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
  27 kmem_zone_t     *xfs_log_ticket_zone;
  28
  29 /* Local miscellaneous function prototypes */
  30 STATIC int
  31 xlog_commit_record(
  32         struct xlog             *log,
  33         struct xlog_ticket      *ticket,
  34         struct xlog_in_core     **iclog,
  35         xfs_lsn_t               *commitlsnp);
  36
  37 STATIC struct xlog *
  38 xlog_alloc_log(
  39         struct xfs_mount        *mp,
  40         struct xfs_buftarg      *log_target,
  41         xfs_daddr_t             blk_offset,
  42         int                     num_bblks);
  43 STATIC int
  44 xlog_space_left(
  45         struct xlog             *log,
  46         atomic64_t              *head);
  47 STATIC int
  48 xlog_sync(
  49         struct xlog             *log,
  50         struct xlog_in_core     *iclog);
  51 STATIC void
  52 xlog_dealloc_log(
  53         struct xlog             *log);
  54
  55 /* local state machine functions */
  56 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  57 STATIC void
  58 xlog_state_do_callback(
  59         struct xlog             *log,
  60         int                     aborted,
  61         struct xlog_in_core     *iclog);
  62 STATIC int
  63 xlog_state_get_iclog_space(
  64         struct xlog             *log,
  65         int                     len,
  66         struct xlog_in_core     **iclog,
  67         struct xlog_ticket      *ticket,
  68         int                     *continued_write,
  69         int                     *logoffsetp);
  70 STATIC int
  71 xlog_state_release_iclog(
  72         struct xlog             *log,
  73         struct xlog_in_core     *iclog);
  74 STATIC void
  75 xlog_state_switch_iclogs(
  76         struct xlog             *log,
  77         struct xlog_in_core     *iclog,
  78         int                     eventual_size);
  79 STATIC void
  80 xlog_state_want_sync(
  81         struct xlog             *log,
  82         struct xlog_in_core     *iclog);
  83
  84 STATIC void
  85 xlog_grant_push_ail(
  86         struct xlog             *log,
  87         int                     need_bytes);
  88 STATIC void
  89 xlog_regrant_reserve_log_space(
  90         struct xlog             *log,
  91         struct xlog_ticket      *ticket);
  92 STATIC void
  93 xlog_ungrant_log_space(
  94         struct xlog             *log,
  95         struct xlog_ticket      *ticket);
  96
  97 #if defined(DEBUG)
  98 STATIC void
  99 xlog_verify_dest_ptr(
 100         struct xlog             *log,
 101         void                    *ptr);
 102 STATIC void
 103 xlog_verify_grant_tail(
 104         struct xlog *log);
 105 STATIC void
 106 xlog_verify_iclog(
 107         struct xlog             *log,
 108         struct xlog_in_core     *iclog,
 109         int                     count,
 110         bool                    syncing);
 111 STATIC void
 112 xlog_verify_tail_lsn(
 113         struct xlog             *log,
 114         struct xlog_in_core     *iclog,
 115         xfs_lsn_t               tail_lsn);
 116 #else
 117 #define xlog_verify_dest_ptr(a,b)
 118 #define xlog_verify_grant_tail(a)
 119 #define xlog_verify_iclog(a,b,c,d)
 120 #define xlog_verify_tail_lsn(a,b,c)
 121 #endif
 122
 123 STATIC int
 124 xlog_iclogs_empty(
 125         struct xlog             *log);
 126
 127 static void
 128 xlog_grant_sub_space(
 129         struct xlog             *log,
 130         atomic64_t              *head,
 131         int                     bytes)
 132 {
 133         int64_t head_val = atomic64_read(head);
 134         int64_t new, old;
 135
 136         do {
 137                 int     cycle, space;
 138
 139                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 140
 141                 space -= bytes;
 142                 if (space < 0) {
 143                         space += log->l_logsize;
 144                         cycle--;
 145                 }
 146
 147                 old = head_val;
 148                 new = xlog_assign_grant_head_val(cycle, space);
 149                 head_val = atomic64_cmpxchg(head, old, new);
 150         } while (head_val != old);
 151 }
 152
 153 static void
 154 xlog_grant_add_space(
 155         struct xlog             *log,
 156         atomic64_t              *head,
 157         int                     bytes)
 158 {
 159         int64_t head_val = atomic64_read(head);
 160         int64_t new, old;
 161
 162         do {
 163                 int             tmp;
 164                 int             cycle, space;
 165
 166                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 167
 168                 tmp = log->l_logsize - space;
 169                 if (tmp > bytes)
 170                         space += bytes;
 171                 else {
 172                         space = bytes - tmp;
 173                         cycle++;
 174                 }
 175
 176                 old = head_val;
 177                 new = xlog_assign_grant_head_val(cycle, space);
 178                 head_val = atomic64_cmpxchg(head, old, new);
 179         } while (head_val != old);
 180 }
 181
 182 STATIC void
 183 xlog_grant_head_init(
 184         struct xlog_grant_head  *head)
 185 {
 186         xlog_assign_grant_head(&head->grant, 1, 0);
 187         INIT_LIST_HEAD(&head->waiters);
 188         spin_lock_init(&head->lock);
 189 }
 190
 191 STATIC void
 192 xlog_grant_head_wake_all(
 193         struct xlog_grant_head  *head)
 194 {
 195         struct xlog_ticket      *tic;
 196
 197         spin_lock(&head->lock);
 198         list_for_each_entry(tic, &head->waiters, t_queue)
 199                 wake_up_process(tic->t_task);
 200         spin_unlock(&head->lock);
 201 }
 202
 203 static inline int
 204 xlog_ticket_reservation(
 205         struct xlog             *log,
 206         struct xlog_grant_head  *head,
 207         struct xlog_ticket      *tic)
 208 {
 209         if (head == &log->l_write_head) {
 210                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 211                 return tic->t_unit_res;
 212         } else {
 213                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 214                         return tic->t_unit_res * tic->t_cnt;
 215                 else
 216                         return tic->t_unit_res;
 217         }
 218 }
 219
 220 STATIC bool
 221 xlog_grant_head_wake(
 222         struct xlog             *log,
 223         struct xlog_grant_head  *head,
 224         int                     *free_bytes)
 225 {
 226         struct xlog_ticket      *tic;
 227         int                     need_bytes;
 228
 229         list_for_each_entry(tic, &head->waiters, t_queue) {
 230                 need_bytes = xlog_ticket_reservation(log, head, tic);
 231                 if (*free_bytes < need_bytes)
 232                         return false;
 233
 234                 *free_bytes -= need_bytes;
 235                 trace_xfs_log_grant_wake_up(log, tic);
 236                 wake_up_process(tic->t_task);
 237         }
 238
 239         return true;
 240 }
 241
 242 STATIC int
 243 xlog_grant_head_wait(
 244         struct xlog             *log,
 245         struct xlog_grant_head  *head,
 246         struct xlog_ticket      *tic,
 247         int                     need_bytes) __releases(&head->lock)
 248                                             __acquires(&head->lock)
 249 {
 250         list_add_tail(&tic->t_queue, &head->waiters);
 251
 252         do {
 253                 if (XLOG_FORCED_SHUTDOWN(log))
 254                         goto shutdown;
 255                 xlog_grant_push_ail(log, need_bytes);
 256
 257                 __set_current_state(TASK_UNINTERRUPTIBLE);
 258                 spin_unlock(&head->lock);
 259
 260                 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
 261
 262                 trace_xfs_log_grant_sleep(log, tic);
 263                 schedule();
 264                 trace_xfs_log_grant_wake(log, tic);
 265
 266                 spin_lock(&head->lock);
 267                 if (XLOG_FORCED_SHUTDOWN(log))
 268                         goto shutdown;
 269         } while (xlog_space_left(log, &head->grant) < need_bytes);
 270
 271         list_del_init(&tic->t_queue);
 272         return 0;
 273 shutdown:
 274         list_del_init(&tic->t_queue);
 275         return -EIO;
 276 }
 277
 278 /*
 279  * Atomically get the log space required for a log ticket.
 280  *
 281  * Once a ticket gets put onto head->waiters, it will only return after the
 282  * needed reservation is satisfied.
 283  *
 284  * This function is structured so that it has a lock free fast path. This is
 285  * necessary because every new transaction reservation will come through this
 286  * path. Hence any lock will be globally hot if we take it unconditionally on
 287  * every pass.
 288  *
 289  * As tickets are only ever moved on and off head->waiters under head->lock, we
 290  * only need to take that lock if we are going to add the ticket to the queue
 291  * and sleep. We can avoid taking the lock if the ticket was never added to
 292  * head->waiters because the t_queue list head will be empty and we hold the
 293  * only reference to it so it can safely be checked unlocked.
 294  */
 295 STATIC int
 296 xlog_grant_head_check(
 297         struct xlog             *log,
 298         struct xlog_grant_head  *head,
 299         struct xlog_ticket      *tic,
 300         int                     *need_bytes)
 301 {
 302         int                     free_bytes;
 303         int                     error = 0;
 304
 305         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 306
 307         /*
 308          * If there are other waiters on the queue then give them a chance at
 309          * logspace before us.  Wake up the first waiters, if we do not wake
 310          * up all the waiters then go to sleep waiting for more free space,
 311          * otherwise try to get some space for this transaction.
 312          */
 313         *need_bytes = xlog_ticket_reservation(log, head, tic);
 314         free_bytes = xlog_space_left(log, &head->grant);
 315         if (!list_empty_careful(&head->waiters)) {
 316                 spin_lock(&head->lock);
 317                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 318                     free_bytes < *need_bytes) {
 319                         error = xlog_grant_head_wait(log, head, tic,
 320                                                      *need_bytes);
 321                 }
 322                 spin_unlock(&head->lock);
 323         } else if (free_bytes < *need_bytes) {
 324                 spin_lock(&head->lock);
 325                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 326                 spin_unlock(&head->lock);
 327         }
 328
 329         return error;
 330 }
 331
 332 static void
 333 xlog_tic_reset_res(xlog_ticket_t *tic)
 334 {
 335         tic->t_res_num = 0;
 336         tic->t_res_arr_sum = 0;
 337         tic->t_res_num_ophdrs = 0;
 338 }
 339
 340 static void
 341 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 342 {
 343         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 344                 /* add to overflow and start again */
 345                 tic->t_res_o_flow += tic->t_res_arr_sum;
 346                 tic->t_res_num = 0;
 347                 tic->t_res_arr_sum = 0;
 348         }
 349
 350         tic->t_res_arr[tic->t_res_num].r_len = len;
 351         tic->t_res_arr[tic->t_res_num].r_type = type;
 352         tic->t_res_arr_sum += len;
 353         tic->t_res_num++;
 354 }
 355
 356 /*
 357  * Replenish the byte reservation required by moving the grant write head.
 358  */
 359 int
 360 xfs_log_regrant(
 361         struct xfs_mount        *mp,
 362         struct xlog_ticket      *tic)
 363 {
 364         struct xlog             *log = mp->m_log;
 365         int                     need_bytes;
 366         int                     error = 0;
 367
 368         if (XLOG_FORCED_SHUTDOWN(log))
 369                 return -EIO;
 370
 371         XFS_STATS_INC(mp, xs_try_logspace);
 372
 373         /*
 374          * This is a new transaction on the ticket, so we need to change the
 375          * transaction ID so that the next transaction has a different TID in
 376          * the log. Just add one to the existing tid so that we can see chains
 377          * of rolling transactions in the log easily.
 378          */
 379         tic->t_tid++;
 380
 381         xlog_grant_push_ail(log, tic->t_unit_res);
 382
 383         tic->t_curr_res = tic->t_unit_res;
 384         xlog_tic_reset_res(tic);
 385
 386         if (tic->t_cnt > 0)
 387                 return 0;
 388
 389         trace_xfs_log_regrant(log, tic);
 390
 391         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 392                                       &need_bytes);
 393         if (error)
 394                 goto out_error;
 395
 396         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 397         trace_xfs_log_regrant_exit(log, tic);
 398         xlog_verify_grant_tail(log);
 399         return 0;
 400
 401 out_error:
 402         /*
 403          * If we are failing, make sure the ticket doesn't have any current
 404          * reservations.  We don't want to add this back when the ticket/
 405          * transaction gets cancelled.
 406          */
 407         tic->t_curr_res = 0;
 408         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 409         return error;
 410 }
 411
 412 /*
 413  * Reserve log space and return a ticket corresponding the reservation.
 414  *
 415  * Each reservation is going to reserve extra space for a log record header.
 416  * When writes happen to the on-disk log, we don't subtract the length of the
 417  * log record header from any reservation.  By wasting space in each
 418  * reservation, we prevent over allocation problems.
 419  */
 420 int
 421 xfs_log_reserve(
 422         struct xfs_mount        *mp,
 423         int                     unit_bytes,
 424         int                     cnt,
 425         struct xlog_ticket      **ticp,
 426         uint8_t                 client,
 427         bool                    permanent)
 428 {
 429         struct xlog             *log = mp->m_log;
 430         struct xlog_ticket      *tic;
 431         int                     need_bytes;
 432         int                     error = 0;
 433
 434         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 435
 436         if (XLOG_FORCED_SHUTDOWN(log))
 437                 return -EIO;
 438
 439         XFS_STATS_INC(mp, xs_try_logspace);
 440
 441         ASSERT(*ticp == NULL);
 442         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 443                                 KM_SLEEP | KM_MAYFAIL);
 444         if (!tic)
 445                 return -ENOMEM;
 446
 447         *ticp = tic;
 448
 449         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 450                                             : tic->t_unit_res);
 451
 452         trace_xfs_log_reserve(log, tic);
 453
 454         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 455                                       &need_bytes);
 456         if (error)
 457                 goto out_error;
 458
 459         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 460         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 461         trace_xfs_log_reserve_exit(log, tic);
 462         xlog_verify_grant_tail(log);
 463         return 0;
 464
 465 out_error:
 466         /*
 467          * If we are failing, make sure the ticket doesn't have any current
 468          * reservations.  We don't want to add this back when the ticket/
 469          * transaction gets cancelled.
 470          */
 471         tic->t_curr_res = 0;
 472         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 473         return error;
 474 }
 475
 476
 477 /*
 478  * NOTES:
 479  *
 480  *      1. currblock field gets updated at startup and after in-core logs
 481  *              marked as with WANT_SYNC.
 482  */
 483
 484 /*
 485  * This routine is called when a user of a log manager ticket is done with
 486  * the reservation.  If the ticket was ever used, then a commit record for
 487  * the associated transaction is written out as a log operation header with
 488  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 489  * a given ticket.  If the ticket was one with a permanent reservation, then
 490  * a few operations are done differently.  Permanent reservation tickets by
 491  * default don't release the reservation.  They just commit the current
 492  * transaction with the belief that the reservation is still needed.  A flag
 493  * must be passed in before permanent reservations are actually released.
 494  * When these type of tickets are not released, they need to be set into
 495  * the inited state again.  By doing this, a start record will be written
 496  * out when the next write occurs.
 497  */
 498 xfs_lsn_t
 499 xfs_log_done(
 500         struct xfs_mount        *mp,
 501         struct xlog_ticket      *ticket,
 502         struct xlog_in_core     **iclog,
 503         bool                    regrant)
 504 {
 505         struct xlog             *log = mp->m_log;
 506         xfs_lsn_t               lsn = 0;
 507
 508         if (XLOG_FORCED_SHUTDOWN(log) ||
 509             /*
 510              * If nothing was ever written, don't write out commit record.
 511              * If we get an error, just continue and give back the log ticket.
 512              */
 513             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 514              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 515                 lsn = (xfs_lsn_t) -1;
 516                 regrant = false;
 517         }
 518
 519
 520         if (!regrant) {
 521                 trace_xfs_log_done_nonperm(log, ticket);
 522
 523                 /*
 524                  * Release ticket if not permanent reservation or a specific
 525                  * request has been made to release a permanent reservation.
 526                  */
 527                 xlog_ungrant_log_space(log, ticket);
 528         } else {
 529                 trace_xfs_log_done_perm(log, ticket);
 530
 531                 xlog_regrant_reserve_log_space(log, ticket);
 532                 /* If this ticket was a permanent reservation and we aren't
 533                  * trying to release it, reset the inited flags; so next time
 534                  * we write, a start record will be written out.
 535                  */
 536                 ticket->t_flags |= XLOG_TIC_INITED;
 537         }
 538
 539         xfs_log_ticket_put(ticket);
 540         return lsn;
 541 }
 542
 543 /*
 544  * Attaches a new iclog I/O completion callback routine during
 545  * transaction commit.  If the log is in error state, a non-zero
 546  * return code is handed back and the caller is responsible for
 547  * executing the callback at an appropriate time.
 548  */
 549 int
 550 xfs_log_notify(
 551         struct xlog_in_core     *iclog,
 552         xfs_log_callback_t      *cb)
 553 {
 554         int     abortflg;
 555
 556         spin_lock(&iclog->ic_callback_lock);
 557         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 558         if (!abortflg) {
 559                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 560                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 561                 cb->cb_next = NULL;
 562                 *(iclog->ic_callback_tail) = cb;
 563                 iclog->ic_callback_tail = &(cb->cb_next);
 564         }
 565         spin_unlock(&iclog->ic_callback_lock);
 566         return abortflg;
 567 }
 568
 569 int
 570 xfs_log_release_iclog(
 571         struct xfs_mount        *mp,
 572         struct xlog_in_core     *iclog)
 573 {
 574         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 575                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 576                 return -EIO;
 577         }
 578
 579         return 0;
 580 }
 581
 582 /*
 583  * Mount a log filesystem
 584  *
 585  * mp           - ubiquitous xfs mount point structure
 586  * log_target   - buftarg of on-disk log device
 587  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 588  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 589  *
 590  * Return error or zero.
 591  */
 592 int
 593 xfs_log_mount(
 594         xfs_mount_t     *mp,
 595         xfs_buftarg_t   *log_target,
 596         xfs_daddr_t     blk_offset,
 597         int             num_bblks)
 598 {
 599         bool            fatal = xfs_sb_version_hascrc(&mp->m_sb);
 600         int             error = 0;
 601         int             min_logfsbs;
 602
 603         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 604                 xfs_notice(mp, "Mounting V%d Filesystem",
 605                            XFS_SB_VERSION_NUM(&mp->m_sb));
 606         } else {
 607                 xfs_notice(mp,
 608 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
 609                            XFS_SB_VERSION_NUM(&mp->m_sb));
 610                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 611         }
 612
 613         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 614         if (IS_ERR(mp->m_log)) {
 615                 error = PTR_ERR(mp->m_log);
 616                 goto out;
 617         }
 618
 619         /*
 620          * Validate the given log space and drop a critical message via syslog
 621          * if the log size is too small that would lead to some unexpected
 622          * situations in transaction log space reservation stage.
 623          *
 624          * Note: we can't just reject the mount if the validation fails.  This
 625          * would mean that people would have to downgrade their kernel just to
 626          * remedy the situation as there is no way to grow the log (short of
 627          * black magic surgery with xfs_db).
 628          *
 629          * We can, however, reject mounts for CRC format filesystems, as the
 630          * mkfs binary being used to make the filesystem should never create a
 631          * filesystem with a log that is too small.
 632          */
 633         min_logfsbs = xfs_log_calc_minimum_size(mp);
 634
 635         if (mp->m_sb.sb_logblocks < min_logfsbs) {
 636                 xfs_warn(mp,
 637                 "Log size %d blocks too small, minimum size is %d blocks",
 638                          mp->m_sb.sb_logblocks, min_logfsbs);
 639                 error = -EINVAL;
 640         } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
 641                 xfs_warn(mp,
 642                 "Log size %d blocks too large, maximum size is %lld blocks",
 643                          mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
 644                 error = -EINVAL;
 645         } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
 646                 xfs_warn(mp,
 647                 "log size %lld bytes too large, maximum size is %lld bytes",
 648                          XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
 649                          XFS_MAX_LOG_BYTES);
 650                 error = -EINVAL;
 651         } else if (mp->m_sb.sb_logsunit > 1 &&
 652                    mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
 653                 xfs_warn(mp,
 654                 "log stripe unit %u bytes must be a multiple of block size",
 655                          mp->m_sb.sb_logsunit);
 656                 error = -EINVAL;
 657                 fatal = true;
 658         }
 659         if (error) {
 660                 /*
 661                  * Log check errors are always fatal on v5; or whenever bad
 662                  * metadata leads to a crash.
 663                  */
 664                 if (fatal) {
 665                         xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 666                         ASSERT(0);
 667                         goto out_free_log;
 668                 }
 669                 xfs_crit(mp, "Log size out of supported range.");
 670                 xfs_crit(mp,
 671 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
 672         }
 673
 674         /*
 675          * Initialize the AIL now we have a log.
 676          */
 677         error = xfs_trans_ail_init(mp);
 678         if (error) {
 679                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 680                 goto out_free_log;
 681         }
 682         mp->m_log->l_ailp = mp->m_ail;
 683
 684         /*
 685          * skip log recovery on a norecovery mount.  pretend it all
 686          * just worked.
 687          */
 688         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 689                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 690
 691                 if (readonly)
 692                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 693
 694                 error = xlog_recover(mp->m_log);
 695
 696                 if (readonly)
 697                         mp->m_flags |= XFS_MOUNT_RDONLY;
 698                 if (error) {
 699                         xfs_warn(mp, "log mount/recovery failed: error %d",
 700                                 error);
 701                         xlog_recover_cancel(mp->m_log);
 702                         goto out_destroy_ail;
 703                 }
 704         }
 705
 706         error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 707                                "log");
 708         if (error)
 709                 goto out_destroy_ail;
 710
 711         /* Normal transactions can now occur */
 712         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 713
 714         /*
 715          * Now the log has been fully initialised and we know were our
 716          * space grant counters are, we can initialise the permanent ticket
 717          * needed for delayed logging to work.
 718          */
 719         xlog_cil_init_post_recovery(mp->m_log);
 720
 721         return 0;
 722
 723 out_destroy_ail:
 724         xfs_trans_ail_destroy(mp);
 725 out_free_log:
 726         xlog_dealloc_log(mp->m_log);
 727 out:
 728         return error;
 729 }
 730
 731 /*
 732  * Finish the recovery of the file system.  This is separate from the
 733  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 734  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 735  * here.
 736  *
 737  * If we finish recovery successfully, start the background log work. If we are
 738  * not doing recovery, then we have a RO filesystem and we don't need to start
 739  * it.
 740  */
 741 int
 742 xfs_log_mount_finish(
 743         struct xfs_mount        *mp)
 744 {
 745         int     error = 0;
 746         bool    readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 747         bool    recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
 748
 749         if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
 750                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 751                 return 0;
 752         } else if (readonly) {
 753                 /* Allow unlinked processing to proceed */
 754                 mp->m_flags &= ~XFS_MOUNT_RDONLY;
 755         }
 756
 757         /*
 758          * During the second phase of log recovery, we need iget and
 759          * iput to behave like they do for an active filesystem.
 760          * xfs_fs_drop_inode needs to be able to prevent the deletion
 761          * of inodes before we're done replaying log items on those
 762          * inodes.  Turn it off immediately after recovery finishes
 763          * so that we don't leak the quota inodes if subsequent mount
 764          * activities fail.
 765          *
 766          * We let all inodes involved in redo item processing end up on
 767          * the LRU instead of being evicted immediately so that if we do
 768          * something to an unlinked inode, the irele won't cause
 769          * premature truncation and freeing of the inode, which results
 770          * in log recovery failure.  We have to evict the unreferenced
 771          * lru inodes after clearing SB_ACTIVE because we don't
 772          * otherwise clean up the lru if there's a subsequent failure in
 773          * xfs_mountfs, which leads to us leaking the inodes if nothing
 774          * else (e.g. quotacheck) references the inodes before the
 775          * mount failure occurs.
 776          */
 777         mp->m_super->s_flags |= SB_ACTIVE;
 778         error = xlog_recover_finish(mp->m_log);
 779         if (!error)
 780                 xfs_log_work_queue(mp);
 781         mp->m_super->s_flags &= ~SB_ACTIVE;
 782         evict_inodes(mp->m_super);
 783
 784         /*
 785          * Drain the buffer LRU after log recovery. This is required for v4
 786          * filesystems to avoid leaving around buffers with NULL verifier ops,
 787          * but we do it unconditionally to make sure we're always in a clean
 788          * cache state after mount.
 789          *
 790          * Don't push in the error case because the AIL may have pending intents
 791          * that aren't removed until recovery is cancelled.
 792          */
 793         if (!error && recovered) {
 794                 xfs_log_force(mp, XFS_LOG_SYNC);
 795                 xfs_ail_push_all_sync(mp->m_ail);
 796         }
 797         xfs_wait_buftarg(mp->m_ddev_targp);
 798
 799         if (readonly)
 800                 mp->m_flags |= XFS_MOUNT_RDONLY;
 801
 802         return error;
 803 }
 804
 805 /*
 806  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
 807  * the log.
 808  */
 809 int
 810 xfs_log_mount_cancel(
 811         struct xfs_mount        *mp)
 812 {
 813         int                     error;
 814
 815         error = xlog_recover_cancel(mp->m_log);
 816         xfs_log_unmount(mp);
 817
 818         return error;
 819 }
 820
 821 /*
 822  * Final log writes as part of unmount.
 823  *
 824  * Mark the filesystem clean as unmount happens.  Note that during relocation
 825  * this routine needs to be executed as part of source-bag while the
 826  * deallocation must not be done until source-end.
 827  */
 828
 829 /*
 830  * Unmount record used to have a string "Unmount filesystem--" in the
 831  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 832  * We just write the magic number now since that particular field isn't
 833  * currently architecture converted and "Unmount" is a bit foo.
 834  * As far as I know, there weren't any dependencies on the old behaviour.
 835  */
 836
 837 static int
 838 xfs_log_unmount_write(xfs_mount_t *mp)
 839 {
 840         struct xlog      *log = mp->m_log;
 841         xlog_in_core_t   *iclog;
 842 #ifdef DEBUG
 843         xlog_in_core_t   *first_iclog;
 844 #endif
 845         xlog_ticket_t   *tic = NULL;
 846         xfs_lsn_t        lsn;
 847         int              error;
 848
 849         /*
 850          * Don't write out unmount record on norecovery mounts or ro devices.
 851          * Or, if we are doing a forced umount (typically because of IO errors).
 852          */
 853         if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
 854             xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
 855                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 856                 return 0;
 857         }
 858
 859         error = xfs_log_force(mp, XFS_LOG_SYNC);
 860         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 861
 862 #ifdef DEBUG
 863         first_iclog = iclog = log->l_iclog;
 864         do {
 865                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 866                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 867                         ASSERT(iclog->ic_offset == 0);
 868                 }
 869                 iclog = iclog->ic_next;
 870         } while (iclog != first_iclog);
 871 #endif
 872         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 873                 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
 874                 if (!error) {
 875                         /* the data section must be 32 bit size aligned */
 876                         struct {
 877                             uint16_t magic;
 878                             uint16_t pad1;
 879                             uint32_t pad2; /* may as well make it 64 bits */
 880                         } magic = {
 881                                 .magic = XLOG_UNMOUNT_TYPE,
 882                         };
 883                         struct xfs_log_iovec reg = {
 884                                 .i_addr = &magic,
 885                                 .i_len = sizeof(magic),
 886                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 887                         };
 888                         struct xfs_log_vec vec = {
 889                                 .lv_niovecs = 1,
 890                                 .lv_iovecp = &reg,
 891                         };
 892
 893                         /* remove inited flag, and account for space used */
 894                         tic->t_flags = 0;
 895                         tic->t_curr_res -= sizeof(magic);
 896                         error = xlog_write(log, &vec, tic, &lsn,
 897                                            NULL, XLOG_UNMOUNT_TRANS);
 898                         /*
 899                          * At this point, we're umounting anyway,
 900                          * so there's no point in transitioning log state
 901                          * to IOERROR. Just continue...
 902                          */
 903                 }
 904
 905                 if (error)
 906                         xfs_alert(mp, "%s: unmount record failed", __func__);
 907
 908
 909                 spin_lock(&log->l_icloglock);
 910                 iclog = log->l_iclog;
 911                 atomic_inc(&iclog->ic_refcnt);
 912                 xlog_state_want_sync(log, iclog);
 913                 spin_unlock(&log->l_icloglock);
 914                 error = xlog_state_release_iclog(log, iclog);
 915
 916                 spin_lock(&log->l_icloglock);
 917                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 918                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 919                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 920                                 xlog_wait(&iclog->ic_force_wait,
 921                                                         &log->l_icloglock);
 922                         } else {
 923                                 spin_unlock(&log->l_icloglock);
 924                         }
 925                 } else {
 926                         spin_unlock(&log->l_icloglock);
 927                 }
 928                 if (tic) {
 929                         trace_xfs_log_umount_write(log, tic);
 930                         xlog_ungrant_log_space(log, tic);
 931                         xfs_log_ticket_put(tic);
 932                 }
 933         } else {
 934                 /*
 935                  * We're already in forced_shutdown mode, couldn't
 936                  * even attempt to write out the unmount transaction.
 937                  *
 938                  * Go through the motions of sync'ing and releasing
 939                  * the iclog, even though no I/O will actually happen,
 940                  * we need to wait for other log I/Os that may already
 941                  * be in progress.  Do this as a separate section of
 942                  * code so we'll know if we ever get stuck here that
 943                  * we're in this odd situation of trying to unmount
 944                  * a file system that went into forced_shutdown as
 945                  * the result of an unmount..
 946                  */
 947                 spin_lock(&log->l_icloglock);
 948                 iclog = log->l_iclog;
 949                 atomic_inc(&iclog->ic_refcnt);
 950
 951                 xlog_state_want_sync(log, iclog);
 952                 spin_unlock(&log->l_icloglock);
 953                 error =  xlog_state_release_iclog(log, iclog);
 954
 955                 spin_lock(&log->l_icloglock);
 956
 957                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 958                         || iclog->ic_state == XLOG_STATE_DIRTY
 959                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 960
 961                                 xlog_wait(&iclog->ic_force_wait,
 962                                                         &log->l_icloglock);
 963                 } else {
 964                         spin_unlock(&log->l_icloglock);
 965                 }
 966         }
 967
 968         return error;
 969 }       /* xfs_log_unmount_write */
 970
 971 /*
 972  * Empty the log for unmount/freeze.
 973  *
 974  * To do this, we first need to shut down the background log work so it is not
 975  * trying to cover the log as we clean up. We then need to unpin all objects in
 976  * the log so we can then flush them out. Once they have completed their IO and
 977  * run the callbacks removing themselves from the AIL, we can write the unmount
 978  * record.
 979  */
 980 void
 981 xfs_log_quiesce(
 982         struct xfs_mount        *mp)
 983 {
 984         cancel_delayed_work_sync(&mp->m_log->l_work);
 985         xfs_log_force(mp, XFS_LOG_SYNC);
 986
 987         /*
 988          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
 989          * will push it, xfs_wait_buftarg() will not wait for it. Further,
 990          * xfs_buf_iowait() cannot be used because it was pushed with the
 991          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
 992          * the IO to complete.
 993          */
 994         xfs_ail_push_all_sync(mp->m_ail);
 995         xfs_wait_buftarg(mp->m_ddev_targp);
 996         xfs_buf_lock(mp->m_sb_bp);
 997         xfs_buf_unlock(mp->m_sb_bp);
 998
 999         xfs_log_unmount_write(mp);
1000 }
1001
1002 /*
1003  * Shut down and release the AIL and Log.
1004  *
1005  * During unmount, we need to ensure we flush all the dirty metadata objects
1006  * from the AIL so that the log is empty before we write the unmount record to
1007  * the log. Once this is done, we can tear down the AIL and the log.
1008  */
1009 void
1010 xfs_log_unmount(
1011         struct xfs_mount        *mp)
1012 {
1013         xfs_log_quiesce(mp);
1014
1015         xfs_trans_ail_destroy(mp);
1016
1017         xfs_sysfs_del(&mp->m_log->l_kobj);
1018
1019         xlog_dealloc_log(mp->m_log);
1020 }
1021
1022 void
1023 xfs_log_item_init(
1024         struct xfs_mount        *mp,
1025         struct xfs_log_item     *item,
1026         int                     type,
1027         const struct xfs_item_ops *ops)
1028 {
1029         item->li_mountp = mp;
1030         item->li_ailp = mp->m_ail;
1031         item->li_type = type;
1032         item->li_ops = ops;
1033         item->li_lv = NULL;
1034
1035         INIT_LIST_HEAD(&item->li_ail);
1036         INIT_LIST_HEAD(&item->li_cil);
1037         INIT_LIST_HEAD(&item->li_bio_list);
1038         INIT_LIST_HEAD(&item->li_trans);
1039 }
1040
1041 /*
1042  * Wake up processes waiting for log space after we have moved the log tail.
1043  */
1044 void
1045 xfs_log_space_wake(
1046         struct xfs_mount        *mp)
1047 {
1048         struct xlog             *log = mp->m_log;
1049         int                     free_bytes;
1050
1051         if (XLOG_FORCED_SHUTDOWN(log))
1052                 return;
1053
1054         if (!list_empty_careful(&log->l_write_head.waiters)) {
1055                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1056
1057                 spin_lock(&log->l_write_head.lock);
1058                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1059                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1060                 spin_unlock(&log->l_write_head.lock);
1061         }
1062
1063         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1064                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1065
1066                 spin_lock(&log->l_reserve_head.lock);
1067                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1068                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1069                 spin_unlock(&log->l_reserve_head.lock);
1070         }
1071 }
1072
1073 /*
1074  * Determine if we have a transaction that has gone to disk that needs to be
1075  * covered. To begin the transition to the idle state firstly the log needs to
1076  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1077  * we start attempting to cover the log.
1078  *
1079  * Only if we are then in a state where covering is needed, the caller is
1080  * informed that dummy transactions are required to move the log into the idle
1081  * state.
1082  *
1083  * If there are any items in the AIl or CIL, then we do not want to attempt to
1084  * cover the log as we may be in a situation where there isn't log space
1085  * available to run a dummy transaction and this can lead to deadlocks when the
1086  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1087  * there's no point in running a dummy transaction at this point because we
1088  * can't start trying to idle the log until both the CIL and AIL are empty.
1089  */
1090 static int
1091 xfs_log_need_covered(xfs_mount_t *mp)
1092 {
1093         struct xlog     *log = mp->m_log;
1094         int             needed = 0;
1095
1096         if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1097                 return 0;
1098
1099         if (!xlog_cil_empty(log))
1100                 return 0;
1101
1102         spin_lock(&log->l_icloglock);
1103         switch (log->l_covered_state) {
1104         case XLOG_STATE_COVER_DONE:
1105         case XLOG_STATE_COVER_DONE2:
1106         case XLOG_STATE_COVER_IDLE:
1107                 break;
1108         case XLOG_STATE_COVER_NEED:
1109         case XLOG_STATE_COVER_NEED2:
1110                 if (xfs_ail_min_lsn(log->l_ailp))
1111                         break;
1112                 if (!xlog_iclogs_empty(log))
1113                         break;
1114
1115                 needed = 1;
1116                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1117                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1118                 else
1119                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1120                 break;
1121         default:
1122                 needed = 1;
1123                 break;
1124         }
1125         spin_unlock(&log->l_icloglock);
1126         return needed;
1127 }
1128
1129 /*
1130  * We may be holding the log iclog lock upon entering this routine.
1131  */
1132 xfs_lsn_t
1133 xlog_assign_tail_lsn_locked(
1134         struct xfs_mount        *mp)
1135 {
1136         struct xlog             *log = mp->m_log;
1137         struct xfs_log_item     *lip;
1138         xfs_lsn_t               tail_lsn;
1139
1140         assert_spin_locked(&mp->m_ail->ail_lock);
1141
1142         /*
1143          * To make sure we always have a valid LSN for the log tail we keep
1144          * track of the last LSN which was committed in log->l_last_sync_lsn,
1145          * and use that when the AIL was empty.
1146          */
1147         lip = xfs_ail_min(mp->m_ail);
1148         if (lip)
1149                 tail_lsn = lip->li_lsn;
1150         else
1151                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1152         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1153         atomic64_set(&log->l_tail_lsn, tail_lsn);
1154         return tail_lsn;
1155 }
1156
1157 xfs_lsn_t
1158 xlog_assign_tail_lsn(
1159         struct xfs_mount        *mp)
1160 {
1161         xfs_lsn_t               tail_lsn;
1162
1163         spin_lock(&mp->m_ail->ail_lock);
1164         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1165         spin_unlock(&mp->m_ail->ail_lock);
1166
1167         return tail_lsn;
1168 }
1169
1170 /*
1171  * Return the space in the log between the tail and the head.  The head
1172  * is passed in the cycle/bytes formal parms.  In the special case where
1173  * the reserve head has wrapped passed the tail, this calculation is no
1174  * longer valid.  In this case, just return 0 which means there is no space
1175  * in the log.  This works for all places where this function is called
1176  * with the reserve head.  Of course, if the write head were to ever
1177  * wrap the tail, we should blow up.  Rather than catch this case here,
1178  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1179  *
1180  * This code also handles the case where the reservation head is behind
1181  * the tail.  The details of this case are described below, but the end
1182  * result is that we return the size of the log as the amount of space left.
1183  */
1184 STATIC int
1185 xlog_space_left(
1186         struct xlog     *log,
1187         atomic64_t      *head)
1188 {
1189         int             free_bytes;
1190         int             tail_bytes;
1191         int             tail_cycle;
1192         int             head_cycle;
1193         int             head_bytes;
1194
1195         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1196         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1197         tail_bytes = BBTOB(tail_bytes);
1198         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1199                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1200         else if (tail_cycle + 1 < head_cycle)
1201                 return 0;
1202         else if (tail_cycle < head_cycle) {
1203                 ASSERT(tail_cycle == (head_cycle - 1));
1204                 free_bytes = tail_bytes - head_bytes;
1205         } else {
1206                 /*
1207                  * The reservation head is behind the tail.
1208                  * In this case we just want to return the size of the
1209                  * log as the amount of space left.
1210                  */
1211                 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1212                 xfs_alert(log->l_mp,
1213                           "  tail_cycle = %d, tail_bytes = %d",
1214                           tail_cycle, tail_bytes);
1215                 xfs_alert(log->l_mp,
1216                           "  GH   cycle = %d, GH   bytes = %d",
1217                           head_cycle, head_bytes);
1218                 ASSERT(0);
1219                 free_bytes = log->l_logsize;
1220         }
1221         return free_bytes;
1222 }
1223
1224
1225 /*
1226  * Log function which is called when an io completes.
1227  *
1228  * The log manager needs its own routine, in order to control what
1229  * happens with the buffer after the write completes.
1230  */
1231 static void
1232 xlog_iodone(xfs_buf_t *bp)
1233 {
1234         struct xlog_in_core     *iclog = bp->b_log_item;
1235         struct xlog             *l = iclog->ic_log;
1236         int                     aborted = 0;
1237
1238         /*
1239          * Race to shutdown the filesystem if we see an error or the iclog is in
1240          * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
1241          * CRC errors into log recovery.
1242          */
1243         if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR) ||
1244             iclog->ic_state & XLOG_STATE_IOABORT) {
1245                 if (iclog->ic_state & XLOG_STATE_IOABORT)
1246                         iclog->ic_state &= ~XLOG_STATE_IOABORT;
1247
1248                 xfs_buf_ioerror_alert(bp, __func__);
1249                 xfs_buf_stale(bp);
1250                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1251                 /*
1252                  * This flag will be propagated to the trans-committed
1253                  * callback routines to let them know that the log-commit
1254                  * didn't succeed.
1255                  */
1256                 aborted = XFS_LI_ABORTED;
1257         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1258                 aborted = XFS_LI_ABORTED;
1259         }
1260
1261         /* log I/O is always issued ASYNC */
1262         ASSERT(bp->b_flags & XBF_ASYNC);
1263         xlog_state_done_syncing(iclog, aborted);
1264
1265         /*
1266          * drop the buffer lock now that we are done. Nothing references
1267          * the buffer after this, so an unmount waiting on this lock can now
1268          * tear it down safely. As such, it is unsafe to reference the buffer
1269          * (bp) after the unlock as we could race with it being freed.
1270          */
1271         xfs_buf_unlock(bp);
1272 }
1273
1274 /*
1275  * Return size of each in-core log record buffer.
1276  *
1277  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1278  *
1279  * If the filesystem blocksize is too large, we may need to choose a
1280  * larger size since the directory code currently logs entire blocks.
1281  */
1282
1283 STATIC void
1284 xlog_get_iclog_buffer_size(
1285         struct xfs_mount        *mp,
1286         struct xlog             *log)
1287 {
1288         int size;
1289         int xhdrs;
1290
1291         if (mp->m_logbufs <= 0)
1292                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1293         else
1294                 log->l_iclog_bufs = mp->m_logbufs;
1295
1296         /*
1297          * Buffer size passed in from mount system call.
1298          */
1299         if (mp->m_logbsize > 0) {
1300                 size = log->l_iclog_size = mp->m_logbsize;
1301                 log->l_iclog_size_log = 0;
1302                 while (size != 1) {
1303                         log->l_iclog_size_log++;
1304                         size >>= 1;
1305                 }
1306
1307                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1308                         /* # headers = size / 32k
1309                          * one header holds cycles from 32k of data
1310                          */
1311
1312                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1313                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1314                                 xhdrs++;
1315                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1316                         log->l_iclog_heads = xhdrs;
1317                 } else {
1318                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1319                         log->l_iclog_hsize = BBSIZE;
1320                         log->l_iclog_heads = 1;
1321                 }
1322                 goto done;
1323         }
1324
1325         /* All machines use 32kB buffers by default. */
1326         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1327         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1328
1329         /* the default log size is 16k or 32k which is one header sector */
1330         log->l_iclog_hsize = BBSIZE;
1331         log->l_iclog_heads = 1;
1332
1333 done:
1334         /* are we being asked to make the sizes selected above visible? */
1335         if (mp->m_logbufs == 0)
1336                 mp->m_logbufs = log->l_iclog_bufs;
1337         if (mp->m_logbsize == 0)
1338                 mp->m_logbsize = log->l_iclog_size;
1339 }       /* xlog_get_iclog_buffer_size */
1340
1341
1342 void
1343 xfs_log_work_queue(
1344         struct xfs_mount        *mp)
1345 {
1346         queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1347                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1348 }
1349
1350 /*
1351  * Every sync period we need to unpin all items in the AIL and push them to
1352  * disk. If there is nothing dirty, then we might need to cover the log to
1353  * indicate that the filesystem is idle.
1354  */
1355 static void
1356 xfs_log_worker(
1357         struct work_struct      *work)
1358 {
1359         struct xlog             *log = container_of(to_delayed_work(work),
1360                                                 struct xlog, l_work);
1361         struct xfs_mount        *mp = log->l_mp;
1362
1363         /* dgc: errors ignored - not fatal and nowhere to report them */
1364         if (xfs_log_need_covered(mp)) {
1365                 /*
1366                  * Dump a transaction into the log that contains no real change.
1367                  * This is needed to stamp the current tail LSN into the log
1368                  * during the covering operation.
1369                  *
1370                  * We cannot use an inode here for this - that will push dirty
1371                  * state back up into the VFS and then periodic inode flushing
1372                  * will prevent log covering from making progress. Hence we
1373                  * synchronously log the superblock instead to ensure the
1374                  * superblock is immediately unpinned and can be written back.
1375                  */
1376                 xfs_sync_sb(mp, true);
1377         } else
1378                 xfs_log_force(mp, 0);
1379
1380         /* start pushing all the metadata that is currently dirty */
1381         xfs_ail_push_all(mp->m_ail);
1382
1383         /* queue us up again */
1384         xfs_log_work_queue(mp);
1385 }
1386
1387 /*
1388  * This routine initializes some of the log structure for a given mount point.
1389  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1390  * some other stuff may be filled in too.
1391  */
1392 STATIC struct xlog *
1393 xlog_alloc_log(
1394         struct xfs_mount        *mp,
1395         struct xfs_buftarg      *log_target,
1396         xfs_daddr_t             blk_offset,
1397         int                     num_bblks)
1398 {
1399         struct xlog             *log;
1400         xlog_rec_header_t       *head;
1401         xlog_in_core_t          **iclogp;
1402         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1403         xfs_buf_t               *bp;
1404         int                     i;
1405         int                     error = -ENOMEM;
1406         uint                    log2_size = 0;
1407
1408         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1409         if (!log) {
1410                 xfs_warn(mp, "Log allocation failed: No memory!");
1411                 goto out;
1412         }
1413
1414         log->l_mp          = mp;
1415         log->l_targ        = log_target;
1416         log->l_logsize     = BBTOB(num_bblks);
1417         log->l_logBBstart  = blk_offset;
1418         log->l_logBBsize   = num_bblks;
1419         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1420         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1421         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1422
1423         log->l_prev_block  = -1;
1424         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1425         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1426         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1427         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1428
1429         xlog_grant_head_init(&log->l_reserve_head);
1430         xlog_grant_head_init(&log->l_write_head);
1431
1432         error = -EFSCORRUPTED;
1433         if (xfs_sb_version_hassector(&mp->m_sb)) {
1434                 log2_size = mp->m_sb.sb_logsectlog;
1435                 if (log2_size < BBSHIFT) {
1436                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1437                                 log2_size, BBSHIFT);
1438                         goto out_free_log;
1439                 }
1440
1441                 log2_size -= BBSHIFT;
1442                 if (log2_size > mp->m_sectbb_log) {
1443                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1444                                 log2_size, mp->m_sectbb_log);
1445                         goto out_free_log;
1446                 }
1447
1448                 /* for larger sector sizes, must have v2 or external log */
1449                 if (log2_size && log->l_logBBstart > 0 &&
1450                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1451                         xfs_warn(mp,
1452                 "log sector size (0x%x) invalid for configuration.",
1453                                 log2_size);
1454                         goto out_free_log;
1455                 }
1456         }
1457         log->l_sectBBsize = 1 << log2_size;
1458
1459         xlog_get_iclog_buffer_size(mp, log);
1460
1461         /*
1462          * Use a NULL block for the extra log buffer used during splits so that
1463          * it will trigger errors if we ever try to do IO on it without first
1464          * having set it up properly.
1465          */
1466         error = -ENOMEM;
1467         bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1468                            BTOBB(log->l_iclog_size), XBF_NO_IOACCT);
1469         if (!bp)
1470                 goto out_free_log;
1471
1472         /*
1473          * The iclogbuf buffer locks are held over IO but we are not going to do
1474          * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1475          * when appropriately.
1476          */
1477         ASSERT(xfs_buf_islocked(bp));
1478         xfs_buf_unlock(bp);
1479
1480         /* use high priority wq for log I/O completion */
1481         bp->b_ioend_wq = mp->m_log_workqueue;
1482         bp->b_iodone = xlog_iodone;
1483         log->l_xbuf = bp;
1484
1485         spin_lock_init(&log->l_icloglock);
1486         init_waitqueue_head(&log->l_flush_wait);
1487
1488         iclogp = &log->l_iclog;
1489         /*
1490          * The amount of memory to allocate for the iclog structure is
1491          * rather funky due to the way the structure is defined.  It is
1492          * done this way so that we can use different sizes for machines
1493          * with different amounts of memory.  See the definition of
1494          * xlog_in_core_t in xfs_log_priv.h for details.
1495          */
1496         ASSERT(log->l_iclog_size >= 4096);
1497         for (i=0; i < log->l_iclog_bufs; i++) {
1498                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1499                 if (!*iclogp)
1500                         goto out_free_iclog;
1501
1502                 iclog = *iclogp;
1503                 iclog->ic_prev = prev_iclog;
1504                 prev_iclog = iclog;
1505
1506                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1507                                           BTOBB(log->l_iclog_size),
1508                                           XBF_NO_IOACCT);
1509                 if (!bp)
1510                         goto out_free_iclog;
1511
1512                 ASSERT(xfs_buf_islocked(bp));
1513                 xfs_buf_unlock(bp);
1514
1515                 /* use high priority wq for log I/O completion */
1516                 bp->b_ioend_wq = mp->m_log_workqueue;
1517                 bp->b_iodone = xlog_iodone;
1518                 iclog->ic_bp = bp;
1519                 iclog->ic_data = bp->b_addr;
1520 #ifdef DEBUG
1521                 log->l_iclog_bak[i] = &iclog->ic_header;
1522 #endif
1523                 head = &iclog->ic_header;
1524                 memset(head, 0, sizeof(xlog_rec_header_t));
1525                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1526                 head->h_version = cpu_to_be32(
1527                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1528                 head->h_size = cpu_to_be32(log->l_iclog_size);
1529                 /* new fields */
1530                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1531                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1532
1533                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1534                 iclog->ic_state = XLOG_STATE_ACTIVE;
1535                 iclog->ic_log = log;
1536                 atomic_set(&iclog->ic_refcnt, 0);
1537                 spin_lock_init(&iclog->ic_callback_lock);
1538                 iclog->ic_callback_tail = &(iclog->ic_callback);
1539                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1540
1541                 init_waitqueue_head(&iclog->ic_force_wait);
1542                 init_waitqueue_head(&iclog->ic_write_wait);
1543
1544                 iclogp = &iclog->ic_next;
1545         }
1546         *iclogp = log->l_iclog;                 /* complete ring */
1547         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1548
1549         error = xlog_cil_init(log);
1550         if (error)
1551                 goto out_free_iclog;
1552         return log;
1553
1554 out_free_iclog:
1555         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1556                 prev_iclog = iclog->ic_next;
1557                 if (iclog->ic_bp)
1558                         xfs_buf_free(iclog->ic_bp);
1559                 kmem_free(iclog);
1560         }
1561         spinlock_destroy(&log->l_icloglock);
1562         xfs_buf_free(log->l_xbuf);
1563 out_free_log:
1564         kmem_free(log);
1565 out:
1566         return ERR_PTR(error);
1567 }       /* xlog_alloc_log */
1568
1569
1570 /*
1571  * Write out the commit record of a transaction associated with the given
1572  * ticket.  Return the lsn of the commit record.
1573  */
1574 STATIC int
1575 xlog_commit_record(
1576         struct xlog             *log,
1577         struct xlog_ticket      *ticket,
1578         struct xlog_in_core     **iclog,
1579         xfs_lsn_t               *commitlsnp)
1580 {
1581         struct xfs_mount *mp = log->l_mp;
1582         int     error;
1583         struct xfs_log_iovec reg = {
1584                 .i_addr = NULL,
1585                 .i_len = 0,
1586                 .i_type = XLOG_REG_TYPE_COMMIT,
1587         };
1588         struct xfs_log_vec vec = {
1589                 .lv_niovecs = 1,
1590                 .lv_iovecp = &reg,
1591         };
1592
1593         ASSERT_ALWAYS(iclog);
1594         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1595                                         XLOG_COMMIT_TRANS);
1596         if (error)
1597                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1598         return error;
1599 }
1600
1601 /*
1602  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1603  * log space.  This code pushes on the lsn which would supposedly free up
1604  * the 25% which we want to leave free.  We may need to adopt a policy which
1605  * pushes on an lsn which is further along in the log once we reach the high
1606  * water mark.  In this manner, we would be creating a low water mark.
1607  */
1608 STATIC void
1609 xlog_grant_push_ail(
1610         struct xlog     *log,
1611         int             need_bytes)
1612 {
1613         xfs_lsn_t       threshold_lsn = 0;
1614         xfs_lsn_t       last_sync_lsn;
1615         int             free_blocks;
1616         int             free_bytes;
1617         int             threshold_block;
1618         int             threshold_cycle;
1619         int             free_threshold;
1620
1621         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1622
1623         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1624         free_blocks = BTOBBT(free_bytes);
1625
1626         /*
1627          * Set the threshold for the minimum number of free blocks in the
1628          * log to the maximum of what the caller needs, one quarter of the
1629          * log, and 256 blocks.
1630          */
1631         free_threshold = BTOBB(need_bytes);
1632         free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1633         free_threshold = max(free_threshold, 256);
1634         if (free_blocks >= free_threshold)
1635                 return;
1636
1637         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1638                                                 &threshold_block);
1639         threshold_block += free_threshold;
1640         if (threshold_block >= log->l_logBBsize) {
1641                 threshold_block -= log->l_logBBsize;
1642                 threshold_cycle += 1;
1643         }
1644         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1645                                         threshold_block);
1646         /*
1647          * Don't pass in an lsn greater than the lsn of the last
1648          * log record known to be on disk. Use a snapshot of the last sync lsn
1649          * so that it doesn't change between the compare and the set.
1650          */
1651         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1652         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1653                 threshold_lsn = last_sync_lsn;
1654
1655         /*
1656          * Get the transaction layer to kick the dirty buffers out to
1657          * disk asynchronously. No point in trying to do this if
1658          * the filesystem is shutting down.
1659          */
1660         if (!XLOG_FORCED_SHUTDOWN(log))
1661                 xfs_ail_push(log->l_ailp, threshold_lsn);
1662 }
1663
1664 /*
1665  * Stamp cycle number in every block
1666  */
1667 STATIC void
1668 xlog_pack_data(
1669         struct xlog             *log,
1670         struct xlog_in_core     *iclog,
1671         int                     roundoff)
1672 {
1673         int                     i, j, k;
1674         int                     size = iclog->ic_offset + roundoff;
1675         __be32                  cycle_lsn;
1676         char                    *dp;
1677
1678         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1679
1680         dp = iclog->ic_datap;
1681         for (i = 0; i < BTOBB(size); i++) {
1682                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1683                         break;
1684                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1685                 *(__be32 *)dp = cycle_lsn;
1686                 dp += BBSIZE;
1687         }
1688
1689         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1690                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1691
1692                 for ( ; i < BTOBB(size); i++) {
1693                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1694                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1695                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1696                         *(__be32 *)dp = cycle_lsn;
1697                         dp += BBSIZE;
1698                 }
1699
1700                 for (i = 1; i < log->l_iclog_heads; i++)
1701                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1702         }
1703 }
1704
1705 /*
1706  * Calculate the checksum for a log buffer.
1707  *
1708  * This is a little more complicated than it should be because the various
1709  * headers and the actual data are non-contiguous.
1710  */
1711 __le32
1712 xlog_cksum(
1713         struct xlog             *log,
1714         struct xlog_rec_header  *rhead,
1715         char                    *dp,
1716         int                     size)
1717 {
1718         uint32_t                crc;
1719
1720         /* first generate the crc for the record header ... */
1721         crc = xfs_start_cksum_update((char *)rhead,
1722                               sizeof(struct xlog_rec_header),
1723                               offsetof(struct xlog_rec_header, h_crc));
1724
1725         /* ... then for additional cycle data for v2 logs ... */
1726         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1727                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1728                 int             i;
1729                 int             xheads;
1730
1731                 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1732                 if (size % XLOG_HEADER_CYCLE_SIZE)
1733                         xheads++;
1734
1735                 for (i = 1; i < xheads; i++) {
1736                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1737                                      sizeof(struct xlog_rec_ext_header));
1738                 }
1739         }
1740
1741         /* ... and finally for the payload */
1742         crc = crc32c(crc, dp, size);
1743
1744         return xfs_end_cksum(crc);
1745 }
1746
1747 /*
1748  * The bdstrat callback function for log bufs. This gives us a central
1749  * place to trap bufs in case we get hit by a log I/O error and need to
1750  * shutdown. Actually, in practice, even when we didn't get a log error,
1751  * we transition the iclogs to IOERROR state *after* flushing all existing
1752  * iclogs to disk. This is because we don't want anymore new transactions to be
1753  * started or completed afterwards.
1754  *
1755  * We lock the iclogbufs here so that we can serialise against IO completion
1756  * during unmount. We might be processing a shutdown triggered during unmount,
1757  * and that can occur asynchronously to the unmount thread, and hence we need to
1758  * ensure that completes before tearing down the iclogbufs. Hence we need to
1759  * hold the buffer lock across the log IO to acheive that.
1760  */
1761 STATIC int
1762 xlog_bdstrat(
1763         struct xfs_buf          *bp)
1764 {
1765         struct xlog_in_core     *iclog = bp->b_log_item;
1766
1767         xfs_buf_lock(bp);
1768         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1769                 xfs_buf_ioerror(bp, -EIO);
1770                 xfs_buf_stale(bp);
1771                 xfs_buf_ioend(bp);
1772                 /*
1773                  * It would seem logical to return EIO here, but we rely on
1774                  * the log state machine to propagate I/O errors instead of
1775                  * doing it here. Similarly, IO completion will unlock the
1776                  * buffer, so we don't do it here.
1777                  */
1778                 return 0;
1779         }
1780
1781         xfs_buf_submit(bp);
1782         return 0;
1783 }
1784
1785 /*
1786  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1787  * fashion.  Previously, we should have moved the current iclog
1788  * ptr in the log to point to the next available iclog.  This allows further
1789  * write to continue while this code syncs out an iclog ready to go.
1790  * Before an in-core log can be written out, the data section must be scanned
1791  * to save away the 1st word of each BBSIZE block into the header.  We replace
1792  * it with the current cycle count.  Each BBSIZE block is tagged with the
1793  * cycle count because there in an implicit assumption that drives will
1794  * guarantee that entire 512 byte blocks get written at once.  In other words,
1795  * we can't have part of a 512 byte block written and part not written.  By
1796  * tagging each block, we will know which blocks are valid when recovering
1797  * after an unclean shutdown.
1798  *
1799  * This routine is single threaded on the iclog.  No other thread can be in
1800  * this routine with the same iclog.  Changing contents of iclog can there-
1801  * fore be done without grabbing the state machine lock.  Updating the global
1802  * log will require grabbing the lock though.
1803  *
1804  * The entire log manager uses a logical block numbering scheme.  Only
1805  * log_sync (and then only bwrite()) know about the fact that the log may
1806  * not start with block zero on a given device.  The log block start offset
1807  * is added immediately before calling bwrite().
1808  */
1809
1810 STATIC int
1811 xlog_sync(
1812         struct xlog             *log,
1813         struct xlog_in_core     *iclog)
1814 {
1815         xfs_buf_t       *bp;
1816         int             i;
1817         uint            count;          /* byte count of bwrite */
1818         uint            count_init;     /* initial count before roundup */
1819         int             roundoff;       /* roundoff to BB or stripe */
1820         int             split = 0;      /* split write into two regions */
1821         int             error;
1822         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1823         int             size;
1824
1825         XFS_STATS_INC(log->l_mp, xs_log_writes);
1826         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1827
1828         /* Add for LR header */
1829         count_init = log->l_iclog_hsize + iclog->ic_offset;
1830
1831         /* Round out the log write size */
1832         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1833                 /* we have a v2 stripe unit to use */
1834                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1835         } else {
1836                 count = BBTOB(BTOBB(count_init));
1837         }
1838         roundoff = count - count_init;
1839         ASSERT(roundoff >= 0);
1840         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1841                 roundoff < log->l_mp->m_sb.sb_logsunit)
1842                 ||
1843                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1844                  roundoff < BBTOB(1)));
1845
1846         /* move grant heads by roundoff in sync */
1847         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1848         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1849
1850         /* put cycle number in every block */
1851         xlog_pack_data(log, iclog, roundoff);
1852
1853         /* real byte length */
1854         size = iclog->ic_offset;
1855         if (v2)
1856                 size += roundoff;
1857         iclog->ic_header.h_len = cpu_to_be32(size);
1858
1859         bp = iclog->ic_bp;
1860         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1861
1862         XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1863
1864         /* Do we need to split this write into 2 parts? */
1865         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1866                 char            *dptr;
1867
1868                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1869                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1870                 iclog->ic_bwritecnt = 2;
1871
1872                 /*
1873                  * Bump the cycle numbers at the start of each block in the
1874                  * part of the iclog that ends up in the buffer that gets
1875                  * written to the start of the log.
1876                  *
1877                  * Watch out for the header magic number case, though.
1878                  */
1879                 dptr = (char *)&iclog->ic_header + count;
1880                 for (i = 0; i < split; i += BBSIZE) {
1881                         uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1882                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1883                                 cycle++;
1884                         *(__be32 *)dptr = cpu_to_be32(cycle);
1885
1886                         dptr += BBSIZE;
1887                 }
1888         } else {
1889                 iclog->ic_bwritecnt = 1;
1890         }
1891
1892         /* calculcate the checksum */
1893         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1894                                             iclog->ic_datap, size);
1895         /*
1896          * Intentionally corrupt the log record CRC based on the error injection
1897          * frequency, if defined. This facilitates testing log recovery in the
1898          * event of torn writes. Hence, set the IOABORT state to abort the log
1899          * write on I/O completion and shutdown the fs. The subsequent mount
1900          * detects the bad CRC and attempts to recover.
1901          */
1902         if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1903                 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1904                 iclog->ic_state |= XLOG_STATE_IOABORT;
1905                 xfs_warn(log->l_mp,
1906         "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1907                          be64_to_cpu(iclog->ic_header.h_lsn));
1908         }
1909
1910         bp->b_io_length = BTOBB(count);
1911         bp->b_log_item = iclog;
1912         bp->b_flags &= ~XBF_FLUSH;
1913         bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1914
1915         /*
1916          * Flush the data device before flushing the log to make sure all meta
1917          * data written back from the AIL actually made it to disk before
1918          * stamping the new log tail LSN into the log buffer.  For an external
1919          * log we need to issue the flush explicitly, and unfortunately
1920          * synchronously here; for an internal log we can simply use the block
1921          * layer state machine for preflushes.
1922          */
1923         if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1924                 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1925         else
1926                 bp->b_flags |= XBF_FLUSH;
1927
1928         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1929         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1930
1931         xlog_verify_iclog(log, iclog, count, true);
1932
1933         /* account for log which doesn't start at block #0 */
1934         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1935
1936         /*
1937          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1938          * is shutting down.
1939          */
1940         error = xlog_bdstrat(bp);
1941         if (error) {
1942                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1943                 return error;
1944         }
1945         if (split) {
1946                 bp = iclog->ic_log->l_xbuf;
1947                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1948                 xfs_buf_associate_memory(bp,
1949                                 (char *)&iclog->ic_header + count, split);
1950                 bp->b_log_item = iclog;
1951                 bp->b_flags &= ~XBF_FLUSH;
1952                 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1953
1954                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1955                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1956
1957                 /* account for internal log which doesn't start at block #0 */
1958                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1959                 error = xlog_bdstrat(bp);
1960                 if (error) {
1961                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1962                         return error;
1963                 }
1964         }
1965         return 0;
1966 }       /* xlog_sync */
1967
1968 /*
1969  * Deallocate a log structure
1970  */
1971 STATIC void
1972 xlog_dealloc_log(
1973         struct xlog     *log)
1974 {
1975         xlog_in_core_t  *iclog, *next_iclog;
1976         int             i;
1977
1978         xlog_cil_destroy(log);
1979
1980         /*
1981          * Cycle all the iclogbuf locks to make sure all log IO completion
1982          * is done before we tear down these buffers.
1983          */
1984         iclog = log->l_iclog;
1985         for (i = 0; i < log->l_iclog_bufs; i++) {
1986                 xfs_buf_lock(iclog->ic_bp);
1987                 xfs_buf_unlock(iclog->ic_bp);
1988                 iclog = iclog->ic_next;
1989         }
1990
1991         /*
1992          * Always need to ensure that the extra buffer does not point to memory
1993          * owned by another log buffer before we free it. Also, cycle the lock
1994          * first to ensure we've completed IO on it.
1995          */
1996         xfs_buf_lock(log->l_xbuf);
1997         xfs_buf_unlock(log->l_xbuf);
1998         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1999         xfs_buf_free(log->l_xbuf);
2000
2001         iclog = log->l_iclog;
2002         for (i = 0; i < log->l_iclog_bufs; i++) {
2003                 xfs_buf_free(iclog->ic_bp);
2004                 next_iclog = iclog->ic_next;
2005                 kmem_free(iclog);
2006                 iclog = next_iclog;
2007         }
2008         spinlock_destroy(&log->l_icloglock);
2009
2010         log->l_mp->m_log = NULL;
2011         kmem_free(log);
2012 }       /* xlog_dealloc_log */
2013
2014 /*
2015  * Update counters atomically now that memcpy is done.
2016  */
2017 /* ARGSUSED */
2018 static inline void
2019 xlog_state_finish_copy(
2020         struct xlog             *log,
2021         struct xlog_in_core     *iclog,
2022         int                     record_cnt,
2023         int                     copy_bytes)
2024 {
2025         spin_lock(&log->l_icloglock);
2026
2027         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2028         iclog->ic_offset += copy_bytes;
2029
2030         spin_unlock(&log->l_icloglock);
2031 }       /* xlog_state_finish_copy */
2032
2033
2034
2035
2036 /*
2037  * print out info relating to regions written which consume
2038  * the reservation
2039  */
2040 void
2041 xlog_print_tic_res(
2042         struct xfs_mount        *mp,
2043         struct xlog_ticket      *ticket)
2044 {
2045         uint i;
2046         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2047
2048         /* match with XLOG_REG_TYPE_* in xfs_log.h */
2049 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2050         static char *res_type_str[XLOG_REG_TYPE_MAX + 1] = {
2051             REG_TYPE_STR(BFORMAT, "bformat"),
2052             REG_TYPE_STR(BCHUNK, "bchunk"),
2053             REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2054             REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2055             REG_TYPE_STR(IFORMAT, "iformat"),
2056             REG_TYPE_STR(ICORE, "icore"),
2057             REG_TYPE_STR(IEXT, "iext"),
2058             REG_TYPE_STR(IBROOT, "ibroot"),
2059             REG_TYPE_STR(ILOCAL, "ilocal"),
2060             REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2061             REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2062             REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2063             REG_TYPE_STR(QFORMAT, "qformat"),
2064             REG_TYPE_STR(DQUOT, "dquot"),
2065             REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2066             REG_TYPE_STR(LRHEADER, "LR header"),
2067             REG_TYPE_STR(UNMOUNT, "unmount"),
2068             REG_TYPE_STR(COMMIT, "commit"),
2069             REG_TYPE_STR(TRANSHDR, "trans header"),
2070             REG_TYPE_STR(ICREATE, "inode create")
2071         };
2072 #undef REG_TYPE_STR
2073
2074         xfs_warn(mp, "ticket reservation summary:");
2075         xfs_warn(mp, "  unit res    = %d bytes",
2076                  ticket->t_unit_res);
2077         xfs_warn(mp, "  current res = %d bytes",
2078                  ticket->t_curr_res);
2079         xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2080                  ticket->t_res_arr_sum, ticket->t_res_o_flow);
2081         xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2082                  ticket->t_res_num_ophdrs, ophdr_spc);
2083         xfs_warn(mp, "  ophdr + reg = %u bytes",
2084                  ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2085         xfs_warn(mp, "  num regions = %u",
2086                  ticket->t_res_num);
2087
2088         for (i = 0; i < ticket->t_res_num; i++) {
2089                 uint r_type = ticket->t_res_arr[i].r_type;
2090                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2091                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2092                             "bad-rtype" : res_type_str[r_type]),
2093                             ticket->t_res_arr[i].r_len);
2094         }
2095 }
2096
2097 /*
2098  * Print a summary of the transaction.
2099  */
2100 void
2101 xlog_print_trans(
2102         struct xfs_trans        *tp)
2103 {
2104         struct xfs_mount        *mp = tp->t_mountp;
2105         struct xfs_log_item     *lip;
2106
2107         /* dump core transaction and ticket info */
2108         xfs_warn(mp, "transaction summary:");
2109         xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2110         xfs_warn(mp, "  log count = %d", tp->t_log_count);
2111         xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2112
2113         xlog_print_tic_res(mp, tp->t_ticket);
2114
2115         /* dump each log item */
2116         list_for_each_entry(lip, &tp->t_items, li_trans) {
2117                 struct xfs_log_vec      *lv = lip->li_lv;
2118                 struct xfs_log_iovec    *vec;
2119                 int                     i;
2120
2121                 xfs_warn(mp, "log item: ");
2122                 xfs_warn(mp, "  type    = 0x%x", lip->li_type);
2123                 xfs_warn(mp, "  flags   = 0x%lx", lip->li_flags);
2124                 if (!lv)
2125                         continue;
2126                 xfs_warn(mp, "  niovecs = %d", lv->lv_niovecs);
2127                 xfs_warn(mp, "  size    = %d", lv->lv_size);
2128                 xfs_warn(mp, "  bytes   = %d", lv->lv_bytes);
2129                 xfs_warn(mp, "  buf len = %d", lv->lv_buf_len);
2130
2131                 /* dump each iovec for the log item */
2132                 vec = lv->lv_iovecp;
2133                 for (i = 0; i < lv->lv_niovecs; i++) {
2134                         int dumplen = min(vec->i_len, 32);
2135
2136                         xfs_warn(mp, "  iovec[%d]", i);
2137                         xfs_warn(mp, "    type  = 0x%x", vec->i_type);
2138                         xfs_warn(mp, "    len   = %d", vec->i_len);
2139                         xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2140                         xfs_hex_dump(vec->i_addr, dumplen);
2141
2142                         vec++;
2143                 }
2144         }
2145 }
2146
2147 /*
2148  * Calculate the potential space needed by the log vector.  Each region gets
2149  * its own xlog_op_header_t and may need to be double word aligned.
2150  */
2151 static int
2152 xlog_write_calc_vec_length(
2153         struct xlog_ticket      *ticket,
2154         struct xfs_log_vec      *log_vector)
2155 {
2156         struct xfs_log_vec      *lv;
2157         int                     headers = 0;
2158         int                     len = 0;
2159         int                     i;
2160
2161         /* acct for start rec of xact */
2162         if (ticket->t_flags & XLOG_TIC_INITED)
2163                 headers++;
2164
2165         for (lv = log_vector; lv; lv = lv->lv_next) {
2166                 /* we don't write ordered log vectors */
2167                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2168                         continue;
2169
2170                 headers += lv->lv_niovecs;
2171
2172                 for (i = 0; i < lv->lv_niovecs; i++) {
2173                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2174
2175                         len += vecp->i_len;
2176                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2177                 }
2178         }
2179
2180         ticket->t_res_num_ophdrs += headers;
2181         len += headers * sizeof(struct xlog_op_header);
2182
2183         return len;
2184 }
2185
2186 /*
2187  * If first write for transaction, insert start record  We can't be trying to
2188  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2189  */
2190 static int
2191 xlog_write_start_rec(
2192         struct xlog_op_header   *ophdr,
2193         struct xlog_ticket      *ticket)
2194 {
2195         if (!(ticket->t_flags & XLOG_TIC_INITED))
2196                 return 0;
2197
2198         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2199         ophdr->oh_clientid = ticket->t_clientid;
2200         ophdr->oh_len = 0;
2201         ophdr->oh_flags = XLOG_START_TRANS;
2202         ophdr->oh_res2 = 0;
2203
2204         ticket->t_flags &= ~XLOG_TIC_INITED;
2205
2206         return sizeof(struct xlog_op_header);
2207 }
2208
2209 static xlog_op_header_t *
2210 xlog_write_setup_ophdr(
2211         struct xlog             *log,
2212         struct xlog_op_header   *ophdr,
2213         struct xlog_ticket      *ticket,
2214         uint                    flags)
2215 {
2216         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2217         ophdr->oh_clientid = ticket->t_clientid;
2218         ophdr->oh_res2 = 0;
2219
2220         /* are we copying a commit or unmount record? */
2221         ophdr->oh_flags = flags;
2222
2223         /*
2224          * We've seen logs corrupted with bad transaction client ids.  This
2225          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2226          * and shut down the filesystem.
2227          */
2228         switch (ophdr->oh_clientid)  {
2229         case XFS_TRANSACTION:
2230         case XFS_VOLUME:
2231         case XFS_LOG:
2232                 break;
2233         default:
2234                 xfs_warn(log->l_mp,
2235                         "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2236                         ophdr->oh_clientid, ticket);
2237                 return NULL;
2238         }
2239
2240         return ophdr;
2241 }
2242
2243 /*
2244  * Set up the parameters of the region copy into the log. This has
2245  * to handle region write split across multiple log buffers - this
2246  * state is kept external to this function so that this code can
2247  * be written in an obvious, self documenting manner.
2248  */
2249 static int
2250 xlog_write_setup_copy(
2251         struct xlog_ticket      *ticket,
2252         struct xlog_op_header   *ophdr,
2253         int                     space_available,
2254         int                     space_required,
2255         int                     *copy_off,
2256         int                     *copy_len,
2257         int                     *last_was_partial_copy,
2258         int                     *bytes_consumed)
2259 {
2260         int                     still_to_copy;
2261
2262         still_to_copy = space_required - *bytes_consumed;
2263         *copy_off = *bytes_consumed;
2264
2265         if (still_to_copy <= space_available) {
2266                 /* write of region completes here */
2267                 *copy_len = still_to_copy;
2268                 ophdr->oh_len = cpu_to_be32(*copy_len);
2269                 if (*last_was_partial_copy)
2270                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2271                 *last_was_partial_copy = 0;
2272                 *bytes_consumed = 0;
2273                 return 0;
2274         }
2275
2276         /* partial write of region, needs extra log op header reservation */
2277         *copy_len = space_available;
2278         ophdr->oh_len = cpu_to_be32(*copy_len);
2279         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2280         if (*last_was_partial_copy)
2281                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2282         *bytes_consumed += *copy_len;
2283         (*last_was_partial_copy)++;
2284
2285         /* account for new log op header */
2286         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2287         ticket->t_res_num_ophdrs++;
2288
2289         return sizeof(struct xlog_op_header);
2290 }
2291
2292 static int
2293 xlog_write_copy_finish(
2294         struct xlog             *log,
2295         struct xlog_in_core     *iclog,
2296         uint                    flags,
2297         int                     *record_cnt,
2298         int                     *data_cnt,
2299         int                     *partial_copy,
2300         int                     *partial_copy_len,
2301         int                     log_offset,
2302         struct xlog_in_core     **commit_iclog)
2303 {
2304         if (*partial_copy) {
2305                 /*
2306                  * This iclog has already been marked WANT_SYNC by
2307                  * xlog_state_get_iclog_space.
2308                  */
2309                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2310                 *record_cnt = 0;
2311                 *data_cnt = 0;
2312                 return xlog_state_release_iclog(log, iclog);
2313         }
2314
2315         *partial_copy = 0;
2316         *partial_copy_len = 0;
2317
2318         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2319                 /* no more space in this iclog - push it. */
2320                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2321                 *record_cnt = 0;
2322                 *data_cnt = 0;
2323
2324                 spin_lock(&log->l_icloglock);
2325                 xlog_state_want_sync(log, iclog);
2326                 spin_unlock(&log->l_icloglock);
2327
2328                 if (!commit_iclog)
2329                         return xlog_state_release_iclog(log, iclog);
2330                 ASSERT(flags & XLOG_COMMIT_TRANS);
2331                 *commit_iclog = iclog;
2332         }
2333
2334         return 0;
2335 }
2336
2337 /*
2338  * Write some region out to in-core log
2339  *
2340  * This will be called when writing externally provided regions or when
2341  * writing out a commit record for a given transaction.
2342  *
2343  * General algorithm:
2344  *      1. Find total length of this write.  This may include adding to the
2345  *              lengths passed in.
2346  *      2. Check whether we violate the tickets reservation.
2347  *      3. While writing to this iclog
2348  *          A. Reserve as much space in this iclog as can get
2349  *          B. If this is first write, save away start lsn
2350  *          C. While writing this region:
2351  *              1. If first write of transaction, write start record
2352  *              2. Write log operation header (header per region)
2353  *              3. Find out if we can fit entire region into this iclog
2354  *              4. Potentially, verify destination memcpy ptr
2355  *              5. Memcpy (partial) region
2356  *              6. If partial copy, release iclog; otherwise, continue
2357  *                      copying more regions into current iclog
2358  *      4. Mark want sync bit (in simulation mode)
2359  *      5. Release iclog for potential flush to on-disk log.
2360  *
2361  * ERRORS:
2362  * 1.   Panic if reservation is overrun.  This should never happen since
2363  *      reservation amounts are generated internal to the filesystem.
2364  * NOTES:
2365  * 1. Tickets are single threaded data structures.
2366  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2367  *      syncing routine.  When a single log_write region needs to span
2368  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2369  *      on all log operation writes which don't contain the end of the
2370  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2371  *      operation which contains the end of the continued log_write region.
2372  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2373  *      we don't really know exactly how much space will be used.  As a result,
2374  *      we don't update ic_offset until the end when we know exactly how many
2375  *      bytes have been written out.
2376  */
2377 int
2378 xlog_write(
2379         struct xlog             *log,
2380         struct xfs_log_vec      *log_vector,
2381         struct xlog_ticket      *ticket,
2382         xfs_lsn_t               *start_lsn,
2383         struct xlog_in_core     **commit_iclog,
2384         uint                    flags)
2385 {
2386         struct xlog_in_core     *iclog = NULL;
2387         struct xfs_log_iovec    *vecp;
2388         struct xfs_log_vec      *lv;
2389         int                     len;
2390         int                     index;
2391         int                     partial_copy = 0;
2392         int                     partial_copy_len = 0;
2393         int                     contwr = 0;
2394         int                     record_cnt = 0;
2395         int                     data_cnt = 0;
2396         int                     error;
2397
2398         *start_lsn = 0;
2399
2400         len = xlog_write_calc_vec_length(ticket, log_vector);
2401
2402         /*
2403          * Region headers and bytes are already accounted for.
2404          * We only need to take into account start records and
2405          * split regions in this function.
2406          */
2407         if (ticket->t_flags & XLOG_TIC_INITED)
2408                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2409
2410         /*
2411          * Commit record headers need to be accounted for. These
2412          * come in as separate writes so are easy to detect.
2413          */
2414         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2415                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2416
2417         if (ticket->t_curr_res < 0) {
2418                 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2419                      "ctx ticket reservation ran out. Need to up reservation");
2420                 xlog_print_tic_res(log->l_mp, ticket);
2421                 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2422         }
2423
2424         index = 0;
2425         lv = log_vector;
2426         vecp = lv->lv_iovecp;
2427         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2428                 void            *ptr;
2429                 int             log_offset;
2430
2431                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2432                                                    &contwr, &log_offset);
2433                 if (error)
2434                         return error;
2435
2436                 ASSERT(log_offset <= iclog->ic_size - 1);
2437                 ptr = iclog->ic_datap + log_offset;
2438
2439                 /* start_lsn is the first lsn written to. That's all we need. */
2440                 if (!*start_lsn)
2441                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2442
2443                 /*
2444                  * This loop writes out as many regions as can fit in the amount
2445                  * of space which was allocated by xlog_state_get_iclog_space().
2446                  */
2447                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2448                         struct xfs_log_iovec    *reg;
2449                         struct xlog_op_header   *ophdr;
2450                         int                     start_rec_copy;
2451                         int                     copy_len;
2452                         int                     copy_off;
2453                         bool                    ordered = false;
2454
2455                         /* ordered log vectors have no regions to write */
2456                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2457                                 ASSERT(lv->lv_niovecs == 0);
2458                                 ordered = true;
2459                                 goto next_lv;
2460                         }
2461
2462                         reg = &vecp[index];
2463                         ASSERT(reg->i_len % sizeof(int32_t) == 0);
2464                         ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2465
2466                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2467                         if (start_rec_copy) {
2468                                 record_cnt++;
2469                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2470                                                    start_rec_copy);
2471                         }
2472
2473                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2474                         if (!ophdr)
2475                                 return -EIO;
2476
2477                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2478                                            sizeof(struct xlog_op_header));
2479
2480                         len += xlog_write_setup_copy(ticket, ophdr,
2481                                                      iclog->ic_size-log_offset,
2482                                                      reg->i_len,
2483                                                      &copy_off, &copy_len,
2484                                                      &partial_copy,
2485                                                      &partial_copy_len);
2486                         xlog_verify_dest_ptr(log, ptr);
2487
2488                         /*
2489                          * Copy region.
2490                          *
2491                          * Unmount records just log an opheader, so can have
2492                          * empty payloads with no data region to copy. Hence we
2493                          * only copy the payload if the vector says it has data
2494                          * to copy.
2495                          */
2496                         ASSERT(copy_len >= 0);
2497                         if (copy_len > 0) {
2498                                 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2499                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2500                                                    copy_len);
2501                         }
2502                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2503                         record_cnt++;
2504                         data_cnt += contwr ? copy_len : 0;
2505
2506                         error = xlog_write_copy_finish(log, iclog, flags,
2507                                                        &record_cnt, &data_cnt,
2508                                                        &partial_copy,
2509                                                        &partial_copy_len,
2510                                                        log_offset,
2511                                                        commit_iclog);
2512                         if (error)
2513                                 return error;
2514
2515                         /*
2516                          * if we had a partial copy, we need to get more iclog
2517                          * space but we don't want to increment the region
2518                          * index because there is still more is this region to
2519                          * write.
2520                          *
2521                          * If we completed writing this region, and we flushed
2522                          * the iclog (indicated by resetting of the record
2523                          * count), then we also need to get more log space. If
2524                          * this was the last record, though, we are done and
2525                          * can just return.
2526                          */
2527                         if (partial_copy)
2528                                 break;
2529
2530                         if (++index == lv->lv_niovecs) {
2531 next_lv:
2532                                 lv = lv->lv_next;
2533                                 index = 0;
2534                                 if (lv)
2535                                         vecp = lv->lv_iovecp;
2536                         }
2537                         if (record_cnt == 0 && !ordered) {
2538                                 if (!lv)
2539                                         return 0;
2540                                 break;
2541                         }
2542                 }
2543         }
2544
2545         ASSERT(len == 0);
2546
2547         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2548         if (!commit_iclog)
2549                 return xlog_state_release_iclog(log, iclog);
2550
2551         ASSERT(flags & XLOG_COMMIT_TRANS);
2552         *commit_iclog = iclog;
2553         return 0;
2554 }
2555
2556
2557 /*****************************************************************************
2558  *
2559  *              State Machine functions
2560  *
2561  *****************************************************************************
2562  */
2563
2564 /* Clean iclogs starting from the head.  This ordering must be
2565  * maintained, so an iclog doesn't become ACTIVE beyond one that
2566  * is SYNCING.  This is also required to maintain the notion that we use
2567  * a ordered wait queue to hold off would be writers to the log when every
2568  * iclog is trying to sync to disk.
2569  *
2570  * State Change: DIRTY -> ACTIVE
2571  */
2572 STATIC void
2573 xlog_state_clean_log(
2574         struct xlog *log)
2575 {
2576         xlog_in_core_t  *iclog;
2577         int changed = 0;
2578
2579         iclog = log->l_iclog;
2580         do {
2581                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2582                         iclog->ic_state = XLOG_STATE_ACTIVE;
2583                         iclog->ic_offset       = 0;
2584                         ASSERT(iclog->ic_callback == NULL);
2585                         /*
2586                          * If the number of ops in this iclog indicate it just
2587                          * contains the dummy transaction, we can
2588                          * change state into IDLE (the second time around).
2589                          * Otherwise we should change the state into
2590                          * NEED a dummy.
2591                          * We don't need to cover the dummy.
2592                          */
2593                         if (!changed &&
2594                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2595                                         XLOG_COVER_OPS)) {
2596                                 changed = 1;
2597                         } else {
2598                                 /*
2599                                  * We have two dirty iclogs so start over
2600                                  * This could also be num of ops indicates
2601                                  * this is not the dummy going out.
2602                                  */
2603                                 changed = 2;
2604                         }
2605                         iclog->ic_header.h_num_logops = 0;
2606                         memset(iclog->ic_header.h_cycle_data, 0,
2607                               sizeof(iclog->ic_header.h_cycle_data));
2608                         iclog->ic_header.h_lsn = 0;
2609                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2610                         /* do nothing */;
2611                 else
2612                         break;  /* stop cleaning */
2613                 iclog = iclog->ic_next;
2614         } while (iclog != log->l_iclog);
2615
2616         /* log is locked when we are called */
2617         /*
2618          * Change state for the dummy log recording.
2619          * We usually go to NEED. But we go to NEED2 if the changed indicates
2620          * we are done writing the dummy record.
2621          * If we are done with the second dummy recored (DONE2), then
2622          * we go to IDLE.
2623          */
2624         if (changed) {
2625                 switch (log->l_covered_state) {
2626                 case XLOG_STATE_COVER_IDLE:
2627                 case XLOG_STATE_COVER_NEED:
2628                 case XLOG_STATE_COVER_NEED2:
2629                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2630                         break;
2631
2632                 case XLOG_STATE_COVER_DONE:
2633                         if (changed == 1)
2634                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2635                         else
2636                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2637                         break;
2638
2639                 case XLOG_STATE_COVER_DONE2:
2640                         if (changed == 1)
2641                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2642                         else
2643                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2644                         break;
2645
2646                 default:
2647                         ASSERT(0);
2648                 }
2649         }
2650 }       /* xlog_state_clean_log */
2651
2652 STATIC xfs_lsn_t
2653 xlog_get_lowest_lsn(
2654         struct xlog     *log)
2655 {
2656         xlog_in_core_t  *lsn_log;
2657         xfs_lsn_t       lowest_lsn, lsn;
2658
2659         lsn_log = log->l_iclog;
2660         lowest_lsn = 0;
2661         do {
2662             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2663                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2664                 if ((lsn && !lowest_lsn) ||
2665                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2666                         lowest_lsn = lsn;
2667                 }
2668             }
2669             lsn_log = lsn_log->ic_next;
2670         } while (lsn_log != log->l_iclog);
2671         return lowest_lsn;
2672 }
2673
2674
2675 STATIC void
2676 xlog_state_do_callback(
2677         struct xlog             *log,
2678         int                     aborted,
2679         struct xlog_in_core     *ciclog)
2680 {
2681         xlog_in_core_t     *iclog;
2682         xlog_in_core_t     *first_iclog;        /* used to know when we've
2683                                                  * processed all iclogs once */
2684         xfs_log_callback_t *cb, *cb_next;
2685         int                flushcnt = 0;
2686         xfs_lsn_t          lowest_lsn;
2687         int                ioerrors;    /* counter: iclogs with errors */
2688         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2689         int                funcdidcallbacks; /* flag: function did callbacks */
2690         int                repeats;     /* for issuing console warnings if
2691                                          * looping too many times */
2692         int                wake = 0;
2693
2694         spin_lock(&log->l_icloglock);
2695         first_iclog = iclog = log->l_iclog;
2696         ioerrors = 0;
2697         funcdidcallbacks = 0;
2698         repeats = 0;
2699
2700         do {
2701                 /*
2702                  * Scan all iclogs starting with the one pointed to by the
2703                  * log.  Reset this starting point each time the log is
2704                  * unlocked (during callbacks).
2705                  *
2706                  * Keep looping through iclogs until one full pass is made
2707                  * without running any callbacks.
2708                  */
2709                 first_iclog = log->l_iclog;
2710                 iclog = log->l_iclog;
2711                 loopdidcallbacks = 0;
2712                 repeats++;
2713
2714                 do {
2715
2716                         /* skip all iclogs in the ACTIVE & DIRTY states */
2717                         if (iclog->ic_state &
2718                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2719                                 iclog = iclog->ic_next;
2720                                 continue;
2721                         }
2722
2723                         /*
2724                          * Between marking a filesystem SHUTDOWN and stopping
2725                          * the log, we do flush all iclogs to disk (if there
2726                          * wasn't a log I/O error). So, we do want things to
2727                          * go smoothly in case of just a SHUTDOWN  w/o a
2728                          * LOG_IO_ERROR.
2729                          */
2730                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2731                                 /*
2732                                  * Can only perform callbacks in order.  Since
2733                                  * this iclog is not in the DONE_SYNC/
2734                                  * DO_CALLBACK state, we skip the rest and
2735                                  * just try to clean up.  If we set our iclog
2736                                  * to DO_CALLBACK, we will not process it when
2737                                  * we retry since a previous iclog is in the
2738                                  * CALLBACK and the state cannot change since
2739                                  * we are holding the l_icloglock.
2740                                  */
2741                                 if (!(iclog->ic_state &
2742                                         (XLOG_STATE_DONE_SYNC |
2743                                                  XLOG_STATE_DO_CALLBACK))) {
2744                                         if (ciclog && (ciclog->ic_state ==
2745                                                         XLOG_STATE_DONE_SYNC)) {
2746                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2747                                         }
2748                                         break;
2749                                 }
2750                                 /*
2751                                  * We now have an iclog that is in either the
2752                                  * DO_CALLBACK or DONE_SYNC states. The other
2753                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2754                                  * caught by the above if and are going to
2755                                  * clean (i.e. we aren't doing their callbacks)
2756                                  * see the above if.
2757                                  */
2758
2759                                 /*
2760                                  * We will do one more check here to see if we
2761                                  * have chased our tail around.
2762                                  */
2763
2764                                 lowest_lsn = xlog_get_lowest_lsn(log);
2765                                 if (lowest_lsn &&
2766                                     XFS_LSN_CMP(lowest_lsn,
2767                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2768                                         iclog = iclog->ic_next;
2769                                         continue; /* Leave this iclog for
2770                                                    * another thread */
2771                                 }
2772
2773                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2774
2775
2776                                 /*
2777                                  * Completion of a iclog IO does not imply that
2778                                  * a transaction has completed, as transactions
2779                                  * can be large enough to span many iclogs. We
2780                                  * cannot change the tail of the log half way
2781                                  * through a transaction as this may be the only
2782                                  * transaction in the log and moving th etail to
2783                                  * point to the middle of it will prevent
2784                                  * recovery from finding the start of the
2785                                  * transaction. Hence we should only update the
2786                                  * last_sync_lsn if this iclog contains
2787                                  * transaction completion callbacks on it.
2788                                  *
2789                                  * We have to do this before we drop the
2790                                  * icloglock to ensure we are the only one that
2791                                  * can update it.
2792                                  */
2793                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2794                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2795                                 if (iclog->ic_callback)
2796                                         atomic64_set(&log->l_last_sync_lsn,
2797                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2798
2799                         } else
2800                                 ioerrors++;
2801
2802                         spin_unlock(&log->l_icloglock);
2803
2804                         /*
2805                          * Keep processing entries in the callback list until
2806                          * we come around and it is empty.  We need to
2807                          * atomically see that the list is empty and change the
2808                          * state to DIRTY so that we don't miss any more
2809                          * callbacks being added.
2810                          */
2811                         spin_lock(&iclog->ic_callback_lock);
2812                         cb = iclog->ic_callback;
2813                         while (cb) {
2814                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2815                                 iclog->ic_callback = NULL;
2816                                 spin_unlock(&iclog->ic_callback_lock);
2817
2818                                 /* perform callbacks in the order given */
2819                                 for (; cb; cb = cb_next) {
2820                                         cb_next = cb->cb_next;
2821                                         cb->cb_func(cb->cb_arg, aborted);
2822                                 }
2823                                 spin_lock(&iclog->ic_callback_lock);
2824                                 cb = iclog->ic_callback;
2825                         }
2826
2827                         loopdidcallbacks++;
2828                         funcdidcallbacks++;
2829
2830                         spin_lock(&log->l_icloglock);
2831                         ASSERT(iclog->ic_callback == NULL);
2832                         spin_unlock(&iclog->ic_callback_lock);
2833                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2834                                 iclog->ic_state = XLOG_STATE_DIRTY;
2835
2836                         /*
2837                          * Transition from DIRTY to ACTIVE if applicable.
2838                          * NOP if STATE_IOERROR.
2839                          */
2840                         xlog_state_clean_log(log);
2841
2842                         /* wake up threads waiting in xfs_log_force() */
2843                         wake_up_all(&iclog->ic_force_wait);
2844
2845                         iclog = iclog->ic_next;
2846                 } while (first_iclog != iclog);
2847
2848                 if (repeats > 5000) {
2849                         flushcnt += repeats;
2850                         repeats = 0;
2851                         xfs_warn(log->l_mp,
2852                                 "%s: possible infinite loop (%d iterations)",
2853                                 __func__, flushcnt);
2854                 }
2855         } while (!ioerrors && loopdidcallbacks);
2856
2857 #ifdef DEBUG
2858         /*
2859          * Make one last gasp attempt to see if iclogs are being left in limbo.
2860          * If the above loop finds an iclog earlier than the current iclog and
2861          * in one of the syncing states, the current iclog is put into
2862          * DO_CALLBACK and the callbacks are deferred to the completion of the
2863          * earlier iclog. Walk the iclogs in order and make sure that no iclog
2864          * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2865          * states.
2866          *
2867          * Note that SYNCING|IOABORT is a valid state so we cannot just check
2868          * for ic_state == SYNCING.
2869          */
2870         if (funcdidcallbacks) {
2871                 first_iclog = iclog = log->l_iclog;
2872                 do {
2873                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2874                         /*
2875                          * Terminate the loop if iclogs are found in states
2876                          * which will cause other threads to clean up iclogs.
2877                          *
2878                          * SYNCING - i/o completion will go through logs
2879                          * DONE_SYNC - interrupt thread should be waiting for
2880                          *              l_icloglock
2881                          * IOERROR - give up hope all ye who enter here
2882                          */
2883                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2884                             iclog->ic_state & XLOG_STATE_SYNCING ||
2885                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2886                             iclog->ic_state == XLOG_STATE_IOERROR )
2887                                 break;
2888                         iclog = iclog->ic_next;
2889                 } while (first_iclog != iclog);
2890         }
2891 #endif
2892
2893         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2894                 wake = 1;
2895         spin_unlock(&log->l_icloglock);
2896
2897         if (wake)
2898                 wake_up_all(&log->l_flush_wait);
2899 }
2900
2901
2902 /*
2903  * Finish transitioning this iclog to the dirty state.
2904  *
2905  * Make sure that we completely execute this routine only when this is
2906  * the last call to the iclog.  There is a good chance that iclog flushes,
2907  * when we reach the end of the physical log, get turned into 2 separate
2908  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2909  * routine.  By using the reference count bwritecnt, we guarantee that only
2910  * the second completion goes through.
2911  *
2912  * Callbacks could take time, so they are done outside the scope of the
2913  * global state machine log lock.
2914  */
2915 STATIC void
2916 xlog_state_done_syncing(
2917         xlog_in_core_t  *iclog,
2918         int             aborted)
2919 {
2920         struct xlog        *log = iclog->ic_log;
2921
2922         spin_lock(&log->l_icloglock);
2923
2924         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2925                iclog->ic_state == XLOG_STATE_IOERROR);
2926         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2927         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2928
2929
2930         /*
2931          * If we got an error, either on the first buffer, or in the case of
2932          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2933          * and none should ever be attempted to be written to disk
2934          * again.
2935          */
2936         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2937                 if (--iclog->ic_bwritecnt == 1) {
2938                         spin_unlock(&log->l_icloglock);
2939                         return;
2940                 }
2941                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2942         }
2943
2944         /*
2945          * Someone could be sleeping prior to writing out the next
2946          * iclog buffer, we wake them all, one will get to do the
2947          * I/O, the others get to wait for the result.
2948          */
2949         wake_up_all(&iclog->ic_write_wait);
2950         spin_unlock(&log->l_icloglock);
2951         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2952 }       /* xlog_state_done_syncing */
2953
2954
2955 /*
2956  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2957  * sleep.  We wait on the flush queue on the head iclog as that should be
2958  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2959  * we will wait here and all new writes will sleep until a sync completes.
2960  *
2961  * The in-core logs are used in a circular fashion. They are not used
2962  * out-of-order even when an iclog past the head is free.
2963  *
2964  * return:
2965  *      * log_offset where xlog_write() can start writing into the in-core
2966  *              log's data space.
2967  *      * in-core log pointer to which xlog_write() should write.
2968  *      * boolean indicating this is a continued write to an in-core log.
2969  *              If this is the last write, then the in-core log's offset field
2970  *              needs to be incremented, depending on the amount of data which
2971  *              is copied.
2972  */
2973 STATIC int
2974 xlog_state_get_iclog_space(
2975         struct xlog             *log,
2976         int                     len,
2977         struct xlog_in_core     **iclogp,
2978         struct xlog_ticket      *ticket,
2979         int                     *continued_write,
2980         int                     *logoffsetp)
2981 {
2982         int               log_offset;
2983         xlog_rec_header_t *head;
2984         xlog_in_core_t    *iclog;
2985         int               error;
2986
2987 restart:
2988         spin_lock(&log->l_icloglock);
2989         if (XLOG_FORCED_SHUTDOWN(log)) {
2990                 spin_unlock(&log->l_icloglock);
2991                 return -EIO;
2992         }
2993
2994         iclog = log->l_iclog;
2995         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2996                 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2997
2998                 /* Wait for log writes to have flushed */
2999                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
3000                 goto restart;
3001         }
3002
3003         head = &iclog->ic_header;
3004
3005         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
3006         log_offset = iclog->ic_offset;
3007
3008         /* On the 1st write to an iclog, figure out lsn.  This works
3009          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3010          * committing to.  If the offset is set, that's how many blocks
3011          * must be written.
3012          */
3013         if (log_offset == 0) {
3014                 ticket->t_curr_res -= log->l_iclog_hsize;
3015                 xlog_tic_add_region(ticket,
3016                                     log->l_iclog_hsize,
3017                                     XLOG_REG_TYPE_LRHEADER);
3018                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3019                 head->h_lsn = cpu_to_be64(
3020                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3021                 ASSERT(log->l_curr_block >= 0);
3022         }
3023
3024         /* If there is enough room to write everything, then do it.  Otherwise,
3025          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3026          * bit is on, so this will get flushed out.  Don't update ic_offset
3027          * until you know exactly how many bytes get copied.  Therefore, wait
3028          * until later to update ic_offset.
3029          *
3030          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3031          * can fit into remaining data section.
3032          */
3033         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3034                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3035
3036                 /*
3037                  * If I'm the only one writing to this iclog, sync it to disk.
3038                  * We need to do an atomic compare and decrement here to avoid
3039                  * racing with concurrent atomic_dec_and_lock() calls in
3040                  * xlog_state_release_iclog() when there is more than one
3041                  * reference to the iclog.
3042                  */
3043                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3044                         /* we are the only one */
3045                         spin_unlock(&log->l_icloglock);
3046                         error = xlog_state_release_iclog(log, iclog);
3047                         if (error)
3048                                 return error;
3049                 } else {
3050                         spin_unlock(&log->l_icloglock);
3051                 }
3052                 goto restart;
3053         }
3054
3055         /* Do we have enough room to write the full amount in the remainder
3056          * of this iclog?  Or must we continue a write on the next iclog and
3057          * mark this iclog as completely taken?  In the case where we switch
3058          * iclogs (to mark it taken), this particular iclog will release/sync
3059          * to disk in xlog_write().
3060          */
3061         if (len <= iclog->ic_size - iclog->ic_offset) {
3062                 *continued_write = 0;
3063                 iclog->ic_offset += len;
3064         } else {
3065                 *continued_write = 1;
3066                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3067         }
3068         *iclogp = iclog;
3069
3070         ASSERT(iclog->ic_offset <= iclog->ic_size);
3071         spin_unlock(&log->l_icloglock);
3072
3073         *logoffsetp = log_offset;
3074         return 0;
3075 }       /* xlog_state_get_iclog_space */
3076
3077 /* The first cnt-1 times through here we don't need to
3078  * move the grant write head because the permanent
3079  * reservation has reserved cnt times the unit amount.
3080  * Release part of current permanent unit reservation and
3081  * reset current reservation to be one units worth.  Also
3082  * move grant reservation head forward.
3083  */
3084 STATIC void
3085 xlog_regrant_reserve_log_space(
3086         struct xlog             *log,
3087         struct xlog_ticket      *ticket)
3088 {
3089         trace_xfs_log_regrant_reserve_enter(log, ticket);
3090
3091         if (ticket->t_cnt > 0)
3092                 ticket->t_cnt--;
3093
3094         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3095                                         ticket->t_curr_res);
3096         xlog_grant_sub_space(log, &log->l_write_head.grant,
3097                                         ticket->t_curr_res);
3098         ticket->t_curr_res = ticket->t_unit_res;
3099         xlog_tic_reset_res(ticket);
3100
3101         trace_xfs_log_regrant_reserve_sub(log, ticket);
3102
3103         /* just return if we still have some of the pre-reserved space */
3104         if (ticket->t_cnt > 0)
3105                 return;
3106
3107         xlog_grant_add_space(log, &log->l_reserve_head.grant,
3108                                         ticket->t_unit_res);
3109
3110         trace_xfs_log_regrant_reserve_exit(log, ticket);
3111
3112         ticket->t_curr_res = ticket->t_unit_res;
3113         xlog_tic_reset_res(ticket);
3114 }       /* xlog_regrant_reserve_log_space */
3115
3116
3117 /*
3118  * Give back the space left from a reservation.
3119  *
3120  * All the information we need to make a correct determination of space left
3121  * is present.  For non-permanent reservations, things are quite easy.  The
3122  * count should have been decremented to zero.  We only need to deal with the
3123  * space remaining in the current reservation part of the ticket.  If the
3124  * ticket contains a permanent reservation, there may be left over space which
3125  * needs to be released.  A count of N means that N-1 refills of the current
3126  * reservation can be done before we need to ask for more space.  The first
3127  * one goes to fill up the first current reservation.  Once we run out of
3128  * space, the count will stay at zero and the only space remaining will be
3129  * in the current reservation field.
3130  */
3131 STATIC void
3132 xlog_ungrant_log_space(
3133         struct xlog             *log,
3134         struct xlog_ticket      *ticket)
3135 {
3136         int     bytes;
3137
3138         if (ticket->t_cnt > 0)
3139                 ticket->t_cnt--;
3140
3141         trace_xfs_log_ungrant_enter(log, ticket);
3142         trace_xfs_log_ungrant_sub(log, ticket);
3143
3144         /*
3145          * If this is a permanent reservation ticket, we may be able to free
3146          * up more space based on the remaining count.
3147          */
3148         bytes = ticket->t_curr_res;
3149         if (ticket->t_cnt > 0) {
3150                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3151                 bytes += ticket->t_unit_res*ticket->t_cnt;
3152         }
3153
3154         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3155         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3156
3157         trace_xfs_log_ungrant_exit(log, ticket);
3158
3159         xfs_log_space_wake(log->l_mp);
3160 }
3161
3162 /*
3163  * Flush iclog to disk if this is the last reference to the given iclog and
3164  * the WANT_SYNC bit is set.
3165  *
3166  * When this function is entered, the iclog is not necessarily in the
3167  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3168  *
3169  *
3170  */
3171 STATIC int
3172 xlog_state_release_iclog(
3173         struct xlog             *log,
3174         struct xlog_in_core     *iclog)
3175 {
3176         int             sync = 0;       /* do we sync? */
3177
3178         if (iclog->ic_state & XLOG_STATE_IOERROR)
3179                 return -EIO;
3180
3181         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3182         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3183                 return 0;
3184
3185         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3186                 spin_unlock(&log->l_icloglock);
3187                 return -EIO;
3188         }
3189         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3190                iclog->ic_state == XLOG_STATE_WANT_SYNC);
3191
3192         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3193                 /* update tail before writing to iclog */
3194                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3195                 sync++;
3196                 iclog->ic_state = XLOG_STATE_SYNCING;
3197                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3198                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3199                 /* cycle incremented when incrementing curr_block */
3200         }
3201         spin_unlock(&log->l_icloglock);
3202
3203         /*
3204          * We let the log lock go, so it's possible that we hit a log I/O
3205          * error or some other SHUTDOWN condition that marks the iclog
3206          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3207          * this iclog has consistent data, so we ignore IOERROR
3208          * flags after this point.
3209          */
3210         if (sync)
3211                 return xlog_sync(log, iclog);
3212         return 0;
3213 }       /* xlog_state_release_iclog */
3214
3215
3216 /*
3217  * This routine will mark the current iclog in the ring as WANT_SYNC
3218  * and move the current iclog pointer to the next iclog in the ring.
3219  * When this routine is called from xlog_state_get_iclog_space(), the
3220  * exact size of the iclog has not yet been determined.  All we know is
3221  * that every data block.  We have run out of space in this log record.
3222  */
3223 STATIC void
3224 xlog_state_switch_iclogs(
3225         struct xlog             *log,
3226         struct xlog_in_core     *iclog,
3227         int                     eventual_size)
3228 {
3229         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3230         if (!eventual_size)
3231                 eventual_size = iclog->ic_offset;
3232         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3233         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3234         log->l_prev_block = log->l_curr_block;
3235         log->l_prev_cycle = log->l_curr_cycle;
3236
3237         /* roll log?: ic_offset changed later */
3238         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3239
3240         /* Round up to next log-sunit */
3241         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3242             log->l_mp->m_sb.sb_logsunit > 1) {
3243                 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3244                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3245         }
3246
3247         if (log->l_curr_block >= log->l_logBBsize) {
3248                 /*
3249                  * Rewind the current block before the cycle is bumped to make
3250                  * sure that the combined LSN never transiently moves forward
3251                  * when the log wraps to the next cycle. This is to support the
3252                  * unlocked sample of these fields from xlog_valid_lsn(). Most
3253                  * other cases should acquire l_icloglock.
3254                  */
3255                 log->l_curr_block -= log->l_logBBsize;
3256                 ASSERT(log->l_curr_block >= 0);
3257                 smp_wmb();
3258                 log->l_curr_cycle++;
3259                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3260                         log->l_curr_cycle++;
3261         }
3262         ASSERT(iclog == log->l_iclog);
3263         log->l_iclog = iclog->ic_next;
3264 }       /* xlog_state_switch_iclogs */
3265
3266 /*
3267  * Write out all data in the in-core log as of this exact moment in time.
3268  *
3269  * Data may be written to the in-core log during this call.  However,
3270  * we don't guarantee this data will be written out.  A change from past
3271  * implementation means this routine will *not* write out zero length LRs.
3272  *
3273  * Basically, we try and perform an intelligent scan of the in-core logs.
3274  * If we determine there is no flushable data, we just return.  There is no
3275  * flushable data if:
3276  *
3277  *      1. the current iclog is active and has no data; the previous iclog
3278  *              is in the active or dirty state.
3279  *      2. the current iclog is drity, and the previous iclog is in the
3280  *              active or dirty state.
3281  *
3282  * We may sleep if:
3283  *
3284  *      1. the current iclog is not in the active nor dirty state.
3285  *      2. the current iclog dirty, and the previous iclog is not in the
3286  *              active nor dirty state.
3287  *      3. the current iclog is active, and there is another thread writing
3288  *              to this particular iclog.
3289  *      4. a) the current iclog is active and has no other writers
3290  *         b) when we return from flushing out this iclog, it is still
3291  *              not in the active nor dirty state.
3292  */
3293 int
3294 xfs_log_force(
3295         struct xfs_mount        *mp,
3296         uint                    flags)
3297 {
3298         struct xlog             *log = mp->m_log;
3299         struct xlog_in_core     *iclog;
3300         xfs_lsn_t               lsn;
3301
3302         XFS_STATS_INC(mp, xs_log_force);
3303         trace_xfs_log_force(mp, 0, _RET_IP_);
3304
3305         xlog_cil_force(log);
3306
3307         spin_lock(&log->l_icloglock);
3308         iclog = log->l_iclog;
3309         if (iclog->ic_state & XLOG_STATE_IOERROR)
3310                 goto out_error;
3311
3312         if (iclog->ic_state == XLOG_STATE_DIRTY ||
3313             (iclog->ic_state == XLOG_STATE_ACTIVE &&
3314              atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3315                 /*
3316                  * If the head is dirty or (active and empty), then we need to
3317                  * look at the previous iclog.
3318                  *
3319                  * If the previous iclog is active or dirty we are done.  There
3320                  * is nothing to sync out. Otherwise, we attach ourselves to the
3321                  * previous iclog and go to sleep.
3322                  */
3323                 iclog = iclog->ic_prev;
3324                 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3325                     iclog->ic_state == XLOG_STATE_DIRTY)
3326                         goto out_unlock;
3327         } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3328                 if (atomic_read(&iclog->ic_refcnt) == 0) {
3329                         /*
3330                          * We are the only one with access to this iclog.
3331                          *
3332                          * Flush it out now.  There should be a roundoff of zero
3333                          * to show that someone has already taken care of the
3334                          * roundoff from the previous sync.
3335                          */
3336                         atomic_inc(&iclog->ic_refcnt);
3337                         lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3338                         xlog_state_switch_iclogs(log, iclog, 0);
3339                         spin_unlock(&log->l_icloglock);
3340
3341                         if (xlog_state_release_iclog(log, iclog))
3342                                 return -EIO;
3343
3344                         spin_lock(&log->l_icloglock);
3345                         if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3346                             iclog->ic_state == XLOG_STATE_DIRTY)
3347                                 goto out_unlock;
3348                 } else {
3349                         /*
3350                          * Someone else is writing to this iclog.
3351                          *
3352                          * Use its call to flush out the data.  However, the
3353                          * other thread may not force out this LR, so we mark
3354                          * it WANT_SYNC.
3355                          */
3356                         xlog_state_switch_iclogs(log, iclog, 0);
3357                 }
3358         } else {
3359                 /*
3360                  * If the head iclog is not active nor dirty, we just attach
3361                  * ourselves to the head and go to sleep if necessary.
3362                  */
3363                 ;
3364         }
3365
3366         if (!(flags & XFS_LOG_SYNC))
3367                 goto out_unlock;
3368
3369         if (iclog->ic_state & XLOG_STATE_IOERROR)
3370                 goto out_error;
3371         XFS_STATS_INC(mp, xs_log_force_sleep);
3372         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3373         if (iclog->ic_state & XLOG_STATE_IOERROR)
3374                 return -EIO;
3375         return 0;
3376
3377 out_unlock:
3378         spin_unlock(&log->l_icloglock);
3379         return 0;
3380 out_error:
3381         spin_unlock(&log->l_icloglock);
3382         return -EIO;
3383 }
3384
3385 static int
3386 __xfs_log_force_lsn(
3387         struct xfs_mount        *mp,
3388         xfs_lsn_t               lsn,
3389         uint                    flags,
3390         int                     *log_flushed,
3391         bool                    already_slept)
3392 {
3393         struct xlog             *log = mp->m_log;
3394         struct xlog_in_core     *iclog;
3395
3396         spin_lock(&log->l_icloglock);
3397         iclog = log->l_iclog;
3398         if (iclog->ic_state & XLOG_STATE_IOERROR)
3399                 goto out_error;
3400
3401         while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3402                 iclog = iclog->ic_next;
3403                 if (iclog == log->l_iclog)
3404                         goto out_unlock;
3405         }
3406
3407         if (iclog->ic_state == XLOG_STATE_DIRTY)
3408                 goto out_unlock;
3409
3410         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3411                 /*
3412                  * We sleep here if we haven't already slept (e.g. this is the
3413                  * first time we've looked at the correct iclog buf) and the
3414                  * buffer before us is going to be sync'ed.  The reason for this
3415                  * is that if we are doing sync transactions here, by waiting
3416                  * for the previous I/O to complete, we can allow a few more
3417                  * transactions into this iclog before we close it down.
3418                  *
3419                  * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3420                  * refcnt so we can release the log (which drops the ref count).
3421                  * The state switch keeps new transaction commits from using
3422                  * this buffer.  When the current commits finish writing into
3423                  * the buffer, the refcount will drop to zero and the buffer
3424                  * will go out then.
3425                  */
3426                 if (!already_slept &&
3427                     (iclog->ic_prev->ic_state &
3428                      (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3429                         ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3430
3431                         XFS_STATS_INC(mp, xs_log_force_sleep);
3432
3433                         xlog_wait(&iclog->ic_prev->ic_write_wait,
3434                                         &log->l_icloglock);
3435                         return -EAGAIN;
3436                 }
3437                 atomic_inc(&iclog->ic_refcnt);
3438                 xlog_state_switch_iclogs(log, iclog, 0);
3439                 spin_unlock(&log->l_icloglock);
3440                 if (xlog_state_release_iclog(log, iclog))
3441                         return -EIO;
3442                 if (log_flushed)
3443                         *log_flushed = 1;
3444                 spin_lock(&log->l_icloglock);
3445         }
3446
3447         if (!(flags & XFS_LOG_SYNC) ||
3448             (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3449                 goto out_unlock;
3450
3451         if (iclog->ic_state & XLOG_STATE_IOERROR)
3452                 goto out_error;
3453
3454         XFS_STATS_INC(mp, xs_log_force_sleep);
3455         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3456         if (iclog->ic_state & XLOG_STATE_IOERROR)
3457                 return -EIO;
3458         return 0;
3459
3460 out_unlock:
3461         spin_unlock(&log->l_icloglock);
3462         return 0;
3463 out_error:
3464         spin_unlock(&log->l_icloglock);
3465         return -EIO;
3466 }
3467
3468 /*
3469  * Force the in-core log to disk for a specific LSN.
3470  *
3471  * Find in-core log with lsn.
3472  *      If it is in the DIRTY state, just return.
3473  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3474  *              state and go to sleep or return.
3475  *      If it is in any other state, go to sleep or return.
3476  *
3477  * Synchronous forces are implemented with a wait queue.  All callers trying
3478  * to force a given lsn to disk must wait on the queue attached to the
3479  * specific in-core log.  When given in-core log finally completes its write
3480  * to disk, that thread will wake up all threads waiting on the queue.
3481  */
3482 int
3483 xfs_log_force_lsn(
3484         struct xfs_mount        *mp,
3485         xfs_lsn_t               lsn,
3486         uint                    flags,
3487         int                     *log_flushed)
3488 {
3489         int                     ret;
3490         ASSERT(lsn != 0);
3491
3492         XFS_STATS_INC(mp, xs_log_force);
3493         trace_xfs_log_force(mp, lsn, _RET_IP_);
3494
3495         lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3496         if (lsn == NULLCOMMITLSN)
3497                 return 0;
3498
3499         ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3500         if (ret == -EAGAIN)
3501                 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3502         return ret;
3503 }
3504
3505 /*
3506  * Called when we want to mark the current iclog as being ready to sync to
3507  * disk.
3508  */
3509 STATIC void
3510 xlog_state_want_sync(
3511         struct xlog             *log,
3512         struct xlog_in_core     *iclog)
3513 {
3514         assert_spin_locked(&log->l_icloglock);
3515
3516         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3517                 xlog_state_switch_iclogs(log, iclog, 0);
3518         } else {
3519                 ASSERT(iclog->ic_state &
3520                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3521         }
3522 }
3523
3524
3525 /*****************************************************************************
3526  *
3527  *              TICKET functions
3528  *
3529  *****************************************************************************
3530  */
3531
3532 /*
3533  * Free a used ticket when its refcount falls to zero.
3534  */
3535 void
3536 xfs_log_ticket_put(
3537         xlog_ticket_t   *ticket)
3538 {
3539         ASSERT(atomic_read(&ticket->t_ref) > 0);
3540         if (atomic_dec_and_test(&ticket->t_ref))
3541                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3542 }
3543
3544 xlog_ticket_t *
3545 xfs_log_ticket_get(
3546         xlog_ticket_t   *ticket)
3547 {
3548         ASSERT(atomic_read(&ticket->t_ref) > 0);
3549         atomic_inc(&ticket->t_ref);
3550         return ticket;
3551 }
3552
3553 /*
3554  * Figure out the total log space unit (in bytes) that would be
3555  * required for a log ticket.
3556  */
3557 int
3558 xfs_log_calc_unit_res(
3559         struct xfs_mount        *mp,
3560         int                     unit_bytes)
3561 {
3562         struct xlog             *log = mp->m_log;
3563         int                     iclog_space;
3564         uint                    num_headers;
3565
3566         /*
3567          * Permanent reservations have up to 'cnt'-1 active log operations
3568          * in the log.  A unit in this case is the amount of space for one
3569          * of these log operations.  Normal reservations have a cnt of 1
3570          * and their unit amount is the total amount of space required.
3571          *
3572          * The following lines of code account for non-transaction data
3573          * which occupy space in the on-disk log.
3574          *
3575          * Normal form of a transaction is:
3576          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3577          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3578          *
3579          * We need to account for all the leadup data and trailer data
3580          * around the transaction data.
3581          * And then we need to account for the worst case in terms of using
3582          * more space.
3583          * The worst case will happen if:
3584          * - the placement of the transaction happens to be such that the
3585          *   roundoff is at its maximum
3586          * - the transaction data is synced before the commit record is synced
3587          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3588          *   Therefore the commit record is in its own Log Record.
3589          *   This can happen as the commit record is called with its
3590          *   own region to xlog_write().
3591          *   This then means that in the worst case, roundoff can happen for
3592          *   the commit-rec as well.
3593          *   The commit-rec is smaller than padding in this scenario and so it is
3594          *   not added separately.
3595          */
3596
3597         /* for trans header */
3598         unit_bytes += sizeof(xlog_op_header_t);
3599         unit_bytes += sizeof(xfs_trans_header_t);
3600
3601         /* for start-rec */
3602         unit_bytes += sizeof(xlog_op_header_t);
3603
3604         /*
3605          * for LR headers - the space for data in an iclog is the size minus
3606          * the space used for the headers. If we use the iclog size, then we
3607          * undercalculate the number of headers required.
3608          *
3609          * Furthermore - the addition of op headers for split-recs might
3610          * increase the space required enough to require more log and op
3611          * headers, so take that into account too.
3612          *
3613          * IMPORTANT: This reservation makes the assumption that if this
3614          * transaction is the first in an iclog and hence has the LR headers
3615          * accounted to it, then the remaining space in the iclog is
3616          * exclusively for this transaction.  i.e. if the transaction is larger
3617          * than the iclog, it will be the only thing in that iclog.
3618          * Fundamentally, this means we must pass the entire log vector to
3619          * xlog_write to guarantee this.
3620          */
3621         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3622         num_headers = howmany(unit_bytes, iclog_space);
3623
3624         /* for split-recs - ophdrs added when data split over LRs */
3625         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3626
3627         /* add extra header reservations if we overrun */
3628         while (!num_headers ||
3629                howmany(unit_bytes, iclog_space) > num_headers) {
3630                 unit_bytes += sizeof(xlog_op_header_t);
3631                 num_headers++;
3632         }
3633         unit_bytes += log->l_iclog_hsize * num_headers;
3634
3635         /* for commit-rec LR header - note: padding will subsume the ophdr */
3636         unit_bytes += log->l_iclog_hsize;
3637
3638         /* for roundoff padding for transaction data and one for commit record */
3639         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3640                 /* log su roundoff */
3641                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3642         } else {
3643                 /* BB roundoff */
3644                 unit_bytes += 2 * BBSIZE;
3645         }
3646
3647         return unit_bytes;
3648 }
3649
3650 /*
3651  * Allocate and initialise a new log ticket.
3652  */
3653 struct xlog_ticket *
3654 xlog_ticket_alloc(
3655         struct xlog             *log,
3656         int                     unit_bytes,
3657         int                     cnt,
3658         char                    client,
3659         bool                    permanent,
3660         xfs_km_flags_t          alloc_flags)
3661 {
3662         struct xlog_ticket      *tic;
3663         int                     unit_res;
3664
3665         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3666         if (!tic)
3667                 return NULL;
3668
3669         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3670
3671         atomic_set(&tic->t_ref, 1);
3672         tic->t_task             = current;
3673         INIT_LIST_HEAD(&tic->t_queue);
3674         tic->t_unit_res         = unit_res;
3675         tic->t_curr_res         = unit_res;
3676         tic->t_cnt              = cnt;
3677         tic->t_ocnt             = cnt;
3678         tic->t_tid              = prandom_u32();
3679         tic->t_clientid         = client;
3680         tic->t_flags            = XLOG_TIC_INITED;
3681         if (permanent)
3682                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3683
3684         xlog_tic_reset_res(tic);
3685
3686         return tic;
3687 }
3688
3689
3690 /******************************************************************************
3691  *
3692  *              Log debug routines
3693  *
3694  ******************************************************************************
3695  */
3696 #if defined(DEBUG)
3697 /*
3698  * Make sure that the destination ptr is within the valid data region of
3699  * one of the iclogs.  This uses backup pointers stored in a different
3700  * part of the log in case we trash the log structure.
3701  */
3702 STATIC void
3703 xlog_verify_dest_ptr(
3704         struct xlog     *log,
3705         void            *ptr)
3706 {
3707         int i;
3708         int good_ptr = 0;
3709
3710         for (i = 0; i < log->l_iclog_bufs; i++) {
3711                 if (ptr >= log->l_iclog_bak[i] &&
3712                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3713                         good_ptr++;
3714         }
3715
3716         if (!good_ptr)
3717                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3718 }
3719
3720 /*
3721  * Check to make sure the grant write head didn't just over lap the tail.  If
3722  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3723  * the cycles differ by exactly one and check the byte count.
3724  *
3725  * This check is run unlocked, so can give false positives. Rather than assert
3726  * on failures, use a warn-once flag and a panic tag to allow the admin to
3727  * determine if they want to panic the machine when such an error occurs. For
3728  * debug kernels this will have the same effect as using an assert but, unlinke
3729  * an assert, it can be turned off at runtime.
3730  */
3731 STATIC void
3732 xlog_verify_grant_tail(
3733         struct xlog     *log)
3734 {
3735         int             tail_cycle, tail_blocks;
3736         int             cycle, space;
3737
3738         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3739         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3740         if (tail_cycle != cycle) {
3741                 if (cycle - 1 != tail_cycle &&
3742                     !(log->l_flags & XLOG_TAIL_WARN)) {
3743                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3744                                 "%s: cycle - 1 != tail_cycle", __func__);
3745                         log->l_flags |= XLOG_TAIL_WARN;
3746                 }
3747
3748                 if (space > BBTOB(tail_blocks) &&
3749                     !(log->l_flags & XLOG_TAIL_WARN)) {
3750                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3751                                 "%s: space > BBTOB(tail_blocks)", __func__);
3752                         log->l_flags |= XLOG_TAIL_WARN;
3753                 }
3754         }
3755 }
3756
3757 /* check if it will fit */
3758 STATIC void
3759 xlog_verify_tail_lsn(
3760         struct xlog             *log,
3761         struct xlog_in_core     *iclog,
3762         xfs_lsn_t               tail_lsn)
3763 {
3764     int blocks;
3765
3766     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3767         blocks =
3768             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3769         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3770                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3771     } else {
3772         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3773
3774         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3775                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3776
3777         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3778         if (blocks < BTOBB(iclog->ic_offset) + 1)
3779                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3780     }
3781 }       /* xlog_verify_tail_lsn */
3782
3783 /*
3784  * Perform a number of checks on the iclog before writing to disk.
3785  *
3786  * 1. Make sure the iclogs are still circular
3787  * 2. Make sure we have a good magic number
3788  * 3. Make sure we don't have magic numbers in the data
3789  * 4. Check fields of each log operation header for:
3790  *      A. Valid client identifier
3791  *      B. tid ptr value falls in valid ptr space (user space code)
3792  *      C. Length in log record header is correct according to the
3793  *              individual operation headers within record.
3794  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3795  *      log, check the preceding blocks of the physical log to make sure all
3796  *      the cycle numbers agree with the current cycle number.
3797  */
3798 STATIC void
3799 xlog_verify_iclog(
3800         struct xlog             *log,
3801         struct xlog_in_core     *iclog,
3802         int                     count,
3803         bool                    syncing)
3804 {
3805         xlog_op_header_t        *ophead;
3806         xlog_in_core_t          *icptr;
3807         xlog_in_core_2_t        *xhdr;
3808         void                    *base_ptr, *ptr, *p;
3809         ptrdiff_t               field_offset;
3810         uint8_t                 clientid;
3811         int                     len, i, j, k, op_len;
3812         int                     idx;
3813
3814         /* check validity of iclog pointers */
3815         spin_lock(&log->l_icloglock);
3816         icptr = log->l_iclog;
3817         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3818                 ASSERT(icptr);
3819
3820         if (icptr != log->l_iclog)
3821                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3822         spin_unlock(&log->l_icloglock);
3823
3824         /* check log magic numbers */
3825         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3826                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3827
3828         base_ptr = ptr = &iclog->ic_header;
3829         p = &iclog->ic_header;
3830         for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3831                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3832                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3833                                 __func__);
3834         }
3835
3836         /* check fields */
3837         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3838         base_ptr = ptr = iclog->ic_datap;
3839         ophead = ptr;
3840         xhdr = iclog->ic_data;
3841         for (i = 0; i < len; i++) {
3842                 ophead = ptr;
3843
3844                 /* clientid is only 1 byte */
3845                 p = &ophead->oh_clientid;
3846                 field_offset = p - base_ptr;
3847                 if (!syncing || (field_offset & 0x1ff)) {
3848                         clientid = ophead->oh_clientid;
3849                 } else {
3850                         idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3851                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3852                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3853                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3854                                 clientid = xlog_get_client_id(
3855                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3856                         } else {
3857                                 clientid = xlog_get_client_id(
3858                                         iclog->ic_header.h_cycle_data[idx]);
3859                         }
3860                 }
3861                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3862                         xfs_warn(log->l_mp,
3863                                 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3864                                 __func__, clientid, ophead,
3865                                 (unsigned long)field_offset);
3866
3867                 /* check length */
3868                 p = &ophead->oh_len;
3869                 field_offset = p - base_ptr;
3870                 if (!syncing || (field_offset & 0x1ff)) {
3871                         op_len = be32_to_cpu(ophead->oh_len);
3872                 } else {
3873                         idx = BTOBBT((uintptr_t)&ophead->oh_len -
3874                                     (uintptr_t)iclog->ic_datap);
3875                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3876                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3877                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3878                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3879                         } else {
3880                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3881                         }
3882                 }
3883                 ptr += sizeof(xlog_op_header_t) + op_len;
3884         }
3885 }       /* xlog_verify_iclog */
3886 #endif
3887
3888 /*
3889  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3890  */
3891 STATIC int
3892 xlog_state_ioerror(
3893         struct xlog     *log)
3894 {
3895         xlog_in_core_t  *iclog, *ic;
3896
3897         iclog = log->l_iclog;
3898         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3899                 /*
3900                  * Mark all the incore logs IOERROR.
3901                  * From now on, no log flushes will result.
3902                  */
3903                 ic = iclog;
3904                 do {
3905                         ic->ic_state = XLOG_STATE_IOERROR;
3906                         ic = ic->ic_next;
3907                 } while (ic != iclog);
3908                 return 0;
3909         }
3910         /*
3911          * Return non-zero, if state transition has already happened.
3912          */
3913         return 1;
3914 }
3915
3916 /*
3917  * This is called from xfs_force_shutdown, when we're forcibly
3918  * shutting down the filesystem, typically because of an IO error.
3919  * Our main objectives here are to make sure that:
3920  *      a. if !logerror, flush the logs to disk. Anything modified
3921  *         after this is ignored.
3922  *      b. the filesystem gets marked 'SHUTDOWN' for all interested
3923  *         parties to find out, 'atomically'.
3924  *      c. those who're sleeping on log reservations, pinned objects and
3925  *          other resources get woken up, and be told the bad news.
3926  *      d. nothing new gets queued up after (b) and (c) are done.
3927  *
3928  * Note: for the !logerror case we need to flush the regions held in memory out
3929  * to disk first. This needs to be done before the log is marked as shutdown,
3930  * otherwise the iclog writes will fail.
3931  */
3932 int
3933 xfs_log_force_umount(
3934         struct xfs_mount        *mp,
3935         int                     logerror)
3936 {
3937         struct xlog     *log;
3938         int             retval;
3939
3940         log = mp->m_log;
3941
3942         /*
3943          * If this happens during log recovery, don't worry about
3944          * locking; the log isn't open for business yet.
3945          */
3946         if (!log ||
3947             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3948                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3949                 if (mp->m_sb_bp)
3950                         mp->m_sb_bp->b_flags |= XBF_DONE;
3951                 return 0;
3952         }
3953
3954         /*
3955          * Somebody could've already done the hard work for us.
3956          * No need to get locks for this.
3957          */
3958         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3959                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3960                 return 1;
3961         }
3962
3963         /*
3964          * Flush all the completed transactions to disk before marking the log
3965          * being shut down. We need to do it in this order to ensure that
3966          * completed operations are safely on disk before we shut down, and that
3967          * we don't have to issue any buffer IO after the shutdown flags are set
3968          * to guarantee this.
3969          */
3970         if (!logerror)
3971                 xfs_log_force(mp, XFS_LOG_SYNC);
3972
3973         /*
3974          * mark the filesystem and the as in a shutdown state and wake
3975          * everybody up to tell them the bad news.
3976          */
3977         spin_lock(&log->l_icloglock);
3978         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3979         if (mp->m_sb_bp)
3980                 mp->m_sb_bp->b_flags |= XBF_DONE;
3981
3982         /*
3983          * Mark the log and the iclogs with IO error flags to prevent any
3984          * further log IO from being issued or completed.
3985          */
3986         log->l_flags |= XLOG_IO_ERROR;
3987         retval = xlog_state_ioerror(log);
3988         spin_unlock(&log->l_icloglock);
3989
3990         /*
3991          * We don't want anybody waiting for log reservations after this. That
3992          * means we have to wake up everybody queued up on reserveq as well as
3993          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3994          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3995          * action is protected by the grant locks.
3996          */
3997         xlog_grant_head_wake_all(&log->l_reserve_head);
3998         xlog_grant_head_wake_all(&log->l_write_head);
3999
4000         /*
4001          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4002          * as if the log writes were completed. The abort handling in the log
4003          * item committed callback functions will do this again under lock to
4004          * avoid races.
4005          */
4006         wake_up_all(&log->l_cilp->xc_commit_wait);
4007         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
4008
4009 #ifdef XFSERRORDEBUG
4010         {
4011                 xlog_in_core_t  *iclog;
4012
4013                 spin_lock(&log->l_icloglock);
4014                 iclog = log->l_iclog;
4015                 do {
4016                         ASSERT(iclog->ic_callback == 0);
4017                         iclog = iclog->ic_next;
4018                 } while (iclog != log->l_iclog);
4019                 spin_unlock(&log->l_icloglock);
4020         }
4021 #endif
4022         /* return non-zero if log IOERROR transition had already happened */
4023         return retval;
4024 }
4025
4026 STATIC int
4027 xlog_iclogs_empty(
4028         struct xlog     *log)
4029 {
4030         xlog_in_core_t  *iclog;
4031
4032         iclog = log->l_iclog;
4033         do {
4034                 /* endianness does not matter here, zero is zero in
4035                  * any language.
4036                  */
4037                 if (iclog->ic_header.h_num_logops)
4038                         return 0;
4039                 iclog = iclog->ic_next;
4040         } while (iclog != log->l_iclog);
4041         return 1;
4042 }
4043
4044 /*
4045  * Verify that an LSN stamped into a piece of metadata is valid. This is
4046  * intended for use in read verifiers on v5 superblocks.
4047  */
4048 bool
4049 xfs_log_check_lsn(
4050         struct xfs_mount        *mp,
4051         xfs_lsn_t               lsn)
4052 {
4053         struct xlog             *log = mp->m_log;
4054         bool                    valid;
4055
4056         /*
4057          * norecovery mode skips mount-time log processing and unconditionally
4058          * resets the in-core LSN. We can't validate in this mode, but
4059          * modifications are not allowed anyways so just return true.
4060          */
4061         if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4062                 return true;
4063
4064         /*
4065          * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4066          * handled by recovery and thus safe to ignore here.
4067          */
4068         if (lsn == NULLCOMMITLSN)
4069                 return true;
4070
4071         valid = xlog_valid_lsn(mp->m_log, lsn);
4072
4073         /* warn the user about what's gone wrong before verifier failure */
4074         if (!valid) {
4075                 spin_lock(&log->l_icloglock);
4076                 xfs_warn(mp,
4077 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4078 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4079                          CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4080                          log->l_curr_cycle, log->l_curr_block);
4081                 spin_unlock(&log->l_icloglock);
4082         }
4083
4084         return valid;
4085 }