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