fs/xfs/xfs_log.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   4  * All Rights Reserved.
   5  */
   6 #include "xfs.h"
   7 #include "xfs_fs.h"
   8 #include "xfs_shared.h"
   9 #include "xfs_format.h"
  10 #include "xfs_log_format.h"
  11 #include "xfs_trans_resv.h"
  12 #include "xfs_mount.h"
  13 #include "xfs_errortag.h"
  14 #include "xfs_error.h"
  15 #include "xfs_trans.h"
  16 #include "xfs_trans_priv.h"
  17 #include "xfs_log.h"
  18 #include "xfs_log_priv.h"
  19 #include "xfs_log_recover.h"
  20 #include "xfs_inode.h"
  21 #include "xfs_trace.h"
  22 #include "xfs_fsops.h"
  23 #include "xfs_cksum.h"
  24 #include "xfs_sysfs.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_health.h"
  27
  28 kmem_zone_t     *xfs_log_ticket_zone;
  29
  30 /* Local miscellaneous function prototypes */
  31 STATIC int
  32 xlog_commit_record(
  33         struct xlog             *log,
  34         struct xlog_ticket      *ticket,
  35         struct xlog_in_core     **iclog,
  36         xfs_lsn_t               *commitlsnp);
  37
  38 STATIC struct xlog *
  39 xlog_alloc_log(
  40         struct xfs_mount        *mp,
  41         struct xfs_buftarg      *log_target,
  42         xfs_daddr_t             blk_offset,
  43         int                     num_bblks);
  44 STATIC int
  45 xlog_space_left(
  46         struct xlog             *log,
  47         atomic64_t              *head);
  48 STATIC int
  49 xlog_sync(
  50         struct xlog             *log,
  51         struct xlog_in_core     *iclog);
  52 STATIC void
  53 xlog_dealloc_log(
  54         struct xlog             *log);
  55
  56 /* local state machine functions */
  57 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  58 STATIC void
  59 xlog_state_do_callback(
  60         struct xlog             *log,
  61         int                     aborted,
  62         struct xlog_in_core     *iclog);
  63 STATIC int
  64 xlog_state_get_iclog_space(
  65         struct xlog             *log,
  66         int                     len,
  67         struct xlog_in_core     **iclog,
  68         struct xlog_ticket      *ticket,
  69         int                     *continued_write,
  70         int                     *logoffsetp);
  71 STATIC int
  72 xlog_state_release_iclog(
  73         struct xlog             *log,
  74         struct xlog_in_core     *iclog);
  75 STATIC void
  76 xlog_state_switch_iclogs(
  77         struct xlog             *log,
  78         struct xlog_in_core     *iclog,
  79         int                     eventual_size);
  80 STATIC void
  81 xlog_state_want_sync(
  82         struct xlog             *log,
  83         struct xlog_in_core     *iclog);
  84
  85 STATIC void
  86 xlog_grant_push_ail(
  87         struct xlog             *log,
  88         int                     need_bytes);
  89 STATIC void
  90 xlog_regrant_reserve_log_space(
  91         struct xlog             *log,
  92         struct xlog_ticket      *ticket);
  93 STATIC void
  94 xlog_ungrant_log_space(
  95         struct xlog             *log,
  96         struct xlog_ticket      *ticket);
  97
  98 #if defined(DEBUG)
  99 STATIC void
 100 xlog_verify_dest_ptr(
 101         struct xlog             *log,
 102         void                    *ptr);
 103 STATIC void
 104 xlog_verify_grant_tail(
 105         struct xlog *log);
 106 STATIC void
 107 xlog_verify_iclog(
 108         struct xlog             *log,
 109         struct xlog_in_core     *iclog,
 110         int                     count,
 111         bool                    syncing);
 112 STATIC void
 113 xlog_verify_tail_lsn(
 114         struct xlog             *log,
 115         struct xlog_in_core     *iclog,
 116         xfs_lsn_t               tail_lsn);
 117 #else
 118 #define xlog_verify_dest_ptr(a,b)
 119 #define xlog_verify_grant_tail(a)
 120 #define xlog_verify_iclog(a,b,c,d)
 121 #define xlog_verify_tail_lsn(a,b,c)
 122 #endif
 123
 124 STATIC int
 125 xlog_iclogs_empty(
 126         struct xlog             *log);
 127
 128 static void
 129 xlog_grant_sub_space(
 130         struct xlog             *log,
 131         atomic64_t              *head,
 132         int                     bytes)
 133 {
 134         int64_t head_val = atomic64_read(head);
 135         int64_t new, old;
 136
 137         do {
 138                 int     cycle, space;
 139
 140                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 141
 142                 space -= bytes;
 143                 if (space < 0) {
 144                         space += log->l_logsize;
 145                         cycle--;
 146                 }
 147
 148                 old = head_val;
 149                 new = xlog_assign_grant_head_val(cycle, space);
 150                 head_val = atomic64_cmpxchg(head, old, new);
 151         } while (head_val != old);
 152 }
 153
 154 static void
 155 xlog_grant_add_space(
 156         struct xlog             *log,
 157         atomic64_t              *head,
 158         int                     bytes)
 159 {
 160         int64_t head_val = atomic64_read(head);
 161         int64_t new, old;
 162
 163         do {
 164                 int             tmp;
 165                 int             cycle, space;
 166
 167                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 168
 169                 tmp = log->l_logsize - space;
 170                 if (tmp > bytes)
 171                         space += bytes;
 172                 else {
 173                         space = bytes - tmp;
 174                         cycle++;
 175                 }
 176
 177                 old = head_val;
 178                 new = xlog_assign_grant_head_val(cycle, space);
 179                 head_val = atomic64_cmpxchg(head, old, new);
 180         } while (head_val != old);
 181 }
 182
 183 STATIC void
 184 xlog_grant_head_init(
 185         struct xlog_grant_head  *head)
 186 {
 187         xlog_assign_grant_head(&head->grant, 1, 0);
 188         INIT_LIST_HEAD(&head->waiters);
 189         spin_lock_init(&head->lock);
 190 }
 191
 192 STATIC void
 193 xlog_grant_head_wake_all(
 194         struct xlog_grant_head  *head)
 195 {
 196         struct xlog_ticket      *tic;
 197
 198         spin_lock(&head->lock);
 199         list_for_each_entry(tic, &head->waiters, t_queue)
 200                 wake_up_process(tic->t_task);
 201         spin_unlock(&head->lock);
 202 }
 203
 204 static inline int
 205 xlog_ticket_reservation(
 206         struct xlog             *log,
 207         struct xlog_grant_head  *head,
 208         struct xlog_ticket      *tic)
 209 {
 210         if (head == &log->l_write_head) {
 211                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 212                 return tic->t_unit_res;
 213         } else {
 214                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 215                         return tic->t_unit_res * tic->t_cnt;
 216                 else
 217                         return tic->t_unit_res;
 218         }
 219 }
 220
 221 STATIC bool
 222 xlog_grant_head_wake(
 223         struct xlog             *log,
 224         struct xlog_grant_head  *head,
 225         int                     *free_bytes)
 226 {
 227         struct xlog_ticket      *tic;
 228         int                     need_bytes;
 229
 230         list_for_each_entry(tic, &head->waiters, t_queue) {
 231                 need_bytes = xlog_ticket_reservation(log, head, tic);
 232                 if (*free_bytes < need_bytes)
 233                         return false;
 234
 235                 *free_bytes -= need_bytes;
 236                 trace_xfs_log_grant_wake_up(log, tic);
 237                 wake_up_process(tic->t_task);
 238         }
 239
 240         return true;
 241 }
 242
 243 STATIC int
 244 xlog_grant_head_wait(
 245         struct xlog             *log,
 246         struct xlog_grant_head  *head,
 247         struct xlog_ticket      *tic,
 248         int                     need_bytes) __releases(&head->lock)
 249                                             __acquires(&head->lock)
 250 {
 251         list_add_tail(&tic->t_queue, &head->waiters);
 252
 253         do {
 254                 if (XLOG_FORCED_SHUTDOWN(log))
 255                         goto shutdown;
 256                 xlog_grant_push_ail(log, need_bytes);
 257
 258                 __set_current_state(TASK_UNINTERRUPTIBLE);
 259                 spin_unlock(&head->lock);
 260
 261                 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
 262
 263                 trace_xfs_log_grant_sleep(log, tic);
 264                 schedule();
 265                 trace_xfs_log_grant_wake(log, tic);
 266
 267                 spin_lock(&head->lock);
 268                 if (XLOG_FORCED_SHUTDOWN(log))
 269                         goto shutdown;
 270         } while (xlog_space_left(log, &head->grant) < need_bytes);
 271
 272         list_del_init(&tic->t_queue);
 273         return 0;
 274 shutdown:
 275         list_del_init(&tic->t_queue);
 276         return -EIO;
 277 }
 278
 279 /*
 280  * Atomically get the log space required for a log ticket.
 281  *
 282  * Once a ticket gets put onto head->waiters, it will only return after the
 283  * needed reservation is satisfied.
 284  *
 285  * This function is structured so that it has a lock free fast path. This is
 286  * necessary because every new transaction reservation will come through this
 287  * path. Hence any lock will be globally hot if we take it unconditionally on
 288  * every pass.
 289  *
 290  * As tickets are only ever moved on and off head->waiters under head->lock, we
 291  * only need to take that lock if we are going to add the ticket to the queue
 292  * and sleep. We can avoid taking the lock if the ticket was never added to
 293  * head->waiters because the t_queue list head will be empty and we hold the
 294  * only reference to it so it can safely be checked unlocked.
 295  */
 296 STATIC int
 297 xlog_grant_head_check(
 298         struct xlog             *log,
 299         struct xlog_grant_head  *head,
 300         struct xlog_ticket      *tic,
 301         int                     *need_bytes)
 302 {
 303         int                     free_bytes;
 304         int                     error = 0;
 305
 306         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 307
 308         /*
 309          * If there are other waiters on the queue then give them a chance at
 310          * logspace before us.  Wake up the first waiters, if we do not wake
 311          * up all the waiters then go to sleep waiting for more free space,
 312          * otherwise try to get some space for this transaction.
 313          */
 314         *need_bytes = xlog_ticket_reservation(log, head, tic);
 315         free_bytes = xlog_space_left(log, &head->grant);
 316         if (!list_empty_careful(&head->waiters)) {
 317                 spin_lock(&head->lock);
 318                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 319                     free_bytes < *need_bytes) {
 320                         error = xlog_grant_head_wait(log, head, tic,
 321                                                      *need_bytes);
 322                 }
 323                 spin_unlock(&head->lock);
 324         } else if (free_bytes < *need_bytes) {
 325                 spin_lock(&head->lock);
 326                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 327                 spin_unlock(&head->lock);
 328         }
 329
 330         return error;
 331 }
 332
 333 static void
 334 xlog_tic_reset_res(xlog_ticket_t *tic)
 335 {
 336         tic->t_res_num = 0;
 337         tic->t_res_arr_sum = 0;
 338         tic->t_res_num_ophdrs = 0;
 339 }
 340
 341 static void
 342 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 343 {
 344         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 345                 /* add to overflow and start again */
 346                 tic->t_res_o_flow += tic->t_res_arr_sum;
 347                 tic->t_res_num = 0;
 348                 tic->t_res_arr_sum = 0;
 349         }
 350
 351         tic->t_res_arr[tic->t_res_num].r_len = len;
 352         tic->t_res_arr[tic->t_res_num].r_type = type;
 353         tic->t_res_arr_sum += len;
 354         tic->t_res_num++;
 355 }
 356
 357 /*
 358  * Replenish the byte reservation required by moving the grant write head.
 359  */
 360 int
 361 xfs_log_regrant(
 362         struct xfs_mount        *mp,
 363         struct xlog_ticket      *tic)
 364 {
 365         struct xlog             *log = mp->m_log;
 366         int                     need_bytes;
 367         int                     error = 0;
 368
 369         if (XLOG_FORCED_SHUTDOWN(log))
 370                 return -EIO;
 371
 372         XFS_STATS_INC(mp, xs_try_logspace);
 373
 374         /*
 375          * This is a new transaction on the ticket, so we need to change the
 376          * transaction ID so that the next transaction has a different TID in
 377          * the log. Just add one to the existing tid so that we can see chains
 378          * of rolling transactions in the log easily.
 379          */
 380         tic->t_tid++;
 381
 382         xlog_grant_push_ail(log, tic->t_unit_res);
 383
 384         tic->t_curr_res = tic->t_unit_res;
 385         xlog_tic_reset_res(tic);
 386
 387         if (tic->t_cnt > 0)
 388                 return 0;
 389
 390         trace_xfs_log_regrant(log, tic);
 391
 392         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 393                                       &need_bytes);
 394         if (error)
 395                 goto out_error;
 396
 397         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 398         trace_xfs_log_regrant_exit(log, tic);
 399         xlog_verify_grant_tail(log);
 400         return 0;
 401
 402 out_error:
 403         /*
 404          * If we are failing, make sure the ticket doesn't have any current
 405          * reservations.  We don't want to add this back when the ticket/
 406          * transaction gets cancelled.
 407          */
 408         tic->t_curr_res = 0;
 409         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 410         return error;
 411 }
 412
 413 /*
 414  * Reserve log space and return a ticket corresponding to the reservation.
 415  *
 416  * Each reservation is going to reserve extra space for a log record header.
 417  * When writes happen to the on-disk log, we don't subtract the length of the
 418  * log record header from any reservation.  By wasting space in each
 419  * reservation, we prevent over allocation problems.
 420  */
 421 int
 422 xfs_log_reserve(
 423         struct xfs_mount        *mp,
 424         int                     unit_bytes,
 425         int                     cnt,
 426         struct xlog_ticket      **ticp,
 427         uint8_t                 client,
 428         bool                    permanent)
 429 {
 430         struct xlog             *log = mp->m_log;
 431         struct xlog_ticket      *tic;
 432         int                     need_bytes;
 433         int                     error = 0;
 434
 435         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 436
 437         if (XLOG_FORCED_SHUTDOWN(log))
 438                 return -EIO;
 439
 440         XFS_STATS_INC(mp, xs_try_logspace);
 441
 442         ASSERT(*ticp == NULL);
 443         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 444                                 KM_SLEEP | KM_MAYFAIL);
 445         if (!tic)
 446                 return -ENOMEM;
 447
 448         *ticp = tic;
 449
 450         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 451                                             : tic->t_unit_res);
 452
 453         trace_xfs_log_reserve(log, tic);
 454
 455         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 456                                       &need_bytes);
 457         if (error)
 458                 goto out_error;
 459
 460         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 461         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 462         trace_xfs_log_reserve_exit(log, tic);
 463         xlog_verify_grant_tail(log);
 464         return 0;
 465
 466 out_error:
 467         /*
 468          * If we are failing, make sure the ticket doesn't have any current
 469          * reservations.  We don't want to add this back when the ticket/
 470          * transaction gets cancelled.
 471          */
 472         tic->t_curr_res = 0;
 473         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 474         return error;
 475 }
 476
 477
 478 /*
 479  * NOTES:
 480  *
 481  *      1. currblock field gets updated at startup and after in-core logs
 482  *              marked as with WANT_SYNC.
 483  */
 484
 485 /*
 486  * This routine is called when a user of a log manager ticket is done with
 487  * the reservation.  If the ticket was ever used, then a commit record for
 488  * the associated transaction is written out as a log operation header with
 489  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 490  * a given ticket.  If the ticket was one with a permanent reservation, then
 491  * a few operations are done differently.  Permanent reservation tickets by
 492  * default don't release the reservation.  They just commit the current
 493  * transaction with the belief that the reservation is still needed.  A flag
 494  * must be passed in before permanent reservations are actually released.
 495  * When these type of tickets are not released, they need to be set into
 496  * the inited state again.  By doing this, a start record will be written
 497  * out when the next write occurs.
 498  */
 499 xfs_lsn_t
 500 xfs_log_done(
 501         struct xfs_mount        *mp,
 502         struct xlog_ticket      *ticket,
 503         struct xlog_in_core     **iclog,
 504         bool                    regrant)
 505 {
 506         struct xlog             *log = mp->m_log;
 507         xfs_lsn_t               lsn = 0;
 508
 509         if (XLOG_FORCED_SHUTDOWN(log) ||
 510             /*
 511              * If nothing was ever written, don't write out commit record.
 512              * If we get an error, just continue and give back the log ticket.
 513              */
 514             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 515              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 516                 lsn = (xfs_lsn_t) -1;
 517                 regrant = false;
 518         }
 519
 520
 521         if (!regrant) {
 522                 trace_xfs_log_done_nonperm(log, ticket);
 523
 524                 /*
 525                  * Release ticket if not permanent reservation or a specific
 526                  * request has been made to release a permanent reservation.
 527                  */
 528                 xlog_ungrant_log_space(log, ticket);
 529         } else {
 530                 trace_xfs_log_done_perm(log, ticket);
 531
 532                 xlog_regrant_reserve_log_space(log, ticket);
 533                 /* If this ticket was a permanent reservation and we aren't
 534                  * trying to release it, reset the inited flags; so next time
 535                  * we write, a start record will be written out.
 536                  */
 537                 ticket->t_flags |= XLOG_TIC_INITED;
 538         }
 539
 540         xfs_log_ticket_put(ticket);
 541         return lsn;
 542 }
 543
 544 /*
 545  * Attaches a new iclog I/O completion callback routine during
 546  * transaction commit.  If the log is in error state, a non-zero
 547  * return code is handed back and the caller is responsible for
 548  * executing the callback at an appropriate time.
 549  */
 550 int
 551 xfs_log_notify(
 552         struct xlog_in_core     *iclog,
 553         xfs_log_callback_t      *cb)
 554 {
 555         int     abortflg;
 556
 557         spin_lock(&iclog->ic_callback_lock);
 558         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 559         if (!abortflg) {
 560                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 561                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 562                 cb->cb_next = NULL;
 563                 *(iclog->ic_callback_tail) = cb;
 564                 iclog->ic_callback_tail = &(cb->cb_next);
 565         }
 566         spin_unlock(&iclog->ic_callback_lock);
 567         return abortflg;
 568 }
 569
 570 int
 571 xfs_log_release_iclog(
 572         struct xfs_mount        *mp,
 573         struct xlog_in_core     *iclog)
 574 {
 575         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 576                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 577                 return -EIO;
 578         }
 579
 580         return 0;
 581 }
 582
 583 /*
 584  * Mount a log filesystem
 585  *
 586  * mp           - ubiquitous xfs mount point structure
 587  * log_target   - buftarg of on-disk log device
 588  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 589  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 590  *
 591  * Return error or zero.
 592  */
 593 int
 594 xfs_log_mount(
 595         xfs_mount_t     *mp,
 596         xfs_buftarg_t   *log_target,
 597         xfs_daddr_t     blk_offset,
 598         int             num_bblks)
 599 {
 600         bool            fatal = xfs_sb_version_hascrc(&mp->m_sb);
 601         int             error = 0;
 602         int             min_logfsbs;
 603
 604         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 605                 xfs_notice(mp, "Mounting V%d Filesystem",
 606                            XFS_SB_VERSION_NUM(&mp->m_sb));
 607         } else {
 608                 xfs_notice(mp,
 609 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
 610                            XFS_SB_VERSION_NUM(&mp->m_sb));
 611                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 612         }
 613
 614         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 615         if (IS_ERR(mp->m_log)) {
 616                 error = PTR_ERR(mp->m_log);
 617                 goto out;
 618         }
 619
 620         /*
 621          * Validate the given log space and drop a critical message via syslog
 622          * if the log size is too small that would lead to some unexpected
 623          * situations in transaction log space reservation stage.
 624          *
 625          * Note: we can't just reject the mount if the validation fails.  This
 626          * would mean that people would have to downgrade their kernel just to
 627          * remedy the situation as there is no way to grow the log (short of
 628          * black magic surgery with xfs_db).
 629          *
 630          * We can, however, reject mounts for CRC format filesystems, as the
 631          * mkfs binary being used to make the filesystem should never create a
 632          * filesystem with a log that is too small.
 633          */
 634         min_logfsbs = xfs_log_calc_minimum_size(mp);
 635
 636         if (mp->m_sb.sb_logblocks < min_logfsbs) {
 637                 xfs_warn(mp,
 638                 "Log size %d blocks too small, minimum size is %d blocks",
 639                          mp->m_sb.sb_logblocks, min_logfsbs);
 640                 error = -EINVAL;
 641         } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
 642                 xfs_warn(mp,
 643                 "Log size %d blocks too large, maximum size is %lld blocks",
 644                          mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
 645                 error = -EINVAL;
 646         } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
 647                 xfs_warn(mp,
 648                 "log size %lld bytes too large, maximum size is %lld bytes",
 649                          XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
 650                          XFS_MAX_LOG_BYTES);
 651                 error = -EINVAL;
 652         } else if (mp->m_sb.sb_logsunit > 1 &&
 653                    mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
 654                 xfs_warn(mp,
 655                 "log stripe unit %u bytes must be a multiple of block size",
 656                          mp->m_sb.sb_logsunit);
 657                 error = -EINVAL;
 658                 fatal = true;
 659         }
 660         if (error) {
 661                 /*
 662                  * Log check errors are always fatal on v5; or whenever bad
 663                  * metadata leads to a crash.
 664                  */
 665                 if (fatal) {
 666                         xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 667                         ASSERT(0);
 668                         goto out_free_log;
 669                 }
 670                 xfs_crit(mp, "Log size out of supported range.");
 671                 xfs_crit(mp,
 672 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
 673         }
 674
 675         /*
 676          * Initialize the AIL now we have a log.
 677          */
 678         error = xfs_trans_ail_init(mp);
 679         if (error) {
 680                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 681                 goto out_free_log;
 682         }
 683         mp->m_log->l_ailp = mp->m_ail;
 684
 685         /*
 686          * skip log recovery on a norecovery mount.  pretend it all
 687          * just worked.
 688          */
 689         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 690                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 691
 692                 if (readonly)
 693                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 694
 695                 error = xlog_recover(mp->m_log);
 696
 697                 if (readonly)
 698                         mp->m_flags |= XFS_MOUNT_RDONLY;
 699                 if (error) {
 700                         xfs_warn(mp, "log mount/recovery failed: error %d",
 701                                 error);
 702                         xlog_recover_cancel(mp->m_log);
 703                         goto out_destroy_ail;
 704                 }
 705         }
 706
 707         error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 708                                "log");
 709         if (error)
 710                 goto out_destroy_ail;
 711
 712         /* Normal transactions can now occur */
 713         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 714
 715         /*
 716          * Now the log has been fully initialised and we know were our
 717          * space grant counters are, we can initialise the permanent ticket
 718          * needed for delayed logging to work.
 719          */
 720         xlog_cil_init_post_recovery(mp->m_log);
 721
 722         return 0;
 723
 724 out_destroy_ail:
 725         xfs_trans_ail_destroy(mp);
 726 out_free_log:
 727         xlog_dealloc_log(mp->m_log);
 728 out:
 729         return error;
 730 }
 731
 732 /*
 733  * Finish the recovery of the file system.  This is separate from the
 734  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 735  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 736  * here.
 737  *
 738  * If we finish recovery successfully, start the background log work. If we are
 739  * not doing recovery, then we have a RO filesystem and we don't need to start
 740  * it.
 741  */
 742 int
 743 xfs_log_mount_finish(
 744         struct xfs_mount        *mp)
 745 {
 746         int     error = 0;
 747         bool    readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 748         bool    recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
 749
 750         if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
 751                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 752                 return 0;
 753         } else if (readonly) {
 754                 /* Allow unlinked processing to proceed */
 755                 mp->m_flags &= ~XFS_MOUNT_RDONLY;
 756         }
 757
 758         /*
 759          * During the second phase of log recovery, we need iget and
 760          * iput to behave like they do for an active filesystem.
 761          * xfs_fs_drop_inode needs to be able to prevent the deletion
 762          * of inodes before we're done replaying log items on those
 763          * inodes.  Turn it off immediately after recovery finishes
 764          * so that we don't leak the quota inodes if subsequent mount
 765          * activities fail.
 766          *
 767          * We let all inodes involved in redo item processing end up on
 768          * the LRU instead of being evicted immediately so that if we do
 769          * something to an unlinked inode, the irele won't cause
 770          * premature truncation and freeing of the inode, which results
 771          * in log recovery failure.  We have to evict the unreferenced
 772          * lru inodes after clearing SB_ACTIVE because we don't
 773          * otherwise clean up the lru if there's a subsequent failure in
 774          * xfs_mountfs, which leads to us leaking the inodes if nothing
 775          * else (e.g. quotacheck) references the inodes before the
 776          * mount failure occurs.
 777          */
 778         mp->m_super->s_flags |= SB_ACTIVE;
 779         error = xlog_recover_finish(mp->m_log);
 780         if (!error)
 781                 xfs_log_work_queue(mp);
 782         mp->m_super->s_flags &= ~SB_ACTIVE;
 783         evict_inodes(mp->m_super);
 784
 785         /*
 786          * Drain the buffer LRU after log recovery. This is required for v4
 787          * filesystems to avoid leaving around buffers with NULL verifier ops,
 788          * but we do it unconditionally to make sure we're always in a clean
 789          * cache state after mount.
 790          *
 791          * Don't push in the error case because the AIL may have pending intents
 792          * that aren't removed until recovery is cancelled.
 793          */
 794         if (!error && recovered) {
 795                 xfs_log_force(mp, XFS_LOG_SYNC);
 796                 xfs_ail_push_all_sync(mp->m_ail);
 797         }
 798         xfs_wait_buftarg(mp->m_ddev_targp);
 799
 800         if (readonly)
 801                 mp->m_flags |= XFS_MOUNT_RDONLY;
 802
 803         return error;
 804 }
 805
 806 /*
 807  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
 808  * the log.
 809  */
 810 int
 811 xfs_log_mount_cancel(
 812         struct xfs_mount        *mp)
 813 {
 814         int                     error;
 815
 816         error = xlog_recover_cancel(mp->m_log);
 817         xfs_log_unmount(mp);
 818
 819         return error;
 820 }
 821
 822 /*
 823  * Final log writes as part of unmount.
 824  *
 825  * Mark the filesystem clean as unmount happens.  Note that during relocation
 826  * this routine needs to be executed as part of source-bag while the
 827  * deallocation must not be done until source-end.
 828  */
 829
 830 /* Actually write the unmount record to disk. */
 831 static void
 832 xfs_log_write_unmount_record(
 833         struct xfs_mount        *mp)
 834 {
 835         /* the data section must be 32 bit size aligned */
 836         struct xfs_unmount_log_format magic = {
 837                 .magic = XLOG_UNMOUNT_TYPE,
 838         };
 839         struct xfs_log_iovec reg = {
 840                 .i_addr = &magic,
 841                 .i_len = sizeof(magic),
 842                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 843         };
 844         struct xfs_log_vec vec = {
 845                 .lv_niovecs = 1,
 846                 .lv_iovecp = &reg,
 847         };
 848         struct xlog             *log = mp->m_log;
 849         struct xlog_in_core     *iclog;
 850         struct xlog_ticket      *tic = NULL;
 851         xfs_lsn_t               lsn;
 852         uint                    flags = XLOG_UNMOUNT_TRANS;
 853         int                     error;
 854
 855         error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
 856         if (error)
 857                 goto out_err;
 858
 859         /*
 860          * If we think the summary counters are bad, clear the unmount header
 861          * flag in the unmount record so that the summary counters will be
 862          * recalculated during log recovery at next mount.  Refer to
 863          * xlog_check_unmount_rec for more details.
 864          */
 865         if (XFS_TEST_ERROR(xfs_fs_has_sickness(mp, XFS_SICK_FS_COUNTERS), mp,
 866                         XFS_ERRTAG_FORCE_SUMMARY_RECALC)) {
 867                 xfs_alert(mp, "%s: will fix summary counters at next mount",
 868                                 __func__);
 869                 flags &= ~XLOG_UNMOUNT_TRANS;
 870         }
 871
 872         /* remove inited flag, and account for space used */
 873         tic->t_flags = 0;
 874         tic->t_curr_res -= sizeof(magic);
 875         error = xlog_write(log, &vec, tic, &lsn, NULL, flags);
 876         /*
 877          * At this point, we're umounting anyway, so there's no point in
 878          * transitioning log state to IOERROR. Just continue...
 879          */
 880 out_err:
 881         if (error)
 882                 xfs_alert(mp, "%s: unmount record failed", __func__);
 883
 884         spin_lock(&log->l_icloglock);
 885         iclog = log->l_iclog;
 886         atomic_inc(&iclog->ic_refcnt);
 887         xlog_state_want_sync(log, iclog);
 888         spin_unlock(&log->l_icloglock);
 889         error = xlog_state_release_iclog(log, iclog);
 890
 891         spin_lock(&log->l_icloglock);
 892         switch (iclog->ic_state) {
 893         default:
 894                 if (!XLOG_FORCED_SHUTDOWN(log)) {
 895                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
 896                         break;
 897                 }
 898                 /* fall through */
 899         case XLOG_STATE_ACTIVE:
 900         case XLOG_STATE_DIRTY:
 901                 spin_unlock(&log->l_icloglock);
 902                 break;
 903         }
 904
 905         if (tic) {
 906                 trace_xfs_log_umount_write(log, tic);
 907                 xlog_ungrant_log_space(log, tic);
 908                 xfs_log_ticket_put(tic);
 909         }
 910 }
 911
 912 /*
 913  * Unmount record used to have a string "Unmount filesystem--" in the
 914  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 915  * We just write the magic number now since that particular field isn't
 916  * currently architecture converted and "Unmount" is a bit foo.
 917  * As far as I know, there weren't any dependencies on the old behaviour.
 918  */
 919
 920 static int
 921 xfs_log_unmount_write(xfs_mount_t *mp)
 922 {
 923         struct xlog      *log = mp->m_log;
 924         xlog_in_core_t   *iclog;
 925 #ifdef DEBUG
 926         xlog_in_core_t   *first_iclog;
 927 #endif
 928         int              error;
 929
 930         /*
 931          * Don't write out unmount record on norecovery mounts or ro devices.
 932          * Or, if we are doing a forced umount (typically because of IO errors).
 933          */
 934         if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
 935             xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
 936                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 937                 return 0;
 938         }
 939
 940         error = xfs_log_force(mp, XFS_LOG_SYNC);
 941         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 942
 943 #ifdef DEBUG
 944         first_iclog = iclog = log->l_iclog;
 945         do {
 946                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 947                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 948                         ASSERT(iclog->ic_offset == 0);
 949                 }
 950                 iclog = iclog->ic_next;
 951         } while (iclog != first_iclog);
 952 #endif
 953         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 954                 xfs_log_write_unmount_record(mp);
 955         } else {
 956                 /*
 957                  * We're already in forced_shutdown mode, couldn't
 958                  * even attempt to write out the unmount transaction.
 959                  *
 960                  * Go through the motions of sync'ing and releasing
 961                  * the iclog, even though no I/O will actually happen,
 962                  * we need to wait for other log I/Os that may already
 963                  * be in progress.  Do this as a separate section of
 964                  * code so we'll know if we ever get stuck here that
 965                  * we're in this odd situation of trying to unmount
 966                  * a file system that went into forced_shutdown as
 967                  * the result of an unmount..
 968                  */
 969                 spin_lock(&log->l_icloglock);
 970                 iclog = log->l_iclog;
 971                 atomic_inc(&iclog->ic_refcnt);
 972
 973                 xlog_state_want_sync(log, iclog);
 974                 spin_unlock(&log->l_icloglock);
 975                 error =  xlog_state_release_iclog(log, iclog);
 976
 977                 spin_lock(&log->l_icloglock);
 978
 979                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 980                         || iclog->ic_state == XLOG_STATE_DIRTY
 981                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 982
 983                                 xlog_wait(&iclog->ic_force_wait,
 984                                                         &log->l_icloglock);
 985                 } else {
 986                         spin_unlock(&log->l_icloglock);
 987                 }
 988         }
 989
 990         return error;
 991 }       /* xfs_log_unmount_write */
 992
 993 /*
 994  * Empty the log for unmount/freeze.
 995  *
 996  * To do this, we first need to shut down the background log work so it is not
 997  * trying to cover the log as we clean up. We then need to unpin all objects in
 998  * the log so we can then flush them out. Once they have completed their IO and
 999  * run the callbacks removing themselves from the AIL, we can write the unmount
1000  * record.
1001  */
1002 void
1003 xfs_log_quiesce(
1004         struct xfs_mount        *mp)
1005 {
1006         cancel_delayed_work_sync(&mp->m_log->l_work);
1007         xfs_log_force(mp, XFS_LOG_SYNC);
1008
1009         /*
1010          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1011          * will push it, xfs_wait_buftarg() will not wait for it. Further,
1012          * xfs_buf_iowait() cannot be used because it was pushed with the
1013          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1014          * the IO to complete.
1015          */
1016         xfs_ail_push_all_sync(mp->m_ail);
1017         xfs_wait_buftarg(mp->m_ddev_targp);
1018         xfs_buf_lock(mp->m_sb_bp);
1019         xfs_buf_unlock(mp->m_sb_bp);
1020
1021         xfs_log_unmount_write(mp);
1022 }
1023
1024 /*
1025  * Shut down and release the AIL and Log.
1026  *
1027  * During unmount, we need to ensure we flush all the dirty metadata objects
1028  * from the AIL so that the log is empty before we write the unmount record to
1029  * the log. Once this is done, we can tear down the AIL and the log.
1030  */
1031 void
1032 xfs_log_unmount(
1033         struct xfs_mount        *mp)
1034 {
1035         xfs_log_quiesce(mp);
1036
1037         xfs_trans_ail_destroy(mp);
1038
1039         xfs_sysfs_del(&mp->m_log->l_kobj);
1040
1041         xlog_dealloc_log(mp->m_log);
1042 }
1043
1044 void
1045 xfs_log_item_init(
1046         struct xfs_mount        *mp,
1047         struct xfs_log_item     *item,
1048         int                     type,
1049         const struct xfs_item_ops *ops)
1050 {
1051         item->li_mountp = mp;
1052         item->li_ailp = mp->m_ail;
1053         item->li_type = type;
1054         item->li_ops = ops;
1055         item->li_lv = NULL;
1056
1057         INIT_LIST_HEAD(&item->li_ail);
1058         INIT_LIST_HEAD(&item->li_cil);
1059         INIT_LIST_HEAD(&item->li_bio_list);
1060         INIT_LIST_HEAD(&item->li_trans);
1061 }
1062
1063 /*
1064  * Wake up processes waiting for log space after we have moved the log tail.
1065  */
1066 void
1067 xfs_log_space_wake(
1068         struct xfs_mount        *mp)
1069 {
1070         struct xlog             *log = mp->m_log;
1071         int                     free_bytes;
1072
1073         if (XLOG_FORCED_SHUTDOWN(log))
1074                 return;
1075
1076         if (!list_empty_careful(&log->l_write_head.waiters)) {
1077                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1078
1079                 spin_lock(&log->l_write_head.lock);
1080                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1081                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1082                 spin_unlock(&log->l_write_head.lock);
1083         }
1084
1085         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1086                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1087
1088                 spin_lock(&log->l_reserve_head.lock);
1089                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1090                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1091                 spin_unlock(&log->l_reserve_head.lock);
1092         }
1093 }
1094
1095 /*
1096  * Determine if we have a transaction that has gone to disk that needs to be
1097  * covered. To begin the transition to the idle state firstly the log needs to
1098  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1099  * we start attempting to cover the log.
1100  *
1101  * Only if we are then in a state where covering is needed, the caller is
1102  * informed that dummy transactions are required to move the log into the idle
1103  * state.
1104  *
1105  * If there are any items in the AIl or CIL, then we do not want to attempt to
1106  * cover the log as we may be in a situation where there isn't log space
1107  * available to run a dummy transaction and this can lead to deadlocks when the
1108  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1109  * there's no point in running a dummy transaction at this point because we
1110  * can't start trying to idle the log until both the CIL and AIL are empty.
1111  */
1112 static int
1113 xfs_log_need_covered(xfs_mount_t *mp)
1114 {
1115         struct xlog     *log = mp->m_log;
1116         int             needed = 0;
1117
1118         if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1119                 return 0;
1120
1121         if (!xlog_cil_empty(log))
1122                 return 0;
1123
1124         spin_lock(&log->l_icloglock);
1125         switch (log->l_covered_state) {
1126         case XLOG_STATE_COVER_DONE:
1127         case XLOG_STATE_COVER_DONE2:
1128         case XLOG_STATE_COVER_IDLE:
1129                 break;
1130         case XLOG_STATE_COVER_NEED:
1131         case XLOG_STATE_COVER_NEED2:
1132                 if (xfs_ail_min_lsn(log->l_ailp))
1133                         break;
1134                 if (!xlog_iclogs_empty(log))
1135                         break;
1136
1137                 needed = 1;
1138                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1139                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1140                 else
1141                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1142                 break;
1143         default:
1144                 needed = 1;
1145                 break;
1146         }
1147         spin_unlock(&log->l_icloglock);
1148         return needed;
1149 }
1150
1151 /*
1152  * We may be holding the log iclog lock upon entering this routine.
1153  */
1154 xfs_lsn_t
1155 xlog_assign_tail_lsn_locked(
1156         struct xfs_mount        *mp)
1157 {
1158         struct xlog             *log = mp->m_log;
1159         struct xfs_log_item     *lip;
1160         xfs_lsn_t               tail_lsn;
1161
1162         assert_spin_locked(&mp->m_ail->ail_lock);
1163
1164         /*
1165          * To make sure we always have a valid LSN for the log tail we keep
1166          * track of the last LSN which was committed in log->l_last_sync_lsn,
1167          * and use that when the AIL was empty.
1168          */
1169         lip = xfs_ail_min(mp->m_ail);
1170         if (lip)
1171                 tail_lsn = lip->li_lsn;
1172         else
1173                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1174         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1175         atomic64_set(&log->l_tail_lsn, tail_lsn);
1176         return tail_lsn;
1177 }
1178
1179 xfs_lsn_t
1180 xlog_assign_tail_lsn(
1181         struct xfs_mount        *mp)
1182 {
1183         xfs_lsn_t               tail_lsn;
1184
1185         spin_lock(&mp->m_ail->ail_lock);
1186         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1187         spin_unlock(&mp->m_ail->ail_lock);
1188
1189         return tail_lsn;
1190 }
1191
1192 /*
1193  * Return the space in the log between the tail and the head.  The head
1194  * is passed in the cycle/bytes formal parms.  In the special case where
1195  * the reserve head has wrapped passed the tail, this calculation is no
1196  * longer valid.  In this case, just return 0 which means there is no space
1197  * in the log.  This works for all places where this function is called
1198  * with the reserve head.  Of course, if the write head were to ever
1199  * wrap the tail, we should blow up.  Rather than catch this case here,
1200  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1201  *
1202  * This code also handles the case where the reservation head is behind
1203  * the tail.  The details of this case are described below, but the end
1204  * result is that we return the size of the log as the amount of space left.
1205  */
1206 STATIC int
1207 xlog_space_left(
1208         struct xlog     *log,
1209         atomic64_t      *head)
1210 {
1211         int             free_bytes;
1212         int             tail_bytes;
1213         int             tail_cycle;
1214         int             head_cycle;
1215         int             head_bytes;
1216
1217         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1218         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1219         tail_bytes = BBTOB(tail_bytes);
1220         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1221                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1222         else if (tail_cycle + 1 < head_cycle)
1223                 return 0;
1224         else if (tail_cycle < head_cycle) {
1225                 ASSERT(tail_cycle == (head_cycle - 1));
1226                 free_bytes = tail_bytes - head_bytes;
1227         } else {
1228                 /*
1229                  * The reservation head is behind the tail.
1230                  * In this case we just want to return the size of the
1231                  * log as the amount of space left.
1232                  */
1233                 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1234                 xfs_alert(log->l_mp,
1235                           "  tail_cycle = %d, tail_bytes = %d",
1236                           tail_cycle, tail_bytes);
1237                 xfs_alert(log->l_mp,
1238                           "  GH   cycle = %d, GH   bytes = %d",
1239                           head_cycle, head_bytes);
1240                 ASSERT(0);
1241                 free_bytes = log->l_logsize;
1242         }
1243         return free_bytes;
1244 }
1245
1246
1247 /*
1248  * Log function which is called when an io completes.
1249  *
1250  * The log manager needs its own routine, in order to control what
1251  * happens with the buffer after the write completes.
1252  */
1253 static void
1254 xlog_iodone(xfs_buf_t *bp)
1255 {
1256         struct xlog_in_core     *iclog = bp->b_log_item;
1257         struct xlog             *l = iclog->ic_log;
1258         int                     aborted = 0;
1259
1260 #ifdef DEBUG
1261         /* treat writes with injected CRC errors as failed */
1262         if (iclog->ic_fail_crc)
1263                 bp->b_error = -EIO;
1264 #endif
1265
1266         /*
1267          * Race to shutdown the filesystem if we see an error.
1268          */
1269         if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
1270                 xfs_buf_ioerror_alert(bp, __func__);
1271                 xfs_buf_stale(bp);
1272                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1273                 /*
1274                  * This flag will be propagated to the trans-committed
1275                  * callback routines to let them know that the log-commit
1276                  * didn't succeed.
1277                  */
1278                 aborted = XFS_LI_ABORTED;
1279         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1280                 aborted = XFS_LI_ABORTED;
1281         }
1282
1283         /* log I/O is always issued ASYNC */
1284         ASSERT(bp->b_flags & XBF_ASYNC);
1285         xlog_state_done_syncing(iclog, aborted);
1286
1287         /*
1288          * drop the buffer lock now that we are done. Nothing references
1289          * the buffer after this, so an unmount waiting on this lock can now
1290          * tear it down safely. As such, it is unsafe to reference the buffer
1291          * (bp) after the unlock as we could race with it being freed.
1292          */
1293         xfs_buf_unlock(bp);
1294 }
1295
1296 /*
1297  * Return size of each in-core log record buffer.
1298  *
1299  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1300  *
1301  * If the filesystem blocksize is too large, we may need to choose a
1302  * larger size since the directory code currently logs entire blocks.
1303  */
1304 STATIC void
1305 xlog_get_iclog_buffer_size(
1306         struct xfs_mount        *mp,
1307         struct xlog             *log)
1308 {
1309         if (mp->m_logbufs <= 0)
1310                 mp->m_logbufs = XLOG_MAX_ICLOGS;
1311         if (mp->m_logbsize <= 0)
1312                 mp->m_logbsize = XLOG_BIG_RECORD_BSIZE;
1313
1314         log->l_iclog_bufs = mp->m_logbufs;
1315         log->l_iclog_size = mp->m_logbsize;
1316
1317         /*
1318          * # headers = size / 32k - one header holds cycles from 32k of data.
1319          */
1320         log->l_iclog_heads =
1321                 DIV_ROUND_UP(mp->m_logbsize, XLOG_HEADER_CYCLE_SIZE);
1322         log->l_iclog_hsize = log->l_iclog_heads << BBSHIFT;
1323 }
1324
1325 void
1326 xfs_log_work_queue(
1327         struct xfs_mount        *mp)
1328 {
1329         queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1330                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1331 }
1332
1333 /*
1334  * Every sync period we need to unpin all items in the AIL and push them to
1335  * disk. If there is nothing dirty, then we might need to cover the log to
1336  * indicate that the filesystem is idle.
1337  */
1338 static void
1339 xfs_log_worker(
1340         struct work_struct      *work)
1341 {
1342         struct xlog             *log = container_of(to_delayed_work(work),
1343                                                 struct xlog, l_work);
1344         struct xfs_mount        *mp = log->l_mp;
1345
1346         /* dgc: errors ignored - not fatal and nowhere to report them */
1347         if (xfs_log_need_covered(mp)) {
1348                 /*
1349                  * Dump a transaction into the log that contains no real change.
1350                  * This is needed to stamp the current tail LSN into the log
1351                  * during the covering operation.
1352                  *
1353                  * We cannot use an inode here for this - that will push dirty
1354                  * state back up into the VFS and then periodic inode flushing
1355                  * will prevent log covering from making progress. Hence we
1356                  * synchronously log the superblock instead to ensure the
1357                  * superblock is immediately unpinned and can be written back.
1358                  */
1359                 xfs_sync_sb(mp, true);
1360         } else
1361                 xfs_log_force(mp, 0);
1362
1363         /* start pushing all the metadata that is currently dirty */
1364         xfs_ail_push_all(mp->m_ail);
1365
1366         /* queue us up again */
1367         xfs_log_work_queue(mp);
1368 }
1369
1370 /*
1371  * This routine initializes some of the log structure for a given mount point.
1372  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1373  * some other stuff may be filled in too.
1374  */
1375 STATIC struct xlog *
1376 xlog_alloc_log(
1377         struct xfs_mount        *mp,
1378         struct xfs_buftarg      *log_target,
1379         xfs_daddr_t             blk_offset,
1380         int                     num_bblks)
1381 {
1382         struct xlog             *log;
1383         xlog_rec_header_t       *head;
1384         xlog_in_core_t          **iclogp;
1385         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1386         xfs_buf_t               *bp;
1387         int                     i;
1388         int                     error = -ENOMEM;
1389         uint                    log2_size = 0;
1390
1391         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1392         if (!log) {
1393                 xfs_warn(mp, "Log allocation failed: No memory!");
1394                 goto out;
1395         }
1396
1397         log->l_mp          = mp;
1398         log->l_targ        = log_target;
1399         log->l_logsize     = BBTOB(num_bblks);
1400         log->l_logBBstart  = blk_offset;
1401         log->l_logBBsize   = num_bblks;
1402         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1403         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1404         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1405
1406         log->l_prev_block  = -1;
1407         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1408         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1409         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1410         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1411
1412         xlog_grant_head_init(&log->l_reserve_head);
1413         xlog_grant_head_init(&log->l_write_head);
1414
1415         error = -EFSCORRUPTED;
1416         if (xfs_sb_version_hassector(&mp->m_sb)) {
1417                 log2_size = mp->m_sb.sb_logsectlog;
1418                 if (log2_size < BBSHIFT) {
1419                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1420                                 log2_size, BBSHIFT);
1421                         goto out_free_log;
1422                 }
1423
1424                 log2_size -= BBSHIFT;
1425                 if (log2_size > mp->m_sectbb_log) {
1426                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1427                                 log2_size, mp->m_sectbb_log);
1428                         goto out_free_log;
1429                 }
1430
1431                 /* for larger sector sizes, must have v2 or external log */
1432                 if (log2_size && log->l_logBBstart > 0 &&
1433                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1434                         xfs_warn(mp,
1435                 "log sector size (0x%x) invalid for configuration.",
1436                                 log2_size);
1437                         goto out_free_log;
1438                 }
1439         }
1440         log->l_sectBBsize = 1 << log2_size;
1441
1442         xlog_get_iclog_buffer_size(mp, log);
1443
1444         /*
1445          * Use a NULL block for the extra log buffer used during splits so that
1446          * it will trigger errors if we ever try to do IO on it without first
1447          * having set it up properly.
1448          */
1449         error = -ENOMEM;
1450         bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1451                            BTOBB(log->l_iclog_size), XBF_NO_IOACCT);
1452         if (!bp)
1453                 goto out_free_log;
1454
1455         /*
1456          * The iclogbuf buffer locks are held over IO but we are not going to do
1457          * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1458          * when appropriately.
1459          */
1460         ASSERT(xfs_buf_islocked(bp));
1461         xfs_buf_unlock(bp);
1462
1463         /* use high priority wq for log I/O completion */
1464         bp->b_ioend_wq = mp->m_log_workqueue;
1465         bp->b_iodone = xlog_iodone;
1466         log->l_xbuf = bp;
1467
1468         spin_lock_init(&log->l_icloglock);
1469         init_waitqueue_head(&log->l_flush_wait);
1470
1471         iclogp = &log->l_iclog;
1472         /*
1473          * The amount of memory to allocate for the iclog structure is
1474          * rather funky due to the way the structure is defined.  It is
1475          * done this way so that we can use different sizes for machines
1476          * with different amounts of memory.  See the definition of
1477          * xlog_in_core_t in xfs_log_priv.h for details.
1478          */
1479         ASSERT(log->l_iclog_size >= 4096);
1480         for (i=0; i < log->l_iclog_bufs; i++) {
1481                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1482                 if (!*iclogp)
1483                         goto out_free_iclog;
1484
1485                 iclog = *iclogp;
1486                 iclog->ic_prev = prev_iclog;
1487                 prev_iclog = iclog;
1488
1489                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1490                                           BTOBB(log->l_iclog_size),
1491                                           XBF_NO_IOACCT);
1492                 if (!bp)
1493                         goto out_free_iclog;
1494
1495                 ASSERT(xfs_buf_islocked(bp));
1496                 xfs_buf_unlock(bp);
1497
1498                 /* use high priority wq for log I/O completion */
1499                 bp->b_ioend_wq = mp->m_log_workqueue;
1500                 bp->b_iodone = xlog_iodone;
1501                 iclog->ic_bp = bp;
1502                 iclog->ic_data = bp->b_addr;
1503 #ifdef DEBUG
1504                 log->l_iclog_bak[i] = &iclog->ic_header;
1505 #endif
1506                 head = &iclog->ic_header;
1507                 memset(head, 0, sizeof(xlog_rec_header_t));
1508                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1509                 head->h_version = cpu_to_be32(
1510                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1511                 head->h_size = cpu_to_be32(log->l_iclog_size);
1512                 /* new fields */
1513                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1514                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1515
1516                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1517                 iclog->ic_state = XLOG_STATE_ACTIVE;
1518                 iclog->ic_log = log;
1519                 atomic_set(&iclog->ic_refcnt, 0);
1520                 spin_lock_init(&iclog->ic_callback_lock);
1521                 iclog->ic_callback_tail = &(iclog->ic_callback);
1522                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1523
1524                 init_waitqueue_head(&iclog->ic_force_wait);
1525                 init_waitqueue_head(&iclog->ic_write_wait);
1526
1527                 iclogp = &iclog->ic_next;
1528         }
1529         *iclogp = log->l_iclog;                 /* complete ring */
1530         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1531
1532         error = xlog_cil_init(log);
1533         if (error)
1534                 goto out_free_iclog;
1535         return log;
1536
1537 out_free_iclog:
1538         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1539                 prev_iclog = iclog->ic_next;
1540                 if (iclog->ic_bp)
1541                         xfs_buf_free(iclog->ic_bp);
1542                 kmem_free(iclog);
1543         }
1544         xfs_buf_free(log->l_xbuf);
1545 out_free_log:
1546         kmem_free(log);
1547 out:
1548         return ERR_PTR(error);
1549 }       /* xlog_alloc_log */
1550
1551
1552 /*
1553  * Write out the commit record of a transaction associated with the given
1554  * ticket.  Return the lsn of the commit record.
1555  */
1556 STATIC int
1557 xlog_commit_record(
1558         struct xlog             *log,
1559         struct xlog_ticket      *ticket,
1560         struct xlog_in_core     **iclog,
1561         xfs_lsn_t               *commitlsnp)
1562 {
1563         struct xfs_mount *mp = log->l_mp;
1564         int     error;
1565         struct xfs_log_iovec reg = {
1566                 .i_addr = NULL,
1567                 .i_len = 0,
1568                 .i_type = XLOG_REG_TYPE_COMMIT,
1569         };
1570         struct xfs_log_vec vec = {
1571                 .lv_niovecs = 1,
1572                 .lv_iovecp = &reg,
1573         };
1574
1575         ASSERT_ALWAYS(iclog);
1576         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1577                                         XLOG_COMMIT_TRANS);
1578         if (error)
1579                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1580         return error;
1581 }
1582
1583 /*
1584  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1585  * log space.  This code pushes on the lsn which would supposedly free up
1586  * the 25% which we want to leave free.  We may need to adopt a policy which
1587  * pushes on an lsn which is further along in the log once we reach the high
1588  * water mark.  In this manner, we would be creating a low water mark.
1589  */
1590 STATIC void
1591 xlog_grant_push_ail(
1592         struct xlog     *log,
1593         int             need_bytes)
1594 {
1595         xfs_lsn_t       threshold_lsn = 0;
1596         xfs_lsn_t       last_sync_lsn;
1597         int             free_blocks;
1598         int             free_bytes;
1599         int             threshold_block;
1600         int             threshold_cycle;
1601         int             free_threshold;
1602
1603         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1604
1605         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1606         free_blocks = BTOBBT(free_bytes);
1607
1608         /*
1609          * Set the threshold for the minimum number of free blocks in the
1610          * log to the maximum of what the caller needs, one quarter of the
1611          * log, and 256 blocks.
1612          */
1613         free_threshold = BTOBB(need_bytes);
1614         free_threshold = max(free_threshold, (log->l_logBBsize >> 2));
1615         free_threshold = max(free_threshold, 256);
1616         if (free_blocks >= free_threshold)
1617                 return;
1618
1619         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1620                                                 &threshold_block);
1621         threshold_block += free_threshold;
1622         if (threshold_block >= log->l_logBBsize) {
1623                 threshold_block -= log->l_logBBsize;
1624                 threshold_cycle += 1;
1625         }
1626         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1627                                         threshold_block);
1628         /*
1629          * Don't pass in an lsn greater than the lsn of the last
1630          * log record known to be on disk. Use a snapshot of the last sync lsn
1631          * so that it doesn't change between the compare and the set.
1632          */
1633         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1634         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1635                 threshold_lsn = last_sync_lsn;
1636
1637         /*
1638          * Get the transaction layer to kick the dirty buffers out to
1639          * disk asynchronously. No point in trying to do this if
1640          * the filesystem is shutting down.
1641          */
1642         if (!XLOG_FORCED_SHUTDOWN(log))
1643                 xfs_ail_push(log->l_ailp, threshold_lsn);
1644 }
1645
1646 /*
1647  * Stamp cycle number in every block
1648  */
1649 STATIC void
1650 xlog_pack_data(
1651         struct xlog             *log,
1652         struct xlog_in_core     *iclog,
1653         int                     roundoff)
1654 {
1655         int                     i, j, k;
1656         int                     size = iclog->ic_offset + roundoff;
1657         __be32                  cycle_lsn;
1658         char                    *dp;
1659
1660         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1661
1662         dp = iclog->ic_datap;
1663         for (i = 0; i < BTOBB(size); i++) {
1664                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1665                         break;
1666                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1667                 *(__be32 *)dp = cycle_lsn;
1668                 dp += BBSIZE;
1669         }
1670
1671         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1672                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1673
1674                 for ( ; i < BTOBB(size); i++) {
1675                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1676                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1677                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1678                         *(__be32 *)dp = cycle_lsn;
1679                         dp += BBSIZE;
1680                 }
1681
1682                 for (i = 1; i < log->l_iclog_heads; i++)
1683                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1684         }
1685 }
1686
1687 /*
1688  * Calculate the checksum for a log buffer.
1689  *
1690  * This is a little more complicated than it should be because the various
1691  * headers and the actual data are non-contiguous.
1692  */
1693 __le32
1694 xlog_cksum(
1695         struct xlog             *log,
1696         struct xlog_rec_header  *rhead,
1697         char                    *dp,
1698         int                     size)
1699 {
1700         uint32_t                crc;
1701
1702         /* first generate the crc for the record header ... */
1703         crc = xfs_start_cksum_update((char *)rhead,
1704                               sizeof(struct xlog_rec_header),
1705                               offsetof(struct xlog_rec_header, h_crc));
1706
1707         /* ... then for additional cycle data for v2 logs ... */
1708         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1709                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1710                 int             i;
1711                 int             xheads;
1712
1713                 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1714                 if (size % XLOG_HEADER_CYCLE_SIZE)
1715                         xheads++;
1716
1717                 for (i = 1; i < xheads; i++) {
1718                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1719                                      sizeof(struct xlog_rec_ext_header));
1720                 }
1721         }
1722
1723         /* ... and finally for the payload */
1724         crc = crc32c(crc, dp, size);
1725
1726         return xfs_end_cksum(crc);
1727 }
1728
1729 /*
1730  * The bdstrat callback function for log bufs. This gives us a central
1731  * place to trap bufs in case we get hit by a log I/O error and need to
1732  * shutdown. Actually, in practice, even when we didn't get a log error,
1733  * we transition the iclogs to IOERROR state *after* flushing all existing
1734  * iclogs to disk. This is because we don't want anymore new transactions to be
1735  * started or completed afterwards.
1736  *
1737  * We lock the iclogbufs here so that we can serialise against IO completion
1738  * during unmount. We might be processing a shutdown triggered during unmount,
1739  * and that can occur asynchronously to the unmount thread, and hence we need to
1740  * ensure that completes before tearing down the iclogbufs. Hence we need to
1741  * hold the buffer lock across the log IO to acheive that.
1742  */
1743 STATIC int
1744 xlog_bdstrat(
1745         struct xfs_buf          *bp)
1746 {
1747         struct xlog_in_core     *iclog = bp->b_log_item;
1748
1749         xfs_buf_lock(bp);
1750         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1751                 xfs_buf_ioerror(bp, -EIO);
1752                 xfs_buf_stale(bp);
1753                 xfs_buf_ioend(bp);
1754                 /*
1755                  * It would seem logical to return EIO here, but we rely on
1756                  * the log state machine to propagate I/O errors instead of
1757                  * doing it here. Similarly, IO completion will unlock the
1758                  * buffer, so we don't do it here.
1759                  */
1760                 return 0;
1761         }
1762
1763         xfs_buf_submit(bp);
1764         return 0;
1765 }
1766
1767 /*
1768  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1769  * fashion.  Previously, we should have moved the current iclog
1770  * ptr in the log to point to the next available iclog.  This allows further
1771  * write to continue while this code syncs out an iclog ready to go.
1772  * Before an in-core log can be written out, the data section must be scanned
1773  * to save away the 1st word of each BBSIZE block into the header.  We replace
1774  * it with the current cycle count.  Each BBSIZE block is tagged with the
1775  * cycle count because there in an implicit assumption that drives will
1776  * guarantee that entire 512 byte blocks get written at once.  In other words,
1777  * we can't have part of a 512 byte block written and part not written.  By
1778  * tagging each block, we will know which blocks are valid when recovering
1779  * after an unclean shutdown.
1780  *
1781  * This routine is single threaded on the iclog.  No other thread can be in
1782  * this routine with the same iclog.  Changing contents of iclog can there-
1783  * fore be done without grabbing the state machine lock.  Updating the global
1784  * log will require grabbing the lock though.
1785  *
1786  * The entire log manager uses a logical block numbering scheme.  Only
1787  * log_sync (and then only bwrite()) know about the fact that the log may
1788  * not start with block zero on a given device.  The log block start offset
1789  * is added immediately before calling bwrite().
1790  */
1791
1792 STATIC int
1793 xlog_sync(
1794         struct xlog             *log,
1795         struct xlog_in_core     *iclog)
1796 {
1797         xfs_buf_t       *bp;
1798         int             i;
1799         uint            count;          /* byte count of bwrite */
1800         uint            count_init;     /* initial count before roundup */
1801         int             roundoff;       /* roundoff to BB or stripe */
1802         int             split = 0;      /* split write into two regions */
1803         int             error;
1804         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1805         int             size;
1806
1807         XFS_STATS_INC(log->l_mp, xs_log_writes);
1808         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1809
1810         /* Add for LR header */
1811         count_init = log->l_iclog_hsize + iclog->ic_offset;
1812
1813         /* Round out the log write size */
1814         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1815                 /* we have a v2 stripe unit to use */
1816                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1817         } else {
1818                 count = BBTOB(BTOBB(count_init));
1819         }
1820         roundoff = count - count_init;
1821         ASSERT(roundoff >= 0);
1822         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1823                 roundoff < log->l_mp->m_sb.sb_logsunit)
1824                 ||
1825                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1826                  roundoff < BBTOB(1)));
1827
1828         /* move grant heads by roundoff in sync */
1829         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1830         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1831
1832         /* put cycle number in every block */
1833         xlog_pack_data(log, iclog, roundoff);
1834
1835         /* real byte length */
1836         size = iclog->ic_offset;
1837         if (v2)
1838                 size += roundoff;
1839         iclog->ic_header.h_len = cpu_to_be32(size);
1840
1841         bp = iclog->ic_bp;
1842         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1843
1844         XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1845
1846         /* Do we need to split this write into 2 parts? */
1847         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1848                 char            *dptr;
1849
1850                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1851                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1852                 iclog->ic_bwritecnt = 2;
1853
1854                 /*
1855                  * Bump the cycle numbers at the start of each block in the
1856                  * part of the iclog that ends up in the buffer that gets
1857                  * written to the start of the log.
1858                  *
1859                  * Watch out for the header magic number case, though.
1860                  */
1861                 dptr = (char *)&iclog->ic_header + count;
1862                 for (i = 0; i < split; i += BBSIZE) {
1863                         uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1864                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1865                                 cycle++;
1866                         *(__be32 *)dptr = cpu_to_be32(cycle);
1867
1868                         dptr += BBSIZE;
1869                 }
1870         } else {
1871                 iclog->ic_bwritecnt = 1;
1872         }
1873
1874         /* calculcate the checksum */
1875         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1876                                             iclog->ic_datap, size);
1877         /*
1878          * Intentionally corrupt the log record CRC based on the error injection
1879          * frequency, if defined. This facilitates testing log recovery in the
1880          * event of torn writes. Hence, set the IOABORT state to abort the log
1881          * write on I/O completion and shutdown the fs. The subsequent mount
1882          * detects the bad CRC and attempts to recover.
1883          */
1884 #ifdef DEBUG
1885         if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1886                 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1887                 iclog->ic_fail_crc = true;
1888                 xfs_warn(log->l_mp,
1889         "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1890                          be64_to_cpu(iclog->ic_header.h_lsn));
1891         }
1892 #endif
1893
1894         bp->b_io_length = BTOBB(count);
1895         bp->b_log_item = iclog;
1896         bp->b_flags &= ~XBF_FLUSH;
1897         bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1898
1899         /*
1900          * Flush the data device before flushing the log to make sure all meta
1901          * data written back from the AIL actually made it to disk before
1902          * stamping the new log tail LSN into the log buffer.  For an external
1903          * log we need to issue the flush explicitly, and unfortunately
1904          * synchronously here; for an internal log we can simply use the block
1905          * layer state machine for preflushes.
1906          */
1907         if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp || split)
1908                 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1909         else
1910                 bp->b_flags |= XBF_FLUSH;
1911
1912         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1913         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1914
1915         xlog_verify_iclog(log, iclog, count, true);
1916
1917         /* account for log which doesn't start at block #0 */
1918         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1919
1920         /*
1921          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1922          * is shutting down.
1923          */
1924         error = xlog_bdstrat(bp);
1925         if (error) {
1926                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1927                 return error;
1928         }
1929         if (split) {
1930                 bp = iclog->ic_log->l_xbuf;
1931                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1932                 xfs_buf_associate_memory(bp,
1933                                 (char *)&iclog->ic_header + count, split);
1934                 bp->b_log_item = iclog;
1935                 bp->b_flags &= ~XBF_FLUSH;
1936                 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1937
1938                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1939                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1940
1941                 /* account for internal log which doesn't start at block #0 */
1942                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1943                 error = xlog_bdstrat(bp);
1944                 if (error) {
1945                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1946                         return error;
1947                 }
1948         }
1949         return 0;
1950 }       /* xlog_sync */
1951
1952 /*
1953  * Deallocate a log structure
1954  */
1955 STATIC void
1956 xlog_dealloc_log(
1957         struct xlog     *log)
1958 {
1959         xlog_in_core_t  *iclog, *next_iclog;
1960         int             i;
1961
1962         xlog_cil_destroy(log);
1963
1964         /*
1965          * Cycle all the iclogbuf locks to make sure all log IO completion
1966          * is done before we tear down these buffers.
1967          */
1968         iclog = log->l_iclog;
1969         for (i = 0; i < log->l_iclog_bufs; i++) {
1970                 xfs_buf_lock(iclog->ic_bp);
1971                 xfs_buf_unlock(iclog->ic_bp);
1972                 iclog = iclog->ic_next;
1973         }
1974
1975         /*
1976          * Always need to ensure that the extra buffer does not point to memory
1977          * owned by another log buffer before we free it. Also, cycle the lock
1978          * first to ensure we've completed IO on it.
1979          */
1980         xfs_buf_lock(log->l_xbuf);
1981         xfs_buf_unlock(log->l_xbuf);
1982         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1983         xfs_buf_free(log->l_xbuf);
1984
1985         iclog = log->l_iclog;
1986         for (i = 0; i < log->l_iclog_bufs; i++) {
1987                 xfs_buf_free(iclog->ic_bp);
1988                 next_iclog = iclog->ic_next;
1989                 kmem_free(iclog);
1990                 iclog = next_iclog;
1991         }
1992
1993         log->l_mp->m_log = NULL;
1994         kmem_free(log);
1995 }       /* xlog_dealloc_log */
1996
1997 /*
1998  * Update counters atomically now that memcpy is done.
1999  */
2000 /* ARGSUSED */
2001 static inline void
2002 xlog_state_finish_copy(
2003         struct xlog             *log,
2004         struct xlog_in_core     *iclog,
2005         int                     record_cnt,
2006         int                     copy_bytes)
2007 {
2008         spin_lock(&log->l_icloglock);
2009
2010         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2011         iclog->ic_offset += copy_bytes;
2012
2013         spin_unlock(&log->l_icloglock);
2014 }       /* xlog_state_finish_copy */
2015
2016
2017
2018
2019 /*
2020  * print out info relating to regions written which consume
2021  * the reservation
2022  */
2023 void
2024 xlog_print_tic_res(
2025         struct xfs_mount        *mp,
2026         struct xlog_ticket      *ticket)
2027 {
2028         uint i;
2029         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2030
2031         /* match with XLOG_REG_TYPE_* in xfs_log.h */
2032 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2033         static char *res_type_str[] = {
2034             REG_TYPE_STR(BFORMAT, "bformat"),
2035             REG_TYPE_STR(BCHUNK, "bchunk"),
2036             REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2037             REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2038             REG_TYPE_STR(IFORMAT, "iformat"),
2039             REG_TYPE_STR(ICORE, "icore"),
2040             REG_TYPE_STR(IEXT, "iext"),
2041             REG_TYPE_STR(IBROOT, "ibroot"),
2042             REG_TYPE_STR(ILOCAL, "ilocal"),
2043             REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2044             REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2045             REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2046             REG_TYPE_STR(QFORMAT, "qformat"),
2047             REG_TYPE_STR(DQUOT, "dquot"),
2048             REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2049             REG_TYPE_STR(LRHEADER, "LR header"),
2050             REG_TYPE_STR(UNMOUNT, "unmount"),
2051             REG_TYPE_STR(COMMIT, "commit"),
2052             REG_TYPE_STR(TRANSHDR, "trans header"),
2053             REG_TYPE_STR(ICREATE, "inode create"),
2054             REG_TYPE_STR(RUI_FORMAT, "rui_format"),
2055             REG_TYPE_STR(RUD_FORMAT, "rud_format"),
2056             REG_TYPE_STR(CUI_FORMAT, "cui_format"),
2057             REG_TYPE_STR(CUD_FORMAT, "cud_format"),
2058             REG_TYPE_STR(BUI_FORMAT, "bui_format"),
2059             REG_TYPE_STR(BUD_FORMAT, "bud_format"),
2060         };
2061         BUILD_BUG_ON(ARRAY_SIZE(res_type_str) != XLOG_REG_TYPE_MAX + 1);
2062 #undef REG_TYPE_STR
2063
2064         xfs_warn(mp, "ticket reservation summary:");
2065         xfs_warn(mp, "  unit res    = %d bytes",
2066                  ticket->t_unit_res);
2067         xfs_warn(mp, "  current res = %d bytes",
2068                  ticket->t_curr_res);
2069         xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2070                  ticket->t_res_arr_sum, ticket->t_res_o_flow);
2071         xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2072                  ticket->t_res_num_ophdrs, ophdr_spc);
2073         xfs_warn(mp, "  ophdr + reg = %u bytes",
2074                  ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2075         xfs_warn(mp, "  num regions = %u",
2076                  ticket->t_res_num);
2077
2078         for (i = 0; i < ticket->t_res_num; i++) {
2079                 uint r_type = ticket->t_res_arr[i].r_type;
2080                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2081                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2082                             "bad-rtype" : res_type_str[r_type]),
2083                             ticket->t_res_arr[i].r_len);
2084         }
2085 }
2086
2087 /*
2088  * Print a summary of the transaction.
2089  */
2090 void
2091 xlog_print_trans(
2092         struct xfs_trans        *tp)
2093 {
2094         struct xfs_mount        *mp = tp->t_mountp;
2095         struct xfs_log_item     *lip;
2096
2097         /* dump core transaction and ticket info */
2098         xfs_warn(mp, "transaction summary:");
2099         xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2100         xfs_warn(mp, "  log count = %d", tp->t_log_count);
2101         xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2102
2103         xlog_print_tic_res(mp, tp->t_ticket);
2104
2105         /* dump each log item */
2106         list_for_each_entry(lip, &tp->t_items, li_trans) {
2107                 struct xfs_log_vec      *lv = lip->li_lv;
2108                 struct xfs_log_iovec    *vec;
2109                 int                     i;
2110
2111                 xfs_warn(mp, "log item: ");
2112                 xfs_warn(mp, "  type    = 0x%x", lip->li_type);
2113                 xfs_warn(mp, "  flags   = 0x%lx", lip->li_flags);
2114                 if (!lv)
2115                         continue;
2116                 xfs_warn(mp, "  niovecs = %d", lv->lv_niovecs);
2117                 xfs_warn(mp, "  size    = %d", lv->lv_size);
2118                 xfs_warn(mp, "  bytes   = %d", lv->lv_bytes);
2119                 xfs_warn(mp, "  buf len = %d", lv->lv_buf_len);
2120
2121                 /* dump each iovec for the log item */
2122                 vec = lv->lv_iovecp;
2123                 for (i = 0; i < lv->lv_niovecs; i++) {
2124                         int dumplen = min(vec->i_len, 32);
2125
2126                         xfs_warn(mp, "  iovec[%d]", i);
2127                         xfs_warn(mp, "    type  = 0x%x", vec->i_type);
2128                         xfs_warn(mp, "    len   = %d", vec->i_len);
2129                         xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2130                         xfs_hex_dump(vec->i_addr, dumplen);
2131
2132                         vec++;
2133                 }
2134         }
2135 }
2136
2137 /*
2138  * Calculate the potential space needed by the log vector.  Each region gets
2139  * its own xlog_op_header_t and may need to be double word aligned.
2140  */
2141 static int
2142 xlog_write_calc_vec_length(
2143         struct xlog_ticket      *ticket,
2144         struct xfs_log_vec      *log_vector)
2145 {
2146         struct xfs_log_vec      *lv;
2147         int                     headers = 0;
2148         int                     len = 0;
2149         int                     i;
2150
2151         /* acct for start rec of xact */
2152         if (ticket->t_flags & XLOG_TIC_INITED)
2153                 headers++;
2154
2155         for (lv = log_vector; lv; lv = lv->lv_next) {
2156                 /* we don't write ordered log vectors */
2157                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2158                         continue;
2159
2160                 headers += lv->lv_niovecs;
2161
2162                 for (i = 0; i < lv->lv_niovecs; i++) {
2163                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2164
2165                         len += vecp->i_len;
2166                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2167                 }
2168         }
2169
2170         ticket->t_res_num_ophdrs += headers;
2171         len += headers * sizeof(struct xlog_op_header);
2172
2173         return len;
2174 }
2175
2176 /*
2177  * If first write for transaction, insert start record  We can't be trying to
2178  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2179  */
2180 static int
2181 xlog_write_start_rec(
2182         struct xlog_op_header   *ophdr,
2183         struct xlog_ticket      *ticket)
2184 {
2185         if (!(ticket->t_flags & XLOG_TIC_INITED))
2186                 return 0;
2187
2188         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2189         ophdr->oh_clientid = ticket->t_clientid;
2190         ophdr->oh_len = 0;
2191         ophdr->oh_flags = XLOG_START_TRANS;
2192         ophdr->oh_res2 = 0;
2193
2194         ticket->t_flags &= ~XLOG_TIC_INITED;
2195
2196         return sizeof(struct xlog_op_header);
2197 }
2198
2199 static xlog_op_header_t *
2200 xlog_write_setup_ophdr(
2201         struct xlog             *log,
2202         struct xlog_op_header   *ophdr,
2203         struct xlog_ticket      *ticket,
2204         uint                    flags)
2205 {
2206         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2207         ophdr->oh_clientid = ticket->t_clientid;
2208         ophdr->oh_res2 = 0;
2209
2210         /* are we copying a commit or unmount record? */
2211         ophdr->oh_flags = flags;
2212
2213         /*
2214          * We've seen logs corrupted with bad transaction client ids.  This
2215          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2216          * and shut down the filesystem.
2217          */
2218         switch (ophdr->oh_clientid)  {
2219         case XFS_TRANSACTION:
2220         case XFS_VOLUME:
2221         case XFS_LOG:
2222                 break;
2223         default:
2224                 xfs_warn(log->l_mp,
2225                         "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2226                         ophdr->oh_clientid, ticket);
2227                 return NULL;
2228         }
2229
2230         return ophdr;
2231 }
2232
2233 /*
2234  * Set up the parameters of the region copy into the log. This has
2235  * to handle region write split across multiple log buffers - this
2236  * state is kept external to this function so that this code can
2237  * be written in an obvious, self documenting manner.
2238  */
2239 static int
2240 xlog_write_setup_copy(
2241         struct xlog_ticket      *ticket,
2242         struct xlog_op_header   *ophdr,
2243         int                     space_available,
2244         int                     space_required,
2245         int                     *copy_off,
2246         int                     *copy_len,
2247         int                     *last_was_partial_copy,
2248         int                     *bytes_consumed)
2249 {
2250         int                     still_to_copy;
2251
2252         still_to_copy = space_required - *bytes_consumed;
2253         *copy_off = *bytes_consumed;
2254
2255         if (still_to_copy <= space_available) {
2256                 /* write of region completes here */
2257                 *copy_len = still_to_copy;
2258                 ophdr->oh_len = cpu_to_be32(*copy_len);
2259                 if (*last_was_partial_copy)
2260                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2261                 *last_was_partial_copy = 0;
2262                 *bytes_consumed = 0;
2263                 return 0;
2264         }
2265
2266         /* partial write of region, needs extra log op header reservation */
2267         *copy_len = space_available;
2268         ophdr->oh_len = cpu_to_be32(*copy_len);
2269         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2270         if (*last_was_partial_copy)
2271                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2272         *bytes_consumed += *copy_len;
2273         (*last_was_partial_copy)++;
2274
2275         /* account for new log op header */
2276         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2277         ticket->t_res_num_ophdrs++;
2278
2279         return sizeof(struct xlog_op_header);
2280 }
2281
2282 static int
2283 xlog_write_copy_finish(
2284         struct xlog             *log,
2285         struct xlog_in_core     *iclog,
2286         uint                    flags,
2287         int                     *record_cnt,
2288         int                     *data_cnt,
2289         int                     *partial_copy,
2290         int                     *partial_copy_len,
2291         int                     log_offset,
2292         struct xlog_in_core     **commit_iclog)
2293 {
2294         if (*partial_copy) {
2295                 /*
2296                  * This iclog has already been marked WANT_SYNC by
2297                  * xlog_state_get_iclog_space.
2298                  */
2299                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2300                 *record_cnt = 0;
2301                 *data_cnt = 0;
2302                 return xlog_state_release_iclog(log, iclog);
2303         }
2304
2305         *partial_copy = 0;
2306         *partial_copy_len = 0;
2307
2308         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2309                 /* no more space in this iclog - push it. */
2310                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2311                 *record_cnt = 0;
2312                 *data_cnt = 0;
2313
2314                 spin_lock(&log->l_icloglock);
2315                 xlog_state_want_sync(log, iclog);
2316                 spin_unlock(&log->l_icloglock);
2317
2318                 if (!commit_iclog)
2319                         return xlog_state_release_iclog(log, iclog);
2320                 ASSERT(flags & XLOG_COMMIT_TRANS);
2321                 *commit_iclog = iclog;
2322         }
2323
2324         return 0;
2325 }
2326
2327 /*
2328  * Write some region out to in-core log
2329  *
2330  * This will be called when writing externally provided regions or when
2331  * writing out a commit record for a given transaction.
2332  *
2333  * General algorithm:
2334  *      1. Find total length of this write.  This may include adding to the
2335  *              lengths passed in.
2336  *      2. Check whether we violate the tickets reservation.
2337  *      3. While writing to this iclog
2338  *          A. Reserve as much space in this iclog as can get
2339  *          B. If this is first write, save away start lsn
2340  *          C. While writing this region:
2341  *              1. If first write of transaction, write start record
2342  *              2. Write log operation header (header per region)
2343  *              3. Find out if we can fit entire region into this iclog
2344  *              4. Potentially, verify destination memcpy ptr
2345  *              5. Memcpy (partial) region
2346  *              6. If partial copy, release iclog; otherwise, continue
2347  *                      copying more regions into current iclog
2348  *      4. Mark want sync bit (in simulation mode)
2349  *      5. Release iclog for potential flush to on-disk log.
2350  *
2351  * ERRORS:
2352  * 1.   Panic if reservation is overrun.  This should never happen since
2353  *      reservation amounts are generated internal to the filesystem.
2354  * NOTES:
2355  * 1. Tickets are single threaded data structures.
2356  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2357  *      syncing routine.  When a single log_write region needs to span
2358  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2359  *      on all log operation writes which don't contain the end of the
2360  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2361  *      operation which contains the end of the continued log_write region.
2362  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2363  *      we don't really know exactly how much space will be used.  As a result,
2364  *      we don't update ic_offset until the end when we know exactly how many
2365  *      bytes have been written out.
2366  */
2367 int
2368 xlog_write(
2369         struct xlog             *log,
2370         struct xfs_log_vec      *log_vector,
2371         struct xlog_ticket      *ticket,
2372         xfs_lsn_t               *start_lsn,
2373         struct xlog_in_core     **commit_iclog,
2374         uint                    flags)
2375 {
2376         struct xlog_in_core     *iclog = NULL;
2377         struct xfs_log_iovec    *vecp;
2378         struct xfs_log_vec      *lv;
2379         int                     len;
2380         int                     index;
2381         int                     partial_copy = 0;
2382         int                     partial_copy_len = 0;
2383         int                     contwr = 0;
2384         int                     record_cnt = 0;
2385         int                     data_cnt = 0;
2386         int                     error;
2387
2388         *start_lsn = 0;
2389
2390         len = xlog_write_calc_vec_length(ticket, log_vector);
2391
2392         /*
2393          * Region headers and bytes are already accounted for.
2394          * We only need to take into account start records and
2395          * split regions in this function.
2396          */
2397         if (ticket->t_flags & XLOG_TIC_INITED)
2398                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2399
2400         /*
2401          * Commit record headers need to be accounted for. These
2402          * come in as separate writes so are easy to detect.
2403          */
2404         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2405                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2406
2407         if (ticket->t_curr_res < 0) {
2408                 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2409                      "ctx ticket reservation ran out. Need to up reservation");
2410                 xlog_print_tic_res(log->l_mp, ticket);
2411                 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2412         }
2413
2414         index = 0;
2415         lv = log_vector;
2416         vecp = lv->lv_iovecp;
2417         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2418                 void            *ptr;
2419                 int             log_offset;
2420
2421                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2422                                                    &contwr, &log_offset);
2423                 if (error)
2424                         return error;
2425
2426                 ASSERT(log_offset <= iclog->ic_size - 1);
2427                 ptr = iclog->ic_datap + log_offset;
2428
2429                 /* start_lsn is the first lsn written to. That's all we need. */
2430                 if (!*start_lsn)
2431                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2432
2433                 /*
2434                  * This loop writes out as many regions as can fit in the amount
2435                  * of space which was allocated by xlog_state_get_iclog_space().
2436                  */
2437                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2438                         struct xfs_log_iovec    *reg;
2439                         struct xlog_op_header   *ophdr;
2440                         int                     start_rec_copy;
2441                         int                     copy_len;
2442                         int                     copy_off;
2443                         bool                    ordered = false;
2444
2445                         /* ordered log vectors have no regions to write */
2446                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2447                                 ASSERT(lv->lv_niovecs == 0);
2448                                 ordered = true;
2449                                 goto next_lv;
2450                         }
2451
2452                         reg = &vecp[index];
2453                         ASSERT(reg->i_len % sizeof(int32_t) == 0);
2454                         ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2455
2456                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2457                         if (start_rec_copy) {
2458                                 record_cnt++;
2459                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2460                                                    start_rec_copy);
2461                         }
2462
2463                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2464                         if (!ophdr)
2465                                 return -EIO;
2466
2467                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2468                                            sizeof(struct xlog_op_header));
2469
2470                         len += xlog_write_setup_copy(ticket, ophdr,
2471                                                      iclog->ic_size-log_offset,
2472                                                      reg->i_len,
2473                                                      &copy_off, &copy_len,
2474                                                      &partial_copy,
2475                                                      &partial_copy_len);
2476                         xlog_verify_dest_ptr(log, ptr);
2477
2478                         /*
2479                          * Copy region.
2480                          *
2481                          * Unmount records just log an opheader, so can have
2482                          * empty payloads with no data region to copy. Hence we
2483                          * only copy the payload if the vector says it has data
2484                          * to copy.
2485                          */
2486                         ASSERT(copy_len >= 0);
2487                         if (copy_len > 0) {
2488                                 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2489                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2490                                                    copy_len);
2491                         }
2492                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2493                         record_cnt++;
2494                         data_cnt += contwr ? copy_len : 0;
2495
2496                         error = xlog_write_copy_finish(log, iclog, flags,
2497                                                        &record_cnt, &data_cnt,
2498                                                        &partial_copy,
2499                                                        &partial_copy_len,
2500                                                        log_offset,
2501                                                        commit_iclog);
2502                         if (error)
2503                                 return error;
2504
2505                         /*
2506                          * if we had a partial copy, we need to get more iclog
2507                          * space but we don't want to increment the region
2508                          * index because there is still more is this region to
2509                          * write.
2510                          *
2511                          * If we completed writing this region, and we flushed
2512                          * the iclog (indicated by resetting of the record
2513                          * count), then we also need to get more log space. If
2514                          * this was the last record, though, we are done and
2515                          * can just return.
2516                          */
2517                         if (partial_copy)
2518                                 break;
2519
2520                         if (++index == lv->lv_niovecs) {
2521 next_lv:
2522                                 lv = lv->lv_next;
2523                                 index = 0;
2524                                 if (lv)
2525                                         vecp = lv->lv_iovecp;
2526                         }
2527                         if (record_cnt == 0 && !ordered) {
2528                                 if (!lv)
2529                                         return 0;
2530                                 break;
2531                         }
2532                 }
2533         }
2534
2535         ASSERT(len == 0);
2536
2537         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2538         if (!commit_iclog)
2539                 return xlog_state_release_iclog(log, iclog);
2540
2541         ASSERT(flags & XLOG_COMMIT_TRANS);
2542         *commit_iclog = iclog;
2543         return 0;
2544 }
2545
2546
2547 /*****************************************************************************
2548  *
2549  *              State Machine functions
2550  *
2551  *****************************************************************************
2552  */
2553
2554 /* Clean iclogs starting from the head.  This ordering must be
2555  * maintained, so an iclog doesn't become ACTIVE beyond one that
2556  * is SYNCING.  This is also required to maintain the notion that we use
2557  * a ordered wait queue to hold off would be writers to the log when every
2558  * iclog is trying to sync to disk.
2559  *
2560  * State Change: DIRTY -> ACTIVE
2561  */
2562 STATIC void
2563 xlog_state_clean_log(
2564         struct xlog *log)
2565 {
2566         xlog_in_core_t  *iclog;
2567         int changed = 0;
2568
2569         iclog = log->l_iclog;
2570         do {
2571                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2572                         iclog->ic_state = XLOG_STATE_ACTIVE;
2573                         iclog->ic_offset       = 0;
2574                         ASSERT(iclog->ic_callback == NULL);
2575                         /*
2576                          * If the number of ops in this iclog indicate it just
2577                          * contains the dummy transaction, we can
2578                          * change state into IDLE (the second time around).
2579                          * Otherwise we should change the state into
2580                          * NEED a dummy.
2581                          * We don't need to cover the dummy.
2582                          */
2583                         if (!changed &&
2584                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2585                                         XLOG_COVER_OPS)) {
2586                                 changed = 1;
2587                         } else {
2588                                 /*
2589                                  * We have two dirty iclogs so start over
2590                                  * This could also be num of ops indicates
2591                                  * this is not the dummy going out.
2592                                  */
2593                                 changed = 2;
2594                         }
2595                         iclog->ic_header.h_num_logops = 0;
2596                         memset(iclog->ic_header.h_cycle_data, 0,
2597                               sizeof(iclog->ic_header.h_cycle_data));
2598                         iclog->ic_header.h_lsn = 0;
2599                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2600                         /* do nothing */;
2601                 else
2602                         break;  /* stop cleaning */
2603                 iclog = iclog->ic_next;
2604         } while (iclog != log->l_iclog);
2605
2606         /* log is locked when we are called */
2607         /*
2608          * Change state for the dummy log recording.
2609          * We usually go to NEED. But we go to NEED2 if the changed indicates
2610          * we are done writing the dummy record.
2611          * If we are done with the second dummy recored (DONE2), then
2612          * we go to IDLE.
2613          */
2614         if (changed) {
2615                 switch (log->l_covered_state) {
2616                 case XLOG_STATE_COVER_IDLE:
2617                 case XLOG_STATE_COVER_NEED:
2618                 case XLOG_STATE_COVER_NEED2:
2619                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2620                         break;
2621
2622                 case XLOG_STATE_COVER_DONE:
2623                         if (changed == 1)
2624                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2625                         else
2626                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2627                         break;
2628
2629                 case XLOG_STATE_COVER_DONE2:
2630                         if (changed == 1)
2631                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2632                         else
2633                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2634                         break;
2635
2636                 default:
2637                         ASSERT(0);
2638                 }
2639         }
2640 }       /* xlog_state_clean_log */
2641
2642 STATIC xfs_lsn_t
2643 xlog_get_lowest_lsn(
2644         struct xlog             *log)
2645 {
2646         struct xlog_in_core     *iclog = log->l_iclog;
2647         xfs_lsn_t               lowest_lsn = 0, lsn;
2648
2649         do {
2650                 if (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))
2651                         continue;
2652
2653                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2654                 if ((lsn && !lowest_lsn) || XFS_LSN_CMP(lsn, lowest_lsn) < 0)
2655                         lowest_lsn = lsn;
2656         } while ((iclog = iclog->ic_next) != log->l_iclog);
2657
2658         return lowest_lsn;
2659 }
2660
2661 STATIC void
2662 xlog_state_do_callback(
2663         struct xlog             *log,
2664         int                     aborted,
2665         struct xlog_in_core     *ciclog)
2666 {
2667         xlog_in_core_t     *iclog;
2668         xlog_in_core_t     *first_iclog;        /* used to know when we've
2669                                                  * processed all iclogs once */
2670         xfs_log_callback_t *cb, *cb_next;
2671         int                flushcnt = 0;
2672         xfs_lsn_t          lowest_lsn;
2673         int                ioerrors;    /* counter: iclogs with errors */
2674         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2675         int                funcdidcallbacks; /* flag: function did callbacks */
2676         int                repeats;     /* for issuing console warnings if
2677                                          * looping too many times */
2678         int                wake = 0;
2679
2680         spin_lock(&log->l_icloglock);
2681         first_iclog = iclog = log->l_iclog;
2682         ioerrors = 0;
2683         funcdidcallbacks = 0;
2684         repeats = 0;
2685
2686         do {
2687                 /*
2688                  * Scan all iclogs starting with the one pointed to by the
2689                  * log.  Reset this starting point each time the log is
2690                  * unlocked (during callbacks).
2691                  *
2692                  * Keep looping through iclogs until one full pass is made
2693                  * without running any callbacks.
2694                  */
2695                 first_iclog = log->l_iclog;
2696                 iclog = log->l_iclog;
2697                 loopdidcallbacks = 0;
2698                 repeats++;
2699
2700                 do {
2701
2702                         /* skip all iclogs in the ACTIVE & DIRTY states */
2703                         if (iclog->ic_state &
2704                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2705                                 iclog = iclog->ic_next;
2706                                 continue;
2707                         }
2708
2709                         /*
2710                          * Between marking a filesystem SHUTDOWN and stopping
2711                          * the log, we do flush all iclogs to disk (if there
2712                          * wasn't a log I/O error). So, we do want things to
2713                          * go smoothly in case of just a SHUTDOWN  w/o a
2714                          * LOG_IO_ERROR.
2715                          */
2716                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2717                                 /*
2718                                  * Can only perform callbacks in order.  Since
2719                                  * this iclog is not in the DONE_SYNC/
2720                                  * DO_CALLBACK state, we skip the rest and
2721                                  * just try to clean up.  If we set our iclog
2722                                  * to DO_CALLBACK, we will not process it when
2723                                  * we retry since a previous iclog is in the
2724                                  * CALLBACK and the state cannot change since
2725                                  * we are holding the l_icloglock.
2726                                  */
2727                                 if (!(iclog->ic_state &
2728                                         (XLOG_STATE_DONE_SYNC |
2729                                                  XLOG_STATE_DO_CALLBACK))) {
2730                                         if (ciclog && (ciclog->ic_state ==
2731                                                         XLOG_STATE_DONE_SYNC)) {
2732                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2733                                         }
2734                                         break;
2735                                 }
2736                                 /*
2737                                  * We now have an iclog that is in either the
2738                                  * DO_CALLBACK or DONE_SYNC states. The other
2739                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2740                                  * caught by the above if and are going to
2741                                  * clean (i.e. we aren't doing their callbacks)
2742                                  * see the above if.
2743                                  */
2744
2745                                 /*
2746                                  * We will do one more check here to see if we
2747                                  * have chased our tail around.
2748                                  */
2749
2750                                 lowest_lsn = xlog_get_lowest_lsn(log);
2751                                 if (lowest_lsn &&
2752                                     XFS_LSN_CMP(lowest_lsn,
2753                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2754                                         iclog = iclog->ic_next;
2755                                         continue; /* Leave this iclog for
2756                                                    * another thread */
2757                                 }
2758
2759                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2760
2761
2762                                 /*
2763                                  * Completion of a iclog IO does not imply that
2764                                  * a transaction has completed, as transactions
2765                                  * can be large enough to span many iclogs. We
2766                                  * cannot change the tail of the log half way
2767                                  * through a transaction as this may be the only
2768                                  * transaction in the log and moving th etail to
2769                                  * point to the middle of it will prevent
2770                                  * recovery from finding the start of the
2771                                  * transaction. Hence we should only update the
2772                                  * last_sync_lsn if this iclog contains
2773                                  * transaction completion callbacks on it.
2774                                  *
2775                                  * We have to do this before we drop the
2776                                  * icloglock to ensure we are the only one that
2777                                  * can update it.
2778                                  */
2779                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2780                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2781                                 if (iclog->ic_callback)
2782                                         atomic64_set(&log->l_last_sync_lsn,
2783                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2784
2785                         } else
2786                                 ioerrors++;
2787
2788                         spin_unlock(&log->l_icloglock);
2789
2790                         /*
2791                          * Keep processing entries in the callback list until
2792                          * we come around and it is empty.  We need to
2793                          * atomically see that the list is empty and change the
2794                          * state to DIRTY so that we don't miss any more
2795                          * callbacks being added.
2796                          */
2797                         spin_lock(&iclog->ic_callback_lock);
2798                         cb = iclog->ic_callback;
2799                         while (cb) {
2800                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2801                                 iclog->ic_callback = NULL;
2802                                 spin_unlock(&iclog->ic_callback_lock);
2803
2804                                 /* perform callbacks in the order given */
2805                                 for (; cb; cb = cb_next) {
2806                                         cb_next = cb->cb_next;
2807                                         cb->cb_func(cb->cb_arg, aborted);
2808                                 }
2809                                 spin_lock(&iclog->ic_callback_lock);
2810                                 cb = iclog->ic_callback;
2811                         }
2812
2813                         loopdidcallbacks++;
2814                         funcdidcallbacks++;
2815
2816                         spin_lock(&log->l_icloglock);
2817                         ASSERT(iclog->ic_callback == NULL);
2818                         spin_unlock(&iclog->ic_callback_lock);
2819                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2820                                 iclog->ic_state = XLOG_STATE_DIRTY;
2821
2822                         /*
2823                          * Transition from DIRTY to ACTIVE if applicable.
2824                          * NOP if STATE_IOERROR.
2825                          */
2826                         xlog_state_clean_log(log);
2827
2828                         /* wake up threads waiting in xfs_log_force() */
2829                         wake_up_all(&iclog->ic_force_wait);
2830
2831                         iclog = iclog->ic_next;
2832                 } while (first_iclog != iclog);
2833
2834                 if (repeats > 5000) {
2835                         flushcnt += repeats;
2836                         repeats = 0;
2837                         xfs_warn(log->l_mp,
2838                                 "%s: possible infinite loop (%d iterations)",
2839                                 __func__, flushcnt);
2840                 }
2841         } while (!ioerrors && loopdidcallbacks);
2842
2843 #ifdef DEBUG
2844         /*
2845          * Make one last gasp attempt to see if iclogs are being left in limbo.
2846          * If the above loop finds an iclog earlier than the current iclog and
2847          * in one of the syncing states, the current iclog is put into
2848          * DO_CALLBACK and the callbacks are deferred to the completion of the
2849          * earlier iclog. Walk the iclogs in order and make sure that no iclog
2850          * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2851          * states.
2852          *
2853          * Note that SYNCING|IOABORT is a valid state so we cannot just check
2854          * for ic_state == SYNCING.
2855          */
2856         if (funcdidcallbacks) {
2857                 first_iclog = iclog = log->l_iclog;
2858                 do {
2859                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2860                         /*
2861                          * Terminate the loop if iclogs are found in states
2862                          * which will cause other threads to clean up iclogs.
2863                          *
2864                          * SYNCING - i/o completion will go through logs
2865                          * DONE_SYNC - interrupt thread should be waiting for
2866                          *              l_icloglock
2867                          * IOERROR - give up hope all ye who enter here
2868                          */
2869                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2870                             iclog->ic_state & XLOG_STATE_SYNCING ||
2871                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2872                             iclog->ic_state == XLOG_STATE_IOERROR )
2873                                 break;
2874                         iclog = iclog->ic_next;
2875                 } while (first_iclog != iclog);
2876         }
2877 #endif
2878
2879         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2880                 wake = 1;
2881         spin_unlock(&log->l_icloglock);
2882
2883         if (wake)
2884                 wake_up_all(&log->l_flush_wait);
2885 }
2886
2887
2888 /*
2889  * Finish transitioning this iclog to the dirty state.
2890  *
2891  * Make sure that we completely execute this routine only when this is
2892  * the last call to the iclog.  There is a good chance that iclog flushes,
2893  * when we reach the end of the physical log, get turned into 2 separate
2894  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2895  * routine.  By using the reference count bwritecnt, we guarantee that only
2896  * the second completion goes through.
2897  *
2898  * Callbacks could take time, so they are done outside the scope of the
2899  * global state machine log lock.
2900  */
2901 STATIC void
2902 xlog_state_done_syncing(
2903         xlog_in_core_t  *iclog,
2904         int             aborted)
2905 {
2906         struct xlog        *log = iclog->ic_log;
2907
2908         spin_lock(&log->l_icloglock);
2909
2910         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2911                iclog->ic_state == XLOG_STATE_IOERROR);
2912         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2913         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2914
2915
2916         /*
2917          * If we got an error, either on the first buffer, or in the case of
2918          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2919          * and none should ever be attempted to be written to disk
2920          * again.
2921          */
2922         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2923                 if (--iclog->ic_bwritecnt == 1) {
2924                         spin_unlock(&log->l_icloglock);
2925                         return;
2926                 }
2927                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2928         }
2929
2930         /*
2931          * Someone could be sleeping prior to writing out the next
2932          * iclog buffer, we wake them all, one will get to do the
2933          * I/O, the others get to wait for the result.
2934          */
2935         wake_up_all(&iclog->ic_write_wait);
2936         spin_unlock(&log->l_icloglock);
2937         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2938 }       /* xlog_state_done_syncing */
2939
2940
2941 /*
2942  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2943  * sleep.  We wait on the flush queue on the head iclog as that should be
2944  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2945  * we will wait here and all new writes will sleep until a sync completes.
2946  *
2947  * The in-core logs are used in a circular fashion. They are not used
2948  * out-of-order even when an iclog past the head is free.
2949  *
2950  * return:
2951  *      * log_offset where xlog_write() can start writing into the in-core
2952  *              log's data space.
2953  *      * in-core log pointer to which xlog_write() should write.
2954  *      * boolean indicating this is a continued write to an in-core log.
2955  *              If this is the last write, then the in-core log's offset field
2956  *              needs to be incremented, depending on the amount of data which
2957  *              is copied.
2958  */
2959 STATIC int
2960 xlog_state_get_iclog_space(
2961         struct xlog             *log,
2962         int                     len,
2963         struct xlog_in_core     **iclogp,
2964         struct xlog_ticket      *ticket,
2965         int                     *continued_write,
2966         int                     *logoffsetp)
2967 {
2968         int               log_offset;
2969         xlog_rec_header_t *head;
2970         xlog_in_core_t    *iclog;
2971         int               error;
2972
2973 restart:
2974         spin_lock(&log->l_icloglock);
2975         if (XLOG_FORCED_SHUTDOWN(log)) {
2976                 spin_unlock(&log->l_icloglock);
2977                 return -EIO;
2978         }
2979
2980         iclog = log->l_iclog;
2981         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2982                 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2983
2984                 /* Wait for log writes to have flushed */
2985                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2986                 goto restart;
2987         }
2988
2989         head = &iclog->ic_header;
2990
2991         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2992         log_offset = iclog->ic_offset;
2993
2994         /* On the 1st write to an iclog, figure out lsn.  This works
2995          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2996          * committing to.  If the offset is set, that's how many blocks
2997          * must be written.
2998          */
2999         if (log_offset == 0) {
3000                 ticket->t_curr_res -= log->l_iclog_hsize;
3001                 xlog_tic_add_region(ticket,
3002                                     log->l_iclog_hsize,
3003                                     XLOG_REG_TYPE_LRHEADER);
3004                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3005                 head->h_lsn = cpu_to_be64(
3006                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3007                 ASSERT(log->l_curr_block >= 0);
3008         }
3009
3010         /* If there is enough room to write everything, then do it.  Otherwise,
3011          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3012          * bit is on, so this will get flushed out.  Don't update ic_offset
3013          * until you know exactly how many bytes get copied.  Therefore, wait
3014          * until later to update ic_offset.
3015          *
3016          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3017          * can fit into remaining data section.
3018          */
3019         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3020                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3021
3022                 /*
3023                  * If I'm the only one writing to this iclog, sync it to disk.
3024                  * We need to do an atomic compare and decrement here to avoid
3025                  * racing with concurrent atomic_dec_and_lock() calls in
3026                  * xlog_state_release_iclog() when there is more than one
3027                  * reference to the iclog.
3028                  */
3029                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3030                         /* we are the only one */
3031                         spin_unlock(&log->l_icloglock);
3032                         error = xlog_state_release_iclog(log, iclog);
3033                         if (error)
3034                                 return error;
3035                 } else {
3036                         spin_unlock(&log->l_icloglock);
3037                 }
3038                 goto restart;
3039         }
3040
3041         /* Do we have enough room to write the full amount in the remainder
3042          * of this iclog?  Or must we continue a write on the next iclog and
3043          * mark this iclog as completely taken?  In the case where we switch
3044          * iclogs (to mark it taken), this particular iclog will release/sync
3045          * to disk in xlog_write().
3046          */
3047         if (len <= iclog->ic_size - iclog->ic_offset) {
3048                 *continued_write = 0;
3049                 iclog->ic_offset += len;
3050         } else {
3051                 *continued_write = 1;
3052                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3053         }
3054         *iclogp = iclog;
3055
3056         ASSERT(iclog->ic_offset <= iclog->ic_size);
3057         spin_unlock(&log->l_icloglock);
3058
3059         *logoffsetp = log_offset;
3060         return 0;
3061 }       /* xlog_state_get_iclog_space */
3062
3063 /* The first cnt-1 times through here we don't need to
3064  * move the grant write head because the permanent
3065  * reservation has reserved cnt times the unit amount.
3066  * Release part of current permanent unit reservation and
3067  * reset current reservation to be one units worth.  Also
3068  * move grant reservation head forward.
3069  */
3070 STATIC void
3071 xlog_regrant_reserve_log_space(
3072         struct xlog             *log,
3073         struct xlog_ticket      *ticket)
3074 {
3075         trace_xfs_log_regrant_reserve_enter(log, ticket);
3076
3077         if (ticket->t_cnt > 0)
3078                 ticket->t_cnt--;
3079
3080         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3081                                         ticket->t_curr_res);
3082         xlog_grant_sub_space(log, &log->l_write_head.grant,
3083                                         ticket->t_curr_res);
3084         ticket->t_curr_res = ticket->t_unit_res;
3085         xlog_tic_reset_res(ticket);
3086
3087         trace_xfs_log_regrant_reserve_sub(log, ticket);
3088
3089         /* just return if we still have some of the pre-reserved space */
3090         if (ticket->t_cnt > 0)
3091                 return;
3092
3093         xlog_grant_add_space(log, &log->l_reserve_head.grant,
3094                                         ticket->t_unit_res);
3095
3096         trace_xfs_log_regrant_reserve_exit(log, ticket);
3097
3098         ticket->t_curr_res = ticket->t_unit_res;
3099         xlog_tic_reset_res(ticket);
3100 }       /* xlog_regrant_reserve_log_space */
3101
3102
3103 /*
3104  * Give back the space left from a reservation.
3105  *
3106  * All the information we need to make a correct determination of space left
3107  * is present.  For non-permanent reservations, things are quite easy.  The
3108  * count should have been decremented to zero.  We only need to deal with the
3109  * space remaining in the current reservation part of the ticket.  If the
3110  * ticket contains a permanent reservation, there may be left over space which
3111  * needs to be released.  A count of N means that N-1 refills of the current
3112  * reservation can be done before we need to ask for more space.  The first
3113  * one goes to fill up the first current reservation.  Once we run out of
3114  * space, the count will stay at zero and the only space remaining will be
3115  * in the current reservation field.
3116  */
3117 STATIC void
3118 xlog_ungrant_log_space(
3119         struct xlog             *log,
3120         struct xlog_ticket      *ticket)
3121 {
3122         int     bytes;
3123
3124         if (ticket->t_cnt > 0)
3125                 ticket->t_cnt--;
3126
3127         trace_xfs_log_ungrant_enter(log, ticket);
3128         trace_xfs_log_ungrant_sub(log, ticket);
3129
3130         /*
3131          * If this is a permanent reservation ticket, we may be able to free
3132          * up more space based on the remaining count.
3133          */
3134         bytes = ticket->t_curr_res;
3135         if (ticket->t_cnt > 0) {
3136                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3137                 bytes += ticket->t_unit_res*ticket->t_cnt;
3138         }
3139
3140         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3141         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3142
3143         trace_xfs_log_ungrant_exit(log, ticket);
3144
3145         xfs_log_space_wake(log->l_mp);
3146 }
3147
3148 /*
3149  * Flush iclog to disk if this is the last reference to the given iclog and
3150  * the WANT_SYNC bit is set.
3151  *
3152  * When this function is entered, the iclog is not necessarily in the
3153  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3154  *
3155  *
3156  */
3157 STATIC int
3158 xlog_state_release_iclog(
3159         struct xlog             *log,
3160         struct xlog_in_core     *iclog)
3161 {
3162         int             sync = 0;       /* do we sync? */
3163
3164         if (iclog->ic_state & XLOG_STATE_IOERROR)
3165                 return -EIO;
3166
3167         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3168         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3169                 return 0;
3170
3171         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3172                 spin_unlock(&log->l_icloglock);
3173                 return -EIO;
3174         }
3175         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3176                iclog->ic_state == XLOG_STATE_WANT_SYNC);
3177
3178         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3179                 /* update tail before writing to iclog */
3180                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3181                 sync++;
3182                 iclog->ic_state = XLOG_STATE_SYNCING;
3183                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3184                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3185                 /* cycle incremented when incrementing curr_block */
3186         }
3187         spin_unlock(&log->l_icloglock);
3188
3189         /*
3190          * We let the log lock go, so it's possible that we hit a log I/O
3191          * error or some other SHUTDOWN condition that marks the iclog
3192          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3193          * this iclog has consistent data, so we ignore IOERROR
3194          * flags after this point.
3195          */
3196         if (sync)
3197                 return xlog_sync(log, iclog);
3198         return 0;
3199 }       /* xlog_state_release_iclog */
3200
3201
3202 /*
3203  * This routine will mark the current iclog in the ring as WANT_SYNC
3204  * and move the current iclog pointer to the next iclog in the ring.
3205  * When this routine is called from xlog_state_get_iclog_space(), the
3206  * exact size of the iclog has not yet been determined.  All we know is
3207  * that every data block.  We have run out of space in this log record.
3208  */
3209 STATIC void
3210 xlog_state_switch_iclogs(
3211         struct xlog             *log,
3212         struct xlog_in_core     *iclog,
3213         int                     eventual_size)
3214 {
3215         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3216         if (!eventual_size)
3217                 eventual_size = iclog->ic_offset;
3218         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3219         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3220         log->l_prev_block = log->l_curr_block;
3221         log->l_prev_cycle = log->l_curr_cycle;
3222
3223         /* roll log?: ic_offset changed later */
3224         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3225
3226         /* Round up to next log-sunit */
3227         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3228             log->l_mp->m_sb.sb_logsunit > 1) {
3229                 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3230                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3231         }
3232
3233         if (log->l_curr_block >= log->l_logBBsize) {
3234                 /*
3235                  * Rewind the current block before the cycle is bumped to make
3236                  * sure that the combined LSN never transiently moves forward
3237                  * when the log wraps to the next cycle. This is to support the
3238                  * unlocked sample of these fields from xlog_valid_lsn(). Most
3239                  * other cases should acquire l_icloglock.
3240                  */
3241                 log->l_curr_block -= log->l_logBBsize;
3242                 ASSERT(log->l_curr_block >= 0);
3243                 smp_wmb();
3244                 log->l_curr_cycle++;
3245                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3246                         log->l_curr_cycle++;
3247         }
3248         ASSERT(iclog == log->l_iclog);
3249         log->l_iclog = iclog->ic_next;
3250 }       /* xlog_state_switch_iclogs */
3251
3252 /*
3253  * Write out all data in the in-core log as of this exact moment in time.
3254  *
3255  * Data may be written to the in-core log during this call.  However,
3256  * we don't guarantee this data will be written out.  A change from past
3257  * implementation means this routine will *not* write out zero length LRs.
3258  *
3259  * Basically, we try and perform an intelligent scan of the in-core logs.
3260  * If we determine there is no flushable data, we just return.  There is no
3261  * flushable data if:
3262  *
3263  *      1. the current iclog is active and has no data; the previous iclog
3264  *              is in the active or dirty state.
3265  *      2. the current iclog is drity, and the previous iclog is in the
3266  *              active or dirty state.
3267  *
3268  * We may sleep if:
3269  *
3270  *      1. the current iclog is not in the active nor dirty state.
3271  *      2. the current iclog dirty, and the previous iclog is not in the
3272  *              active nor dirty state.
3273  *      3. the current iclog is active, and there is another thread writing
3274  *              to this particular iclog.
3275  *      4. a) the current iclog is active and has no other writers
3276  *         b) when we return from flushing out this iclog, it is still
3277  *              not in the active nor dirty state.
3278  */
3279 int
3280 xfs_log_force(
3281         struct xfs_mount        *mp,
3282         uint                    flags)
3283 {
3284         struct xlog             *log = mp->m_log;
3285         struct xlog_in_core     *iclog;
3286         xfs_lsn_t               lsn;
3287
3288         XFS_STATS_INC(mp, xs_log_force);
3289         trace_xfs_log_force(mp, 0, _RET_IP_);
3290
3291         xlog_cil_force(log);
3292
3293         spin_lock(&log->l_icloglock);
3294         iclog = log->l_iclog;
3295         if (iclog->ic_state & XLOG_STATE_IOERROR)
3296                 goto out_error;
3297
3298         if (iclog->ic_state == XLOG_STATE_DIRTY ||
3299             (iclog->ic_state == XLOG_STATE_ACTIVE &&
3300              atomic_read(&iclog->ic_refcnt) == 0 && iclog->ic_offset == 0)) {
3301                 /*
3302                  * If the head is dirty or (active and empty), then we need to
3303                  * look at the previous iclog.
3304                  *
3305                  * If the previous iclog is active or dirty we are done.  There
3306                  * is nothing to sync out. Otherwise, we attach ourselves to the
3307                  * previous iclog and go to sleep.
3308                  */
3309                 iclog = iclog->ic_prev;
3310                 if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3311                     iclog->ic_state == XLOG_STATE_DIRTY)
3312                         goto out_unlock;
3313         } else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3314                 if (atomic_read(&iclog->ic_refcnt) == 0) {
3315                         /*
3316                          * We are the only one with access to this iclog.
3317                          *
3318                          * Flush it out now.  There should be a roundoff of zero
3319                          * to show that someone has already taken care of the
3320                          * roundoff from the previous sync.
3321                          */
3322                         atomic_inc(&iclog->ic_refcnt);
3323                         lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3324                         xlog_state_switch_iclogs(log, iclog, 0);
3325                         spin_unlock(&log->l_icloglock);
3326
3327                         if (xlog_state_release_iclog(log, iclog))
3328                                 return -EIO;
3329
3330                         spin_lock(&log->l_icloglock);
3331                         if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn ||
3332                             iclog->ic_state == XLOG_STATE_DIRTY)
3333                                 goto out_unlock;
3334                 } else {
3335                         /*
3336                          * Someone else is writing to this iclog.
3337                          *
3338                          * Use its call to flush out the data.  However, the
3339                          * other thread may not force out this LR, so we mark
3340                          * it WANT_SYNC.
3341                          */
3342                         xlog_state_switch_iclogs(log, iclog, 0);
3343                 }
3344         } else {
3345                 /*
3346                  * If the head iclog is not active nor dirty, we just attach
3347                  * ourselves to the head and go to sleep if necessary.
3348                  */
3349                 ;
3350         }
3351
3352         if (!(flags & XFS_LOG_SYNC))
3353                 goto out_unlock;
3354
3355         if (iclog->ic_state & XLOG_STATE_IOERROR)
3356                 goto out_error;
3357         XFS_STATS_INC(mp, xs_log_force_sleep);
3358         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3359         if (iclog->ic_state & XLOG_STATE_IOERROR)
3360                 return -EIO;
3361         return 0;
3362
3363 out_unlock:
3364         spin_unlock(&log->l_icloglock);
3365         return 0;
3366 out_error:
3367         spin_unlock(&log->l_icloglock);
3368         return -EIO;
3369 }
3370
3371 static int
3372 __xfs_log_force_lsn(
3373         struct xfs_mount        *mp,
3374         xfs_lsn_t               lsn,
3375         uint                    flags,
3376         int                     *log_flushed,
3377         bool                    already_slept)
3378 {
3379         struct xlog             *log = mp->m_log;
3380         struct xlog_in_core     *iclog;
3381
3382         spin_lock(&log->l_icloglock);
3383         iclog = log->l_iclog;
3384         if (iclog->ic_state & XLOG_STATE_IOERROR)
3385                 goto out_error;
3386
3387         while (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3388                 iclog = iclog->ic_next;
3389                 if (iclog == log->l_iclog)
3390                         goto out_unlock;
3391         }
3392
3393         if (iclog->ic_state == XLOG_STATE_DIRTY)
3394                 goto out_unlock;
3395
3396         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3397                 /*
3398                  * We sleep here if we haven't already slept (e.g. this is the
3399                  * first time we've looked at the correct iclog buf) and the
3400                  * buffer before us is going to be sync'ed.  The reason for this
3401                  * is that if we are doing sync transactions here, by waiting
3402                  * for the previous I/O to complete, we can allow a few more
3403                  * transactions into this iclog before we close it down.
3404                  *
3405                  * Otherwise, we mark the buffer WANT_SYNC, and bump up the
3406                  * refcnt so we can release the log (which drops the ref count).
3407                  * The state switch keeps new transaction commits from using
3408                  * this buffer.  When the current commits finish writing into
3409                  * the buffer, the refcount will drop to zero and the buffer
3410                  * will go out then.
3411                  */
3412                 if (!already_slept &&
3413                     (iclog->ic_prev->ic_state &
3414                      (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3415                         ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3416
3417                         XFS_STATS_INC(mp, xs_log_force_sleep);
3418
3419                         xlog_wait(&iclog->ic_prev->ic_write_wait,
3420                                         &log->l_icloglock);
3421                         return -EAGAIN;
3422                 }
3423                 atomic_inc(&iclog->ic_refcnt);
3424                 xlog_state_switch_iclogs(log, iclog, 0);
3425                 spin_unlock(&log->l_icloglock);
3426                 if (xlog_state_release_iclog(log, iclog))
3427                         return -EIO;
3428                 if (log_flushed)
3429                         *log_flushed = 1;
3430                 spin_lock(&log->l_icloglock);
3431         }
3432
3433         if (!(flags & XFS_LOG_SYNC) ||
3434             (iclog->ic_state & (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY)))
3435                 goto out_unlock;
3436
3437         if (iclog->ic_state & XLOG_STATE_IOERROR)
3438                 goto out_error;
3439
3440         XFS_STATS_INC(mp, xs_log_force_sleep);
3441         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3442         if (iclog->ic_state & XLOG_STATE_IOERROR)
3443                 return -EIO;
3444         return 0;
3445
3446 out_unlock:
3447         spin_unlock(&log->l_icloglock);
3448         return 0;
3449 out_error:
3450         spin_unlock(&log->l_icloglock);
3451         return -EIO;
3452 }
3453
3454 /*
3455  * Force the in-core log to disk for a specific LSN.
3456  *
3457  * Find in-core log with lsn.
3458  *      If it is in the DIRTY state, just return.
3459  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3460  *              state and go to sleep or return.
3461  *      If it is in any other state, go to sleep or return.
3462  *
3463  * Synchronous forces are implemented with a wait queue.  All callers trying
3464  * to force a given lsn to disk must wait on the queue attached to the
3465  * specific in-core log.  When given in-core log finally completes its write
3466  * to disk, that thread will wake up all threads waiting on the queue.
3467  */
3468 int
3469 xfs_log_force_lsn(
3470         struct xfs_mount        *mp,
3471         xfs_lsn_t               lsn,
3472         uint                    flags,
3473         int                     *log_flushed)
3474 {
3475         int                     ret;
3476         ASSERT(lsn != 0);
3477
3478         XFS_STATS_INC(mp, xs_log_force);
3479         trace_xfs_log_force(mp, lsn, _RET_IP_);
3480
3481         lsn = xlog_cil_force_lsn(mp->m_log, lsn);
3482         if (lsn == NULLCOMMITLSN)
3483                 return 0;
3484
3485         ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, false);
3486         if (ret == -EAGAIN)
3487                 ret = __xfs_log_force_lsn(mp, lsn, flags, log_flushed, true);
3488         return ret;
3489 }
3490
3491 /*
3492  * Called when we want to mark the current iclog as being ready to sync to
3493  * disk.
3494  */
3495 STATIC void
3496 xlog_state_want_sync(
3497         struct xlog             *log,
3498         struct xlog_in_core     *iclog)
3499 {
3500         assert_spin_locked(&log->l_icloglock);
3501
3502         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3503                 xlog_state_switch_iclogs(log, iclog, 0);
3504         } else {
3505                 ASSERT(iclog->ic_state &
3506                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3507         }
3508 }
3509
3510
3511 /*****************************************************************************
3512  *
3513  *              TICKET functions
3514  *
3515  *****************************************************************************
3516  */
3517
3518 /*
3519  * Free a used ticket when its refcount falls to zero.
3520  */
3521 void
3522 xfs_log_ticket_put(
3523         xlog_ticket_t   *ticket)
3524 {
3525         ASSERT(atomic_read(&ticket->t_ref) > 0);
3526         if (atomic_dec_and_test(&ticket->t_ref))
3527                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3528 }
3529
3530 xlog_ticket_t *
3531 xfs_log_ticket_get(
3532         xlog_ticket_t   *ticket)
3533 {
3534         ASSERT(atomic_read(&ticket->t_ref) > 0);
3535         atomic_inc(&ticket->t_ref);
3536         return ticket;
3537 }
3538
3539 /*
3540  * Figure out the total log space unit (in bytes) that would be
3541  * required for a log ticket.
3542  */
3543 int
3544 xfs_log_calc_unit_res(
3545         struct xfs_mount        *mp,
3546         int                     unit_bytes)
3547 {
3548         struct xlog             *log = mp->m_log;
3549         int                     iclog_space;
3550         uint                    num_headers;
3551
3552         /*
3553          * Permanent reservations have up to 'cnt'-1 active log operations
3554          * in the log.  A unit in this case is the amount of space for one
3555          * of these log operations.  Normal reservations have a cnt of 1
3556          * and their unit amount is the total amount of space required.
3557          *
3558          * The following lines of code account for non-transaction data
3559          * which occupy space in the on-disk log.
3560          *
3561          * Normal form of a transaction is:
3562          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3563          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3564          *
3565          * We need to account for all the leadup data and trailer data
3566          * around the transaction data.
3567          * And then we need to account for the worst case in terms of using
3568          * more space.
3569          * The worst case will happen if:
3570          * - the placement of the transaction happens to be such that the
3571          *   roundoff is at its maximum
3572          * - the transaction data is synced before the commit record is synced
3573          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3574          *   Therefore the commit record is in its own Log Record.
3575          *   This can happen as the commit record is called with its
3576          *   own region to xlog_write().
3577          *   This then means that in the worst case, roundoff can happen for
3578          *   the commit-rec as well.
3579          *   The commit-rec is smaller than padding in this scenario and so it is
3580          *   not added separately.
3581          */
3582
3583         /* for trans header */
3584         unit_bytes += sizeof(xlog_op_header_t);
3585         unit_bytes += sizeof(xfs_trans_header_t);
3586
3587         /* for start-rec */
3588         unit_bytes += sizeof(xlog_op_header_t);
3589
3590         /*
3591          * for LR headers - the space for data in an iclog is the size minus
3592          * the space used for the headers. If we use the iclog size, then we
3593          * undercalculate the number of headers required.
3594          *
3595          * Furthermore - the addition of op headers for split-recs might
3596          * increase the space required enough to require more log and op
3597          * headers, so take that into account too.
3598          *
3599          * IMPORTANT: This reservation makes the assumption that if this
3600          * transaction is the first in an iclog and hence has the LR headers
3601          * accounted to it, then the remaining space in the iclog is
3602          * exclusively for this transaction.  i.e. if the transaction is larger
3603          * than the iclog, it will be the only thing in that iclog.
3604          * Fundamentally, this means we must pass the entire log vector to
3605          * xlog_write to guarantee this.
3606          */
3607         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3608         num_headers = howmany(unit_bytes, iclog_space);
3609
3610         /* for split-recs - ophdrs added when data split over LRs */
3611         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3612
3613         /* add extra header reservations if we overrun */
3614         while (!num_headers ||
3615                howmany(unit_bytes, iclog_space) > num_headers) {
3616                 unit_bytes += sizeof(xlog_op_header_t);
3617                 num_headers++;
3618         }
3619         unit_bytes += log->l_iclog_hsize * num_headers;
3620
3621         /* for commit-rec LR header - note: padding will subsume the ophdr */
3622         unit_bytes += log->l_iclog_hsize;
3623
3624         /* for roundoff padding for transaction data and one for commit record */
3625         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3626                 /* log su roundoff */
3627                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3628         } else {
3629                 /* BB roundoff */
3630                 unit_bytes += 2 * BBSIZE;
3631         }
3632
3633         return unit_bytes;
3634 }
3635
3636 /*
3637  * Allocate and initialise a new log ticket.
3638  */
3639 struct xlog_ticket *
3640 xlog_ticket_alloc(
3641         struct xlog             *log,
3642         int                     unit_bytes,
3643         int                     cnt,
3644         char                    client,
3645         bool                    permanent,
3646         xfs_km_flags_t          alloc_flags)
3647 {
3648         struct xlog_ticket      *tic;
3649         int                     unit_res;
3650
3651         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3652         if (!tic)
3653                 return NULL;
3654
3655         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3656
3657         atomic_set(&tic->t_ref, 1);
3658         tic->t_task             = current;
3659         INIT_LIST_HEAD(&tic->t_queue);
3660         tic->t_unit_res         = unit_res;
3661         tic->t_curr_res         = unit_res;
3662         tic->t_cnt              = cnt;
3663         tic->t_ocnt             = cnt;
3664         tic->t_tid              = prandom_u32();
3665         tic->t_clientid         = client;
3666         tic->t_flags            = XLOG_TIC_INITED;
3667         if (permanent)
3668                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3669
3670         xlog_tic_reset_res(tic);
3671
3672         return tic;
3673 }
3674
3675
3676 /******************************************************************************
3677  *
3678  *              Log debug routines
3679  *
3680  ******************************************************************************
3681  */
3682 #if defined(DEBUG)
3683 /*
3684  * Make sure that the destination ptr is within the valid data region of
3685  * one of the iclogs.  This uses backup pointers stored in a different
3686  * part of the log in case we trash the log structure.
3687  */
3688 STATIC void
3689 xlog_verify_dest_ptr(
3690         struct xlog     *log,
3691         void            *ptr)
3692 {
3693         int i;
3694         int good_ptr = 0;
3695
3696         for (i = 0; i < log->l_iclog_bufs; i++) {
3697                 if (ptr >= log->l_iclog_bak[i] &&
3698                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3699                         good_ptr++;
3700         }
3701
3702         if (!good_ptr)
3703                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3704 }
3705
3706 /*
3707  * Check to make sure the grant write head didn't just over lap the tail.  If
3708  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3709  * the cycles differ by exactly one and check the byte count.
3710  *
3711  * This check is run unlocked, so can give false positives. Rather than assert
3712  * on failures, use a warn-once flag and a panic tag to allow the admin to
3713  * determine if they want to panic the machine when such an error occurs. For
3714  * debug kernels this will have the same effect as using an assert but, unlinke
3715  * an assert, it can be turned off at runtime.
3716  */
3717 STATIC void
3718 xlog_verify_grant_tail(
3719         struct xlog     *log)
3720 {
3721         int             tail_cycle, tail_blocks;
3722         int             cycle, space;
3723
3724         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3725         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3726         if (tail_cycle != cycle) {
3727                 if (cycle - 1 != tail_cycle &&
3728                     !(log->l_flags & XLOG_TAIL_WARN)) {
3729                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3730                                 "%s: cycle - 1 != tail_cycle", __func__);
3731                         log->l_flags |= XLOG_TAIL_WARN;
3732                 }
3733
3734                 if (space > BBTOB(tail_blocks) &&
3735                     !(log->l_flags & XLOG_TAIL_WARN)) {
3736                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3737                                 "%s: space > BBTOB(tail_blocks)", __func__);
3738                         log->l_flags |= XLOG_TAIL_WARN;
3739                 }
3740         }
3741 }
3742
3743 /* check if it will fit */
3744 STATIC void
3745 xlog_verify_tail_lsn(
3746         struct xlog             *log,
3747         struct xlog_in_core     *iclog,
3748         xfs_lsn_t               tail_lsn)
3749 {
3750     int blocks;
3751
3752     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3753         blocks =
3754             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3755         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3756                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3757     } else {
3758         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3759
3760         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3761                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3762
3763         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3764         if (blocks < BTOBB(iclog->ic_offset) + 1)
3765                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3766     }
3767 }       /* xlog_verify_tail_lsn */
3768
3769 /*
3770  * Perform a number of checks on the iclog before writing to disk.
3771  *
3772  * 1. Make sure the iclogs are still circular
3773  * 2. Make sure we have a good magic number
3774  * 3. Make sure we don't have magic numbers in the data
3775  * 4. Check fields of each log operation header for:
3776  *      A. Valid client identifier
3777  *      B. tid ptr value falls in valid ptr space (user space code)
3778  *      C. Length in log record header is correct according to the
3779  *              individual operation headers within record.
3780  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3781  *      log, check the preceding blocks of the physical log to make sure all
3782  *      the cycle numbers agree with the current cycle number.
3783  */
3784 STATIC void
3785 xlog_verify_iclog(
3786         struct xlog             *log,
3787         struct xlog_in_core     *iclog,
3788         int                     count,
3789         bool                    syncing)
3790 {
3791         xlog_op_header_t        *ophead;
3792         xlog_in_core_t          *icptr;
3793         xlog_in_core_2_t        *xhdr;
3794         void                    *base_ptr, *ptr, *p;
3795         ptrdiff_t               field_offset;
3796         uint8_t                 clientid;
3797         int                     len, i, j, k, op_len;
3798         int                     idx;
3799
3800         /* check validity of iclog pointers */
3801         spin_lock(&log->l_icloglock);
3802         icptr = log->l_iclog;
3803         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3804                 ASSERT(icptr);
3805
3806         if (icptr != log->l_iclog)
3807                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3808         spin_unlock(&log->l_icloglock);
3809
3810         /* check log magic numbers */
3811         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3812                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3813
3814         base_ptr = ptr = &iclog->ic_header;
3815         p = &iclog->ic_header;
3816         for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3817                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3818                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3819                                 __func__);
3820         }
3821
3822         /* check fields */
3823         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3824         base_ptr = ptr = iclog->ic_datap;
3825         ophead = ptr;
3826         xhdr = iclog->ic_data;
3827         for (i = 0; i < len; i++) {
3828                 ophead = ptr;
3829
3830                 /* clientid is only 1 byte */
3831                 p = &ophead->oh_clientid;
3832                 field_offset = p - base_ptr;
3833                 if (!syncing || (field_offset & 0x1ff)) {
3834                         clientid = ophead->oh_clientid;
3835                 } else {
3836                         idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3837                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3838                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3839                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3840                                 clientid = xlog_get_client_id(
3841                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3842                         } else {
3843                                 clientid = xlog_get_client_id(
3844                                         iclog->ic_header.h_cycle_data[idx]);
3845                         }
3846                 }
3847                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3848                         xfs_warn(log->l_mp,
3849                                 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3850                                 __func__, clientid, ophead,
3851                                 (unsigned long)field_offset);
3852
3853                 /* check length */
3854                 p = &ophead->oh_len;
3855                 field_offset = p - base_ptr;
3856                 if (!syncing || (field_offset & 0x1ff)) {
3857                         op_len = be32_to_cpu(ophead->oh_len);
3858                 } else {
3859                         idx = BTOBBT((uintptr_t)&ophead->oh_len -
3860                                     (uintptr_t)iclog->ic_datap);
3861                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3862                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3863                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3864                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3865                         } else {
3866                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3867                         }
3868                 }
3869                 ptr += sizeof(xlog_op_header_t) + op_len;
3870         }
3871 }       /* xlog_verify_iclog */
3872 #endif
3873
3874 /*
3875  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3876  */
3877 STATIC int
3878 xlog_state_ioerror(
3879         struct xlog     *log)
3880 {
3881         xlog_in_core_t  *iclog, *ic;
3882
3883         iclog = log->l_iclog;
3884         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3885                 /*
3886                  * Mark all the incore logs IOERROR.
3887                  * From now on, no log flushes will result.
3888                  */
3889                 ic = iclog;
3890                 do {
3891                         ic->ic_state = XLOG_STATE_IOERROR;
3892                         ic = ic->ic_next;
3893                 } while (ic != iclog);
3894                 return 0;
3895         }
3896         /*
3897          * Return non-zero, if state transition has already happened.
3898          */
3899         return 1;
3900 }
3901
3902 /*
3903  * This is called from xfs_force_shutdown, when we're forcibly
3904  * shutting down the filesystem, typically because of an IO error.
3905  * Our main objectives here are to make sure that:
3906  *      a. if !logerror, flush the logs to disk. Anything modified
3907  *         after this is ignored.
3908  *      b. the filesystem gets marked 'SHUTDOWN' for all interested
3909  *         parties to find out, 'atomically'.
3910  *      c. those who're sleeping on log reservations, pinned objects and
3911  *          other resources get woken up, and be told the bad news.
3912  *      d. nothing new gets queued up after (b) and (c) are done.
3913  *
3914  * Note: for the !logerror case we need to flush the regions held in memory out
3915  * to disk first. This needs to be done before the log is marked as shutdown,
3916  * otherwise the iclog writes will fail.
3917  */
3918 int
3919 xfs_log_force_umount(
3920         struct xfs_mount        *mp,
3921         int                     logerror)
3922 {
3923         struct xlog     *log;
3924         int             retval;
3925
3926         log = mp->m_log;
3927
3928         /*
3929          * If this happens during log recovery, don't worry about
3930          * locking; the log isn't open for business yet.
3931          */
3932         if (!log ||
3933             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3934                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3935                 if (mp->m_sb_bp)
3936                         mp->m_sb_bp->b_flags |= XBF_DONE;
3937                 return 0;
3938         }
3939
3940         /*
3941          * Somebody could've already done the hard work for us.
3942          * No need to get locks for this.
3943          */
3944         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3945                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3946                 return 1;
3947         }
3948
3949         /*
3950          * Flush all the completed transactions to disk before marking the log
3951          * being shut down. We need to do it in this order to ensure that
3952          * completed operations are safely on disk before we shut down, and that
3953          * we don't have to issue any buffer IO after the shutdown flags are set
3954          * to guarantee this.
3955          */
3956         if (!logerror)
3957                 xfs_log_force(mp, XFS_LOG_SYNC);
3958
3959         /*
3960          * mark the filesystem and the as in a shutdown state and wake
3961          * everybody up to tell them the bad news.
3962          */
3963         spin_lock(&log->l_icloglock);
3964         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3965         if (mp->m_sb_bp)
3966                 mp->m_sb_bp->b_flags |= XBF_DONE;
3967
3968         /*
3969          * Mark the log and the iclogs with IO error flags to prevent any
3970          * further log IO from being issued or completed.
3971          */
3972         log->l_flags |= XLOG_IO_ERROR;
3973         retval = xlog_state_ioerror(log);
3974         spin_unlock(&log->l_icloglock);
3975
3976         /*
3977          * We don't want anybody waiting for log reservations after this. That
3978          * means we have to wake up everybody queued up on reserveq as well as
3979          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3980          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3981          * action is protected by the grant locks.
3982          */
3983         xlog_grant_head_wake_all(&log->l_reserve_head);
3984         xlog_grant_head_wake_all(&log->l_write_head);
3985
3986         /*
3987          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3988          * as if the log writes were completed. The abort handling in the log
3989          * item committed callback functions will do this again under lock to
3990          * avoid races.
3991          */
3992         wake_up_all(&log->l_cilp->xc_commit_wait);
3993         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3994
3995 #ifdef XFSERRORDEBUG
3996         {
3997                 xlog_in_core_t  *iclog;
3998
3999                 spin_lock(&log->l_icloglock);
4000                 iclog = log->l_iclog;
4001                 do {
4002                         ASSERT(iclog->ic_callback == 0);
4003                         iclog = iclog->ic_next;
4004                 } while (iclog != log->l_iclog);
4005                 spin_unlock(&log->l_icloglock);
4006         }
4007 #endif
4008         /* return non-zero if log IOERROR transition had already happened */
4009         return retval;
4010 }
4011
4012 STATIC int
4013 xlog_iclogs_empty(
4014         struct xlog     *log)
4015 {
4016         xlog_in_core_t  *iclog;
4017
4018         iclog = log->l_iclog;
4019         do {
4020                 /* endianness does not matter here, zero is zero in
4021                  * any language.
4022                  */
4023                 if (iclog->ic_header.h_num_logops)
4024                         return 0;
4025                 iclog = iclog->ic_next;
4026         } while (iclog != log->l_iclog);
4027         return 1;
4028 }
4029
4030 /*
4031  * Verify that an LSN stamped into a piece of metadata is valid. This is
4032  * intended for use in read verifiers on v5 superblocks.
4033  */
4034 bool
4035 xfs_log_check_lsn(
4036         struct xfs_mount        *mp,
4037         xfs_lsn_t               lsn)
4038 {
4039         struct xlog             *log = mp->m_log;
4040         bool                    valid;
4041
4042         /*
4043          * norecovery mode skips mount-time log processing and unconditionally
4044          * resets the in-core LSN. We can't validate in this mode, but
4045          * modifications are not allowed anyways so just return true.
4046          */
4047         if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4048                 return true;
4049
4050         /*
4051          * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4052          * handled by recovery and thus safe to ignore here.
4053          */
4054         if (lsn == NULLCOMMITLSN)
4055                 return true;
4056
4057         valid = xlog_valid_lsn(mp->m_log, lsn);
4058
4059         /* warn the user about what's gone wrong before verifier failure */
4060         if (!valid) {
4061                 spin_lock(&log->l_icloglock);
4062                 xfs_warn(mp,
4063 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4064 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4065                          CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4066                          log->l_curr_cycle, log->l_curr_block);
4067                 spin_unlock(&log->l_icloglock);
4068         }
4069
4070         return valid;
4071 }
4072
4073 bool
4074 xfs_log_in_recovery(
4075         struct xfs_mount        *mp)
4076 {
4077         struct xlog             *log = mp->m_log;
4078
4079         return log->l_flags & XLOG_ACTIVE_RECOVERY;
4080 }