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 to 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 /* Actually write the unmount record to disk. */
 830 static void
 831 xfs_log_write_unmount_record(
 832         struct xfs_mount        *mp)
 833 {
 834         /* the data section must be 32 bit size aligned */
 835         struct xfs_unmount_log_format magic = {
 836                 .magic = XLOG_UNMOUNT_TYPE,
 837         };
 838         struct xfs_log_iovec reg = {
 839                 .i_addr = &magic,
 840                 .i_len = sizeof(magic),
 841                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 842         };
 843         struct xfs_log_vec vec = {
 844                 .lv_niovecs = 1,
 845                 .lv_iovecp = &reg,
 846         };
 847         struct xlog             *log = mp->m_log;
 848         struct xlog_in_core     *iclog;
 849         struct xlog_ticket      *tic = NULL;
 850         xfs_lsn_t               lsn;
 851         uint                    flags = XLOG_UNMOUNT_TRANS;
 852         int                     error;
 853
 854         error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
 855         if (error)
 856                 goto out_err;
 857
 858         /*
 859          * If we think the summary counters are bad, clear the unmount header
 860          * flag in the unmount record so that the summary counters will be
 861          * recalculated during log recovery at next mount.  Refer to
 862          * xlog_check_unmount_rec for more details.
 863          */
 864         if (XFS_TEST_ERROR((mp->m_flags & XFS_MOUNT_BAD_SUMMARY), mp,
 865                         XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
 866                 xfs_alert(mp, "%s: will fix summary counters at next mount",
 867                                 __func__);
 868                 flags &= ~XLOG_UNMOUNT_TRANS;
 869         }
 870
 871         /* remove inited flag, and account for space used */
 872         tic->t_flags = 0;
 873         tic->t_curr_res -= sizeof(magic);
 874         error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
 875         /*
 876          * At this point, we're umounting anyway, so there's no point in
 877          * transitioning log state to IOERROR. Just continue...
 878          */
 879 out_err:
 880         if (error)
 881                 xfs_alert(mp, "%s: unmount record failed", __func__);
 882
 883         spin_lock(&log->l_icloglock);
 884         iclog = log->l_iclog;
 885         atomic_inc(&iclog->ic_refcnt);
 886         xlog_state_want_sync(log, iclog);
 887         spin_unlock(&log->l_icloglock);
 888         error = xlog_state_release_iclog(log, iclog);
 889
 890         spin_lock(&log->l_icloglock);
 891         switch (iclog->ic_state) {
 892         default:
 893                 if (!XLOG_FORCED_SHUTDOWN(log)) {
 894                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
 895                         break;
 896                 }
 897                 /* fall through */
 898         case XLOG_STATE_ACTIVE:
 899         case XLOG_STATE_DIRTY:
 900                 spin_unlock(&log->l_icloglock);
 901                 break;
 902         }
 903
 904         if (tic) {
 905                 trace_xfs_log_umount_write(log, tic);
 906                 xlog_ungrant_log_space(log, tic);
 907                 xfs_log_ticket_put(tic);
 908         }
 909 }
 910
 911 /*
 912  * Unmount record used to have a string "Unmount filesystem--" in the
 913  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 914  * We just write the magic number now since that particular field isn't
 915  * currently architecture converted and "Unmount" is a bit foo.
 916  * As far as I know, there weren't any dependencies on the old behaviour.
 917  */
 918
 919 static int
 920 xfs_log_unmount_write(xfs_mount_t *mp)
 921 {
 922         struct xlog      *log = mp->m_log;
 923         xlog_in_core_t   *iclog;
 924 #ifdef DEBUG
 925         xlog_in_core_t   *first_iclog;
 926 #endif
 927         int              error;
 928
 929         /*
 930          * Don't write out unmount record on norecovery mounts or ro devices.
 931          * Or, if we are doing a forced umount (typically because of IO errors).
 932          */
 933         if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
 934             xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
 935                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 936                 return 0;
 937         }
 938
 939         error = xfs_log_force(mp, XFS_LOG_SYNC);
 940         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 941
 942 #ifdef DEBUG
 943         first_iclog = iclog = log->l_iclog;
 944         do {
 945                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 946                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 947                         ASSERT(iclog->ic_offset == 0);
 948                 }
 949                 iclog = iclog->ic_next;
 950         } while (iclog != first_iclog);
 951 #endif
 952         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 953                 xfs_log_write_unmount_record(mp);
 954         } else {
 955                 /*
 956                  * We're already in forced_shutdown mode, couldn't
 957                  * even attempt to write out the unmount transaction.
 958                  *
 959                  * Go through the motions of sync'ing and releasing
 960                  * the iclog, even though no I/O will actually happen,
 961                  * we need to wait for other log I/Os that may already
 962                  * be in progress.  Do this as a separate section of
 963                  * code so we'll know if we ever get stuck here that
 964                  * we're in this odd situation of trying to unmount
 965                  * a file system that went into forced_shutdown as
 966                  * the result of an unmount..
 967                  */
 968                 spin_lock(&log->l_icloglock);
 969                 iclog = log->l_iclog;
 970                 atomic_inc(&iclog->ic_refcnt);
 971
 972                 xlog_state_want_sync(log, iclog);
 973                 spin_unlock(&log->l_icloglock);
 974                 error =  xlog_state_release_iclog(log, iclog);
 975
 976                 spin_lock(&log->l_icloglock);
 977
 978                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 979                         || iclog->ic_state == XLOG_STATE_DIRTY
 980                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 981
 982                                 xlog_wait(&iclog->ic_force_wait,
 983                                                         &log->l_icloglock);
 984                 } else {
 985                         spin_unlock(&log->l_icloglock);
 986                 }
 987         }
 988
 989         return error;
 990 }       /* xfs_log_unmount_write */
 991
 992 /*
 993  * Empty the log for unmount/freeze.
 994  *
 995  * To do this, we first need to shut down the background log work so it is not
 996  * trying to cover the log as we clean up. We then need to unpin all objects in
 997  * the log so we can then flush them out. Once they have completed their IO and
 998  * run the callbacks removing themselves from the AIL, we can write the unmount
 999  * record.
1000  */
1001 void
1002 xfs_log_quiesce(
1003         struct xfs_mount        *mp)
1004 {
1005         cancel_delayed_work_sync(&mp->m_log->l_work);
1006         xfs_log_force(mp, XFS_LOG_SYNC);
1007
1008         /*
1009          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1010          * will push it, xfs_wait_buftarg() will not wait for it. Further,
1011          * xfs_buf_iowait() cannot be used because it was pushed with the
1012          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1013          * the IO to complete.
1014          */
1015         xfs_ail_push_all_sync(mp->m_ail);
1016         xfs_wait_buftarg(mp->m_ddev_targp);
1017         xfs_buf_lock(mp->m_sb_bp);
1018         xfs_buf_unlock(mp->m_sb_bp);
1019
1020         xfs_log_unmount_write(mp);
1021 }
1022
1023 /*
1024  * Shut down and release the AIL and Log.
1025  *
1026  * During unmount, we need to ensure we flush all the dirty metadata objects
1027  * from the AIL so that the log is empty before we write the unmount record to
1028  * the log. Once this is done, we can tear down the AIL and the log.
1029  */
1030 void
1031 xfs_log_unmount(
1032         struct xfs_mount        *mp)
1033 {
1034         xfs_log_quiesce(mp);
1035
1036         xfs_trans_ail_destroy(mp);
1037
1038         xfs_sysfs_del(&mp->m_log->l_kobj);
1039
1040         xlog_dealloc_log(mp->m_log);
1041 }
1042
1043 void
1044 xfs_log_item_init(
1045         struct xfs_mount        *mp,
1046         struct xfs_log_item     *item,
1047         int                     type,
1048         const struct xfs_item_ops *ops)
1049 {
1050         item->li_mountp = mp;
1051         item->li_ailp = mp->m_ail;
1052         item->li_type = type;
1053         item->li_ops = ops;
1054         item->li_lv = NULL;
1055
1056         INIT_LIST_HEAD(&item->li_ail);
1057         INIT_LIST_HEAD(&item->li_cil);
1058         INIT_LIST_HEAD(&item->li_bio_list);
1059         INIT_LIST_HEAD(&item->li_trans);
1060 }
1061
1062 /*
1063  * Wake up processes waiting for log space after we have moved the log tail.
1064  */
1065 void
1066 xfs_log_space_wake(
1067         struct xfs_mount        *mp)
1068 {
1069         struct xlog             *log = mp->m_log;
1070         int                     free_bytes;
1071
1072         if (XLOG_FORCED_SHUTDOWN(log))
1073                 return;
1074
1075         if (!list_empty_careful(&log->l_write_head.waiters)) {
1076                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1077
1078                 spin_lock(&log->l_write_head.lock);
1079                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1080                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1081                 spin_unlock(&log->l_write_head.lock);
1082         }
1083
1084         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1085                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1086
1087                 spin_lock(&log->l_reserve_head.lock);
1088                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1089                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1090                 spin_unlock(&log->l_reserve_head.lock);
1091         }
1092 }
1093
1094 /*
1095  * Determine if we have a transaction that has gone to disk that needs to be
1096  * covered. To begin the transition to the idle state firstly the log needs to
1097  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1098  * we start attempting to cover the log.
1099  *
1100  * Only if we are then in a state where covering is needed, the caller is
1101  * informed that dummy transactions are required to move the log into the idle
1102  * state.
1103  *
1104  * If there are any items in the AIl or CIL, then we do not want to attempt to
1105  * cover the log as we may be in a situation where there isn't log space
1106  * available to run a dummy transaction and this can lead to deadlocks when the
1107  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1108  * there's no point in running a dummy transaction at this point because we
1109  * can't start trying to idle the log until both the CIL and AIL are empty.
1110  */
1111 static int
1112 xfs_log_need_covered(xfs_mount_t *mp)
1113 {
1114         struct xlog     *log = mp->m_log;
1115         int             needed = 0;
1116
1117         if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1118                 return 0;
1119
1120         if (!xlog_cil_empty(log))
1121                 return 0;
1122
1123         spin_lock(&log->l_icloglock);
1124         switch (log->l_covered_state) {
1125         case XLOG_STATE_COVER_DONE:
1126         case XLOG_STATE_COVER_DONE2:
1127         case XLOG_STATE_COVER_IDLE:
1128                 break;
1129         case XLOG_STATE_COVER_NEED:
1130         case XLOG_STATE_COVER_NEED2:
1131                 if (xfs_ail_min_lsn(log->l_ailp))
1132                         break;
1133                 if (!xlog_iclogs_empty(log))
1134                         break;
1135
1136                 needed = 1;
1137                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1138                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1139                 else
1140                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1141                 break;
1142         default:
1143                 needed = 1;
1144                 break;
1145         }
1146         spin_unlock(&log->l_icloglock);
1147         return needed;
1148 }
1149
1150 /*
1151  * We may be holding the log iclog lock upon entering this routine.
1152  */
1153 xfs_lsn_t
1154 xlog_assign_tail_lsn_locked(
1155         struct xfs_mount        *mp)
1156 {
1157         struct xlog             *log = mp->m_log;
1158         struct xfs_log_item     *lip;
1159         xfs_lsn_t               tail_lsn;
1160
1161         assert_spin_locked(&mp->m_ail->ail_lock);
1162
1163         /*
1164          * To make sure we always have a valid LSN for the log tail we keep
1165          * track of the last LSN which was committed in log->l_last_sync_lsn,
1166          * and use that when the AIL was empty.
1167          */
1168         lip = xfs_ail_min(mp->m_ail);
1169         if (lip)
1170                 tail_lsn = lip->li_lsn;
1171         else
1172                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1173         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1174         atomic64_set(&log->l_tail_lsn, tail_lsn);
1175         return tail_lsn;
1176 }
1177
1178 xfs_lsn_t
1179 xlog_assign_tail_lsn(
1180         struct xfs_mount        *mp)
1181 {
1182         xfs_lsn_t               tail_lsn;
1183
1184         spin_lock(&mp->m_ail->ail_lock);
1185         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1186         spin_unlock(&mp->m_ail->ail_lock);
1187
1188         return tail_lsn;
1189 }
1190
1191 /*
1192  * Return the space in the log between the tail and the head.  The head
1193  * is passed in the cycle/bytes formal parms.  In the special case where
1194  * the reserve head has wrapped passed the tail, this calculation is no
1195  * longer valid.  In this case, just return 0 which means there is no space
1196  * in the log.  This works for all places where this function is called
1197  * with the reserve head.  Of course, if the write head were to ever
1198  * wrap the tail, we should blow up.  Rather than catch this case here,
1199  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1200  *
1201  * This code also handles the case where the reservation head is behind
1202  * the tail.  The details of this case are described below, but the end
1203  * result is that we return the size of the log as the amount of space left.
1204  */
1205 STATIC int
1206 xlog_space_left(
1207         struct xlog     *log,
1208         atomic64_t      *head)
1209 {
1210         int             free_bytes;
1211         int             tail_bytes;
1212         int             tail_cycle;
1213         int             head_cycle;
1214         int             head_bytes;
1215
1216         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1217         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1218         tail_bytes = BBTOB(tail_bytes);
1219         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1220                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1221         else if (tail_cycle + 1 < head_cycle)
1222                 return 0;
1223         else if (tail_cycle < head_cycle) {
1224                 ASSERT(tail_cycle == (head_cycle - 1));
1225                 free_bytes = tail_bytes - head_bytes;
1226         } else {
1227                 /*
1228                  * The reservation head is behind the tail.
1229                  * In this case we just want to return the size of the
1230                  * log as the amount of space left.
1231                  */
1232                 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1233                 xfs_alert(log->l_mp,
1234                           "  tail_cycle = %d, tail_bytes = %d",
1235                           tail_cycle, tail_bytes);
1236                 xfs_alert(log->l_mp,
1237                           "  GH   cycle = %d, GH   bytes = %d",
1238                           head_cycle, head_bytes);
1239                 ASSERT(0);
1240                 free_bytes = log->l_logsize;
1241         }
1242         return free_bytes;
1243 }
1244
1245
1246 /*
1247  * Log function which is called when an io completes.
1248  *
1249  * The log manager needs its own routine, in order to control what
1250  * happens with the buffer after the write completes.
1251  */
1252 static void
1253 xlog_iodone(xfs_buf_t *bp)
1254 {
1255         struct xlog_in_core     *iclog = bp->b_log_item;
1256         struct xlog             *l = iclog->ic_log;
1257         int                     aborted = 0;
1258
1259         /*
1260          * Race to shutdown the filesystem if we see an error or the iclog is in
1261          * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
1262          * CRC errors into log recovery.
1263          */
1264         if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR) ||
1265             iclog->ic_state & XLOG_STATE_IOABORT) {
1266                 if (iclog->ic_state & XLOG_STATE_IOABORT)
1267                         iclog->ic_state &= ~XLOG_STATE_IOABORT;
1268
1269                 xfs_buf_ioerror_alert(bp, __func__);
1270                 xfs_buf_stale(bp);
1271                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1272                 /*
1273                  * This flag will be propagated to the trans-committed
1274                  * callback routines to let them know that the log-commit
1275                  * didn't succeed.
1276                  */
1277                 aborted = XFS_LI_ABORTED;
1278         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1279                 aborted = XFS_LI_ABORTED;
1280         }
1281
1282         /* log I/O is always issued ASYNC */
1283         ASSERT(bp->b_flags & XBF_ASYNC);
1284         xlog_state_done_syncing(iclog, aborted);
1285
1286         /*
1287          * drop the buffer lock now that we are done. Nothing references
1288          * the buffer after this, so an unmount waiting on this lock can now
1289          * tear it down safely. As such, it is unsafe to reference the buffer
1290          * (bp) after the unlock as we could race with it being freed.
1291          */
1292         xfs_buf_unlock(bp);
1293 }
1294
1295 /*
1296  * Return size of each in-core log record buffer.
1297  *
1298  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1299  *
1300  * If the filesystem blocksize is too large, we may need to choose a
1301  * larger size since the directory code currently logs entire blocks.
1302  */
1303
1304 STATIC void
1305 xlog_get_iclog_buffer_size(
1306         struct xfs_mount        *mp,
1307         struct xlog             *log)
1308 {
1309         int size;
1310         int xhdrs;
1311
1312         if (mp->m_logbufs <= 0)
1313                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1314         else
1315                 log->l_iclog_bufs = mp->m_logbufs;
1316
1317         /*
1318          * Buffer size passed in from mount system call.
1319          */
1320         if (mp->m_logbsize > 0) {
1321                 size = log->l_iclog_size = mp->m_logbsize;
1322                 log->l_iclog_size_log = 0;
1323                 while (size != 1) {
1324                         log->l_iclog_size_log++;
1325                         size >>= 1;
1326                 }
1327
1328                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1329                         /* # headers = size / 32k
1330                          * one header holds cycles from 32k of data
1331                          */
1332
1333                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1334                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1335                                 xhdrs++;
1336                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1337                         log->l_iclog_heads = xhdrs;
1338                 } else {
1339                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1340                         log->l_iclog_hsize = BBSIZE;
1341                         log->l_iclog_heads = 1;
1342                 }
1343                 goto done;
1344         }
1345
1346         /* All machines use 32kB buffers by default. */
1347         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1348         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1349
1350         /* the default log size is 16k or 32k which is one header sector */
1351         log->l_iclog_hsize = BBSIZE;
1352         log->l_iclog_heads = 1;
1353
1354 done:
1355         /* are we being asked to make the sizes selected above visible? */
1356         if (mp->m_logbufs == 0)
1357                 mp->m_logbufs = log->l_iclog_bufs;
1358         if (mp->m_logbsize == 0)
1359                 mp->m_logbsize = log->l_iclog_size;
1360 }       /* xlog_get_iclog_buffer_size */
1361
1362
1363 void
1364 xfs_log_work_queue(
1365         struct xfs_mount        *mp)
1366 {
1367         queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1368                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1369 }
1370
1371 /*
1372  * Every sync period we need to unpin all items in the AIL and push them to
1373  * disk. If there is nothing dirty, then we might need to cover the log to
1374  * indicate that the filesystem is idle.
1375  */
1376 static void
1377 xfs_log_worker(
1378         struct work_struct      *work)
1379 {
1380         struct xlog             *log = container_of(to_delayed_work(work),
1381                                                 struct xlog, l_work);
1382         struct xfs_mount        *mp = log->l_mp;
1383
1384         /* dgc: errors ignored - not fatal and nowhere to report them */
1385         if (xfs_log_need_covered(mp)) {
1386                 /*
1387                  * Dump a transaction into the log that contains no real change.
1388                  * This is needed to stamp the current tail LSN into the log
1389                  * during the covering operation.
1390                  *
1391                  * We cannot use an inode here for this - that will push dirty
1392                  * state back up into the VFS and then periodic inode flushing
1393                  * will prevent log covering from making progress. Hence we
1394                  * synchronously log the superblock instead to ensure the
1395                  * superblock is immediately unpinned and can be written back.
1396                  */
1397                 xfs_sync_sb(mp, true);
1398         } else
1399                 xfs_log_force(mp, 0);
1400
1401         /* start pushing all the metadata that is currently dirty */
1402         xfs_ail_push_all(mp->m_ail);
1403
1404         /* queue us up again */
1405         xfs_log_work_queue(mp);
1406 }
1407
1408 /*
1409  * This routine initializes some of the log structure for a given mount point.
1410  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1411  * some other stuff may be filled in too.
1412  */
1413 STATIC struct xlog *
1414 xlog_alloc_log(
1415         struct xfs_mount        *mp,
1416         struct xfs_buftarg      *log_target,
1417         xfs_daddr_t             blk_offset,
1418         int                     num_bblks)
1419 {
1420         struct xlog             *log;
1421         xlog_rec_header_t       *head;
1422         xlog_in_core_t          **iclogp;
1423         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1424         xfs_buf_t               *bp;
1425         int                     i;
1426         int                     error = -ENOMEM;
1427         uint                    log2_size = 0;
1428
1429         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1430         if (!log) {
1431                 xfs_warn(mp, "Log allocation failed: No memory!");
1432                 goto out;
1433         }
1434
1435         log->l_mp          = mp;
1436         log->l_targ        = log_target;
1437         log->l_logsize     = BBTOB(num_bblks);
1438         log->l_logBBstart  = blk_offset;
1439         log->l_logBBsize   = num_bblks;
1440         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1441         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1442         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1443
1444         log->l_prev_block  = -1;
1445         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1446         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1447         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1448         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1449
1450         xlog_grant_head_init(&log->l_reserve_head);
1451         xlog_grant_head_init(&log->l_write_head);
1452
1453         error = -EFSCORRUPTED;
1454         if (xfs_sb_version_hassector(&mp->m_sb)) {
1455                 log2_size = mp->m_sb.sb_logsectlog;
1456                 if (log2_size < BBSHIFT) {
1457                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1458                                 log2_size, BBSHIFT);
1459                         goto out_free_log;
1460                 }
1461
1462                 log2_size -= BBSHIFT;
1463                 if (log2_size > mp->m_sectbb_log) {
1464                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1465                                 log2_size, mp->m_sectbb_log);
1466                         goto out_free_log;
1467                 }
1468
1469                 /* for larger sector sizes, must have v2 or external log */
1470                 if (log2_size && log->l_logBBstart > 0 &&
1471                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1472                         xfs_warn(mp,
1473                 "log sector size (0x%x) invalid for configuration.",
1474                                 log2_size);
1475                         goto out_free_log;
1476                 }
1477         }
1478         log->l_sectBBsize = 1 << log2_size;
1479
1480         xlog_get_iclog_buffer_size(mp, log);
1481
1482         /*
1483          * Use a NULL block for the extra log buffer used during splits so that
1484          * it will trigger errors if we ever try to do IO on it without first
1485          * having set it up properly.
1486          */
1487         error = -ENOMEM;
1488         bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1489                            BTOBB(log->l_iclog_size), XBF_NO_IOACCT);
1490         if (!bp)
1491                 goto out_free_log;
1492
1493         /*
1494          * The iclogbuf buffer locks are held over IO but we are not going to do
1495          * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1496          * when appropriately.
1497          */
1498         ASSERT(xfs_buf_islocked(bp));
1499         xfs_buf_unlock(bp);
1500
1501         /* use high priority wq for log I/O completion */
1502         bp->b_ioend_wq = mp->m_log_workqueue;
1503         bp->b_iodone = xlog_iodone;
1504         log->l_xbuf = bp;
1505
1506         spin_lock_init(&log->l_icloglock);
1507         init_waitqueue_head(&log->l_flush_wait);
1508
1509         iclogp = &log->l_iclog;
1510         /*
1511          * The amount of memory to allocate for the iclog structure is
1512          * rather funky due to the way the structure is defined.  It is
1513          * done this way so that we can use different sizes for machines
1514          * with different amounts of memory.  See the definition of
1515          * xlog_in_core_t in xfs_log_priv.h for details.
1516          */
1517         ASSERT(log->l_iclog_size >= 4096);
1518         for (i=0; i < log->l_iclog_bufs; i++) {
1519                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1520                 if (!*iclogp)
1521                         goto out_free_iclog;
1522
1523                 iclog = *iclogp;
1524                 iclog->ic_prev = prev_iclog;
1525                 prev_iclog = iclog;
1526
1527                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1528                                           BTOBB(log->l_iclog_size),
1529                                           XBF_NO_IOACCT);
1530                 if (!bp)
1531                         goto out_free_iclog;
1532
1533                 ASSERT(xfs_buf_islocked(bp));
1534                 xfs_buf_unlock(bp);
1535
1536                 /* use high priority wq for log I/O completion */
1537                 bp->b_ioend_wq = mp->m_log_workqueue;
1538                 bp->b_iodone = xlog_iodone;
1539                 iclog->ic_bp = bp;
1540                 iclog->ic_data = bp->b_addr;
1541 #ifdef DEBUG
1542                 log->l_iclog_bak[i] = &iclog->ic_header;
1543 #endif
1544                 head = &iclog->ic_header;
1545                 memset(head, 0, sizeof(xlog_rec_header_t));
1546                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1547                 head->h_version = cpu_to_be32(
1548                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1549                 head->h_size = cpu_to_be32(log->l_iclog_size);
1550                 /* new fields */
1551                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1552                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1553
1554                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1555                 iclog->ic_state = XLOG_STATE_ACTIVE;
1556                 iclog->ic_log = log;
1557                 atomic_set(&iclog->ic_refcnt, 0);
1558                 spin_lock_init(&iclog->ic_callback_lock);
1559                 iclog->ic_callback_tail = &(iclog->ic_callback);
1560                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1561
1562                 init_waitqueue_head(&iclog->ic_force_wait);
1563                 init_waitqueue_head(&iclog->ic_write_wait);
1564
1565                 iclogp = &iclog->ic_next;
1566         }
1567         *iclogp = log->l_iclog;                 /* complete ring */
1568         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1569
1570         error = xlog_cil_init(log);
1571         if (error)
1572                 goto out_free_iclog;
1573         return log;
1574
1575 out_free_iclog:
1576         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1577                 prev_iclog = iclog->ic_next;
1578                 if (iclog->ic_bp)
1579                         xfs_buf_free(iclog->ic_bp);
1580                 kmem_free(iclog);
1581         }
1582         spinlock_destroy(&log->l_icloglock);
1583         xfs_buf_free(log->l_xbuf);
1584 out_free_log:
1585         kmem_free(log);
1586 out:
1587         return ERR_PTR(error);
1588 }       /* xlog_alloc_log */
1589
1590
1591 /*
1592  * Write out the commit record of a transaction associated with the given
1593  * ticket.  Return the lsn of the commit record.
1594  */
1595 STATIC int
1596 xlog_commit_record(
1597         struct xlog             *log,
1598         struct xlog_ticket      *ticket,
1599         struct xlog_in_core     **iclog,
1600         xfs_lsn_t               *commitlsnp)
1601 {
1602         struct xfs_mount *mp = log->l_mp;
1603         int     error;
1604         struct xfs_log_iovec reg = {
1605                 .i_addr = NULL,
1606                 .i_len = 0,
1607                 .i_type = XLOG_REG_TYPE_COMMIT,
1608         };
1609         struct xfs_log_vec vec = {
1610                 .lv_niovecs = 1,
1611                 .lv_iovecp = &reg,
1612         };
1613
1614         ASSERT_ALWAYS(iclog);
1615         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1616                                         XLOG_COMMIT_TRANS);
1617         if (error)
1618                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1619         return error;
1620 }
1621
1622 /*
1623  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1624  * log space.  This code pushes on the lsn which would supposedly free up
1625  * the 25% which we want to leave free.  We may need to adopt a policy which
1626  * pushes on an lsn which is further along in the log once we reach the high
1627  * water mark.  In this manner, we would be creating a low water mark.
1628  */
1629 STATIC void
1630 xlog_grant_push_ail(
1631         struct xlog     *log,
1632         int             need_bytes)
1633 {
1634         xfs_lsn_t       threshold_lsn = 0;
1635         xfs_lsn_t       last_sync_lsn;
1636         int             free_blocks;
1637         int             free_bytes;
1638         int             threshold_block;
1639         int             threshold_cycle;
1640         int             free_threshold;
1641
1642         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1643
1644         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1645         free_blocks = BTOBBT(free_bytes);
1646
1647         /*
1648          * Set the threshold for the minimum number of free blocks in the
1649          * log to the maximum of what the caller needs, one quarter of the
1650          * log, and 256 blocks.
1651          */
1652         free_threshold = BTOBB(need_bytes);
1653         free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1654         free_threshold = max(free_threshold, 256);
1655         if (free_blocks >= free_threshold)
1656                 return;
1657
1658         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1659                                                 &threshold_block);
1660         threshold_block += free_threshold;
1661         if (threshold_block >= log->l_logBBsize) {
1662                 threshold_block -= log->l_logBBsize;
1663                 threshold_cycle += 1;
1664         }
1665         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1666                                         threshold_block);
1667         /*
1668          * Don't pass in an lsn greater than the lsn of the last
1669          * log record known to be on disk. Use a snapshot of the last sync lsn
1670          * so that it doesn't change between the compare and the set.
1671          */
1672         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1673         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1674                 threshold_lsn = last_sync_lsn;
1675
1676         /*
1677          * Get the transaction layer to kick the dirty buffers out to
1678          * disk asynchronously. No point in trying to do this if
1679          * the filesystem is shutting down.
1680          */
1681         if (!XLOG_FORCED_SHUTDOWN(log))
1682                 xfs_ail_push(log->l_ailp, threshold_lsn);
1683 }
1684
1685 /*
1686  * Stamp cycle number in every block
1687  */
1688 STATIC void
1689 xlog_pack_data(
1690         struct xlog             *log,
1691         struct xlog_in_core     *iclog,
1692         int                     roundoff)
1693 {
1694         int                     i, j, k;
1695         int                     size = iclog->ic_offset + roundoff;
1696         __be32                  cycle_lsn;
1697         char                    *dp;
1698
1699         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1700
1701         dp = iclog->ic_datap;
1702         for (i = 0; i < BTOBB(size); i++) {
1703                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1704                         break;
1705                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1706                 *(__be32 *)dp = cycle_lsn;
1707                 dp += BBSIZE;
1708         }
1709
1710         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1711                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1712
1713                 for ( ; i < BTOBB(size); i++) {
1714                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1715                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1716                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1717                         *(__be32 *)dp = cycle_lsn;
1718                         dp += BBSIZE;
1719                 }
1720
1721                 for (i = 1; i < log->l_iclog_heads; i++)
1722                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1723         }
1724 }
1725
1726 /*
1727  * Calculate the checksum for a log buffer.
1728  *
1729  * This is a little more complicated than it should be because the various
1730  * headers and the actual data are non-contiguous.
1731  */
1732 __le32
1733 xlog_cksum(
1734         struct xlog             *log,
1735         struct xlog_rec_header  *rhead,
1736         char                    *dp,
1737         int                     size)
1738 {
1739         uint32_t                crc;
1740
1741         /* first generate the crc for the record header ... */
1742         crc = xfs_start_cksum_update((char *)rhead,
1743                               sizeof(struct xlog_rec_header),
1744                               offsetof(struct xlog_rec_header, h_crc));
1745
1746         /* ... then for additional cycle data for v2 logs ... */
1747         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1748                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1749                 int             i;
1750                 int             xheads;
1751
1752                 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1753                 if (size % XLOG_HEADER_CYCLE_SIZE)
1754                         xheads++;
1755
1756                 for (i = 1; i < xheads; i++) {
1757                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1758                                      sizeof(struct xlog_rec_ext_header));
1759                 }
1760         }
1761
1762         /* ... and finally for the payload */
1763         crc = crc32c(crc, dp, size);
1764
1765         return xfs_end_cksum(crc);
1766 }
1767
1768 /*
1769  * The bdstrat callback function for log bufs. This gives us a central
1770  * place to trap bufs in case we get hit by a log I/O error and need to
1771  * shutdown. Actually, in practice, even when we didn't get a log error,
1772  * we transition the iclogs to IOERROR state *after* flushing all existing
1773  * iclogs to disk. This is because we don't want anymore new transactions to be
1774  * started or completed afterwards.
1775  *
1776  * We lock the iclogbufs here so that we can serialise against IO completion
1777  * during unmount. We might be processing a shutdown triggered during unmount,
1778  * and that can occur asynchronously to the unmount thread, and hence we need to
1779  * ensure that completes before tearing down the iclogbufs. Hence we need to
1780  * hold the buffer lock across the log IO to acheive that.
1781  */
1782 STATIC int
1783 xlog_bdstrat(
1784         struct xfs_buf          *bp)
1785 {
1786         struct xlog_in_core     *iclog = bp->b_log_item;
1787
1788         xfs_buf_lock(bp);
1789         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1790                 xfs_buf_ioerror(bp, -EIO);
1791                 xfs_buf_stale(bp);
1792                 xfs_buf_ioend(bp);
1793                 /*
1794                  * It would seem logical to return EIO here, but we rely on
1795                  * the log state machine to propagate I/O errors instead of
1796                  * doing it here. Similarly, IO completion will unlock the
1797                  * buffer, so we don't do it here.
1798                  */
1799                 return 0;
1800         }
1801
1802         xfs_buf_submit(bp);
1803         return 0;
1804 }
1805
1806 /*
1807  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1808  * fashion.  Previously, we should have moved the current iclog
1809  * ptr in the log to point to the next available iclog.  This allows further
1810  * write to continue while this code syncs out an iclog ready to go.
1811  * Before an in-core log can be written out, the data section must be scanned
1812  * to save away the 1st word of each BBSIZE block into the header.  We replace
1813  * it with the current cycle count.  Each BBSIZE block is tagged with the
1814  * cycle count because there in an implicit assumption that drives will
1815  * guarantee that entire 512 byte blocks get written at once.  In other words,
1816  * we can't have part of a 512 byte block written and part not written.  By
1817  * tagging each block, we will know which blocks are valid when recovering
1818  * after an unclean shutdown.
1819  *
1820  * This routine is single threaded on the iclog.  No other thread can be in
1821  * this routine with the same iclog.  Changing contents of iclog can there-
1822  * fore be done without grabbing the state machine lock.  Updating the global
1823  * log will require grabbing the lock though.
1824  *
1825  * The entire log manager uses a logical block numbering scheme.  Only
1826  * log_sync (and then only bwrite()) know about the fact that the log may
1827  * not start with block zero on a given device.  The log block start offset
1828  * is added immediately before calling bwrite().
1829  */
1830
1831 STATIC int
1832 xlog_sync(
1833         struct xlog             *log,
1834         struct xlog_in_core     *iclog)
1835 {
1836         xfs_buf_t       *bp;
1837         int             i;
1838         uint            count;          /* byte count of bwrite */
1839         uint            count_init;     /* initial count before roundup */
1840         int             roundoff;       /* roundoff to BB or stripe */
1841         int             split = 0;      /* split write into two regions */
1842         int             error;
1843         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1844         int             size;
1845
1846         XFS_STATS_INC(log->l_mp, xs_log_writes);
1847         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1848
1849         /* Add for LR header */
1850         count_init = log->l_iclog_hsize + iclog->ic_offset;
1851
1852         /* Round out the log write size */
1853         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1854                 /* we have a v2 stripe unit to use */
1855                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1856         } else {
1857                 count = BBTOB(BTOBB(count_init));
1858         }
1859         roundoff = count - count_init;
1860         ASSERT(roundoff >= 0);
1861         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1862                 roundoff < log->l_mp->m_sb.sb_logsunit)
1863                 ||
1864                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1865                  roundoff < BBTOB(1)));
1866
1867         /* move grant heads by roundoff in sync */
1868         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1869         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1870
1871         /* put cycle number in every block */
1872         xlog_pack_data(log, iclog, roundoff);
1873
1874         /* real byte length */
1875         size = iclog->ic_offset;
1876         if (v2)
1877                 size += roundoff;
1878         iclog->ic_header.h_len = cpu_to_be32(size);
1879
1880         bp = iclog->ic_bp;
1881         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1882
1883         XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1884
1885         /* Do we need to split this write into 2 parts? */
1886         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1887                 char            *dptr;
1888
1889                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1890                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1891                 iclog->ic_bwritecnt = 2;
1892
1893                 /*
1894                  * Bump the cycle numbers at the start of each block in the
1895                  * part of the iclog that ends up in the buffer that gets
1896                  * written to the start of the log.
1897                  *
1898                  * Watch out for the header magic number case, though.
1899                  */
1900                 dptr = (char *)&iclog->ic_header + count;
1901                 for (i = 0; i < split; i += BBSIZE) {
1902                         uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1903                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1904                                 cycle++;
1905                         *(__be32 *)dptr = cpu_to_be32(cycle);
1906
1907                         dptr += BBSIZE;
1908                 }
1909         } else {
1910                 iclog->ic_bwritecnt = 1;
1911         }
1912
1913         /* calculcate the checksum */
1914         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1915                                             iclog->ic_datap, size);
1916         /*
1917          * Intentionally corrupt the log record CRC based on the error injection
1918          * frequency, if defined. This facilitates testing log recovery in the
1919          * event of torn writes. Hence, set the IOABORT state to abort the log
1920          * write on I/O completion and shutdown the fs. The subsequent mount
1921          * detects the bad CRC and attempts to recover.
1922          */
1923         if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1924                 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1925                 iclog->ic_state |= XLOG_STATE_IOABORT;
1926                 xfs_warn(log->l_mp,
1927         "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1928                          be64_to_cpu(iclog->ic_header.h_lsn));
1929         }
1930
1931         bp->b_io_length = BTOBB(count);
1932         bp->b_log_item = iclog;
1933         bp->b_flags &= ~XBF_FLUSH;
1934         bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1935
1936         /*
1937          * Flush the data device before flushing the log to make sure all meta
1938          * data written back from the AIL actually made it to disk before
1939          * stamping the new log tail LSN into the log buffer.  For an external
1940          * log we need to issue the flush explicitly, and unfortunately
1941          * synchronously here; for an internal log we can simply use the block
1942          * layer state machine for preflushes.
1943          */
1944         if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1945                 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1946         else
1947                 bp->b_flags |= XBF_FLUSH;
1948
1949         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1950         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1951
1952         xlog_verify_iclog(log, iclog, count, true);
1953
1954         /* account for log which doesn't start at block #0 */
1955         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1956
1957         /*
1958          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1959          * is shutting down.
1960          */
1961         error = xlog_bdstrat(bp);
1962         if (error) {
1963                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1964                 return error;
1965         }
1966         if (split) {
1967                 bp = iclog->ic_log->l_xbuf;
1968                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1969                 xfs_buf_associate_memory(bp,
1970                                 (char *)&iclog->ic_header + count, split);
1971                 bp->b_log_item = iclog;
1972                 bp->b_flags &= ~XBF_FLUSH;
1973                 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1974
1975                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1976                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1977
1978                 /* account for internal log which doesn't start at block #0 */
1979                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1980                 error = xlog_bdstrat(bp);
1981                 if (error) {
1982                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1983                         return error;
1984                 }
1985         }
1986         return 0;
1987 }       /* xlog_sync */
1988
1989 /*
1990  * Deallocate a log structure
1991  */
1992 STATIC void
1993 xlog_dealloc_log(
1994         struct xlog     *log)
1995 {
1996         xlog_in_core_t  *iclog, *next_iclog;
1997         int             i;
1998
1999         xlog_cil_destroy(log);
2000
2001         /*
2002          * Cycle all the iclogbuf locks to make sure all log IO completion
2003          * is done before we tear down these buffers.
2004          */
2005         iclog = log->l_iclog;
2006         for (i = 0; i < log->l_iclog_bufs; i++) {
2007                 xfs_buf_lock(iclog->ic_bp);
2008                 xfs_buf_unlock(iclog->ic_bp);
2009                 iclog = iclog->ic_next;
2010         }
2011
2012         /*
2013          * Always need to ensure that the extra buffer does not point to memory
2014          * owned by another log buffer before we free it. Also, cycle the lock
2015          * first to ensure we've completed IO on it.
2016          */
2017         xfs_buf_lock(log->l_xbuf);
2018         xfs_buf_unlock(log->l_xbuf);
2019         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
2020         xfs_buf_free(log->l_xbuf);
2021
2022         iclog = log->l_iclog;
2023         for (i = 0; i < log->l_iclog_bufs; i++) {
2024                 xfs_buf_free(iclog->ic_bp);
2025                 next_iclog = iclog->ic_next;
2026                 kmem_free(iclog);
2027                 iclog = next_iclog;
2028         }
2029         spinlock_destroy(&log->l_icloglock);
2030
2031         log->l_mp->m_log = NULL;
2032         kmem_free(log);
2033 }       /* xlog_dealloc_log */
2034
2035 /*
2036  * Update counters atomically now that memcpy is done.
2037  */
2038 /* ARGSUSED */
2039 static inline void
2040 xlog_state_finish_copy(
2041         struct xlog             *log,
2042         struct xlog_in_core     *iclog,
2043         int                     record_cnt,
2044         int                     copy_bytes)
2045 {
2046         spin_lock(&log->l_icloglock);
2047
2048         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2049         iclog->ic_offset += copy_bytes;
2050
2051         spin_unlock(&log->l_icloglock);
2052 }       /* xlog_state_finish_copy */
2053
2054
2055
2056
2057 /*
2058  * print out info relating to regions written which consume
2059  * the reservation
2060  */
2061 void
2062 xlog_print_tic_res(
2063         struct xfs_mount        *mp,
2064         struct xlog_ticket      *ticket)
2065 {
2066         uint i;
2067         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2068
2069         /* match with XLOG_REG_TYPE_* in xfs_log.h */
2070 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2071         static char *res_type_str[XLOG_REG_TYPE_MAX + 1] = {
2072             REG_TYPE_STR(BFORMAT, "bformat"),
2073             REG_TYPE_STR(BCHUNK, "bchunk"),
2074             REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2075             REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2076             REG_TYPE_STR(IFORMAT, "iformat"),
2077             REG_TYPE_STR(ICORE, "icore"),
2078             REG_TYPE_STR(IEXT, "iext"),
2079             REG_TYPE_STR(IBROOT, "ibroot"),
2080             REG_TYPE_STR(ILOCAL, "ilocal"),
2081             REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2082             REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2083             REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2084             REG_TYPE_STR(QFORMAT, "qformat"),
2085             REG_TYPE_STR(DQUOT, "dquot"),
2086             REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2087             REG_TYPE_STR(LRHEADER, "LR header"),
2088             REG_TYPE_STR(UNMOUNT, "unmount"),
2089             REG_TYPE_STR(COMMIT, "commit"),
2090             REG_TYPE_STR(TRANSHDR, "trans header"),
2091             REG_TYPE_STR(ICREATE, "inode create")
2092         };
2093 #undef REG_TYPE_STR
2094
2095         xfs_warn(mp, "ticket reservation summary:");
2096         xfs_warn(mp, "  unit res    = %d bytes",
2097                  ticket->t_unit_res);
2098         xfs_warn(mp, "  current res = %d bytes",
2099                  ticket->t_curr_res);
2100         xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2101                  ticket->t_res_arr_sum, ticket->t_res_o_flow);
2102         xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2103                  ticket->t_res_num_ophdrs, ophdr_spc);
2104         xfs_warn(mp, "  ophdr + reg = %u bytes",
2105                  ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2106         xfs_warn(mp, "  num regions = %u",
2107                  ticket->t_res_num);
2108
2109         for (i = 0; i < ticket->t_res_num; i++) {
2110                 uint r_type = ticket->t_res_arr[i].r_type;
2111                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2112                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2113                             "bad-rtype" : res_type_str[r_type]),
2114                             ticket->t_res_arr[i].r_len);
2115         }
2116 }
2117
2118 /*
2119  * Print a summary of the transaction.
2120  */
2121 void
2122 xlog_print_trans(
2123         struct xfs_trans        *tp)
2124 {
2125         struct xfs_mount        *mp = tp->t_mountp;
2126         struct xfs_log_item     *lip;
2127
2128         /* dump core transaction and ticket info */
2129         xfs_warn(mp, "transaction summary:");
2130         xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2131         xfs_warn(mp, "  log count = %d", tp->t_log_count);
2132         xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2133
2134         xlog_print_tic_res(mp, tp->t_ticket);
2135
2136         /* dump each log item */
2137         list_for_each_entry(lip, &tp->t_items, li_trans) {
2138                 struct xfs_log_vec      *lv = lip->li_lv;
2139                 struct xfs_log_iovec    *vec;
2140                 int                     i;
2141
2142                 xfs_warn(mp, "log item: ");
2143                 xfs_warn(mp, "  type    = 0x%x", lip->li_type);
2144                 xfs_warn(mp, "  flags   = 0x%lx", lip->li_flags);
2145                 if (!lv)
2146                         continue;
2147                 xfs_warn(mp, "  niovecs = %d", lv->lv_niovecs);
2148                 xfs_warn(mp, "  size    = %d", lv->lv_size);
2149                 xfs_warn(mp, "  bytes   = %d", lv->lv_bytes);
2150                 xfs_warn(mp, "  buf len = %d", lv->lv_buf_len);
2151
2152                 /* dump each iovec for the log item */
2153                 vec = lv->lv_iovecp;
2154                 for (i = 0; i < lv->lv_niovecs; i++) {
2155                         int dumplen = min(vec->i_len, 32);
2156
2157                         xfs_warn(mp, "  iovec[%d]", i);
2158                         xfs_warn(mp, "    type  = 0x%x", vec->i_type);
2159                         xfs_warn(mp, "    len   = %d", vec->i_len);
2160                         xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2161                         xfs_hex_dump(vec->i_addr, dumplen);
2162
2163                         vec++;
2164                 }
2165         }
2166 }
2167
2168 /*
2169  * Calculate the potential space needed by the log vector.  Each region gets
2170  * its own xlog_op_header_t and may need to be double word aligned.
2171  */
2172 static int
2173 xlog_write_calc_vec_length(
2174         struct xlog_ticket      *ticket,
2175         struct xfs_log_vec      *log_vector)
2176 {
2177         struct xfs_log_vec      *lv;
2178         int                     headers = 0;
2179         int                     len = 0;
2180         int                     i;
2181
2182         /* acct for start rec of xact */
2183         if (ticket->t_flags & XLOG_TIC_INITED)
2184                 headers++;
2185
2186         for (lv = log_vector; lv; lv = lv->lv_next) {
2187                 /* we don't write ordered log vectors */
2188                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2189                         continue;
2190
2191                 headers += lv->lv_niovecs;
2192
2193                 for (i = 0; i < lv->lv_niovecs; i++) {
2194                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2195
2196                         len += vecp->i_len;
2197                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2198                 }
2199         }
2200
2201         ticket->t_res_num_ophdrs += headers;
2202         len += headers * sizeof(struct xlog_op_header);
2203
2204         return len;
2205 }
2206
2207 /*
2208  * If first write for transaction, insert start record  We can't be trying to
2209  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2210  */
2211 static int
2212 xlog_write_start_rec(
2213         struct xlog_op_header   *ophdr,
2214         struct xlog_ticket      *ticket)
2215 {
2216         if (!(ticket->t_flags & XLOG_TIC_INITED))
2217                 return 0;
2218
2219         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2220         ophdr->oh_clientid = ticket->t_clientid;
2221         ophdr->oh_len = 0;
2222         ophdr->oh_flags = XLOG_START_TRANS;
2223         ophdr->oh_res2 = 0;
2224
2225         ticket->t_flags &= ~XLOG_TIC_INITED;
2226
2227         return sizeof(struct xlog_op_header);
2228 }
2229
2230 static xlog_op_header_t *
2231 xlog_write_setup_ophdr(
2232         struct xlog             *log,
2233         struct xlog_op_header   *ophdr,
2234         struct xlog_ticket      *ticket,
2235         uint                    flags)
2236 {
2237         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2238         ophdr->oh_clientid = ticket->t_clientid;
2239         ophdr->oh_res2 = 0;
2240
2241         /* are we copying a commit or unmount record? */
2242         ophdr->oh_flags = flags;
2243
2244         /*
2245          * We've seen logs corrupted with bad transaction client ids.  This
2246          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2247          * and shut down the filesystem.
2248          */
2249         switch (ophdr->oh_clientid)  {
2250         case XFS_TRANSACTION:
2251         case XFS_VOLUME:
2252         case XFS_LOG:
2253                 break;
2254         default:
2255                 xfs_warn(log->l_mp,
2256                         "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2257                         ophdr->oh_clientid, ticket);
2258                 return NULL;
2259         }
2260
2261         return ophdr;
2262 }
2263
2264 /*
2265  * Set up the parameters of the region copy into the log. This has
2266  * to handle region write split across multiple log buffers - this
2267  * state is kept external to this function so that this code can
2268  * be written in an obvious, self documenting manner.
2269  */
2270 static int
2271 xlog_write_setup_copy(
2272         struct xlog_ticket      *ticket,
2273         struct xlog_op_header   *ophdr,
2274         int                     space_available,
2275         int                     space_required,
2276         int                     *copy_off,
2277         int                     *copy_len,
2278         int                     *last_was_partial_copy,
2279         int                     *bytes_consumed)
2280 {
2281         int                     still_to_copy;
2282
2283         still_to_copy = space_required - *bytes_consumed;
2284         *copy_off = *bytes_consumed;
2285
2286         if (still_to_copy <= space_available) {
2287                 /* write of region completes here */
2288                 *copy_len = still_to_copy;
2289                 ophdr->oh_len = cpu_to_be32(*copy_len);
2290                 if (*last_was_partial_copy)
2291                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2292                 *last_was_partial_copy = 0;
2293                 *bytes_consumed = 0;
2294                 return 0;
2295         }
2296
2297         /* partial write of region, needs extra log op header reservation */
2298         *copy_len = space_available;
2299         ophdr->oh_len = cpu_to_be32(*copy_len);
2300         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2301         if (*last_was_partial_copy)
2302                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2303         *bytes_consumed += *copy_len;
2304         (*last_was_partial_copy)++;
2305
2306         /* account for new log op header */
2307         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2308         ticket->t_res_num_ophdrs++;
2309
2310         return sizeof(struct xlog_op_header);
2311 }
2312
2313 static int
2314 xlog_write_copy_finish(
2315         struct xlog             *log,
2316         struct xlog_in_core     *iclog,
2317         uint                    flags,
2318         int                     *record_cnt,
2319         int                     *data_cnt,
2320         int                     *partial_copy,
2321         int                     *partial_copy_len,
2322         int                     log_offset,
2323         struct xlog_in_core     **commit_iclog)
2324 {
2325         if (*partial_copy) {
2326                 /*
2327                  * This iclog has already been marked WANT_SYNC by
2328                  * xlog_state_get_iclog_space.
2329                  */
2330                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2331                 *record_cnt = 0;
2332                 *data_cnt = 0;
2333                 return xlog_state_release_iclog(log, iclog);
2334         }
2335
2336         *partial_copy = 0;
2337         *partial_copy_len = 0;
2338
2339         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2340                 /* no more space in this iclog - push it. */
2341                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2342                 *record_cnt = 0;
2343                 *data_cnt = 0;
2344
2345                 spin_lock(&log->l_icloglock);
2346                 xlog_state_want_sync(log, iclog);
2347                 spin_unlock(&log->l_icloglock);
2348
2349                 if (!commit_iclog)
2350                         return xlog_state_release_iclog(log, iclog);
2351                 ASSERT(flags & XLOG_COMMIT_TRANS);
2352                 *commit_iclog = iclog;
2353         }
2354
2355         return 0;
2356 }
2357
2358 /*
2359  * Write some region out to in-core log
2360  *
2361  * This will be called when writing externally provided regions or when
2362  * writing out a commit record for a given transaction.
2363  *
2364  * General algorithm:
2365  *      1. Find total length of this write.  This may include adding to the
2366  *              lengths passed in.
2367  *      2. Check whether we violate the tickets reservation.
2368  *      3. While writing to this iclog
2369  *          A. Reserve as much space in this iclog as can get
2370  *          B. If this is first write, save away start lsn
2371  *          C. While writing this region:
2372  *              1. If first write of transaction, write start record
2373  *              2. Write log operation header (header per region)
2374  *              3. Find out if we can fit entire region into this iclog
2375  *              4. Potentially, verify destination memcpy ptr
2376  *              5. Memcpy (partial) region
2377  *              6. If partial copy, release iclog; otherwise, continue
2378  *                      copying more regions into current iclog
2379  *      4. Mark want sync bit (in simulation mode)
2380  *      5. Release iclog for potential flush to on-disk log.
2381  *
2382  * ERRORS:
2383  * 1.   Panic if reservation is overrun.  This should never happen since
2384  *      reservation amounts are generated internal to the filesystem.
2385  * NOTES:
2386  * 1. Tickets are single threaded data structures.
2387  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2388  *      syncing routine.  When a single log_write region needs to span
2389  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2390  *      on all log operation writes which don't contain the end of the
2391  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2392  *      operation which contains the end of the continued log_write region.
2393  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2394  *      we don't really know exactly how much space will be used.  As a result,
2395  *      we don't update ic_offset until the end when we know exactly how many
2396  *      bytes have been written out.
2397  */
2398 int
2399 xlog_write(
2400         struct xlog             *log,
2401         struct xfs_log_vec      *log_vector,
2402         struct xlog_ticket      *ticket,
2403         xfs_lsn_t               *start_lsn,
2404         struct xlog_in_core     **commit_iclog,
2405         uint                    flags)
2406 {
2407         struct xlog_in_core     *iclog = NULL;
2408         struct xfs_log_iovec    *vecp;
2409         struct xfs_log_vec      *lv;
2410         int                     len;
2411         int                     index;
2412         int                     partial_copy = 0;
2413         int                     partial_copy_len = 0;
2414         int                     contwr = 0;
2415         int                     record_cnt = 0;
2416         int                     data_cnt = 0;
2417         int                     error;
2418
2419         *start_lsn = 0;
2420
2421         len = xlog_write_calc_vec_length(ticket, log_vector);
2422
2423         /*
2424          * Region headers and bytes are already accounted for.
2425          * We only need to take into account start records and
2426          * split regions in this function.
2427          */
2428         if (ticket->t_flags & XLOG_TIC_INITED)
2429                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2430
2431         /*
2432          * Commit record headers need to be accounted for. These
2433          * come in as separate writes so are easy to detect.
2434          */
2435         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2436                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2437
2438         if (ticket->t_curr_res < 0) {
2439                 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2440                      "ctx ticket reservation ran out. Need to up reservation");
2441                 xlog_print_tic_res(log->l_mp, ticket);
2442                 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2443         }
2444
2445         index = 0;
2446         lv = log_vector;
2447         vecp = lv->lv_iovecp;
2448         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2449                 void            *ptr;
2450                 int             log_offset;
2451
2452                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2453                                                    &contwr, &log_offset);
2454                 if (error)
2455                         return error;
2456
2457                 ASSERT(log_offset <= iclog->ic_size - 1);
2458                 ptr = iclog->ic_datap + log_offset;
2459
2460                 /* start_lsn is the first lsn written to. That's all we need. */
2461                 if (!*start_lsn)
2462                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2463
2464                 /*
2465                  * This loop writes out as many regions as can fit in the amount
2466                  * of space which was allocated by xlog_state_get_iclog_space().
2467                  */
2468                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2469                         struct xfs_log_iovec    *reg;
2470                         struct xlog_op_header   *ophdr;
2471                         int                     start_rec_copy;
2472                         int                     copy_len;
2473                         int                     copy_off;
2474                         bool                    ordered = false;
2475
2476                         /* ordered log vectors have no regions to write */
2477                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2478                                 ASSERT(lv->lv_niovecs == 0);
2479                                 ordered = true;
2480                                 goto next_lv;
2481                         }
2482
2483                         reg = &vecp[index];
2484                         ASSERT(reg->i_len % sizeof(int32_t) == 0);
2485                         ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2486
2487                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2488                         if (start_rec_copy) {
2489                                 record_cnt++;
2490                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2491                                                    start_rec_copy);
2492                         }
2493
2494                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2495                         if (!ophdr)
2496                                 return -EIO;
2497
2498                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2499                                            sizeof(struct xlog_op_header));
2500
2501                         len += xlog_write_setup_copy(ticket, ophdr,
2502                                                      iclog->ic_size-log_offset,
2503                                                      reg->i_len,
2504                                                      &copy_off, &copy_len,
2505                                                      &partial_copy,
2506                                                      &partial_copy_len);
2507                         xlog_verify_dest_ptr(log, ptr);
2508
2509                         /*
2510                          * Copy region.
2511                          *
2512                          * Unmount records just log an opheader, so can have
2513                          * empty payloads with no data region to copy. Hence we
2514                          * only copy the payload if the vector says it has data
2515                          * to copy.
2516                          */
2517                         ASSERT(copy_len >= 0);
2518                         if (copy_len > 0) {
2519                                 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2520                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2521                                                    copy_len);
2522                         }
2523                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2524                         record_cnt++;
2525                         data_cnt += contwr ? copy_len : 0;
2526
2527                         error = xlog_write_copy_finish(log, iclog, flags,
2528                                                        &record_cnt, &data_cnt,
2529                                                        &partial_copy,
2530                                                        &partial_copy_len,
2531                                                        log_offset,
2532                                                        commit_iclog);
2533                         if (error)
2534                                 return error;
2535
2536                         /*
2537                          * if we had a partial copy, we need to get more iclog
2538                          * space but we don't want to increment the region
2539                          * index because there is still more is this region to
2540                          * write.
2541                          *
2542                          * If we completed writing this region, and we flushed
2543                          * the iclog (indicated by resetting of the record
2544                          * count), then we also need to get more log space. If
2545                          * this was the last record, though, we are done and
2546                          * can just return.
2547                          */
2548                         if (partial_copy)
2549                                 break;
2550
2551                         if (++index == lv->lv_niovecs) {
2552 next_lv:
2553                                 lv = lv->lv_next;
2554                                 index = 0;
2555                                 if (lv)
2556                                         vecp = lv->lv_iovecp;
2557                         }
2558                         if (record_cnt == 0 && !ordered) {
2559                                 if (!lv)
2560                                         return 0;
2561                                 break;
2562                         }
2563                 }
2564         }
2565
2566         ASSERT(len == 0);
2567
2568         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2569         if (!commit_iclog)
2570                 return xlog_state_release_iclog(log, iclog);
2571
2572         ASSERT(flags & XLOG_COMMIT_TRANS);
2573         *commit_iclog = iclog;
2574         return 0;
2575 }
2576
2577
2578 /*****************************************************************************
2579  *
2580  *              State Machine functions
2581  *
2582  *****************************************************************************
2583  */
2584
2585 /* Clean iclogs starting from the head.  This ordering must be
2586  * maintained, so an iclog doesn't become ACTIVE beyond one that
2587  * is SYNCING.  This is also required to maintain the notion that we use
2588  * a ordered wait queue to hold off would be writers to the log when every
2589  * iclog is trying to sync to disk.
2590  *
2591  * State Change: DIRTY -> ACTIVE
2592  */
2593 STATIC void
2594 xlog_state_clean_log(
2595         struct xlog *log)
2596 {
2597         xlog_in_core_t  *iclog;
2598         int changed = 0;
2599
2600         iclog = log->l_iclog;
2601         do {
2602                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2603                         iclog->ic_state = XLOG_STATE_ACTIVE;
2604                         iclog->ic_offset       = 0;
2605                         ASSERT(iclog->ic_callback == NULL);
2606                         /*
2607                          * If the number of ops in this iclog indicate it just
2608                          * contains the dummy transaction, we can
2609                          * change state into IDLE (the second time around).
2610                          * Otherwise we should change the state into
2611                          * NEED a dummy.
2612                          * We don't need to cover the dummy.
2613                          */
2614                         if (!changed &&
2615                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2616                                         XLOG_COVER_OPS)) {
2617                                 changed = 1;
2618                         } else {
2619                                 /*
2620                                  * We have two dirty iclogs so start over
2621                                  * This could also be num of ops indicates
2622                                  * this is not the dummy going out.
2623                                  */
2624                                 changed = 2;
2625                         }
2626                         iclog->ic_header.h_num_logops = 0;
2627                         memset(iclog->ic_header.h_cycle_data, 0,
2628                               sizeof(iclog->ic_header.h_cycle_data));
2629                         iclog->ic_header.h_lsn = 0;
2630                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2631                         /* do nothing */;
2632                 else
2633                         break;  /* stop cleaning */
2634                 iclog = iclog->ic_next;
2635         } while (iclog != log->l_iclog);
2636
2637         /* log is locked when we are called */
2638         /*
2639          * Change state for the dummy log recording.
2640          * We usually go to NEED. But we go to NEED2 if the changed indicates
2641          * we are done writing the dummy record.
2642          * If we are done with the second dummy recored (DONE2), then
2643          * we go to IDLE.
2644          */
2645         if (changed) {
2646                 switch (log->l_covered_state) {
2647                 case XLOG_STATE_COVER_IDLE:
2648                 case XLOG_STATE_COVER_NEED:
2649                 case XLOG_STATE_COVER_NEED2:
2650                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2651                         break;
2652
2653                 case XLOG_STATE_COVER_DONE:
2654                         if (changed == 1)
2655                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2656                         else
2657                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2658                         break;
2659
2660                 case XLOG_STATE_COVER_DONE2:
2661                         if (changed == 1)
2662                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2663                         else
2664                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2665                         break;
2666
2667                 default:
2668                         ASSERT(0);
2669                 }
2670         }
2671 }       /* xlog_state_clean_log */
2672
2673 STATIC xfs_lsn_t
2674 xlog_get_lowest_lsn(
2675         struct xlog     *log)
2676 {
2677         xlog_in_core_t  *lsn_log;
2678         xfs_lsn_t       lowest_lsn, lsn;
2679
2680         lsn_log = log->l_iclog;
2681         lowest_lsn = 0;
2682         do {
2683             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2684                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2685                 if ((lsn && !lowest_lsn) ||
2686                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2687                         lowest_lsn = lsn;
2688                 }
2689             }
2690             lsn_log = lsn_log->ic_next;
2691         } while (lsn_log != log->l_iclog);
2692         return lowest_lsn;
2693 }
2694
2695
2696 STATIC void
2697 xlog_state_do_callback(
2698         struct xlog             *log,
2699         int                     aborted,
2700         struct xlog_in_core     *ciclog)
2701 {
2702         xlog_in_core_t     *iclog;
2703         xlog_in_core_t     *first_iclog;        /* used to know when we've
2704                                                  * processed all iclogs once */
2705         xfs_log_callback_t *cb, *cb_next;
2706         int                flushcnt = 0;
2707         xfs_lsn_t          lowest_lsn;
2708         int                ioerrors;    /* counter: iclogs with errors */
2709         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2710         int                funcdidcallbacks; /* flag: function did callbacks */
2711         int                repeats;     /* for issuing console warnings if
2712                                          * looping too many times */
2713         int                wake = 0;
2714
2715         spin_lock(&log->l_icloglock);
2716         first_iclog = iclog = log->l_iclog;
2717         ioerrors = 0;
2718         funcdidcallbacks = 0;
2719         repeats = 0;
2720
2721         do {
2722                 /*
2723                  * Scan all iclogs starting with the one pointed to by the
2724                  * log.  Reset this starting point each time the log is
2725                  * unlocked (during callbacks).
2726                  *
2727                  * Keep looping through iclogs until one full pass is made
2728                  * without running any callbacks.
2729                  */
2730                 first_iclog = log->l_iclog;
2731                 iclog = log->l_iclog;
2732                 loopdidcallbacks = 0;
2733                 repeats++;
2734
2735                 do {
2736
2737                         /* skip all iclogs in the ACTIVE & DIRTY states */
2738                         if (iclog->ic_state &
2739                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2740                                 iclog = iclog->ic_next;
2741                                 continue;
2742                         }
2743
2744                         /*
2745                          * Between marking a filesystem SHUTDOWN and stopping
2746                          * the log, we do flush all iclogs to disk (if there
2747                          * wasn't a log I/O error). So, we do want things to
2748                          * go smoothly in case of just a SHUTDOWN  w/o a
2749                          * LOG_IO_ERROR.
2750                          */
2751                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2752                                 /*
2753                                  * Can only perform callbacks in order.  Since
2754                                  * this iclog is not in the DONE_SYNC/
2755                                  * DO_CALLBACK state, we skip the rest and
2756                                  * just try to clean up.  If we set our iclog
2757                                  * to DO_CALLBACK, we will not process it when
2758                                  * we retry since a previous iclog is in the
2759                                  * CALLBACK and the state cannot change since
2760                                  * we are holding the l_icloglock.
2761                                  */
2762                                 if (!(iclog->ic_state &
2763                                         (XLOG_STATE_DONE_SYNC |
2764                                                  XLOG_STATE_DO_CALLBACK))) {
2765                                         if (ciclog && (ciclog->ic_state ==
2766                                                         XLOG_STATE_DONE_SYNC)) {
2767                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2768                                         }
2769                                         break;
2770                                 }
2771                                 /*
2772                                  * We now have an iclog that is in either the
2773                                  * DO_CALLBACK or DONE_SYNC states. The other
2774                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2775                                  * caught by the above if and are going to
2776                                  * clean (i.e. we aren't doing their callbacks)
2777                                  * see the above if.
2778                                  */
2779
2780                                 /*
2781                                  * We will do one more check here to see if we
2782                                  * have chased our tail around.
2783                                  */
2784
2785                                 lowest_lsn = xlog_get_lowest_lsn(log);
2786                                 if (lowest_lsn &&
2787                                     XFS_LSN_CMP(lowest_lsn,
2788                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2789                                         iclog = iclog->ic_next;
2790                                         continue; /* Leave this iclog for
2791                                                    * another thread */
2792                                 }
2793
2794                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2795
2796
2797                                 /*
2798                                  * Completion of a iclog IO does not imply that
2799                                  * a transaction has completed, as transactions
2800                                  * can be large enough to span many iclogs. We
2801                                  * cannot change the tail of the log half way
2802                                  * through a transaction as this may be the only
2803                                  * transaction in the log and moving th etail to
2804                                  * point to the middle of it will prevent
2805                                  * recovery from finding the start of the
2806                                  * transaction. Hence we should only update the
2807                                  * last_sync_lsn if this iclog contains
2808                                  * transaction completion callbacks on it.
2809                                  *
2810                                  * We have to do this before we drop the
2811                                  * icloglock to ensure we are the only one that
2812                                  * can update it.
2813                                  */
2814                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2815                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2816                                 if (iclog->ic_callback)
2817                                         atomic64_set(&log->l_last_sync_lsn,
2818                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2819
2820                         } else
2821                                 ioerrors++;
2822
2823                         spin_unlock(&log->l_icloglock);
2824
2825                         /*
2826                          * Keep processing entries in the callback list until
2827                          * we come around and it is empty.  We need to
2828                          * atomically see that the list is empty and change the
2829                          * state to DIRTY so that we don't miss any more
2830                          * callbacks being added.
2831                          */
2832                         spin_lock(&iclog->ic_callback_lock);
2833                         cb = iclog->ic_callback;
2834                         while (cb) {
2835                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2836                                 iclog->ic_callback = NULL;
2837                                 spin_unlock(&iclog->ic_callback_lock);
2838
2839                                 /* perform callbacks in the order given */
2840                                 for (; cb; cb = cb_next) {
2841                                         cb_next = cb->cb_next;
2842                                         cb->cb_func(cb->cb_arg, aborted);
2843                                 }
2844                                 spin_lock(&iclog->ic_callback_lock);
2845                                 cb = iclog->ic_callback;
2846                         }
2847
2848                         loopdidcallbacks++;
2849                         funcdidcallbacks++;
2850
2851                         spin_lock(&log->l_icloglock);
2852                         ASSERT(iclog->ic_callback == NULL);
2853                         spin_unlock(&iclog->ic_callback_lock);
2854                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2855                                 iclog->ic_state = XLOG_STATE_DIRTY;
2856
2857                         /*
2858                          * Transition from DIRTY to ACTIVE if applicable.
2859                          * NOP if STATE_IOERROR.
2860                          */
2861                         xlog_state_clean_log(log);
2862
2863                         /* wake up threads waiting in xfs_log_force() */
2864                         wake_up_all(&iclog->ic_force_wait);
2865
2866                         iclog = iclog->ic_next;
2867                 } while (first_iclog != iclog);
2868
2869                 if (repeats > 5000) {
2870                         flushcnt += repeats;
2871                         repeats = 0;
2872                         xfs_warn(log->l_mp,
2873                                 "%s: possible infinite loop (%d iterations)",
2874                                 __func__, flushcnt);
2875                 }
2876         } while (!ioerrors && loopdidcallbacks);
2877
2878 #ifdef DEBUG
2879         /*
2880          * Make one last gasp attempt to see if iclogs are being left in limbo.
2881          * If the above loop finds an iclog earlier than the current iclog and
2882          * in one of the syncing states, the current iclog is put into
2883          * DO_CALLBACK and the callbacks are deferred to the completion of the
2884          * earlier iclog. Walk the iclogs in order and make sure that no iclog
2885          * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2886          * states.
2887          *
2888          * Note that SYNCING|IOABORT is a valid state so we cannot just check
2889          * for ic_state == SYNCING.
2890          */
2891         if (funcdidcallbacks) {
2892                 first_iclog = iclog = log->l_iclog;
2893                 do {
2894                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2895                         /*
2896                          * Terminate the loop if iclogs are found in states
2897                          * which will cause other threads to clean up iclogs.
2898                          *
2899                          * SYNCING - i/o completion will go through logs
2900                          * DONE_SYNC - interrupt thread should be waiting for
2901                          *              l_icloglock
2902                          * IOERROR - give up hope all ye who enter here
2903                          */
2904                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2905                             iclog->ic_state & XLOG_STATE_SYNCING ||
2906                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2907                             iclog->ic_state == XLOG_STATE_IOERROR )
2908                                 break;
2909                         iclog = iclog->ic_next;
2910                 } while (first_iclog != iclog);
2911         }
2912 #endif
2913
2914         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2915                 wake = 1;
2916         spin_unlock(&log->l_icloglock);
2917
2918         if (wake)
2919                 wake_up_all(&log->l_flush_wait);
2920 }
2921
2922
2923 /*
2924  * Finish transitioning this iclog to the dirty state.
2925  *
2926  * Make sure that we completely execute this routine only when this is
2927  * the last call to the iclog.  There is a good chance that iclog flushes,
2928  * when we reach the end of the physical log, get turned into 2 separate
2929  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2930  * routine.  By using the reference count bwritecnt, we guarantee that only
2931  * the second completion goes through.
2932  *
2933  * Callbacks could take time, so they are done outside the scope of the
2934  * global state machine log lock.
2935  */
2936 STATIC void
2937 xlog_state_done_syncing(
2938         xlog_in_core_t  *iclog,
2939         int             aborted)
2940 {
2941         struct xlog        *log = iclog->ic_log;
2942
2943         spin_lock(&log->l_icloglock);
2944
2945         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2946                iclog->ic_state == XLOG_STATE_IOERROR);
2947         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2948         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2949
2950
2951         /*
2952          * If we got an error, either on the first buffer, or in the case of
2953          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2954          * and none should ever be attempted to be written to disk
2955          * again.
2956          */
2957         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2958                 if (--iclog->ic_bwritecnt == 1) {
2959                         spin_unlock(&log->l_icloglock);
2960                         return;
2961                 }
2962                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2963         }
2964
2965         /*
2966          * Someone could be sleeping prior to writing out the next
2967          * iclog buffer, we wake them all, one will get to do the
2968          * I/O, the others get to wait for the result.
2969          */
2970         wake_up_all(&iclog->ic_write_wait);
2971         spin_unlock(&log->l_icloglock);
2972         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2973 }       /* xlog_state_done_syncing */
2974
2975
2976 /*
2977  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2978  * sleep.  We wait on the flush queue on the head iclog as that should be
2979  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2980  * we will wait here and all new writes will sleep until a sync completes.
2981  *
2982  * The in-core logs are used in a circular fashion. They are not used
2983  * out-of-order even when an iclog past the head is free.
2984  *
2985  * return:
2986  *      * log_offset where xlog_write() can start writing into the in-core
2987  *              log's data space.
2988  *      * in-core log pointer to which xlog_write() should write.
2989  *      * boolean indicating this is a continued write to an in-core log.
2990  *              If this is the last write, then the in-core log's offset field
2991  *              needs to be incremented, depending on the amount of data which
2992  *              is copied.
2993  */
2994 STATIC int
2995 xlog_state_get_iclog_space(
2996         struct xlog             *log,
2997         int                     len,
2998         struct xlog_in_core     **iclogp,
2999         struct xlog_ticket      *ticket,
3000         int                     *continued_write,
3001         int                     *logoffsetp)
3002 {
3003         int               log_offset;
3004         xlog_rec_header_t *head;
3005         xlog_in_core_t    *iclog;
3006         int               error;
3007
3008 restart:
3009         spin_lock(&log->l_icloglock);
3010         if (XLOG_FORCED_SHUTDOWN(log)) {
3011                 spin_unlock(&log->l_icloglock);
3012                 return -EIO;
3013         }
3014
3015         iclog = log->l_iclog;
3016         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
3017                 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
3018
3019                 /* Wait for log writes to have flushed */
3020                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
3021                 goto restart;
3022         }
3023
3024         head = &iclog->ic_header;
3025
3026         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
3027         log_offset = iclog->ic_offset;
3028
3029         /* On the 1st write to an iclog, figure out lsn.  This works
3030          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3031          * committing to.  If the offset is set, that's how many blocks
3032          * must be written.
3033          */
3034         if (log_offset == 0) {
3035                 ticket->t_curr_res -= log->l_iclog_hsize;
3036                 xlog_tic_add_region(ticket,
3037                                     log->l_iclog_hsize,
3038                                     XLOG_REG_TYPE_LRHEADER);
3039                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3040                 head->h_lsn = cpu_to_be64(
3041                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3042                 ASSERT(log->l_curr_block >= 0);
3043         }
3044
3045         /* If there is enough room to write everything, then do it.  Otherwise,
3046          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3047          * bit is on, so this will get flushed out.  Don't update ic_offset
3048          * until you know exactly how many bytes get copied.  Therefore, wait
3049          * until later to update ic_offset.
3050          *
3051          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3052          * can fit into remaining data section.
3053          */
3054         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3055                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3056
3057                 /*
3058                  * If I'm the only one writing to this iclog, sync it to disk.
3059                  * We need to do an atomic compare and decrement here to avoid
3060                  * racing with concurrent atomic_dec_and_lock() calls in
3061                  * xlog_state_release_iclog() when there is more than one
3062                  * reference to the iclog.
3063                  */
3064                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3065                         /* we are the only one */
3066                         spin_unlock(&log->l_icloglock);
3067                         error = xlog_state_release_iclog(log, iclog);
3068                         if (error)
3069                                 return error;
3070                 } else {
3071                         spin_unlock(&log->l_icloglock);
3072                 }
3073                 goto restart;
3074         }
3075
3076         /* Do we have enough room to write the full amount in the remainder
3077          * of this iclog?  Or must we continue a write on the next iclog and
3078          * mark this iclog as completely taken?  In the case where we switch
3079          * iclogs (to mark it taken), this particular iclog will release/sync
3080          * to disk in xlog_write().
3081          */
3082         if (len <= iclog->ic_size - iclog->ic_offset) {
3083                 *continued_write = 0;
3084                 iclog->ic_offset += len;
3085         } else {
3086                 *continued_write = 1;
3087                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3088         }
3089         *iclogp = iclog;
3090
3091         ASSERT(iclog->ic_offset <= iclog->ic_size);
3092         spin_unlock(&log->l_icloglock);
3093
3094         *logoffsetp = log_offset;
3095         return 0;
3096 }       /* xlog_state_get_iclog_space */
3097
3098 /* The first cnt-1 times through here we don't need to
3099  * move the grant write head because the permanent
3100  * reservation has reserved cnt times the unit amount.
3101  * Release part of current permanent unit reservation and
3102  * reset current reservation to be one units worth.  Also
3103  * move grant reservation head forward.
3104  */
3105 STATIC void
3106 xlog_regrant_reserve_log_space(
3107         struct xlog             *log,
3108         struct xlog_ticket      *ticket)
3109 {
3110         trace_xfs_log_regrant_reserve_enter(log, ticket);
3111
3112         if (ticket->t_cnt > 0)
3113                 ticket->t_cnt--;
3114
3115         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3116                                         ticket->t_curr_res);
3117         xlog_grant_sub_space(log, &log->l_write_head.grant,
3118                                         ticket->t_curr_res);
3119         ticket->t_curr_res = ticket->t_unit_res;
3120         xlog_tic_reset_res(ticket);
3121
3122         trace_xfs_log_regrant_reserve_sub(log, ticket);
3123
3124         /* just return if we still have some of the pre-reserved space */
3125         if (ticket->t_cnt > 0)
3126                 return;
3127
3128         xlog_grant_add_space(log, &log->l_reserve_head.grant,
3129                                         ticket->t_unit_res);
3130
3131         trace_xfs_log_regrant_reserve_exit(log, ticket);
3132
3133         ticket->t_curr_res = ticket->t_unit_res;
3134         xlog_tic_reset_res(ticket);
3135 }       /* xlog_regrant_reserve_log_space */
3136
3137
3138 /*
3139  * Give back the space left from a reservation.
3140  *
3141  * All the information we need to make a correct determination of space left
3142  * is present.  For non-permanent reservations, things are quite easy.  The
3143  * count should have been decremented to zero.  We only need to deal with the
3144  * space remaining in the current reservation part of the ticket.  If the
3145  * ticket contains a permanent reservation, there may be left over space which
3146  * needs to be released.  A count of N means that N-1 refills of the current
3147  * reservation can be done before we need to ask for more space.  The first
3148  * one goes to fill up the first current reservation.  Once we run out of
3149  * space, the count will stay at zero and the only space remaining will be
3150  * in the current reservation field.
3151  */
3152 STATIC void
3153 xlog_ungrant_log_space(
3154         struct xlog             *log,
3155         struct xlog_ticket      *ticket)
3156 {
3157         int     bytes;
3158
3159         if (ticket->t_cnt > 0)
3160                 ticket->t_cnt--;
3161
3162         trace_xfs_log_ungrant_enter(log, ticket);
3163         trace_xfs_log_ungrant_sub(log, ticket);
3164
3165         /*
3166          * If this is a permanent reservation ticket, we may be able to free
3167          * up more space based on the remaining count.
3168          */
3169         bytes = ticket->t_curr_res;
3170         if (ticket->t_cnt > 0) {
3171                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3172                 bytes += ticket->t_unit_res*ticket->t_cnt;
3173         }
3174
3175         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3176         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3177
3178         trace_xfs_log_ungrant_exit(log, ticket);
3179
3180         xfs_log_space_wake(log->l_mp);
3181 }
3182
3183 /*
3184  * Flush iclog to disk if this is the last reference to the given iclog and
3185  * the WANT_SYNC bit is set.
3186  *
3187  * When this function is entered, the iclog is not necessarily in the
3188  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3189  *
3190  *
3191  */
3192 STATIC int
3193 xlog_state_release_iclog(
3194         struct xlog             *log,
3195         struct xlog_in_core     *iclog)
3196 {
3197         int             sync = 0;       /* do we sync? */
3198
3199         if (iclog->ic_state & XLOG_STATE_IOERROR)
3200                 return -EIO;
3201
3202         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3203         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3204                 return 0;
3205
3206         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3207                 spin_unlock(&log->l_icloglock);
3208                 return -EIO;
3209         }
3210         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3211                iclog->ic_state == XLOG_STATE_WANT_SYNC);
3212
3213         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3214                 /* update tail before writing to iclog */
3215                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3216                 sync++;
3217                 iclog->ic_state = XLOG_STATE_SYNCING;
3218                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3219                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3220                 /* cycle incremented when incrementing curr_block */
3221         }
3222         spin_unlock(&log->l_icloglock);
3223
3224         /*
3225          * We let the log lock go, so it's possible that we hit a log I/O
3226          * error or some other SHUTDOWN condition that marks the iclog
3227          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3228          * this iclog has consistent data, so we ignore IOERROR
3229          * flags after this point.
3230          */
3231         if (sync)
3232                 return xlog_sync(log, iclog);
3233         return 0;
3234 }       /* xlog_state_release_iclog */
3235
3236
3237 /*
3238  * This routine will mark the current iclog in the ring as WANT_SYNC
3239  * and move the current iclog pointer to the next iclog in the ring.
3240  * When this routine is called from xlog_state_get_iclog_space(), the
3241  * exact size of the iclog has not yet been determined.  All we know is
3242  * that every data block.  We have run out of space in this log record.
3243  */
3244 STATIC void
3245 xlog_state_switch_iclogs(
3246         struct xlog             *log,
3247         struct xlog_in_core     *iclog,
3248         int                     eventual_size)
3249 {
3250         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3251         if (!eventual_size)
3252                 eventual_size = iclog->ic_offset;
3253         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3254         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3255         log->l_prev_block = log->l_curr_block;
3256         log->l_prev_cycle = log->l_curr_cycle;
3257
3258         /* roll log?: ic_offset changed later */
3259         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3260
3261         /* Round up to next log-sunit */
3262         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3263             log->l_mp->m_sb.sb_logsunit > 1) {
3264                 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3265                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3266         }
3267
3268         if (log->l_curr_block >= log->l_logBBsize) {
3269                 /*
3270                  * Rewind the current block before the cycle is bumped to make
3271                  * sure that the combined LSN never transiently moves forward
3272                  * when the log wraps to the next cycle. This is to support the
3273                  * unlocked sample of these fields from xlog_valid_lsn(). Most
3274                  * other cases should acquire l_icloglock.
3275                  */
3276                 log->l_curr_block -= log->l_logBBsize;
3277                 ASSERT(log->l_curr_block >= 0);
3278                 smp_wmb();
3279                 log->l_curr_cycle++;
3280                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3281                         log->l_curr_cycle++;
3282         }
3283         ASSERT(iclog == log->l_iclog);
3284         log->l_iclog = iclog->ic_next;
3285 }       /* xlog_state_switch_iclogs */
3286
3287 /*
3288  * Write out all data in the in-core log as of this exact moment in time.
3289  *
3290  * Data may be written to the in-core log during this call.  However,
3291  * we don't guarantee this data will be written out.  A change from past
3292  * implementation means this routine will *not* write out zero length LRs.
3293  *
3294  * Basically, we try and perform an intelligent scan of the in-core logs.
3295  * If we determine there is no flushable data, we just return.  There is no
3296  * flushable data if:
3297  *
3298  *      1. the current iclog is active and has no data; the previous iclog
3299  *              is in the active or dirty state.
3300  *      2. the current iclog is drity, and the previous iclog is in the
3301  *              active or dirty state.
3302  *
3303  * We may sleep if:
3304  *
3305  *      1. the current iclog is not in the active nor dirty state.
3306  *      2. the current iclog dirty, and the previous iclog is not in the
3307  *              active nor dirty state.
3308  *      3. the current iclog is active, and there is another thread writing
3309  *              to this particular iclog.
3310  *      4. a) the current iclog is active and has no other writers
3311  *         b) when we return from flushing out this iclog, it is still
3312  *              not in the active nor dirty state.
3313  */
3314 int
3315 xfs_log_force(
3316         struct xfs_mount        *mp,
3317         uint                    flags)
3318 {
3319         struct xlog             *log = mp->m_log;
3320         struct xlog_in_core     *iclog;
3321         xfs_lsn_t               lsn;
3322
3323         XFS_STATS_INC(mp, xs_log_force);
3324         trace_xfs_log_force(mp, 0, _RET_IP_);
3325
3326         xlog_cil_force(log);
3327
3328         spin_lock(&log->l_icloglock);
3329         iclog = log->l_iclog;
3330         if (iclog->ic_state & XLOG_STATE_IOERROR)
3331                 goto out_error;
3332
3333         if (iclog->ic_state == XLOG_STATE_DIRTY ||
3334             (iclog->ic_state == XLOG_STATE_ACTIVE &&
3335              atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3336                 /*
3337                  * If the head is dirty or (active and empty), then we need to
3338                  * look at the previous iclog.
3339                  *
3340                  * If the previous iclog is active or dirty we are done.  There
3341                  * is nothing to sync out. Otherwise, we attach ourselves to the
3342                  * previous iclog and go to sleep.
3343                  */
3344                 iclog = iclog->ic_prev;
3345                 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3346                     iclog->ic_state == XLOG_STATE_DIRTY)
3347                         goto out_unlock;
3348         } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3349                 if (atomic_read(&iclog->ic_refcnt) == 0) {
3350                         /*
3351                          * We are the only one with access to this iclog.
3352                          *
3353                          * Flush it out now.  There should be a roundoff of zero
3354                          * to show that someone has already taken care of the
3355                          * roundoff from the previous sync.
3356                          */
3357                         atomic_inc(&iclog->ic_refcnt);
3358                         lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3359                         xlog_state_switch_iclogs(log, iclog, 0);
3360                         spin_unlock(&log->l_icloglock);
3361
3362                         if (xlog_state_release_iclog(log, iclog))
3363                                 return -EIO;
3364
3365                         spin_lock(&log->l_icloglock);
3366                         if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3367                             iclog->ic_state == XLOG_STATE_DIRTY)
3368                                 goto out_unlock;
3369                 } else {
3370                         /*
3371                          * Someone else is writing to this iclog.
3372                          *
3373                          * Use its call to flush out the data.  However, the
3374                          * other thread may not force out this LR, so we mark
3375                          * it WANT_SYNC.
3376                          */
3377                         xlog_state_switch_iclogs(log, iclog, 0);
3378                 }
3379         } else {
3380                 /*
3381                  * If the head iclog is not active nor dirty, we just attach
3382                  * ourselves to the head and go to sleep if necessary.
3383                  */
3384                 ;
3385         }
3386
3387         if (!(flags & XFS_LOG_SYNC))
3388                 goto out_unlock;
3389
3390         if (iclog->ic_state & XLOG_STATE_IOERROR)
3391                 goto out_error;
3392         XFS_STATS_INC(mp, xs_log_force_sleep);
3393         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3394         if (iclog->ic_state & XLOG_STATE_IOERROR)
3395                 return -EIO;
3396         return 0;
3397
3398 out_unlock:
3399         spin_unlock(&log->l_icloglock);
3400         return 0;
3401 out_error:
3402         spin_unlock(&log->l_icloglock);
3403         return -EIO;
3404 }
3405
3406 static int
3407 __xfs_log_force_lsn(
3408         struct xfs_mount        *mp,
3409         xfs_lsn_t               lsn,
3410         uint                    flags,
3411         int                     *log_flushed,
3412         bool                    already_slept)
3413 {
3414         struct xlog             *log = mp->m_log;
3415         struct xlog_in_core     *iclog;
3416
3417         spin_lock(&log->l_icloglock);
3418         iclog = log->l_iclog;
3419         if (iclog->ic_state & XLOG_STATE_IOERROR)
3420                 goto out_error;
3421
3422         while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3423                 iclog = iclog->ic_next;
3424                 if (iclog == log->l_iclog)
3425                         goto out_unlock;
3426         }
3427
3428         if (iclog->ic_state == XLOG_STATE_DIRTY)
3429                 goto out_unlock;
3430
3431         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3432                 /*
3433                  * We sleep here if we haven't already slept (e.g. this is the
3434                  * first time we've looked at the correct iclog buf) and the
3435                  * buffer before us is going to be sync'ed.  The reason for this
3436                  * is that if we are doing sync transactions here, by waiting
3437                  * for the previous I/O to complete, we can allow a few more
3438                  * transactions into this iclog before we close it down.
3439                  *
3440                  * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3441                  * refcnt so we can release the log (which drops the ref count).
3442                  * The state switch keeps new transaction commits from using
3443                  * this buffer.  When the current commits finish writing into
3444                  * the buffer, the refcount will drop to zero and the buffer
3445                  * will go out then.
3446                  */
3447                 if (!already_slept &&
3448                     (iclog->ic_prev->ic_state &
3449                      (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3450                         ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3451
3452                         XFS_STATS_INC(mp, xs_log_force_sleep);
3453
3454                         xlog_wait(&iclog->ic_prev->ic_write_wait,
3455                                         &log->l_icloglock);
3456                         return -EAGAIN;
3457                 }
3458                 atomic_inc(&iclog->ic_refcnt);
3459                 xlog_state_switch_iclogs(log, iclog, 0);
3460                 spin_unlock(&log->l_icloglock);
3461                 if (xlog_state_release_iclog(log, iclog))
3462                         return -EIO;
3463                 if (log_flushed)
3464                         *log_flushed = 1;
3465                 spin_lock(&log->l_icloglock);
3466         }
3467
3468         if (!(flags & XFS_LOG_SYNC) ||
3469             (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3470                 goto out_unlock;
3471
3472         if (iclog->ic_state & XLOG_STATE_IOERROR)
3473                 goto out_error;
3474
3475         XFS_STATS_INC(mp, xs_log_force_sleep);
3476         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3477         if (iclog->ic_state & XLOG_STATE_IOERROR)
3478                 return -EIO;
3479         return 0;
3480
3481 out_unlock:
3482         spin_unlock(&log->l_icloglock);
3483         return 0;
3484 out_error:
3485         spin_unlock(&log->l_icloglock);
3486         return -EIO;
3487 }
3488
3489 /*
3490  * Force the in-core log to disk for a specific LSN.
3491  *
3492  * Find in-core log with lsn.
3493  *      If it is in the DIRTY state, just return.
3494  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3495  *              state and go to sleep or return.
3496  *      If it is in any other state, go to sleep or return.
3497  *
3498  * Synchronous forces are implemented with a wait queue.  All callers trying
3499  * to force a given lsn to disk must wait on the queue attached to the
3500  * specific in-core log.  When given in-core log finally completes its write
3501  * to disk, that thread will wake up all threads waiting on the queue.
3502  */
3503 int
3504 xfs_log_force_lsn(
3505         struct xfs_mount        *mp,
3506         xfs_lsn_t               lsn,
3507         uint                    flags,
3508         int                     *log_flushed)
3509 {
3510         int                     ret;
3511         ASSERT(lsn != 0);
3512
3513         XFS_STATS_INC(mp, xs_log_force);
3514         trace_xfs_log_force(mp, lsn, _RET_IP_);
3515
3516         lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3517         if (lsn == NULLCOMMITLSN)
3518                 return 0;
3519
3520         ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3521         if (ret == -EAGAIN)
3522                 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3523         return ret;
3524 }
3525
3526 /*
3527  * Called when we want to mark the current iclog as being ready to sync to
3528  * disk.
3529  */
3530 STATIC void
3531 xlog_state_want_sync(
3532         struct xlog             *log,
3533         struct xlog_in_core     *iclog)
3534 {
3535         assert_spin_locked(&log->l_icloglock);
3536
3537         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3538                 xlog_state_switch_iclogs(log, iclog, 0);
3539         } else {
3540                 ASSERT(iclog->ic_state &
3541                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3542         }
3543 }
3544
3545
3546 /*****************************************************************************
3547  *
3548  *              TICKET functions
3549  *
3550  *****************************************************************************
3551  */
3552
3553 /*
3554  * Free a used ticket when its refcount falls to zero.
3555  */
3556 void
3557 xfs_log_ticket_put(
3558         xlog_ticket_t   *ticket)
3559 {
3560         ASSERT(atomic_read(&ticket->t_ref) > 0);
3561         if (atomic_dec_and_test(&ticket->t_ref))
3562                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3563 }
3564
3565 xlog_ticket_t *
3566 xfs_log_ticket_get(
3567         xlog_ticket_t   *ticket)
3568 {
3569         ASSERT(atomic_read(&ticket->t_ref) > 0);
3570         atomic_inc(&ticket->t_ref);
3571         return ticket;
3572 }
3573
3574 /*
3575  * Figure out the total log space unit (in bytes) that would be
3576  * required for a log ticket.
3577  */
3578 int
3579 xfs_log_calc_unit_res(
3580         struct xfs_mount        *mp,
3581         int                     unit_bytes)
3582 {
3583         struct xlog             *log = mp->m_log;
3584         int                     iclog_space;
3585         uint                    num_headers;
3586
3587         /*
3588          * Permanent reservations have up to 'cnt'-1 active log operations
3589          * in the log.  A unit in this case is the amount of space for one
3590          * of these log operations.  Normal reservations have a cnt of 1
3591          * and their unit amount is the total amount of space required.
3592          *
3593          * The following lines of code account for non-transaction data
3594          * which occupy space in the on-disk log.
3595          *
3596          * Normal form of a transaction is:
3597          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3598          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3599          *
3600          * We need to account for all the leadup data and trailer data
3601          * around the transaction data.
3602          * And then we need to account for the worst case in terms of using
3603          * more space.
3604          * The worst case will happen if:
3605          * - the placement of the transaction happens to be such that the
3606          *   roundoff is at its maximum
3607          * - the transaction data is synced before the commit record is synced
3608          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3609          *   Therefore the commit record is in its own Log Record.
3610          *   This can happen as the commit record is called with its
3611          *   own region to xlog_write().
3612          *   This then means that in the worst case, roundoff can happen for
3613          *   the commit-rec as well.
3614          *   The commit-rec is smaller than padding in this scenario and so it is
3615          *   not added separately.
3616          */
3617
3618         /* for trans header */
3619         unit_bytes += sizeof(xlog_op_header_t);
3620         unit_bytes += sizeof(xfs_trans_header_t);
3621
3622         /* for start-rec */
3623         unit_bytes += sizeof(xlog_op_header_t);
3624
3625         /*
3626          * for LR headers - the space for data in an iclog is the size minus
3627          * the space used for the headers. If we use the iclog size, then we
3628          * undercalculate the number of headers required.
3629          *
3630          * Furthermore - the addition of op headers for split-recs might
3631          * increase the space required enough to require more log and op
3632          * headers, so take that into account too.
3633          *
3634          * IMPORTANT: This reservation makes the assumption that if this
3635          * transaction is the first in an iclog and hence has the LR headers
3636          * accounted to it, then the remaining space in the iclog is
3637          * exclusively for this transaction.  i.e. if the transaction is larger
3638          * than the iclog, it will be the only thing in that iclog.
3639          * Fundamentally, this means we must pass the entire log vector to
3640          * xlog_write to guarantee this.
3641          */
3642         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3643         num_headers = howmany(unit_bytes, iclog_space);
3644
3645         /* for split-recs - ophdrs added when data split over LRs */
3646         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3647
3648         /* add extra header reservations if we overrun */
3649         while (!num_headers ||
3650                howmany(unit_bytes, iclog_space) > num_headers) {
3651                 unit_bytes += sizeof(xlog_op_header_t);
3652                 num_headers++;
3653         }
3654         unit_bytes += log->l_iclog_hsize * num_headers;
3655
3656         /* for commit-rec LR header - note: padding will subsume the ophdr */
3657         unit_bytes += log->l_iclog_hsize;
3658
3659         /* for roundoff padding for transaction data and one for commit record */
3660         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3661                 /* log su roundoff */
3662                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3663         } else {
3664                 /* BB roundoff */
3665                 unit_bytes += 2 * BBSIZE;
3666         }
3667
3668         return unit_bytes;
3669 }
3670
3671 /*
3672  * Allocate and initialise a new log ticket.
3673  */
3674 struct xlog_ticket *
3675 xlog_ticket_alloc(
3676         struct xlog             *log,
3677         int                     unit_bytes,
3678         int                     cnt,
3679         char                    client,
3680         bool                    permanent,
3681         xfs_km_flags_t          alloc_flags)
3682 {
3683         struct xlog_ticket      *tic;
3684         int                     unit_res;
3685
3686         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3687         if (!tic)
3688                 return NULL;
3689
3690         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3691
3692         atomic_set(&tic->t_ref, 1);
3693         tic->t_task             = current;
3694         INIT_LIST_HEAD(&tic->t_queue);
3695         tic->t_unit_res         = unit_res;
3696         tic->t_curr_res         = unit_res;
3697         tic->t_cnt              = cnt;
3698         tic->t_ocnt             = cnt;
3699         tic->t_tid              = prandom_u32();
3700         tic->t_clientid         = client;
3701         tic->t_flags            = XLOG_TIC_INITED;
3702         if (permanent)
3703                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3704
3705         xlog_tic_reset_res(tic);
3706
3707         return tic;
3708 }
3709
3710
3711 /******************************************************************************
3712  *
3713  *              Log debug routines
3714  *
3715  ******************************************************************************
3716  */
3717 #if defined(DEBUG)
3718 /*
3719  * Make sure that the destination ptr is within the valid data region of
3720  * one of the iclogs.  This uses backup pointers stored in a different
3721  * part of the log in case we trash the log structure.
3722  */
3723 STATIC void
3724 xlog_verify_dest_ptr(
3725         struct xlog     *log,
3726         void            *ptr)
3727 {
3728         int i;
3729         int good_ptr = 0;
3730
3731         for (i = 0; i < log->l_iclog_bufs; i++) {
3732                 if (ptr >= log->l_iclog_bak[i] &&
3733                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3734                         good_ptr++;
3735         }
3736
3737         if (!good_ptr)
3738                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3739 }
3740
3741 /*
3742  * Check to make sure the grant write head didn't just over lap the tail.  If
3743  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3744  * the cycles differ by exactly one and check the byte count.
3745  *
3746  * This check is run unlocked, so can give false positives. Rather than assert
3747  * on failures, use a warn-once flag and a panic tag to allow the admin to
3748  * determine if they want to panic the machine when such an error occurs. For
3749  * debug kernels this will have the same effect as using an assert but, unlinke
3750  * an assert, it can be turned off at runtime.
3751  */
3752 STATIC void
3753 xlog_verify_grant_tail(
3754         struct xlog     *log)
3755 {
3756         int             tail_cycle, tail_blocks;
3757         int             cycle, space;
3758
3759         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3760         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3761         if (tail_cycle != cycle) {
3762                 if (cycle - 1 != tail_cycle &&
3763                     !(log->l_flags & XLOG_TAIL_WARN)) {
3764                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3765                                 "%s: cycle - 1 != tail_cycle", __func__);
3766                         log->l_flags |= XLOG_TAIL_WARN;
3767                 }
3768
3769                 if (space > BBTOB(tail_blocks) &&
3770                     !(log->l_flags & XLOG_TAIL_WARN)) {
3771                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3772                                 "%s: space > BBTOB(tail_blocks)", __func__);
3773                         log->l_flags |= XLOG_TAIL_WARN;
3774                 }
3775         }
3776 }
3777
3778 /* check if it will fit */
3779 STATIC void
3780 xlog_verify_tail_lsn(
3781         struct xlog             *log,
3782         struct xlog_in_core     *iclog,
3783         xfs_lsn_t               tail_lsn)
3784 {
3785     int blocks;
3786
3787     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3788         blocks =
3789             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3790         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3791                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3792     } else {
3793         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3794
3795         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3796                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3797
3798         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3799         if (blocks < BTOBB(iclog->ic_offset) + 1)
3800                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3801     }
3802 }       /* xlog_verify_tail_lsn */
3803
3804 /*
3805  * Perform a number of checks on the iclog before writing to disk.
3806  *
3807  * 1. Make sure the iclogs are still circular
3808  * 2. Make sure we have a good magic number
3809  * 3. Make sure we don't have magic numbers in the data
3810  * 4. Check fields of each log operation header for:
3811  *      A. Valid client identifier
3812  *      B. tid ptr value falls in valid ptr space (user space code)
3813  *      C. Length in log record header is correct according to the
3814  *              individual operation headers within record.
3815  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3816  *      log, check the preceding blocks of the physical log to make sure all
3817  *      the cycle numbers agree with the current cycle number.
3818  */
3819 STATIC void
3820 xlog_verify_iclog(
3821         struct xlog             *log,
3822         struct xlog_in_core     *iclog,
3823         int                     count,
3824         bool                    syncing)
3825 {
3826         xlog_op_header_t        *ophead;
3827         xlog_in_core_t          *icptr;
3828         xlog_in_core_2_t        *xhdr;
3829         void                    *base_ptr, *ptr, *p;
3830         ptrdiff_t               field_offset;
3831         uint8_t                 clientid;
3832         int                     len, i, j, k, op_len;
3833         int                     idx;
3834
3835         /* check validity of iclog pointers */
3836         spin_lock(&log->l_icloglock);
3837         icptr = log->l_iclog;
3838         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3839                 ASSERT(icptr);
3840
3841         if (icptr != log->l_iclog)
3842                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3843         spin_unlock(&log->l_icloglock);
3844
3845         /* check log magic numbers */
3846         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3847                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3848
3849         base_ptr = ptr = &iclog->ic_header;
3850         p = &iclog->ic_header;
3851         for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3852                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3853                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3854                                 __func__);
3855         }
3856
3857         /* check fields */
3858         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3859         base_ptr = ptr = iclog->ic_datap;
3860         ophead = ptr;
3861         xhdr = iclog->ic_data;
3862         for (i = 0; i < len; i++) {
3863                 ophead = ptr;
3864
3865                 /* clientid is only 1 byte */
3866                 p = &ophead->oh_clientid;
3867                 field_offset = p - base_ptr;
3868                 if (!syncing || (field_offset & 0x1ff)) {
3869                         clientid = ophead->oh_clientid;
3870                 } else {
3871                         idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3872                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3873                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3874                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3875                                 clientid = xlog_get_client_id(
3876                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3877                         } else {
3878                                 clientid = xlog_get_client_id(
3879                                         iclog->ic_header.h_cycle_data[idx]);
3880                         }
3881                 }
3882                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3883                         xfs_warn(log->l_mp,
3884                                 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3885                                 __func__, clientid, ophead,
3886                                 (unsigned long)field_offset);
3887
3888                 /* check length */
3889                 p = &ophead->oh_len;
3890                 field_offset = p - base_ptr;
3891                 if (!syncing || (field_offset & 0x1ff)) {
3892                         op_len = be32_to_cpu(ophead->oh_len);
3893                 } else {
3894                         idx = BTOBBT((uintptr_t)&ophead->oh_len -
3895                                     (uintptr_t)iclog->ic_datap);
3896                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3897                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3898                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3899                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3900                         } else {
3901                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3902                         }
3903                 }
3904                 ptr += sizeof(xlog_op_header_t) + op_len;
3905         }
3906 }       /* xlog_verify_iclog */
3907 #endif
3908
3909 /*
3910  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3911  */
3912 STATIC int
3913 xlog_state_ioerror(
3914         struct xlog     *log)
3915 {
3916         xlog_in_core_t  *iclog, *ic;
3917
3918         iclog = log->l_iclog;
3919         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3920                 /*
3921                  * Mark all the incore logs IOERROR.
3922                  * From now on, no log flushes will result.
3923                  */
3924                 ic = iclog;
3925                 do {
3926                         ic->ic_state = XLOG_STATE_IOERROR;
3927                         ic = ic->ic_next;
3928                 } while (ic != iclog);
3929                 return 0;
3930         }
3931         /*
3932          * Return non-zero, if state transition has already happened.
3933          */
3934         return 1;
3935 }
3936
3937 /*
3938  * This is called from xfs_force_shutdown, when we're forcibly
3939  * shutting down the filesystem, typically because of an IO error.
3940  * Our main objectives here are to make sure that:
3941  *      a. if !logerror, flush the logs to disk. Anything modified
3942  *         after this is ignored.
3943  *      b. the filesystem gets marked 'SHUTDOWN' for all interested
3944  *         parties to find out, 'atomically'.
3945  *      c. those who're sleeping on log reservations, pinned objects and
3946  *          other resources get woken up, and be told the bad news.
3947  *      d. nothing new gets queued up after (b) and (c) are done.
3948  *
3949  * Note: for the !logerror case we need to flush the regions held in memory out
3950  * to disk first. This needs to be done before the log is marked as shutdown,
3951  * otherwise the iclog writes will fail.
3952  */
3953 int
3954 xfs_log_force_umount(
3955         struct xfs_mount        *mp,
3956         int                     logerror)
3957 {
3958         struct xlog     *log;
3959         int             retval;
3960
3961         log = mp->m_log;
3962
3963         /*
3964          * If this happens during log recovery, don't worry about
3965          * locking; the log isn't open for business yet.
3966          */
3967         if (!log ||
3968             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3969                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3970                 if (mp->m_sb_bp)
3971                         mp->m_sb_bp->b_flags |= XBF_DONE;
3972                 return 0;
3973         }
3974
3975         /*
3976          * Somebody could've already done the hard work for us.
3977          * No need to get locks for this.
3978          */
3979         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3980                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3981                 return 1;
3982         }
3983
3984         /*
3985          * Flush all the completed transactions to disk before marking the log
3986          * being shut down. We need to do it in this order to ensure that
3987          * completed operations are safely on disk before we shut down, and that
3988          * we don't have to issue any buffer IO after the shutdown flags are set
3989          * to guarantee this.
3990          */
3991         if (!logerror)
3992                 xfs_log_force(mp, XFS_LOG_SYNC);
3993
3994         /*
3995          * mark the filesystem and the as in a shutdown state and wake
3996          * everybody up to tell them the bad news.
3997          */
3998         spin_lock(&log->l_icloglock);
3999         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
4000         if (mp->m_sb_bp)
4001                 mp->m_sb_bp->b_flags |= XBF_DONE;
4002
4003         /*
4004          * Mark the log and the iclogs with IO error flags to prevent any
4005          * further log IO from being issued or completed.
4006          */
4007         log->l_flags |= XLOG_IO_ERROR;
4008         retval = xlog_state_ioerror(log);
4009         spin_unlock(&log->l_icloglock);
4010
4011         /*
4012          * We don't want anybody waiting for log reservations after this. That
4013          * means we have to wake up everybody queued up on reserveq as well as
4014          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
4015          * we don't enqueue anything once the SHUTDOWN flag is set, and this
4016          * action is protected by the grant locks.
4017          */
4018         xlog_grant_head_wake_all(&log->l_reserve_head);
4019         xlog_grant_head_wake_all(&log->l_write_head);
4020
4021         /*
4022          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4023          * as if the log writes were completed. The abort handling in the log
4024          * item committed callback functions will do this again under lock to
4025          * avoid races.
4026          */
4027         wake_up_all(&log->l_cilp->xc_commit_wait);
4028         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
4029
4030 #ifdef XFSERRORDEBUG
4031         {
4032                 xlog_in_core_t  *iclog;
4033
4034                 spin_lock(&log->l_icloglock);
4035                 iclog = log->l_iclog;
4036                 do {
4037                         ASSERT(iclog->ic_callback == 0);
4038                         iclog = iclog->ic_next;
4039                 } while (iclog != log->l_iclog);
4040                 spin_unlock(&log->l_icloglock);
4041         }
4042 #endif
4043         /* return non-zero if log IOERROR transition had already happened */
4044         return retval;
4045 }
4046
4047 STATIC int
4048 xlog_iclogs_empty(
4049         struct xlog     *log)
4050 {
4051         xlog_in_core_t  *iclog;
4052
4053         iclog = log->l_iclog;
4054         do {
4055                 /* endianness does not matter here, zero is zero in
4056                  * any language.
4057                  */
4058                 if (iclog->ic_header.h_num_logops)
4059                         return 0;
4060                 iclog = iclog->ic_next;
4061         } while (iclog != log->l_iclog);
4062         return 1;
4063 }
4064
4065 /*
4066  * Verify that an LSN stamped into a piece of metadata is valid. This is
4067  * intended for use in read verifiers on v5 superblocks.
4068  */
4069 bool
4070 xfs_log_check_lsn(
4071         struct xfs_mount        *mp,
4072         xfs_lsn_t               lsn)
4073 {
4074         struct xlog             *log = mp->m_log;
4075         bool                    valid;
4076
4077         /*
4078          * norecovery mode skips mount-time log processing and unconditionally
4079          * resets the in-core LSN. We can't validate in this mode, but
4080          * modifications are not allowed anyways so just return true.
4081          */
4082         if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4083                 return true;
4084
4085         /*
4086          * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4087          * handled by recovery and thus safe to ignore here.
4088          */
4089         if (lsn == NULLCOMMITLSN)
4090                 return true;
4091
4092         valid = xlog_valid_lsn(mp->m_log, lsn);
4093
4094         /* warn the user about what's gone wrong before verifier failure */
4095         if (!valid) {
4096                 spin_lock(&log->l_icloglock);
4097                 xfs_warn(mp,
4098 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4099 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4100                          CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4101                          log->l_curr_cycle, log->l_curr_block);
4102                 spin_unlock(&log->l_icloglock);
4103         }
4104
4105         return valid;
4106 }
4107
4108 bool
4109 xfs_log_in_recovery(
4110         struct xfs_mount        *mp)
4111 {
4112         struct xlog             *log = mp->m_log;
4113
4114         return log->l_flags & XLOG_ACTIVE_RECOVERY;
4115 }