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