fs/xfs/xfs_log.c

   1 /*
   2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_types.h"
  21 #include "xfs_log.h"
  22 #include "xfs_trans.h"
  23 #include "xfs_sb.h"
  24 #include "xfs_ag.h"
  25 #include "xfs_mount.h"
  26 #include "xfs_error.h"
  27 #include "xfs_log_priv.h"
  28 #include "xfs_buf_item.h"
  29 #include "xfs_bmap_btree.h"
  30 #include "xfs_alloc_btree.h"
  31 #include "xfs_ialloc_btree.h"
  32 #include "xfs_log_recover.h"
  33 #include "xfs_trans_priv.h"
  34 #include "xfs_dinode.h"
  35 #include "xfs_inode.h"
  36 #include "xfs_trace.h"
  37 #include "xfs_fsops.h"
  38 #include "xfs_cksum.h"
  39
  40 kmem_zone_t     *xfs_log_ticket_zone;
  41
  42 /* Local miscellaneous function prototypes */
  43 STATIC int
  44 xlog_commit_record(
  45         struct xlog             *log,
  46         struct xlog_ticket      *ticket,
  47         struct xlog_in_core     **iclog,
  48         xfs_lsn_t               *commitlsnp);
  49
  50 STATIC struct xlog *
  51 xlog_alloc_log(
  52         struct xfs_mount        *mp,
  53         struct xfs_buftarg      *log_target,
  54         xfs_daddr_t             blk_offset,
  55         int                     num_bblks);
  56 STATIC int
  57 xlog_space_left(
  58         struct xlog             *log,
  59         atomic64_t              *head);
  60 STATIC int
  61 xlog_sync(
  62         struct xlog             *log,
  63         struct xlog_in_core     *iclog);
  64 STATIC void
  65 xlog_dealloc_log(
  66         struct xlog             *log);
  67
  68 /* local state machine functions */
  69 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  70 STATIC void
  71 xlog_state_do_callback(
  72         struct xlog             *log,
  73         int                     aborted,
  74         struct xlog_in_core     *iclog);
  75 STATIC int
  76 xlog_state_get_iclog_space(
  77         struct xlog             *log,
  78         int                     len,
  79         struct xlog_in_core     **iclog,
  80         struct xlog_ticket      *ticket,
  81         int                     *continued_write,
  82         int                     *logoffsetp);
  83 STATIC int
  84 xlog_state_release_iclog(
  85         struct xlog             *log,
  86         struct xlog_in_core     *iclog);
  87 STATIC void
  88 xlog_state_switch_iclogs(
  89         struct xlog             *log,
  90         struct xlog_in_core     *iclog,
  91         int                     eventual_size);
  92 STATIC void
  93 xlog_state_want_sync(
  94         struct xlog             *log,
  95         struct xlog_in_core     *iclog);
  96
  97 STATIC void
  98 xlog_grant_push_ail(
  99         struct xlog             *log,
 100         int                     need_bytes);
 101 STATIC void
 102 xlog_regrant_reserve_log_space(
 103         struct xlog             *log,
 104         struct xlog_ticket      *ticket);
 105 STATIC void
 106 xlog_ungrant_log_space(
 107         struct xlog             *log,
 108         struct xlog_ticket      *ticket);
 109
 110 #if defined(DEBUG)
 111 STATIC void
 112 xlog_verify_dest_ptr(
 113         struct xlog             *log,
 114         char                    *ptr);
 115 STATIC void
 116 xlog_verify_grant_tail(
 117         struct xlog *log);
 118 STATIC void
 119 xlog_verify_iclog(
 120         struct xlog             *log,
 121         struct xlog_in_core     *iclog,
 122         int                     count,
 123         bool                    syncing);
 124 STATIC void
 125 xlog_verify_tail_lsn(
 126         struct xlog             *log,
 127         struct xlog_in_core     *iclog,
 128         xfs_lsn_t               tail_lsn);
 129 #else
 130 #define xlog_verify_dest_ptr(a,b)
 131 #define xlog_verify_grant_tail(a)
 132 #define xlog_verify_iclog(a,b,c,d)
 133 #define xlog_verify_tail_lsn(a,b,c)
 134 #endif
 135
 136 STATIC int
 137 xlog_iclogs_empty(
 138         struct xlog             *log);
 139
 140 static void
 141 xlog_grant_sub_space(
 142         struct xlog             *log,
 143         atomic64_t              *head,
 144         int                     bytes)
 145 {
 146         int64_t head_val = atomic64_read(head);
 147         int64_t new, old;
 148
 149         do {
 150                 int     cycle, space;
 151
 152                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 153
 154                 space -= bytes;
 155                 if (space < 0) {
 156                         space += log->l_logsize;
 157                         cycle--;
 158                 }
 159
 160                 old = head_val;
 161                 new = xlog_assign_grant_head_val(cycle, space);
 162                 head_val = atomic64_cmpxchg(head, old, new);
 163         } while (head_val != old);
 164 }
 165
 166 static void
 167 xlog_grant_add_space(
 168         struct xlog             *log,
 169         atomic64_t              *head,
 170         int                     bytes)
 171 {
 172         int64_t head_val = atomic64_read(head);
 173         int64_t new, old;
 174
 175         do {
 176                 int             tmp;
 177                 int             cycle, space;
 178
 179                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 180
 181                 tmp = log->l_logsize - space;
 182                 if (tmp > bytes)
 183                         space += bytes;
 184                 else {
 185                         space = bytes - tmp;
 186                         cycle++;
 187                 }
 188
 189                 old = head_val;
 190                 new = xlog_assign_grant_head_val(cycle, space);
 191                 head_val = atomic64_cmpxchg(head, old, new);
 192         } while (head_val != old);
 193 }
 194
 195 STATIC void
 196 xlog_grant_head_init(
 197         struct xlog_grant_head  *head)
 198 {
 199         xlog_assign_grant_head(&head->grant, 1, 0);
 200         INIT_LIST_HEAD(&head->waiters);
 201         spin_lock_init(&head->lock);
 202 }
 203
 204 STATIC void
 205 xlog_grant_head_wake_all(
 206         struct xlog_grant_head  *head)
 207 {
 208         struct xlog_ticket      *tic;
 209
 210         spin_lock(&head->lock);
 211         list_for_each_entry(tic, &head->waiters, t_queue)
 212                 wake_up_process(tic->t_task);
 213         spin_unlock(&head->lock);
 214 }
 215
 216 static inline int
 217 xlog_ticket_reservation(
 218         struct xlog             *log,
 219         struct xlog_grant_head  *head,
 220         struct xlog_ticket      *tic)
 221 {
 222         if (head == &log->l_write_head) {
 223                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 224                 return tic->t_unit_res;
 225         } else {
 226                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 227                         return tic->t_unit_res * tic->t_cnt;
 228                 else
 229                         return tic->t_unit_res;
 230         }
 231 }
 232
 233 STATIC bool
 234 xlog_grant_head_wake(
 235         struct xlog             *log,
 236         struct xlog_grant_head  *head,
 237         int                     *free_bytes)
 238 {
 239         struct xlog_ticket      *tic;
 240         int                     need_bytes;
 241
 242         list_for_each_entry(tic, &head->waiters, t_queue) {
 243                 need_bytes = xlog_ticket_reservation(log, head, tic);
 244                 if (*free_bytes < need_bytes)
 245                         return false;
 246
 247                 *free_bytes -= need_bytes;
 248                 trace_xfs_log_grant_wake_up(log, tic);
 249                 wake_up_process(tic->t_task);
 250         }
 251
 252         return true;
 253 }
 254
 255 STATIC int
 256 xlog_grant_head_wait(
 257         struct xlog             *log,
 258         struct xlog_grant_head  *head,
 259         struct xlog_ticket      *tic,
 260         int                     need_bytes)
 261 {
 262         list_add_tail(&tic->t_queue, &head->waiters);
 263
 264         do {
 265                 if (XLOG_FORCED_SHUTDOWN(log))
 266                         goto shutdown;
 267                 xlog_grant_push_ail(log, need_bytes);
 268
 269                 __set_current_state(TASK_UNINTERRUPTIBLE);
 270                 spin_unlock(&head->lock);
 271
 272                 XFS_STATS_INC(xs_sleep_logspace);
 273
 274                 trace_xfs_log_grant_sleep(log, tic);
 275                 schedule();
 276                 trace_xfs_log_grant_wake(log, tic);
 277
 278                 spin_lock(&head->lock);
 279                 if (XLOG_FORCED_SHUTDOWN(log))
 280                         goto shutdown;
 281         } while (xlog_space_left(log, &head->grant) < need_bytes);
 282
 283         list_del_init(&tic->t_queue);
 284         return 0;
 285 shutdown:
 286         list_del_init(&tic->t_queue);
 287         return XFS_ERROR(EIO);
 288 }
 289
 290 /*
 291  * Atomically get the log space required for a log ticket.
 292  *
 293  * Once a ticket gets put onto head->waiters, it will only return after the
 294  * needed reservation is satisfied.
 295  *
 296  * This function is structured so that it has a lock free fast path. This is
 297  * necessary because every new transaction reservation will come through this
 298  * path. Hence any lock will be globally hot if we take it unconditionally on
 299  * every pass.
 300  *
 301  * As tickets are only ever moved on and off head->waiters under head->lock, we
 302  * only need to take that lock if we are going to add the ticket to the queue
 303  * and sleep. We can avoid taking the lock if the ticket was never added to
 304  * head->waiters because the t_queue list head will be empty and we hold the
 305  * only reference to it so it can safely be checked unlocked.
 306  */
 307 STATIC int
 308 xlog_grant_head_check(
 309         struct xlog             *log,
 310         struct xlog_grant_head  *head,
 311         struct xlog_ticket      *tic,
 312         int                     *need_bytes)
 313 {
 314         int                     free_bytes;
 315         int                     error = 0;
 316
 317         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 318
 319         /*
 320          * If there are other waiters on the queue then give them a chance at
 321          * logspace before us.  Wake up the first waiters, if we do not wake
 322          * up all the waiters then go to sleep waiting for more free space,
 323          * otherwise try to get some space for this transaction.
 324          */
 325         *need_bytes = xlog_ticket_reservation(log, head, tic);
 326         free_bytes = xlog_space_left(log, &head->grant);
 327         if (!list_empty_careful(&head->waiters)) {
 328                 spin_lock(&head->lock);
 329                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 330                     free_bytes < *need_bytes) {
 331                         error = xlog_grant_head_wait(log, head, tic,
 332                                                      *need_bytes);
 333                 }
 334                 spin_unlock(&head->lock);
 335         } else if (free_bytes < *need_bytes) {
 336                 spin_lock(&head->lock);
 337                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 338                 spin_unlock(&head->lock);
 339         }
 340
 341         return error;
 342 }
 343
 344 static void
 345 xlog_tic_reset_res(xlog_ticket_t *tic)
 346 {
 347         tic->t_res_num = 0;
 348         tic->t_res_arr_sum = 0;
 349         tic->t_res_num_ophdrs = 0;
 350 }
 351
 352 static void
 353 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 354 {
 355         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 356                 /* add to overflow and start again */
 357                 tic->t_res_o_flow += tic->t_res_arr_sum;
 358                 tic->t_res_num = 0;
 359                 tic->t_res_arr_sum = 0;
 360         }
 361
 362         tic->t_res_arr[tic->t_res_num].r_len = len;
 363         tic->t_res_arr[tic->t_res_num].r_type = type;
 364         tic->t_res_arr_sum += len;
 365         tic->t_res_num++;
 366 }
 367
 368 /*
 369  * Replenish the byte reservation required by moving the grant write head.
 370  */
 371 int
 372 xfs_log_regrant(
 373         struct xfs_mount        *mp,
 374         struct xlog_ticket      *tic)
 375 {
 376         struct xlog             *log = mp->m_log;
 377         int                     need_bytes;
 378         int                     error = 0;
 379
 380         if (XLOG_FORCED_SHUTDOWN(log))
 381                 return XFS_ERROR(EIO);
 382
 383         XFS_STATS_INC(xs_try_logspace);
 384
 385         /*
 386          * This is a new transaction on the ticket, so we need to change the
 387          * transaction ID so that the next transaction has a different TID in
 388          * the log. Just add one to the existing tid so that we can see chains
 389          * of rolling transactions in the log easily.
 390          */
 391         tic->t_tid++;
 392
 393         xlog_grant_push_ail(log, tic->t_unit_res);
 394
 395         tic->t_curr_res = tic->t_unit_res;
 396         xlog_tic_reset_res(tic);
 397
 398         if (tic->t_cnt > 0)
 399                 return 0;
 400
 401         trace_xfs_log_regrant(log, tic);
 402
 403         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 404                                       &need_bytes);
 405         if (error)
 406                 goto out_error;
 407
 408         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 409         trace_xfs_log_regrant_exit(log, tic);
 410         xlog_verify_grant_tail(log);
 411         return 0;
 412
 413 out_error:
 414         /*
 415          * If we are failing, make sure the ticket doesn't have any current
 416          * reservations.  We don't want to add this back when the ticket/
 417          * transaction gets cancelled.
 418          */
 419         tic->t_curr_res = 0;
 420         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 421         return error;
 422 }
 423
 424 /*
 425  * Reserve log space and return a ticket corresponding the reservation.
 426  *
 427  * Each reservation is going to reserve extra space for a log record header.
 428  * When writes happen to the on-disk log, we don't subtract the length of the
 429  * log record header from any reservation.  By wasting space in each
 430  * reservation, we prevent over allocation problems.
 431  */
 432 int
 433 xfs_log_reserve(
 434         struct xfs_mount        *mp,
 435         int                     unit_bytes,
 436         int                     cnt,
 437         struct xlog_ticket      **ticp,
 438         __uint8_t               client,
 439         bool                    permanent,
 440         uint                    t_type)
 441 {
 442         struct xlog             *log = mp->m_log;
 443         struct xlog_ticket      *tic;
 444         int                     need_bytes;
 445         int                     error = 0;
 446
 447         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 448
 449         if (XLOG_FORCED_SHUTDOWN(log))
 450                 return XFS_ERROR(EIO);
 451
 452         XFS_STATS_INC(xs_try_logspace);
 453
 454         ASSERT(*ticp == NULL);
 455         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 456                                 KM_SLEEP | KM_MAYFAIL);
 457         if (!tic)
 458                 return XFS_ERROR(ENOMEM);
 459
 460         tic->t_trans_type = t_type;
 461         *ticp = tic;
 462
 463         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 464                                             : tic->t_unit_res);
 465
 466         trace_xfs_log_reserve(log, tic);
 467
 468         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 469                                       &need_bytes);
 470         if (error)
 471                 goto out_error;
 472
 473         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 474         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 475         trace_xfs_log_reserve_exit(log, tic);
 476         xlog_verify_grant_tail(log);
 477         return 0;
 478
 479 out_error:
 480         /*
 481          * If we are failing, make sure the ticket doesn't have any current
 482          * reservations.  We don't want to add this back when the ticket/
 483          * transaction gets cancelled.
 484          */
 485         tic->t_curr_res = 0;
 486         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 487         return error;
 488 }
 489
 490
 491 /*
 492  * NOTES:
 493  *
 494  *      1. currblock field gets updated at startup and after in-core logs
 495  *              marked as with WANT_SYNC.
 496  */
 497
 498 /*
 499  * This routine is called when a user of a log manager ticket is done with
 500  * the reservation.  If the ticket was ever used, then a commit record for
 501  * the associated transaction is written out as a log operation header with
 502  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 503  * a given ticket.  If the ticket was one with a permanent reservation, then
 504  * a few operations are done differently.  Permanent reservation tickets by
 505  * default don't release the reservation.  They just commit the current
 506  * transaction with the belief that the reservation is still needed.  A flag
 507  * must be passed in before permanent reservations are actually released.
 508  * When these type of tickets are not released, they need to be set into
 509  * the inited state again.  By doing this, a start record will be written
 510  * out when the next write occurs.
 511  */
 512 xfs_lsn_t
 513 xfs_log_done(
 514         struct xfs_mount        *mp,
 515         struct xlog_ticket      *ticket,
 516         struct xlog_in_core     **iclog,
 517         uint                    flags)
 518 {
 519         struct xlog             *log = mp->m_log;
 520         xfs_lsn_t               lsn = 0;
 521
 522         if (XLOG_FORCED_SHUTDOWN(log) ||
 523             /*
 524              * If nothing was ever written, don't write out commit record.
 525              * If we get an error, just continue and give back the log ticket.
 526              */
 527             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 528              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 529                 lsn = (xfs_lsn_t) -1;
 530                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
 531                         flags |= XFS_LOG_REL_PERM_RESERV;
 532                 }
 533         }
 534
 535
 536         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
 537             (flags & XFS_LOG_REL_PERM_RESERV)) {
 538                 trace_xfs_log_done_nonperm(log, ticket);
 539
 540                 /*
 541                  * Release ticket if not permanent reservation or a specific
 542                  * request has been made to release a permanent reservation.
 543                  */
 544                 xlog_ungrant_log_space(log, ticket);
 545                 xfs_log_ticket_put(ticket);
 546         } else {
 547                 trace_xfs_log_done_perm(log, ticket);
 548
 549                 xlog_regrant_reserve_log_space(log, ticket);
 550                 /* If this ticket was a permanent reservation and we aren't
 551                  * trying to release it, reset the inited flags; so next time
 552                  * we write, a start record will be written out.
 553                  */
 554                 ticket->t_flags |= XLOG_TIC_INITED;
 555         }
 556
 557         return lsn;
 558 }
 559
 560 /*
 561  * Attaches a new iclog I/O completion callback routine during
 562  * transaction commit.  If the log is in error state, a non-zero
 563  * return code is handed back and the caller is responsible for
 564  * executing the callback at an appropriate time.
 565  */
 566 int
 567 xfs_log_notify(
 568         struct xfs_mount        *mp,
 569         struct xlog_in_core     *iclog,
 570         xfs_log_callback_t      *cb)
 571 {
 572         int     abortflg;
 573
 574         spin_lock(&iclog->ic_callback_lock);
 575         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 576         if (!abortflg) {
 577                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 578                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 579                 cb->cb_next = NULL;
 580                 *(iclog->ic_callback_tail) = cb;
 581                 iclog->ic_callback_tail = &(cb->cb_next);
 582         }
 583         spin_unlock(&iclog->ic_callback_lock);
 584         return abortflg;
 585 }
 586
 587 int
 588 xfs_log_release_iclog(
 589         struct xfs_mount        *mp,
 590         struct xlog_in_core     *iclog)
 591 {
 592         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 593                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 594                 return EIO;
 595         }
 596
 597         return 0;
 598 }
 599
 600 /*
 601  * Mount a log filesystem
 602  *
 603  * mp           - ubiquitous xfs mount point structure
 604  * log_target   - buftarg of on-disk log device
 605  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 606  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 607  *
 608  * Return error or zero.
 609  */
 610 int
 611 xfs_log_mount(
 612         xfs_mount_t     *mp,
 613         xfs_buftarg_t   *log_target,
 614         xfs_daddr_t     blk_offset,
 615         int             num_bblks)
 616 {
 617         int             error;
 618
 619         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
 620                 xfs_notice(mp, "Mounting Filesystem");
 621         else {
 622                 xfs_notice(mp,
 623 "Mounting filesystem in no-recovery mode.  Filesystem will be inconsistent.");
 624                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 625         }
 626
 627         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 628         if (IS_ERR(mp->m_log)) {
 629                 error = -PTR_ERR(mp->m_log);
 630                 goto out;
 631         }
 632
 633         /*
 634          * Initialize the AIL now we have a log.
 635          */
 636         error = xfs_trans_ail_init(mp);
 637         if (error) {
 638                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 639                 goto out_free_log;
 640         }
 641         mp->m_log->l_ailp = mp->m_ail;
 642
 643         /*
 644          * skip log recovery on a norecovery mount.  pretend it all
 645          * just worked.
 646          */
 647         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 648                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 649
 650                 if (readonly)
 651                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 652
 653                 error = xlog_recover(mp->m_log);
 654
 655                 if (readonly)
 656                         mp->m_flags |= XFS_MOUNT_RDONLY;
 657                 if (error) {
 658                         xfs_warn(mp, "log mount/recovery failed: error %d",
 659                                 error);
 660                         goto out_destroy_ail;
 661                 }
 662         }
 663
 664         /* Normal transactions can now occur */
 665         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 666
 667         /*
 668          * Now the log has been fully initialised and we know were our
 669          * space grant counters are, we can initialise the permanent ticket
 670          * needed for delayed logging to work.
 671          */
 672         xlog_cil_init_post_recovery(mp->m_log);
 673
 674         return 0;
 675
 676 out_destroy_ail:
 677         xfs_trans_ail_destroy(mp);
 678 out_free_log:
 679         xlog_dealloc_log(mp->m_log);
 680 out:
 681         return error;
 682 }
 683
 684 /*
 685  * Finish the recovery of the file system.  This is separate from the
 686  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 687  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 688  * here.
 689  *
 690  * If we finish recovery successfully, start the background log work. If we are
 691  * not doing recovery, then we have a RO filesystem and we don't need to start
 692  * it.
 693  */
 694 int
 695 xfs_log_mount_finish(xfs_mount_t *mp)
 696 {
 697         int     error = 0;
 698
 699         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 700                 error = xlog_recover_finish(mp->m_log);
 701                 if (!error)
 702                         xfs_log_work_queue(mp);
 703         } else {
 704                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 705         }
 706
 707
 708         return error;
 709 }
 710
 711 /*
 712  * Final log writes as part of unmount.
 713  *
 714  * Mark the filesystem clean as unmount happens.  Note that during relocation
 715  * this routine needs to be executed as part of source-bag while the
 716  * deallocation must not be done until source-end.
 717  */
 718
 719 /*
 720  * Unmount record used to have a string "Unmount filesystem--" in the
 721  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 722  * We just write the magic number now since that particular field isn't
 723  * currently architecture converted and "nUmount" is a bit foo.
 724  * As far as I know, there weren't any dependencies on the old behaviour.
 725  */
 726
 727 int
 728 xfs_log_unmount_write(xfs_mount_t *mp)
 729 {
 730         struct xlog      *log = mp->m_log;
 731         xlog_in_core_t   *iclog;
 732 #ifdef DEBUG
 733         xlog_in_core_t   *first_iclog;
 734 #endif
 735         xlog_ticket_t   *tic = NULL;
 736         xfs_lsn_t        lsn;
 737         int              error;
 738
 739         /*
 740          * Don't write out unmount record on read-only mounts.
 741          * Or, if we are doing a forced umount (typically because of IO errors).
 742          */
 743         if (mp->m_flags & XFS_MOUNT_RDONLY)
 744                 return 0;
 745
 746         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
 747         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 748
 749 #ifdef DEBUG
 750         first_iclog = iclog = log->l_iclog;
 751         do {
 752                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 753                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 754                         ASSERT(iclog->ic_offset == 0);
 755                 }
 756                 iclog = iclog->ic_next;
 757         } while (iclog != first_iclog);
 758 #endif
 759         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 760                 error = xfs_log_reserve(mp, 600, 1, &tic,
 761                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
 762                 if (!error) {
 763                         /* the data section must be 32 bit size aligned */
 764                         struct {
 765                             __uint16_t magic;
 766                             __uint16_t pad1;
 767                             __uint32_t pad2; /* may as well make it 64 bits */
 768                         } magic = {
 769                                 .magic = XLOG_UNMOUNT_TYPE,
 770                         };
 771                         struct xfs_log_iovec reg = {
 772                                 .i_addr = &magic,
 773                                 .i_len = sizeof(magic),
 774                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 775                         };
 776                         struct xfs_log_vec vec = {
 777                                 .lv_niovecs = 1,
 778                                 .lv_iovecp = &reg,
 779                         };
 780
 781                         /* remove inited flag, and account for space used */
 782                         tic->t_flags = 0;
 783                         tic->t_curr_res -= sizeof(magic);
 784                         error = xlog_write(log, &vec, tic, &lsn,
 785                                            NULL, XLOG_UNMOUNT_TRANS);
 786                         /*
 787                          * At this point, we're umounting anyway,
 788                          * so there's no point in transitioning log state
 789                          * to IOERROR. Just continue...
 790                          */
 791                 }
 792
 793                 if (error)
 794                         xfs_alert(mp, "%s: unmount record failed", __func__);
 795
 796
 797                 spin_lock(&log->l_icloglock);
 798                 iclog = log->l_iclog;
 799                 atomic_inc(&iclog->ic_refcnt);
 800                 xlog_state_want_sync(log, iclog);
 801                 spin_unlock(&log->l_icloglock);
 802                 error = xlog_state_release_iclog(log, iclog);
 803
 804                 spin_lock(&log->l_icloglock);
 805                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 806                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 807                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 808                                 xlog_wait(&iclog->ic_force_wait,
 809                                                         &log->l_icloglock);
 810                         } else {
 811                                 spin_unlock(&log->l_icloglock);
 812                         }
 813                 } else {
 814                         spin_unlock(&log->l_icloglock);
 815                 }
 816                 if (tic) {
 817                         trace_xfs_log_umount_write(log, tic);
 818                         xlog_ungrant_log_space(log, tic);
 819                         xfs_log_ticket_put(tic);
 820                 }
 821         } else {
 822                 /*
 823                  * We're already in forced_shutdown mode, couldn't
 824                  * even attempt to write out the unmount transaction.
 825                  *
 826                  * Go through the motions of sync'ing and releasing
 827                  * the iclog, even though no I/O will actually happen,
 828                  * we need to wait for other log I/Os that may already
 829                  * be in progress.  Do this as a separate section of
 830                  * code so we'll know if we ever get stuck here that
 831                  * we're in this odd situation of trying to unmount
 832                  * a file system that went into forced_shutdown as
 833                  * the result of an unmount..
 834                  */
 835                 spin_lock(&log->l_icloglock);
 836                 iclog = log->l_iclog;
 837                 atomic_inc(&iclog->ic_refcnt);
 838
 839                 xlog_state_want_sync(log, iclog);
 840                 spin_unlock(&log->l_icloglock);
 841                 error =  xlog_state_release_iclog(log, iclog);
 842
 843                 spin_lock(&log->l_icloglock);
 844
 845                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 846                         || iclog->ic_state == XLOG_STATE_DIRTY
 847                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 848
 849                                 xlog_wait(&iclog->ic_force_wait,
 850                                                         &log->l_icloglock);
 851                 } else {
 852                         spin_unlock(&log->l_icloglock);
 853                 }
 854         }
 855
 856         return error;
 857 }       /* xfs_log_unmount_write */
 858
 859 /*
 860  * Empty the log for unmount/freeze.
 861  *
 862  * To do this, we first need to shut down the background log work so it is not
 863  * trying to cover the log as we clean up. We then need to unpin all objects in
 864  * the log so we can then flush them out. Once they have completed their IO and
 865  * run the callbacks removing themselves from the AIL, we can write the unmount
 866  * record.
 867  */
 868 void
 869 xfs_log_quiesce(
 870         struct xfs_mount        *mp)
 871 {
 872         cancel_delayed_work_sync(&mp->m_log->l_work);
 873         xfs_log_force(mp, XFS_LOG_SYNC);
 874
 875         /*
 876          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
 877          * will push it, xfs_wait_buftarg() will not wait for it. Further,
 878          * xfs_buf_iowait() cannot be used because it was pushed with the
 879          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
 880          * the IO to complete.
 881          */
 882         xfs_ail_push_all_sync(mp->m_ail);
 883         xfs_wait_buftarg(mp->m_ddev_targp);
 884         xfs_buf_lock(mp->m_sb_bp);
 885         xfs_buf_unlock(mp->m_sb_bp);
 886
 887         xfs_log_unmount_write(mp);
 888 }
 889
 890 /*
 891  * Shut down and release the AIL and Log.
 892  *
 893  * During unmount, we need to ensure we flush all the dirty metadata objects
 894  * from the AIL so that the log is empty before we write the unmount record to
 895  * the log. Once this is done, we can tear down the AIL and the log.
 896  */
 897 void
 898 xfs_log_unmount(
 899         struct xfs_mount        *mp)
 900 {
 901         xfs_log_quiesce(mp);
 902
 903         xfs_trans_ail_destroy(mp);
 904         xlog_dealloc_log(mp->m_log);
 905 }
 906
 907 void
 908 xfs_log_item_init(
 909         struct xfs_mount        *mp,
 910         struct xfs_log_item     *item,
 911         int                     type,
 912         const struct xfs_item_ops *ops)
 913 {
 914         item->li_mountp = mp;
 915         item->li_ailp = mp->m_ail;
 916         item->li_type = type;
 917         item->li_ops = ops;
 918         item->li_lv = NULL;
 919
 920         INIT_LIST_HEAD(&item->li_ail);
 921         INIT_LIST_HEAD(&item->li_cil);
 922 }
 923
 924 /*
 925  * Wake up processes waiting for log space after we have moved the log tail.
 926  */
 927 void
 928 xfs_log_space_wake(
 929         struct xfs_mount        *mp)
 930 {
 931         struct xlog             *log = mp->m_log;
 932         int                     free_bytes;
 933
 934         if (XLOG_FORCED_SHUTDOWN(log))
 935                 return;
 936
 937         if (!list_empty_careful(&log->l_write_head.waiters)) {
 938                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 939
 940                 spin_lock(&log->l_write_head.lock);
 941                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
 942                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
 943                 spin_unlock(&log->l_write_head.lock);
 944         }
 945
 946         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
 947                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 948
 949                 spin_lock(&log->l_reserve_head.lock);
 950                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
 951                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
 952                 spin_unlock(&log->l_reserve_head.lock);
 953         }
 954 }
 955
 956 /*
 957  * Determine if we have a transaction that has gone to disk
 958  * that needs to be covered. To begin the transition to the idle state
 959  * firstly the log needs to be idle (no AIL and nothing in the iclogs).
 960  * If we are then in a state where covering is needed, the caller is informed
 961  * that dummy transactions are required to move the log into the idle state.
 962  *
 963  * Because this is called as part of the sync process, we should also indicate
 964  * that dummy transactions should be issued in anything but the covered or
 965  * idle states. This ensures that the log tail is accurately reflected in
 966  * the log at the end of the sync, hence if a crash occurrs avoids replay
 967  * of transactions where the metadata is already on disk.
 968  */
 969 int
 970 xfs_log_need_covered(xfs_mount_t *mp)
 971 {
 972         int             needed = 0;
 973         struct xlog     *log = mp->m_log;
 974
 975         if (!xfs_fs_writable(mp))
 976                 return 0;
 977
 978         spin_lock(&log->l_icloglock);
 979         switch (log->l_covered_state) {
 980         case XLOG_STATE_COVER_DONE:
 981         case XLOG_STATE_COVER_DONE2:
 982         case XLOG_STATE_COVER_IDLE:
 983                 break;
 984         case XLOG_STATE_COVER_NEED:
 985         case XLOG_STATE_COVER_NEED2:
 986                 if (!xfs_ail_min_lsn(log->l_ailp) &&
 987                     xlog_iclogs_empty(log)) {
 988                         if (log->l_covered_state == XLOG_STATE_COVER_NEED)
 989                                 log->l_covered_state = XLOG_STATE_COVER_DONE;
 990                         else
 991                                 log->l_covered_state = XLOG_STATE_COVER_DONE2;
 992                 }
 993                 /* FALLTHRU */
 994         default:
 995                 needed = 1;
 996                 break;
 997         }
 998         spin_unlock(&log->l_icloglock);
 999         return needed;
1000 }
1001
1002 /*
1003  * We may be holding the log iclog lock upon entering this routine.
1004  */
1005 xfs_lsn_t
1006 xlog_assign_tail_lsn_locked(
1007         struct xfs_mount        *mp)
1008 {
1009         struct xlog             *log = mp->m_log;
1010         struct xfs_log_item     *lip;
1011         xfs_lsn_t               tail_lsn;
1012
1013         assert_spin_locked(&mp->m_ail->xa_lock);
1014
1015         /*
1016          * To make sure we always have a valid LSN for the log tail we keep
1017          * track of the last LSN which was committed in log->l_last_sync_lsn,
1018          * and use that when the AIL was empty.
1019          */
1020         lip = xfs_ail_min(mp->m_ail);
1021         if (lip)
1022                 tail_lsn = lip->li_lsn;
1023         else
1024                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1025         atomic64_set(&log->l_tail_lsn, tail_lsn);
1026         return tail_lsn;
1027 }
1028
1029 xfs_lsn_t
1030 xlog_assign_tail_lsn(
1031         struct xfs_mount        *mp)
1032 {
1033         xfs_lsn_t               tail_lsn;
1034
1035         spin_lock(&mp->m_ail->xa_lock);
1036         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1037         spin_unlock(&mp->m_ail->xa_lock);
1038
1039         return tail_lsn;
1040 }
1041
1042 /*
1043  * Return the space in the log between the tail and the head.  The head
1044  * is passed in the cycle/bytes formal parms.  In the special case where
1045  * the reserve head has wrapped passed the tail, this calculation is no
1046  * longer valid.  In this case, just return 0 which means there is no space
1047  * in the log.  This works for all places where this function is called
1048  * with the reserve head.  Of course, if the write head were to ever
1049  * wrap the tail, we should blow up.  Rather than catch this case here,
1050  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1051  *
1052  * This code also handles the case where the reservation head is behind
1053  * the tail.  The details of this case are described below, but the end
1054  * result is that we return the size of the log as the amount of space left.
1055  */
1056 STATIC int
1057 xlog_space_left(
1058         struct xlog     *log,
1059         atomic64_t      *head)
1060 {
1061         int             free_bytes;
1062         int             tail_bytes;
1063         int             tail_cycle;
1064         int             head_cycle;
1065         int             head_bytes;
1066
1067         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1068         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1069         tail_bytes = BBTOB(tail_bytes);
1070         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1071                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1072         else if (tail_cycle + 1 < head_cycle)
1073                 return 0;
1074         else if (tail_cycle < head_cycle) {
1075                 ASSERT(tail_cycle == (head_cycle - 1));
1076                 free_bytes = tail_bytes - head_bytes;
1077         } else {
1078                 /*
1079                  * The reservation head is behind the tail.
1080                  * In this case we just want to return the size of the
1081                  * log as the amount of space left.
1082                  */
1083                 xfs_alert(log->l_mp,
1084                         "xlog_space_left: head behind tail\n"
1085                         "  tail_cycle = %d, tail_bytes = %d\n"
1086                         "  GH   cycle = %d, GH   bytes = %d",
1087                         tail_cycle, tail_bytes, head_cycle, head_bytes);
1088                 ASSERT(0);
1089                 free_bytes = log->l_logsize;
1090         }
1091         return free_bytes;
1092 }
1093
1094
1095 /*
1096  * Log function which is called when an io completes.
1097  *
1098  * The log manager needs its own routine, in order to control what
1099  * happens with the buffer after the write completes.
1100  */
1101 void
1102 xlog_iodone(xfs_buf_t *bp)
1103 {
1104         struct xlog_in_core     *iclog = bp->b_fspriv;
1105         struct xlog             *l = iclog->ic_log;
1106         int                     aborted = 0;
1107
1108         /*
1109          * Race to shutdown the filesystem if we see an error.
1110          */
1111         if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1112                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1113                 xfs_buf_ioerror_alert(bp, __func__);
1114                 xfs_buf_stale(bp);
1115                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1116                 /*
1117                  * This flag will be propagated to the trans-committed
1118                  * callback routines to let them know that the log-commit
1119                  * didn't succeed.
1120                  */
1121                 aborted = XFS_LI_ABORTED;
1122         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1123                 aborted = XFS_LI_ABORTED;
1124         }
1125
1126         /* log I/O is always issued ASYNC */
1127         ASSERT(XFS_BUF_ISASYNC(bp));
1128         xlog_state_done_syncing(iclog, aborted);
1129         /*
1130          * do not reference the buffer (bp) here as we could race
1131          * with it being freed after writing the unmount record to the
1132          * log.
1133          */
1134 }
1135
1136 /*
1137  * Return size of each in-core log record buffer.
1138  *
1139  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1140  *
1141  * If the filesystem blocksize is too large, we may need to choose a
1142  * larger size since the directory code currently logs entire blocks.
1143  */
1144
1145 STATIC void
1146 xlog_get_iclog_buffer_size(
1147         struct xfs_mount        *mp,
1148         struct xlog             *log)
1149 {
1150         int size;
1151         int xhdrs;
1152
1153         if (mp->m_logbufs <= 0)
1154                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1155         else
1156                 log->l_iclog_bufs = mp->m_logbufs;
1157
1158         /*
1159          * Buffer size passed in from mount system call.
1160          */
1161         if (mp->m_logbsize > 0) {
1162                 size = log->l_iclog_size = mp->m_logbsize;
1163                 log->l_iclog_size_log = 0;
1164                 while (size != 1) {
1165                         log->l_iclog_size_log++;
1166                         size >>= 1;
1167                 }
1168
1169                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1170                         /* # headers = size / 32k
1171                          * one header holds cycles from 32k of data
1172                          */
1173
1174                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1175                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1176                                 xhdrs++;
1177                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1178                         log->l_iclog_heads = xhdrs;
1179                 } else {
1180                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1181                         log->l_iclog_hsize = BBSIZE;
1182                         log->l_iclog_heads = 1;
1183                 }
1184                 goto done;
1185         }
1186
1187         /* All machines use 32kB buffers by default. */
1188         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1189         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1190
1191         /* the default log size is 16k or 32k which is one header sector */
1192         log->l_iclog_hsize = BBSIZE;
1193         log->l_iclog_heads = 1;
1194
1195 done:
1196         /* are we being asked to make the sizes selected above visible? */
1197         if (mp->m_logbufs == 0)
1198                 mp->m_logbufs = log->l_iclog_bufs;
1199         if (mp->m_logbsize == 0)
1200                 mp->m_logbsize = log->l_iclog_size;
1201 }       /* xlog_get_iclog_buffer_size */
1202
1203
1204 void
1205 xfs_log_work_queue(
1206         struct xfs_mount        *mp)
1207 {
1208         queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1209                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1210 }
1211
1212 /*
1213  * Every sync period we need to unpin all items in the AIL and push them to
1214  * disk. If there is nothing dirty, then we might need to cover the log to
1215  * indicate that the filesystem is idle.
1216  */
1217 void
1218 xfs_log_worker(
1219         struct work_struct      *work)
1220 {
1221         struct xlog             *log = container_of(to_delayed_work(work),
1222                                                 struct xlog, l_work);
1223         struct xfs_mount        *mp = log->l_mp;
1224
1225         /* dgc: errors ignored - not fatal and nowhere to report them */
1226         if (xfs_log_need_covered(mp))
1227                 xfs_fs_log_dummy(mp);
1228         else
1229                 xfs_log_force(mp, 0);
1230
1231         /* start pushing all the metadata that is currently dirty */
1232         xfs_ail_push_all(mp->m_ail);
1233
1234         /* queue us up again */
1235         xfs_log_work_queue(mp);
1236 }
1237
1238 /*
1239  * This routine initializes some of the log structure for a given mount point.
1240  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1241  * some other stuff may be filled in too.
1242  */
1243 STATIC struct xlog *
1244 xlog_alloc_log(
1245         struct xfs_mount        *mp,
1246         struct xfs_buftarg      *log_target,
1247         xfs_daddr_t             blk_offset,
1248         int                     num_bblks)
1249 {
1250         struct xlog             *log;
1251         xlog_rec_header_t       *head;
1252         xlog_in_core_t          **iclogp;
1253         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1254         xfs_buf_t               *bp;
1255         int                     i;
1256         int                     error = ENOMEM;
1257         uint                    log2_size = 0;
1258
1259         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1260         if (!log) {
1261                 xfs_warn(mp, "Log allocation failed: No memory!");
1262                 goto out;
1263         }
1264
1265         log->l_mp          = mp;
1266         log->l_targ        = log_target;
1267         log->l_logsize     = BBTOB(num_bblks);
1268         log->l_logBBstart  = blk_offset;
1269         log->l_logBBsize   = num_bblks;
1270         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1271         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1272         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1273
1274         log->l_prev_block  = -1;
1275         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1276         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1277         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1278         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1279
1280         xlog_grant_head_init(&log->l_reserve_head);
1281         xlog_grant_head_init(&log->l_write_head);
1282
1283         error = EFSCORRUPTED;
1284         if (xfs_sb_version_hassector(&mp->m_sb)) {
1285                 log2_size = mp->m_sb.sb_logsectlog;
1286                 if (log2_size < BBSHIFT) {
1287                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1288                                 log2_size, BBSHIFT);
1289                         goto out_free_log;
1290                 }
1291
1292                 log2_size -= BBSHIFT;
1293                 if (log2_size > mp->m_sectbb_log) {
1294                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1295                                 log2_size, mp->m_sectbb_log);
1296                         goto out_free_log;
1297                 }
1298
1299                 /* for larger sector sizes, must have v2 or external log */
1300                 if (log2_size && log->l_logBBstart > 0 &&
1301                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1302                         xfs_warn(mp,
1303                 "log sector size (0x%x) invalid for configuration.",
1304                                 log2_size);
1305                         goto out_free_log;
1306                 }
1307         }
1308         log->l_sectBBsize = 1 << log2_size;
1309
1310         xlog_get_iclog_buffer_size(mp, log);
1311
1312         error = ENOMEM;
1313         bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1314         if (!bp)
1315                 goto out_free_log;
1316         bp->b_iodone = xlog_iodone;
1317         ASSERT(xfs_buf_islocked(bp));
1318         log->l_xbuf = bp;
1319
1320         spin_lock_init(&log->l_icloglock);
1321         init_waitqueue_head(&log->l_flush_wait);
1322
1323         iclogp = &log->l_iclog;
1324         /*
1325          * The amount of memory to allocate for the iclog structure is
1326          * rather funky due to the way the structure is defined.  It is
1327          * done this way so that we can use different sizes for machines
1328          * with different amounts of memory.  See the definition of
1329          * xlog_in_core_t in xfs_log_priv.h for details.
1330          */
1331         ASSERT(log->l_iclog_size >= 4096);
1332         for (i=0; i < log->l_iclog_bufs; i++) {
1333                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1334                 if (!*iclogp)
1335                         goto out_free_iclog;
1336
1337                 iclog = *iclogp;
1338                 iclog->ic_prev = prev_iclog;
1339                 prev_iclog = iclog;
1340
1341                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1342                                                 BTOBB(log->l_iclog_size), 0);
1343                 if (!bp)
1344                         goto out_free_iclog;
1345
1346                 bp->b_iodone = xlog_iodone;
1347                 iclog->ic_bp = bp;
1348                 iclog->ic_data = bp->b_addr;
1349 #ifdef DEBUG
1350                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1351 #endif
1352                 head = &iclog->ic_header;
1353                 memset(head, 0, sizeof(xlog_rec_header_t));
1354                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1355                 head->h_version = cpu_to_be32(
1356                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1357                 head->h_size = cpu_to_be32(log->l_iclog_size);
1358                 /* new fields */
1359                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1360                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1361
1362                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1363                 iclog->ic_state = XLOG_STATE_ACTIVE;
1364                 iclog->ic_log = log;
1365                 atomic_set(&iclog->ic_refcnt, 0);
1366                 spin_lock_init(&iclog->ic_callback_lock);
1367                 iclog->ic_callback_tail = &(iclog->ic_callback);
1368                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1369
1370                 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1371                 init_waitqueue_head(&iclog->ic_force_wait);
1372                 init_waitqueue_head(&iclog->ic_write_wait);
1373
1374                 iclogp = &iclog->ic_next;
1375         }
1376         *iclogp = log->l_iclog;                 /* complete ring */
1377         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1378
1379         error = xlog_cil_init(log);
1380         if (error)
1381                 goto out_free_iclog;
1382         return log;
1383
1384 out_free_iclog:
1385         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1386                 prev_iclog = iclog->ic_next;
1387                 if (iclog->ic_bp)
1388                         xfs_buf_free(iclog->ic_bp);
1389                 kmem_free(iclog);
1390         }
1391         spinlock_destroy(&log->l_icloglock);
1392         xfs_buf_free(log->l_xbuf);
1393 out_free_log:
1394         kmem_free(log);
1395 out:
1396         return ERR_PTR(-error);
1397 }       /* xlog_alloc_log */
1398
1399
1400 /*
1401  * Write out the commit record of a transaction associated with the given
1402  * ticket.  Return the lsn of the commit record.
1403  */
1404 STATIC int
1405 xlog_commit_record(
1406         struct xlog             *log,
1407         struct xlog_ticket      *ticket,
1408         struct xlog_in_core     **iclog,
1409         xfs_lsn_t               *commitlsnp)
1410 {
1411         struct xfs_mount *mp = log->l_mp;
1412         int     error;
1413         struct xfs_log_iovec reg = {
1414                 .i_addr = NULL,
1415                 .i_len = 0,
1416                 .i_type = XLOG_REG_TYPE_COMMIT,
1417         };
1418         struct xfs_log_vec vec = {
1419                 .lv_niovecs = 1,
1420                 .lv_iovecp = &reg,
1421         };
1422
1423         ASSERT_ALWAYS(iclog);
1424         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1425                                         XLOG_COMMIT_TRANS);
1426         if (error)
1427                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1428         return error;
1429 }
1430
1431 /*
1432  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1433  * log space.  This code pushes on the lsn which would supposedly free up
1434  * the 25% which we want to leave free.  We may need to adopt a policy which
1435  * pushes on an lsn which is further along in the log once we reach the high
1436  * water mark.  In this manner, we would be creating a low water mark.
1437  */
1438 STATIC void
1439 xlog_grant_push_ail(
1440         struct xlog     *log,
1441         int             need_bytes)
1442 {
1443         xfs_lsn_t       threshold_lsn = 0;
1444         xfs_lsn_t       last_sync_lsn;
1445         int             free_blocks;
1446         int             free_bytes;
1447         int             threshold_block;
1448         int             threshold_cycle;
1449         int             free_threshold;
1450
1451         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1452
1453         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1454         free_blocks = BTOBBT(free_bytes);
1455
1456         /*
1457          * Set the threshold for the minimum number of free blocks in the
1458          * log to the maximum of what the caller needs, one quarter of the
1459          * log, and 256 blocks.
1460          */
1461         free_threshold = BTOBB(need_bytes);
1462         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1463         free_threshold = MAX(free_threshold, 256);
1464         if (free_blocks >= free_threshold)
1465                 return;
1466
1467         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1468                                                 &threshold_block);
1469         threshold_block += free_threshold;
1470         if (threshold_block >= log->l_logBBsize) {
1471                 threshold_block -= log->l_logBBsize;
1472                 threshold_cycle += 1;
1473         }
1474         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1475                                         threshold_block);
1476         /*
1477          * Don't pass in an lsn greater than the lsn of the last
1478          * log record known to be on disk. Use a snapshot of the last sync lsn
1479          * so that it doesn't change between the compare and the set.
1480          */
1481         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1482         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1483                 threshold_lsn = last_sync_lsn;
1484
1485         /*
1486          * Get the transaction layer to kick the dirty buffers out to
1487          * disk asynchronously. No point in trying to do this if
1488          * the filesystem is shutting down.
1489          */
1490         if (!XLOG_FORCED_SHUTDOWN(log))
1491                 xfs_ail_push(log->l_ailp, threshold_lsn);
1492 }
1493
1494 /*
1495  * Stamp cycle number in every block
1496  */
1497 STATIC void
1498 xlog_pack_data(
1499         struct xlog             *log,
1500         struct xlog_in_core     *iclog,
1501         int                     roundoff)
1502 {
1503         int                     i, j, k;
1504         int                     size = iclog->ic_offset + roundoff;
1505         __be32                  cycle_lsn;
1506         xfs_caddr_t             dp;
1507
1508         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1509
1510         dp = iclog->ic_datap;
1511         for (i = 0; i < BTOBB(size); i++) {
1512                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1513                         break;
1514                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1515                 *(__be32 *)dp = cycle_lsn;
1516                 dp += BBSIZE;
1517         }
1518
1519         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1520                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1521
1522                 for ( ; i < BTOBB(size); i++) {
1523                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1524                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1525                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1526                         *(__be32 *)dp = cycle_lsn;
1527                         dp += BBSIZE;
1528                 }
1529
1530                 for (i = 1; i < log->l_iclog_heads; i++)
1531                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1532         }
1533 }
1534
1535 /*
1536  * Calculate the checksum for a log buffer.
1537  *
1538  * This is a little more complicated than it should be because the various
1539  * headers and the actual data are non-contiguous.
1540  */
1541 __le32
1542 xlog_cksum(
1543         struct xlog             *log,
1544         struct xlog_rec_header  *rhead,
1545         char                    *dp,
1546         int                     size)
1547 {
1548         __uint32_t              crc;
1549
1550         /* first generate the crc for the record header ... */
1551         crc = xfs_start_cksum((char *)rhead,
1552                               sizeof(struct xlog_rec_header),
1553                               offsetof(struct xlog_rec_header, h_crc));
1554
1555         /* ... then for additional cycle data for v2 logs ... */
1556         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1557                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1558                 int             i;
1559
1560                 for (i = 1; i < log->l_iclog_heads; i++) {
1561                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1562                                      sizeof(struct xlog_rec_ext_header));
1563                 }
1564         }
1565
1566         /* ... and finally for the payload */
1567         crc = crc32c(crc, dp, size);
1568
1569         return xfs_end_cksum(crc);
1570 }
1571
1572 /*
1573  * The bdstrat callback function for log bufs. This gives us a central
1574  * place to trap bufs in case we get hit by a log I/O error and need to
1575  * shutdown. Actually, in practice, even when we didn't get a log error,
1576  * we transition the iclogs to IOERROR state *after* flushing all existing
1577  * iclogs to disk. This is because we don't want anymore new transactions to be
1578  * started or completed afterwards.
1579  */
1580 STATIC int
1581 xlog_bdstrat(
1582         struct xfs_buf          *bp)
1583 {
1584         struct xlog_in_core     *iclog = bp->b_fspriv;
1585
1586         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1587                 xfs_buf_ioerror(bp, EIO);
1588                 xfs_buf_stale(bp);
1589                 xfs_buf_ioend(bp, 0);
1590                 /*
1591                  * It would seem logical to return EIO here, but we rely on
1592                  * the log state machine to propagate I/O errors instead of
1593                  * doing it here.
1594                  */
1595                 return 0;
1596         }
1597
1598         xfs_buf_iorequest(bp);
1599         return 0;
1600 }
1601
1602 /*
1603  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1604  * fashion.  Previously, we should have moved the current iclog
1605  * ptr in the log to point to the next available iclog.  This allows further
1606  * write to continue while this code syncs out an iclog ready to go.
1607  * Before an in-core log can be written out, the data section must be scanned
1608  * to save away the 1st word of each BBSIZE block into the header.  We replace
1609  * it with the current cycle count.  Each BBSIZE block is tagged with the
1610  * cycle count because there in an implicit assumption that drives will
1611  * guarantee that entire 512 byte blocks get written at once.  In other words,
1612  * we can't have part of a 512 byte block written and part not written.  By
1613  * tagging each block, we will know which blocks are valid when recovering
1614  * after an unclean shutdown.
1615  *
1616  * This routine is single threaded on the iclog.  No other thread can be in
1617  * this routine with the same iclog.  Changing contents of iclog can there-
1618  * fore be done without grabbing the state machine lock.  Updating the global
1619  * log will require grabbing the lock though.
1620  *
1621  * The entire log manager uses a logical block numbering scheme.  Only
1622  * log_sync (and then only bwrite()) know about the fact that the log may
1623  * not start with block zero on a given device.  The log block start offset
1624  * is added immediately before calling bwrite().
1625  */
1626
1627 STATIC int
1628 xlog_sync(
1629         struct xlog             *log,
1630         struct xlog_in_core     *iclog)
1631 {
1632         xfs_buf_t       *bp;
1633         int             i;
1634         uint            count;          /* byte count of bwrite */
1635         uint            count_init;     /* initial count before roundup */
1636         int             roundoff;       /* roundoff to BB or stripe */
1637         int             split = 0;      /* split write into two regions */
1638         int             error;
1639         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1640         int             size;
1641
1642         XFS_STATS_INC(xs_log_writes);
1643         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1644
1645         /* Add for LR header */
1646         count_init = log->l_iclog_hsize + iclog->ic_offset;
1647
1648         /* Round out the log write size */
1649         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1650                 /* we have a v2 stripe unit to use */
1651                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1652         } else {
1653                 count = BBTOB(BTOBB(count_init));
1654         }
1655         roundoff = count - count_init;
1656         ASSERT(roundoff >= 0);
1657         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1658                 roundoff < log->l_mp->m_sb.sb_logsunit)
1659                 ||
1660                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1661                  roundoff < BBTOB(1)));
1662
1663         /* move grant heads by roundoff in sync */
1664         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1665         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1666
1667         /* put cycle number in every block */
1668         xlog_pack_data(log, iclog, roundoff);
1669
1670         /* real byte length */
1671         size = iclog->ic_offset;
1672         if (v2)
1673                 size += roundoff;
1674         iclog->ic_header.h_len = cpu_to_be32(size);
1675
1676         bp = iclog->ic_bp;
1677         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1678
1679         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1680
1681         /* Do we need to split this write into 2 parts? */
1682         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1683                 char            *dptr;
1684
1685                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1686                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1687                 iclog->ic_bwritecnt = 2;
1688
1689                 /*
1690                  * Bump the cycle numbers at the start of each block in the
1691                  * part of the iclog that ends up in the buffer that gets
1692                  * written to the start of the log.
1693                  *
1694                  * Watch out for the header magic number case, though.
1695                  */
1696                 dptr = (char *)&iclog->ic_header + count;
1697                 for (i = 0; i < split; i += BBSIZE) {
1698                         __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1699                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1700                                 cycle++;
1701                         *(__be32 *)dptr = cpu_to_be32(cycle);
1702
1703                         dptr += BBSIZE;
1704                 }
1705         } else {
1706                 iclog->ic_bwritecnt = 1;
1707         }
1708
1709         /* calculcate the checksum */
1710         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1711                                             iclog->ic_datap, size);
1712
1713         bp->b_io_length = BTOBB(count);
1714         bp->b_fspriv = iclog;
1715         XFS_BUF_ZEROFLAGS(bp);
1716         XFS_BUF_ASYNC(bp);
1717         bp->b_flags |= XBF_SYNCIO;
1718
1719         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1720                 bp->b_flags |= XBF_FUA;
1721
1722                 /*
1723                  * Flush the data device before flushing the log to make
1724                  * sure all meta data written back from the AIL actually made
1725                  * it to disk before stamping the new log tail LSN into the
1726                  * log buffer.  For an external log we need to issue the
1727                  * flush explicitly, and unfortunately synchronously here;
1728                  * for an internal log we can simply use the block layer
1729                  * state machine for preflushes.
1730                  */
1731                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1732                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1733                 else
1734                         bp->b_flags |= XBF_FLUSH;
1735         }
1736
1737         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1738         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1739
1740         xlog_verify_iclog(log, iclog, count, true);
1741
1742         /* account for log which doesn't start at block #0 */
1743         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1744         /*
1745          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1746          * is shutting down.
1747          */
1748         XFS_BUF_WRITE(bp);
1749
1750         error = xlog_bdstrat(bp);
1751         if (error) {
1752                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1753                 return error;
1754         }
1755         if (split) {
1756                 bp = iclog->ic_log->l_xbuf;
1757                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1758                 xfs_buf_associate_memory(bp,
1759                                 (char *)&iclog->ic_header + count, split);
1760                 bp->b_fspriv = iclog;
1761                 XFS_BUF_ZEROFLAGS(bp);
1762                 XFS_BUF_ASYNC(bp);
1763                 bp->b_flags |= XBF_SYNCIO;
1764                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1765                         bp->b_flags |= XBF_FUA;
1766
1767                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1768                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1769
1770                 /* account for internal log which doesn't start at block #0 */
1771                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1772                 XFS_BUF_WRITE(bp);
1773                 error = xlog_bdstrat(bp);
1774                 if (error) {
1775                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1776                         return error;
1777                 }
1778         }
1779         return 0;
1780 }       /* xlog_sync */
1781
1782 /*
1783  * Deallocate a log structure
1784  */
1785 STATIC void
1786 xlog_dealloc_log(
1787         struct xlog     *log)
1788 {
1789         xlog_in_core_t  *iclog, *next_iclog;
1790         int             i;
1791
1792         xlog_cil_destroy(log);
1793
1794         /*
1795          * always need to ensure that the extra buffer does not point to memory
1796          * owned by another log buffer before we free it.
1797          */
1798         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1799         xfs_buf_free(log->l_xbuf);
1800
1801         iclog = log->l_iclog;
1802         for (i=0; i<log->l_iclog_bufs; i++) {
1803                 xfs_buf_free(iclog->ic_bp);
1804                 next_iclog = iclog->ic_next;
1805                 kmem_free(iclog);
1806                 iclog = next_iclog;
1807         }
1808         spinlock_destroy(&log->l_icloglock);
1809
1810         log->l_mp->m_log = NULL;
1811         kmem_free(log);
1812 }       /* xlog_dealloc_log */
1813
1814 /*
1815  * Update counters atomically now that memcpy is done.
1816  */
1817 /* ARGSUSED */
1818 static inline void
1819 xlog_state_finish_copy(
1820         struct xlog             *log,
1821         struct xlog_in_core     *iclog,
1822         int                     record_cnt,
1823         int                     copy_bytes)
1824 {
1825         spin_lock(&log->l_icloglock);
1826
1827         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1828         iclog->ic_offset += copy_bytes;
1829
1830         spin_unlock(&log->l_icloglock);
1831 }       /* xlog_state_finish_copy */
1832
1833
1834
1835
1836 /*
1837  * print out info relating to regions written which consume
1838  * the reservation
1839  */
1840 void
1841 xlog_print_tic_res(
1842         struct xfs_mount        *mp,
1843         struct xlog_ticket      *ticket)
1844 {
1845         uint i;
1846         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1847
1848         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1849         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1850             "bformat",
1851             "bchunk",
1852             "efi_format",
1853             "efd_format",
1854             "iformat",
1855             "icore",
1856             "iext",
1857             "ibroot",
1858             "ilocal",
1859             "iattr_ext",
1860             "iattr_broot",
1861             "iattr_local",
1862             "qformat",
1863             "dquot",
1864             "quotaoff",
1865             "LR header",
1866             "unmount",
1867             "commit",
1868             "trans header"
1869         };
1870         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1871             "SETATTR_NOT_SIZE",
1872             "SETATTR_SIZE",
1873             "INACTIVE",
1874             "CREATE",
1875             "CREATE_TRUNC",
1876             "TRUNCATE_FILE",
1877             "REMOVE",
1878             "LINK",
1879             "RENAME",
1880             "MKDIR",
1881             "RMDIR",
1882             "SYMLINK",
1883             "SET_DMATTRS",
1884             "GROWFS",
1885             "STRAT_WRITE",
1886             "DIOSTRAT",
1887             "WRITE_SYNC",
1888             "WRITEID",
1889             "ADDAFORK",
1890             "ATTRINVAL",
1891             "ATRUNCATE",
1892             "ATTR_SET",
1893             "ATTR_RM",
1894             "ATTR_FLAG",
1895             "CLEAR_AGI_BUCKET",
1896             "QM_SBCHANGE",
1897             "DUMMY1",
1898             "DUMMY2",
1899             "QM_QUOTAOFF",
1900             "QM_DQALLOC",
1901             "QM_SETQLIM",
1902             "QM_DQCLUSTER",
1903             "QM_QINOCREATE",
1904             "QM_QUOTAOFF_END",
1905             "SB_UNIT",
1906             "FSYNC_TS",
1907             "GROWFSRT_ALLOC",
1908             "GROWFSRT_ZERO",
1909             "GROWFSRT_FREE",
1910             "SWAPEXT"
1911         };
1912
1913         xfs_warn(mp,
1914                 "xlog_write: reservation summary:\n"
1915                 "  trans type  = %s (%u)\n"
1916                 "  unit res    = %d bytes\n"
1917                 "  current res = %d bytes\n"
1918                 "  total reg   = %u bytes (o/flow = %u bytes)\n"
1919                 "  ophdrs      = %u (ophdr space = %u bytes)\n"
1920                 "  ophdr + reg = %u bytes\n"
1921                 "  num regions = %u\n",
1922                 ((ticket->t_trans_type <= 0 ||
1923                   ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1924                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1925                 ticket->t_trans_type,
1926                 ticket->t_unit_res,
1927                 ticket->t_curr_res,
1928                 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1929                 ticket->t_res_num_ophdrs, ophdr_spc,
1930                 ticket->t_res_arr_sum +
1931                 ticket->t_res_o_flow + ophdr_spc,
1932                 ticket->t_res_num);
1933
1934         for (i = 0; i < ticket->t_res_num; i++) {
1935                 uint r_type = ticket->t_res_arr[i].r_type;
1936                 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1937                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1938                             "bad-rtype" : res_type_str[r_type-1]),
1939                             ticket->t_res_arr[i].r_len);
1940         }
1941
1942         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1943                 "xlog_write: reservation ran out. Need to up reservation");
1944         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1945 }
1946
1947 /*
1948  * Calculate the potential space needed by the log vector.  Each region gets
1949  * its own xlog_op_header_t and may need to be double word aligned.
1950  */
1951 static int
1952 xlog_write_calc_vec_length(
1953         struct xlog_ticket      *ticket,
1954         struct xfs_log_vec      *log_vector)
1955 {
1956         struct xfs_log_vec      *lv;
1957         int                     headers = 0;
1958         int                     len = 0;
1959         int                     i;
1960
1961         /* acct for start rec of xact */
1962         if (ticket->t_flags & XLOG_TIC_INITED)
1963                 headers++;
1964
1965         for (lv = log_vector; lv; lv = lv->lv_next) {
1966                 /* we don't write ordered log vectors */
1967                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
1968                         continue;
1969
1970                 headers += lv->lv_niovecs;
1971
1972                 for (i = 0; i < lv->lv_niovecs; i++) {
1973                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
1974
1975                         len += vecp->i_len;
1976                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1977                 }
1978         }
1979
1980         ticket->t_res_num_ophdrs += headers;
1981         len += headers * sizeof(struct xlog_op_header);
1982
1983         return len;
1984 }
1985
1986 /*
1987  * If first write for transaction, insert start record  We can't be trying to
1988  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
1989  */
1990 static int
1991 xlog_write_start_rec(
1992         struct xlog_op_header   *ophdr,
1993         struct xlog_ticket      *ticket)
1994 {
1995         if (!(ticket->t_flags & XLOG_TIC_INITED))
1996                 return 0;
1997
1998         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
1999         ophdr->oh_clientid = ticket->t_clientid;
2000         ophdr->oh_len = 0;
2001         ophdr->oh_flags = XLOG_START_TRANS;
2002         ophdr->oh_res2 = 0;
2003
2004         ticket->t_flags &= ~XLOG_TIC_INITED;
2005
2006         return sizeof(struct xlog_op_header);
2007 }
2008
2009 static xlog_op_header_t *
2010 xlog_write_setup_ophdr(
2011         struct xlog             *log,
2012         struct xlog_op_header   *ophdr,
2013         struct xlog_ticket      *ticket,
2014         uint                    flags)
2015 {
2016         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2017         ophdr->oh_clientid = ticket->t_clientid;
2018         ophdr->oh_res2 = 0;
2019
2020         /* are we copying a commit or unmount record? */
2021         ophdr->oh_flags = flags;
2022
2023         /*
2024          * We've seen logs corrupted with bad transaction client ids.  This
2025          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2026          * and shut down the filesystem.
2027          */
2028         switch (ophdr->oh_clientid)  {
2029         case XFS_TRANSACTION:
2030         case XFS_VOLUME:
2031         case XFS_LOG:
2032                 break;
2033         default:
2034                 xfs_warn(log->l_mp,
2035                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2036                         ophdr->oh_clientid, ticket);
2037                 return NULL;
2038         }
2039
2040         return ophdr;
2041 }
2042
2043 /*
2044  * Set up the parameters of the region copy into the log. This has
2045  * to handle region write split across multiple log buffers - this
2046  * state is kept external to this function so that this code can
2047  * can be written in an obvious, self documenting manner.
2048  */
2049 static int
2050 xlog_write_setup_copy(
2051         struct xlog_ticket      *ticket,
2052         struct xlog_op_header   *ophdr,
2053         int                     space_available,
2054         int                     space_required,
2055         int                     *copy_off,
2056         int                     *copy_len,
2057         int                     *last_was_partial_copy,
2058         int                     *bytes_consumed)
2059 {
2060         int                     still_to_copy;
2061
2062         still_to_copy = space_required - *bytes_consumed;
2063         *copy_off = *bytes_consumed;
2064
2065         if (still_to_copy <= space_available) {
2066                 /* write of region completes here */
2067                 *copy_len = still_to_copy;
2068                 ophdr->oh_len = cpu_to_be32(*copy_len);
2069                 if (*last_was_partial_copy)
2070                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2071                 *last_was_partial_copy = 0;
2072                 *bytes_consumed = 0;
2073                 return 0;
2074         }
2075
2076         /* partial write of region, needs extra log op header reservation */
2077         *copy_len = space_available;
2078         ophdr->oh_len = cpu_to_be32(*copy_len);
2079         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2080         if (*last_was_partial_copy)
2081                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2082         *bytes_consumed += *copy_len;
2083         (*last_was_partial_copy)++;
2084
2085         /* account for new log op header */
2086         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2087         ticket->t_res_num_ophdrs++;
2088
2089         return sizeof(struct xlog_op_header);
2090 }
2091
2092 static int
2093 xlog_write_copy_finish(
2094         struct xlog             *log,
2095         struct xlog_in_core     *iclog,
2096         uint                    flags,
2097         int                     *record_cnt,
2098         int                     *data_cnt,
2099         int                     *partial_copy,
2100         int                     *partial_copy_len,
2101         int                     log_offset,
2102         struct xlog_in_core     **commit_iclog)
2103 {
2104         if (*partial_copy) {
2105                 /*
2106                  * This iclog has already been marked WANT_SYNC by
2107                  * xlog_state_get_iclog_space.
2108                  */
2109                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2110                 *record_cnt = 0;
2111                 *data_cnt = 0;
2112                 return xlog_state_release_iclog(log, iclog);
2113         }
2114
2115         *partial_copy = 0;
2116         *partial_copy_len = 0;
2117
2118         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2119                 /* no more space in this iclog - push it. */
2120                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2121                 *record_cnt = 0;
2122                 *data_cnt = 0;
2123
2124                 spin_lock(&log->l_icloglock);
2125                 xlog_state_want_sync(log, iclog);
2126                 spin_unlock(&log->l_icloglock);
2127
2128                 if (!commit_iclog)
2129                         return xlog_state_release_iclog(log, iclog);
2130                 ASSERT(flags & XLOG_COMMIT_TRANS);
2131                 *commit_iclog = iclog;
2132         }
2133
2134         return 0;
2135 }
2136
2137 /*
2138  * Write some region out to in-core log
2139  *
2140  * This will be called when writing externally provided regions or when
2141  * writing out a commit record for a given transaction.
2142  *
2143  * General algorithm:
2144  *      1. Find total length of this write.  This may include adding to the
2145  *              lengths passed in.
2146  *      2. Check whether we violate the tickets reservation.
2147  *      3. While writing to this iclog
2148  *          A. Reserve as much space in this iclog as can get
2149  *          B. If this is first write, save away start lsn
2150  *          C. While writing this region:
2151  *              1. If first write of transaction, write start record
2152  *              2. Write log operation header (header per region)
2153  *              3. Find out if we can fit entire region into this iclog
2154  *              4. Potentially, verify destination memcpy ptr
2155  *              5. Memcpy (partial) region
2156  *              6. If partial copy, release iclog; otherwise, continue
2157  *                      copying more regions into current iclog
2158  *      4. Mark want sync bit (in simulation mode)
2159  *      5. Release iclog for potential flush to on-disk log.
2160  *
2161  * ERRORS:
2162  * 1.   Panic if reservation is overrun.  This should never happen since
2163  *      reservation amounts are generated internal to the filesystem.
2164  * NOTES:
2165  * 1. Tickets are single threaded data structures.
2166  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2167  *      syncing routine.  When a single log_write region needs to span
2168  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2169  *      on all log operation writes which don't contain the end of the
2170  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2171  *      operation which contains the end of the continued log_write region.
2172  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2173  *      we don't really know exactly how much space will be used.  As a result,
2174  *      we don't update ic_offset until the end when we know exactly how many
2175  *      bytes have been written out.
2176  */
2177 int
2178 xlog_write(
2179         struct xlog             *log,
2180         struct xfs_log_vec      *log_vector,
2181         struct xlog_ticket      *ticket,
2182         xfs_lsn_t               *start_lsn,
2183         struct xlog_in_core     **commit_iclog,
2184         uint                    flags)
2185 {
2186         struct xlog_in_core     *iclog = NULL;
2187         struct xfs_log_iovec    *vecp;
2188         struct xfs_log_vec      *lv;
2189         int                     len;
2190         int                     index;
2191         int                     partial_copy = 0;
2192         int                     partial_copy_len = 0;
2193         int                     contwr = 0;
2194         int                     record_cnt = 0;
2195         int                     data_cnt = 0;
2196         int                     error;
2197
2198         *start_lsn = 0;
2199
2200         len = xlog_write_calc_vec_length(ticket, log_vector);
2201
2202         /*
2203          * Region headers and bytes are already accounted for.
2204          * We only need to take into account start records and
2205          * split regions in this function.
2206          */
2207         if (ticket->t_flags & XLOG_TIC_INITED)
2208                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2209
2210         /*
2211          * Commit record headers need to be accounted for. These
2212          * come in as separate writes so are easy to detect.
2213          */
2214         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2215                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2216
2217         if (ticket->t_curr_res < 0)
2218                 xlog_print_tic_res(log->l_mp, ticket);
2219
2220         index = 0;
2221         lv = log_vector;
2222         vecp = lv->lv_iovecp;
2223         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2224                 void            *ptr;
2225                 int             log_offset;
2226
2227                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2228                                                    &contwr, &log_offset);
2229                 if (error)
2230                         return error;
2231
2232                 ASSERT(log_offset <= iclog->ic_size - 1);
2233                 ptr = iclog->ic_datap + log_offset;
2234
2235                 /* start_lsn is the first lsn written to. That's all we need. */
2236                 if (!*start_lsn)
2237                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2238
2239                 /*
2240                  * This loop writes out as many regions as can fit in the amount
2241                  * of space which was allocated by xlog_state_get_iclog_space().
2242                  */
2243                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2244                         struct xfs_log_iovec    *reg;
2245                         struct xlog_op_header   *ophdr;
2246                         int                     start_rec_copy;
2247                         int                     copy_len;
2248                         int                     copy_off;
2249                         bool                    ordered = false;
2250
2251                         /* ordered log vectors have no regions to write */
2252                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2253                                 ASSERT(lv->lv_niovecs == 0);
2254                                 ordered = true;
2255                                 goto next_lv;
2256                         }
2257
2258                         reg = &vecp[index];
2259                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2260                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2261
2262                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2263                         if (start_rec_copy) {
2264                                 record_cnt++;
2265                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2266                                                    start_rec_copy);
2267                         }
2268
2269                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2270                         if (!ophdr)
2271                                 return XFS_ERROR(EIO);
2272
2273                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2274                                            sizeof(struct xlog_op_header));
2275
2276                         len += xlog_write_setup_copy(ticket, ophdr,
2277                                                      iclog->ic_size-log_offset,
2278                                                      reg->i_len,
2279                                                      &copy_off, &copy_len,
2280                                                      &partial_copy,
2281                                                      &partial_copy_len);
2282                         xlog_verify_dest_ptr(log, ptr);
2283
2284                         /* copy region */
2285                         ASSERT(copy_len >= 0);
2286                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
2287                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2288
2289                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2290                         record_cnt++;
2291                         data_cnt += contwr ? copy_len : 0;
2292
2293                         error = xlog_write_copy_finish(log, iclog, flags,
2294                                                        &record_cnt, &data_cnt,
2295                                                        &partial_copy,
2296                                                        &partial_copy_len,
2297                                                        log_offset,
2298                                                        commit_iclog);
2299                         if (error)
2300                                 return error;
2301
2302                         /*
2303                          * if we had a partial copy, we need to get more iclog
2304                          * space but we don't want to increment the region
2305                          * index because there is still more is this region to
2306                          * write.
2307                          *
2308                          * If we completed writing this region, and we flushed
2309                          * the iclog (indicated by resetting of the record
2310                          * count), then we also need to get more log space. If
2311                          * this was the last record, though, we are done and
2312                          * can just return.
2313                          */
2314                         if (partial_copy)
2315                                 break;
2316
2317                         if (++index == lv->lv_niovecs) {
2318 next_lv:
2319                                 lv = lv->lv_next;
2320                                 index = 0;
2321                                 if (lv)
2322                                         vecp = lv->lv_iovecp;
2323                         }
2324                         if (record_cnt == 0 && ordered == false) {
2325                                 if (!lv)
2326                                         return 0;
2327                                 break;
2328                         }
2329                 }
2330         }
2331
2332         ASSERT(len == 0);
2333
2334         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2335         if (!commit_iclog)
2336                 return xlog_state_release_iclog(log, iclog);
2337
2338         ASSERT(flags & XLOG_COMMIT_TRANS);
2339         *commit_iclog = iclog;
2340         return 0;
2341 }
2342
2343
2344 /*****************************************************************************
2345  *
2346  *              State Machine functions
2347  *
2348  *****************************************************************************
2349  */
2350
2351 /* Clean iclogs starting from the head.  This ordering must be
2352  * maintained, so an iclog doesn't become ACTIVE beyond one that
2353  * is SYNCING.  This is also required to maintain the notion that we use
2354  * a ordered wait queue to hold off would be writers to the log when every
2355  * iclog is trying to sync to disk.
2356  *
2357  * State Change: DIRTY -> ACTIVE
2358  */
2359 STATIC void
2360 xlog_state_clean_log(
2361         struct xlog *log)
2362 {
2363         xlog_in_core_t  *iclog;
2364         int changed = 0;
2365
2366         iclog = log->l_iclog;
2367         do {
2368                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2369                         iclog->ic_state = XLOG_STATE_ACTIVE;
2370                         iclog->ic_offset       = 0;
2371                         ASSERT(iclog->ic_callback == NULL);
2372                         /*
2373                          * If the number of ops in this iclog indicate it just
2374                          * contains the dummy transaction, we can
2375                          * change state into IDLE (the second time around).
2376                          * Otherwise we should change the state into
2377                          * NEED a dummy.
2378                          * We don't need to cover the dummy.
2379                          */
2380                         if (!changed &&
2381                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2382                                         XLOG_COVER_OPS)) {
2383                                 changed = 1;
2384                         } else {
2385                                 /*
2386                                  * We have two dirty iclogs so start over
2387                                  * This could also be num of ops indicates
2388                                  * this is not the dummy going out.
2389                                  */
2390                                 changed = 2;
2391                         }
2392                         iclog->ic_header.h_num_logops = 0;
2393                         memset(iclog->ic_header.h_cycle_data, 0,
2394                               sizeof(iclog->ic_header.h_cycle_data));
2395                         iclog->ic_header.h_lsn = 0;
2396                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2397                         /* do nothing */;
2398                 else
2399                         break;  /* stop cleaning */
2400                 iclog = iclog->ic_next;
2401         } while (iclog != log->l_iclog);
2402
2403         /* log is locked when we are called */
2404         /*
2405          * Change state for the dummy log recording.
2406          * We usually go to NEED. But we go to NEED2 if the changed indicates
2407          * we are done writing the dummy record.
2408          * If we are done with the second dummy recored (DONE2), then
2409          * we go to IDLE.
2410          */
2411         if (changed) {
2412                 switch (log->l_covered_state) {
2413                 case XLOG_STATE_COVER_IDLE:
2414                 case XLOG_STATE_COVER_NEED:
2415                 case XLOG_STATE_COVER_NEED2:
2416                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2417                         break;
2418
2419                 case XLOG_STATE_COVER_DONE:
2420                         if (changed == 1)
2421                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2422                         else
2423                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2424                         break;
2425
2426                 case XLOG_STATE_COVER_DONE2:
2427                         if (changed == 1)
2428                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2429                         else
2430                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2431                         break;
2432
2433                 default:
2434                         ASSERT(0);
2435                 }
2436         }
2437 }       /* xlog_state_clean_log */
2438
2439 STATIC xfs_lsn_t
2440 xlog_get_lowest_lsn(
2441         struct xlog     *log)
2442 {
2443         xlog_in_core_t  *lsn_log;
2444         xfs_lsn_t       lowest_lsn, lsn;
2445
2446         lsn_log = log->l_iclog;
2447         lowest_lsn = 0;
2448         do {
2449             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2450                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2451                 if ((lsn && !lowest_lsn) ||
2452                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2453                         lowest_lsn = lsn;
2454                 }
2455             }
2456             lsn_log = lsn_log->ic_next;
2457         } while (lsn_log != log->l_iclog);
2458         return lowest_lsn;
2459 }
2460
2461
2462 STATIC void
2463 xlog_state_do_callback(
2464         struct xlog             *log,
2465         int                     aborted,
2466         struct xlog_in_core     *ciclog)
2467 {
2468         xlog_in_core_t     *iclog;
2469         xlog_in_core_t     *first_iclog;        /* used to know when we've
2470                                                  * processed all iclogs once */
2471         xfs_log_callback_t *cb, *cb_next;
2472         int                flushcnt = 0;
2473         xfs_lsn_t          lowest_lsn;
2474         int                ioerrors;    /* counter: iclogs with errors */
2475         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2476         int                funcdidcallbacks; /* flag: function did callbacks */
2477         int                repeats;     /* for issuing console warnings if
2478                                          * looping too many times */
2479         int                wake = 0;
2480
2481         spin_lock(&log->l_icloglock);
2482         first_iclog = iclog = log->l_iclog;
2483         ioerrors = 0;
2484         funcdidcallbacks = 0;
2485         repeats = 0;
2486
2487         do {
2488                 /*
2489                  * Scan all iclogs starting with the one pointed to by the
2490                  * log.  Reset this starting point each time the log is
2491                  * unlocked (during callbacks).
2492                  *
2493                  * Keep looping through iclogs until one full pass is made
2494                  * without running any callbacks.
2495                  */
2496                 first_iclog = log->l_iclog;
2497                 iclog = log->l_iclog;
2498                 loopdidcallbacks = 0;
2499                 repeats++;
2500
2501                 do {
2502
2503                         /* skip all iclogs in the ACTIVE & DIRTY states */
2504                         if (iclog->ic_state &
2505                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2506                                 iclog = iclog->ic_next;
2507                                 continue;
2508                         }
2509
2510                         /*
2511                          * Between marking a filesystem SHUTDOWN and stopping
2512                          * the log, we do flush all iclogs to disk (if there
2513                          * wasn't a log I/O error). So, we do want things to
2514                          * go smoothly in case of just a SHUTDOWN  w/o a
2515                          * LOG_IO_ERROR.
2516                          */
2517                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2518                                 /*
2519                                  * Can only perform callbacks in order.  Since
2520                                  * this iclog is not in the DONE_SYNC/
2521                                  * DO_CALLBACK state, we skip the rest and
2522                                  * just try to clean up.  If we set our iclog
2523                                  * to DO_CALLBACK, we will not process it when
2524                                  * we retry since a previous iclog is in the
2525                                  * CALLBACK and the state cannot change since
2526                                  * we are holding the l_icloglock.
2527                                  */
2528                                 if (!(iclog->ic_state &
2529                                         (XLOG_STATE_DONE_SYNC |
2530                                                  XLOG_STATE_DO_CALLBACK))) {
2531                                         if (ciclog && (ciclog->ic_state ==
2532                                                         XLOG_STATE_DONE_SYNC)) {
2533                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2534                                         }
2535                                         break;
2536                                 }
2537                                 /*
2538                                  * We now have an iclog that is in either the
2539                                  * DO_CALLBACK or DONE_SYNC states. The other
2540                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2541                                  * caught by the above if and are going to
2542                                  * clean (i.e. we aren't doing their callbacks)
2543                                  * see the above if.
2544                                  */
2545
2546                                 /*
2547                                  * We will do one more check here to see if we
2548                                  * have chased our tail around.
2549                                  */
2550
2551                                 lowest_lsn = xlog_get_lowest_lsn(log);
2552                                 if (lowest_lsn &&
2553                                     XFS_LSN_CMP(lowest_lsn,
2554                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2555                                         iclog = iclog->ic_next;
2556                                         continue; /* Leave this iclog for
2557                                                    * another thread */
2558                                 }
2559
2560                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2561
2562
2563                                 /*
2564                                  * Completion of a iclog IO does not imply that
2565                                  * a transaction has completed, as transactions
2566                                  * can be large enough to span many iclogs. We
2567                                  * cannot change the tail of the log half way
2568                                  * through a transaction as this may be the only
2569                                  * transaction in the log and moving th etail to
2570                                  * point to the middle of it will prevent
2571                                  * recovery from finding the start of the
2572                                  * transaction. Hence we should only update the
2573                                  * last_sync_lsn if this iclog contains
2574                                  * transaction completion callbacks on it.
2575                                  *
2576                                  * We have to do this before we drop the
2577                                  * icloglock to ensure we are the only one that
2578                                  * can update it.
2579                                  */
2580                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2581                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2582                                 if (iclog->ic_callback)
2583                                         atomic64_set(&log->l_last_sync_lsn,
2584                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2585
2586                         } else
2587                                 ioerrors++;
2588
2589                         spin_unlock(&log->l_icloglock);
2590
2591                         /*
2592                          * Keep processing entries in the callback list until
2593                          * we come around and it is empty.  We need to
2594                          * atomically see that the list is empty and change the
2595                          * state to DIRTY so that we don't miss any more
2596                          * callbacks being added.
2597                          */
2598                         spin_lock(&iclog->ic_callback_lock);
2599                         cb = iclog->ic_callback;
2600                         while (cb) {
2601                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2602                                 iclog->ic_callback = NULL;
2603                                 spin_unlock(&iclog->ic_callback_lock);
2604
2605                                 /* perform callbacks in the order given */
2606                                 for (; cb; cb = cb_next) {
2607                                         cb_next = cb->cb_next;
2608                                         cb->cb_func(cb->cb_arg, aborted);
2609                                 }
2610                                 spin_lock(&iclog->ic_callback_lock);
2611                                 cb = iclog->ic_callback;
2612                         }
2613
2614                         loopdidcallbacks++;
2615                         funcdidcallbacks++;
2616
2617                         spin_lock(&log->l_icloglock);
2618                         ASSERT(iclog->ic_callback == NULL);
2619                         spin_unlock(&iclog->ic_callback_lock);
2620                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2621                                 iclog->ic_state = XLOG_STATE_DIRTY;
2622
2623                         /*
2624                          * Transition from DIRTY to ACTIVE if applicable.
2625                          * NOP if STATE_IOERROR.
2626                          */
2627                         xlog_state_clean_log(log);
2628
2629                         /* wake up threads waiting in xfs_log_force() */
2630                         wake_up_all(&iclog->ic_force_wait);
2631
2632                         iclog = iclog->ic_next;
2633                 } while (first_iclog != iclog);
2634
2635                 if (repeats > 5000) {
2636                         flushcnt += repeats;
2637                         repeats = 0;
2638                         xfs_warn(log->l_mp,
2639                                 "%s: possible infinite loop (%d iterations)",
2640                                 __func__, flushcnt);
2641                 }
2642         } while (!ioerrors && loopdidcallbacks);
2643
2644         /*
2645          * make one last gasp attempt to see if iclogs are being left in
2646          * limbo..
2647          */
2648 #ifdef DEBUG
2649         if (funcdidcallbacks) {
2650                 first_iclog = iclog = log->l_iclog;
2651                 do {
2652                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2653                         /*
2654                          * Terminate the loop if iclogs are found in states
2655                          * which will cause other threads to clean up iclogs.
2656                          *
2657                          * SYNCING - i/o completion will go through logs
2658                          * DONE_SYNC - interrupt thread should be waiting for
2659                          *              l_icloglock
2660                          * IOERROR - give up hope all ye who enter here
2661                          */
2662                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2663                             iclog->ic_state == XLOG_STATE_SYNCING ||
2664                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2665                             iclog->ic_state == XLOG_STATE_IOERROR )
2666                                 break;
2667                         iclog = iclog->ic_next;
2668                 } while (first_iclog != iclog);
2669         }
2670 #endif
2671
2672         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2673                 wake = 1;
2674         spin_unlock(&log->l_icloglock);
2675
2676         if (wake)
2677                 wake_up_all(&log->l_flush_wait);
2678 }
2679
2680
2681 /*
2682  * Finish transitioning this iclog to the dirty state.
2683  *
2684  * Make sure that we completely execute this routine only when this is
2685  * the last call to the iclog.  There is a good chance that iclog flushes,
2686  * when we reach the end of the physical log, get turned into 2 separate
2687  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2688  * routine.  By using the reference count bwritecnt, we guarantee that only
2689  * the second completion goes through.
2690  *
2691  * Callbacks could take time, so they are done outside the scope of the
2692  * global state machine log lock.
2693  */
2694 STATIC void
2695 xlog_state_done_syncing(
2696         xlog_in_core_t  *iclog,
2697         int             aborted)
2698 {
2699         struct xlog        *log = iclog->ic_log;
2700
2701         spin_lock(&log->l_icloglock);
2702
2703         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2704                iclog->ic_state == XLOG_STATE_IOERROR);
2705         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2706         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2707
2708
2709         /*
2710          * If we got an error, either on the first buffer, or in the case of
2711          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2712          * and none should ever be attempted to be written to disk
2713          * again.
2714          */
2715         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2716                 if (--iclog->ic_bwritecnt == 1) {
2717                         spin_unlock(&log->l_icloglock);
2718                         return;
2719                 }
2720                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2721         }
2722
2723         /*
2724          * Someone could be sleeping prior to writing out the next
2725          * iclog buffer, we wake them all, one will get to do the
2726          * I/O, the others get to wait for the result.
2727          */
2728         wake_up_all(&iclog->ic_write_wait);
2729         spin_unlock(&log->l_icloglock);
2730         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2731 }       /* xlog_state_done_syncing */
2732
2733
2734 /*
2735  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2736  * sleep.  We wait on the flush queue on the head iclog as that should be
2737  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2738  * we will wait here and all new writes will sleep until a sync completes.
2739  *
2740  * The in-core logs are used in a circular fashion. They are not used
2741  * out-of-order even when an iclog past the head is free.
2742  *
2743  * return:
2744  *      * log_offset where xlog_write() can start writing into the in-core
2745  *              log's data space.
2746  *      * in-core log pointer to which xlog_write() should write.
2747  *      * boolean indicating this is a continued write to an in-core log.
2748  *              If this is the last write, then the in-core log's offset field
2749  *              needs to be incremented, depending on the amount of data which
2750  *              is copied.
2751  */
2752 STATIC int
2753 xlog_state_get_iclog_space(
2754         struct xlog             *log,
2755         int                     len,
2756         struct xlog_in_core     **iclogp,
2757         struct xlog_ticket      *ticket,
2758         int                     *continued_write,
2759         int                     *logoffsetp)
2760 {
2761         int               log_offset;
2762         xlog_rec_header_t *head;
2763         xlog_in_core_t    *iclog;
2764         int               error;
2765
2766 restart:
2767         spin_lock(&log->l_icloglock);
2768         if (XLOG_FORCED_SHUTDOWN(log)) {
2769                 spin_unlock(&log->l_icloglock);
2770                 return XFS_ERROR(EIO);
2771         }
2772
2773         iclog = log->l_iclog;
2774         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2775                 XFS_STATS_INC(xs_log_noiclogs);
2776
2777                 /* Wait for log writes to have flushed */
2778                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2779                 goto restart;
2780         }
2781
2782         head = &iclog->ic_header;
2783
2784         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2785         log_offset = iclog->ic_offset;
2786
2787         /* On the 1st write to an iclog, figure out lsn.  This works
2788          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2789          * committing to.  If the offset is set, that's how many blocks
2790          * must be written.
2791          */
2792         if (log_offset == 0) {
2793                 ticket->t_curr_res -= log->l_iclog_hsize;
2794                 xlog_tic_add_region(ticket,
2795                                     log->l_iclog_hsize,
2796                                     XLOG_REG_TYPE_LRHEADER);
2797                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2798                 head->h_lsn = cpu_to_be64(
2799                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2800                 ASSERT(log->l_curr_block >= 0);
2801         }
2802
2803         /* If there is enough room to write everything, then do it.  Otherwise,
2804          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2805          * bit is on, so this will get flushed out.  Don't update ic_offset
2806          * until you know exactly how many bytes get copied.  Therefore, wait
2807          * until later to update ic_offset.
2808          *
2809          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2810          * can fit into remaining data section.
2811          */
2812         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2813                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2814
2815                 /*
2816                  * If I'm the only one writing to this iclog, sync it to disk.
2817                  * We need to do an atomic compare and decrement here to avoid
2818                  * racing with concurrent atomic_dec_and_lock() calls in
2819                  * xlog_state_release_iclog() when there is more than one
2820                  * reference to the iclog.
2821                  */
2822                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2823                         /* we are the only one */
2824                         spin_unlock(&log->l_icloglock);
2825                         error = xlog_state_release_iclog(log, iclog);
2826                         if (error)
2827                                 return error;
2828                 } else {
2829                         spin_unlock(&log->l_icloglock);
2830                 }
2831                 goto restart;
2832         }
2833
2834         /* Do we have enough room to write the full amount in the remainder
2835          * of this iclog?  Or must we continue a write on the next iclog and
2836          * mark this iclog as completely taken?  In the case where we switch
2837          * iclogs (to mark it taken), this particular iclog will release/sync
2838          * to disk in xlog_write().
2839          */
2840         if (len <= iclog->ic_size - iclog->ic_offset) {
2841                 *continued_write = 0;
2842                 iclog->ic_offset += len;
2843         } else {
2844                 *continued_write = 1;
2845                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2846         }
2847         *iclogp = iclog;
2848
2849         ASSERT(iclog->ic_offset <= iclog->ic_size);
2850         spin_unlock(&log->l_icloglock);
2851
2852         *logoffsetp = log_offset;
2853         return 0;
2854 }       /* xlog_state_get_iclog_space */
2855
2856 /* The first cnt-1 times through here we don't need to
2857  * move the grant write head because the permanent
2858  * reservation has reserved cnt times the unit amount.
2859  * Release part of current permanent unit reservation and
2860  * reset current reservation to be one units worth.  Also
2861  * move grant reservation head forward.
2862  */
2863 STATIC void
2864 xlog_regrant_reserve_log_space(
2865         struct xlog             *log,
2866         struct xlog_ticket      *ticket)
2867 {
2868         trace_xfs_log_regrant_reserve_enter(log, ticket);
2869
2870         if (ticket->t_cnt > 0)
2871                 ticket->t_cnt--;
2872
2873         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2874                                         ticket->t_curr_res);
2875         xlog_grant_sub_space(log, &log->l_write_head.grant,
2876                                         ticket->t_curr_res);
2877         ticket->t_curr_res = ticket->t_unit_res;
2878         xlog_tic_reset_res(ticket);
2879
2880         trace_xfs_log_regrant_reserve_sub(log, ticket);
2881
2882         /* just return if we still have some of the pre-reserved space */
2883         if (ticket->t_cnt > 0)
2884                 return;
2885
2886         xlog_grant_add_space(log, &log->l_reserve_head.grant,
2887                                         ticket->t_unit_res);
2888
2889         trace_xfs_log_regrant_reserve_exit(log, ticket);
2890
2891         ticket->t_curr_res = ticket->t_unit_res;
2892         xlog_tic_reset_res(ticket);
2893 }       /* xlog_regrant_reserve_log_space */
2894
2895
2896 /*
2897  * Give back the space left from a reservation.
2898  *
2899  * All the information we need to make a correct determination of space left
2900  * is present.  For non-permanent reservations, things are quite easy.  The
2901  * count should have been decremented to zero.  We only need to deal with the
2902  * space remaining in the current reservation part of the ticket.  If the
2903  * ticket contains a permanent reservation, there may be left over space which
2904  * needs to be released.  A count of N means that N-1 refills of the current
2905  * reservation can be done before we need to ask for more space.  The first
2906  * one goes to fill up the first current reservation.  Once we run out of
2907  * space, the count will stay at zero and the only space remaining will be
2908  * in the current reservation field.
2909  */
2910 STATIC void
2911 xlog_ungrant_log_space(
2912         struct xlog             *log,
2913         struct xlog_ticket      *ticket)
2914 {
2915         int     bytes;
2916
2917         if (ticket->t_cnt > 0)
2918                 ticket->t_cnt--;
2919
2920         trace_xfs_log_ungrant_enter(log, ticket);
2921         trace_xfs_log_ungrant_sub(log, ticket);
2922
2923         /*
2924          * If this is a permanent reservation ticket, we may be able to free
2925          * up more space based on the remaining count.
2926          */
2927         bytes = ticket->t_curr_res;
2928         if (ticket->t_cnt > 0) {
2929                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2930                 bytes += ticket->t_unit_res*ticket->t_cnt;
2931         }
2932
2933         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2934         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2935
2936         trace_xfs_log_ungrant_exit(log, ticket);
2937
2938         xfs_log_space_wake(log->l_mp);
2939 }
2940
2941 /*
2942  * Flush iclog to disk if this is the last reference to the given iclog and
2943  * the WANT_SYNC bit is set.
2944  *
2945  * When this function is entered, the iclog is not necessarily in the
2946  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2947  *
2948  *
2949  */
2950 STATIC int
2951 xlog_state_release_iclog(
2952         struct xlog             *log,
2953         struct xlog_in_core     *iclog)
2954 {
2955         int             sync = 0;       /* do we sync? */
2956
2957         if (iclog->ic_state & XLOG_STATE_IOERROR)
2958                 return XFS_ERROR(EIO);
2959
2960         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2961         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2962                 return 0;
2963
2964         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2965                 spin_unlock(&log->l_icloglock);
2966                 return XFS_ERROR(EIO);
2967         }
2968         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2969                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2970
2971         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2972                 /* update tail before writing to iclog */
2973                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2974                 sync++;
2975                 iclog->ic_state = XLOG_STATE_SYNCING;
2976                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2977                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2978                 /* cycle incremented when incrementing curr_block */
2979         }
2980         spin_unlock(&log->l_icloglock);
2981
2982         /*
2983          * We let the log lock go, so it's possible that we hit a log I/O
2984          * error or some other SHUTDOWN condition that marks the iclog
2985          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2986          * this iclog has consistent data, so we ignore IOERROR
2987          * flags after this point.
2988          */
2989         if (sync)
2990                 return xlog_sync(log, iclog);
2991         return 0;
2992 }       /* xlog_state_release_iclog */
2993
2994
2995 /*
2996  * This routine will mark the current iclog in the ring as WANT_SYNC
2997  * and move the current iclog pointer to the next iclog in the ring.
2998  * When this routine is called from xlog_state_get_iclog_space(), the
2999  * exact size of the iclog has not yet been determined.  All we know is
3000  * that every data block.  We have run out of space in this log record.
3001  */
3002 STATIC void
3003 xlog_state_switch_iclogs(
3004         struct xlog             *log,
3005         struct xlog_in_core     *iclog,
3006         int                     eventual_size)
3007 {
3008         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3009         if (!eventual_size)
3010                 eventual_size = iclog->ic_offset;
3011         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3012         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3013         log->l_prev_block = log->l_curr_block;
3014         log->l_prev_cycle = log->l_curr_cycle;
3015
3016         /* roll log?: ic_offset changed later */
3017         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3018
3019         /* Round up to next log-sunit */
3020         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3021             log->l_mp->m_sb.sb_logsunit > 1) {
3022                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3023                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3024         }
3025
3026         if (log->l_curr_block >= log->l_logBBsize) {
3027                 log->l_curr_cycle++;
3028                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3029                         log->l_curr_cycle++;
3030                 log->l_curr_block -= log->l_logBBsize;
3031                 ASSERT(log->l_curr_block >= 0);
3032         }
3033         ASSERT(iclog == log->l_iclog);
3034         log->l_iclog = iclog->ic_next;
3035 }       /* xlog_state_switch_iclogs */
3036
3037 /*
3038  * Write out all data in the in-core log as of this exact moment in time.
3039  *
3040  * Data may be written to the in-core log during this call.  However,
3041  * we don't guarantee this data will be written out.  A change from past
3042  * implementation means this routine will *not* write out zero length LRs.
3043  *
3044  * Basically, we try and perform an intelligent scan of the in-core logs.
3045  * If we determine there is no flushable data, we just return.  There is no
3046  * flushable data if:
3047  *
3048  *      1. the current iclog is active and has no data; the previous iclog
3049  *              is in the active or dirty state.
3050  *      2. the current iclog is drity, and the previous iclog is in the
3051  *              active or dirty state.
3052  *
3053  * We may sleep if:
3054  *
3055  *      1. the current iclog is not in the active nor dirty state.
3056  *      2. the current iclog dirty, and the previous iclog is not in the
3057  *              active nor dirty state.
3058  *      3. the current iclog is active, and there is another thread writing
3059  *              to this particular iclog.
3060  *      4. a) the current iclog is active and has no other writers
3061  *         b) when we return from flushing out this iclog, it is still
3062  *              not in the active nor dirty state.
3063  */
3064 int
3065 _xfs_log_force(
3066         struct xfs_mount        *mp,
3067         uint                    flags,
3068         int                     *log_flushed)
3069 {
3070         struct xlog             *log = mp->m_log;
3071         struct xlog_in_core     *iclog;
3072         xfs_lsn_t               lsn;
3073
3074         XFS_STATS_INC(xs_log_force);
3075
3076         xlog_cil_force(log);
3077
3078         spin_lock(&log->l_icloglock);
3079
3080         iclog = log->l_iclog;
3081         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3082                 spin_unlock(&log->l_icloglock);
3083                 return XFS_ERROR(EIO);
3084         }
3085
3086         /* If the head iclog is not active nor dirty, we just attach
3087          * ourselves to the head and go to sleep.
3088          */
3089         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3090             iclog->ic_state == XLOG_STATE_DIRTY) {
3091                 /*
3092                  * If the head is dirty or (active and empty), then
3093                  * we need to look at the previous iclog.  If the previous
3094                  * iclog is active or dirty we are done.  There is nothing
3095                  * to sync out.  Otherwise, we attach ourselves to the
3096                  * previous iclog and go to sleep.
3097                  */
3098                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3099                     (atomic_read(&iclog->ic_refcnt) == 0
3100                      && iclog->ic_offset == 0)) {
3101                         iclog = iclog->ic_prev;
3102                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3103                             iclog->ic_state == XLOG_STATE_DIRTY)
3104                                 goto no_sleep;
3105                         else
3106                                 goto maybe_sleep;
3107                 } else {
3108                         if (atomic_read(&iclog->ic_refcnt) == 0) {
3109                                 /* We are the only one with access to this
3110                                  * iclog.  Flush it out now.  There should
3111                                  * be a roundoff of zero to show that someone
3112                                  * has already taken care of the roundoff from
3113                                  * the previous sync.
3114                                  */
3115                                 atomic_inc(&iclog->ic_refcnt);
3116                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3117                                 xlog_state_switch_iclogs(log, iclog, 0);
3118                                 spin_unlock(&log->l_icloglock);
3119
3120                                 if (xlog_state_release_iclog(log, iclog))
3121                                         return XFS_ERROR(EIO);
3122
3123                                 if (log_flushed)
3124                                         *log_flushed = 1;
3125                                 spin_lock(&log->l_icloglock);
3126                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3127                                     iclog->ic_state != XLOG_STATE_DIRTY)
3128                                         goto maybe_sleep;
3129                                 else
3130                                         goto no_sleep;
3131                         } else {
3132                                 /* Someone else is writing to this iclog.
3133                                  * Use its call to flush out the data.  However,
3134                                  * the other thread may not force out this LR,
3135                                  * so we mark it WANT_SYNC.
3136                                  */
3137                                 xlog_state_switch_iclogs(log, iclog, 0);
3138                                 goto maybe_sleep;
3139                         }
3140                 }
3141         }
3142
3143         /* By the time we come around again, the iclog could've been filled
3144          * which would give it another lsn.  If we have a new lsn, just
3145          * return because the relevant data has been flushed.
3146          */
3147 maybe_sleep:
3148         if (flags & XFS_LOG_SYNC) {
3149                 /*
3150                  * We must check if we're shutting down here, before
3151                  * we wait, while we're holding the l_icloglock.
3152                  * Then we check again after waking up, in case our
3153                  * sleep was disturbed by a bad news.
3154                  */
3155                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3156                         spin_unlock(&log->l_icloglock);
3157                         return XFS_ERROR(EIO);
3158                 }
3159                 XFS_STATS_INC(xs_log_force_sleep);
3160                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3161                 /*
3162                  * No need to grab the log lock here since we're
3163                  * only deciding whether or not to return EIO
3164                  * and the memory read should be atomic.
3165                  */
3166                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3167                         return XFS_ERROR(EIO);
3168                 if (log_flushed)
3169                         *log_flushed = 1;
3170         } else {
3171
3172 no_sleep:
3173                 spin_unlock(&log->l_icloglock);
3174         }
3175         return 0;
3176 }
3177
3178 /*
3179  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3180  * about errors or whether the log was flushed or not. This is the normal
3181  * interface to use when trying to unpin items or move the log forward.
3182  */
3183 void
3184 xfs_log_force(
3185         xfs_mount_t     *mp,
3186         uint            flags)
3187 {
3188         int     error;
3189
3190         trace_xfs_log_force(mp, 0);
3191         error = _xfs_log_force(mp, flags, NULL);
3192         if (error)
3193                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3194 }
3195
3196 /*
3197  * Force the in-core log to disk for a specific LSN.
3198  *
3199  * Find in-core log with lsn.
3200  *      If it is in the DIRTY state, just return.
3201  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3202  *              state and go to sleep or return.
3203  *      If it is in any other state, go to sleep or return.
3204  *
3205  * Synchronous forces are implemented with a signal variable. All callers
3206  * to force a given lsn to disk will wait on a the sv attached to the
3207  * specific in-core log.  When given in-core log finally completes its
3208  * write to disk, that thread will wake up all threads waiting on the
3209  * sv.
3210  */
3211 int
3212 _xfs_log_force_lsn(
3213         struct xfs_mount        *mp,
3214         xfs_lsn_t               lsn,
3215         uint                    flags,
3216         int                     *log_flushed)
3217 {
3218         struct xlog             *log = mp->m_log;
3219         struct xlog_in_core     *iclog;
3220         int                     already_slept = 0;
3221
3222         ASSERT(lsn != 0);
3223
3224         XFS_STATS_INC(xs_log_force);
3225
3226         lsn = xlog_cil_force_lsn(log, lsn);
3227         if (lsn == NULLCOMMITLSN)
3228                 return 0;
3229
3230 try_again:
3231         spin_lock(&log->l_icloglock);
3232         iclog = log->l_iclog;
3233         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3234                 spin_unlock(&log->l_icloglock);
3235                 return XFS_ERROR(EIO);
3236         }
3237
3238         do {
3239                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3240                         iclog = iclog->ic_next;
3241                         continue;
3242                 }
3243
3244                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3245                         spin_unlock(&log->l_icloglock);
3246                         return 0;
3247                 }
3248
3249                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3250                         /*
3251                          * We sleep here if we haven't already slept (e.g.
3252                          * this is the first time we've looked at the correct
3253                          * iclog buf) and the buffer before us is going to
3254                          * be sync'ed. The reason for this is that if we
3255                          * are doing sync transactions here, by waiting for
3256                          * the previous I/O to complete, we can allow a few
3257                          * more transactions into this iclog before we close
3258                          * it down.
3259                          *
3260                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3261                          * up the refcnt so we can release the log (which
3262                          * drops the ref count).  The state switch keeps new
3263                          * transaction commits from using this buffer.  When
3264                          * the current commits finish writing into the buffer,
3265                          * the refcount will drop to zero and the buffer will
3266                          * go out then.
3267                          */
3268                         if (!already_slept &&
3269                             (iclog->ic_prev->ic_state &
3270                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3271                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3272
3273                                 XFS_STATS_INC(xs_log_force_sleep);
3274
3275                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3276                                                         &log->l_icloglock);
3277                                 if (log_flushed)
3278                                         *log_flushed = 1;
3279                                 already_slept = 1;
3280                                 goto try_again;
3281                         }
3282                         atomic_inc(&iclog->ic_refcnt);
3283                         xlog_state_switch_iclogs(log, iclog, 0);
3284                         spin_unlock(&log->l_icloglock);
3285                         if (xlog_state_release_iclog(log, iclog))
3286                                 return XFS_ERROR(EIO);
3287                         if (log_flushed)
3288                                 *log_flushed = 1;
3289                         spin_lock(&log->l_icloglock);
3290                 }
3291
3292                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3293                     !(iclog->ic_state &
3294                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3295                         /*
3296                          * Don't wait on completion if we know that we've
3297                          * gotten a log write error.
3298                          */
3299                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3300                                 spin_unlock(&log->l_icloglock);
3301                                 return XFS_ERROR(EIO);
3302                         }
3303                         XFS_STATS_INC(xs_log_force_sleep);
3304                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3305                         /*
3306                          * No need to grab the log lock here since we're
3307                          * only deciding whether or not to return EIO
3308                          * and the memory read should be atomic.
3309                          */
3310                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3311                                 return XFS_ERROR(EIO);
3312
3313                         if (log_flushed)
3314                                 *log_flushed = 1;
3315                 } else {                /* just return */
3316                         spin_unlock(&log->l_icloglock);
3317                 }
3318
3319                 return 0;
3320         } while (iclog != log->l_iclog);
3321
3322         spin_unlock(&log->l_icloglock);
3323         return 0;
3324 }
3325
3326 /*
3327  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3328  * about errors or whether the log was flushed or not. This is the normal
3329  * interface to use when trying to unpin items or move the log forward.
3330  */
3331 void
3332 xfs_log_force_lsn(
3333         xfs_mount_t     *mp,
3334         xfs_lsn_t       lsn,
3335         uint            flags)
3336 {
3337         int     error;
3338
3339         trace_xfs_log_force(mp, lsn);
3340         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3341         if (error)
3342                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3343 }
3344
3345 /*
3346  * Called when we want to mark the current iclog as being ready to sync to
3347  * disk.
3348  */
3349 STATIC void
3350 xlog_state_want_sync(
3351         struct xlog             *log,
3352         struct xlog_in_core     *iclog)
3353 {
3354         assert_spin_locked(&log->l_icloglock);
3355
3356         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3357                 xlog_state_switch_iclogs(log, iclog, 0);
3358         } else {
3359                 ASSERT(iclog->ic_state &
3360                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3361         }
3362 }
3363
3364
3365 /*****************************************************************************
3366  *
3367  *              TICKET functions
3368  *
3369  *****************************************************************************
3370  */
3371
3372 /*
3373  * Free a used ticket when its refcount falls to zero.
3374  */
3375 void
3376 xfs_log_ticket_put(
3377         xlog_ticket_t   *ticket)
3378 {
3379         ASSERT(atomic_read(&ticket->t_ref) > 0);
3380         if (atomic_dec_and_test(&ticket->t_ref))
3381                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3382 }
3383
3384 xlog_ticket_t *
3385 xfs_log_ticket_get(
3386         xlog_ticket_t   *ticket)
3387 {
3388         ASSERT(atomic_read(&ticket->t_ref) > 0);
3389         atomic_inc(&ticket->t_ref);
3390         return ticket;
3391 }
3392
3393 /*
3394  * Allocate and initialise a new log ticket.
3395  */
3396 struct xlog_ticket *
3397 xlog_ticket_alloc(
3398         struct xlog     *log,
3399         int             unit_bytes,
3400         int             cnt,
3401         char            client,
3402         bool            permanent,
3403         xfs_km_flags_t  alloc_flags)
3404 {
3405         struct xlog_ticket *tic;
3406         uint            num_headers;
3407         int             iclog_space;
3408
3409         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3410         if (!tic)
3411                 return NULL;
3412
3413         /*
3414          * Permanent reservations have up to 'cnt'-1 active log operations
3415          * in the log.  A unit in this case is the amount of space for one
3416          * of these log operations.  Normal reservations have a cnt of 1
3417          * and their unit amount is the total amount of space required.
3418          *
3419          * The following lines of code account for non-transaction data
3420          * which occupy space in the on-disk log.
3421          *
3422          * Normal form of a transaction is:
3423          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3424          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3425          *
3426          * We need to account for all the leadup data and trailer data
3427          * around the transaction data.
3428          * And then we need to account for the worst case in terms of using
3429          * more space.
3430          * The worst case will happen if:
3431          * - the placement of the transaction happens to be such that the
3432          *   roundoff is at its maximum
3433          * - the transaction data is synced before the commit record is synced
3434          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3435          *   Therefore the commit record is in its own Log Record.
3436          *   This can happen as the commit record is called with its
3437          *   own region to xlog_write().
3438          *   This then means that in the worst case, roundoff can happen for
3439          *   the commit-rec as well.
3440          *   The commit-rec is smaller than padding in this scenario and so it is
3441          *   not added separately.
3442          */
3443
3444         /* for trans header */
3445         unit_bytes += sizeof(xlog_op_header_t);
3446         unit_bytes += sizeof(xfs_trans_header_t);
3447
3448         /* for start-rec */
3449         unit_bytes += sizeof(xlog_op_header_t);
3450
3451         /*
3452          * for LR headers - the space for data in an iclog is the size minus
3453          * the space used for the headers. If we use the iclog size, then we
3454          * undercalculate the number of headers required.
3455          *
3456          * Furthermore - the addition of op headers for split-recs might
3457          * increase the space required enough to require more log and op
3458          * headers, so take that into account too.
3459          *
3460          * IMPORTANT: This reservation makes the assumption that if this
3461          * transaction is the first in an iclog and hence has the LR headers
3462          * accounted to it, then the remaining space in the iclog is
3463          * exclusively for this transaction.  i.e. if the transaction is larger
3464          * than the iclog, it will be the only thing in that iclog.
3465          * Fundamentally, this means we must pass the entire log vector to
3466          * xlog_write to guarantee this.
3467          */
3468         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3469         num_headers = howmany(unit_bytes, iclog_space);
3470
3471         /* for split-recs - ophdrs added when data split over LRs */
3472         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3473
3474         /* add extra header reservations if we overrun */
3475         while (!num_headers ||
3476                howmany(unit_bytes, iclog_space) > num_headers) {
3477                 unit_bytes += sizeof(xlog_op_header_t);
3478                 num_headers++;
3479         }
3480         unit_bytes += log->l_iclog_hsize * num_headers;
3481
3482         /* for commit-rec LR header - note: padding will subsume the ophdr */
3483         unit_bytes += log->l_iclog_hsize;
3484
3485         /* for roundoff padding for transaction data and one for commit record */
3486         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3487             log->l_mp->m_sb.sb_logsunit > 1) {
3488                 /* log su roundoff */
3489                 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3490         } else {
3491                 /* BB roundoff */
3492                 unit_bytes += 2*BBSIZE;
3493         }
3494
3495         atomic_set(&tic->t_ref, 1);
3496         tic->t_task             = current;
3497         INIT_LIST_HEAD(&tic->t_queue);
3498         tic->t_unit_res         = unit_bytes;
3499         tic->t_curr_res         = unit_bytes;
3500         tic->t_cnt              = cnt;
3501         tic->t_ocnt             = cnt;
3502         tic->t_tid              = prandom_u32();
3503         tic->t_clientid         = client;
3504         tic->t_flags            = XLOG_TIC_INITED;
3505         tic->t_trans_type       = 0;
3506         if (permanent)
3507                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3508
3509         xlog_tic_reset_res(tic);
3510
3511         return tic;
3512 }
3513
3514
3515 /******************************************************************************
3516  *
3517  *              Log debug routines
3518  *
3519  ******************************************************************************
3520  */
3521 #if defined(DEBUG)
3522 /*
3523  * Make sure that the destination ptr is within the valid data region of
3524  * one of the iclogs.  This uses backup pointers stored in a different
3525  * part of the log in case we trash the log structure.
3526  */
3527 void
3528 xlog_verify_dest_ptr(
3529         struct xlog     *log,
3530         char            *ptr)
3531 {
3532         int i;
3533         int good_ptr = 0;
3534
3535         for (i = 0; i < log->l_iclog_bufs; i++) {
3536                 if (ptr >= log->l_iclog_bak[i] &&
3537                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3538                         good_ptr++;
3539         }
3540
3541         if (!good_ptr)
3542                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3543 }
3544
3545 /*
3546  * Check to make sure the grant write head didn't just over lap the tail.  If
3547  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3548  * the cycles differ by exactly one and check the byte count.
3549  *
3550  * This check is run unlocked, so can give false positives. Rather than assert
3551  * on failures, use a warn-once flag and a panic tag to allow the admin to
3552  * determine if they want to panic the machine when such an error occurs. For
3553  * debug kernels this will have the same effect as using an assert but, unlinke
3554  * an assert, it can be turned off at runtime.
3555  */
3556 STATIC void
3557 xlog_verify_grant_tail(
3558         struct xlog     *log)
3559 {
3560         int             tail_cycle, tail_blocks;
3561         int             cycle, space;
3562
3563         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3564         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3565         if (tail_cycle != cycle) {
3566                 if (cycle - 1 != tail_cycle &&
3567                     !(log->l_flags & XLOG_TAIL_WARN)) {
3568                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3569                                 "%s: cycle - 1 != tail_cycle", __func__);
3570                         log->l_flags |= XLOG_TAIL_WARN;
3571                 }
3572
3573                 if (space > BBTOB(tail_blocks) &&
3574                     !(log->l_flags & XLOG_TAIL_WARN)) {
3575                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3576                                 "%s: space > BBTOB(tail_blocks)", __func__);
3577                         log->l_flags |= XLOG_TAIL_WARN;
3578                 }
3579         }
3580 }
3581
3582 /* check if it will fit */
3583 STATIC void
3584 xlog_verify_tail_lsn(
3585         struct xlog             *log,
3586         struct xlog_in_core     *iclog,
3587         xfs_lsn_t               tail_lsn)
3588 {
3589     int blocks;
3590
3591     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3592         blocks =
3593             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3594         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3595                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3596     } else {
3597         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3598
3599         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3600                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3601
3602         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3603         if (blocks < BTOBB(iclog->ic_offset) + 1)
3604                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3605     }
3606 }       /* xlog_verify_tail_lsn */
3607
3608 /*
3609  * Perform a number of checks on the iclog before writing to disk.
3610  *
3611  * 1. Make sure the iclogs are still circular
3612  * 2. Make sure we have a good magic number
3613  * 3. Make sure we don't have magic numbers in the data
3614  * 4. Check fields of each log operation header for:
3615  *      A. Valid client identifier
3616  *      B. tid ptr value falls in valid ptr space (user space code)
3617  *      C. Length in log record header is correct according to the
3618  *              individual operation headers within record.
3619  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3620  *      log, check the preceding blocks of the physical log to make sure all
3621  *      the cycle numbers agree with the current cycle number.
3622  */
3623 STATIC void
3624 xlog_verify_iclog(
3625         struct xlog             *log,
3626         struct xlog_in_core     *iclog,
3627         int                     count,
3628         bool                    syncing)
3629 {
3630         xlog_op_header_t        *ophead;
3631         xlog_in_core_t          *icptr;
3632         xlog_in_core_2_t        *xhdr;
3633         xfs_caddr_t             ptr;
3634         xfs_caddr_t             base_ptr;
3635         __psint_t               field_offset;
3636         __uint8_t               clientid;
3637         int                     len, i, j, k, op_len;
3638         int                     idx;
3639
3640         /* check validity of iclog pointers */
3641         spin_lock(&log->l_icloglock);
3642         icptr = log->l_iclog;
3643         for (i=0; i < log->l_iclog_bufs; i++) {
3644                 if (icptr == NULL)
3645                         xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3646                 icptr = icptr->ic_next;
3647         }
3648         if (icptr != log->l_iclog)
3649                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3650         spin_unlock(&log->l_icloglock);
3651
3652         /* check log magic numbers */
3653         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3654                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3655
3656         ptr = (xfs_caddr_t) &iclog->ic_header;
3657         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3658              ptr += BBSIZE) {
3659                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3660                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3661                                 __func__);
3662         }
3663
3664         /* check fields */
3665         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3666         ptr = iclog->ic_datap;
3667         base_ptr = ptr;
3668         ophead = (xlog_op_header_t *)ptr;
3669         xhdr = iclog->ic_data;
3670         for (i = 0; i < len; i++) {
3671                 ophead = (xlog_op_header_t *)ptr;
3672
3673                 /* clientid is only 1 byte */
3674                 field_offset = (__psint_t)
3675                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3676                 if (!syncing || (field_offset & 0x1ff)) {
3677                         clientid = ophead->oh_clientid;
3678                 } else {
3679                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3680                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3681                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3682                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3683                                 clientid = xlog_get_client_id(
3684                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3685                         } else {
3686                                 clientid = xlog_get_client_id(
3687                                         iclog->ic_header.h_cycle_data[idx]);
3688                         }
3689                 }
3690                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3691                         xfs_warn(log->l_mp,
3692                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3693                                 __func__, clientid, ophead,
3694                                 (unsigned long)field_offset);
3695
3696                 /* check length */
3697                 field_offset = (__psint_t)
3698                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3699                 if (!syncing || (field_offset & 0x1ff)) {
3700                         op_len = be32_to_cpu(ophead->oh_len);
3701                 } else {
3702                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3703                                     (__psint_t)iclog->ic_datap);
3704                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3705                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3706                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3707                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3708                         } else {
3709                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3710                         }
3711                 }
3712                 ptr += sizeof(xlog_op_header_t) + op_len;
3713         }
3714 }       /* xlog_verify_iclog */
3715 #endif
3716
3717 /*
3718  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3719  */
3720 STATIC int
3721 xlog_state_ioerror(
3722         struct xlog     *log)
3723 {
3724         xlog_in_core_t  *iclog, *ic;
3725
3726         iclog = log->l_iclog;
3727         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3728                 /*
3729                  * Mark all the incore logs IOERROR.
3730                  * From now on, no log flushes will result.
3731                  */
3732                 ic = iclog;
3733                 do {
3734                         ic->ic_state = XLOG_STATE_IOERROR;
3735                         ic = ic->ic_next;
3736                 } while (ic != iclog);
3737                 return 0;
3738         }
3739         /*
3740          * Return non-zero, if state transition has already happened.
3741          */
3742         return 1;
3743 }
3744
3745 /*
3746  * This is called from xfs_force_shutdown, when we're forcibly
3747  * shutting down the filesystem, typically because of an IO error.
3748  * Our main objectives here are to make sure that:
3749  *      a. the filesystem gets marked 'SHUTDOWN' for all interested
3750  *         parties to find out, 'atomically'.
3751  *      b. those who're sleeping on log reservations, pinned objects and
3752  *          other resources get woken up, and be told the bad news.
3753  *      c. nothing new gets queued up after (a) and (b) are done.
3754  *      d. if !logerror, flush the iclogs to disk, then seal them off
3755  *         for business.
3756  *
3757  * Note: for delayed logging the !logerror case needs to flush the regions
3758  * held in memory out to the iclogs before flushing them to disk. This needs
3759  * to be done before the log is marked as shutdown, otherwise the flush to the
3760  * iclogs will fail.
3761  */
3762 int
3763 xfs_log_force_umount(
3764         struct xfs_mount        *mp,
3765         int                     logerror)
3766 {
3767         struct xlog     *log;
3768         int             retval;
3769
3770         log = mp->m_log;
3771
3772         /*
3773          * If this happens during log recovery, don't worry about
3774          * locking; the log isn't open for business yet.
3775          */
3776         if (!log ||
3777             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3778                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3779                 if (mp->m_sb_bp)
3780                         XFS_BUF_DONE(mp->m_sb_bp);
3781                 return 0;
3782         }
3783
3784         /*
3785          * Somebody could've already done the hard work for us.
3786          * No need to get locks for this.
3787          */
3788         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3789                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3790                 return 1;
3791         }
3792         retval = 0;
3793
3794         /*
3795          * Flush the in memory commit item list before marking the log as
3796          * being shut down. We need to do it in this order to ensure all the
3797          * completed transactions are flushed to disk with the xfs_log_force()
3798          * call below.
3799          */
3800         if (!logerror)
3801                 xlog_cil_force(log);
3802
3803         /*
3804          * mark the filesystem and the as in a shutdown state and wake
3805          * everybody up to tell them the bad news.
3806          */
3807         spin_lock(&log->l_icloglock);
3808         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3809         if (mp->m_sb_bp)
3810                 XFS_BUF_DONE(mp->m_sb_bp);
3811
3812         /*
3813          * This flag is sort of redundant because of the mount flag, but
3814          * it's good to maintain the separation between the log and the rest
3815          * of XFS.
3816          */
3817         log->l_flags |= XLOG_IO_ERROR;
3818
3819         /*
3820          * If we hit a log error, we want to mark all the iclogs IOERROR
3821          * while we're still holding the loglock.
3822          */
3823         if (logerror)
3824                 retval = xlog_state_ioerror(log);
3825         spin_unlock(&log->l_icloglock);
3826
3827         /*
3828          * We don't want anybody waiting for log reservations after this. That
3829          * means we have to wake up everybody queued up on reserveq as well as
3830          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3831          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3832          * action is protected by the grant locks.
3833          */
3834         xlog_grant_head_wake_all(&log->l_reserve_head);
3835         xlog_grant_head_wake_all(&log->l_write_head);
3836
3837         if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3838                 ASSERT(!logerror);
3839                 /*
3840                  * Force the incore logs to disk before shutting the
3841                  * log down completely.
3842                  */
3843                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3844
3845                 spin_lock(&log->l_icloglock);
3846                 retval = xlog_state_ioerror(log);
3847                 spin_unlock(&log->l_icloglock);
3848         }
3849         /*
3850          * Wake up everybody waiting on xfs_log_force.
3851          * Callback all log item committed functions as if the
3852          * log writes were completed.
3853          */
3854         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3855
3856 #ifdef XFSERRORDEBUG
3857         {
3858                 xlog_in_core_t  *iclog;
3859
3860                 spin_lock(&log->l_icloglock);
3861                 iclog = log->l_iclog;
3862                 do {
3863                         ASSERT(iclog->ic_callback == 0);
3864                         iclog = iclog->ic_next;
3865                 } while (iclog != log->l_iclog);
3866                 spin_unlock(&log->l_icloglock);
3867         }
3868 #endif
3869         /* return non-zero if log IOERROR transition had already happened */
3870         return retval;
3871 }
3872
3873 STATIC int
3874 xlog_iclogs_empty(
3875         struct xlog     *log)
3876 {
3877         xlog_in_core_t  *iclog;
3878
3879         iclog = log->l_iclog;
3880         do {
3881                 /* endianness does not matter here, zero is zero in
3882                  * any language.
3883                  */
3884                 if (iclog->ic_header.h_num_logops)
3885                         return 0;
3886                 iclog = iclog->ic_next;
3887         } while (iclog != log->l_iclog);
3888         return 1;
3889 }
3890