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_shared.h"
  21 #include "xfs_format.h"
  22 #include "xfs_log_format.h"
  23 #include "xfs_trans_resv.h"
  24 #include "xfs_mount.h"
  25 #include "xfs_errortag.h"
  26 #include "xfs_error.h"
  27 #include "xfs_trans.h"
  28 #include "xfs_trans_priv.h"
  29 #include "xfs_log.h"
  30 #include "xfs_log_priv.h"
  31 #include "xfs_log_recover.h"
  32 #include "xfs_inode.h"
  33 #include "xfs_trace.h"
  34 #include "xfs_fsops.h"
  35 #include "xfs_cksum.h"
  36 #include "xfs_sysfs.h"
  37 #include "xfs_sb.h"
  38
  39 kmem_zone_t     *xfs_log_ticket_zone;
  40
  41 /* Local miscellaneous function prototypes */
  42 STATIC int
  43 xlog_commit_record(
  44         struct xlog             *log,
  45         struct xlog_ticket      *ticket,
  46         struct xlog_in_core     **iclog,
  47         xfs_lsn_t               *commitlsnp);
  48
  49 STATIC struct xlog *
  50 xlog_alloc_log(
  51         struct xfs_mount        *mp,
  52         struct xfs_buftarg      *log_target,
  53         xfs_daddr_t             blk_offset,
  54         int                     num_bblks);
  55 STATIC int
  56 xlog_space_left(
  57         struct xlog             *log,
  58         atomic64_t              *head);
  59 STATIC int
  60 xlog_sync(
  61         struct xlog             *log,
  62         struct xlog_in_core     *iclog);
  63 STATIC void
  64 xlog_dealloc_log(
  65         struct xlog             *log);
  66
  67 /* local state machine functions */
  68 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  69 STATIC void
  70 xlog_state_do_callback(
  71         struct xlog             *log,
  72         int                     aborted,
  73         struct xlog_in_core     *iclog);
  74 STATIC int
  75 xlog_state_get_iclog_space(
  76         struct xlog             *log,
  77         int                     len,
  78         struct xlog_in_core     **iclog,
  79         struct xlog_ticket      *ticket,
  80         int                     *continued_write,
  81         int                     *logoffsetp);
  82 STATIC int
  83 xlog_state_release_iclog(
  84         struct xlog             *log,
  85         struct xlog_in_core     *iclog);
  86 STATIC void
  87 xlog_state_switch_iclogs(
  88         struct xlog             *log,
  89         struct xlog_in_core     *iclog,
  90         int                     eventual_size);
  91 STATIC void
  92 xlog_state_want_sync(
  93         struct xlog             *log,
  94         struct xlog_in_core     *iclog);
  95
  96 STATIC void
  97 xlog_grant_push_ail(
  98         struct xlog             *log,
  99         int                     need_bytes);
 100 STATIC void
 101 xlog_regrant_reserve_log_space(
 102         struct xlog             *log,
 103         struct xlog_ticket      *ticket);
 104 STATIC void
 105 xlog_ungrant_log_space(
 106         struct xlog             *log,
 107         struct xlog_ticket      *ticket);
 108
 109 #if defined(DEBUG)
 110 STATIC void
 111 xlog_verify_dest_ptr(
 112         struct xlog             *log,
 113         void                    *ptr);
 114 STATIC void
 115 xlog_verify_grant_tail(
 116         struct xlog *log);
 117 STATIC void
 118 xlog_verify_iclog(
 119         struct xlog             *log,
 120         struct xlog_in_core     *iclog,
 121         int                     count,
 122         bool                    syncing);
 123 STATIC void
 124 xlog_verify_tail_lsn(
 125         struct xlog             *log,
 126         struct xlog_in_core     *iclog,
 127         xfs_lsn_t               tail_lsn);
 128 #else
 129 #define xlog_verify_dest_ptr(a,b)
 130 #define xlog_verify_grant_tail(a)
 131 #define xlog_verify_iclog(a,b,c,d)
 132 #define xlog_verify_tail_lsn(a,b,c)
 133 #endif
 134
 135 STATIC int
 136 xlog_iclogs_empty(
 137         struct xlog             *log);
 138
 139 static void
 140 xlog_grant_sub_space(
 141         struct xlog             *log,
 142         atomic64_t              *head,
 143         int                     bytes)
 144 {
 145         int64_t head_val = atomic64_read(head);
 146         int64_t new, old;
 147
 148         do {
 149                 int     cycle, space;
 150
 151                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 152
 153                 space -= bytes;
 154                 if (space < 0) {
 155                         space += log->l_logsize;
 156                         cycle--;
 157                 }
 158
 159                 old = head_val;
 160                 new = xlog_assign_grant_head_val(cycle, space);
 161                 head_val = atomic64_cmpxchg(head, old, new);
 162         } while (head_val != old);
 163 }
 164
 165 static void
 166 xlog_grant_add_space(
 167         struct xlog             *log,
 168         atomic64_t              *head,
 169         int                     bytes)
 170 {
 171         int64_t head_val = atomic64_read(head);
 172         int64_t new, old;
 173
 174         do {
 175                 int             tmp;
 176                 int             cycle, space;
 177
 178                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 179
 180                 tmp = log->l_logsize - space;
 181                 if (tmp > bytes)
 182                         space += bytes;
 183                 else {
 184                         space = bytes - tmp;
 185                         cycle++;
 186                 }
 187
 188                 old = head_val;
 189                 new = xlog_assign_grant_head_val(cycle, space);
 190                 head_val = atomic64_cmpxchg(head, old, new);
 191         } while (head_val != old);
 192 }
 193
 194 STATIC void
 195 xlog_grant_head_init(
 196         struct xlog_grant_head  *head)
 197 {
 198         xlog_assign_grant_head(&head->grant, 1, 0);
 199         INIT_LIST_HEAD(&head->waiters);
 200         spin_lock_init(&head->lock);
 201 }
 202
 203 STATIC void
 204 xlog_grant_head_wake_all(
 205         struct xlog_grant_head  *head)
 206 {
 207         struct xlog_ticket      *tic;
 208
 209         spin_lock(&head->lock);
 210         list_for_each_entry(tic, &head->waiters, t_queue)
 211                 wake_up_process(tic->t_task);
 212         spin_unlock(&head->lock);
 213 }
 214
 215 static inline int
 216 xlog_ticket_reservation(
 217         struct xlog             *log,
 218         struct xlog_grant_head  *head,
 219         struct xlog_ticket      *tic)
 220 {
 221         if (head == &log->l_write_head) {
 222                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 223                 return tic->t_unit_res;
 224         } else {
 225                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 226                         return tic->t_unit_res * tic->t_cnt;
 227                 else
 228                         return tic->t_unit_res;
 229         }
 230 }
 231
 232 STATIC bool
 233 xlog_grant_head_wake(
 234         struct xlog             *log,
 235         struct xlog_grant_head  *head,
 236         int                     *free_bytes)
 237 {
 238         struct xlog_ticket      *tic;
 239         int                     need_bytes;
 240
 241         list_for_each_entry(tic, &head->waiters, t_queue) {
 242                 need_bytes = xlog_ticket_reservation(log, head, tic);
 243                 if (*free_bytes < need_bytes)
 244                         return false;
 245
 246                 *free_bytes -= need_bytes;
 247                 trace_xfs_log_grant_wake_up(log, tic);
 248                 wake_up_process(tic->t_task);
 249         }
 250
 251         return true;
 252 }
 253
 254 STATIC int
 255 xlog_grant_head_wait(
 256         struct xlog             *log,
 257         struct xlog_grant_head  *head,
 258         struct xlog_ticket      *tic,
 259         int                     need_bytes) __releases(&head->lock)
 260                                             __acquires(&head->lock)
 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(log->l_mp, 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 -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 -EIO;
 382
 383         XFS_STATS_INC(mp, 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 {
 441         struct xlog             *log = mp->m_log;
 442         struct xlog_ticket      *tic;
 443         int                     need_bytes;
 444         int                     error = 0;
 445
 446         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 447
 448         if (XLOG_FORCED_SHUTDOWN(log))
 449                 return -EIO;
 450
 451         XFS_STATS_INC(mp, xs_try_logspace);
 452
 453         ASSERT(*ticp == NULL);
 454         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 455                                 KM_SLEEP | KM_MAYFAIL);
 456         if (!tic)
 457                 return -ENOMEM;
 458
 459         *ticp = tic;
 460
 461         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 462                                             : tic->t_unit_res);
 463
 464         trace_xfs_log_reserve(log, tic);
 465
 466         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 467                                       &need_bytes);
 468         if (error)
 469                 goto out_error;
 470
 471         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 472         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 473         trace_xfs_log_reserve_exit(log, tic);
 474         xlog_verify_grant_tail(log);
 475         return 0;
 476
 477 out_error:
 478         /*
 479          * If we are failing, make sure the ticket doesn't have any current
 480          * reservations.  We don't want to add this back when the ticket/
 481          * transaction gets cancelled.
 482          */
 483         tic->t_curr_res = 0;
 484         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 485         return error;
 486 }
 487
 488
 489 /*
 490  * NOTES:
 491  *
 492  *      1. currblock field gets updated at startup and after in-core logs
 493  *              marked as with WANT_SYNC.
 494  */
 495
 496 /*
 497  * This routine is called when a user of a log manager ticket is done with
 498  * the reservation.  If the ticket was ever used, then a commit record for
 499  * the associated transaction is written out as a log operation header with
 500  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 501  * a given ticket.  If the ticket was one with a permanent reservation, then
 502  * a few operations are done differently.  Permanent reservation tickets by
 503  * default don't release the reservation.  They just commit the current
 504  * transaction with the belief that the reservation is still needed.  A flag
 505  * must be passed in before permanent reservations are actually released.
 506  * When these type of tickets are not released, they need to be set into
 507  * the inited state again.  By doing this, a start record will be written
 508  * out when the next write occurs.
 509  */
 510 xfs_lsn_t
 511 xfs_log_done(
 512         struct xfs_mount        *mp,
 513         struct xlog_ticket      *ticket,
 514         struct xlog_in_core     **iclog,
 515         bool                    regrant)
 516 {
 517         struct xlog             *log = mp->m_log;
 518         xfs_lsn_t               lsn = 0;
 519
 520         if (XLOG_FORCED_SHUTDOWN(log) ||
 521             /*
 522              * If nothing was ever written, don't write out commit record.
 523              * If we get an error, just continue and give back the log ticket.
 524              */
 525             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 526              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 527                 lsn = (xfs_lsn_t) -1;
 528                 regrant = false;
 529         }
 530
 531
 532         if (!regrant) {
 533                 trace_xfs_log_done_nonperm(log, ticket);
 534
 535                 /*
 536                  * Release ticket if not permanent reservation or a specific
 537                  * request has been made to release a permanent reservation.
 538                  */
 539                 xlog_ungrant_log_space(log, ticket);
 540         } else {
 541                 trace_xfs_log_done_perm(log, ticket);
 542
 543                 xlog_regrant_reserve_log_space(log, ticket);
 544                 /* If this ticket was a permanent reservation and we aren't
 545                  * trying to release it, reset the inited flags; so next time
 546                  * we write, a start record will be written out.
 547                  */
 548                 ticket->t_flags |= XLOG_TIC_INITED;
 549         }
 550
 551         xfs_log_ticket_put(ticket);
 552         return lsn;
 553 }
 554
 555 /*
 556  * Attaches a new iclog I/O completion callback routine during
 557  * transaction commit.  If the log is in error state, a non-zero
 558  * return code is handed back and the caller is responsible for
 559  * executing the callback at an appropriate time.
 560  */
 561 int
 562 xfs_log_notify(
 563         struct xfs_mount        *mp,
 564         struct xlog_in_core     *iclog,
 565         xfs_log_callback_t      *cb)
 566 {
 567         int     abortflg;
 568
 569         spin_lock(&iclog->ic_callback_lock);
 570         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 571         if (!abortflg) {
 572                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 573                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 574                 cb->cb_next = NULL;
 575                 *(iclog->ic_callback_tail) = cb;
 576                 iclog->ic_callback_tail = &(cb->cb_next);
 577         }
 578         spin_unlock(&iclog->ic_callback_lock);
 579         return abortflg;
 580 }
 581
 582 int
 583 xfs_log_release_iclog(
 584         struct xfs_mount        *mp,
 585         struct xlog_in_core     *iclog)
 586 {
 587         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 588                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 589                 return -EIO;
 590         }
 591
 592         return 0;
 593 }
 594
 595 /*
 596  * Mount a log filesystem
 597  *
 598  * mp           - ubiquitous xfs mount point structure
 599  * log_target   - buftarg of on-disk log device
 600  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 601  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 602  *
 603  * Return error or zero.
 604  */
 605 int
 606 xfs_log_mount(
 607         xfs_mount_t     *mp,
 608         xfs_buftarg_t   *log_target,
 609         xfs_daddr_t     blk_offset,
 610         int             num_bblks)
 611 {
 612         bool            fatal = xfs_sb_version_hascrc(&mp->m_sb);
 613         int             error = 0;
 614         int             min_logfsbs;
 615
 616         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 617                 xfs_notice(mp, "Mounting V%d Filesystem",
 618                            XFS_SB_VERSION_NUM(&mp->m_sb));
 619         } else {
 620                 xfs_notice(mp,
 621 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
 622                            XFS_SB_VERSION_NUM(&mp->m_sb));
 623                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 624         }
 625
 626         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 627         if (IS_ERR(mp->m_log)) {
 628                 error = PTR_ERR(mp->m_log);
 629                 goto out;
 630         }
 631
 632         /*
 633          * Validate the given log space and drop a critical message via syslog
 634          * if the log size is too small that would lead to some unexpected
 635          * situations in transaction log space reservation stage.
 636          *
 637          * Note: we can't just reject the mount if the validation fails.  This
 638          * would mean that people would have to downgrade their kernel just to
 639          * remedy the situation as there is no way to grow the log (short of
 640          * black magic surgery with xfs_db).
 641          *
 642          * We can, however, reject mounts for CRC format filesystems, as the
 643          * mkfs binary being used to make the filesystem should never create a
 644          * filesystem with a log that is too small.
 645          */
 646         min_logfsbs = xfs_log_calc_minimum_size(mp);
 647
 648         if (mp->m_sb.sb_logblocks < min_logfsbs) {
 649                 xfs_warn(mp,
 650                 "Log size %d blocks too small, minimum size is %d blocks",
 651                          mp->m_sb.sb_logblocks, min_logfsbs);
 652                 error = -EINVAL;
 653         } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
 654                 xfs_warn(mp,
 655                 "Log size %d blocks too large, maximum size is %lld blocks",
 656                          mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
 657                 error = -EINVAL;
 658         } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
 659                 xfs_warn(mp,
 660                 "log size %lld bytes too large, maximum size is %lld bytes",
 661                          XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
 662                          XFS_MAX_LOG_BYTES);
 663                 error = -EINVAL;
 664         } else if (mp->m_sb.sb_logsunit > 1 &&
 665                    mp->m_sb.sb_logsunit % mp->m_sb.sb_blocksize) {
 666                 xfs_warn(mp,
 667                 "log stripe unit %u bytes must be a multiple of block size",
 668                          mp->m_sb.sb_logsunit);
 669                 error = -EINVAL;
 670                 fatal = true;
 671         }
 672         if (error) {
 673                 /*
 674                  * Log check errors are always fatal on v5; or whenever bad
 675                  * metadata leads to a crash.
 676                  */
 677                 if (fatal) {
 678                         xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 679                         ASSERT(0);
 680                         goto out_free_log;
 681                 }
 682                 xfs_crit(mp, "Log size out of supported range.");
 683                 xfs_crit(mp,
 684 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
 685         }
 686
 687         /*
 688          * Initialize the AIL now we have a log.
 689          */
 690         error = xfs_trans_ail_init(mp);
 691         if (error) {
 692                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 693                 goto out_free_log;
 694         }
 695         mp->m_log->l_ailp = mp->m_ail;
 696
 697         /*
 698          * skip log recovery on a norecovery mount.  pretend it all
 699          * just worked.
 700          */
 701         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 702                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 703
 704                 if (readonly)
 705                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 706
 707                 error = xlog_recover(mp->m_log);
 708
 709                 if (readonly)
 710                         mp->m_flags |= XFS_MOUNT_RDONLY;
 711                 if (error) {
 712                         xfs_warn(mp, "log mount/recovery failed: error %d",
 713                                 error);
 714                         xlog_recover_cancel(mp->m_log);
 715                         goto out_destroy_ail;
 716                 }
 717         }
 718
 719         error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 720                                "log");
 721         if (error)
 722                 goto out_destroy_ail;
 723
 724         /* Normal transactions can now occur */
 725         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 726
 727         /*
 728          * Now the log has been fully initialised and we know were our
 729          * space grant counters are, we can initialise the permanent ticket
 730          * needed for delayed logging to work.
 731          */
 732         xlog_cil_init_post_recovery(mp->m_log);
 733
 734         return 0;
 735
 736 out_destroy_ail:
 737         xfs_trans_ail_destroy(mp);
 738 out_free_log:
 739         xlog_dealloc_log(mp->m_log);
 740 out:
 741         return error;
 742 }
 743
 744 /*
 745  * Finish the recovery of the file system.  This is separate from the
 746  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 747  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 748  * here.
 749  *
 750  * If we finish recovery successfully, start the background log work. If we are
 751  * not doing recovery, then we have a RO filesystem and we don't need to start
 752  * it.
 753  */
 754 int
 755 xfs_log_mount_finish(
 756         struct xfs_mount        *mp)
 757 {
 758         int     error = 0;
 759         bool    readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 760         bool    recovered = mp->m_log->l_flags & XLOG_RECOVERY_NEEDED;
 761
 762         if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
 763                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 764                 return 0;
 765         } else if (readonly) {
 766                 /* Allow unlinked processing to proceed */
 767                 mp->m_flags &= ~XFS_MOUNT_RDONLY;
 768         }
 769
 770         /*
 771          * During the second phase of log recovery, we need iget and
 772          * iput to behave like they do for an active filesystem.
 773          * xfs_fs_drop_inode needs to be able to prevent the deletion
 774          * of inodes before we're done replaying log items on those
 775          * inodes.  Turn it off immediately after recovery finishes
 776          * so that we don't leak the quota inodes if subsequent mount
 777          * activities fail.
 778          *
 779          * We let all inodes involved in redo item processing end up on
 780          * the LRU instead of being evicted immediately so that if we do
 781          * something to an unlinked inode, the irele won't cause
 782          * premature truncation and freeing of the inode, which results
 783          * in log recovery failure.  We have to evict the unreferenced
 784          * lru inodes after clearing SB_ACTIVE because we don't
 785          * otherwise clean up the lru if there's a subsequent failure in
 786          * xfs_mountfs, which leads to us leaking the inodes if nothing
 787          * else (e.g. quotacheck) references the inodes before the
 788          * mount failure occurs.
 789          */
 790         mp->m_super->s_flags |= SB_ACTIVE;
 791         error = xlog_recover_finish(mp->m_log);
 792         if (!error)
 793                 xfs_log_work_queue(mp);
 794         mp->m_super->s_flags &= ~SB_ACTIVE;
 795         evict_inodes(mp->m_super);
 796
 797         /*
 798          * Drain the buffer LRU after log recovery. This is required for v4
 799          * filesystems to avoid leaving around buffers with NULL verifier ops,
 800          * but we do it unconditionally to make sure we're always in a clean
 801          * cache state after mount.
 802          *
 803          * Don't push in the error case because the AIL may have pending intents
 804          * that aren't removed until recovery is cancelled.
 805          */
 806         if (!error && recovered) {
 807                 xfs_log_force(mp, XFS_LOG_SYNC);
 808                 xfs_ail_push_all_sync(mp->m_ail);
 809         }
 810         xfs_wait_buftarg(mp->m_ddev_targp);
 811
 812         if (readonly)
 813                 mp->m_flags |= XFS_MOUNT_RDONLY;
 814
 815         return error;
 816 }
 817
 818 /*
 819  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
 820  * the log.
 821  */
 822 int
 823 xfs_log_mount_cancel(
 824         struct xfs_mount        *mp)
 825 {
 826         int                     error;
 827
 828         error = xlog_recover_cancel(mp->m_log);
 829         xfs_log_unmount(mp);
 830
 831         return error;
 832 }
 833
 834 /*
 835  * Final log writes as part of unmount.
 836  *
 837  * Mark the filesystem clean as unmount happens.  Note that during relocation
 838  * this routine needs to be executed as part of source-bag while the
 839  * deallocation must not be done until source-end.
 840  */
 841
 842 /*
 843  * Unmount record used to have a string "Unmount filesystem--" in the
 844  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 845  * We just write the magic number now since that particular field isn't
 846  * currently architecture converted and "Unmount" is a bit foo.
 847  * As far as I know, there weren't any dependencies on the old behaviour.
 848  */
 849
 850 static int
 851 xfs_log_unmount_write(xfs_mount_t *mp)
 852 {
 853         struct xlog      *log = mp->m_log;
 854         xlog_in_core_t   *iclog;
 855 #ifdef DEBUG
 856         xlog_in_core_t   *first_iclog;
 857 #endif
 858         xlog_ticket_t   *tic = NULL;
 859         xfs_lsn_t        lsn;
 860         int              error;
 861
 862         /*
 863          * Don't write out unmount record on norecovery mounts or ro devices.
 864          * Or, if we are doing a forced umount (typically because of IO errors).
 865          */
 866         if (mp->m_flags & XFS_MOUNT_NORECOVERY ||
 867             xfs_readonly_buftarg(log->l_mp->m_logdev_targp)) {
 868                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 869                 return 0;
 870         }
 871
 872         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
 873         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 874
 875 #ifdef DEBUG
 876         first_iclog = iclog = log->l_iclog;
 877         do {
 878                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 879                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 880                         ASSERT(iclog->ic_offset == 0);
 881                 }
 882                 iclog = iclog->ic_next;
 883         } while (iclog != first_iclog);
 884 #endif
 885         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 886                 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
 887                 if (!error) {
 888                         /* the data section must be 32 bit size aligned */
 889                         struct {
 890                             uint16_t magic;
 891                             uint16_t pad1;
 892                             uint32_t pad2; /* may as well make it 64 bits */
 893                         } magic = {
 894                                 .magic = XLOG_UNMOUNT_TYPE,
 895                         };
 896                         struct xfs_log_iovec reg = {
 897                                 .i_addr = &magic,
 898                                 .i_len = sizeof(magic),
 899                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 900                         };
 901                         struct xfs_log_vec vec = {
 902                                 .lv_niovecs = 1,
 903                                 .lv_iovecp = &reg,
 904                         };
 905
 906                         /* remove inited flag, and account for space used */
 907                         tic->t_flags = 0;
 908                         tic->t_curr_res -= sizeof(magic);
 909                         error = xlog_write(log, &vec, tic, &lsn,
 910                                            NULL, XLOG_UNMOUNT_TRANS);
 911                         /*
 912                          * At this point, we're umounting anyway,
 913                          * so there's no point in transitioning log state
 914                          * to IOERROR. Just continue...
 915                          */
 916                 }
 917
 918                 if (error)
 919                         xfs_alert(mp, "%s: unmount record failed", __func__);
 920
 921
 922                 spin_lock(&log->l_icloglock);
 923                 iclog = log->l_iclog;
 924                 atomic_inc(&iclog->ic_refcnt);
 925                 xlog_state_want_sync(log, iclog);
 926                 spin_unlock(&log->l_icloglock);
 927                 error = xlog_state_release_iclog(log, iclog);
 928
 929                 spin_lock(&log->l_icloglock);
 930                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 931                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 932                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 933                                 xlog_wait(&iclog->ic_force_wait,
 934                                                         &log->l_icloglock);
 935                         } else {
 936                                 spin_unlock(&log->l_icloglock);
 937                         }
 938                 } else {
 939                         spin_unlock(&log->l_icloglock);
 940                 }
 941                 if (tic) {
 942                         trace_xfs_log_umount_write(log, tic);
 943                         xlog_ungrant_log_space(log, tic);
 944                         xfs_log_ticket_put(tic);
 945                 }
 946         } else {
 947                 /*
 948                  * We're already in forced_shutdown mode, couldn't
 949                  * even attempt to write out the unmount transaction.
 950                  *
 951                  * Go through the motions of sync'ing and releasing
 952                  * the iclog, even though no I/O will actually happen,
 953                  * we need to wait for other log I/Os that may already
 954                  * be in progress.  Do this as a separate section of
 955                  * code so we'll know if we ever get stuck here that
 956                  * we're in this odd situation of trying to unmount
 957                  * a file system that went into forced_shutdown as
 958                  * the result of an unmount..
 959                  */
 960                 spin_lock(&log->l_icloglock);
 961                 iclog = log->l_iclog;
 962                 atomic_inc(&iclog->ic_refcnt);
 963
 964                 xlog_state_want_sync(log, iclog);
 965                 spin_unlock(&log->l_icloglock);
 966                 error =  xlog_state_release_iclog(log, iclog);
 967
 968                 spin_lock(&log->l_icloglock);
 969
 970                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 971                         || iclog->ic_state == XLOG_STATE_DIRTY
 972                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 973
 974                                 xlog_wait(&iclog->ic_force_wait,
 975                                                         &log->l_icloglock);
 976                 } else {
 977                         spin_unlock(&log->l_icloglock);
 978                 }
 979         }
 980
 981         return error;
 982 }       /* xfs_log_unmount_write */
 983
 984 /*
 985  * Empty the log for unmount/freeze.
 986  *
 987  * To do this, we first need to shut down the background log work so it is not
 988  * trying to cover the log as we clean up. We then need to unpin all objects in
 989  * the log so we can then flush them out. Once they have completed their IO and
 990  * run the callbacks removing themselves from the AIL, we can write the unmount
 991  * record.
 992  */
 993 void
 994 xfs_log_quiesce(
 995         struct xfs_mount        *mp)
 996 {
 997         cancel_delayed_work_sync(&mp->m_log->l_work);
 998         xfs_log_force(mp, XFS_LOG_SYNC);
 999
1000         /*
1001          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
1002          * will push it, xfs_wait_buftarg() will not wait for it. Further,
1003          * xfs_buf_iowait() cannot be used because it was pushed with the
1004          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
1005          * the IO to complete.
1006          */
1007         xfs_ail_push_all_sync(mp->m_ail);
1008         xfs_wait_buftarg(mp->m_ddev_targp);
1009         xfs_buf_lock(mp->m_sb_bp);
1010         xfs_buf_unlock(mp->m_sb_bp);
1011
1012         xfs_log_unmount_write(mp);
1013 }
1014
1015 /*
1016  * Shut down and release the AIL and Log.
1017  *
1018  * During unmount, we need to ensure we flush all the dirty metadata objects
1019  * from the AIL so that the log is empty before we write the unmount record to
1020  * the log. Once this is done, we can tear down the AIL and the log.
1021  */
1022 void
1023 xfs_log_unmount(
1024         struct xfs_mount        *mp)
1025 {
1026         xfs_log_quiesce(mp);
1027
1028         xfs_trans_ail_destroy(mp);
1029
1030         xfs_sysfs_del(&mp->m_log->l_kobj);
1031
1032         xlog_dealloc_log(mp->m_log);
1033 }
1034
1035 void
1036 xfs_log_item_init(
1037         struct xfs_mount        *mp,
1038         struct xfs_log_item     *item,
1039         int                     type,
1040         const struct xfs_item_ops *ops)
1041 {
1042         item->li_mountp = mp;
1043         item->li_ailp = mp->m_ail;
1044         item->li_type = type;
1045         item->li_ops = ops;
1046         item->li_lv = NULL;
1047
1048         INIT_LIST_HEAD(&item->li_ail);
1049         INIT_LIST_HEAD(&item->li_cil);
1050         INIT_LIST_HEAD(&item->li_bio_list);
1051 }
1052
1053 /*
1054  * Wake up processes waiting for log space after we have moved the log tail.
1055  */
1056 void
1057 xfs_log_space_wake(
1058         struct xfs_mount        *mp)
1059 {
1060         struct xlog             *log = mp->m_log;
1061         int                     free_bytes;
1062
1063         if (XLOG_FORCED_SHUTDOWN(log))
1064                 return;
1065
1066         if (!list_empty_careful(&log->l_write_head.waiters)) {
1067                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1068
1069                 spin_lock(&log->l_write_head.lock);
1070                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1071                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1072                 spin_unlock(&log->l_write_head.lock);
1073         }
1074
1075         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1076                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1077
1078                 spin_lock(&log->l_reserve_head.lock);
1079                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1080                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1081                 spin_unlock(&log->l_reserve_head.lock);
1082         }
1083 }
1084
1085 /*
1086  * Determine if we have a transaction that has gone to disk that needs to be
1087  * covered. To begin the transition to the idle state firstly the log needs to
1088  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1089  * we start attempting to cover the log.
1090  *
1091  * Only if we are then in a state where covering is needed, the caller is
1092  * informed that dummy transactions are required to move the log into the idle
1093  * state.
1094  *
1095  * If there are any items in the AIl or CIL, then we do not want to attempt to
1096  * cover the log as we may be in a situation where there isn't log space
1097  * available to run a dummy transaction and this can lead to deadlocks when the
1098  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1099  * there's no point in running a dummy transaction at this point because we
1100  * can't start trying to idle the log until both the CIL and AIL are empty.
1101  */
1102 static int
1103 xfs_log_need_covered(xfs_mount_t *mp)
1104 {
1105         struct xlog     *log = mp->m_log;
1106         int             needed = 0;
1107
1108         if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1109                 return 0;
1110
1111         if (!xlog_cil_empty(log))
1112                 return 0;
1113
1114         spin_lock(&log->l_icloglock);
1115         switch (log->l_covered_state) {
1116         case XLOG_STATE_COVER_DONE:
1117         case XLOG_STATE_COVER_DONE2:
1118         case XLOG_STATE_COVER_IDLE:
1119                 break;
1120         case XLOG_STATE_COVER_NEED:
1121         case XLOG_STATE_COVER_NEED2:
1122                 if (xfs_ail_min_lsn(log->l_ailp))
1123                         break;
1124                 if (!xlog_iclogs_empty(log))
1125                         break;
1126
1127                 needed = 1;
1128                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1129                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1130                 else
1131                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1132                 break;
1133         default:
1134                 needed = 1;
1135                 break;
1136         }
1137         spin_unlock(&log->l_icloglock);
1138         return needed;
1139 }
1140
1141 /*
1142  * We may be holding the log iclog lock upon entering this routine.
1143  */
1144 xfs_lsn_t
1145 xlog_assign_tail_lsn_locked(
1146         struct xfs_mount        *mp)
1147 {
1148         struct xlog             *log = mp->m_log;
1149         struct xfs_log_item     *lip;
1150         xfs_lsn_t               tail_lsn;
1151
1152         assert_spin_locked(&mp->m_ail->xa_lock);
1153
1154         /*
1155          * To make sure we always have a valid LSN for the log tail we keep
1156          * track of the last LSN which was committed in log->l_last_sync_lsn,
1157          * and use that when the AIL was empty.
1158          */
1159         lip = xfs_ail_min(mp->m_ail);
1160         if (lip)
1161                 tail_lsn = lip->li_lsn;
1162         else
1163                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1164         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1165         atomic64_set(&log->l_tail_lsn, tail_lsn);
1166         return tail_lsn;
1167 }
1168
1169 xfs_lsn_t
1170 xlog_assign_tail_lsn(
1171         struct xfs_mount        *mp)
1172 {
1173         xfs_lsn_t               tail_lsn;
1174
1175         spin_lock(&mp->m_ail->xa_lock);
1176         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1177         spin_unlock(&mp->m_ail->xa_lock);
1178
1179         return tail_lsn;
1180 }
1181
1182 /*
1183  * Return the space in the log between the tail and the head.  The head
1184  * is passed in the cycle/bytes formal parms.  In the special case where
1185  * the reserve head has wrapped passed the tail, this calculation is no
1186  * longer valid.  In this case, just return 0 which means there is no space
1187  * in the log.  This works for all places where this function is called
1188  * with the reserve head.  Of course, if the write head were to ever
1189  * wrap the tail, we should blow up.  Rather than catch this case here,
1190  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1191  *
1192  * This code also handles the case where the reservation head is behind
1193  * the tail.  The details of this case are described below, but the end
1194  * result is that we return the size of the log as the amount of space left.
1195  */
1196 STATIC int
1197 xlog_space_left(
1198         struct xlog     *log,
1199         atomic64_t      *head)
1200 {
1201         int             free_bytes;
1202         int             tail_bytes;
1203         int             tail_cycle;
1204         int             head_cycle;
1205         int             head_bytes;
1206
1207         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1208         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1209         tail_bytes = BBTOB(tail_bytes);
1210         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1211                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1212         else if (tail_cycle + 1 < head_cycle)
1213                 return 0;
1214         else if (tail_cycle < head_cycle) {
1215                 ASSERT(tail_cycle == (head_cycle - 1));
1216                 free_bytes = tail_bytes - head_bytes;
1217         } else {
1218                 /*
1219                  * The reservation head is behind the tail.
1220                  * In this case we just want to return the size of the
1221                  * log as the amount of space left.
1222                  */
1223                 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1224                 xfs_alert(log->l_mp,
1225                           "  tail_cycle = %d, tail_bytes = %d",
1226                           tail_cycle, tail_bytes);
1227                 xfs_alert(log->l_mp,
1228                           "  GH   cycle = %d, GH   bytes = %d",
1229                           head_cycle, head_bytes);
1230                 ASSERT(0);
1231                 free_bytes = log->l_logsize;
1232         }
1233         return free_bytes;
1234 }
1235
1236
1237 /*
1238  * Log function which is called when an io completes.
1239  *
1240  * The log manager needs its own routine, in order to control what
1241  * happens with the buffer after the write completes.
1242  */
1243 static void
1244 xlog_iodone(xfs_buf_t *bp)
1245 {
1246         struct xlog_in_core     *iclog = bp->b_log_item;
1247         struct xlog             *l = iclog->ic_log;
1248         int                     aborted = 0;
1249
1250         /*
1251          * Race to shutdown the filesystem if we see an error or the iclog is in
1252          * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
1253          * CRC errors into log recovery.
1254          */
1255         if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR) ||
1256             iclog->ic_state & XLOG_STATE_IOABORT) {
1257                 if (iclog->ic_state & XLOG_STATE_IOABORT)
1258                         iclog->ic_state &= ~XLOG_STATE_IOABORT;
1259
1260                 xfs_buf_ioerror_alert(bp, __func__);
1261                 xfs_buf_stale(bp);
1262                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1263                 /*
1264                  * This flag will be propagated to the trans-committed
1265                  * callback routines to let them know that the log-commit
1266                  * didn't succeed.
1267                  */
1268                 aborted = XFS_LI_ABORTED;
1269         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1270                 aborted = XFS_LI_ABORTED;
1271         }
1272
1273         /* log I/O is always issued ASYNC */
1274         ASSERT(bp->b_flags & XBF_ASYNC);
1275         xlog_state_done_syncing(iclog, aborted);
1276
1277         /*
1278          * drop the buffer lock now that we are done. Nothing references
1279          * the buffer after this, so an unmount waiting on this lock can now
1280          * tear it down safely. As such, it is unsafe to reference the buffer
1281          * (bp) after the unlock as we could race with it being freed.
1282          */
1283         xfs_buf_unlock(bp);
1284 }
1285
1286 /*
1287  * Return size of each in-core log record buffer.
1288  *
1289  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1290  *
1291  * If the filesystem blocksize is too large, we may need to choose a
1292  * larger size since the directory code currently logs entire blocks.
1293  */
1294
1295 STATIC void
1296 xlog_get_iclog_buffer_size(
1297         struct xfs_mount        *mp,
1298         struct xlog             *log)
1299 {
1300         int size;
1301         int xhdrs;
1302
1303         if (mp->m_logbufs <= 0)
1304                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1305         else
1306                 log->l_iclog_bufs = mp->m_logbufs;
1307
1308         /*
1309          * Buffer size passed in from mount system call.
1310          */
1311         if (mp->m_logbsize > 0) {
1312                 size = log->l_iclog_size = mp->m_logbsize;
1313                 log->l_iclog_size_log = 0;
1314                 while (size != 1) {
1315                         log->l_iclog_size_log++;
1316                         size >>= 1;
1317                 }
1318
1319                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1320                         /* # headers = size / 32k
1321                          * one header holds cycles from 32k of data
1322                          */
1323
1324                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1325                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1326                                 xhdrs++;
1327                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1328                         log->l_iclog_heads = xhdrs;
1329                 } else {
1330                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1331                         log->l_iclog_hsize = BBSIZE;
1332                         log->l_iclog_heads = 1;
1333                 }
1334                 goto done;
1335         }
1336
1337         /* All machines use 32kB buffers by default. */
1338         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1339         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1340
1341         /* the default log size is 16k or 32k which is one header sector */
1342         log->l_iclog_hsize = BBSIZE;
1343         log->l_iclog_heads = 1;
1344
1345 done:
1346         /* are we being asked to make the sizes selected above visible? */
1347         if (mp->m_logbufs == 0)
1348                 mp->m_logbufs = log->l_iclog_bufs;
1349         if (mp->m_logbsize == 0)
1350                 mp->m_logbsize = log->l_iclog_size;
1351 }       /* xlog_get_iclog_buffer_size */
1352
1353
1354 void
1355 xfs_log_work_queue(
1356         struct xfs_mount        *mp)
1357 {
1358         queue_delayed_work(mp->m_sync_workqueue, &mp->m_log->l_work,
1359                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1360 }
1361
1362 /*
1363  * Every sync period we need to unpin all items in the AIL and push them to
1364  * disk. If there is nothing dirty, then we might need to cover the log to
1365  * indicate that the filesystem is idle.
1366  */
1367 static void
1368 xfs_log_worker(
1369         struct work_struct      *work)
1370 {
1371         struct xlog             *log = container_of(to_delayed_work(work),
1372                                                 struct xlog, l_work);
1373         struct xfs_mount        *mp = log->l_mp;
1374
1375         /* dgc: errors ignored - not fatal and nowhere to report them */
1376         if (xfs_log_need_covered(mp)) {
1377                 /*
1378                  * Dump a transaction into the log that contains no real change.
1379                  * This is needed to stamp the current tail LSN into the log
1380                  * during the covering operation.
1381                  *
1382                  * We cannot use an inode here for this - that will push dirty
1383                  * state back up into the VFS and then periodic inode flushing
1384                  * will prevent log covering from making progress. Hence we
1385                  * synchronously log the superblock instead to ensure the
1386                  * superblock is immediately unpinned and can be written back.
1387                  */
1388                 xfs_sync_sb(mp, true);
1389         } else
1390                 xfs_log_force(mp, 0);
1391
1392         /* start pushing all the metadata that is currently dirty */
1393         xfs_ail_push_all(mp->m_ail);
1394
1395         /* queue us up again */
1396         xfs_log_work_queue(mp);
1397 }
1398
1399 /*
1400  * This routine initializes some of the log structure for a given mount point.
1401  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1402  * some other stuff may be filled in too.
1403  */
1404 STATIC struct xlog *
1405 xlog_alloc_log(
1406         struct xfs_mount        *mp,
1407         struct xfs_buftarg      *log_target,
1408         xfs_daddr_t             blk_offset,
1409         int                     num_bblks)
1410 {
1411         struct xlog             *log;
1412         xlog_rec_header_t       *head;
1413         xlog_in_core_t          **iclogp;
1414         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1415         xfs_buf_t               *bp;
1416         int                     i;
1417         int                     error = -ENOMEM;
1418         uint                    log2_size = 0;
1419
1420         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1421         if (!log) {
1422                 xfs_warn(mp, "Log allocation failed: No memory!");
1423                 goto out;
1424         }
1425
1426         log->l_mp          = mp;
1427         log->l_targ        = log_target;
1428         log->l_logsize     = BBTOB(num_bblks);
1429         log->l_logBBstart  = blk_offset;
1430         log->l_logBBsize   = num_bblks;
1431         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1432         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1433         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1434
1435         log->l_prev_block  = -1;
1436         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1437         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1438         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1439         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1440
1441         xlog_grant_head_init(&log->l_reserve_head);
1442         xlog_grant_head_init(&log->l_write_head);
1443
1444         error = -EFSCORRUPTED;
1445         if (xfs_sb_version_hassector(&mp->m_sb)) {
1446                 log2_size = mp->m_sb.sb_logsectlog;
1447                 if (log2_size < BBSHIFT) {
1448                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1449                                 log2_size, BBSHIFT);
1450                         goto out_free_log;
1451                 }
1452
1453                 log2_size -= BBSHIFT;
1454                 if (log2_size > mp->m_sectbb_log) {
1455                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1456                                 log2_size, mp->m_sectbb_log);
1457                         goto out_free_log;
1458                 }
1459
1460                 /* for larger sector sizes, must have v2 or external log */
1461                 if (log2_size && log->l_logBBstart > 0 &&
1462                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1463                         xfs_warn(mp,
1464                 "log sector size (0x%x) invalid for configuration.",
1465                                 log2_size);
1466                         goto out_free_log;
1467                 }
1468         }
1469         log->l_sectBBsize = 1 << log2_size;
1470
1471         xlog_get_iclog_buffer_size(mp, log);
1472
1473         /*
1474          * Use a NULL block for the extra log buffer used during splits so that
1475          * it will trigger errors if we ever try to do IO on it without first
1476          * having set it up properly.
1477          */
1478         error = -ENOMEM;
1479         bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1480                            BTOBB(log->l_iclog_size), XBF_NO_IOACCT);
1481         if (!bp)
1482                 goto out_free_log;
1483
1484         /*
1485          * The iclogbuf buffer locks are held over IO but we are not going to do
1486          * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1487          * when appropriately.
1488          */
1489         ASSERT(xfs_buf_islocked(bp));
1490         xfs_buf_unlock(bp);
1491
1492         /* use high priority wq for log I/O completion */
1493         bp->b_ioend_wq = mp->m_log_workqueue;
1494         bp->b_iodone = xlog_iodone;
1495         log->l_xbuf = bp;
1496
1497         spin_lock_init(&log->l_icloglock);
1498         init_waitqueue_head(&log->l_flush_wait);
1499
1500         iclogp = &log->l_iclog;
1501         /*
1502          * The amount of memory to allocate for the iclog structure is
1503          * rather funky due to the way the structure is defined.  It is
1504          * done this way so that we can use different sizes for machines
1505          * with different amounts of memory.  See the definition of
1506          * xlog_in_core_t in xfs_log_priv.h for details.
1507          */
1508         ASSERT(log->l_iclog_size >= 4096);
1509         for (i=0; i < log->l_iclog_bufs; i++) {
1510                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1511                 if (!*iclogp)
1512                         goto out_free_iclog;
1513
1514                 iclog = *iclogp;
1515                 iclog->ic_prev = prev_iclog;
1516                 prev_iclog = iclog;
1517
1518                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1519                                           BTOBB(log->l_iclog_size),
1520                                           XBF_NO_IOACCT);
1521                 if (!bp)
1522                         goto out_free_iclog;
1523
1524                 ASSERT(xfs_buf_islocked(bp));
1525                 xfs_buf_unlock(bp);
1526
1527                 /* use high priority wq for log I/O completion */
1528                 bp->b_ioend_wq = mp->m_log_workqueue;
1529                 bp->b_iodone = xlog_iodone;
1530                 iclog->ic_bp = bp;
1531                 iclog->ic_data = bp->b_addr;
1532 #ifdef DEBUG
1533                 log->l_iclog_bak[i] = &iclog->ic_header;
1534 #endif
1535                 head = &iclog->ic_header;
1536                 memset(head, 0, sizeof(xlog_rec_header_t));
1537                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1538                 head->h_version = cpu_to_be32(
1539                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1540                 head->h_size = cpu_to_be32(log->l_iclog_size);
1541                 /* new fields */
1542                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1543                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1544
1545                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1546                 iclog->ic_state = XLOG_STATE_ACTIVE;
1547                 iclog->ic_log = log;
1548                 atomic_set(&iclog->ic_refcnt, 0);
1549                 spin_lock_init(&iclog->ic_callback_lock);
1550                 iclog->ic_callback_tail = &(iclog->ic_callback);
1551                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1552
1553                 init_waitqueue_head(&iclog->ic_force_wait);
1554                 init_waitqueue_head(&iclog->ic_write_wait);
1555
1556                 iclogp = &iclog->ic_next;
1557         }
1558         *iclogp = log->l_iclog;                 /* complete ring */
1559         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1560
1561         error = xlog_cil_init(log);
1562         if (error)
1563                 goto out_free_iclog;
1564         return log;
1565
1566 out_free_iclog:
1567         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1568                 prev_iclog = iclog->ic_next;
1569                 if (iclog->ic_bp)
1570                         xfs_buf_free(iclog->ic_bp);
1571                 kmem_free(iclog);
1572         }
1573         spinlock_destroy(&log->l_icloglock);
1574         xfs_buf_free(log->l_xbuf);
1575 out_free_log:
1576         kmem_free(log);
1577 out:
1578         return ERR_PTR(error);
1579 }       /* xlog_alloc_log */
1580
1581
1582 /*
1583  * Write out the commit record of a transaction associated with the given
1584  * ticket.  Return the lsn of the commit record.
1585  */
1586 STATIC int
1587 xlog_commit_record(
1588         struct xlog             *log,
1589         struct xlog_ticket      *ticket,
1590         struct xlog_in_core     **iclog,
1591         xfs_lsn_t               *commitlsnp)
1592 {
1593         struct xfs_mount *mp = log->l_mp;
1594         int     error;
1595         struct xfs_log_iovec reg = {
1596                 .i_addr = NULL,
1597                 .i_len = 0,
1598                 .i_type = XLOG_REG_TYPE_COMMIT,
1599         };
1600         struct xfs_log_vec vec = {
1601                 .lv_niovecs = 1,
1602                 .lv_iovecp = &reg,
1603         };
1604
1605         ASSERT_ALWAYS(iclog);
1606         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1607                                         XLOG_COMMIT_TRANS);
1608         if (error)
1609                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1610         return error;
1611 }
1612
1613 /*
1614  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1615  * log space.  This code pushes on the lsn which would supposedly free up
1616  * the 25% which we want to leave free.  We may need to adopt a policy which
1617  * pushes on an lsn which is further along in the log once we reach the high
1618  * water mark.  In this manner, we would be creating a low water mark.
1619  */
1620 STATIC void
1621 xlog_grant_push_ail(
1622         struct xlog     *log,
1623         int             need_bytes)
1624 {
1625         xfs_lsn_t       threshold_lsn = 0;
1626         xfs_lsn_t       last_sync_lsn;
1627         int             free_blocks;
1628         int             free_bytes;
1629         int             threshold_block;
1630         int             threshold_cycle;
1631         int             free_threshold;
1632
1633         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1634
1635         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1636         free_blocks = BTOBBT(free_bytes);
1637
1638         /*
1639          * Set the threshold for the minimum number of free blocks in the
1640          * log to the maximum of what the caller needs, one quarter of the
1641          * log, and 256 blocks.
1642          */
1643         free_threshold = BTOBB(need_bytes);
1644         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1645         free_threshold = MAX(free_threshold, 256);
1646         if (free_blocks >= free_threshold)
1647                 return;
1648
1649         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1650                                                 &threshold_block);
1651         threshold_block += free_threshold;
1652         if (threshold_block >= log->l_logBBsize) {
1653                 threshold_block -= log->l_logBBsize;
1654                 threshold_cycle += 1;
1655         }
1656         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1657                                         threshold_block);
1658         /*
1659          * Don't pass in an lsn greater than the lsn of the last
1660          * log record known to be on disk. Use a snapshot of the last sync lsn
1661          * so that it doesn't change between the compare and the set.
1662          */
1663         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1664         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1665                 threshold_lsn = last_sync_lsn;
1666
1667         /*
1668          * Get the transaction layer to kick the dirty buffers out to
1669          * disk asynchronously. No point in trying to do this if
1670          * the filesystem is shutting down.
1671          */
1672         if (!XLOG_FORCED_SHUTDOWN(log))
1673                 xfs_ail_push(log->l_ailp, threshold_lsn);
1674 }
1675
1676 /*
1677  * Stamp cycle number in every block
1678  */
1679 STATIC void
1680 xlog_pack_data(
1681         struct xlog             *log,
1682         struct xlog_in_core     *iclog,
1683         int                     roundoff)
1684 {
1685         int                     i, j, k;
1686         int                     size = iclog->ic_offset + roundoff;
1687         __be32                  cycle_lsn;
1688         char                    *dp;
1689
1690         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1691
1692         dp = iclog->ic_datap;
1693         for (i = 0; i < BTOBB(size); i++) {
1694                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1695                         break;
1696                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1697                 *(__be32 *)dp = cycle_lsn;
1698                 dp += BBSIZE;
1699         }
1700
1701         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1702                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1703
1704                 for ( ; i < BTOBB(size); i++) {
1705                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1706                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1707                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1708                         *(__be32 *)dp = cycle_lsn;
1709                         dp += BBSIZE;
1710                 }
1711
1712                 for (i = 1; i < log->l_iclog_heads; i++)
1713                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1714         }
1715 }
1716
1717 /*
1718  * Calculate the checksum for a log buffer.
1719  *
1720  * This is a little more complicated than it should be because the various
1721  * headers and the actual data are non-contiguous.
1722  */
1723 __le32
1724 xlog_cksum(
1725         struct xlog             *log,
1726         struct xlog_rec_header  *rhead,
1727         char                    *dp,
1728         int                     size)
1729 {
1730         uint32_t                crc;
1731
1732         /* first generate the crc for the record header ... */
1733         crc = xfs_start_cksum_update((char *)rhead,
1734                               sizeof(struct xlog_rec_header),
1735                               offsetof(struct xlog_rec_header, h_crc));
1736
1737         /* ... then for additional cycle data for v2 logs ... */
1738         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1739                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1740                 int             i;
1741                 int             xheads;
1742
1743                 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1744                 if (size % XLOG_HEADER_CYCLE_SIZE)
1745                         xheads++;
1746
1747                 for (i = 1; i < xheads; i++) {
1748                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1749                                      sizeof(struct xlog_rec_ext_header));
1750                 }
1751         }
1752
1753         /* ... and finally for the payload */
1754         crc = crc32c(crc, dp, size);
1755
1756         return xfs_end_cksum(crc);
1757 }
1758
1759 /*
1760  * The bdstrat callback function for log bufs. This gives us a central
1761  * place to trap bufs in case we get hit by a log I/O error and need to
1762  * shutdown. Actually, in practice, even when we didn't get a log error,
1763  * we transition the iclogs to IOERROR state *after* flushing all existing
1764  * iclogs to disk. This is because we don't want anymore new transactions to be
1765  * started or completed afterwards.
1766  *
1767  * We lock the iclogbufs here so that we can serialise against IO completion
1768  * during unmount. We might be processing a shutdown triggered during unmount,
1769  * and that can occur asynchronously to the unmount thread, and hence we need to
1770  * ensure that completes before tearing down the iclogbufs. Hence we need to
1771  * hold the buffer lock across the log IO to acheive that.
1772  */
1773 STATIC int
1774 xlog_bdstrat(
1775         struct xfs_buf          *bp)
1776 {
1777         struct xlog_in_core     *iclog = bp->b_log_item;
1778
1779         xfs_buf_lock(bp);
1780         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1781                 xfs_buf_ioerror(bp, -EIO);
1782                 xfs_buf_stale(bp);
1783                 xfs_buf_ioend(bp);
1784                 /*
1785                  * It would seem logical to return EIO here, but we rely on
1786                  * the log state machine to propagate I/O errors instead of
1787                  * doing it here. Similarly, IO completion will unlock the
1788                  * buffer, so we don't do it here.
1789                  */
1790                 return 0;
1791         }
1792
1793         xfs_buf_submit(bp);
1794         return 0;
1795 }
1796
1797 /*
1798  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1799  * fashion.  Previously, we should have moved the current iclog
1800  * ptr in the log to point to the next available iclog.  This allows further
1801  * write to continue while this code syncs out an iclog ready to go.
1802  * Before an in-core log can be written out, the data section must be scanned
1803  * to save away the 1st word of each BBSIZE block into the header.  We replace
1804  * it with the current cycle count.  Each BBSIZE block is tagged with the
1805  * cycle count because there in an implicit assumption that drives will
1806  * guarantee that entire 512 byte blocks get written at once.  In other words,
1807  * we can't have part of a 512 byte block written and part not written.  By
1808  * tagging each block, we will know which blocks are valid when recovering
1809  * after an unclean shutdown.
1810  *
1811  * This routine is single threaded on the iclog.  No other thread can be in
1812  * this routine with the same iclog.  Changing contents of iclog can there-
1813  * fore be done without grabbing the state machine lock.  Updating the global
1814  * log will require grabbing the lock though.
1815  *
1816  * The entire log manager uses a logical block numbering scheme.  Only
1817  * log_sync (and then only bwrite()) know about the fact that the log may
1818  * not start with block zero on a given device.  The log block start offset
1819  * is added immediately before calling bwrite().
1820  */
1821
1822 STATIC int
1823 xlog_sync(
1824         struct xlog             *log,
1825         struct xlog_in_core     *iclog)
1826 {
1827         xfs_buf_t       *bp;
1828         int             i;
1829         uint            count;          /* byte count of bwrite */
1830         uint            count_init;     /* initial count before roundup */
1831         int             roundoff;       /* roundoff to BB or stripe */
1832         int             split = 0;      /* split write into two regions */
1833         int             error;
1834         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1835         int             size;
1836
1837         XFS_STATS_INC(log->l_mp, xs_log_writes);
1838         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1839
1840         /* Add for LR header */
1841         count_init = log->l_iclog_hsize + iclog->ic_offset;
1842
1843         /* Round out the log write size */
1844         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1845                 /* we have a v2 stripe unit to use */
1846                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1847         } else {
1848                 count = BBTOB(BTOBB(count_init));
1849         }
1850         roundoff = count - count_init;
1851         ASSERT(roundoff >= 0);
1852         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1853                 roundoff < log->l_mp->m_sb.sb_logsunit)
1854                 ||
1855                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1856                  roundoff < BBTOB(1)));
1857
1858         /* move grant heads by roundoff in sync */
1859         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1860         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1861
1862         /* put cycle number in every block */
1863         xlog_pack_data(log, iclog, roundoff);
1864
1865         /* real byte length */
1866         size = iclog->ic_offset;
1867         if (v2)
1868                 size += roundoff;
1869         iclog->ic_header.h_len = cpu_to_be32(size);
1870
1871         bp = iclog->ic_bp;
1872         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1873
1874         XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1875
1876         /* Do we need to split this write into 2 parts? */
1877         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1878                 char            *dptr;
1879
1880                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1881                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1882                 iclog->ic_bwritecnt = 2;
1883
1884                 /*
1885                  * Bump the cycle numbers at the start of each block in the
1886                  * part of the iclog that ends up in the buffer that gets
1887                  * written to the start of the log.
1888                  *
1889                  * Watch out for the header magic number case, though.
1890                  */
1891                 dptr = (char *)&iclog->ic_header + count;
1892                 for (i = 0; i < split; i += BBSIZE) {
1893                         uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1894                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1895                                 cycle++;
1896                         *(__be32 *)dptr = cpu_to_be32(cycle);
1897
1898                         dptr += BBSIZE;
1899                 }
1900         } else {
1901                 iclog->ic_bwritecnt = 1;
1902         }
1903
1904         /* calculcate the checksum */
1905         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1906                                             iclog->ic_datap, size);
1907         /*
1908          * Intentionally corrupt the log record CRC based on the error injection
1909          * frequency, if defined. This facilitates testing log recovery in the
1910          * event of torn writes. Hence, set the IOABORT state to abort the log
1911          * write on I/O completion and shutdown the fs. The subsequent mount
1912          * detects the bad CRC and attempts to recover.
1913          */
1914         if (XFS_TEST_ERROR(false, log->l_mp, XFS_ERRTAG_LOG_BAD_CRC)) {
1915                 iclog->ic_header.h_crc &= cpu_to_le32(0xAAAAAAAA);
1916                 iclog->ic_state |= XLOG_STATE_IOABORT;
1917                 xfs_warn(log->l_mp,
1918         "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1919                          be64_to_cpu(iclog->ic_header.h_lsn));
1920         }
1921
1922         bp->b_io_length = BTOBB(count);
1923         bp->b_log_item = iclog;
1924         bp->b_flags &= ~XBF_FLUSH;
1925         bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1926
1927         /*
1928          * Flush the data device before flushing the log to make sure all meta
1929          * data written back from the AIL actually made it to disk before
1930          * stamping the new log tail LSN into the log buffer.  For an external
1931          * log we need to issue the flush explicitly, and unfortunately
1932          * synchronously here; for an internal log we can simply use the block
1933          * layer state machine for preflushes.
1934          */
1935         if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1936                 xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1937         else
1938                 bp->b_flags |= XBF_FLUSH;
1939
1940         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1941         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1942
1943         xlog_verify_iclog(log, iclog, count, true);
1944
1945         /* account for log which doesn't start at block #0 */
1946         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1947
1948         /*
1949          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1950          * is shutting down.
1951          */
1952         error = xlog_bdstrat(bp);
1953         if (error) {
1954                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1955                 return error;
1956         }
1957         if (split) {
1958                 bp = iclog->ic_log->l_xbuf;
1959                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1960                 xfs_buf_associate_memory(bp,
1961                                 (char *)&iclog->ic_header + count, split);
1962                 bp->b_log_item = iclog;
1963                 bp->b_flags &= ~XBF_FLUSH;
1964                 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE | XBF_FUA);
1965
1966                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1967                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1968
1969                 /* account for internal log which doesn't start at block #0 */
1970                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1971                 error = xlog_bdstrat(bp);
1972                 if (error) {
1973                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1974                         return error;
1975                 }
1976         }
1977         return 0;
1978 }       /* xlog_sync */
1979
1980 /*
1981  * Deallocate a log structure
1982  */
1983 STATIC void
1984 xlog_dealloc_log(
1985         struct xlog     *log)
1986 {
1987         xlog_in_core_t  *iclog, *next_iclog;
1988         int             i;
1989
1990         xlog_cil_destroy(log);
1991
1992         /*
1993          * Cycle all the iclogbuf locks to make sure all log IO completion
1994          * is done before we tear down these buffers.
1995          */
1996         iclog = log->l_iclog;
1997         for (i = 0; i < log->l_iclog_bufs; i++) {
1998                 xfs_buf_lock(iclog->ic_bp);
1999                 xfs_buf_unlock(iclog->ic_bp);
2000                 iclog = iclog->ic_next;
2001         }
2002
2003         /*
2004          * Always need to ensure that the extra buffer does not point to memory
2005          * owned by another log buffer before we free it. Also, cycle the lock
2006          * first to ensure we've completed IO on it.
2007          */
2008         xfs_buf_lock(log->l_xbuf);
2009         xfs_buf_unlock(log->l_xbuf);
2010         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
2011         xfs_buf_free(log->l_xbuf);
2012
2013         iclog = log->l_iclog;
2014         for (i = 0; i < log->l_iclog_bufs; i++) {
2015                 xfs_buf_free(iclog->ic_bp);
2016                 next_iclog = iclog->ic_next;
2017                 kmem_free(iclog);
2018                 iclog = next_iclog;
2019         }
2020         spinlock_destroy(&log->l_icloglock);
2021
2022         log->l_mp->m_log = NULL;
2023         kmem_free(log);
2024 }       /* xlog_dealloc_log */
2025
2026 /*
2027  * Update counters atomically now that memcpy is done.
2028  */
2029 /* ARGSUSED */
2030 static inline void
2031 xlog_state_finish_copy(
2032         struct xlog             *log,
2033         struct xlog_in_core     *iclog,
2034         int                     record_cnt,
2035         int                     copy_bytes)
2036 {
2037         spin_lock(&log->l_icloglock);
2038
2039         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
2040         iclog->ic_offset += copy_bytes;
2041
2042         spin_unlock(&log->l_icloglock);
2043 }       /* xlog_state_finish_copy */
2044
2045
2046
2047
2048 /*
2049  * print out info relating to regions written which consume
2050  * the reservation
2051  */
2052 void
2053 xlog_print_tic_res(
2054         struct xfs_mount        *mp,
2055         struct xlog_ticket      *ticket)
2056 {
2057         uint i;
2058         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2059
2060         /* match with XLOG_REG_TYPE_* in xfs_log.h */
2061 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2062         static char *res_type_str[XLOG_REG_TYPE_MAX + 1] = {
2063             REG_TYPE_STR(BFORMAT, "bformat"),
2064             REG_TYPE_STR(BCHUNK, "bchunk"),
2065             REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2066             REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2067             REG_TYPE_STR(IFORMAT, "iformat"),
2068             REG_TYPE_STR(ICORE, "icore"),
2069             REG_TYPE_STR(IEXT, "iext"),
2070             REG_TYPE_STR(IBROOT, "ibroot"),
2071             REG_TYPE_STR(ILOCAL, "ilocal"),
2072             REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2073             REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2074             REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2075             REG_TYPE_STR(QFORMAT, "qformat"),
2076             REG_TYPE_STR(DQUOT, "dquot"),
2077             REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2078             REG_TYPE_STR(LRHEADER, "LR header"),
2079             REG_TYPE_STR(UNMOUNT, "unmount"),
2080             REG_TYPE_STR(COMMIT, "commit"),
2081             REG_TYPE_STR(TRANSHDR, "trans header"),
2082             REG_TYPE_STR(ICREATE, "inode create")
2083         };
2084 #undef REG_TYPE_STR
2085
2086         xfs_warn(mp, "ticket reservation summary:");
2087         xfs_warn(mp, "  unit res    = %d bytes",
2088                  ticket->t_unit_res);
2089         xfs_warn(mp, "  current res = %d bytes",
2090                  ticket->t_curr_res);
2091         xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2092                  ticket->t_res_arr_sum, ticket->t_res_o_flow);
2093         xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2094                  ticket->t_res_num_ophdrs, ophdr_spc);
2095         xfs_warn(mp, "  ophdr + reg = %u bytes",
2096                  ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2097         xfs_warn(mp, "  num regions = %u",
2098                  ticket->t_res_num);
2099
2100         for (i = 0; i < ticket->t_res_num; i++) {
2101                 uint r_type = ticket->t_res_arr[i].r_type;
2102                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2103                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2104                             "bad-rtype" : res_type_str[r_type]),
2105                             ticket->t_res_arr[i].r_len);
2106         }
2107 }
2108
2109 /*
2110  * Print a summary of the transaction.
2111  */
2112 void
2113 xlog_print_trans(
2114         struct xfs_trans                *tp)
2115 {
2116         struct xfs_mount                *mp = tp->t_mountp;
2117         struct xfs_log_item_desc        *lidp;
2118
2119         /* dump core transaction and ticket info */
2120         xfs_warn(mp, "transaction summary:");
2121         xfs_warn(mp, "  log res   = %d", tp->t_log_res);
2122         xfs_warn(mp, "  log count = %d", tp->t_log_count);
2123         xfs_warn(mp, "  flags     = 0x%x", tp->t_flags);
2124
2125         xlog_print_tic_res(mp, tp->t_ticket);
2126
2127         /* dump each log item */
2128         list_for_each_entry(lidp, &tp->t_items, lid_trans) {
2129                 struct xfs_log_item     *lip = lidp->lid_item;
2130                 struct xfs_log_vec      *lv = lip->li_lv;
2131                 struct xfs_log_iovec    *vec;
2132                 int                     i;
2133
2134                 xfs_warn(mp, "log item: ");
2135                 xfs_warn(mp, "  type    = 0x%x", lip->li_type);
2136                 xfs_warn(mp, "  flags   = 0x%x", lip->li_flags);
2137                 if (!lv)
2138                         continue;
2139                 xfs_warn(mp, "  niovecs = %d", lv->lv_niovecs);
2140                 xfs_warn(mp, "  size    = %d", lv->lv_size);
2141                 xfs_warn(mp, "  bytes   = %d", lv->lv_bytes);
2142                 xfs_warn(mp, "  buf len = %d", lv->lv_buf_len);
2143
2144                 /* dump each iovec for the log item */
2145                 vec = lv->lv_iovecp;
2146                 for (i = 0; i < lv->lv_niovecs; i++) {
2147                         int dumplen = min(vec->i_len, 32);
2148
2149                         xfs_warn(mp, "  iovec[%d]", i);
2150                         xfs_warn(mp, "    type  = 0x%x", vec->i_type);
2151                         xfs_warn(mp, "    len   = %d", vec->i_len);
2152                         xfs_warn(mp, "    first %d bytes of iovec[%d]:", dumplen, i);
2153                         xfs_hex_dump(vec->i_addr, dumplen);
2154
2155                         vec++;
2156                 }
2157         }
2158 }
2159
2160 /*
2161  * Calculate the potential space needed by the log vector.  Each region gets
2162  * its own xlog_op_header_t and may need to be double word aligned.
2163  */
2164 static int
2165 xlog_write_calc_vec_length(
2166         struct xlog_ticket      *ticket,
2167         struct xfs_log_vec      *log_vector)
2168 {
2169         struct xfs_log_vec      *lv;
2170         int                     headers = 0;
2171         int                     len = 0;
2172         int                     i;
2173
2174         /* acct for start rec of xact */
2175         if (ticket->t_flags & XLOG_TIC_INITED)
2176                 headers++;
2177
2178         for (lv = log_vector; lv; lv = lv->lv_next) {
2179                 /* we don't write ordered log vectors */
2180                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2181                         continue;
2182
2183                 headers += lv->lv_niovecs;
2184
2185                 for (i = 0; i < lv->lv_niovecs; i++) {
2186                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2187
2188                         len += vecp->i_len;
2189                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2190                 }
2191         }
2192
2193         ticket->t_res_num_ophdrs += headers;
2194         len += headers * sizeof(struct xlog_op_header);
2195
2196         return len;
2197 }
2198
2199 /*
2200  * If first write for transaction, insert start record  We can't be trying to
2201  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2202  */
2203 static int
2204 xlog_write_start_rec(
2205         struct xlog_op_header   *ophdr,
2206         struct xlog_ticket      *ticket)
2207 {
2208         if (!(ticket->t_flags & XLOG_TIC_INITED))
2209                 return 0;
2210
2211         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2212         ophdr->oh_clientid = ticket->t_clientid;
2213         ophdr->oh_len = 0;
2214         ophdr->oh_flags = XLOG_START_TRANS;
2215         ophdr->oh_res2 = 0;
2216
2217         ticket->t_flags &= ~XLOG_TIC_INITED;
2218
2219         return sizeof(struct xlog_op_header);
2220 }
2221
2222 static xlog_op_header_t *
2223 xlog_write_setup_ophdr(
2224         struct xlog             *log,
2225         struct xlog_op_header   *ophdr,
2226         struct xlog_ticket      *ticket,
2227         uint                    flags)
2228 {
2229         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2230         ophdr->oh_clientid = ticket->t_clientid;
2231         ophdr->oh_res2 = 0;
2232
2233         /* are we copying a commit or unmount record? */
2234         ophdr->oh_flags = flags;
2235
2236         /*
2237          * We've seen logs corrupted with bad transaction client ids.  This
2238          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2239          * and shut down the filesystem.
2240          */
2241         switch (ophdr->oh_clientid)  {
2242         case XFS_TRANSACTION:
2243         case XFS_VOLUME:
2244         case XFS_LOG:
2245                 break;
2246         default:
2247                 xfs_warn(log->l_mp,
2248                         "Bad XFS transaction clientid 0x%x in ticket "PTR_FMT,
2249                         ophdr->oh_clientid, ticket);
2250                 return NULL;
2251         }
2252
2253         return ophdr;
2254 }
2255
2256 /*
2257  * Set up the parameters of the region copy into the log. This has
2258  * to handle region write split across multiple log buffers - this
2259  * state is kept external to this function so that this code can
2260  * be written in an obvious, self documenting manner.
2261  */
2262 static int
2263 xlog_write_setup_copy(
2264         struct xlog_ticket      *ticket,
2265         struct xlog_op_header   *ophdr,
2266         int                     space_available,
2267         int                     space_required,
2268         int                     *copy_off,
2269         int                     *copy_len,
2270         int                     *last_was_partial_copy,
2271         int                     *bytes_consumed)
2272 {
2273         int                     still_to_copy;
2274
2275         still_to_copy = space_required - *bytes_consumed;
2276         *copy_off = *bytes_consumed;
2277
2278         if (still_to_copy <= space_available) {
2279                 /* write of region completes here */
2280                 *copy_len = still_to_copy;
2281                 ophdr->oh_len = cpu_to_be32(*copy_len);
2282                 if (*last_was_partial_copy)
2283                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2284                 *last_was_partial_copy = 0;
2285                 *bytes_consumed = 0;
2286                 return 0;
2287         }
2288
2289         /* partial write of region, needs extra log op header reservation */
2290         *copy_len = space_available;
2291         ophdr->oh_len = cpu_to_be32(*copy_len);
2292         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2293         if (*last_was_partial_copy)
2294                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2295         *bytes_consumed += *copy_len;
2296         (*last_was_partial_copy)++;
2297
2298         /* account for new log op header */
2299         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2300         ticket->t_res_num_ophdrs++;
2301
2302         return sizeof(struct xlog_op_header);
2303 }
2304
2305 static int
2306 xlog_write_copy_finish(
2307         struct xlog             *log,
2308         struct xlog_in_core     *iclog,
2309         uint                    flags,
2310         int                     *record_cnt,
2311         int                     *data_cnt,
2312         int                     *partial_copy,
2313         int                     *partial_copy_len,
2314         int                     log_offset,
2315         struct xlog_in_core     **commit_iclog)
2316 {
2317         if (*partial_copy) {
2318                 /*
2319                  * This iclog has already been marked WANT_SYNC by
2320                  * xlog_state_get_iclog_space.
2321                  */
2322                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2323                 *record_cnt = 0;
2324                 *data_cnt = 0;
2325                 return xlog_state_release_iclog(log, iclog);
2326         }
2327
2328         *partial_copy = 0;
2329         *partial_copy_len = 0;
2330
2331         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2332                 /* no more space in this iclog - push it. */
2333                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2334                 *record_cnt = 0;
2335                 *data_cnt = 0;
2336
2337                 spin_lock(&log->l_icloglock);
2338                 xlog_state_want_sync(log, iclog);
2339                 spin_unlock(&log->l_icloglock);
2340
2341                 if (!commit_iclog)
2342                         return xlog_state_release_iclog(log, iclog);
2343                 ASSERT(flags & XLOG_COMMIT_TRANS);
2344                 *commit_iclog = iclog;
2345         }
2346
2347         return 0;
2348 }
2349
2350 /*
2351  * Write some region out to in-core log
2352  *
2353  * This will be called when writing externally provided regions or when
2354  * writing out a commit record for a given transaction.
2355  *
2356  * General algorithm:
2357  *      1. Find total length of this write.  This may include adding to the
2358  *              lengths passed in.
2359  *      2. Check whether we violate the tickets reservation.
2360  *      3. While writing to this iclog
2361  *          A. Reserve as much space in this iclog as can get
2362  *          B. If this is first write, save away start lsn
2363  *          C. While writing this region:
2364  *              1. If first write of transaction, write start record
2365  *              2. Write log operation header (header per region)
2366  *              3. Find out if we can fit entire region into this iclog
2367  *              4. Potentially, verify destination memcpy ptr
2368  *              5. Memcpy (partial) region
2369  *              6. If partial copy, release iclog; otherwise, continue
2370  *                      copying more regions into current iclog
2371  *      4. Mark want sync bit (in simulation mode)
2372  *      5. Release iclog for potential flush to on-disk log.
2373  *
2374  * ERRORS:
2375  * 1.   Panic if reservation is overrun.  This should never happen since
2376  *      reservation amounts are generated internal to the filesystem.
2377  * NOTES:
2378  * 1. Tickets are single threaded data structures.
2379  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2380  *      syncing routine.  When a single log_write region needs to span
2381  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2382  *      on all log operation writes which don't contain the end of the
2383  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2384  *      operation which contains the end of the continued log_write region.
2385  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2386  *      we don't really know exactly how much space will be used.  As a result,
2387  *      we don't update ic_offset until the end when we know exactly how many
2388  *      bytes have been written out.
2389  */
2390 int
2391 xlog_write(
2392         struct xlog             *log,
2393         struct xfs_log_vec      *log_vector,
2394         struct xlog_ticket      *ticket,
2395         xfs_lsn_t               *start_lsn,
2396         struct xlog_in_core     **commit_iclog,
2397         uint                    flags)
2398 {
2399         struct xlog_in_core     *iclog = NULL;
2400         struct xfs_log_iovec    *vecp;
2401         struct xfs_log_vec      *lv;
2402         int                     len;
2403         int                     index;
2404         int                     partial_copy = 0;
2405         int                     partial_copy_len = 0;
2406         int                     contwr = 0;
2407         int                     record_cnt = 0;
2408         int                     data_cnt = 0;
2409         int                     error;
2410
2411         *start_lsn = 0;
2412
2413         len = xlog_write_calc_vec_length(ticket, log_vector);
2414
2415         /*
2416          * Region headers and bytes are already accounted for.
2417          * We only need to take into account start records and
2418          * split regions in this function.
2419          */
2420         if (ticket->t_flags & XLOG_TIC_INITED)
2421                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2422
2423         /*
2424          * Commit record headers need to be accounted for. These
2425          * come in as separate writes so are easy to detect.
2426          */
2427         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2428                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2429
2430         if (ticket->t_curr_res < 0) {
2431                 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2432                      "ctx ticket reservation ran out. Need to up reservation");
2433                 xlog_print_tic_res(log->l_mp, ticket);
2434                 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2435         }
2436
2437         index = 0;
2438         lv = log_vector;
2439         vecp = lv->lv_iovecp;
2440         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2441                 void            *ptr;
2442                 int             log_offset;
2443
2444                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2445                                                    &contwr, &log_offset);
2446                 if (error)
2447                         return error;
2448
2449                 ASSERT(log_offset <= iclog->ic_size - 1);
2450                 ptr = iclog->ic_datap + log_offset;
2451
2452                 /* start_lsn is the first lsn written to. That's all we need. */
2453                 if (!*start_lsn)
2454                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2455
2456                 /*
2457                  * This loop writes out as many regions as can fit in the amount
2458                  * of space which was allocated by xlog_state_get_iclog_space().
2459                  */
2460                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2461                         struct xfs_log_iovec    *reg;
2462                         struct xlog_op_header   *ophdr;
2463                         int                     start_rec_copy;
2464                         int                     copy_len;
2465                         int                     copy_off;
2466                         bool                    ordered = false;
2467
2468                         /* ordered log vectors have no regions to write */
2469                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2470                                 ASSERT(lv->lv_niovecs == 0);
2471                                 ordered = true;
2472                                 goto next_lv;
2473                         }
2474
2475                         reg = &vecp[index];
2476                         ASSERT(reg->i_len % sizeof(int32_t) == 0);
2477                         ASSERT((unsigned long)ptr % sizeof(int32_t) == 0);
2478
2479                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2480                         if (start_rec_copy) {
2481                                 record_cnt++;
2482                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2483                                                    start_rec_copy);
2484                         }
2485
2486                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2487                         if (!ophdr)
2488                                 return -EIO;
2489
2490                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2491                                            sizeof(struct xlog_op_header));
2492
2493                         len += xlog_write_setup_copy(ticket, ophdr,
2494                                                      iclog->ic_size-log_offset,
2495                                                      reg->i_len,
2496                                                      &copy_off, &copy_len,
2497                                                      &partial_copy,
2498                                                      &partial_copy_len);
2499                         xlog_verify_dest_ptr(log, ptr);
2500
2501                         /*
2502                          * Copy region.
2503                          *
2504                          * Unmount records just log an opheader, so can have
2505                          * empty payloads with no data region to copy. Hence we
2506                          * only copy the payload if the vector says it has data
2507                          * to copy.
2508                          */
2509                         ASSERT(copy_len >= 0);
2510                         if (copy_len > 0) {
2511                                 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2512                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2513                                                    copy_len);
2514                         }
2515                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2516                         record_cnt++;
2517                         data_cnt += contwr ? copy_len : 0;
2518
2519                         error = xlog_write_copy_finish(log, iclog, flags,
2520                                                        &record_cnt, &data_cnt,
2521                                                        &partial_copy,
2522                                                        &partial_copy_len,
2523                                                        log_offset,
2524                                                        commit_iclog);
2525                         if (error)
2526                                 return error;
2527
2528                         /*
2529                          * if we had a partial copy, we need to get more iclog
2530                          * space but we don't want to increment the region
2531                          * index because there is still more is this region to
2532                          * write.
2533                          *
2534                          * If we completed writing this region, and we flushed
2535                          * the iclog (indicated by resetting of the record
2536                          * count), then we also need to get more log space. If
2537                          * this was the last record, though, we are done and
2538                          * can just return.
2539                          */
2540                         if (partial_copy)
2541                                 break;
2542
2543                         if (++index == lv->lv_niovecs) {
2544 next_lv:
2545                                 lv = lv->lv_next;
2546                                 index = 0;
2547                                 if (lv)
2548                                         vecp = lv->lv_iovecp;
2549                         }
2550                         if (record_cnt == 0 && !ordered) {
2551                                 if (!lv)
2552                                         return 0;
2553                                 break;
2554                         }
2555                 }
2556         }
2557
2558         ASSERT(len == 0);
2559
2560         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2561         if (!commit_iclog)
2562                 return xlog_state_release_iclog(log, iclog);
2563
2564         ASSERT(flags & XLOG_COMMIT_TRANS);
2565         *commit_iclog = iclog;
2566         return 0;
2567 }
2568
2569
2570 /*****************************************************************************
2571  *
2572  *              State Machine functions
2573  *
2574  *****************************************************************************
2575  */
2576
2577 /* Clean iclogs starting from the head.  This ordering must be
2578  * maintained, so an iclog doesn't become ACTIVE beyond one that
2579  * is SYNCING.  This is also required to maintain the notion that we use
2580  * a ordered wait queue to hold off would be writers to the log when every
2581  * iclog is trying to sync to disk.
2582  *
2583  * State Change: DIRTY -> ACTIVE
2584  */
2585 STATIC void
2586 xlog_state_clean_log(
2587         struct xlog *log)
2588 {
2589         xlog_in_core_t  *iclog;
2590         int changed = 0;
2591
2592         iclog = log->l_iclog;
2593         do {
2594                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2595                         iclog->ic_state = XLOG_STATE_ACTIVE;
2596                         iclog->ic_offset       = 0;
2597                         ASSERT(iclog->ic_callback == NULL);
2598                         /*
2599                          * If the number of ops in this iclog indicate it just
2600                          * contains the dummy transaction, we can
2601                          * change state into IDLE (the second time around).
2602                          * Otherwise we should change the state into
2603                          * NEED a dummy.
2604                          * We don't need to cover the dummy.
2605                          */
2606                         if (!changed &&
2607                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2608                                         XLOG_COVER_OPS)) {
2609                                 changed = 1;
2610                         } else {
2611                                 /*
2612                                  * We have two dirty iclogs so start over
2613                                  * This could also be num of ops indicates
2614                                  * this is not the dummy going out.
2615                                  */
2616                                 changed = 2;
2617                         }
2618                         iclog->ic_header.h_num_logops = 0;
2619                         memset(iclog->ic_header.h_cycle_data, 0,
2620                               sizeof(iclog->ic_header.h_cycle_data));
2621                         iclog->ic_header.h_lsn = 0;
2622                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2623                         /* do nothing */;
2624                 else
2625                         break;  /* stop cleaning */
2626                 iclog = iclog->ic_next;
2627         } while (iclog != log->l_iclog);
2628
2629         /* log is locked when we are called */
2630         /*
2631          * Change state for the dummy log recording.
2632          * We usually go to NEED. But we go to NEED2 if the changed indicates
2633          * we are done writing the dummy record.
2634          * If we are done with the second dummy recored (DONE2), then
2635          * we go to IDLE.
2636          */
2637         if (changed) {
2638                 switch (log->l_covered_state) {
2639                 case XLOG_STATE_COVER_IDLE:
2640                 case XLOG_STATE_COVER_NEED:
2641                 case XLOG_STATE_COVER_NEED2:
2642                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2643                         break;
2644
2645                 case XLOG_STATE_COVER_DONE:
2646                         if (changed == 1)
2647                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2648                         else
2649                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2650                         break;
2651
2652                 case XLOG_STATE_COVER_DONE2:
2653                         if (changed == 1)
2654                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2655                         else
2656                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2657                         break;
2658
2659                 default:
2660                         ASSERT(0);
2661                 }
2662         }
2663 }       /* xlog_state_clean_log */
2664
2665 STATIC xfs_lsn_t
2666 xlog_get_lowest_lsn(
2667         struct xlog     *log)
2668 {
2669         xlog_in_core_t  *lsn_log;
2670         xfs_lsn_t       lowest_lsn, lsn;
2671
2672         lsn_log = log->l_iclog;
2673         lowest_lsn = 0;
2674         do {
2675             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2676                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2677                 if ((lsn && !lowest_lsn) ||
2678                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2679                         lowest_lsn = lsn;
2680                 }
2681             }
2682             lsn_log = lsn_log->ic_next;
2683         } while (lsn_log != log->l_iclog);
2684         return lowest_lsn;
2685 }
2686
2687
2688 STATIC void
2689 xlog_state_do_callback(
2690         struct xlog             *log,
2691         int                     aborted,
2692         struct xlog_in_core     *ciclog)
2693 {
2694         xlog_in_core_t     *iclog;
2695         xlog_in_core_t     *first_iclog;        /* used to know when we've
2696                                                  * processed all iclogs once */
2697         xfs_log_callback_t *cb, *cb_next;
2698         int                flushcnt = 0;
2699         xfs_lsn_t          lowest_lsn;
2700         int                ioerrors;    /* counter: iclogs with errors */
2701         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2702         int                funcdidcallbacks; /* flag: function did callbacks */
2703         int                repeats;     /* for issuing console warnings if
2704                                          * looping too many times */
2705         int                wake = 0;
2706
2707         spin_lock(&log->l_icloglock);
2708         first_iclog = iclog = log->l_iclog;
2709         ioerrors = 0;
2710         funcdidcallbacks = 0;
2711         repeats = 0;
2712
2713         do {
2714                 /*
2715                  * Scan all iclogs starting with the one pointed to by the
2716                  * log.  Reset this starting point each time the log is
2717                  * unlocked (during callbacks).
2718                  *
2719                  * Keep looping through iclogs until one full pass is made
2720                  * without running any callbacks.
2721                  */
2722                 first_iclog = log->l_iclog;
2723                 iclog = log->l_iclog;
2724                 loopdidcallbacks = 0;
2725                 repeats++;
2726
2727                 do {
2728
2729                         /* skip all iclogs in the ACTIVE & DIRTY states */
2730                         if (iclog->ic_state &
2731                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2732                                 iclog = iclog->ic_next;
2733                                 continue;
2734                         }
2735
2736                         /*
2737                          * Between marking a filesystem SHUTDOWN and stopping
2738                          * the log, we do flush all iclogs to disk (if there
2739                          * wasn't a log I/O error). So, we do want things to
2740                          * go smoothly in case of just a SHUTDOWN  w/o a
2741                          * LOG_IO_ERROR.
2742                          */
2743                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2744                                 /*
2745                                  * Can only perform callbacks in order.  Since
2746                                  * this iclog is not in the DONE_SYNC/
2747                                  * DO_CALLBACK state, we skip the rest and
2748                                  * just try to clean up.  If we set our iclog
2749                                  * to DO_CALLBACK, we will not process it when
2750                                  * we retry since a previous iclog is in the
2751                                  * CALLBACK and the state cannot change since
2752                                  * we are holding the l_icloglock.
2753                                  */
2754                                 if (!(iclog->ic_state &
2755                                         (XLOG_STATE_DONE_SYNC |
2756                                                  XLOG_STATE_DO_CALLBACK))) {
2757                                         if (ciclog && (ciclog->ic_state ==
2758                                                         XLOG_STATE_DONE_SYNC)) {
2759                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2760                                         }
2761                                         break;
2762                                 }
2763                                 /*
2764                                  * We now have an iclog that is in either the
2765                                  * DO_CALLBACK or DONE_SYNC states. The other
2766                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2767                                  * caught by the above if and are going to
2768                                  * clean (i.e. we aren't doing their callbacks)
2769                                  * see the above if.
2770                                  */
2771
2772                                 /*
2773                                  * We will do one more check here to see if we
2774                                  * have chased our tail around.
2775                                  */
2776
2777                                 lowest_lsn = xlog_get_lowest_lsn(log);
2778                                 if (lowest_lsn &&
2779                                     XFS_LSN_CMP(lowest_lsn,
2780                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2781                                         iclog = iclog->ic_next;
2782                                         continue; /* Leave this iclog for
2783                                                    * another thread */
2784                                 }
2785
2786                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2787
2788
2789                                 /*
2790                                  * Completion of a iclog IO does not imply that
2791                                  * a transaction has completed, as transactions
2792                                  * can be large enough to span many iclogs. We
2793                                  * cannot change the tail of the log half way
2794                                  * through a transaction as this may be the only
2795                                  * transaction in the log and moving th etail to
2796                                  * point to the middle of it will prevent
2797                                  * recovery from finding the start of the
2798                                  * transaction. Hence we should only update the
2799                                  * last_sync_lsn if this iclog contains
2800                                  * transaction completion callbacks on it.
2801                                  *
2802                                  * We have to do this before we drop the
2803                                  * icloglock to ensure we are the only one that
2804                                  * can update it.
2805                                  */
2806                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2807                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2808                                 if (iclog->ic_callback)
2809                                         atomic64_set(&log->l_last_sync_lsn,
2810                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2811
2812                         } else
2813                                 ioerrors++;
2814
2815                         spin_unlock(&log->l_icloglock);
2816
2817                         /*
2818                          * Keep processing entries in the callback list until
2819                          * we come around and it is empty.  We need to
2820                          * atomically see that the list is empty and change the
2821                          * state to DIRTY so that we don't miss any more
2822                          * callbacks being added.
2823                          */
2824                         spin_lock(&iclog->ic_callback_lock);
2825                         cb = iclog->ic_callback;
2826                         while (cb) {
2827                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2828                                 iclog->ic_callback = NULL;
2829                                 spin_unlock(&iclog->ic_callback_lock);
2830
2831                                 /* perform callbacks in the order given */
2832                                 for (; cb; cb = cb_next) {
2833                                         cb_next = cb->cb_next;
2834                                         cb->cb_func(cb->cb_arg, aborted);
2835                                 }
2836                                 spin_lock(&iclog->ic_callback_lock);
2837                                 cb = iclog->ic_callback;
2838                         }
2839
2840                         loopdidcallbacks++;
2841                         funcdidcallbacks++;
2842
2843                         spin_lock(&log->l_icloglock);
2844                         ASSERT(iclog->ic_callback == NULL);
2845                         spin_unlock(&iclog->ic_callback_lock);
2846                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2847                                 iclog->ic_state = XLOG_STATE_DIRTY;
2848
2849                         /*
2850                          * Transition from DIRTY to ACTIVE if applicable.
2851                          * NOP if STATE_IOERROR.
2852                          */
2853                         xlog_state_clean_log(log);
2854
2855                         /* wake up threads waiting in xfs_log_force() */
2856                         wake_up_all(&iclog->ic_force_wait);
2857
2858                         iclog = iclog->ic_next;
2859                 } while (first_iclog != iclog);
2860
2861                 if (repeats > 5000) {
2862                         flushcnt += repeats;
2863                         repeats = 0;
2864                         xfs_warn(log->l_mp,
2865                                 "%s: possible infinite loop (%d iterations)",
2866                                 __func__, flushcnt);
2867                 }
2868         } while (!ioerrors && loopdidcallbacks);
2869
2870 #ifdef DEBUG
2871         /*
2872          * Make one last gasp attempt to see if iclogs are being left in limbo.
2873          * If the above loop finds an iclog earlier than the current iclog and
2874          * in one of the syncing states, the current iclog is put into
2875          * DO_CALLBACK and the callbacks are deferred to the completion of the
2876          * earlier iclog. Walk the iclogs in order and make sure that no iclog
2877          * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2878          * states.
2879          *
2880          * Note that SYNCING|IOABORT is a valid state so we cannot just check
2881          * for ic_state == SYNCING.
2882          */
2883         if (funcdidcallbacks) {
2884                 first_iclog = iclog = log->l_iclog;
2885                 do {
2886                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2887                         /*
2888                          * Terminate the loop if iclogs are found in states
2889                          * which will cause other threads to clean up iclogs.
2890                          *
2891                          * SYNCING - i/o completion will go through logs
2892                          * DONE_SYNC - interrupt thread should be waiting for
2893                          *              l_icloglock
2894                          * IOERROR - give up hope all ye who enter here
2895                          */
2896                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2897                             iclog->ic_state & XLOG_STATE_SYNCING ||
2898                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2899                             iclog->ic_state == XLOG_STATE_IOERROR )
2900                                 break;
2901                         iclog = iclog->ic_next;
2902                 } while (first_iclog != iclog);
2903         }
2904 #endif
2905
2906         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2907                 wake = 1;
2908         spin_unlock(&log->l_icloglock);
2909
2910         if (wake)
2911                 wake_up_all(&log->l_flush_wait);
2912 }
2913
2914
2915 /*
2916  * Finish transitioning this iclog to the dirty state.
2917  *
2918  * Make sure that we completely execute this routine only when this is
2919  * the last call to the iclog.  There is a good chance that iclog flushes,
2920  * when we reach the end of the physical log, get turned into 2 separate
2921  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2922  * routine.  By using the reference count bwritecnt, we guarantee that only
2923  * the second completion goes through.
2924  *
2925  * Callbacks could take time, so they are done outside the scope of the
2926  * global state machine log lock.
2927  */
2928 STATIC void
2929 xlog_state_done_syncing(
2930         xlog_in_core_t  *iclog,
2931         int             aborted)
2932 {
2933         struct xlog        *log = iclog->ic_log;
2934
2935         spin_lock(&log->l_icloglock);
2936
2937         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2938                iclog->ic_state == XLOG_STATE_IOERROR);
2939         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2940         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2941
2942
2943         /*
2944          * If we got an error, either on the first buffer, or in the case of
2945          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2946          * and none should ever be attempted to be written to disk
2947          * again.
2948          */
2949         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2950                 if (--iclog->ic_bwritecnt == 1) {
2951                         spin_unlock(&log->l_icloglock);
2952                         return;
2953                 }
2954                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2955         }
2956
2957         /*
2958          * Someone could be sleeping prior to writing out the next
2959          * iclog buffer, we wake them all, one will get to do the
2960          * I/O, the others get to wait for the result.
2961          */
2962         wake_up_all(&iclog->ic_write_wait);
2963         spin_unlock(&log->l_icloglock);
2964         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2965 }       /* xlog_state_done_syncing */
2966
2967
2968 /*
2969  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2970  * sleep.  We wait on the flush queue on the head iclog as that should be
2971  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2972  * we will wait here and all new writes will sleep until a sync completes.
2973  *
2974  * The in-core logs are used in a circular fashion. They are not used
2975  * out-of-order even when an iclog past the head is free.
2976  *
2977  * return:
2978  *      * log_offset where xlog_write() can start writing into the in-core
2979  *              log's data space.
2980  *      * in-core log pointer to which xlog_write() should write.
2981  *      * boolean indicating this is a continued write to an in-core log.
2982  *              If this is the last write, then the in-core log's offset field
2983  *              needs to be incremented, depending on the amount of data which
2984  *              is copied.
2985  */
2986 STATIC int
2987 xlog_state_get_iclog_space(
2988         struct xlog             *log,
2989         int                     len,
2990         struct xlog_in_core     **iclogp,
2991         struct xlog_ticket      *ticket,
2992         int                     *continued_write,
2993         int                     *logoffsetp)
2994 {
2995         int               log_offset;
2996         xlog_rec_header_t *head;
2997         xlog_in_core_t    *iclog;
2998         int               error;
2999
3000 restart:
3001         spin_lock(&log->l_icloglock);
3002         if (XLOG_FORCED_SHUTDOWN(log)) {
3003                 spin_unlock(&log->l_icloglock);
3004                 return -EIO;
3005         }
3006
3007         iclog = log->l_iclog;
3008         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
3009                 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
3010
3011                 /* Wait for log writes to have flushed */
3012                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
3013                 goto restart;
3014         }
3015
3016         head = &iclog->ic_header;
3017
3018         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
3019         log_offset = iclog->ic_offset;
3020
3021         /* On the 1st write to an iclog, figure out lsn.  This works
3022          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
3023          * committing to.  If the offset is set, that's how many blocks
3024          * must be written.
3025          */
3026         if (log_offset == 0) {
3027                 ticket->t_curr_res -= log->l_iclog_hsize;
3028                 xlog_tic_add_region(ticket,
3029                                     log->l_iclog_hsize,
3030                                     XLOG_REG_TYPE_LRHEADER);
3031                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
3032                 head->h_lsn = cpu_to_be64(
3033                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
3034                 ASSERT(log->l_curr_block >= 0);
3035         }
3036
3037         /* If there is enough room to write everything, then do it.  Otherwise,
3038          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
3039          * bit is on, so this will get flushed out.  Don't update ic_offset
3040          * until you know exactly how many bytes get copied.  Therefore, wait
3041          * until later to update ic_offset.
3042          *
3043          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
3044          * can fit into remaining data section.
3045          */
3046         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
3047                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3048
3049                 /*
3050                  * If I'm the only one writing to this iclog, sync it to disk.
3051                  * We need to do an atomic compare and decrement here to avoid
3052                  * racing with concurrent atomic_dec_and_lock() calls in
3053                  * xlog_state_release_iclog() when there is more than one
3054                  * reference to the iclog.
3055                  */
3056                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
3057                         /* we are the only one */
3058                         spin_unlock(&log->l_icloglock);
3059                         error = xlog_state_release_iclog(log, iclog);
3060                         if (error)
3061                                 return error;
3062                 } else {
3063                         spin_unlock(&log->l_icloglock);
3064                 }
3065                 goto restart;
3066         }
3067
3068         /* Do we have enough room to write the full amount in the remainder
3069          * of this iclog?  Or must we continue a write on the next iclog and
3070          * mark this iclog as completely taken?  In the case where we switch
3071          * iclogs (to mark it taken), this particular iclog will release/sync
3072          * to disk in xlog_write().
3073          */
3074         if (len <= iclog->ic_size - iclog->ic_offset) {
3075                 *continued_write = 0;
3076                 iclog->ic_offset += len;
3077         } else {
3078                 *continued_write = 1;
3079                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
3080         }
3081         *iclogp = iclog;
3082
3083         ASSERT(iclog->ic_offset <= iclog->ic_size);
3084         spin_unlock(&log->l_icloglock);
3085
3086         *logoffsetp = log_offset;
3087         return 0;
3088 }       /* xlog_state_get_iclog_space */
3089
3090 /* The first cnt-1 times through here we don't need to
3091  * move the grant write head because the permanent
3092  * reservation has reserved cnt times the unit amount.
3093  * Release part of current permanent unit reservation and
3094  * reset current reservation to be one units worth.  Also
3095  * move grant reservation head forward.
3096  */
3097 STATIC void
3098 xlog_regrant_reserve_log_space(
3099         struct xlog             *log,
3100         struct xlog_ticket      *ticket)
3101 {
3102         trace_xfs_log_regrant_reserve_enter(log, ticket);
3103
3104         if (ticket->t_cnt > 0)
3105                 ticket->t_cnt--;
3106
3107         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3108                                         ticket->t_curr_res);
3109         xlog_grant_sub_space(log, &log->l_write_head.grant,
3110                                         ticket->t_curr_res);
3111         ticket->t_curr_res = ticket->t_unit_res;
3112         xlog_tic_reset_res(ticket);
3113
3114         trace_xfs_log_regrant_reserve_sub(log, ticket);
3115
3116         /* just return if we still have some of the pre-reserved space */
3117         if (ticket->t_cnt > 0)
3118                 return;
3119
3120         xlog_grant_add_space(log, &log->l_reserve_head.grant,
3121                                         ticket->t_unit_res);
3122
3123         trace_xfs_log_regrant_reserve_exit(log, ticket);
3124
3125         ticket->t_curr_res = ticket->t_unit_res;
3126         xlog_tic_reset_res(ticket);
3127 }       /* xlog_regrant_reserve_log_space */
3128
3129
3130 /*
3131  * Give back the space left from a reservation.
3132  *
3133  * All the information we need to make a correct determination of space left
3134  * is present.  For non-permanent reservations, things are quite easy.  The
3135  * count should have been decremented to zero.  We only need to deal with the
3136  * space remaining in the current reservation part of the ticket.  If the
3137  * ticket contains a permanent reservation, there may be left over space which
3138  * needs to be released.  A count of N means that N-1 refills of the current
3139  * reservation can be done before we need to ask for more space.  The first
3140  * one goes to fill up the first current reservation.  Once we run out of
3141  * space, the count will stay at zero and the only space remaining will be
3142  * in the current reservation field.
3143  */
3144 STATIC void
3145 xlog_ungrant_log_space(
3146         struct xlog             *log,
3147         struct xlog_ticket      *ticket)
3148 {
3149         int     bytes;
3150
3151         if (ticket->t_cnt > 0)
3152                 ticket->t_cnt--;
3153
3154         trace_xfs_log_ungrant_enter(log, ticket);
3155         trace_xfs_log_ungrant_sub(log, ticket);
3156
3157         /*
3158          * If this is a permanent reservation ticket, we may be able to free
3159          * up more space based on the remaining count.
3160          */
3161         bytes = ticket->t_curr_res;
3162         if (ticket->t_cnt > 0) {
3163                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3164                 bytes += ticket->t_unit_res*ticket->t_cnt;
3165         }
3166
3167         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3168         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3169
3170         trace_xfs_log_ungrant_exit(log, ticket);
3171
3172         xfs_log_space_wake(log->l_mp);
3173 }
3174
3175 /*
3176  * Flush iclog to disk if this is the last reference to the given iclog and
3177  * the WANT_SYNC bit is set.
3178  *
3179  * When this function is entered, the iclog is not necessarily in the
3180  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3181  *
3182  *
3183  */
3184 STATIC int
3185 xlog_state_release_iclog(
3186         struct xlog             *log,
3187         struct xlog_in_core     *iclog)
3188 {
3189         int             sync = 0;       /* do we sync? */
3190
3191         if (iclog->ic_state & XLOG_STATE_IOERROR)
3192                 return -EIO;
3193
3194         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3195         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3196                 return 0;
3197
3198         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3199                 spin_unlock(&log->l_icloglock);
3200                 return -EIO;
3201         }
3202         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3203                iclog->ic_state == XLOG_STATE_WANT_SYNC);
3204
3205         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3206                 /* update tail before writing to iclog */
3207                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3208                 sync++;
3209                 iclog->ic_state = XLOG_STATE_SYNCING;
3210                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3211                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3212                 /* cycle incremented when incrementing curr_block */
3213         }
3214         spin_unlock(&log->l_icloglock);
3215
3216         /*
3217          * We let the log lock go, so it's possible that we hit a log I/O
3218          * error or some other SHUTDOWN condition that marks the iclog
3219          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3220          * this iclog has consistent data, so we ignore IOERROR
3221          * flags after this point.
3222          */
3223         if (sync)
3224                 return xlog_sync(log, iclog);
3225         return 0;
3226 }       /* xlog_state_release_iclog */
3227
3228
3229 /*
3230  * This routine will mark the current iclog in the ring as WANT_SYNC
3231  * and move the current iclog pointer to the next iclog in the ring.
3232  * When this routine is called from xlog_state_get_iclog_space(), the
3233  * exact size of the iclog has not yet been determined.  All we know is
3234  * that every data block.  We have run out of space in this log record.
3235  */
3236 STATIC void
3237 xlog_state_switch_iclogs(
3238         struct xlog             *log,
3239         struct xlog_in_core     *iclog,
3240         int                     eventual_size)
3241 {
3242         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3243         if (!eventual_size)
3244                 eventual_size = iclog->ic_offset;
3245         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3246         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3247         log->l_prev_block = log->l_curr_block;
3248         log->l_prev_cycle = log->l_curr_cycle;
3249
3250         /* roll log?: ic_offset changed later */
3251         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3252
3253         /* Round up to next log-sunit */
3254         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3255             log->l_mp->m_sb.sb_logsunit > 1) {
3256                 uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3257                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3258         }
3259
3260         if (log->l_curr_block >= log->l_logBBsize) {
3261                 /*
3262                  * Rewind the current block before the cycle is bumped to make
3263                  * sure that the combined LSN never transiently moves forward
3264                  * when the log wraps to the next cycle. This is to support the
3265                  * unlocked sample of these fields from xlog_valid_lsn(). Most
3266                  * other cases should acquire l_icloglock.
3267                  */
3268                 log->l_curr_block -= log->l_logBBsize;
3269                 ASSERT(log->l_curr_block >= 0);
3270                 smp_wmb();
3271                 log->l_curr_cycle++;
3272                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3273                         log->l_curr_cycle++;
3274         }
3275         ASSERT(iclog == log->l_iclog);
3276         log->l_iclog = iclog->ic_next;
3277 }       /* xlog_state_switch_iclogs */
3278
3279 /*
3280  * Write out all data in the in-core log as of this exact moment in time.
3281  *
3282  * Data may be written to the in-core log during this call.  However,
3283  * we don't guarantee this data will be written out.  A change from past
3284  * implementation means this routine will *not* write out zero length LRs.
3285  *
3286  * Basically, we try and perform an intelligent scan of the in-core logs.
3287  * If we determine there is no flushable data, we just return.  There is no
3288  * flushable data if:
3289  *
3290  *      1. the current iclog is active and has no data; the previous iclog
3291  *              is in the active or dirty state.
3292  *      2. the current iclog is drity, and the previous iclog is in the
3293  *              active or dirty state.
3294  *
3295  * We may sleep if:
3296  *
3297  *      1. the current iclog is not in the active nor dirty state.
3298  *      2. the current iclog dirty, and the previous iclog is not in the
3299  *              active nor dirty state.
3300  *      3. the current iclog is active, and there is another thread writing
3301  *              to this particular iclog.
3302  *      4. a) the current iclog is active and has no other writers
3303  *         b) when we return from flushing out this iclog, it is still
3304  *              not in the active nor dirty state.
3305  */
3306 int
3307 _xfs_log_force(
3308         struct xfs_mount        *mp,
3309         uint                    flags,
3310         int                     *log_flushed)
3311 {
3312         struct xlog             *log = mp->m_log;
3313         struct xlog_in_core     *iclog;
3314         xfs_lsn_t               lsn;
3315
3316         XFS_STATS_INC(mp, xs_log_force);
3317
3318         xlog_cil_force(log);
3319
3320         spin_lock(&log->l_icloglock);
3321
3322         iclog = log->l_iclog;
3323         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3324                 spin_unlock(&log->l_icloglock);
3325                 return -EIO;
3326         }
3327
3328         /* If the head iclog is not active nor dirty, we just attach
3329          * ourselves to the head and go to sleep.
3330          */
3331         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3332             iclog->ic_state == XLOG_STATE_DIRTY) {
3333                 /*
3334                  * If the head is dirty or (active and empty), then
3335                  * we need to look at the previous iclog.  If the previous
3336                  * iclog is active or dirty we are done.  There is nothing
3337                  * to sync out.  Otherwise, we attach ourselves to the
3338                  * previous iclog and go to sleep.
3339                  */
3340                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3341                     (atomic_read(&iclog->ic_refcnt) == 0
3342                      && iclog->ic_offset == 0)) {
3343                         iclog = iclog->ic_prev;
3344                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3345                             iclog->ic_state == XLOG_STATE_DIRTY)
3346                                 goto no_sleep;
3347                         else
3348                                 goto maybe_sleep;
3349                 } else {
3350                         if (atomic_read(&iclog->ic_refcnt) == 0) {
3351                                 /* We are the only one with access to this
3352                                  * iclog.  Flush it out now.  There should
3353                                  * be a roundoff of zero to show that someone
3354                                  * has already taken care of the roundoff from
3355                                  * the previous sync.
3356                                  */
3357                                 atomic_inc(&iclog->ic_refcnt);
3358                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3359                                 xlog_state_switch_iclogs(log, iclog, 0);
3360                                 spin_unlock(&log->l_icloglock);
3361
3362                                 if (xlog_state_release_iclog(log, iclog))
3363                                         return -EIO;
3364
3365                                 if (log_flushed)
3366                                         *log_flushed = 1;
3367                                 spin_lock(&log->l_icloglock);
3368                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3369                                     iclog->ic_state != XLOG_STATE_DIRTY)
3370                                         goto maybe_sleep;
3371                                 else
3372                                         goto no_sleep;
3373                         } else {
3374                                 /* Someone else is writing to this iclog.
3375                                  * Use its call to flush out the data.  However,
3376                                  * the other thread may not force out this LR,
3377                                  * so we mark it WANT_SYNC.
3378                                  */
3379                                 xlog_state_switch_iclogs(log, iclog, 0);
3380                                 goto maybe_sleep;
3381                         }
3382                 }
3383         }
3384
3385         /* By the time we come around again, the iclog could've been filled
3386          * which would give it another lsn.  If we have a new lsn, just
3387          * return because the relevant data has been flushed.
3388          */
3389 maybe_sleep:
3390         if (flags & XFS_LOG_SYNC) {
3391                 /*
3392                  * We must check if we're shutting down here, before
3393                  * we wait, while we're holding the l_icloglock.
3394                  * Then we check again after waking up, in case our
3395                  * sleep was disturbed by a bad news.
3396                  */
3397                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3398                         spin_unlock(&log->l_icloglock);
3399                         return -EIO;
3400                 }
3401                 XFS_STATS_INC(mp, xs_log_force_sleep);
3402                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3403                 /*
3404                  * No need to grab the log lock here since we're
3405                  * only deciding whether or not to return EIO
3406                  * and the memory read should be atomic.
3407                  */
3408                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3409                         return -EIO;
3410         } else {
3411
3412 no_sleep:
3413                 spin_unlock(&log->l_icloglock);
3414         }
3415         return 0;
3416 }
3417
3418 /*
3419  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3420  * about errors or whether the log was flushed or not. This is the normal
3421  * interface to use when trying to unpin items or move the log forward.
3422  */
3423 void
3424 xfs_log_force(
3425         xfs_mount_t     *mp,
3426         uint            flags)
3427 {
3428         trace_xfs_log_force(mp, 0, _RET_IP_);
3429         _xfs_log_force(mp, flags, NULL);
3430 }
3431
3432 /*
3433  * Force the in-core log to disk for a specific LSN.
3434  *
3435  * Find in-core log with lsn.
3436  *      If it is in the DIRTY state, just return.
3437  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3438  *              state and go to sleep or return.
3439  *      If it is in any other state, go to sleep or return.
3440  *
3441  * Synchronous forces are implemented with a signal variable. All callers
3442  * to force a given lsn to disk will wait on a the sv attached to the
3443  * specific in-core log.  When given in-core log finally completes its
3444  * write to disk, that thread will wake up all threads waiting on the
3445  * sv.
3446  */
3447 int
3448 _xfs_log_force_lsn(
3449         struct xfs_mount        *mp,
3450         xfs_lsn_t               lsn,
3451         uint                    flags,
3452         int                     *log_flushed)
3453 {
3454         struct xlog             *log = mp->m_log;
3455         struct xlog_in_core     *iclog;
3456         int                     already_slept = 0;
3457
3458         ASSERT(lsn != 0);
3459
3460         XFS_STATS_INC(mp, xs_log_force);
3461
3462         lsn = xlog_cil_force_lsn(log, lsn);
3463         if (lsn == NULLCOMMITLSN)
3464                 return 0;
3465
3466 try_again:
3467         spin_lock(&log->l_icloglock);
3468         iclog = log->l_iclog;
3469         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3470                 spin_unlock(&log->l_icloglock);
3471                 return -EIO;
3472         }
3473
3474         do {
3475                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3476                         iclog = iclog->ic_next;
3477                         continue;
3478                 }
3479
3480                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3481                         spin_unlock(&log->l_icloglock);
3482                         return 0;
3483                 }
3484
3485                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3486                         /*
3487                          * We sleep here if we haven't already slept (e.g.
3488                          * this is the first time we've looked at the correct
3489                          * iclog buf) and the buffer before us is going to
3490                          * be sync'ed. The reason for this is that if we
3491                          * are doing sync transactions here, by waiting for
3492                          * the previous I/O to complete, we can allow a few
3493                          * more transactions into this iclog before we close
3494                          * it down.
3495                          *
3496                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3497                          * up the refcnt so we can release the log (which
3498                          * drops the ref count).  The state switch keeps new
3499                          * transaction commits from using this buffer.  When
3500                          * the current commits finish writing into the buffer,
3501                          * the refcount will drop to zero and the buffer will
3502                          * go out then.
3503                          */
3504                         if (!already_slept &&
3505                             (iclog->ic_prev->ic_state &
3506                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3507                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3508
3509                                 XFS_STATS_INC(mp, xs_log_force_sleep);
3510
3511                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3512                                                         &log->l_icloglock);
3513                                 already_slept = 1;
3514                                 goto try_again;
3515                         }
3516                         atomic_inc(&iclog->ic_refcnt);
3517                         xlog_state_switch_iclogs(log, iclog, 0);
3518                         spin_unlock(&log->l_icloglock);
3519                         if (xlog_state_release_iclog(log, iclog))
3520                                 return -EIO;
3521                         if (log_flushed)
3522                                 *log_flushed = 1;
3523                         spin_lock(&log->l_icloglock);
3524                 }
3525
3526                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3527                     !(iclog->ic_state &
3528                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3529                         /*
3530                          * Don't wait on completion if we know that we've
3531                          * gotten a log write error.
3532                          */
3533                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3534                                 spin_unlock(&log->l_icloglock);
3535                                 return -EIO;
3536                         }
3537                         XFS_STATS_INC(mp, xs_log_force_sleep);
3538                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3539                         /*
3540                          * No need to grab the log lock here since we're
3541                          * only deciding whether or not to return EIO
3542                          * and the memory read should be atomic.
3543                          */
3544                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3545                                 return -EIO;
3546                 } else {                /* just return */
3547                         spin_unlock(&log->l_icloglock);
3548                 }
3549
3550                 return 0;
3551         } while (iclog != log->l_iclog);
3552
3553         spin_unlock(&log->l_icloglock);
3554         return 0;
3555 }
3556
3557 /*
3558  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3559  * about errors or whether the log was flushed or not. This is the normal
3560  * interface to use when trying to unpin items or move the log forward.
3561  */
3562 void
3563 xfs_log_force_lsn(
3564         xfs_mount_t     *mp,
3565         xfs_lsn_t       lsn,
3566         uint            flags)
3567 {
3568         trace_xfs_log_force(mp, lsn, _RET_IP_);
3569         _xfs_log_force_lsn(mp, lsn, flags, NULL);
3570 }
3571
3572 /*
3573  * Called when we want to mark the current iclog as being ready to sync to
3574  * disk.
3575  */
3576 STATIC void
3577 xlog_state_want_sync(
3578         struct xlog             *log,
3579         struct xlog_in_core     *iclog)
3580 {
3581         assert_spin_locked(&log->l_icloglock);
3582
3583         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3584                 xlog_state_switch_iclogs(log, iclog, 0);
3585         } else {
3586                 ASSERT(iclog->ic_state &
3587                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3588         }
3589 }
3590
3591
3592 /*****************************************************************************
3593  *
3594  *              TICKET functions
3595  *
3596  *****************************************************************************
3597  */
3598
3599 /*
3600  * Free a used ticket when its refcount falls to zero.
3601  */
3602 void
3603 xfs_log_ticket_put(
3604         xlog_ticket_t   *ticket)
3605 {
3606         ASSERT(atomic_read(&ticket->t_ref) > 0);
3607         if (atomic_dec_and_test(&ticket->t_ref))
3608                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3609 }
3610
3611 xlog_ticket_t *
3612 xfs_log_ticket_get(
3613         xlog_ticket_t   *ticket)
3614 {
3615         ASSERT(atomic_read(&ticket->t_ref) > 0);
3616         atomic_inc(&ticket->t_ref);
3617         return ticket;
3618 }
3619
3620 /*
3621  * Figure out the total log space unit (in bytes) that would be
3622  * required for a log ticket.
3623  */
3624 int
3625 xfs_log_calc_unit_res(
3626         struct xfs_mount        *mp,
3627         int                     unit_bytes)
3628 {
3629         struct xlog             *log = mp->m_log;
3630         int                     iclog_space;
3631         uint                    num_headers;
3632
3633         /*
3634          * Permanent reservations have up to 'cnt'-1 active log operations
3635          * in the log.  A unit in this case is the amount of space for one
3636          * of these log operations.  Normal reservations have a cnt of 1
3637          * and their unit amount is the total amount of space required.
3638          *
3639          * The following lines of code account for non-transaction data
3640          * which occupy space in the on-disk log.
3641          *
3642          * Normal form of a transaction is:
3643          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3644          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3645          *
3646          * We need to account for all the leadup data and trailer data
3647          * around the transaction data.
3648          * And then we need to account for the worst case in terms of using
3649          * more space.
3650          * The worst case will happen if:
3651          * - the placement of the transaction happens to be such that the
3652          *   roundoff is at its maximum
3653          * - the transaction data is synced before the commit record is synced
3654          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3655          *   Therefore the commit record is in its own Log Record.
3656          *   This can happen as the commit record is called with its
3657          *   own region to xlog_write().
3658          *   This then means that in the worst case, roundoff can happen for
3659          *   the commit-rec as well.
3660          *   The commit-rec is smaller than padding in this scenario and so it is
3661          *   not added separately.
3662          */
3663
3664         /* for trans header */
3665         unit_bytes += sizeof(xlog_op_header_t);
3666         unit_bytes += sizeof(xfs_trans_header_t);
3667
3668         /* for start-rec */
3669         unit_bytes += sizeof(xlog_op_header_t);
3670
3671         /*
3672          * for LR headers - the space for data in an iclog is the size minus
3673          * the space used for the headers. If we use the iclog size, then we
3674          * undercalculate the number of headers required.
3675          *
3676          * Furthermore - the addition of op headers for split-recs might
3677          * increase the space required enough to require more log and op
3678          * headers, so take that into account too.
3679          *
3680          * IMPORTANT: This reservation makes the assumption that if this
3681          * transaction is the first in an iclog and hence has the LR headers
3682          * accounted to it, then the remaining space in the iclog is
3683          * exclusively for this transaction.  i.e. if the transaction is larger
3684          * than the iclog, it will be the only thing in that iclog.
3685          * Fundamentally, this means we must pass the entire log vector to
3686          * xlog_write to guarantee this.
3687          */
3688         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3689         num_headers = howmany(unit_bytes, iclog_space);
3690
3691         /* for split-recs - ophdrs added when data split over LRs */
3692         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3693
3694         /* add extra header reservations if we overrun */
3695         while (!num_headers ||
3696                howmany(unit_bytes, iclog_space) > num_headers) {
3697                 unit_bytes += sizeof(xlog_op_header_t);
3698                 num_headers++;
3699         }
3700         unit_bytes += log->l_iclog_hsize * num_headers;
3701
3702         /* for commit-rec LR header - note: padding will subsume the ophdr */
3703         unit_bytes += log->l_iclog_hsize;
3704
3705         /* for roundoff padding for transaction data and one for commit record */
3706         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3707                 /* log su roundoff */
3708                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3709         } else {
3710                 /* BB roundoff */
3711                 unit_bytes += 2 * BBSIZE;
3712         }
3713
3714         return unit_bytes;
3715 }
3716
3717 /*
3718  * Allocate and initialise a new log ticket.
3719  */
3720 struct xlog_ticket *
3721 xlog_ticket_alloc(
3722         struct xlog             *log,
3723         int                     unit_bytes,
3724         int                     cnt,
3725         char                    client,
3726         bool                    permanent,
3727         xfs_km_flags_t          alloc_flags)
3728 {
3729         struct xlog_ticket      *tic;
3730         int                     unit_res;
3731
3732         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3733         if (!tic)
3734                 return NULL;
3735
3736         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3737
3738         atomic_set(&tic->t_ref, 1);
3739         tic->t_task             = current;
3740         INIT_LIST_HEAD(&tic->t_queue);
3741         tic->t_unit_res         = unit_res;
3742         tic->t_curr_res         = unit_res;
3743         tic->t_cnt              = cnt;
3744         tic->t_ocnt             = cnt;
3745         tic->t_tid              = prandom_u32();
3746         tic->t_clientid         = client;
3747         tic->t_flags            = XLOG_TIC_INITED;
3748         if (permanent)
3749                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3750
3751         xlog_tic_reset_res(tic);
3752
3753         return tic;
3754 }
3755
3756
3757 /******************************************************************************
3758  *
3759  *              Log debug routines
3760  *
3761  ******************************************************************************
3762  */
3763 #if defined(DEBUG)
3764 /*
3765  * Make sure that the destination ptr is within the valid data region of
3766  * one of the iclogs.  This uses backup pointers stored in a different
3767  * part of the log in case we trash the log structure.
3768  */
3769 STATIC void
3770 xlog_verify_dest_ptr(
3771         struct xlog     *log,
3772         void            *ptr)
3773 {
3774         int i;
3775         int good_ptr = 0;
3776
3777         for (i = 0; i < log->l_iclog_bufs; i++) {
3778                 if (ptr >= log->l_iclog_bak[i] &&
3779                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3780                         good_ptr++;
3781         }
3782
3783         if (!good_ptr)
3784                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3785 }
3786
3787 /*
3788  * Check to make sure the grant write head didn't just over lap the tail.  If
3789  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3790  * the cycles differ by exactly one and check the byte count.
3791  *
3792  * This check is run unlocked, so can give false positives. Rather than assert
3793  * on failures, use a warn-once flag and a panic tag to allow the admin to
3794  * determine if they want to panic the machine when such an error occurs. For
3795  * debug kernels this will have the same effect as using an assert but, unlinke
3796  * an assert, it can be turned off at runtime.
3797  */
3798 STATIC void
3799 xlog_verify_grant_tail(
3800         struct xlog     *log)
3801 {
3802         int             tail_cycle, tail_blocks;
3803         int             cycle, space;
3804
3805         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3806         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3807         if (tail_cycle != cycle) {
3808                 if (cycle - 1 != tail_cycle &&
3809                     !(log->l_flags & XLOG_TAIL_WARN)) {
3810                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3811                                 "%s: cycle - 1 != tail_cycle", __func__);
3812                         log->l_flags |= XLOG_TAIL_WARN;
3813                 }
3814
3815                 if (space > BBTOB(tail_blocks) &&
3816                     !(log->l_flags & XLOG_TAIL_WARN)) {
3817                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3818                                 "%s: space > BBTOB(tail_blocks)", __func__);
3819                         log->l_flags |= XLOG_TAIL_WARN;
3820                 }
3821         }
3822 }
3823
3824 /* check if it will fit */
3825 STATIC void
3826 xlog_verify_tail_lsn(
3827         struct xlog             *log,
3828         struct xlog_in_core     *iclog,
3829         xfs_lsn_t               tail_lsn)
3830 {
3831     int blocks;
3832
3833     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3834         blocks =
3835             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3836         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3837                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3838     } else {
3839         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3840
3841         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3842                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3843
3844         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3845         if (blocks < BTOBB(iclog->ic_offset) + 1)
3846                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3847     }
3848 }       /* xlog_verify_tail_lsn */
3849
3850 /*
3851  * Perform a number of checks on the iclog before writing to disk.
3852  *
3853  * 1. Make sure the iclogs are still circular
3854  * 2. Make sure we have a good magic number
3855  * 3. Make sure we don't have magic numbers in the data
3856  * 4. Check fields of each log operation header for:
3857  *      A. Valid client identifier
3858  *      B. tid ptr value falls in valid ptr space (user space code)
3859  *      C. Length in log record header is correct according to the
3860  *              individual operation headers within record.
3861  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3862  *      log, check the preceding blocks of the physical log to make sure all
3863  *      the cycle numbers agree with the current cycle number.
3864  */
3865 STATIC void
3866 xlog_verify_iclog(
3867         struct xlog             *log,
3868         struct xlog_in_core     *iclog,
3869         int                     count,
3870         bool                    syncing)
3871 {
3872         xlog_op_header_t        *ophead;
3873         xlog_in_core_t          *icptr;
3874         xlog_in_core_2_t        *xhdr;
3875         void                    *base_ptr, *ptr, *p;
3876         ptrdiff_t               field_offset;
3877         uint8_t                 clientid;
3878         int                     len, i, j, k, op_len;
3879         int                     idx;
3880
3881         /* check validity of iclog pointers */
3882         spin_lock(&log->l_icloglock);
3883         icptr = log->l_iclog;
3884         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3885                 ASSERT(icptr);
3886
3887         if (icptr != log->l_iclog)
3888                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3889         spin_unlock(&log->l_icloglock);
3890
3891         /* check log magic numbers */
3892         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3893                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3894
3895         base_ptr = ptr = &iclog->ic_header;
3896         p = &iclog->ic_header;
3897         for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3898                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3899                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3900                                 __func__);
3901         }
3902
3903         /* check fields */
3904         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3905         base_ptr = ptr = iclog->ic_datap;
3906         ophead = ptr;
3907         xhdr = iclog->ic_data;
3908         for (i = 0; i < len; i++) {
3909                 ophead = ptr;
3910
3911                 /* clientid is only 1 byte */
3912                 p = &ophead->oh_clientid;
3913                 field_offset = p - base_ptr;
3914                 if (!syncing || (field_offset & 0x1ff)) {
3915                         clientid = ophead->oh_clientid;
3916                 } else {
3917                         idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3918                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3919                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3920                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3921                                 clientid = xlog_get_client_id(
3922                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3923                         } else {
3924                                 clientid = xlog_get_client_id(
3925                                         iclog->ic_header.h_cycle_data[idx]);
3926                         }
3927                 }
3928                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3929                         xfs_warn(log->l_mp,
3930                                 "%s: invalid clientid %d op "PTR_FMT" offset 0x%lx",
3931                                 __func__, clientid, ophead,
3932                                 (unsigned long)field_offset);
3933
3934                 /* check length */
3935                 p = &ophead->oh_len;
3936                 field_offset = p - base_ptr;
3937                 if (!syncing || (field_offset & 0x1ff)) {
3938                         op_len = be32_to_cpu(ophead->oh_len);
3939                 } else {
3940                         idx = BTOBBT((uintptr_t)&ophead->oh_len -
3941                                     (uintptr_t)iclog->ic_datap);
3942                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3943                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3944                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3945                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3946                         } else {
3947                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3948                         }
3949                 }
3950                 ptr += sizeof(xlog_op_header_t) + op_len;
3951         }
3952 }       /* xlog_verify_iclog */
3953 #endif
3954
3955 /*
3956  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3957  */
3958 STATIC int
3959 xlog_state_ioerror(
3960         struct xlog     *log)
3961 {
3962         xlog_in_core_t  *iclog, *ic;
3963
3964         iclog = log->l_iclog;
3965         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3966                 /*
3967                  * Mark all the incore logs IOERROR.
3968                  * From now on, no log flushes will result.
3969                  */
3970                 ic = iclog;
3971                 do {
3972                         ic->ic_state = XLOG_STATE_IOERROR;
3973                         ic = ic->ic_next;
3974                 } while (ic != iclog);
3975                 return 0;
3976         }
3977         /*
3978          * Return non-zero, if state transition has already happened.
3979          */
3980         return 1;
3981 }
3982
3983 /*
3984  * This is called from xfs_force_shutdown, when we're forcibly
3985  * shutting down the filesystem, typically because of an IO error.
3986  * Our main objectives here are to make sure that:
3987  *      a. if !logerror, flush the logs to disk. Anything modified
3988  *         after this is ignored.
3989  *      b. the filesystem gets marked 'SHUTDOWN' for all interested
3990  *         parties to find out, 'atomically'.
3991  *      c. those who're sleeping on log reservations, pinned objects and
3992  *          other resources get woken up, and be told the bad news.
3993  *      d. nothing new gets queued up after (b) and (c) are done.
3994  *
3995  * Note: for the !logerror case we need to flush the regions held in memory out
3996  * to disk first. This needs to be done before the log is marked as shutdown,
3997  * otherwise the iclog writes will fail.
3998  */
3999 int
4000 xfs_log_force_umount(
4001         struct xfs_mount        *mp,
4002         int                     logerror)
4003 {
4004         struct xlog     *log;
4005         int             retval;
4006
4007         log = mp->m_log;
4008
4009         /*
4010          * If this happens during log recovery, don't worry about
4011          * locking; the log isn't open for business yet.
4012          */
4013         if (!log ||
4014             log->l_flags & XLOG_ACTIVE_RECOVERY) {
4015                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
4016                 if (mp->m_sb_bp)
4017                         mp->m_sb_bp->b_flags |= XBF_DONE;
4018                 return 0;
4019         }
4020
4021         /*
4022          * Somebody could've already done the hard work for us.
4023          * No need to get locks for this.
4024          */
4025         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
4026                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
4027                 return 1;
4028         }
4029
4030         /*
4031          * Flush all the completed transactions to disk before marking the log
4032          * being shut down. We need to do it in this order to ensure that
4033          * completed operations are safely on disk before we shut down, and that
4034          * we don't have to issue any buffer IO after the shutdown flags are set
4035          * to guarantee this.
4036          */
4037         if (!logerror)
4038                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
4039
4040         /*
4041          * mark the filesystem and the as in a shutdown state and wake
4042          * everybody up to tell them the bad news.
4043          */
4044         spin_lock(&log->l_icloglock);
4045         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
4046         if (mp->m_sb_bp)
4047                 mp->m_sb_bp->b_flags |= XBF_DONE;
4048
4049         /*
4050          * Mark the log and the iclogs with IO error flags to prevent any
4051          * further log IO from being issued or completed.
4052          */
4053         log->l_flags |= XLOG_IO_ERROR;
4054         retval = xlog_state_ioerror(log);
4055         spin_unlock(&log->l_icloglock);
4056
4057         /*
4058          * We don't want anybody waiting for log reservations after this. That
4059          * means we have to wake up everybody queued up on reserveq as well as
4060          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
4061          * we don't enqueue anything once the SHUTDOWN flag is set, and this
4062          * action is protected by the grant locks.
4063          */
4064         xlog_grant_head_wake_all(&log->l_reserve_head);
4065         xlog_grant_head_wake_all(&log->l_write_head);
4066
4067         /*
4068          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4069          * as if the log writes were completed. The abort handling in the log
4070          * item committed callback functions will do this again under lock to
4071          * avoid races.
4072          */
4073         wake_up_all(&log->l_cilp->xc_commit_wait);
4074         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
4075
4076 #ifdef XFSERRORDEBUG
4077         {
4078                 xlog_in_core_t  *iclog;
4079
4080                 spin_lock(&log->l_icloglock);
4081                 iclog = log->l_iclog;
4082                 do {
4083                         ASSERT(iclog->ic_callback == 0);
4084                         iclog = iclog->ic_next;
4085                 } while (iclog != log->l_iclog);
4086                 spin_unlock(&log->l_icloglock);
4087         }
4088 #endif
4089         /* return non-zero if log IOERROR transition had already happened */
4090         return retval;
4091 }
4092
4093 STATIC int
4094 xlog_iclogs_empty(
4095         struct xlog     *log)
4096 {
4097         xlog_in_core_t  *iclog;
4098
4099         iclog = log->l_iclog;
4100         do {
4101                 /* endianness does not matter here, zero is zero in
4102                  * any language.
4103                  */
4104                 if (iclog->ic_header.h_num_logops)
4105                         return 0;
4106                 iclog = iclog->ic_next;
4107         } while (iclog != log->l_iclog);
4108         return 1;
4109 }
4110
4111 /*
4112  * Verify that an LSN stamped into a piece of metadata is valid. This is
4113  * intended for use in read verifiers on v5 superblocks.
4114  */
4115 bool
4116 xfs_log_check_lsn(
4117         struct xfs_mount        *mp,
4118         xfs_lsn_t               lsn)
4119 {
4120         struct xlog             *log = mp->m_log;
4121         bool                    valid;
4122
4123         /*
4124          * norecovery mode skips mount-time log processing and unconditionally
4125          * resets the in-core LSN. We can't validate in this mode, but
4126          * modifications are not allowed anyways so just return true.
4127          */
4128         if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4129                 return true;
4130
4131         /*
4132          * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4133          * handled by recovery and thus safe to ignore here.
4134          */
4135         if (lsn == NULLCOMMITLSN)
4136                 return true;
4137
4138         valid = xlog_valid_lsn(mp->m_log, lsn);
4139
4140         /* warn the user about what's gone wrong before verifier failure */
4141         if (!valid) {
4142                 spin_lock(&log->l_icloglock);
4143                 xfs_warn(mp,
4144 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4145 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4146                          CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4147                          log->l_curr_cycle, log->l_curr_block);
4148                 spin_unlock(&log->l_icloglock);
4149         }
4150
4151         return valid;
4152 }