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