIL | KM_ZERO);
1459                 if (!iclog->ic_data)
1460                         goto out_free_iclog;
1461 #ifdef DEBUG
1462                 log->l_iclog_bak[i] = &iclog->ic_header;
1463 #endif
1464                 head = &iclog->ic_header;
1465                 memset(head, 0, sizeof(xlog_rec_header_t));
1466                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1467                 head->h_version = cpu_to_be32(
1468                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1469                 head->h_size = cpu_to_be32(log->l_iclog_size);
1470                 /* new fields */
1471                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1472                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1473
1474                 iclog->ic_size = log->l_iclog_size - log->l_iclog_hsize;
1475                 iclog->ic_state = XLOG_STATE_ACTIVE;
1476                 iclog->ic_log = log;
1477                 atomic_set(&iclog->ic_refcnt, 0);
1478                 spin_lock_init(&iclog->ic_callback_lock);
1479                 INIT_LIST_HEAD(&iclog->ic_callbacks);
1480                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1481
1482                 init_waitqueue_head(&iclog->ic_force_wait);
1483                 init_waitqueue_head(&iclog->ic_write_wait);
1484                 INIT_WORK(&iclog->ic_end_io_work, xlog_ioend_work);
1485                 sema_init(&iclog->ic_sema, 1);
1486
1487                 iclogp = &iclog->ic_next;
1488         }
1489         *iclogp = log->l_iclog;                 /* complete ring */
1490         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1491
1492         log->l_ioend_workqueue = alloc_workqueue("xfs-log/%s",
1493                         WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI, 0,
1494                         mp->m_fsname);
1495         if (!log->l_ioend_workqueue)
1496                 goto out_free_iclog;
1497
1498         error = xlog_cil_init(log);
1499         if (error)
1500                 goto out_destroy_workqueue;
1501         return log;
1502
1503 out_destroy_workqueue:
1504         destroy_workqueue(log->l_ioend_workqueue);
1505 out_free_iclog:
1506         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1507                 prev_iclog = iclog->ic_next;
1508                 kmem_free(iclog->ic_data);
1509                 kmem_free(iclog);
1510                 if (prev_iclog == log->l_iclog)
1511                         break;
1512         }
1513 out_free_log:
1514         kmem_free(log);
1515 out:
1516         return ERR_PTR(error);
1517 }       /* xlog_alloc_log */
1518
1519
1520 /*
1521  * Write out the commit record of a transaction associated with the given
1522  * ticket.  Return the lsn of the commit record.
1523  */
1524 STATIC int
1525 xlog_commit_record(
1526         struct xlog             *log,
1527         struct xlog_ticket      *ticket,
1528         struct xlog_in_core     **iclog,
1529         xfs_lsn_t               *commitlsnp)
1530 {
1531         struct xfs_mount *mp = log->l_mp;
1532         int     error;
1533         struct xfs_log_iovec reg = {
1534                 .i_addr = NULL,
1535                 .i_len = 0,
1536                 .i_type = XLOG_REG_TYPE_COMMIT,
1537         };
1538         struct xfs_log_vec vec = {
1539                 .lv_niovecs = 1,
1540                 .lv_iovecp = &reg,
1541         };
1542
1543         ASSERT_ALWAYS(iclog);
1544         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1545                                         XLOG_COMMIT_TRANS);
1546         if (error)
1547                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1548         return error;
1549 }
1550
1551 /*
1552  * Compute the LSN that we'd need to push the log tail towards in order to have
1553  * (a) enough on-disk log space to log the number of bytes specified, (b) at
1554  * least 25% of the log space free, and (c) at least 256 blocks free.  If the
1555  * log free space already meets all three thresholds, this function returns
1556  * NULLCOMMITLSN.
1557  */
1558 xfs_lsn_t
1559 xlog_grant_push_threshold(
1560         struct xlog     *log,
1561         int             need_bytes)
1562 {
1563         xfs_lsn_t       threshold_lsn = 0;
1564         xfs_lsn_t       last_sync_lsn;
1565         int             free_blocks;
1566         int             free_bytes;
1567         int             threshold_block;
1568         int             threshold_cycle;
1569         int             free_threshold;
1570
1571         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1572
1573         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1574         free_blocks = BTOBBT(free_bytes);
1575
1576         /*
1577          * Set the threshold for the minimum number of free blocks in the
1578          * log to the maximum of what the caller needs, one quarter of the
1579          * log, and 256 blocks.
1580          */
1581         free_threshold = BTOBB(need_bytes);
1582         free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1583         free_threshold = max(free_threshold, 256);
1584         if (free_blocks >= free_threshold)
1585                 return NULLCOMMITLSN;
1586
1587         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1588                                                 &threshold_block);
1589         threshold_block += free_threshold;
1590         if (threshold_block >= log->l_logBBsize) {
1591                 threshold_block -= log->l_logBBsize;
1592                 threshold_cycle += 1;
1593         }
1594         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1595                                         threshold_block);
1596         /*
1597          * Don't pass in an lsn greater than the lsn of the last
1598          * log record known to be on disk. Use a snapshot of the last sync lsn
1599          * so that it doesn't change between the compare and the set.
1600          */
1601         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1602         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1603                 threshold_lsn = last_sync_lsn;
1604
1605         return threshold_lsn;
1606 }
1607
1608 /*
1609  * Push the tail of the log if we need to do so to maintain the free log space
1610  * thresholds set out by xlog_grant_push_threshold.  We may need to adopt a
1611  * policy which pushes on an lsn which is further along in the log once we
1612  * reach the high water mark.  In this manner, we would be creating a low water
1613  * mark.
1614  */
1615 STATIC void
1616 xlog_grant_push_ail(
1617         struct xlog     *log,
1618         int             need_bytes)
1619 {
1620         xfs_lsn_t       threshold_lsn;
1621
1622         threshold_lsn = xlog_grant_push_threshold(log, need_bytes);
1623         if (threshold_lsn == NULLCOMMITLSN || XLOG_FORCED_SHUTDOWN(log))
1624                 return;
1625
1626         /*
1627          * Get the transaction layer to kick the dirty buffers out to
1628          * disk asynchronously. No point in trying to do this if
1629          * the filesystem is shutting down.
1630          */
1631         xfs_ail_push(log->l_ailp, threshold_lsn);
1632 }
1633
1634 /*
1635  * Stamp cycle number in every block
1636  */
1637 STATIC void
1638 xlog_pack_data(
1639         struct xlog             *log,
1640         struct xlog_in_core     *iclog,
1641         int                     roundoff)
1642 {
1643         int                     i, j, k;
1644         int                     size = iclog->ic_offset + roundoff;
1645         __be32                  cycle_lsn;
1646         char                    *dp;
1647
1648         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1649
1650         dp = iclog->ic_datap;
1651         for (i = 0; i < BTOBB(size); i++) {
1652                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1653                         break;
1654                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1655                 *(__be32 *)dp = cycle_lsn;
1656                 dp += BBSIZE;
1657         }
1658
1659         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1660                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1661
1662                 for ( ; i < BTOBB(size); i++) {
1663                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1664                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1665                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1666                         *(__be32 *)dp = cycle_lsn;
1667                         dp += BBSIZE;
1668                 }
1669
1670                 for (i = 1; i < log->l_iclog_heads; i++)
1671                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1672         }
1673 }
1674
1675 /*
1676  * Calculate the checksum for a log buffer.
1677  *
1678  * This is a little more complicated than it should be because the various
1679  * headers and the actual data are non-contiguous.
1680  */
1681 __le32
1682 xlog_cksum(
1683         struct xlog             *log,
1684         struct xlog_rec_header  *rhead,
1685         char                    *dp,
1686         int                     size)
1687 {
1688         uint32_t                crc;
1689
1690         /* first generate the crc for the record header ... */
1691         crc = xfs_start_cksum_update((char *)rhead,
1692                               sizeof(struct xlog_rec_header),
1693                               offsetof(struct xlog_rec_header, h_crc));
1694
1695         /* ... then for additional cycle data for v2 logs ... */
1696         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1697                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1698                 int             i;
1699                 int             xheads;
1700
1701                 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1702                 if (size % XLOG_HEADER_CYCLE_SIZE)
1703                         xheads++;
1704
1705                 for (i = 1; i < xheads; i++) {
1706                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1707                                      sizeof(struct xlog_rec_ext_header));
1708                 }
1709         }
1710
1711         /* ... and finally for the payload */
1712         crc = crc32c(crc, dp, size);
1713
1714         return xfs_end_cksum(crc);
1715 }
1716
1717 static void
1718 xlog_bio_end_io(
1719         struct bio              *bio)
1720 {
1721         struct xlog_in_core     *iclog = bio->bi_private;
1722
1723         queue_work(iclog->ic_log->l_ioend_workqueue,
1724                    &iclog->ic_end_io_work);
1725 }
1726
1727 static void
1728 xlog_map_iclog_data(
1729         struct bio              *bio,
1730         void                    *data,
1731         size_t                  count)
1732 {
1733         do {
1734                 struct page     *page = kmem_to_page(data);
1735                 unsigned int    off = offset_in_page(data);
1736                 size_t          len = min_t(size_t, count, PAGE_SIZE - off);
1737
1738                 WARN_ON_ONCE(bio_add_page(bio, page, len, off) != len);
1739
1740                 data += len;
1741                 count -= len;
1742         } while (count);
1743 }
1744
1745 STATIC void
1746 xlog_write_iclog(
1747         struct xlog             *log,
1748         struct xlog_in_core     *iclog,
1749         uint64_t                bno,
1750         unsigned int            count,
1751         bool                    need_flush)
1752 {
1753         ASSERT(bno < log->l_logBBsize);
1754
1755         /*
1756          * We lock the iclogbufs here so that we can serialise against I/O
1757          * completion during unmount.  We might be processing a shutdown
1758          * triggered during unmount, and that can occur asynchronously to the
1759          * unmount thread, and hence we need to ensure that completes before
1760          * tearing down the iclogbufs.  Hence we need to hold the buffer lock
1761          * across the log IO to archieve that.
1762          */
1763         down(&iclog->ic_sema);
1764         if (unlikely(iclog->ic_state & XLOG_STATE_IOERROR)) {
1765                 /*
1766                  * It would seem logical to return EIO here, but we rely on
1767                  * the log state machine to propagate I/O errors instead of
1768                  * doing it here.  We kick of the state machine and unlock
1769                  * the buffer manually, the code needs to be kept in sync
1770                  * with the I/O completion path.
1771                  */
1772                 xlog_state_done_syncing(iclog, XFS_LI_ABORTED);
1773                 up(&iclog->ic_sema);
1774                 return;
1775         }
1776
1777         iclog->ic_io_size = count;
1778
1779         bio_init(&iclog->ic_bio, iclog->ic_bvec, howmany(count, PAGE_SIZE));
1780         bio_set_dev(&iclog->ic_bio, log->l_targ->bt_bdev);
1781         iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart + bno;
1782         iclog->ic_bio.bi_end_io = xlog_bio_end_io;
1783         iclog->ic_bio.bi_private = iclog;
1784         iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_FUA;
1785         if (need_flush)
1786                 iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
1787
1788         xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, iclog->ic_io_size);
1789         if (is_vmalloc_addr(iclog->ic_data))
1790                 flush_kernel_vmap_range(iclog->ic_data, iclog->ic_io_size);
1791
1792         /*
1793          * If this log buffer would straddle the end of the log we will have
1794          * to split it up into two bios, so that we can continue at the start.
1795          */
1796         if (bno + BTOBB(count) > log->l_logBBsize) {
1797                 struct bio *split;
1798
1799                 split = bio_split(&iclog->ic_bio, log->l_logBBsize - bno,
1800                                   GFP_NOIO, &fs_bio_set);
1801                 bio_chain(split, &iclog->ic_bio);
1802                 submit_bio(split);
1803
1804                 /* restart at logical offset zero for the remainder */
1805                 iclog->ic_bio.bi_iter.bi_sector = log->l_logBBstart;
1806         }
1807
1808         submit_bio(&iclog->ic_bio);
1809 }
1810
1811 /*
1812  * We need to bump cycle number for the part of the iclog that is
1813  * written to the start of the log. Watch out for the header magic
1814  * number case, though.
1815  */
1816 static void
1817 xlog_split_iclog(
1818         struct xlog             *log,
1819         void                    *data,
1820         uint64_t                bno,
1821         unsigned int            count)
1822 {
1823         unsigned int            split_offset = BBTOB(log->l_logBBsize - bno);
1824         unsigned int            i;
1825
1826         for (i = split_offset; i < count; i += BBSIZE) {
1827                 uint32_t cycle = get_unaligned_be32(data + i);
1828
1829                 if (++cycle == XLOG_HEADER_MAGIC_NUM)
1830                         cycle++;
1831                 put_unaligned_be32(cycle, data + i);
1832         }
1833 }
1834
1835 static int
1836 xlog_calc_iclog_size(
1837         struct xlog             *log,
1838         struct xlog_in_core     *iclog,
1839         uint32_t                *roundoff)
1840 {
1841         uint32_t                count_init, count;
1842         bool                    use_lsunit;
1843
1844         use_lsunit = xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
1845                         log->l_mp->m_sb.sb_logsunit > 1;
1846
1847         /* Add for LR header */
1848         count_init = log->l_iclog_hsize + iclog->ic_offset;
1849
1850         /* Round out the log write size */
1851         if (use_lsunit) {
1852                 /* we have a v2 stripe unit to use */
1853                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1854         } else {
1855                 count = BBTOB(BTOBB(count_init));
1856         }
1857
1858         ASSERT(count >= count_init);
1859         *roundoff = count - count_init;
1860
1861         if (use_lsunit)
1862                 ASSERT(*roundoff < log->l_mp->m_sb.sb_logsunit);
1863         else
1864                 ASSERT(*roundoff < BBTOB(1));
1865         return count;
1866 }
1867
1868 /*
1869  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1870  * fashion.  Previously, we should have moved the current iclog
1871  * ptr in the log to point to the next available iclog.  This allows further
1872  * write to continue while this code syncs out an iclog ready to go.
1873  * Before an in-core log can be written out, the data section must be scanned
1874  * to save away the 1st word of each BBSIZE block into the header.  We replace
1875  * it with the current cycle count.  Each BBSIZE block is tagged with the
1876  * cycle count because there in an implicit assumption that drives will
1877  * guarantee that entire 512 byte blocks get written at once.  In other words,
1878  * we can't have part of a 512 byte block written and part not written.  By
1879  * tagging each block, we will know which blocks are valid when recovering
1880  * after an unclean shutdown.
1881  *
1882  * This routine is single threaded on the iclog.  No other thread can be in
1883  * this routine with the same iclog.  Changing contents of iclog can there-
1884  * fore be done without grabbing the state machine lock.  Updating the global
1885  * log will require grabbing the lock though.
1886  *
1887  * The entire log manager uses a logical block numbering scheme.  Only
1888  * xlog_write_iclog knows about the fact that the log may not start with
1889  * block zero on a given device.
1890  */
1891 STATIC void
1892 xlog_sync(
1893         struct xlog             *log,
1894         struct xlog_in_core     *iclog)
1895 {
1896         unsigned int            count;          /* byte count of bwrite */
1897         unsigned int            roundoff;       /* roundoff to BB or stripe */
1898         uint64_t                bno;
1899         unsigned int            size;
1900         bool                    need_flush = true, split = false;
1901
1902         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1903
1904         count = xlog_calc_iclog_size(log, iclog, &roundoff);
1905
1906         /* move grant heads by roundoff in sync */
1907         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1908         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1909
1910         /* put cycle number in every block */
1911         xlog_pack_data(log, iclog, roundoff);
1912
1913         /* real byte length */
1914         size = iclog->ic_offset;
1915         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb))
1916                 size += roundoff;
1917         iclog->ic_header.h_len = cpu_to_be32(size);
1918
1919         XFS_STATS_INC(log->l_mp, xs_log_writes);
1920         XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1921
1922         bno = BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn));
1923
1924         /* Do we need to split this write into 2 parts? */
1925         if (bno + BTOBB(count) > log->l_logBBsize) {
1926                 xlog_split_iclog(log, &iclog->ic_header, bno, count);
1927                 split = true;
1928         }
1929
1930         /* calculcate the checksum */
1931         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1932                                             iclog->ic_datap, size);
1933         /*
1934          * Intentionally corrupt the log record CRC based on the error injection
1935          * frequency, if defined. This facilitates testing log recovery in the
1936          * event of torn writes. Hence, set the IOABORT state to abort the log
1937          * write on I/O completion and shutdown the fs. The subsequent mount
1938          * detects the bad CRC and attempts to recover.
1939          */
1940 #ifdef DEBUG
1941         if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1942                 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1943                 iclog->ic_fail_crc = true;
1944                 xfs_warn(log->l_mp,
1945         "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1946                          be64_to_cpu(iclog->ic_header.h_lsn));
1947         }
1948 #endif
1949
1950         /*
1951          * Flush the data device before flushing the log to make sure all meta
1952          * data written back from the AIL actually made it to disk before
1953          * stamping the new log tail LSN into the log buffer.  For an external
1954          * log we need to issue the flush explicitly, and unfortunately
1955          * synchronously here; for an internal log we can simply use the block
1956          * layer state machine for preflushes.
1957          */
1958         if (log->l_targ != log->l_mp->m_ddev_targp || split) {
1959                 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1960                 need_flush = false;
1961         }
1962
1963         xlog_verify_iclog(log, iclog, count);
1964         xlog_write_iclog(log, iclog, bno, count, need_flush);
1965 }
1966
1967 /*
1968  * Deallocate a log structure
1969  */
1970 STATIC void
1971 xlog_dealloc_log(
1972         struct xlog     *log)
1973 {
1974         xlog_in_core_t  *iclog, *next_iclog;
1975         int             i;
1976
1977         xlog_cil_destroy(log);
1978
1979         /*
1980          * Cycle all the iclogbuf locks to make sure all log IO completion
1981          * is done before we tear down these buffers.
1982          */
1983         iclog = log->l_iclog;
1984         for (i = 0; i < log->l_iclog_bufs; i++) {
1985                 down(&iclog->ic_sema);
1986                 up(&iclog->ic_sema);
1987                 iclog = iclog->ic_next;
1988         }
1989
1990         iclog = log->l_iclog;
1991         for (i = 0; i < log->l_iclog_bufs; i++) {
1992                 next_iclog = iclog->ic_next;
1993                 kmem_free(iclog->ic_data);
1994                 kmem_free(iclog);
1995                 iclog = next_iclog;
1996         }
1997
1998         log->l_mp->m_log = NULL;
1999         destroy_workqueue(log->l_ioend_workqueue);
2000         kmem_free(log);
2001 }       /* xlog_dealloc_log */
2002
2003 /*
2004  * Update counters atomically now that memcpy is done.
2005  */
2006 /* ARGSUSED */
2007 static inline void
2008 xlog_state_finish_copy(
2009         struct xlog             *log,
2010         struct xlog_in_core     *iclog,
2011         int                     record_cnt,
2012         int                     copy_bytes)
2013 {
2014         spin_lock(&log->l_icloglock);
2015
2016         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2017         iclog->ic_offset += copy_bytes;
2018
2019         spin_unlock(&log->l_icloglock);
2020 }       /* xlog_state_finish_copy */
2021
2022
2023
2024
2025 /*
2026  * print out info relating to regions written which consume
2027  * the reservation
2028  */
2029 void
2030 xlog_print_tic_res(
2031         struct xfs_mount        *mp,
2032         struct xlog_ticket      *ticket)
2033 {
2034         uint i;
2035         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2036
2037         /* match with XLOG_REG_TYPE_* in xfs_log.h */
2038 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2039         static char *res_type_str[] = {
2040             REG_TYPE_STR(BFORMAT, "bformat"),
2041             REG_TYPE_STR(BCHUNK, "bchunk"),
2042             REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2043             REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2044             REG_TYPE_STR(IFORMAT, "iformat"),
2045             REG_TYPE_STR(ICORE, "icore"),
2046             REG_TYPE_STR(IEXT, "iext"),
2047             REG_TYPE_STR(IBROOT, "ibroot"),
2048             REG_TYPE_STR(ILOCAL, "ilocal"),
2049             REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2050             REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2051             REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2052             REG_TYPE_STR(QFORMAT, "qformat"),
2053             REG_TYPE_STR(DQUOT, "dquot"),
2054             REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2055             REG_TYPE_STR(LRHEADER, "LR header"),
2056             REG_TYPE_STR(UNMOUNT, "unmount"),
2057             REG_TYPE_STR(COMMIT, "commit"),
2058             REG_TYPE_STR(TRANSHDR, "trans header"),
2059             REG_TYPE_STR(ICREATE, "inode create"),
2060             REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2061             REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2062             REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2063             REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2064             REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2065             REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2066         };
2067         BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2068 #undef REG_TYPE_STR
2069
2070         xfs_warn(mp, "ticket reservation summary:");
2071         xfs_warn(mp, "  unit res    = %d bytes",
2072                  ticket->t_unit_res);
2073         xfs_warn(mp, "  current res = %d bytes",
2074                  ticket->t_curr_res);
2075         xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2076                  ticket->t_res_arr_sum, ticket->t_res_o_flow);
2077         xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2078                  ticket->t_res_num_ophdrs, ophdr_spc);
2079         xfs_warn(mp, "  ophdr + reg = %u bytes",
2080                  ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2081         xfs_warn(mp, "  num regions = %u",
2082                  ticket->t_res_num);
2083
2084         for (i = 0; i < ticket->t_res_num; i++) {
2085                 uint r_type = ticket->t_res_arr[i].r_type;
2086                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2087                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2088                             "bad-rtype" : res_type_str[r_type]),
2089                             ticket->t_res_arr[i].r_len);
2090         }
2091 }
2092
2093 /*
2094  * Print a summary of the transaction.
2095  */
2096 void
2097 xlog_print_trans(
2098         struct xfs_trans        *tp)
2099 {
2100         struct xfs_mount        *mp = tp->t_mountp;
2101         struct xfs_log_item     *lip;
2102
2103         /* dump core transaction and ticket info */
2104         xfs_warn(mp, "transaction summary:");
2105         xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2106         xfs_warn(mp, "  log count = %d", tp->t_log_count);
2107         xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2108
2109         xlog_print_tic_res(mp, tp->t_ticket);
2110
2111         /* dump each log item */
2112         list_for_each_entry(lip, &tp->t_items, li_trans) {
2113                 struct xfs_log_vec      *lv = lip->li_lv;
2114                 struct xfs_log_iovec    *vec;
2115                 int                     i;
2116
2117                 xfs_warn(mp, "log item: ");
2118                 xfs_warn(mp, "  type    = 0x%x", lip->li_type);
2119                 xfs_warn(mp, "  flags   = 0x%lx", lip->li_flags);
2120                 if (!lv)
2121                         continue;
2122                 xfs_warn(mp, "  niovecs = %d", lv->lv_niovecs);
2123                 xfs_warn(mp, "  size    = %d", lv->lv_size);
2124                 xfs_warn(mp, "  bytes   = %d", lv->lv_bytes);
2125                 xfs_warn(mp, "  buf len = %d", lv->lv_buf_len);
2126
2127                 /* dump each iovec for the log item */
2128                 vec = lv->lv_iovecp;
2129                 for (i = 0; i < lv->lv_niovecs; i++) {
2130                         int dumplen = min(vec->i_len, 32);
2131
2132                         xfs_warn(mp, "  iovec[%d]", i);
2133                         xfs_warn(mp, "    type  = 0x%x", vec->i_type);
2134                         xfs_warn(mp, "    len   = %d", vec->i_len);
2135                         xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2136                         xfs_hex_dump(vec->i_addr, dumplen);
2137
2138                         vec++;
2139                 }
2140         }
2141 }
2142
2143 /*
2144  * Calculate the potential space needed by the log vector.  Each region gets
2145  * its own xlog_op_header_t and may need to be double word aligned.
2146  */
2147 static int
2148 xlog_write_calc_vec_length(
2149         struct xlog_ticket      *ticket,
2150         struct xfs_log_vec      *log_vector)
2151 {
2152         struct xfs_log_vec      *lv;
2153         int                     headers = 0;
2154         int                     len = 0;
2155         int                     i;
2156
2157         /* acct for start rec of xact */
2158         if (ticket->t_flags & XLOG_TIC_INITED)
2159                 headers++;
2160
2161         for (lv = log_vector; lv; lv = lv->lv_next) {
2162                 /* we don't write ordered log vectors */
2163                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2164                         continue;
2165
2166                 headers += lv->lv_niovecs;
2167
2168                 for (i = 0; i < lv->lv_niovecs; i++) {
2169                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2170
2171                         len += vecp->i_len;
2172                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2173                 }
2174         }
2175
2176         ticket->t_res_num_ophdrs += headers;
2177         len += headers * sizeof(struct xlog_op_header);
2178
2179         return len;
2180 }
2181
2182 /*
2183  * If first write for transaction, insert start record  We can't be trying to
2184  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2185  */
2186 static int
2187 xlog_write_start_rec(
2188         struct xlog_op_header   *ophdr,
2189         struct xlog_ticket      *ticket)
2190 {
2191         if (!(ticket->t_flags & XLOG_TIC_INITED))
2192                 return 0;
2193
2194         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2195         ophdr->oh_clientid = ticket->t_clientid;
2196         ophdr->oh_len = 0;
2197         ophdr->oh_flags = XLOG_START_TRANS;
2198         ophdr->oh_res2 = 0;
2199
2200         ticket->t_flags &= ~XLOG_TIC_INITED;
2201
2202         return sizeof(struct xlog_op_header);
2203 }
2204
2205 static xlog_op_header_t *
2206 xlog_write_setup_ophdr(
2207         struct xlog             *log,
2208         struct xlog_op_header   *ophdr,
2209         struct xlog_ticket      *ticket,
2210         uint                    flags)
2211 {
2212         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2213         ophdr->oh_clientid = ticket->t_clientid;
2214         ophdr->oh_res2 = 0;
2215
2216         /* are we copying a commit or unmount record? */
2217         ophdr->oh_flags = flags;
2218
2219         /*
2220          * We've seen logs corrupted with bad transaction client ids.  This
2221          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2222          * and shut down the filesystem.
2223          */
2224         switch (ophdr->oh_clientid)  {
2225         case XFS_TRANSACTION:
2226         case XFS_VOLUME:
2227         case XFS_LOG:
2228                 break;
2229         default:
2230                 xfs_warn(log->l_mp,
2231                         "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2232                         ophdr->oh_clientid, ticket);
2233                 return NULL;
2234         }
2235
2236         return ophdr;
2237 }
2238
2239 /*
2240  * Set up the parameters of the region copy into the log. This has
2241  * to handle region write split across multiple log buffers - this
2242  * state is kept external to this function so that this code can
2243  * be written in an obvious, self documenting manner.
2244  */
2245 static int
2246 xlog_write_setup_copy(
2247         struct xlog_ticket      *ticket,
2248         struct xlog_op_header   *ophdr,
2249         int                     space_available,
2250         int                     space_required,
2251         int                     *copy_off,
2252         int                     *copy_len,
2253         int                     *last_was_partial_copy,
2254         int                     *bytes_consumed)
2255 {
2256         int                     still_to_copy;
2257
2258         still_to_copy = space_required - *bytes_consumed;
2259         *copy_off = *bytes_consumed;
2260
2261         if (still_to_copy <= space_available) {
2262                 /* write of region completes here */
2263                 *copy_len = still_to_copy;
2264                 ophdr->oh_len = cpu_to_be32(*copy_len);
2265                 if (*last_was_partial_copy)
2266                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2267                 *last_was_partial_copy = 0;
2268                 *bytes_consumed = 0;
2269                 return 0;
2270         }
2271
2272         /* partial write of region, needs extra log op header reservation */
2273         *copy_len = space_available;
2274         ophdr->oh_len = cpu_to_be32(*copy_len);
2275         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2276         if (*last_was_partial_copy)
2277                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2278         *bytes_consumed += *copy_len;
2279         (*last_was_partial_copy)++;
2280
2281         /* account for new log op header */
2282         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2283         ticket->t_res_num_ophdrs++;
2284
2285         return sizeof(struct xlog_op_header);
2286 }
2287
2288 static int
2289 xlog_write_copy_finish(
2290         struct xlog             *log,
2291         struct xlog_in_core     *iclog,
2292         uint                    flags,
2293         int                     *record_cnt,
2294         int                     *data_cnt,
2295         int                     *partial_copy,
2296         int                     *partial_copy_len,
2297         int                     log_offset,
2298         struct xlog_in_core     **commit_iclog)
2299 {
2300         if (*partial_copy) {
2301                 /*
2302                  * This iclog has already been marked WANT_SYNC by
2303                  * xlog_state_get_iclog_space.
2304                  */
2305                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2306                 *record_cnt = 0;
2307                 *data_cnt = 0;
2308                 return xlog_state_release_iclog(log, iclog);
2309         }
2310
2311         *partial_copy = 0;
2312         *partial_copy_len = 0;
2313
2314         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2315                 /* no more space in this iclog - push it. */
2316                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2317                 *record_cnt = 0;
2318                 *data_cnt = 0;
2319
2320                 spin_lock(&log->l_icloglock);
2321                 xlog_state_want_sync(log, iclog);
2322                 spin_unlock(&log->l_icloglock);
2323
2324                 if (!commit_iclog)
2325                         return xlog_state_release_iclog(log, iclog);
2326                 ASSERT(flags & XLOG_COMMIT_TRANS);
2327                 *commit_iclog = iclog;
2328         }
2329
2330         return 0;
2331 }
2332
2333 /*
2334  * Write some region out to in-core log
2335  *
2336  * This will be called when writing externally provided regions or when
2337  * writing out a commit record for a given transaction.
2338  *
2339  * General algorithm:
2340  *      1. Find total length of this write.  This may include adding to the
2341  *              lengths passed in.
2342  *      2. Check whether we violate the tickets reservation.
2343  *      3. While writing to this iclog
2344  *          A. Reserve as much space in this iclog as can get
2345  *          B. If this is first write, save away start lsn
2346  *          C. While writing this region:
2347  *              1. If first write of transaction, write start record
2348  *              2. Write log operation header (header per region)
2349  *              3. Find out if we can fit entire region into this iclog
2350  *              4. Potentially, verify destination memcpy ptr
2351  *              5. Memcpy (partial) region
2352  *              6. If partial copy, release iclog; otherwise, continue
2353  *                      copying more regions into current iclog
2354  *      4. Mark want sync bit (in simulation mode)
2355  *      5. Release iclog for potential flush to on-disk log.
2356  *
2357  * ERRORS:
2358  * 1.   Panic if reservation is overrun.  This should never happen since
2359  *      reservation amounts are generated internal to the filesystem.
2360  * NOTES:
2361  * 1. Tickets are single threaded data structures.
2362  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2363  *      syncing routine.  When a single log_write region needs to span
2364  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2365  *      on all log operation writes which don't contain the end of the
2366  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2367  *      operation which contains the end of the continued log_write region.
2368  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2369  *      we don't really know exactly how much space will be used.  As a result,
2370  *      we don't update ic_offset until the end when we know exactly how many
2371  *      bytes have been written out.
2372  */
2373 int
2374 xlog_write(
2375         struct xlog             *log,
2376         struct xfs_log_vec      *log_vector,
2377         struct xlog_ticket      *ticket,
2378         xfs_lsn_t               *start_lsn,
2379         struct xlog_in_core     **commit_iclog,
2380         uint                    flags)
2381 {
2382         struct xlog_in_core     *iclog = NULL;
2383         struct xfs_log_iovec    *vecp;
2384         struct xfs_log_vec      *lv;
2385         int                     len;
2386         int                     index;
2387         int                     partial_copy = 0;
2388         int                     partial_copy_len = 0;
2389         int                     contwr = 0;
2390         int                     record_cnt = 0;
2391         int                     data_cnt = 0;
2392         int                     error;
2393
2394         *start_lsn = 0;
2395
2396         len = xlog_write_calc_vec_length(ticket, log_vector);
2397
2398         /*
2399          * Region headers and bytes are already accounted for.
2400          * We only need to take into account start records and
2401          * split regions in this function.
2402          */
2403         if (ticket->t_flags & XLOG_TIC_INITED)
2404                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2405
2406         /*
2407          * Commit record headers need to be accounted for. These
2408          * come in as separate writes so are easy to detect.
2409          */
2410         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2411                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2412
2413         if (ticket->t_curr_res < 0) {
2414                 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2415                      "ctx ticket reservation ran out. Need to up reservation");
2416                 xlog_print_tic_res(log->l_mp, ticket);
2417                 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2418         }
2419
2420         index = 0;
2421         lv = log_vector;
2422         vecp = lv->lv_iovecp;
2423         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2424                 void            *ptr;
2425                 int             log_offset;
2426
2427                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2428                                                    &contwr, &log_offset);
2429                 if (error)
2430                         return error;
2431
2432                 ASSERT(log_offset <= iclog->ic_size - 1);
2433                 ptr = iclog->ic_datap + log_offset;
2434
2435                 /* start_lsn is the first lsn written to. That's all we need. */
2436                 if (!*start_lsn)
2437                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2438
2439                 /*
2440                  * This loop writes out as many regions as can fit in the amount
2441                  * of space which was allocated by xlog_state_get_iclog_space().
2442                  */
2443                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2444                         struct xfs_log_iovec    *reg;
2445                         struct xlog_op_header   *ophdr;
2446                         int                     start_rec_copy;
2447                         int                     copy_len;
2448                         int                     copy_off;
2449                         bool                    ordered = false;
2450
2451                         /* ordered log vectors have no regions to write */
2452                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2453                                 ASSERT(lv->lv_niovecs == 0);
2454                                 ordered = true;
2455                                 goto next_lv;
2456                         }
2457
2458                         reg = &vecp[index];
2459                         ASSERT(reg->i_len % sizeof(int32_t) == 0);
2460                         ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2461
2462                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2463                         if (start_rec_copy) {
2464                                 record_cnt++;
2465                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2466                                                    start_rec_copy);
2467                         }
2468
2469                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2470                         if (!ophdr)
2471                                 return -EIO;
2472
2473                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2474                                            sizeof(struct xlog_op_header));
2475
2476                         len += xlog_write_setup_copy(ticket, ophdr,
2477                                                      iclog->ic_size-log_offset,
2478                                                      reg->i_len,
2479                                                      &copy_off, &copy_len,
2480                                                      &partial_copy,
2481                                                      &partial_copy_len);
2482                         xlog_verify_dest_ptr(log, ptr);
2483
2484                         /*
2485                          * Copy region.
2486                          *
2487                          * Unmount records just log an opheader, so can have
2488                          * empty payloads with no data region to copy. Hence we
2489                          * only copy the payload if the vector says it has data
2490                          * to copy.
2491                          */
2492                         ASSERT(copy_len >= 0);
2493                         if (copy_len > 0) {
2494                                 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2495                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2496                                                    copy_len);
2497                         }
2498                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2499                         record_cnt++;
2500                         data_cnt += contwr ? copy_len : 0;
2501
2502                         error = xlog_write_copy_finish(log, iclog, flags,
2503                                                        &record_cnt, &data_cnt,
2504                                                        &partial_copy,
2505                                                        &partial_copy_len,
2506                                                        log_offset,
2507                                                        commit_iclog);
2508                         if (error)
2509                                 return error;
2510
2511                         /*
2512                          * if we had a partial copy, we need to get more iclog
2513                          * space but we don't want to increment the region
2514                          * index because there is still more is this region to
2515                          * write.
2516                          *
2517                          * If we completed writing this region, and we flushed
2518                          * the iclog (indicated by resetting of the record
2519                          * count), then we also need to get more log space. If
2520                          * this was the last record, though, we are done and
2521                          * can just return.
2522                          */
2523                         if (partial_copy)
2524                                 break;
2525
2526                         if (++index == lv->lv_niovecs) {
2527 next_lv:
2528                                 lv = lv->lv_next;
2529                                 index = 0;
2530                                 if (lv)
2531                                         vecp = lv->lv_iovecp;
2532                         }
2533                         if (record_cnt == 0 && !ordered) {
2534                                 if (!lv)
2535                                         return 0;
2536                                 break;
2537                         }
2538                 }
2539         }
2540
2541         ASSERT(len == 0);
2542
2543         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2544         if (!commit_iclog)
2545                 return xlog_state_release_iclog(log, iclog);
2546
2547         ASSERT(flags & XLOG_COMMIT_TRANS);
2548         *commit_iclog = iclog;
2549         return 0;
2550 }
2551
2552
2553 /*****************************************************************************
2554  *
2555  *              State Machine functions
2556  *
2557  *****************************************************************************
2558  */
2559
2560 /*
2561  * An iclog has just finished IO completion processing, so we need to update
2562  * the iclog state and propagate that up into the overall log state. Hence we
2563  * prepare the iclog for cleaning, and then clean all the pending dirty iclogs
2564  * starting from the head, and then wake up any threads that are waiting for the
2565  * iclog to be marked clean.
2566  *
2567  * The ordering of marking iclogs ACTIVE must be maintained, so an iclog
2568  * doesn't become ACTIVE beyond one that is SYNCING.  This is also required to
2569  * maintain the notion that we use a ordered wait queue to hold off would be
2570  * writers to the log when every iclog is trying to sync to disk.
2571  *
2572  * Caller must hold the icloglock before calling us.
2573  *
2574  * State Change: !IOERROR -> DIRTY -> ACTIVE
2575  */
2576 STATIC void
2577 xlog_state_clean_iclog(
2578         struct xlog             *log,
2579         struct xlog_in_core     *dirty_iclog)
2580 {
2581         struct xlog_in_core     *iclog;
2582         int                     changed = 0;
2583
2584         /* Prepare the completed iclog. */
2585         if (!(dirty_iclog->ic_state & XLOG_STATE_IOERROR))
2586                 dirty_iclog->ic_state = XLOG_STATE_DIRTY;
2587
2588         /* Walk all the iclogs to update the ordered active state. */
2589         iclog = log->l_iclog;
2590         do {
2591                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2592                         iclog->ic_state = XLOG_STATE_ACTIVE;
2593                         iclog->ic_offset       = 0;
2594                         ASSERT(list_empty_careful(&iclog->ic_callbacks));
2595                         /*
2596                          * If the number of ops in this iclog indicate it just
2597                          * contains the dummy transaction, we can
2598                          * change state into IDLE (the second time around).
2599                          * Otherwise we should change the state into
2600                          * NEED a dummy.
2601                          * We don't need to cover the dummy.
2602                          */
2603                         if (!changed &&
2604                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2605                                         XLOG_COVER_OPS)) {
2606                                 changed = 1;
2607                         } else {
2608                                 /*
2609                                  * We have two dirty iclogs so start over
2610                                  * This could also be num of ops indicates
2611                                  * this is not the dummy going out.
2612                                  */
2613                                 changed = 2;
2614                         }
2615                         iclog->ic_header.h_num_logops = 0;
2616                         memset(iclog->ic_header.h_cycle_data, 0,
2617                               sizeof(iclog->ic_header.h_cycle_data));
2618                         iclog->ic_header.h_lsn = 0;
2619                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2620                         /* do nothing */;
2621                 else
2622                         break;  /* stop cleaning */
2623                 iclog = iclog->ic_next;
2624         } while (iclog != log->l_iclog);
2625
2626
2627         /*
2628          * Wake up threads waiting in xfs_log_force() for the dirty iclog
2629          * to be cleaned.
2630          */
2631         wake_up_all(&dirty_iclog->ic_force_wait);
2632
2633         /*
2634          * Change state for the dummy log recording.
2635          * We usually go to NEED. But we go to NEED2 if the changed indicates
2636          * we are done writing the dummy record.
2637          * If we are done with the second dummy recored (DONE2), then
2638          * we go to IDLE.
2639          */
2640         if (changed) {
2641                 switch (log->l_covered_state) {
2642                 case XLOG_STATE_COVER_IDLE:
2643                 case XLOG_STATE_COVER_NEED:
2644                 case XLOG_STATE_COVER_NEED2:
2645                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2646                         break;
2647
2648                 case XLOG_STATE_COVER_DONE:
2649                         if (changed == 1)
2650                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2651                         else
2652                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2653                         break;
2654
2655                 case XLOG_STATE_COVER_DONE2:
2656                         if (changed == 1)
2657                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2658                         else
2659                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2660                         break;
2661
2662                 default:
2663                         ASSERT(0);
2664                 }
2665         }
2666 }
2667
2668 STATIC xfs_lsn_t
2669 xlog_get_lowest_lsn(
2670         struct xlog             *log)
2671 {
2672         struct xlog_in_core     *iclog = log->l_iclog;
2673         xfs_lsn_t               lowest_lsn = 0, lsn;
2674
2675         do {
2676                 if (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))
2677                         continue;
2678
2679                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2680                 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2681                         lowest_lsn = lsn;
2682         } while ((iclog = iclog->ic_next) != log->l_iclog);
2683
2684         return lowest_lsn;
2685 }
2686
2687 /*
2688  * Completion of a iclog IO does not imply that a transaction has completed, as
2689  * transactions can be large enough to span many iclogs. We cannot change the
2690  * tail of the log half way through a transaction as this may be the only
2691  * transaction in the log and moving the tail to point to the middle of it
2692  * will prevent recovery from finding the start of the transaction. Hence we
2693  * should only update the last_sync_lsn if this iclog contains transaction
2694  * completion callbacks on it.
2695  *
2696  * We have to do this before we drop the icloglock to ensure we are the only one
2697  * that can update it.
2698  *
2699  * If we are moving the last_sync_lsn forwards, we also need to ensure we kick
2700  * the reservation grant head pushing. This is due to the fact that the push
2701  * target is bound by the current last_sync_lsn value. Hence if we have a large
2702  * amount of log space bound up in this committing transaction then the
2703  * last_sync_lsn value may be the limiting factor preventing tail pushing from
2704  * freeing space in the log. Hence once we've updated the last_sync_lsn we
2705  * should push the AIL to ensure the push target (and hence the grant head) is
2706  * no longer bound by the old log head location and can move forwards and make
2707  * progress again.
2708  */
2709 static void
2710 xlog_state_set_callback(
2711         struct xlog             *log,
2712         struct xlog_in_core     *iclog,
2713         xfs_lsn_t               header_lsn)
2714 {
2715         iclog->ic_state = XLOG_STATE_CALLBACK;
2716
2717         ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2718                            header_lsn) <= 0);
2719
2720         if (list_empty_careful(&iclog->ic_callbacks))
2721                 return;
2722
2723         atomic64_set(&log->l_last_sync_lsn, header_lsn);
2724         xlog_grant_push_ail(log, 0);
2725 }
2726
2727 /*
2728  * Return true if we need to stop processing, false to continue to the next
2729  * iclog. The caller will need to run callbacks if the iclog is returned in the
2730  * XLOG_STATE_CALLBACK state.
2731  */
2732 static bool
2733 xlog_state_iodone_process_iclog(
2734         struct xlog             *log,
2735         struct xlog_in_core     *iclog,
2736         struct xlog_in_core     *completed_iclog,
2737         bool                    *ioerror)
2738 {
2739         xfs_lsn_t               lowest_lsn;
2740         xfs_lsn_t               header_lsn;
2741
2742         /* Skip all iclogs in the ACTIVE & DIRTY states */
2743         if (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))
2744                 return false;
2745
2746         /*
2747          * Between marking a filesystem SHUTDOWN and stopping the log, we do
2748          * flush all iclogs to disk (if there wasn't a log I/O error). So, we do
2749          * want things to go smoothly in case of just a SHUTDOWN  w/o a
2750          * LOG_IO_ERROR.
2751          */
2752         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2753                 *ioerror = true;
2754                 return false;
2755         }
2756
2757         /*
2758          * Can only perform callbacks in order.  Since this iclog is not in the
2759          * DONE_SYNC/ DO_CALLBACK state, we skip the rest and just try to clean
2760          * up.  If we set our iclog to DO_CALLBACK, we will not process it when
2761          * we retry since a previous iclog is in the CALLBACK and the state
2762          * cannot change since we are holding the l_icloglock.
2763          */
2764         if (!(iclog->ic_state &
2765                         (XLOG_STATE_DONE_SYNC | XLOG_STATE_DO_CALLBACK))) {
2766                 if (completed_iclog &&
2767                     (completed_iclog->ic_state == XLOG_STATE_DONE_SYNC)) {
2768                         completed_iclog->ic_state = XLOG_STATE_DO_CALLBACK;
2769                 }
2770                 return true;
2771         }
2772
2773         /*
2774          * We now have an iclog that is in either the DO_CALLBACK or DONE_SYNC
2775          * states. The other states (WANT_SYNC, SYNCING, or CALLBACK were caught
2776          * by the above if and are going to clean (i.e. we aren't doing their
2777          * callbacks) see the above if.
2778          *
2779          * We will do one more check here to see if we have chased our tail
2780          * around. If this is not the lowest lsn iclog, then we will leave it
2781          * for another completion to process.
2782          */
2783         header_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2784         lowest_lsn = xlog_get_lowest_lsn(log);
2785         if (lowest_lsn && XFS_LSN_CMP(lowest_lsn, header_lsn) < 0)
2786                 return false;
2787
2788         xlog_state_set_callback(log, iclog, header_lsn);
2789         return false;
2790
2791 }
2792
2793 /*
2794  * Keep processing entries in the iclog callback list until we come around and
2795  * it is empty.  We need to atomically see that the list is empty and change the
2796  * state to DIRTY so that we don't miss any more callbacks being added.
2797  *
2798  * This function is called with the icloglock held and returns with it held. We
2799  * drop it while running callbacks, however, as holding it over thousands of
2800  * callbacks is unnecessary and causes excessive contention if we do.
2801  */
2802 static void
2803 xlog_state_do_iclog_callbacks(
2804         struct xlog             *log,
2805         struct xlog_in_core     *iclog,
2806         bool                    aborted)
2807 {
2808         spin_unlock(&log->l_icloglock);
2809         spin_lock(&iclog->ic_callback_lock);
2810         while (!list_empty(&iclog->ic_callbacks)) {
2811                 LIST_HEAD(tmp);
2812
2813                 list_splice_init(&iclog->ic_callbacks, &tmp);
2814
2815                 spin_unlock(&iclog->ic_callback_lock);
2816                 xlog_cil_process_committed(&tmp, aborted);
2817                 spin_lock(&iclog->ic_callback_lock);
2818         }
2819
2820         /*
2821          * Pick up the icloglock while still holding the callback lock so we
2822          * serialise against anyone trying to add more callbacks to this iclog
2823          * now we've finished processing.
2824          */
2825         spin_lock(&log->l_icloglock);
2826         spin_unlock(&iclog->ic_callback_lock);
2827 }
2828
2829 #ifdef DEBUG
2830 /*
2831  * Make one last gasp attempt to see if iclogs are being left in limbo.  If the
2832  * above loop finds an iclog earlier than the current iclog and in one of the
2833  * syncing states, the current iclog is put into DO_CALLBACK and the callbacks
2834  * are deferred to the completion of the earlier iclog. Walk the iclogs in order
2835  * and make sure that no iclog is in DO_CALLBACK unless an earlier iclog is in
2836  * one of the syncing states.
2837  *
2838  * Note that SYNCING|IOERROR is a valid state so we cannot just check for
2839  * ic_state == SYNCING.
2840  */
2841 static void
2842 xlog_state_callback_check_state(
2843         struct xlog             *log)
2844 {
2845         struct xlog_in_core     *first_iclog = log->l_iclog;
2846         struct xlog_in_core     *iclog = first_iclog;
2847
2848         do {
2849                 ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2850                 /*
2851                  * Terminate the loop if iclogs are found in states
2852                  * which will cause other threads to clean up iclogs.
2853                  *
2854                  * SYNCING - i/o completion will go through logs
2855                  * DONE_SYNC - interrupt thread should be waiting for
2856                  *              l_icloglock
2857                  * IOERROR - give up hope all ye who enter here
2858                  */
2859                 if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2860                     iclog->ic_state & XLOG_STATE_SYNCING ||
2861                     iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2862                     iclog->ic_state == XLOG_STATE_IOERROR )
2863                         break;
2864                 iclog = iclog->ic_next;
2865         } while (first_iclog != iclog);
2866 }
2867 #else
2868 #define xlog_state_callback_check_state(l)      ((void)0)
2869 #endif
2870
2871 STATIC void
2872 xlog_state_do_callback(
2873         struct xlog             *log,
2874         bool                    aborted,
2875         struct xlog_in_core     *ciclog)
2876 {
2877         struct xlog_in_core     *iclog;
2878         struct xlog_in_core     *first_iclog;
2879         bool                    did_callbacks = false;
2880         bool                    cycled_icloglock;
2881         bool                    ioerror;
2882         int                     flushcnt = 0;
2883         int                     repeats = 0;
2884
2885         spin_lock(&log->l_icloglock);
2886         do {
2887                 /*
2888                  * Scan all iclogs starting with the one pointed to by the
2889                  * log.  Reset this starting point each time the log is
2890                  * unlocked (during callbacks).
2891                  *
2892                  * Keep looping through iclogs until one full pass is made
2893                  * without running any callbacks.
2894                  */
2895                 first_iclog = log->l_iclog;
2896                 iclog = log->l_iclog;
2897                 cycled_icloglock = false;
2898                 ioerror = false;
2899                 repeats++;
2900
2901                 do {
2902                         if (xlog_state_iodone_process_iclog(log, iclog,
2903                                                         ciclog, &ioerror))
2904                                 break;
2905
2906                         if (!(iclog->ic_state &
2907                               (XLOG_STATE_CALLBACK | XLOG_STATE_IOERROR))) {
2908                                 iclog = iclog->ic_next;
2909                                 continue;
2910                         }
2911
2912                         /*
2913                          * Running callbacks will drop the icloglock which means
2914                          * we'll have to run at least one more complete loop.
2915                          */
2916                         cycled_icloglock = true;
2917                         xlog_state_do_iclog_callbacks(log, iclog, aborted);
2918
2919                         xlog_state_clean_iclog(log, iclog);
2920                         iclog = iclog->ic_next;
2921                 } while (first_iclog != iclog);
2922
2923                 did_callbacks |= cycled_icloglock;
2924
2925                 if (repeats > 5000) {
2926                         flushcnt += repeats;
2927                         repeats = 0;
2928                         xfs_warn(log->l_mp,
2929                                 "%s: possible infinite loop (%d iterations)",
2930                                 __func__, flushcnt);
2931                 }
2932         } while (!ioerror && cycled_icloglock);
2933
2934         if (did_callbacks)
2935                 xlog_state_callback_check_state(log);
2936
2937         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2938                 wake_up_all(&log->l_flush_wait);
2939
2940         spin_unlock(&log->l_icloglock);
2941 }
2942
2943
2944 /*
2945  * Finish transitioning this iclog to the dirty state.
2946  *
2947  * Make sure that we completely execute this routine only when this is
2948  * the last call to the iclog.  There is a good chance that iclog flushes,
2949  * when we reach the end of the physical log, get turned into 2 separate
2950  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2951  * routine.  By using the reference count bwritecnt, we guarantee that only
2952  * the second completion goes through.
2953  *
2954  * Callbacks could take time, so they are done outside the scope of the
2955  * global state machine log lock.
2956  */
2957 STATIC void
2958 xlog_state_done_syncing(
2959         struct xlog_in_core     *iclog,
2960         bool                    aborted)
2961 {
2962         struct xlog             *log = iclog->ic_log;
2963
2964         spin_lock(&log->l_icloglock);
2965
2966         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2967                iclog->ic_state == XLOG_STATE_IOERROR);
2968         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2969
2970         /*
2971          * If we got an error, either on the first buffer, or in the case of
2972          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2973          * and none should ever be attempted to be written to disk
2974          * again.
2975          */
2976         if (iclog->ic_state != XLOG_STATE_IOERROR)
2977                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2978
2979         /*
2980          * Someone could be sleeping prior to writing out the next
2981          * iclog buffer, we wake them all, one will get to do the
2982          * I/O, the others get to wait for the result.
2983          */
2984         wake_up_all(&iclog->ic_write_wait);
2985         spin_unlock(&log->l_icloglock);
2986         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2987 }       /* xlog_state_done_syncing */
2988
2989
2990 /*
2991  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2992  * sleep.  We wait on the flush queue on the head iclog as that should be
2993  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2994  * we will wait here and all new writes will sleep until a sync completes.
2995  *
2996  * The in-core logs are used in a circular fashion. They are not used
2997  * out-of-order even when an iclog past the head is free.
2998  *
2999  * return:
3000  *      * log_offset where xlog_write() can start writing into the in-core
3001  *              log's data space.
3002  *      * in-core log pointer to which xlog_write() should write.
3003  *      * boolean indicating this is a continued write to an in-core log.
3004  *              If this is the last write, then the in-core log's offset field
3005  *              needs to be incremented, depending on the amount of data which
3006  *              is copied.
3007  */
3008 STATIC int
3009 xlog_state_get_iclog_space(
3010         struct xlog             *log,
3011         int                     len,
3012         struct xlog_in_core     **iclogp,
3013         struct xlog_ticket      *ticket,
3014         int                     *continued_write,
3015         int                     *logoffsetp)
3016 {
3017         int               log_offset;
3018         xlog_rec_header_t *head;
3019         xlog_in_core_t    *iclog;
3020         int               error;
3021
3022 restart:
3023         spin_lock(&log->l_icloglock);
3024         if (XLOG_FORCED_SHUTDOWN(log)) {
3025                 spin_unlock(&log->l_icloglock);
3026                 return -EIO;
3027         }
3028
3029         iclog = log->l_iclog;
3030         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
3031                 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
3032
3033                 /* Wait for log writes to have flushed */
3034                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
3035                 goto restart;
3036         }
3037
3038         head = &iclog->ic_header;
3039
3040         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
3041         log_offset = iclog->ic_offset;
3042
3043         /* On the 1st write to an iclog, figure out lsn.  This works
3044          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3045          * committing to.  If the offset is set, that's how many blocks
3046          * must be written.
3047          */
3048         if (log_offset == 0) {
3049                 ticket->t_curr_res -= log->l_iclog_hsize;
3050                 xlog_tic_add_region(ticket,
3051                                     log->l_iclog_hsize,
3052                                     XLOG_REG_TYPE_LRHEADER);
3053                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3054                 head->h_lsn = cpu_to_be64(
3055                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3056                 ASSERT(log->l_curr_block >= 0);
3057         }
3058
3059         /* If there is enough room to write everything, then do it.  Otherwise,
3060          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3061          * bit is on, so this will get flushed out.  Don't update ic_offset
3062          * until you know exactly how many bytes get copied.  Therefore, wait
3063          * until later to update ic_offset.
3064          *
3065          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3066          * can fit into remaining data section.
3067          */
3068         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3069                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3070
3071                 /*
3072                  * If I'm the only one writing to this iclog, sync it to disk.
3073                  * We need to do an atomic compare and decrement here to avoid
3074                  * racing with concurrent atomic_dec_and_lock() calls in
3075                  * xlog_state_release_iclog() when there is more than one
3076                  * reference to the iclog.
3077                  */
3078                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3079                         /* we are the only one */
3080                         spin_unlock(&log->l_icloglock);
3081                         error = xlog_state_release_iclog(log, iclog);
3082                         if (error)
3083                                 return error;
3084                 } else {
3085                         spin_unlock(&log->l_icloglock);
3086                 }
3087                 goto restart;
3088         }
3089
3090         /* Do we have enough room to write the full amount in the remainder
3091          * of this iclog?  Or must we continue a write on the next iclog and
3092          * mark this iclog as completely taken?  In the case where we switch
3093          * iclogs (to mark it taken), this particular iclog will release/sync
3094          * to disk in xlog_write().
3095          */
3096         if (len <= iclog->ic_size - iclog->ic_offset) {
3097                 *continued_write = 0;
3098                 iclog->ic_offset += len;
3099         } else {
3100                 *continued_write = 1;
3101                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3102         }
3103         *iclogp = iclog;
3104
3105         ASSERT(iclog->ic_offset <= iclog->ic_size);
3106         spin_unlock(&log->l_icloglock);
3107
3108         *logoffsetp = log_offset;
3109         return 0;
3110 }       /* xlog_state_get_iclog_space */
3111
3112 /* The first cnt-1 times through here we don't need to
3113  * move the grant write head because the permanent
3114  * reservation has reserved cnt times the unit amount.
3115  * Release part of current permanent unit reservation and
3116  * reset current reservation to be one units worth.  Also
3117  * move grant reservation head forward.
3118  */
3119 STATIC void
3120 xlog_regrant_reserve_log_space(
3121         struct xlog             *log,
3122         struct xlog_ticket      *ticket)
3123 {
3124         trace_xfs_log_regrant_reserve_enter(log, ticket);
3125
3126         if (ticket->t_cnt > 0)
3127                 ticket->t_cnt--;
3128
3129         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3130                                         ticket->t_curr_res);
3131         xlog_grant_sub_space(log, &log->l_write_head.grant,
3132                                         ticket->t_curr_res);
3133         ticket->t_curr_res = ticket->t_unit_res;
3134         xlog_tic_reset_res(ticket);
3135
3136         trace_xfs_log_regrant_reserve_sub(log, ticket);
3137
3138         /* just return if we still have some of the pre-reserved space */
3139         if (ticket->t_cnt > 0)
3140                 return;
3141
3142         xlog_grant_add_space(log, &log->l_reserve_head.grant,
3143                                         ticket->t_unit_res);
3144
3145         trace_xfs_log_regrant_reserve_exit(log, ticket);
3146
3147         ticket->t_curr_res = ticket->t_unit_res;
3148         xlog_tic_reset_res(ticket);
3149 }       /* xlog_regrant_reserve_log_space */
3150
3151
3152 /*
3153  * Give back the space left from a reservation.
3154  *
3155  * All the information we need to make a correct determination of space left
3156  * is present.  For non-permanent reservations, things are quite easy.  The
3157  * count should have been decremented to zero.  We only need to deal with the
3158  * space remaining in the current reservation part of the ticket.  If the
3159  * ticket contains a permanent reservation, there may be left over space which
3160  * needs to be released.  A count of N means that N-1 refills of the current
3161  * reservation can be done before we need to ask for more space.  The first
3162  * one goes to fill up the first current reservation.  Once we run out of
3163  * space, the count will stay at zero and the only space remaining will be
3164  * in the current reservation field.
3165  */
3166 STATIC void
3167 xlog_ungrant_log_space(
3168         struct xlog             *log,
3169         struct xlog_ticket      *ticket)
3170 {
3171         int     bytes;
3172
3173         if (ticket->t_cnt > 0)
3174                 ticket->t_cnt--;
3175
3176         trace_xfs_log_ungrant_enter(log, ticket);
3177         trace_xfs_log_ungrant_sub(log, ticket);
3178
3179         /*
3180          * If this is a permanent reservation ticket, we may be able to free
3181          * up more space based on the remaining count.
3182          */
3183         bytes = ticket->t_curr_res;
3184         if (ticket->t_cnt > 0) {
3185                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3186                 bytes += ticket->t_unit_res*ticket->t_cnt;
3187         }
3188
3189         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3190         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3191
3192         trace_xfs_log_ungrant_exit(log, ticket);
3193
3194         xfs_log_space_wake(log->l_mp);
3195 }
3196
3197 /*
3198  * Flush iclog to disk if this is the last reference to the given iclog and
3199  * the WANT_SYNC bit is set.
3200  *
3201  * When this function is entered, the iclog is not necessarily in the
3202  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3203  *
3204  *
3205  */
3206 STATIC int
3207 xlog_state_release_iclog(
3208         struct xlog             *log,
3209         struct xlog_in_core     *iclog)
3210 {
3211         int             sync = 0;       /* do we sync? */
3212
3213         if (iclog->ic_state & XLOG_STATE_IOERROR)
3214                 return -EIO;
3215
3216         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3217         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3218                 return 0;
3219
3220         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3221                 spin_unlock(&log->l_icloglock);
3222                 return -EIO;
3223         }
3224         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3225                iclog->ic_state == XLOG_STATE_WANT_SYNC);
3226
3227         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3228                 /* update tail before writing to iclog */
3229                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3230                 sync++;
3231                 iclog->ic_state = XLOG_STATE_SYNCING;
3232                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3233                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3234                 /* cycle incremented when incrementing curr_block */
3235         }
3236         spin_unlock(&log->l_icloglock);
3237
3238         /*
3239          * We let the log lock go, so it's possible that we hit a log I/O
3240          * error or some other SHUTDOWN condition that marks the iclog
3241          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3242          * this iclog has consistent data, so we ignore IOERROR
3243          * flags after this point.
3244          */
3245         if (sync)
3246                 xlog_sync(log, iclog);
3247         return 0;
3248 }       /* xlog_state_release_iclog */
3249
3250
3251 /*
3252  * This routine will mark the current iclog in the ring as WANT_SYNC
3253  * and move the current iclog pointer to the next iclog in the ring.
3254  * When this routine is called from xlog_state_get_iclog_space(), the
3255  * exact size of the iclog has not yet been determined.  All we know is
3256  * that every data block.  We have run out of space in this log record.
3257  */
3258 STATIC void
3259 xlog_state_switch_iclogs(
3260         struct xlog             *log,
3261         struct xlog_in_core     *iclog,
3262         int                     eventual_size)
3263 {
3264         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3265         if (!eventual_size)
3266                 eventual_size = iclog->ic_offset;
3267         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3268         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3269         log->l_prev_block = log->l_curr_block;
3270         log->l_prev_cycle = log->l_curr_cycle;
3271
3272         /* roll log?: ic_offset changed later */
3273         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3274
3275         /* Round up to next log-sunit */
3276         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3277             log->l_mp->m_sb.sb_logsunit > 1) {
3278                 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3279                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3280         }
3281
3282         if (log->l_curr_block >= log->l_logBBsize) {
3283                 /*
3284                  * Rewind the current block before the cycle is bumped to make
3285                  * sure that the combined LSN never transiently moves forward
3286                  * when the log wraps to the next cycle. This is to support the
3287                  * unlocked sample of these fields from xlog_valid_lsn(). Most
3288                  * other cases should acquire l_icloglock.
3289                  */
3290                 log->l_curr_block -= log->l_logBBsize;
3291                 ASSERT(log->l_curr_block >= 0);
3292                 smp_wmb();
3293                 log->l_curr_cycle++;
3294                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3295                         log->l_curr_cycle++;
3296         }
3297         ASSERT(iclog == log->l_iclog);
3298         log->l_iclog = iclog->ic_next;
3299 }       /* xlog_state_switch_iclogs */
3300
3301 /*
3302  * Write out all data in the in-core log as of this exact moment in time.
3303  *
3304  * Data may be written to the in-core log during this call.  However,
3305  * we don't guarantee this data will be written out.  A change from past
3306  * implementation means this routine will *not* write out zero length LRs.
3307  *
3308  * Basically, we try and perform an intelligent scan of the in-core logs.
3309  * If we determine there is no flushable data, we just return.  There is no
3310  * flushable data if:
3311  *
3312  *      1. the current iclog is active and has no data; the previous iclog
3313  *              is in the active or dirty state.
3314  *      2. the current iclog is drity, and the previous iclog is in the
3315  *              active or dirty state.
3316  *
3317  * We may sleep if:
3318  *
3319  *      1. the current iclog is not in the active nor dirty state.
3320  *      2. the current iclog dirty, and the previous iclog is not in the
3321  *              active nor dirty state.
3322  *      3. the current iclog is active, and there is another thread writing
3323  *              to this particular iclog.
3324  *      4. a) the current iclog is active and has no other writers
3325  *         b) when we return from flushing out this iclog, it is still
3326  *              not in the active nor dirty state.
3327  */
3328 int
3329 xfs_log_force(
3330         struct xfs_mount        *mp,
3331         uint                    flags)
3332 {
3333         struct xlog             *log = mp->m_log;
3334         struct xlog_in_core     *iclog;
3335         xfs_lsn_t               lsn;
3336
3337         XFS_STATS_INC(mp, xs_log_force);
3338         trace_xfs_log_force(mp, 0, _RET_IP_);
3339
3340         xlog_cil_force(log);
3341
3342         spin_lock(&log->l_icloglock);
3343         iclog = log->l_iclog;
3344         if (iclog->ic_state & XLOG_STATE_IOERROR)
3345                 goto out_error;
3346
3347         if (iclog->ic_state == XLOG_STATE_DIRTY ||
3348             (iclog->ic_state == XLOG_STATE_ACTIVE &&
3349              atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3350                 /*
3351                  * If the head is dirty or (active and empty), then we need to
3352                  * look at the previous iclog.
3353                  *
3354                  * If the previous iclog is active or dirty we are done.  There
3355                  * is nothing to sync out. Otherwise, we attach ourselves to the
3356                  * previous iclog and go to sleep.
3357                  */
3358                 iclog = iclog->ic_prev;
3359                 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3360                     iclog->ic_state == XLOG_STATE_DIRTY)
3361                         goto out_unlock;
3362         } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3363                 if (atomic_read(&iclog->ic_refcnt) == 0) {
3364                         /*
3365                          * We are the only one with access to this iclog.
3366                          *
3367                          * Flush it out now.  There should be a roundoff of zero
3368                          * to show that someone has already taken care of the
3369                          * roundoff from the previous sync.
3370                          */
3371                         atomic_inc(&iclog->ic_refcnt);
3372                         lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3373                         xlog_state_switch_iclogs(log, iclog, 0);
3374                         spin_unlock(&log->l_icloglock);
3375
3376                         if (xlog_state_release_iclog(log, iclog))
3377                                 return -EIO;
3378
3379                         spin_lock(&log->l_icloglock);
3380                         if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3381                             iclog->ic_state == XLOG_STATE_DIRTY)
3382                                 goto out_unlock;
3383                 } else {
3384                         /*
3385                          * Someone else is writing to this iclog.
3386                          *
3387                          * Use its call to flush out the data.  However, the
3388                          * other thread may not force out this LR, so we mark
3389                          * it WANT_SYNC.
3390                          */
3391                         xlog_state_switch_iclogs(log, iclog, 0);
3392                 }
3393         } else {
3394                 /*
3395                  * If the head iclog is not active nor dirty, we just attach
3396                  * ourselves to the head and go to sleep if necessary.
3397                  */
3398                 ;
3399         }
3400
3401         if (!(flags & XFS_LOG_SYNC))
3402                 goto out_unlock;
3403
3404         if (iclog->ic_state & XLOG_STATE_IOERROR)
3405                 goto out_error;
3406         XFS_STATS_INC(mp, xs_log_force_sleep);
3407         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3408         if (iclog->ic_state & XLOG_STATE_IOERROR)
3409                 return -EIO;
3410         return 0;
3411
3412 out_unlock:
3413         spin_unlock(&log->l_icloglock);
3414         return 0;
3415 out_error:
3416         spin_unlock(&log->l_icloglock);
3417         return -EIO;
3418 }
3419
3420 static int
3421 __xfs_log_force_lsn(
3422         struct xfs_mount        *mp,
3423         xfs_lsn_t               lsn,
3424         uint                    flags,
3425         int                     *log_flushed,
3426         bool                    already_slept)
3427 {
3428         struct xlog             *log = mp->m_log;
3429         struct xlog_in_core     *iclog;
3430
3431         spin_lock(&log->l_icloglock);
3432         iclog = log->l_iclog;
3433         if (iclog->ic_state & XLOG_STATE_IOERROR)
3434                 goto out_error;
3435
3436         while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3437                 iclog = iclog->ic_next;
3438                 if (iclog == log->l_iclog)
3439                         goto out_unlock;
3440         }
3441
3442         if (iclog->ic_state == XLOG_STATE_DIRTY)
3443                 goto out_unlock;
3444
3445         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3446                 /*
3447                  * We sleep here if we haven't already slept (e.g. this is the
3448                  * first time we've looked at the correct iclog buf) and the
3449                  * buffer before us is going to be sync'ed.  The reason for this
3450                  * is that if we are doing sync transactions here, by waiting
3451                  * for the previous I/O to complete, we can allow a few more
3452                  * transactions into this iclog before we close it down.
3453                  *
3454                  * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3455                  * refcnt so we can release the log (which drops the ref count).
3456                  * The state switch keeps new transaction commits from using
3457                  * this buffer.  When the current commits finish writing into
3458                  * the buffer, the refcount will drop to zero and the buffer
3459                  * will go out then.
3460                  */
3461                 if (!already_slept &&
3462                     (iclog->ic_prev->ic_state &
3463                      (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3464                         ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3465
3466                         XFS_STATS_INC(mp, xs_log_force_sleep);
3467
3468                         xlog_wait(&iclog->ic_prev->ic_write_wait,
3469                                         &log->l_icloglock);
3470                         return -EAGAIN;
3471                 }
3472                 atomic_inc(&iclog->ic_refcnt);
3473                 xlog_state_switch_iclogs(log, iclog, 0);
3474                 spin_unlock(&log->l_icloglock);
3475                 if (xlog_state_release_iclog(log, iclog))
3476                         return -EIO;
3477                 if (log_flushed)
3478                         *log_flushed = 1;
3479                 spin_lock(&log->l_icloglock);
3480         }
3481
3482         if (!(flags & XFS_LOG_SYNC) ||
3483             (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3484                 goto out_unlock;
3485
3486         if (iclog->ic_state & XLOG_STATE_IOERROR)
3487                 goto out_error;
3488
3489         XFS_STATS_INC(mp, xs_log_force_sleep);
3490         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3491         if (iclog->ic_state & XLOG_STATE_IOERROR)
3492                 return -EIO;
3493         return 0;
3494
3495 out_unlock:
3496         spin_unlock(&log->l_icloglock);
3497         return 0;
3498 out_error:
3499         spin_unlock(&log->l_icloglock);
3500         return -EIO;
3501 }
3502
3503 /*
3504  * Force the in-core log to disk for a specific LSN.
3505  *
3506  * Find in-core log with lsn.
3507  *      If it is in the DIRTY state, just return.
3508  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3509  *              state and go to sleep or return.
3510  *      If it is in any other state, go to sleep or return.
3511  *
3512  * Synchronous forces are implemented with a wait queue.  All callers trying
3513  * to force a given lsn to disk must wait on the queue attached to the
3514  * specific in-core log.  When given in-core log finally completes its write
3515  * to disk, that thread will wake up all threads waiting on the queue.
3516  */
3517 int
3518 xfs_log_force_lsn(
3519         struct xfs_mount        *mp,
3520         xfs_lsn_t               lsn,
3521         uint                    flags,
3522         int                     *log_flushed)
3523 {
3524         int                     ret;
3525         ASSERT(lsn != 0);
3526
3527         XFS_STATS_INC(mp, xs_log_force);
3528         trace_xfs_log_force(mp, lsn, _RET_IP_);
3529
3530         lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3531         if (lsn == NULLCOMMITLSN)
3532                 return 0;
3533
3534         ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3535         if (ret == -EAGAIN)
3536                 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3537         return ret;
3538 }
3539
3540 /*
3541  * Called when we want to mark the current iclog as being ready to sync to
3542  * disk.
3543  */
3544 STATIC void
3545 xlog_state_want_sync(
3546         struct xlog             *log,
3547         struct xlog_in_core     *iclog)
3548 {
3549         assert_spin_locked(&log->l_icloglock);
3550
3551         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3552                 xlog_state_switch_iclogs(log, iclog, 0);
3553         } else {
3554                 ASSERT(iclog->ic_state &
3555                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3556         }
3557 }
3558
3559
3560 /*****************************************************************************
3561  *
3562  *              TICKET functions
3563  *
3564  *****************************************************************************
3565  */
3566
3567 /*
3568  * Free a used ticket when its refcount falls to zero.
3569  */
3570 void
3571 xfs_log_ticket_put(
3572         xlog_ticket_t   *ticket)
3573 {
3574         ASSERT(atomic_read(&ticket->t_ref) > 0);
3575         if (atomic_dec_and_test(&ticket->t_ref))
3576                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3577 }
3578
3579 xlog_ticket_t *
3580 xfs_log_ticket_get(
3581         xlog_ticket_t   *ticket)
3582 {
3583         ASSERT(atomic_read(&ticket->t_ref) > 0);
3584         atomic_inc(&ticket->t_ref);
3585         return ticket;
3586 }
3587
3588 /*
3589  * Figure out the total log space unit (in bytes) that would be
3590  * required for a log ticket.
3591  */
3592 int
3593 xfs_log_calc_unit_res(
3594         struct xfs_mount        *mp,
3595         int                     unit_bytes)
3596 {
3597         struct xlog             *log = mp->m_log;
3598         int                     iclog_space;
3599         uint                    num_headers;
3600
3601         /*
3602          * Permanent reservations have up to 'cnt'-1 active log operations
3603          * in the log.  A unit in this case is the amount of space for one
3604          * of these log operations.  Normal reservations have a cnt of 1
3605          * and their unit amount is the total amount of space required.
3606          *
3607          * The following lines of code account for non-transaction data
3608          * which occupy space in the on-disk log.
3609          *
3610          * Normal form of a transaction is:
3611          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3612          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3613          *
3614          * We need to account for all the leadup data and trailer data
3615          * around the transaction data.
3616          * And then we need to account for the worst case in terms of using
3617          * more space.
3618          * The worst case will happen if:
3619          * - the placement of the transaction happens to be such that the
3620          *   roundoff is at its maximum
3621          * - the transaction data is synced before the commit record is synced
3622          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3623          *   Therefore the commit record is in its own Log Record.
3624          *   This can happen as the commit record is called with its
3625          *   own region to xlog_write().
3626          *   This then means that in the worst case, roundoff can happen for
3627          *   the commit-rec as well.
3628          *   The commit-rec is smaller than padding in this scenario and so it is
3629          *   not added separately.
3630          */
3631
3632         /* for trans header */
3633         unit_bytes += sizeof(xlog_op_header_t);
3634         unit_bytes += sizeof(xfs_trans_header_t);
3635
3636         /* for start-rec */
3637         unit_bytes += sizeof(xlog_op_header_t);
3638
3639         /*
3640          * for LR headers - the space for data in an iclog is the size minus
3641          * the space used for the headers. If we use the iclog size, then we
3642          * undercalculate the number of headers required.
3643          *
3644          * Furthermore - the addition of op headers for split-recs might
3645          * increase the space required enough to require more log and op
3646          * headers, so take that into account too.
3647          *
3648          * IMPORTANT: This reservation makes the assumption that if this
3649          * transaction is the first in an iclog and hence has the LR headers
3650          * accounted to it, then the remaining space in the iclog is
3651          * exclusively for this transaction.  i.e. if the transaction is larger
3652          * than the iclog, it will be the only thing in that iclog.
3653          * Fundamentally, this means we must pass the entire log vector to
3654          * xlog_write to guarantee this.
3655          */
3656         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3657         num_headers = howmany(unit_bytes, iclog_space);
3658
3659         /* for split-recs - ophdrs added when data split over LRs */
3660         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3661
3662         /* add extra header reservations if we overrun */
3663         while (!num_headers ||
3664                howmany(unit_bytes, iclog_space) > num_headers) {
3665                 unit_bytes += sizeof(xlog_op_header_t);
3666                 num_headers++;
3667         }
3668         unit_bytes += log->l_iclog_hsize * num_headers;
3669
3670         /* for commit-rec LR header - note: padding will subsume the ophdr */
3671         unit_bytes += log->l_iclog_hsize;
3672
3673         /* for roundoff padding for transaction data and one for commit record */
3674         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3675                 /* log su roundoff */
3676                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3677         } else {
3678                 /* BB roundoff */
3679                 unit_bytes += 2 * BBSIZE;
3680         }
3681
3682         return unit_bytes;
3683 }
3684
3685 /*
3686  * Allocate and initialise a new log ticket.
3687  */
3688 struct xlog_ticket *
3689 xlog_ticket_alloc(
3690         struct xlog             *log,
3691         int                     unit_bytes,
3692         int                     cnt,
3693         char                    client,
3694         bool                    permanent,
3695         xfs_km_flags_t          alloc_flags)
3696 {
3697         struct xlog_ticket      *tic;
3698         int                     unit_res;
3699
3700         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3701         if (!tic)
3702                 return NULL;
3703
3704         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3705
3706         atomic_set(&tic->t_ref, 1);
3707         tic->t_task             = current;
3708         INIT_LIST_HEAD(&tic->t_queue);
3709         tic->t_unit_res         = unit_res;
3710         tic->t_curr_res         = unit_res;
3711         tic->t_cnt              = cnt;
3712         tic->t_ocnt             = cnt;
3713         tic->t_tid              = prandom_u32();
3714         tic->t_clientid         = client;
3715         tic->t_flags            = XLOG_TIC_INITED;
3716         if (permanent)
3717                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3718
3719         xlog_tic_reset_res(tic);
3720
3721         return tic;
3722 }
3723
3724
3725 /******************************************************************************
3726  *
3727  *              Log debug routines
3728  *
3729  ******************************************************************************
3730  */
3731 #if defined(DEBUG)
3732 /*
3733  * Make sure that the destination ptr is within the valid data region of
3734  * one of the iclogs.  This uses backup pointers stored in a different
3735  * part of the log in case we trash the log structure.
3736  */
3737 STATIC void
3738 xlog_verify_dest_ptr(
3739         struct xlog     *log,
3740         void            *ptr)
3741 {
3742         int i;
3743         int good_ptr = 0;
3744
3745         for (i = 0; i < log->l_iclog_bufs; i++) {
3746                 if (ptr >= log->l_iclog_bak[i] &&
3747                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3748                         good_ptr++;
3749         }
3750
3751         if (!good_ptr)
3752                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3753 }
3754
3755 /*
3756  * Check to make sure the grant write head didn't just over lap the tail.  If
3757  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3758  * the cycles differ by exactly one and check the byte count.
3759  *
3760  * This check is run unlocked, so can give false positives. Rather than assert
3761  * on failures, use a warn-once flag and a panic tag to allow the admin to
3762  * determine if they want to panic the machine when such an error occurs. For
3763  * debug kernels this will have the same effect as using an assert but, unlinke
3764  * an assert, it can be turned off at runtime.
3765  */
3766 STATIC void
3767 xlog_verify_grant_tail(
3768         struct xlog     *log)
3769 {
3770         int             tail_cycle, tail_blocks;
3771         int             cycle, space;
3772
3773         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3774         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3775         if (tail_cycle != cycle) {
3776                 if (cycle - 1 != tail_cycle &&
3777                     !(log->l_flags & XLOG_TAIL_WARN)) {
3778                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3779                                 "%s: cycle - 1 != tail_cycle", __func__);
3780                         log->l_flags |= XLOG_TAIL_WARN;
3781                 }
3782
3783                 if (space > BBTOB(tail_blocks) &&
3784                     !(log->l_flags & XLOG_TAIL_WARN)) {
3785                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3786                                 "%s: space > BBTOB(tail_blocks)", __func__);
3787                         log->l_flags |= XLOG_TAIL_WARN;
3788                 }
3789         }
3790 }
3791
3792 /* check if it will fit */
3793 STATIC void
3794 xlog_verify_tail_lsn(
3795         struct xlog             *log,
3796         struct xlog_in_core     *iclog,
3797         xfs_lsn_t               tail_lsn)
3798 {
3799     int blocks;
3800
3801     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3802         blocks =
3803             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3804         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3805                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3806     } else {
3807         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3808
3809         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3810                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3811
3812         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3813         if (blocks < BTOBB(iclog->ic_offset) + 1)
3814                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3815     }
3816 }       /* xlog_verify_tail_lsn */
3817
3818 /*
3819  * Perform a number of checks on the iclog before writing to disk.
3820  *
3821  * 1. Make sure the iclogs are still circular
3822  * 2. Make sure we have a good magic number
3823  * 3. Make sure we don't have magic numbers in the data
3824  * 4. Check fields of each log operation header for:
3825  *      A. Valid client identifier
3826  *      B. tid ptr value falls in valid ptr space (user space code)
3827  *      C. Length in log record header is correct according to the
3828  *              individual operation headers within record.
3829  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3830  *      log, check the preceding blocks of the physical log to make sure all
3831  *      the cycle numbers agree with the current cycle number.
3832  */
3833 STATIC void
3834 xlog_verify_iclog(
3835         struct xlog             *log,
3836         struct xlog_in_core     *iclog,
3837         int                     count)
3838 {
3839         xlog_op_header_t        *ophead;
3840         xlog_in_core_t          *icptr;
3841         xlog_in_core_2_t        *xhdr;
3842         void                    *base_ptr, *ptr, *p;
3843         ptrdiff_t               field_offset;
3844         uint8_t                 clientid;
3845         int                     len, i, j, k, op_len;
3846         int                     idx;
3847
3848         /* check validity of iclog pointers */
3849         spin_lock(&log->l_icloglock);
3850         icptr = log->l_iclog;
3851         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3852                 ASSERT(icptr);
3853
3854         if (icptr != log->l_iclog)
3855                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3856         spin_unlock(&log->l_icloglock);
3857
3858         /* check log magic numbers */
3859         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3860                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3861
3862         base_ptr = ptr = &iclog->ic_header;
3863         p = &iclog->ic_header;
3864         for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3865                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3866                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3867                                 __func__);
3868         }
3869
3870         /* check fields */
3871         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3872         base_ptr = ptr = iclog->ic_datap;
3873         ophead = ptr;
3874         xhdr = iclog->ic_data;
3875         for (i = 0; i < len; i++) {
3876                 ophead = ptr;
3877
3878                 /* clientid is only 1 byte */
3879                 p = &ophead->oh_clientid;
3880                 field_offset = p - base_ptr;
3881                 if (field_offset & 0x1ff) {
3882                         clientid = ophead->oh_clientid;
3883                 } else {
3884                         idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3885                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3886                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3887                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3888                                 clientid = xlog_get_client_id(
3889                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3890                         } else {
3891                                 clientid = xlog_get_client_id(
3892                                         iclog->ic_header.h_cycle_data[idx]);
3893                         }
3894                 }
3895                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3896                         xfs_warn(log->l_mp,
3897                                 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3898                                 __func__, clientid, ophead,
3899                                 (unsigned long)field_offset);
3900
3901                 /* check length */
3902                 p = &ophead->oh_len;
3903                 field_offset = p - base_ptr;
3904                 if (field_offset & 0x1ff) {
3905                         op_len = be32_to_cpu(ophead->oh_len);
3906                 } else {
3907                         idx = BTOBBT((uintptr_t)&ophead->oh_len -
3908                                     (uintptr_t)iclog->ic_datap);
3909                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3910                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3911                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3912                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3913                         } else {
3914                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3915                         }
3916                 }
3917                 ptr += sizeof(xlog_op_header_t) + op_len;
3918         }
3919 }       /* xlog_verify_iclog */
3920 #endif
3921
3922 /*
3923  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3924  */
3925 STATIC int
3926 xlog_state_ioerror(
3927         struct xlog     *log)
3928 {
3929         xlog_in_core_t  *iclog, *ic;
3930
3931         iclog = log->l_iclog;
3932         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3933                 /*
3934                  * Mark all the incore logs IOERROR.
3935                  * From now on, no log flushes will result.
3936                  */
3937                 ic = iclog;
3938                 do {
3939                         ic->ic_state = XLOG_STATE_IOERROR;
3940                         ic = ic->ic_next;
3941                 } while (ic != iclog);
3942                 return 0;
3943         }
3944         /*
3945          * Return non-zero, if state transition has already happened.
3946          */
3947         return 1;
3948 }
3949
3950 /*
3951  * This is called from xfs_force_shutdown, when we're forcibly
3952  * shutting down the filesystem, typically because of an IO error.
3953  * Our main objectives here are to make sure that:
3954  *      a. if !logerror, flush the logs to disk. Anything modified
3955  *         after this is ignored.
3956  *      b. the filesystem gets marked 'SHUTDOWN' for all interested
3957  *         parties to find out, 'atomically'.
3958  *      c. those who're sleeping on log reservations, pinned objects and
3959  *          other resources get woken up, and be told the bad news.
3960  *      d. nothing new gets queued up after (b) and (c) are done.
3961  *
3962  * Note: for the !logerror case we need to flush the regions held in memory out
3963  * to disk first. This needs to be done before the log is marked as shutdown,
3964  * otherwise the iclog writes will fail.
3965  */
3966 int
3967 xfs_log_force_umount(
3968         struct xfs_mount        *mp,
3969         int                     logerror)
3970 {
3971         struct xlog     *log;
3972         int             retval;
3973
3974         log = mp->m_log;
3975
3976         /*
3977          * If this happens during log recovery, don't worry about
3978          * locking; the log isn't open for business yet.
3979          */
3980         if (!log ||
3981             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3982                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3983                 if (mp->m_sb_bp)
3984                         mp->m_sb_bp->b_flags |= XBF_DONE;
3985                 return 0;
3986         }
3987
3988         /*
3989          * Somebody could've already done the hard work for us.
3990          * No need to get locks for this.
3991          */
3992         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3993                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3994                 return 1;
3995         }
3996
3997         /*
3998          * Flush all the completed transactions to disk before marking the log
3999          * being shut down. We need to do it in this order to ensure that
4000          * completed operations are safely on disk before we shut down, and that
4001          * we don't have to issue any buffer IO after the shutdown flags are set
4002          * to guarantee this.
4003          */
4004         if (!logerror)
4005                 xfs_log_force(mp, XFS_LOG_SYNC);
4006
4007         /*
4008          * mark the filesystem and the as in a shutdown state and wake
4009          * everybody up to tell them the bad news.
4010          */
4011         spin_lock(&log->l_icloglock);
4012         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
4013         if (mp->m_sb_bp)
4014                 mp->m_sb_bp->b_flags |= XBF_DONE;
4015
4016         /*
4017          * Mark the log and the iclogs with IO error flags to prevent any
4018          * further log IO from being issued or completed.
4019          */
4020         log->l_flags |= XLOG_IO_ERROR;
4021         retval = xlog_state_ioerror(log);
4022         spin_unlock(&log->l_icloglock);
4023
4024         /*
4025          * We don't want anybody waiting for log reservations after this. That
4026          * means we have to wake up everybody queued up on reserveq as well as
4027          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
4028          * we don't enqueue anything once the SHUTDOWN flag is set, and this
4029          * action is protected by the grant locks.
4030          */
4031         xlog_grant_head_wake_all(&log->l_reserve_head);
4032         xlog_grant_head_wake_all(&log->l_write_head);
4033
4034         /*
4035          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4036          * as if the log writes were completed. The abort handling in the log
4037          * item committed callback functions will do this again under lock to
4038          * avoid races.
4039          */
4040         spin_lock(&log->l_cilp->xc_push_lock);
4041         wake_up_all(&log->l_cilp->xc_commit_wait);
4042         spin_unlock(&log->l_cilp->xc_push_lock);
4043         xlog_state_do_callback(log, true, NULL);
4044
4045 #ifdef XFSERRORDEBUG
4046         {
4047                 xlog_in_core_t  *iclog;
4048
4049                 spin_lock(&log->l_icloglock);
4050                 iclog = log->l_iclog;
4051                 do {
4052                         ASSERT(iclog->ic_callback == 0);
4053                         iclog = iclog->ic_next;
4054                 } while (iclog != log->l_iclog);
4055                 spin_unlock(&log->l_icloglock);
4056         }
4057 #endif
4058         /* return non-zero if log IOERROR transition had already happened */
4059         return retval;
4060 }
4061
4062 STATIC int
4063 xlog_iclogs_empty(
4064         struct xlog     *log)
4065 {
4066         xlog_in_core_t  *iclog;
4067
4068         iclog = log->l_iclog;
4069         do {
4070                 /* endianness does not matter here, zero is zero in
4071                  * any language.
4072                  */
4073                 if (iclog->ic_header.h_num_logops)
4074                         return 0;
4075                 iclog = iclog->ic_next;
4076         } while (iclog != log->l_iclog);
4077         return 1;
4078 }
4079
4080 /*
4081  * Verify that an LSN stamped into a piece of metadata is valid. This is
4082  * intended for use in read verifiers on v5 superblocks.
4083  */
4084 bool
4085 xfs_log_check_lsn(
4086         struct xfs_mount        *mp,
4087         xfs_lsn_t               lsn)
4088 {
4089         struct xlog             *log = mp->m_log;
4090         bool                    valid;
4091
4092         /*
4093          * norecovery mode skips mount-time log processing and unconditionally
4094          * resets the in-core LSN. We can't validate in this mode, but
4095          * modifications are not allowed anyways so just return true.
4096          */
4097         if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4098                 return true;
4099
4100         /*
4101          * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4102          * handled by recovery and thus safe to ignore here.
4103          */
4104         if (lsn == NULLCOMMITLSN)
4105                 return true;
4106
4107         valid = xlog_valid_lsn(mp->m_log, lsn);
4108
4109         /* warn the user about what's gone wrong before verifier failure */
4110         if (!valid) {
4111                 spin_lock(&log->l_icloglock);
4112                 xfs_warn(mp,
4113 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4114 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4115                          CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4116                          log->l_curr_cycle, log->l_curr_block);
4117                 spin_unlock(&log->l_icloglock);
4118         }
4119
4120         return valid;
4121 }
4122
4123 bool
4124 xfs_log_in_recovery(
4125         struct xfs_mount        *mp)
4126 {
4127         struct xlog             *log = mp->m_log;
4128
4129         return log->l_flags & XLOG_ACTIVE_RECOVERY;
4130 }