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_sb.h"
  25 #include "xfs_ag.h"
  26 #include "xfs_mount.h"
  27 #include "xfs_error.h"
  28 #include "xfs_trans.h"
  29 #include "xfs_trans_priv.h"
  30 #include "xfs_log.h"
  31 #include "xfs_log_priv.h"
  32 #include "xfs_log_recover.h"
  33 #include "xfs_inode.h"
  34 #include "xfs_trace.h"
  35 #include "xfs_fsops.h"
  36 #include "xfs_cksum.h"
  37 #include "xfs_sysfs.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         char                    *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(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(xs_try_logspace);
 384
 385         /*
 386          * This is a new transaction on the ticket, so we need to change the
 387          * transaction ID so that the next transaction has a different TID in
 388          * the log. Just add one to the existing tid so that we can see chains
 389          * of rolling transactions in the log easily.
 390          */
 391         tic->t_tid++;
 392
 393         xlog_grant_push_ail(log, tic->t_unit_res);
 394
 395         tic->t_curr_res = tic->t_unit_res;
 396         xlog_tic_reset_res(tic);
 397
 398         if (tic->t_cnt > 0)
 399                 return 0;
 400
 401         trace_xfs_log_regrant(log, tic);
 402
 403         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 404                                       &need_bytes);
 405         if (error)
 406                 goto out_error;
 407
 408         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 409         trace_xfs_log_regrant_exit(log, tic);
 410         xlog_verify_grant_tail(log);
 411         return 0;
 412
 413 out_error:
 414         /*
 415          * If we are failing, make sure the ticket doesn't have any current
 416          * reservations.  We don't want to add this back when the ticket/
 417          * transaction gets cancelled.
 418          */
 419         tic->t_curr_res = 0;
 420         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 421         return error;
 422 }
 423
 424 /*
 425  * Reserve log space and return a ticket corresponding the reservation.
 426  *
 427  * Each reservation is going to reserve extra space for a log record header.
 428  * When writes happen to the on-disk log, we don't subtract the length of the
 429  * log record header from any reservation.  By wasting space in each
 430  * reservation, we prevent over allocation problems.
 431  */
 432 int
 433 xfs_log_reserve(
 434         struct xfs_mount        *mp,
 435         int                     unit_bytes,
 436         int                     cnt,
 437         struct xlog_ticket      **ticp,
 438         __uint8_t               client,
 439         bool                    permanent,
 440         uint                    t_type)
 441 {
 442         struct xlog             *log = mp->m_log;
 443         struct xlog_ticket      *tic;
 444         int                     need_bytes;
 445         int                     error = 0;
 446
 447         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 448
 449         if (XLOG_FORCED_SHUTDOWN(log))
 450                 return -EIO;
 451
 452         XFS_STATS_INC(xs_try_logspace);
 453
 454         ASSERT(*ticp == NULL);
 455         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 456                                 KM_SLEEP | KM_MAYFAIL);
 457         if (!tic)
 458                 return -ENOMEM;
 459
 460         tic->t_trans_type = t_type;
 461         *ticp = tic;
 462
 463         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 464                                             : tic->t_unit_res);
 465
 466         trace_xfs_log_reserve(log, tic);
 467
 468         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 469                                       &need_bytes);
 470         if (error)
 471                 goto out_error;
 472
 473         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 474         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 475         trace_xfs_log_reserve_exit(log, tic);
 476         xlog_verify_grant_tail(log);
 477         return 0;
 478
 479 out_error:
 480         /*
 481          * If we are failing, make sure the ticket doesn't have any current
 482          * reservations.  We don't want to add this back when the ticket/
 483          * transaction gets cancelled.
 484          */
 485         tic->t_curr_res = 0;
 486         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 487         return error;
 488 }
 489
 490
 491 /*
 492  * NOTES:
 493  *
 494  *      1. currblock field gets updated at startup and after in-core logs
 495  *              marked as with WANT_SYNC.
 496  */
 497
 498 /*
 499  * This routine is called when a user of a log manager ticket is done with
 500  * the reservation.  If the ticket was ever used, then a commit record for
 501  * the associated transaction is written out as a log operation header with
 502  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 503  * a given ticket.  If the ticket was one with a permanent reservation, then
 504  * a few operations are done differently.  Permanent reservation tickets by
 505  * default don't release the reservation.  They just commit the current
 506  * transaction with the belief that the reservation is still needed.  A flag
 507  * must be passed in before permanent reservations are actually released.
 508  * When these type of tickets are not released, they need to be set into
 509  * the inited state again.  By doing this, a start record will be written
 510  * out when the next write occurs.
 511  */
 512 xfs_lsn_t
 513 xfs_log_done(
 514         struct xfs_mount        *mp,
 515         struct xlog_ticket      *ticket,
 516         struct xlog_in_core     **iclog,
 517         uint                    flags)
 518 {
 519         struct xlog             *log = mp->m_log;
 520         xfs_lsn_t               lsn = 0;
 521
 522         if (XLOG_FORCED_SHUTDOWN(log) ||
 523             /*
 524              * If nothing was ever written, don't write out commit record.
 525              * If we get an error, just continue and give back the log ticket.
 526              */
 527             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 528              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 529                 lsn = (xfs_lsn_t) -1;
 530                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
 531                         flags |= XFS_LOG_REL_PERM_RESERV;
 532                 }
 533         }
 534
 535
 536         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
 537             (flags & XFS_LOG_REL_PERM_RESERV)) {
 538                 trace_xfs_log_done_nonperm(log, ticket);
 539
 540                 /*
 541                  * Release ticket if not permanent reservation or a specific
 542                  * request has been made to release a permanent reservation.
 543                  */
 544                 xlog_ungrant_log_space(log, ticket);
 545                 xfs_log_ticket_put(ticket);
 546         } else {
 547                 trace_xfs_log_done_perm(log, ticket);
 548
 549                 xlog_regrant_reserve_log_space(log, ticket);
 550                 /* If this ticket was a permanent reservation and we aren't
 551                  * trying to release it, reset the inited flags; so next time
 552                  * we write, a start record will be written out.
 553                  */
 554                 ticket->t_flags |= XLOG_TIC_INITED;
 555         }
 556
 557         return lsn;
 558 }
 559
 560 /*
 561  * Attaches a new iclog I/O completion callback routine during
 562  * transaction commit.  If the log is in error state, a non-zero
 563  * return code is handed back and the caller is responsible for
 564  * executing the callback at an appropriate time.
 565  */
 566 int
 567 xfs_log_notify(
 568         struct xfs_mount        *mp,
 569         struct xlog_in_core     *iclog,
 570         xfs_log_callback_t      *cb)
 571 {
 572         int     abortflg;
 573
 574         spin_lock(&iclog->ic_callback_lock);
 575         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 576         if (!abortflg) {
 577                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 578                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 579                 cb->cb_next = NULL;
 580                 *(iclog->ic_callback_tail) = cb;
 581                 iclog->ic_callback_tail = &(cb->cb_next);
 582         }
 583         spin_unlock(&iclog->ic_callback_lock);
 584         return abortflg;
 585 }
 586
 587 int
 588 xfs_log_release_iclog(
 589         struct xfs_mount        *mp,
 590         struct xlog_in_core     *iclog)
 591 {
 592         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 593                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 594                 return -EIO;
 595         }
 596
 597         return 0;
 598 }
 599
 600 /*
 601  * Mount a log filesystem
 602  *
 603  * mp           - ubiquitous xfs mount point structure
 604  * log_target   - buftarg of on-disk log device
 605  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 606  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 607  *
 608  * Return error or zero.
 609  */
 610 int
 611 xfs_log_mount(
 612         xfs_mount_t     *mp,
 613         xfs_buftarg_t   *log_target,
 614         xfs_daddr_t     blk_offset,
 615         int             num_bblks)
 616 {
 617         int             error = 0;
 618         int             min_logfsbs;
 619
 620         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 621                 xfs_notice(mp, "Mounting V%d Filesystem",
 622                            XFS_SB_VERSION_NUM(&mp->m_sb));
 623         } else {
 624                 xfs_notice(mp,
 625 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
 626                            XFS_SB_VERSION_NUM(&mp->m_sb));
 627                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 628         }
 629
 630         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 631         if (IS_ERR(mp->m_log)) {
 632                 error = PTR_ERR(mp->m_log);
 633                 goto out;
 634         }
 635
 636         /*
 637          * Validate the given log space and drop a critical message via syslog
 638          * if the log size is too small that would lead to some unexpected
 639          * situations in transaction log space reservation stage.
 640          *
 641          * Note: we can't just reject the mount if the validation fails.  This
 642          * would mean that people would have to downgrade their kernel just to
 643          * remedy the situation as there is no way to grow the log (short of
 644          * black magic surgery with xfs_db).
 645          *
 646          * We can, however, reject mounts for CRC format filesystems, as the
 647          * mkfs binary being used to make the filesystem should never create a
 648          * filesystem with a log that is too small.
 649          */
 650         min_logfsbs = xfs_log_calc_minimum_size(mp);
 651
 652         if (mp->m_sb.sb_logblocks < min_logfsbs) {
 653                 xfs_warn(mp,
 654                 "Log size %d blocks too small, minimum size is %d blocks",
 655                          mp->m_sb.sb_logblocks, min_logfsbs);
 656                 error = -EINVAL;
 657         } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
 658                 xfs_warn(mp,
 659                 "Log size %d blocks too large, maximum size is %lld blocks",
 660                          mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
 661                 error = -EINVAL;
 662         } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
 663                 xfs_warn(mp,
 664                 "log size %lld bytes too large, maximum size is %lld bytes",
 665                          XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
 666                          XFS_MAX_LOG_BYTES);
 667                 error = -EINVAL;
 668         }
 669         if (error) {
 670                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
 671                         xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 672                         ASSERT(0);
 673                         goto out_free_log;
 674                 }
 675                 xfs_crit(mp,
 676 "Log size out of supported range. Continuing onwards, but if log hangs are\n"
 677 "experienced then please report this message in the bug report.");
 678         }
 679
 680         /*
 681          * Initialize the AIL now we have a log.
 682          */
 683         error = xfs_trans_ail_init(mp);
 684         if (error) {
 685                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 686                 goto out_free_log;
 687         }
 688         mp->m_log->l_ailp = mp->m_ail;
 689
 690         /*
 691          * skip log recovery on a norecovery mount.  pretend it all
 692          * just worked.
 693          */
 694         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 695                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 696
 697                 if (readonly)
 698                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 699
 700                 error = xlog_recover(mp->m_log);
 701
 702                 if (readonly)
 703                         mp->m_flags |= XFS_MOUNT_RDONLY;
 704                 if (error) {
 705                         xfs_warn(mp, "log mount/recovery failed: error %d",
 706                                 error);
 707                         goto out_destroy_ail;
 708                 }
 709         }
 710
 711         error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 712                                "log");
 713         if (error)
 714                 goto out_destroy_ail;
 715
 716         /* Normal transactions can now occur */
 717         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 718
 719         /*
 720          * Now the log has been fully initialised and we know were our
 721          * space grant counters are, we can initialise the permanent ticket
 722          * needed for delayed logging to work.
 723          */
 724         xlog_cil_init_post_recovery(mp->m_log);
 725
 726         return 0;
 727
 728 out_destroy_ail:
 729         xfs_trans_ail_destroy(mp);
 730 out_free_log:
 731         xlog_dealloc_log(mp->m_log);
 732 out:
 733         return error;
 734 }
 735
 736 /*
 737  * Finish the recovery of the file system.  This is separate from the
 738  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 739  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 740  * here.
 741  *
 742  * If we finish recovery successfully, start the background log work. If we are
 743  * not doing recovery, then we have a RO filesystem and we don't need to start
 744  * it.
 745  */
 746 int
 747 xfs_log_mount_finish(xfs_mount_t *mp)
 748 {
 749         int     error = 0;
 750
 751         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 752                 error = xlog_recover_finish(mp->m_log);
 753                 if (!error)
 754                         xfs_log_work_queue(mp);
 755         } else {
 756                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 757         }
 758
 759
 760         return error;
 761 }
 762
 763 /*
 764  * Final log writes as part of unmount.
 765  *
 766  * Mark the filesystem clean as unmount happens.  Note that during relocation
 767  * this routine needs to be executed as part of source-bag while the
 768  * deallocation must not be done until source-end.
 769  */
 770
 771 /*
 772  * Unmount record used to have a string "Unmount filesystem--" in the
 773  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 774  * We just write the magic number now since that particular field isn't
 775  * currently architecture converted and "Unmount" is a bit foo.
 776  * As far as I know, there weren't any dependencies on the old behaviour.
 777  */
 778
 779 int
 780 xfs_log_unmount_write(xfs_mount_t *mp)
 781 {
 782         struct xlog      *log = mp->m_log;
 783         xlog_in_core_t   *iclog;
 784 #ifdef DEBUG
 785         xlog_in_core_t   *first_iclog;
 786 #endif
 787         xlog_ticket_t   *tic = NULL;
 788         xfs_lsn_t        lsn;
 789         int              error;
 790
 791         /*
 792          * Don't write out unmount record on read-only mounts.
 793          * Or, if we are doing a forced umount (typically because of IO errors).
 794          */
 795         if (mp->m_flags & XFS_MOUNT_RDONLY)
 796                 return 0;
 797
 798         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
 799         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 800
 801 #ifdef DEBUG
 802         first_iclog = iclog = log->l_iclog;
 803         do {
 804                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 805                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 806                         ASSERT(iclog->ic_offset == 0);
 807                 }
 808                 iclog = iclog->ic_next;
 809         } while (iclog != first_iclog);
 810 #endif
 811         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 812                 error = xfs_log_reserve(mp, 600, 1, &tic,
 813                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
 814                 if (!error) {
 815                         /* the data section must be 32 bit size aligned */
 816                         struct {
 817                             __uint16_t magic;
 818                             __uint16_t pad1;
 819                             __uint32_t pad2; /* may as well make it 64 bits */
 820                         } magic = {
 821                                 .magic = XLOG_UNMOUNT_TYPE,
 822                         };
 823                         struct xfs_log_iovec reg = {
 824                                 .i_addr = &magic,
 825                                 .i_len = sizeof(magic),
 826                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 827                         };
 828                         struct xfs_log_vec vec = {
 829                                 .lv_niovecs = 1,
 830                                 .lv_iovecp = &reg,
 831                         };
 832
 833                         /* remove inited flag, and account for space used */
 834                         tic->t_flags = 0;
 835                         tic->t_curr_res -= sizeof(magic);
 836                         error = xlog_write(log, &vec, tic, &lsn,
 837                                            NULL, XLOG_UNMOUNT_TRANS);
 838                         /*
 839                          * At this point, we're umounting anyway,
 840                          * so there's no point in transitioning log state
 841                          * to IOERROR. Just continue...
 842                          */
 843                 }
 844
 845                 if (error)
 846                         xfs_alert(mp, "%s: unmount record failed", __func__);
 847
 848
 849                 spin_lock(&log->l_icloglock);
 850                 iclog = log->l_iclog;
 851                 atomic_inc(&iclog->ic_refcnt);
 852                 xlog_state_want_sync(log, iclog);
 853                 spin_unlock(&log->l_icloglock);
 854                 error = xlog_state_release_iclog(log, iclog);
 855
 856                 spin_lock(&log->l_icloglock);
 857                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 858                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 859                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 860                                 xlog_wait(&iclog->ic_force_wait,
 861                                                         &log->l_icloglock);
 862                         } else {
 863                                 spin_unlock(&log->l_icloglock);
 864                         }
 865                 } else {
 866                         spin_unlock(&log->l_icloglock);
 867                 }
 868                 if (tic) {
 869                         trace_xfs_log_umount_write(log, tic);
 870                         xlog_ungrant_log_space(log, tic);
 871                         xfs_log_ticket_put(tic);
 872                 }
 873         } else {
 874                 /*
 875                  * We're already in forced_shutdown mode, couldn't
 876                  * even attempt to write out the unmount transaction.
 877                  *
 878                  * Go through the motions of sync'ing and releasing
 879                  * the iclog, even though no I/O will actually happen,
 880                  * we need to wait for other log I/Os that may already
 881                  * be in progress.  Do this as a separate section of
 882                  * code so we'll know if we ever get stuck here that
 883                  * we're in this odd situation of trying to unmount
 884                  * a file system that went into forced_shutdown as
 885                  * the result of an unmount..
 886                  */
 887                 spin_lock(&log->l_icloglock);
 888                 iclog = log->l_iclog;
 889                 atomic_inc(&iclog->ic_refcnt);
 890
 891                 xlog_state_want_sync(log, iclog);
 892                 spin_unlock(&log->l_icloglock);
 893                 error =  xlog_state_release_iclog(log, iclog);
 894
 895                 spin_lock(&log->l_icloglock);
 896
 897                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 898                         || iclog->ic_state == XLOG_STATE_DIRTY
 899                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 900
 901                                 xlog_wait(&iclog->ic_force_wait,
 902                                                         &log->l_icloglock);
 903                 } else {
 904                         spin_unlock(&log->l_icloglock);
 905                 }
 906         }
 907
 908         return error;
 909 }       /* xfs_log_unmount_write */
 910
 911 /*
 912  * Empty the log for unmount/freeze.
 913  *
 914  * To do this, we first need to shut down the background log work so it is not
 915  * trying to cover the log as we clean up. We then need to unpin all objects in
 916  * the log so we can then flush them out. Once they have completed their IO and
 917  * run the callbacks removing themselves from the AIL, we can write the unmount
 918  * record.
 919  */
 920 void
 921 xfs_log_quiesce(
 922         struct xfs_mount        *mp)
 923 {
 924         cancel_delayed_work_sync(&mp->m_log->l_work);
 925         xfs_log_force(mp, XFS_LOG_SYNC);
 926
 927         /*
 928          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
 929          * will push it, xfs_wait_buftarg() will not wait for it. Further,
 930          * xfs_buf_iowait() cannot be used because it was pushed with the
 931          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
 932          * the IO to complete.
 933          */
 934         xfs_ail_push_all_sync(mp->m_ail);
 935         xfs_wait_buftarg(mp->m_ddev_targp);
 936         xfs_buf_lock(mp->m_sb_bp);
 937         xfs_buf_unlock(mp->m_sb_bp);
 938
 939         xfs_log_unmount_write(mp);
 940 }
 941
 942 /*
 943  * Shut down and release the AIL and Log.
 944  *
 945  * During unmount, we need to ensure we flush all the dirty metadata objects
 946  * from the AIL so that the log is empty before we write the unmount record to
 947  * the log. Once this is done, we can tear down the AIL and the log.
 948  */
 949 void
 950 xfs_log_unmount(
 951         struct xfs_mount        *mp)
 952 {
 953         xfs_log_quiesce(mp);
 954
 955         xfs_trans_ail_destroy(mp);
 956
 957         xfs_sysfs_del(&mp->m_log->l_kobj);
 958
 959         xlog_dealloc_log(mp->m_log);
 960 }
 961
 962 void
 963 xfs_log_item_init(
 964         struct xfs_mount        *mp,
 965         struct xfs_log_item     *item,
 966         int                     type,
 967         const struct xfs_item_ops *ops)
 968 {
 969         item->li_mountp = mp;
 970         item->li_ailp = mp->m_ail;
 971         item->li_type = type;
 972         item->li_ops = ops;
 973         item->li_lv = NULL;
 974
 975         INIT_LIST_HEAD(&item->li_ail);
 976         INIT_LIST_HEAD(&item->li_cil);
 977 }
 978
 979 /*
 980  * Wake up processes waiting for log space after we have moved the log tail.
 981  */
 982 void
 983 xfs_log_space_wake(
 984         struct xfs_mount        *mp)
 985 {
 986         struct xlog             *log = mp->m_log;
 987         int                     free_bytes;
 988
 989         if (XLOG_FORCED_SHUTDOWN(log))
 990                 return;
 991
 992         if (!list_empty_careful(&log->l_write_head.waiters)) {
 993                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 994
 995                 spin_lock(&log->l_write_head.lock);
 996                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
 997                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
 998                 spin_unlock(&log->l_write_head.lock);
 999         }
1000
1001         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1002                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1003
1004                 spin_lock(&log->l_reserve_head.lock);
1005                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1006                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1007                 spin_unlock(&log->l_reserve_head.lock);
1008         }
1009 }
1010
1011 /*
1012  * Determine if we have a transaction that has gone to disk that needs to be
1013  * covered. To begin the transition to the idle state firstly the log needs to
1014  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1015  * we start attempting to cover the log.
1016  *
1017  * Only if we are then in a state where covering is needed, the caller is
1018  * informed that dummy transactions are required to move the log into the idle
1019  * state.
1020  *
1021  * If there are any items in the AIl or CIL, then we do not want to attempt to
1022  * cover the log as we may be in a situation where there isn't log space
1023  * available to run a dummy transaction and this can lead to deadlocks when the
1024  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1025  * there's no point in running a dummy transaction at this point because we
1026  * can't start trying to idle the log until both the CIL and AIL are empty.
1027  */
1028 int
1029 xfs_log_need_covered(xfs_mount_t *mp)
1030 {
1031         struct xlog     *log = mp->m_log;
1032         int             needed = 0;
1033
1034         if (!xfs_fs_writable(mp))
1035                 return 0;
1036
1037         if (!xlog_cil_empty(log))
1038                 return 0;
1039
1040         spin_lock(&log->l_icloglock);
1041         switch (log->l_covered_state) {
1042         case XLOG_STATE_COVER_DONE:
1043         case XLOG_STATE_COVER_DONE2:
1044         case XLOG_STATE_COVER_IDLE:
1045                 break;
1046         case XLOG_STATE_COVER_NEED:
1047         case XLOG_STATE_COVER_NEED2:
1048                 if (xfs_ail_min_lsn(log->l_ailp))
1049                         break;
1050                 if (!xlog_iclogs_empty(log))
1051                         break;
1052
1053                 needed = 1;
1054                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1055                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1056                 else
1057                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1058                 break;
1059         default:
1060                 needed = 1;
1061                 break;
1062         }
1063         spin_unlock(&log->l_icloglock);
1064         return needed;
1065 }
1066
1067 /*
1068  * We may be holding the log iclog lock upon entering this routine.
1069  */
1070 xfs_lsn_t
1071 xlog_assign_tail_lsn_locked(
1072         struct xfs_mount        *mp)
1073 {
1074         struct xlog             *log = mp->m_log;
1075         struct xfs_log_item     *lip;
1076         xfs_lsn_t               tail_lsn;
1077
1078         assert_spin_locked(&mp->m_ail->xa_lock);
1079
1080         /*
1081          * To make sure we always have a valid LSN for the log tail we keep
1082          * track of the last LSN which was committed in log->l_last_sync_lsn,
1083          * and use that when the AIL was empty.
1084          */
1085         lip = xfs_ail_min(mp->m_ail);
1086         if (lip)
1087                 tail_lsn = lip->li_lsn;
1088         else
1089                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1090         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1091         atomic64_set(&log->l_tail_lsn, tail_lsn);
1092         return tail_lsn;
1093 }
1094
1095 xfs_lsn_t
1096 xlog_assign_tail_lsn(
1097         struct xfs_mount        *mp)
1098 {
1099         xfs_lsn_t               tail_lsn;
1100
1101         spin_lock(&mp->m_ail->xa_lock);
1102         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1103         spin_unlock(&mp->m_ail->xa_lock);
1104
1105         return tail_lsn;
1106 }
1107
1108 /*
1109  * Return the space in the log between the tail and the head.  The head
1110  * is passed in the cycle/bytes formal parms.  In the special case where
1111  * the reserve head has wrapped passed the tail, this calculation is no
1112  * longer valid.  In this case, just return 0 which means there is no space
1113  * in the log.  This works for all places where this function is called
1114  * with the reserve head.  Of course, if the write head were to ever
1115  * wrap the tail, we should blow up.  Rather than catch this case here,
1116  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1117  *
1118  * This code also handles the case where the reservation head is behind
1119  * the tail.  The details of this case are described below, but the end
1120  * result is that we return the size of the log as the amount of space left.
1121  */
1122 STATIC int
1123 xlog_space_left(
1124         struct xlog     *log,
1125         atomic64_t      *head)
1126 {
1127         int             free_bytes;
1128         int             tail_bytes;
1129         int             tail_cycle;
1130         int             head_cycle;
1131         int             head_bytes;
1132
1133         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1134         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1135         tail_bytes = BBTOB(tail_bytes);
1136         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1137                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1138         else if (tail_cycle + 1 < head_cycle)
1139                 return 0;
1140         else if (tail_cycle < head_cycle) {
1141                 ASSERT(tail_cycle == (head_cycle - 1));
1142                 free_bytes = tail_bytes - head_bytes;
1143         } else {
1144                 /*
1145                  * The reservation head is behind the tail.
1146                  * In this case we just want to return the size of the
1147                  * log as the amount of space left.
1148                  */
1149                 xfs_alert(log->l_mp,
1150                         "xlog_space_left: head behind tail\n"
1151                         "  tail_cycle = %d, tail_bytes = %d\n"
1152                         "  GH   cycle = %d, GH   bytes = %d",
1153                         tail_cycle, tail_bytes, head_cycle, head_bytes);
1154                 ASSERT(0);
1155                 free_bytes = log->l_logsize;
1156         }
1157         return free_bytes;
1158 }
1159
1160
1161 /*
1162  * Log function which is called when an io completes.
1163  *
1164  * The log manager needs its own routine, in order to control what
1165  * happens with the buffer after the write completes.
1166  */
1167 void
1168 xlog_iodone(xfs_buf_t *bp)
1169 {
1170         struct xlog_in_core     *iclog = bp->b_fspriv;
1171         struct xlog             *l = iclog->ic_log;
1172         int                     aborted = 0;
1173
1174         /*
1175          * Race to shutdown the filesystem if we see an error.
1176          */
1177         if (XFS_TEST_ERROR(bp->b_error, l->l_mp,
1178                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1179                 xfs_buf_ioerror_alert(bp, __func__);
1180                 xfs_buf_stale(bp);
1181                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1182                 /*
1183                  * This flag will be propagated to the trans-committed
1184                  * callback routines to let them know that the log-commit
1185                  * didn't succeed.
1186                  */
1187                 aborted = XFS_LI_ABORTED;
1188         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1189                 aborted = XFS_LI_ABORTED;
1190         }
1191
1192         /* log I/O is always issued ASYNC */
1193         ASSERT(XFS_BUF_ISASYNC(bp));
1194         xlog_state_done_syncing(iclog, aborted);
1195
1196         /*
1197          * drop the buffer lock now that we are done. Nothing references
1198          * the buffer after this, so an unmount waiting on this lock can now
1199          * tear it down safely. As such, it is unsafe to reference the buffer
1200          * (bp) after the unlock as we could race with it being freed.
1201          */
1202         xfs_buf_unlock(bp);
1203 }
1204
1205 /*
1206  * Return size of each in-core log record buffer.
1207  *
1208  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1209  *
1210  * If the filesystem blocksize is too large, we may need to choose a
1211  * larger size since the directory code currently logs entire blocks.
1212  */
1213
1214 STATIC void
1215 xlog_get_iclog_buffer_size(
1216         struct xfs_mount        *mp,
1217         struct xlog             *log)
1218 {
1219         int size;
1220         int xhdrs;
1221
1222         if (mp->m_logbufs <= 0)
1223                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1224         else
1225                 log->l_iclog_bufs = mp->m_logbufs;
1226
1227         /*
1228          * Buffer size passed in from mount system call.
1229          */
1230         if (mp->m_logbsize > 0) {
1231                 size = log->l_iclog_size = mp->m_logbsize;
1232                 log->l_iclog_size_log = 0;
1233                 while (size != 1) {
1234                         log->l_iclog_size_log++;
1235                         size >>= 1;
1236                 }
1237
1238                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1239                         /* # headers = size / 32k
1240                          * one header holds cycles from 32k of data
1241                          */
1242
1243                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1244                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1245                                 xhdrs++;
1246                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1247                         log->l_iclog_heads = xhdrs;
1248                 } else {
1249                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1250                         log->l_iclog_hsize = BBSIZE;
1251                         log->l_iclog_heads = 1;
1252                 }
1253                 goto done;
1254         }
1255
1256         /* All machines use 32kB buffers by default. */
1257         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1258         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1259
1260         /* the default log size is 16k or 32k which is one header sector */
1261         log->l_iclog_hsize = BBSIZE;
1262         log->l_iclog_heads = 1;
1263
1264 done:
1265         /* are we being asked to make the sizes selected above visible? */
1266         if (mp->m_logbufs == 0)
1267                 mp->m_logbufs = log->l_iclog_bufs;
1268         if (mp->m_logbsize == 0)
1269                 mp->m_logbsize = log->l_iclog_size;
1270 }       /* xlog_get_iclog_buffer_size */
1271
1272
1273 void
1274 xfs_log_work_queue(
1275         struct xfs_mount        *mp)
1276 {
1277         queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1278                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1279 }
1280
1281 /*
1282  * Every sync period we need to unpin all items in the AIL and push them to
1283  * disk. If there is nothing dirty, then we might need to cover the log to
1284  * indicate that the filesystem is idle.
1285  */
1286 void
1287 xfs_log_worker(
1288         struct work_struct      *work)
1289 {
1290         struct xlog             *log = container_of(to_delayed_work(work),
1291                                                 struct xlog, l_work);
1292         struct xfs_mount        *mp = log->l_mp;
1293
1294         /* dgc: errors ignored - not fatal and nowhere to report them */
1295         if (xfs_log_need_covered(mp))
1296                 xfs_fs_log_dummy(mp);
1297         else
1298                 xfs_log_force(mp, 0);
1299
1300         /* start pushing all the metadata that is currently dirty */
1301         xfs_ail_push_all(mp->m_ail);
1302
1303         /* queue us up again */
1304         xfs_log_work_queue(mp);
1305 }
1306
1307 /*
1308  * This routine initializes some of the log structure for a given mount point.
1309  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1310  * some other stuff may be filled in too.
1311  */
1312 STATIC struct xlog *
1313 xlog_alloc_log(
1314         struct xfs_mount        *mp,
1315         struct xfs_buftarg      *log_target,
1316         xfs_daddr_t             blk_offset,
1317         int                     num_bblks)
1318 {
1319         struct xlog             *log;
1320         xlog_rec_header_t       *head;
1321         xlog_in_core_t          **iclogp;
1322         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1323         xfs_buf_t               *bp;
1324         int                     i;
1325         int                     error = -ENOMEM;
1326         uint                    log2_size = 0;
1327
1328         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1329         if (!log) {
1330                 xfs_warn(mp, "Log allocation failed: No memory!");
1331                 goto out;
1332         }
1333
1334         log->l_mp          = mp;
1335         log->l_targ        = log_target;
1336         log->l_logsize     = BBTOB(num_bblks);
1337         log->l_logBBstart  = blk_offset;
1338         log->l_logBBsize   = num_bblks;
1339         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1340         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1341         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1342
1343         log->l_prev_block  = -1;
1344         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1345         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1346         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1347         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1348
1349         xlog_grant_head_init(&log->l_reserve_head);
1350         xlog_grant_head_init(&log->l_write_head);
1351
1352         error = -EFSCORRUPTED;
1353         if (xfs_sb_version_hassector(&mp->m_sb)) {
1354                 log2_size = mp->m_sb.sb_logsectlog;
1355                 if (log2_size < BBSHIFT) {
1356                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1357                                 log2_size, BBSHIFT);
1358                         goto out_free_log;
1359                 }
1360
1361                 log2_size -= BBSHIFT;
1362                 if (log2_size > mp->m_sectbb_log) {
1363                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1364                                 log2_size, mp->m_sectbb_log);
1365                         goto out_free_log;
1366                 }
1367
1368                 /* for larger sector sizes, must have v2 or external log */
1369                 if (log2_size && log->l_logBBstart > 0 &&
1370                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1371                         xfs_warn(mp,
1372                 "log sector size (0x%x) invalid for configuration.",
1373                                 log2_size);
1374                         goto out_free_log;
1375                 }
1376         }
1377         log->l_sectBBsize = 1 << log2_size;
1378
1379         xlog_get_iclog_buffer_size(mp, log);
1380
1381         error = -ENOMEM;
1382         bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1383         if (!bp)
1384                 goto out_free_log;
1385
1386         /*
1387          * The iclogbuf buffer locks are held over IO but we are not going to do
1388          * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1389          * when appropriately.
1390          */
1391         ASSERT(xfs_buf_islocked(bp));
1392         xfs_buf_unlock(bp);
1393
1394         bp->b_iodone = xlog_iodone;
1395         log->l_xbuf = bp;
1396
1397         spin_lock_init(&log->l_icloglock);
1398         init_waitqueue_head(&log->l_flush_wait);
1399
1400         iclogp = &log->l_iclog;
1401         /*
1402          * The amount of memory to allocate for the iclog structure is
1403          * rather funky due to the way the structure is defined.  It is
1404          * done this way so that we can use different sizes for machines
1405          * with different amounts of memory.  See the definition of
1406          * xlog_in_core_t in xfs_log_priv.h for details.
1407          */
1408         ASSERT(log->l_iclog_size >= 4096);
1409         for (i=0; i < log->l_iclog_bufs; i++) {
1410                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1411                 if (!*iclogp)
1412                         goto out_free_iclog;
1413
1414                 iclog = *iclogp;
1415                 iclog->ic_prev = prev_iclog;
1416                 prev_iclog = iclog;
1417
1418                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1419                                                 BTOBB(log->l_iclog_size), 0);
1420                 if (!bp)
1421                         goto out_free_iclog;
1422
1423                 ASSERT(xfs_buf_islocked(bp));
1424                 xfs_buf_unlock(bp);
1425
1426                 bp->b_iodone = xlog_iodone;
1427                 iclog->ic_bp = bp;
1428                 iclog->ic_data = bp->b_addr;
1429 #ifdef DEBUG
1430                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1431 #endif
1432                 head = &iclog->ic_header;
1433                 memset(head, 0, sizeof(xlog_rec_header_t));
1434                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1435                 head->h_version = cpu_to_be32(
1436                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1437                 head->h_size = cpu_to_be32(log->l_iclog_size);
1438                 /* new fields */
1439                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1440                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1441
1442                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1443                 iclog->ic_state = XLOG_STATE_ACTIVE;
1444                 iclog->ic_log = log;
1445                 atomic_set(&iclog->ic_refcnt, 0);
1446                 spin_lock_init(&iclog->ic_callback_lock);
1447                 iclog->ic_callback_tail = &(iclog->ic_callback);
1448                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1449
1450                 init_waitqueue_head(&iclog->ic_force_wait);
1451                 init_waitqueue_head(&iclog->ic_write_wait);
1452
1453                 iclogp = &iclog->ic_next;
1454         }
1455         *iclogp = log->l_iclog;                 /* complete ring */
1456         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1457
1458         error = xlog_cil_init(log);
1459         if (error)
1460                 goto out_free_iclog;
1461         return log;
1462
1463 out_free_iclog:
1464         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1465                 prev_iclog = iclog->ic_next;
1466                 if (iclog->ic_bp)
1467                         xfs_buf_free(iclog->ic_bp);
1468                 kmem_free(iclog);
1469         }
1470         spinlock_destroy(&log->l_icloglock);
1471         xfs_buf_free(log->l_xbuf);
1472 out_free_log:
1473         kmem_free(log);
1474 out:
1475         return ERR_PTR(error);
1476 }       /* xlog_alloc_log */
1477
1478
1479 /*
1480  * Write out the commit record of a transaction associated with the given
1481  * ticket.  Return the lsn of the commit record.
1482  */
1483 STATIC int
1484 xlog_commit_record(
1485         struct xlog             *log,
1486         struct xlog_ticket      *ticket,
1487         struct xlog_in_core     **iclog,
1488         xfs_lsn_t               *commitlsnp)
1489 {
1490         struct xfs_mount *mp = log->l_mp;
1491         int     error;
1492         struct xfs_log_iovec reg = {
1493                 .i_addr = NULL,
1494                 .i_len = 0,
1495                 .i_type = XLOG_REG_TYPE_COMMIT,
1496         };
1497         struct xfs_log_vec vec = {
1498                 .lv_niovecs = 1,
1499                 .lv_iovecp = &reg,
1500         };
1501
1502         ASSERT_ALWAYS(iclog);
1503         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1504                                         XLOG_COMMIT_TRANS);
1505         if (error)
1506                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1507         return error;
1508 }
1509
1510 /*
1511  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1512  * log space.  This code pushes on the lsn which would supposedly free up
1513  * the 25% which we want to leave free.  We may need to adopt a policy which
1514  * pushes on an lsn which is further along in the log once we reach the high
1515  * water mark.  In this manner, we would be creating a low water mark.
1516  */
1517 STATIC void
1518 xlog_grant_push_ail(
1519         struct xlog     *log,
1520         int             need_bytes)
1521 {
1522         xfs_lsn_t       threshold_lsn = 0;
1523         xfs_lsn_t       last_sync_lsn;
1524         int             free_blocks;
1525         int             free_bytes;
1526         int             threshold_block;
1527         int             threshold_cycle;
1528         int             free_threshold;
1529
1530         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1531
1532         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1533         free_blocks = BTOBBT(free_bytes);
1534
1535         /*
1536          * Set the threshold for the minimum number of free blocks in the
1537          * log to the maximum of what the caller needs, one quarter of the
1538          * log, and 256 blocks.
1539          */
1540         free_threshold = BTOBB(need_bytes);
1541         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1542         free_threshold = MAX(free_threshold, 256);
1543         if (free_blocks >= free_threshold)
1544                 return;
1545
1546         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1547                                                 &threshold_block);
1548         threshold_block += free_threshold;
1549         if (threshold_block >= log->l_logBBsize) {
1550                 threshold_block -= log->l_logBBsize;
1551                 threshold_cycle += 1;
1552         }
1553         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1554                                         threshold_block);
1555         /*
1556          * Don't pass in an lsn greater than the lsn of the last
1557          * log record known to be on disk. Use a snapshot of the last sync lsn
1558          * so that it doesn't change between the compare and the set.
1559          */
1560         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1561         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1562                 threshold_lsn = last_sync_lsn;
1563
1564         /*
1565          * Get the transaction layer to kick the dirty buffers out to
1566          * disk asynchronously. No point in trying to do this if
1567          * the filesystem is shutting down.
1568          */
1569         if (!XLOG_FORCED_SHUTDOWN(log))
1570                 xfs_ail_push(log->l_ailp, threshold_lsn);
1571 }
1572
1573 /*
1574  * Stamp cycle number in every block
1575  */
1576 STATIC void
1577 xlog_pack_data(
1578         struct xlog             *log,
1579         struct xlog_in_core     *iclog,
1580         int                     roundoff)
1581 {
1582         int                     i, j, k;
1583         int                     size = iclog->ic_offset + roundoff;
1584         __be32                  cycle_lsn;
1585         xfs_caddr_t             dp;
1586
1587         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1588
1589         dp = iclog->ic_datap;
1590         for (i = 0; i < BTOBB(size); i++) {
1591                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1592                         break;
1593                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1594                 *(__be32 *)dp = cycle_lsn;
1595                 dp += BBSIZE;
1596         }
1597
1598         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1599                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1600
1601                 for ( ; i < BTOBB(size); i++) {
1602                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1603                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1604                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1605                         *(__be32 *)dp = cycle_lsn;
1606                         dp += BBSIZE;
1607                 }
1608
1609                 for (i = 1; i < log->l_iclog_heads; i++)
1610                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1611         }
1612 }
1613
1614 /*
1615  * Calculate the checksum for a log buffer.
1616  *
1617  * This is a little more complicated than it should be because the various
1618  * headers and the actual data are non-contiguous.
1619  */
1620 __le32
1621 xlog_cksum(
1622         struct xlog             *log,
1623         struct xlog_rec_header  *rhead,
1624         char                    *dp,
1625         int                     size)
1626 {
1627         __uint32_t              crc;
1628
1629         /* first generate the crc for the record header ... */
1630         crc = xfs_start_cksum((char *)rhead,
1631                               sizeof(struct xlog_rec_header),
1632                               offsetof(struct xlog_rec_header, h_crc));
1633
1634         /* ... then for additional cycle data for v2 logs ... */
1635         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1636                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1637                 int             i;
1638
1639                 for (i = 1; i < log->l_iclog_heads; i++) {
1640                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1641                                      sizeof(struct xlog_rec_ext_header));
1642                 }
1643         }
1644
1645         /* ... and finally for the payload */
1646         crc = crc32c(crc, dp, size);
1647
1648         return xfs_end_cksum(crc);
1649 }
1650
1651 /*
1652  * The bdstrat callback function for log bufs. This gives us a central
1653  * place to trap bufs in case we get hit by a log I/O error and need to
1654  * shutdown. Actually, in practice, even when we didn't get a log error,
1655  * we transition the iclogs to IOERROR state *after* flushing all existing
1656  * iclogs to disk. This is because we don't want anymore new transactions to be
1657  * started or completed afterwards.
1658  *
1659  * We lock the iclogbufs here so that we can serialise against IO completion
1660  * during unmount. We might be processing a shutdown triggered during unmount,
1661  * and that can occur asynchronously to the unmount thread, and hence we need to
1662  * ensure that completes before tearing down the iclogbufs. Hence we need to
1663  * hold the buffer lock across the log IO to acheive that.
1664  */
1665 STATIC int
1666 xlog_bdstrat(
1667         struct xfs_buf          *bp)
1668 {
1669         struct xlog_in_core     *iclog = bp->b_fspriv;
1670
1671         xfs_buf_lock(bp);
1672         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1673                 xfs_buf_ioerror(bp, -EIO);
1674                 xfs_buf_stale(bp);
1675                 xfs_buf_ioend(bp, 0);
1676                 /*
1677                  * It would seem logical to return EIO here, but we rely on
1678                  * the log state machine to propagate I/O errors instead of
1679                  * doing it here. Similarly, IO completion will unlock the
1680                  * buffer, so we don't do it here.
1681                  */
1682                 return 0;
1683         }
1684
1685         xfs_buf_iorequest(bp);
1686         return 0;
1687 }
1688
1689 /*
1690  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1691  * fashion.  Previously, we should have moved the current iclog
1692  * ptr in the log to point to the next available iclog.  This allows further
1693  * write to continue while this code syncs out an iclog ready to go.
1694  * Before an in-core log can be written out, the data section must be scanned
1695  * to save away the 1st word of each BBSIZE block into the header.  We replace
1696  * it with the current cycle count.  Each BBSIZE block is tagged with the
1697  * cycle count because there in an implicit assumption that drives will
1698  * guarantee that entire 512 byte blocks get written at once.  In other words,
1699  * we can't have part of a 512 byte block written and part not written.  By
1700  * tagging each block, we will know which blocks are valid when recovering
1701  * after an unclean shutdown.
1702  *
1703  * This routine is single threaded on the iclog.  No other thread can be in
1704  * this routine with the same iclog.  Changing contents of iclog can there-
1705  * fore be done without grabbing the state machine lock.  Updating the global
1706  * log will require grabbing the lock though.
1707  *
1708  * The entire log manager uses a logical block numbering scheme.  Only
1709  * log_sync (and then only bwrite()) know about the fact that the log may
1710  * not start with block zero on a given device.  The log block start offset
1711  * is added immediately before calling bwrite().
1712  */
1713
1714 STATIC int
1715 xlog_sync(
1716         struct xlog             *log,
1717         struct xlog_in_core     *iclog)
1718 {
1719         xfs_buf_t       *bp;
1720         int             i;
1721         uint            count;          /* byte count of bwrite */
1722         uint            count_init;     /* initial count before roundup */
1723         int             roundoff;       /* roundoff to BB or stripe */
1724         int             split = 0;      /* split write into two regions */
1725         int             error;
1726         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1727         int             size;
1728
1729         XFS_STATS_INC(xs_log_writes);
1730         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1731
1732         /* Add for LR header */
1733         count_init = log->l_iclog_hsize + iclog->ic_offset;
1734
1735         /* Round out the log write size */
1736         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1737                 /* we have a v2 stripe unit to use */
1738                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1739         } else {
1740                 count = BBTOB(BTOBB(count_init));
1741         }
1742         roundoff = count - count_init;
1743         ASSERT(roundoff >= 0);
1744         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1745                 roundoff < log->l_mp->m_sb.sb_logsunit)
1746                 ||
1747                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1748                  roundoff < BBTOB(1)));
1749
1750         /* move grant heads by roundoff in sync */
1751         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1752         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1753
1754         /* put cycle number in every block */
1755         xlog_pack_data(log, iclog, roundoff);
1756
1757         /* real byte length */
1758         size = iclog->ic_offset;
1759         if (v2)
1760                 size += roundoff;
1761         iclog->ic_header.h_len = cpu_to_be32(size);
1762
1763         bp = iclog->ic_bp;
1764         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1765
1766         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1767
1768         /* Do we need to split this write into 2 parts? */
1769         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1770                 char            *dptr;
1771
1772                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1773                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1774                 iclog->ic_bwritecnt = 2;
1775
1776                 /*
1777                  * Bump the cycle numbers at the start of each block in the
1778                  * part of the iclog that ends up in the buffer that gets
1779                  * written to the start of the log.
1780                  *
1781                  * Watch out for the header magic number case, though.
1782                  */
1783                 dptr = (char *)&iclog->ic_header + count;
1784                 for (i = 0; i < split; i += BBSIZE) {
1785                         __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1786                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1787                                 cycle++;
1788                         *(__be32 *)dptr = cpu_to_be32(cycle);
1789
1790                         dptr += BBSIZE;
1791                 }
1792         } else {
1793                 iclog->ic_bwritecnt = 1;
1794         }
1795
1796         /* calculcate the checksum */
1797         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1798                                             iclog->ic_datap, size);
1799
1800         bp->b_io_length = BTOBB(count);
1801         bp->b_fspriv = iclog;
1802         XFS_BUF_ZEROFLAGS(bp);
1803         XFS_BUF_ASYNC(bp);
1804         bp->b_flags |= XBF_SYNCIO;
1805
1806         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1807                 bp->b_flags |= XBF_FUA;
1808
1809                 /*
1810                  * Flush the data device before flushing the log to make
1811                  * sure all meta data written back from the AIL actually made
1812                  * it to disk before stamping the new log tail LSN into the
1813                  * log buffer.  For an external log we need to issue the
1814                  * flush explicitly, and unfortunately synchronously here;
1815                  * for an internal log we can simply use the block layer
1816                  * state machine for preflushes.
1817                  */
1818                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1819                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1820                 else
1821                         bp->b_flags |= XBF_FLUSH;
1822         }
1823
1824         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1825         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1826
1827         xlog_verify_iclog(log, iclog, count, true);
1828
1829         /* account for log which doesn't start at block #0 */
1830         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1831         /*
1832          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1833          * is shutting down.
1834          */
1835         XFS_BUF_WRITE(bp);
1836
1837         error = xlog_bdstrat(bp);
1838         if (error) {
1839                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1840                 return error;
1841         }
1842         if (split) {
1843                 bp = iclog->ic_log->l_xbuf;
1844                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1845                 xfs_buf_associate_memory(bp,
1846                                 (char *)&iclog->ic_header + count, split);
1847                 bp->b_fspriv = iclog;
1848                 XFS_BUF_ZEROFLAGS(bp);
1849                 XFS_BUF_ASYNC(bp);
1850                 bp->b_flags |= XBF_SYNCIO;
1851                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1852                         bp->b_flags |= XBF_FUA;
1853
1854                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1855                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1856
1857                 /* account for internal log which doesn't start at block #0 */
1858                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1859                 XFS_BUF_WRITE(bp);
1860                 error = xlog_bdstrat(bp);
1861                 if (error) {
1862                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1863                         return error;
1864                 }
1865         }
1866         return 0;
1867 }       /* xlog_sync */
1868
1869 /*
1870  * Deallocate a log structure
1871  */
1872 STATIC void
1873 xlog_dealloc_log(
1874         struct xlog     *log)
1875 {
1876         xlog_in_core_t  *iclog, *next_iclog;
1877         int             i;
1878
1879         xlog_cil_destroy(log);
1880
1881         /*
1882          * Cycle all the iclogbuf locks to make sure all log IO completion
1883          * is done before we tear down these buffers.
1884          */
1885         iclog = log->l_iclog;
1886         for (i = 0; i < log->l_iclog_bufs; i++) {
1887                 xfs_buf_lock(iclog->ic_bp);
1888                 xfs_buf_unlock(iclog->ic_bp);
1889                 iclog = iclog->ic_next;
1890         }
1891
1892         /*
1893          * Always need to ensure that the extra buffer does not point to memory
1894          * owned by another log buffer before we free it. Also, cycle the lock
1895          * first to ensure we've completed IO on it.
1896          */
1897         xfs_buf_lock(log->l_xbuf);
1898         xfs_buf_unlock(log->l_xbuf);
1899         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1900         xfs_buf_free(log->l_xbuf);
1901
1902         iclog = log->l_iclog;
1903         for (i = 0; i < log->l_iclog_bufs; i++) {
1904                 xfs_buf_free(iclog->ic_bp);
1905                 next_iclog = iclog->ic_next;
1906                 kmem_free(iclog);
1907                 iclog = next_iclog;
1908         }
1909         spinlock_destroy(&log->l_icloglock);
1910
1911         log->l_mp->m_log = NULL;
1912         kmem_free(log);
1913 }       /* xlog_dealloc_log */
1914
1915 /*
1916  * Update counters atomically now that memcpy is done.
1917  */
1918 /* ARGSUSED */
1919 static inline void
1920 xlog_state_finish_copy(
1921         struct xlog             *log,
1922         struct xlog_in_core     *iclog,
1923         int                     record_cnt,
1924         int                     copy_bytes)
1925 {
1926         spin_lock(&log->l_icloglock);
1927
1928         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1929         iclog->ic_offset += copy_bytes;
1930
1931         spin_unlock(&log->l_icloglock);
1932 }       /* xlog_state_finish_copy */
1933
1934
1935
1936
1937 /*
1938  * print out info relating to regions written which consume
1939  * the reservation
1940  */
1941 void
1942 xlog_print_tic_res(
1943         struct xfs_mount        *mp,
1944         struct xlog_ticket      *ticket)
1945 {
1946         uint i;
1947         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1948
1949         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1950         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1951             "bformat",
1952             "bchunk",
1953             "efi_format",
1954             "efd_format",
1955             "iformat",
1956             "icore",
1957             "iext",
1958             "ibroot",
1959             "ilocal",
1960             "iattr_ext",
1961             "iattr_broot",
1962             "iattr_local",
1963             "qformat",
1964             "dquot",
1965             "quotaoff",
1966             "LR header",
1967             "unmount",
1968             "commit",
1969             "trans header"
1970         };
1971         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1972             "SETATTR_NOT_SIZE",
1973             "SETATTR_SIZE",
1974             "INACTIVE",
1975             "CREATE",
1976             "CREATE_TRUNC",
1977             "TRUNCATE_FILE",
1978             "REMOVE",
1979             "LINK",
1980             "RENAME",
1981             "MKDIR",
1982             "RMDIR",
1983             "SYMLINK",
1984             "SET_DMATTRS",
1985             "GROWFS",
1986             "STRAT_WRITE",
1987             "DIOSTRAT",
1988             "WRITE_SYNC",
1989             "WRITEID",
1990             "ADDAFORK",
1991             "ATTRINVAL",
1992             "ATRUNCATE",
1993             "ATTR_SET",
1994             "ATTR_RM",
1995             "ATTR_FLAG",
1996             "CLEAR_AGI_BUCKET",
1997             "QM_SBCHANGE",
1998             "DUMMY1",
1999             "DUMMY2",
2000             "QM_QUOTAOFF",
2001             "QM_DQALLOC",
2002             "QM_SETQLIM",
2003             "QM_DQCLUSTER",
2004             "QM_QINOCREATE",
2005             "QM_QUOTAOFF_END",
2006             "SB_UNIT",
2007             "FSYNC_TS",
2008             "GROWFSRT_ALLOC",
2009             "GROWFSRT_ZERO",
2010             "GROWFSRT_FREE",
2011             "SWAPEXT"
2012         };
2013
2014         xfs_warn(mp,
2015                 "xlog_write: reservation summary:\n"
2016                 "  trans type  = %s (%u)\n"
2017                 "  unit res    = %d bytes\n"
2018                 "  current res = %d bytes\n"
2019                 "  total reg   = %u bytes (o/flow = %u bytes)\n"
2020                 "  ophdrs      = %u (ophdr space = %u bytes)\n"
2021                 "  ophdr + reg = %u bytes\n"
2022                 "  num regions = %u\n",
2023                 ((ticket->t_trans_type <= 0 ||
2024                   ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2025                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
2026                 ticket->t_trans_type,
2027                 ticket->t_unit_res,
2028                 ticket->t_curr_res,
2029                 ticket->t_res_arr_sum, ticket->t_res_o_flow,
2030                 ticket->t_res_num_ophdrs, ophdr_spc,
2031                 ticket->t_res_arr_sum +
2032                 ticket->t_res_o_flow + ophdr_spc,
2033                 ticket->t_res_num);
2034
2035         for (i = 0; i < ticket->t_res_num; i++) {
2036                 uint r_type = ticket->t_res_arr[i].r_type;
2037                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2038                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2039                             "bad-rtype" : res_type_str[r_type-1]),
2040                             ticket->t_res_arr[i].r_len);
2041         }
2042
2043         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2044                 "xlog_write: reservation ran out. Need to up reservation");
2045         xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2046 }
2047
2048 /*
2049  * Calculate the potential space needed by the log vector.  Each region gets
2050  * its own xlog_op_header_t and may need to be double word aligned.
2051  */
2052 static int
2053 xlog_write_calc_vec_length(
2054         struct xlog_ticket      *ticket,
2055         struct xfs_log_vec      *log_vector)
2056 {
2057         struct xfs_log_vec      *lv;
2058         int                     headers = 0;
2059         int                     len = 0;
2060         int                     i;
2061
2062         /* acct for start rec of xact */
2063         if (ticket->t_flags & XLOG_TIC_INITED)
2064                 headers++;
2065
2066         for (lv = log_vector; lv; lv = lv->lv_next) {
2067                 /* we don't write ordered log vectors */
2068                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2069                         continue;
2070
2071                 headers += lv->lv_niovecs;
2072
2073                 for (i = 0; i < lv->lv_niovecs; i++) {
2074                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2075
2076                         len += vecp->i_len;
2077                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2078                 }
2079         }
2080
2081         ticket->t_res_num_ophdrs += headers;
2082         len += headers * sizeof(struct xlog_op_header);
2083
2084         return len;
2085 }
2086
2087 /*
2088  * If first write for transaction, insert start record  We can't be trying to
2089  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2090  */
2091 static int
2092 xlog_write_start_rec(
2093         struct xlog_op_header   *ophdr,
2094         struct xlog_ticket      *ticket)
2095 {
2096         if (!(ticket->t_flags & XLOG_TIC_INITED))
2097                 return 0;
2098
2099         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2100         ophdr->oh_clientid = ticket->t_clientid;
2101         ophdr->oh_len = 0;
2102         ophdr->oh_flags = XLOG_START_TRANS;
2103         ophdr->oh_res2 = 0;
2104
2105         ticket->t_flags &= ~XLOG_TIC_INITED;
2106
2107         return sizeof(struct xlog_op_header);
2108 }
2109
2110 static xlog_op_header_t *
2111 xlog_write_setup_ophdr(
2112         struct xlog             *log,
2113         struct xlog_op_header   *ophdr,
2114         struct xlog_ticket      *ticket,
2115         uint                    flags)
2116 {
2117         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2118         ophdr->oh_clientid = ticket->t_clientid;
2119         ophdr->oh_res2 = 0;
2120
2121         /* are we copying a commit or unmount record? */
2122         ophdr->oh_flags = flags;
2123
2124         /*
2125          * We've seen logs corrupted with bad transaction client ids.  This
2126          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2127          * and shut down the filesystem.
2128          */
2129         switch (ophdr->oh_clientid)  {
2130         case XFS_TRANSACTION:
2131         case XFS_VOLUME:
2132         case XFS_LOG:
2133                 break;
2134         default:
2135                 xfs_warn(log->l_mp,
2136                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2137                         ophdr->oh_clientid, ticket);
2138                 return NULL;
2139         }
2140
2141         return ophdr;
2142 }
2143
2144 /*
2145  * Set up the parameters of the region copy into the log. This has
2146  * to handle region write split across multiple log buffers - this
2147  * state is kept external to this function so that this code can
2148  * be written in an obvious, self documenting manner.
2149  */
2150 static int
2151 xlog_write_setup_copy(
2152         struct xlog_ticket      *ticket,
2153         struct xlog_op_header   *ophdr,
2154         int                     space_available,
2155         int                     space_required,
2156         int                     *copy_off,
2157         int                     *copy_len,
2158         int                     *last_was_partial_copy,
2159         int                     *bytes_consumed)
2160 {
2161         int                     still_to_copy;
2162
2163         still_to_copy = space_required - *bytes_consumed;
2164         *copy_off = *bytes_consumed;
2165
2166         if (still_to_copy <= space_available) {
2167                 /* write of region completes here */
2168                 *copy_len = still_to_copy;
2169                 ophdr->oh_len = cpu_to_be32(*copy_len);
2170                 if (*last_was_partial_copy)
2171                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2172                 *last_was_partial_copy = 0;
2173                 *bytes_consumed = 0;
2174                 return 0;
2175         }
2176
2177         /* partial write of region, needs extra log op header reservation */
2178         *copy_len = space_available;
2179         ophdr->oh_len = cpu_to_be32(*copy_len);
2180         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2181         if (*last_was_partial_copy)
2182                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2183         *bytes_consumed += *copy_len;
2184         (*last_was_partial_copy)++;
2185
2186         /* account for new log op header */
2187         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2188         ticket->t_res_num_ophdrs++;
2189
2190         return sizeof(struct xlog_op_header);
2191 }
2192
2193 static int
2194 xlog_write_copy_finish(
2195         struct xlog             *log,
2196         struct xlog_in_core     *iclog,
2197         uint                    flags,
2198         int                     *record_cnt,
2199         int                     *data_cnt,
2200         int                     *partial_copy,
2201         int                     *partial_copy_len,
2202         int                     log_offset,
2203         struct xlog_in_core     **commit_iclog)
2204 {
2205         if (*partial_copy) {
2206                 /*
2207                  * This iclog has already been marked WANT_SYNC by
2208                  * xlog_state_get_iclog_space.
2209                  */
2210                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2211                 *record_cnt = 0;
2212                 *data_cnt = 0;
2213                 return xlog_state_release_iclog(log, iclog);
2214         }
2215
2216         *partial_copy = 0;
2217         *partial_copy_len = 0;
2218
2219         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2220                 /* no more space in this iclog - push it. */
2221                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2222                 *record_cnt = 0;
2223                 *data_cnt = 0;
2224
2225                 spin_lock(&log->l_icloglock);
2226                 xlog_state_want_sync(log, iclog);
2227                 spin_unlock(&log->l_icloglock);
2228
2229                 if (!commit_iclog)
2230                         return xlog_state_release_iclog(log, iclog);
2231                 ASSERT(flags & XLOG_COMMIT_TRANS);
2232                 *commit_iclog = iclog;
2233         }
2234
2235         return 0;
2236 }
2237
2238 /*
2239  * Write some region out to in-core log
2240  *
2241  * This will be called when writing externally provided regions or when
2242  * writing out a commit record for a given transaction.
2243  *
2244  * General algorithm:
2245  *      1. Find total length of this write.  This may include adding to the
2246  *              lengths passed in.
2247  *      2. Check whether we violate the tickets reservation.
2248  *      3. While writing to this iclog
2249  *          A. Reserve as much space in this iclog as can get
2250  *          B. If this is first write, save away start lsn
2251  *          C. While writing this region:
2252  *              1. If first write of transaction, write start record
2253  *              2. Write log operation header (header per region)
2254  *              3. Find out if we can fit entire region into this iclog
2255  *              4. Potentially, verify destination memcpy ptr
2256  *              5. Memcpy (partial) region
2257  *              6. If partial copy, release iclog; otherwise, continue
2258  *                      copying more regions into current iclog
2259  *      4. Mark want sync bit (in simulation mode)
2260  *      5. Release iclog for potential flush to on-disk log.
2261  *
2262  * ERRORS:
2263  * 1.   Panic if reservation is overrun.  This should never happen since
2264  *      reservation amounts are generated internal to the filesystem.
2265  * NOTES:
2266  * 1. Tickets are single threaded data structures.
2267  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2268  *      syncing routine.  When a single log_write region needs to span
2269  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2270  *      on all log operation writes which don't contain the end of the
2271  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2272  *      operation which contains the end of the continued log_write region.
2273  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2274  *      we don't really know exactly how much space will be used.  As a result,
2275  *      we don't update ic_offset until the end when we know exactly how many
2276  *      bytes have been written out.
2277  */
2278 int
2279 xlog_write(
2280         struct xlog             *log,
2281         struct xfs_log_vec      *log_vector,
2282         struct xlog_ticket      *ticket,
2283         xfs_lsn_t               *start_lsn,
2284         struct xlog_in_core     **commit_iclog,
2285         uint                    flags)
2286 {
2287         struct xlog_in_core     *iclog = NULL;
2288         struct xfs_log_iovec    *vecp;
2289         struct xfs_log_vec      *lv;
2290         int                     len;
2291         int                     index;
2292         int                     partial_copy = 0;
2293         int                     partial_copy_len = 0;
2294         int                     contwr = 0;
2295         int                     record_cnt = 0;
2296         int                     data_cnt = 0;
2297         int                     error;
2298
2299         *start_lsn = 0;
2300
2301         len = xlog_write_calc_vec_length(ticket, log_vector);
2302
2303         /*
2304          * Region headers and bytes are already accounted for.
2305          * We only need to take into account start records and
2306          * split regions in this function.
2307          */
2308         if (ticket->t_flags & XLOG_TIC_INITED)
2309                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2310
2311         /*
2312          * Commit record headers need to be accounted for. These
2313          * come in as separate writes so are easy to detect.
2314          */
2315         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2316                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2317
2318         if (ticket->t_curr_res < 0)
2319                 xlog_print_tic_res(log->l_mp, ticket);
2320
2321         index = 0;
2322         lv = log_vector;
2323         vecp = lv->lv_iovecp;
2324         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2325                 void            *ptr;
2326                 int             log_offset;
2327
2328                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2329                                                    &contwr, &log_offset);
2330                 if (error)
2331                         return error;
2332
2333                 ASSERT(log_offset <= iclog->ic_size - 1);
2334                 ptr = iclog->ic_datap + log_offset;
2335
2336                 /* start_lsn is the first lsn written to. That's all we need. */
2337                 if (!*start_lsn)
2338                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2339
2340                 /*
2341                  * This loop writes out as many regions as can fit in the amount
2342                  * of space which was allocated by xlog_state_get_iclog_space().
2343                  */
2344                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2345                         struct xfs_log_iovec    *reg;
2346                         struct xlog_op_header   *ophdr;
2347                         int                     start_rec_copy;
2348                         int                     copy_len;
2349                         int                     copy_off;
2350                         bool                    ordered = false;
2351
2352                         /* ordered log vectors have no regions to write */
2353                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2354                                 ASSERT(lv->lv_niovecs == 0);
2355                                 ordered = true;
2356                                 goto next_lv;
2357                         }
2358
2359                         reg = &vecp[index];
2360                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2361                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2362
2363                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2364                         if (start_rec_copy) {
2365                                 record_cnt++;
2366                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2367                                                    start_rec_copy);
2368                         }
2369
2370                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2371                         if (!ophdr)
2372                                 return -EIO;
2373
2374                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2375                                            sizeof(struct xlog_op_header));
2376
2377                         len += xlog_write_setup_copy(ticket, ophdr,
2378                                                      iclog->ic_size-log_offset,
2379                                                      reg->i_len,
2380                                                      &copy_off, &copy_len,
2381                                                      &partial_copy,
2382                                                      &partial_copy_len);
2383                         xlog_verify_dest_ptr(log, ptr);
2384
2385                         /* copy region */
2386                         ASSERT(copy_len >= 0);
2387                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
2388                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2389
2390                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2391                         record_cnt++;
2392                         data_cnt += contwr ? copy_len : 0;
2393
2394                         error = xlog_write_copy_finish(log, iclog, flags,
2395                                                        &record_cnt, &data_cnt,
2396                                                        &partial_copy,
2397                                                        &partial_copy_len,
2398                                                        log_offset,
2399                                                        commit_iclog);
2400                         if (error)
2401                                 return error;
2402
2403                         /*
2404                          * if we had a partial copy, we need to get more iclog
2405                          * space but we don't want to increment the region
2406                          * index because there is still more is this region to
2407                          * write.
2408                          *
2409                          * If we completed writing this region, and we flushed
2410                          * the iclog (indicated by resetting of the record
2411                          * count), then we also need to get more log space. If
2412                          * this was the last record, though, we are done and
2413                          * can just return.
2414                          */
2415                         if (partial_copy)
2416                                 break;
2417
2418                         if (++index == lv->lv_niovecs) {
2419 next_lv:
2420                                 lv = lv->lv_next;
2421                                 index = 0;
2422                                 if (lv)
2423                                         vecp = lv->lv_iovecp;
2424                         }
2425                         if (record_cnt == 0 && ordered == false) {
2426                                 if (!lv)
2427                                         return 0;
2428                                 break;
2429                         }
2430                 }
2431         }
2432
2433         ASSERT(len == 0);
2434
2435         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2436         if (!commit_iclog)
2437                 return xlog_state_release_iclog(log, iclog);
2438
2439         ASSERT(flags & XLOG_COMMIT_TRANS);
2440         *commit_iclog = iclog;
2441         return 0;
2442 }
2443
2444
2445 /*****************************************************************************
2446  *
2447  *              State Machine functions
2448  *
2449  *****************************************************************************
2450  */
2451
2452 /* Clean iclogs starting from the head.  This ordering must be
2453  * maintained, so an iclog doesn't become ACTIVE beyond one that
2454  * is SYNCING.  This is also required to maintain the notion that we use
2455  * a ordered wait queue to hold off would be writers to the log when every
2456  * iclog is trying to sync to disk.
2457  *
2458  * State Change: DIRTY -> ACTIVE
2459  */
2460 STATIC void
2461 xlog_state_clean_log(
2462         struct xlog *log)
2463 {
2464         xlog_in_core_t  *iclog;
2465         int changed = 0;
2466
2467         iclog = log->l_iclog;
2468         do {
2469                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2470                         iclog->ic_state = XLOG_STATE_ACTIVE;
2471                         iclog->ic_offset       = 0;
2472                         ASSERT(iclog->ic_callback == NULL);
2473                         /*
2474                          * If the number of ops in this iclog indicate it just
2475                          * contains the dummy transaction, we can
2476                          * change state into IDLE (the second time around).
2477                          * Otherwise we should change the state into
2478                          * NEED a dummy.
2479                          * We don't need to cover the dummy.
2480                          */
2481                         if (!changed &&
2482                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2483                                         XLOG_COVER_OPS)) {
2484                                 changed = 1;
2485                         } else {
2486                                 /*
2487                                  * We have two dirty iclogs so start over
2488                                  * This could also be num of ops indicates
2489                                  * this is not the dummy going out.
2490                                  */
2491                                 changed = 2;
2492                         }
2493                         iclog->ic_header.h_num_logops = 0;
2494                         memset(iclog->ic_header.h_cycle_data, 0,
2495                               sizeof(iclog->ic_header.h_cycle_data));
2496                         iclog->ic_header.h_lsn = 0;
2497                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2498                         /* do nothing */;
2499                 else
2500                         break;  /* stop cleaning */
2501                 iclog = iclog->ic_next;
2502         } while (iclog != log->l_iclog);
2503
2504         /* log is locked when we are called */
2505         /*
2506          * Change state for the dummy log recording.
2507          * We usually go to NEED. But we go to NEED2 if the changed indicates
2508          * we are done writing the dummy record.
2509          * If we are done with the second dummy recored (DONE2), then
2510          * we go to IDLE.
2511          */
2512         if (changed) {
2513                 switch (log->l_covered_state) {
2514                 case XLOG_STATE_COVER_IDLE:
2515                 case XLOG_STATE_COVER_NEED:
2516                 case XLOG_STATE_COVER_NEED2:
2517                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2518                         break;
2519
2520                 case XLOG_STATE_COVER_DONE:
2521                         if (changed == 1)
2522                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2523                         else
2524                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2525                         break;
2526
2527                 case XLOG_STATE_COVER_DONE2:
2528                         if (changed == 1)
2529                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2530                         else
2531                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2532                         break;
2533
2534                 default:
2535                         ASSERT(0);
2536                 }
2537         }
2538 }       /* xlog_state_clean_log */
2539
2540 STATIC xfs_lsn_t
2541 xlog_get_lowest_lsn(
2542         struct xlog     *log)
2543 {
2544         xlog_in_core_t  *lsn_log;
2545         xfs_lsn_t       lowest_lsn, lsn;
2546
2547         lsn_log = log->l_iclog;
2548         lowest_lsn = 0;
2549         do {
2550             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2551                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2552                 if ((lsn && !lowest_lsn) ||
2553                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2554                         lowest_lsn = lsn;
2555                 }
2556             }
2557             lsn_log = lsn_log->ic_next;
2558         } while (lsn_log != log->l_iclog);
2559         return lowest_lsn;
2560 }
2561
2562
2563 STATIC void
2564 xlog_state_do_callback(
2565         struct xlog             *log,
2566         int                     aborted,
2567         struct xlog_in_core     *ciclog)
2568 {
2569         xlog_in_core_t     *iclog;
2570         xlog_in_core_t     *first_iclog;        /* used to know when we've
2571                                                  * processed all iclogs once */
2572         xfs_log_callback_t *cb, *cb_next;
2573         int                flushcnt = 0;
2574         xfs_lsn_t          lowest_lsn;
2575         int                ioerrors;    /* counter: iclogs with errors */
2576         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2577         int                funcdidcallbacks; /* flag: function did callbacks */
2578         int                repeats;     /* for issuing console warnings if
2579                                          * looping too many times */
2580         int                wake = 0;
2581
2582         spin_lock(&log->l_icloglock);
2583         first_iclog = iclog = log->l_iclog;
2584         ioerrors = 0;
2585         funcdidcallbacks = 0;
2586         repeats = 0;
2587
2588         do {
2589                 /*
2590                  * Scan all iclogs starting with the one pointed to by the
2591                  * log.  Reset this starting point each time the log is
2592                  * unlocked (during callbacks).
2593                  *
2594                  * Keep looping through iclogs until one full pass is made
2595                  * without running any callbacks.
2596                  */
2597                 first_iclog = log->l_iclog;
2598                 iclog = log->l_iclog;
2599                 loopdidcallbacks = 0;
2600                 repeats++;
2601
2602                 do {
2603
2604                         /* skip all iclogs in the ACTIVE & DIRTY states */
2605                         if (iclog->ic_state &
2606                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2607                                 iclog = iclog->ic_next;
2608                                 continue;
2609                         }
2610
2611                         /*
2612                          * Between marking a filesystem SHUTDOWN and stopping
2613                          * the log, we do flush all iclogs to disk (if there
2614                          * wasn't a log I/O error). So, we do want things to
2615                          * go smoothly in case of just a SHUTDOWN  w/o a
2616                          * LOG_IO_ERROR.
2617                          */
2618                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2619                                 /*
2620                                  * Can only perform callbacks in order.  Since
2621                                  * this iclog is not in the DONE_SYNC/
2622                                  * DO_CALLBACK state, we skip the rest and
2623                                  * just try to clean up.  If we set our iclog
2624                                  * to DO_CALLBACK, we will not process it when
2625                                  * we retry since a previous iclog is in the
2626                                  * CALLBACK and the state cannot change since
2627                                  * we are holding the l_icloglock.
2628                                  */
2629                                 if (!(iclog->ic_state &
2630                                         (XLOG_STATE_DONE_SYNC |
2631                                                  XLOG_STATE_DO_CALLBACK))) {
2632                                         if (ciclog && (ciclog->ic_state ==
2633                                                         XLOG_STATE_DONE_SYNC)) {
2634                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2635                                         }
2636                                         break;
2637                                 }
2638                                 /*
2639                                  * We now have an iclog that is in either the
2640                                  * DO_CALLBACK or DONE_SYNC states. The other
2641                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2642                                  * caught by the above if and are going to
2643                                  * clean (i.e. we aren't doing their callbacks)
2644                                  * see the above if.
2645                                  */
2646
2647                                 /*
2648                                  * We will do one more check here to see if we
2649                                  * have chased our tail around.
2650                                  */
2651
2652                                 lowest_lsn = xlog_get_lowest_lsn(log);
2653                                 if (lowest_lsn &&
2654                                     XFS_LSN_CMP(lowest_lsn,
2655                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2656                                         iclog = iclog->ic_next;
2657                                         continue; /* Leave this iclog for
2658                                                    * another thread */
2659                                 }
2660
2661                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2662
2663
2664                                 /*
2665                                  * Completion of a iclog IO does not imply that
2666                                  * a transaction has completed, as transactions
2667                                  * can be large enough to span many iclogs. We
2668                                  * cannot change the tail of the log half way
2669                                  * through a transaction as this may be the only
2670                                  * transaction in the log and moving th etail to
2671                                  * point to the middle of it will prevent
2672                                  * recovery from finding the start of the
2673                                  * transaction. Hence we should only update the
2674                                  * last_sync_lsn if this iclog contains
2675                                  * transaction completion callbacks on it.
2676                                  *
2677                                  * We have to do this before we drop the
2678                                  * icloglock to ensure we are the only one that
2679                                  * can update it.
2680                                  */
2681                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2682                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2683                                 if (iclog->ic_callback)
2684                                         atomic64_set(&log->l_last_sync_lsn,
2685                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2686
2687                         } else
2688                                 ioerrors++;
2689
2690                         spin_unlock(&log->l_icloglock);
2691
2692                         /*
2693                          * Keep processing entries in the callback list until
2694                          * we come around and it is empty.  We need to
2695                          * atomically see that the list is empty and change the
2696                          * state to DIRTY so that we don't miss any more
2697                          * callbacks being added.
2698                          */
2699                         spin_lock(&iclog->ic_callback_lock);
2700                         cb = iclog->ic_callback;
2701                         while (cb) {
2702                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2703                                 iclog->ic_callback = NULL;
2704                                 spin_unlock(&iclog->ic_callback_lock);
2705
2706                                 /* perform callbacks in the order given */
2707                                 for (; cb; cb = cb_next) {
2708                                         cb_next = cb->cb_next;
2709                                         cb->cb_func(cb->cb_arg, aborted);
2710                                 }
2711                                 spin_lock(&iclog->ic_callback_lock);
2712                                 cb = iclog->ic_callback;
2713                         }
2714
2715                         loopdidcallbacks++;
2716                         funcdidcallbacks++;
2717
2718                         spin_lock(&log->l_icloglock);
2719                         ASSERT(iclog->ic_callback == NULL);
2720                         spin_unlock(&iclog->ic_callback_lock);
2721                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2722                                 iclog->ic_state = XLOG_STATE_DIRTY;
2723
2724                         /*
2725                          * Transition from DIRTY to ACTIVE if applicable.
2726                          * NOP if STATE_IOERROR.
2727                          */
2728                         xlog_state_clean_log(log);
2729
2730                         /* wake up threads waiting in xfs_log_force() */
2731                         wake_up_all(&iclog->ic_force_wait);
2732
2733                         iclog = iclog->ic_next;
2734                 } while (first_iclog != iclog);
2735
2736                 if (repeats > 5000) {
2737                         flushcnt += repeats;
2738                         repeats = 0;
2739                         xfs_warn(log->l_mp,
2740                                 "%s: possible infinite loop (%d iterations)",
2741                                 __func__, flushcnt);
2742                 }
2743         } while (!ioerrors && loopdidcallbacks);
2744
2745         /*
2746          * make one last gasp attempt to see if iclogs are being left in
2747          * limbo..
2748          */
2749 #ifdef DEBUG
2750         if (funcdidcallbacks) {
2751                 first_iclog = iclog = log->l_iclog;
2752                 do {
2753                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2754                         /*
2755                          * Terminate the loop if iclogs are found in states
2756                          * which will cause other threads to clean up iclogs.
2757                          *
2758                          * SYNCING - i/o completion will go through logs
2759                          * DONE_SYNC - interrupt thread should be waiting for
2760                          *              l_icloglock
2761                          * IOERROR - give up hope all ye who enter here
2762                          */
2763                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2764                             iclog->ic_state == XLOG_STATE_SYNCING ||
2765                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2766                             iclog->ic_state == XLOG_STATE_IOERROR )
2767                                 break;
2768                         iclog = iclog->ic_next;
2769                 } while (first_iclog != iclog);
2770         }
2771 #endif
2772
2773         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2774                 wake = 1;
2775         spin_unlock(&log->l_icloglock);
2776
2777         if (wake)
2778                 wake_up_all(&log->l_flush_wait);
2779 }
2780
2781
2782 /*
2783  * Finish transitioning this iclog to the dirty state.
2784  *
2785  * Make sure that we completely execute this routine only when this is
2786  * the last call to the iclog.  There is a good chance that iclog flushes,
2787  * when we reach the end of the physical log, get turned into 2 separate
2788  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2789  * routine.  By using the reference count bwritecnt, we guarantee that only
2790  * the second completion goes through.
2791  *
2792  * Callbacks could take time, so they are done outside the scope of the
2793  * global state machine log lock.
2794  */
2795 STATIC void
2796 xlog_state_done_syncing(
2797         xlog_in_core_t  *iclog,
2798         int             aborted)
2799 {
2800         struct xlog        *log = iclog->ic_log;
2801
2802         spin_lock(&log->l_icloglock);
2803
2804         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2805                iclog->ic_state == XLOG_STATE_IOERROR);
2806         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2807         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2808
2809
2810         /*
2811          * If we got an error, either on the first buffer, or in the case of
2812          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2813          * and none should ever be attempted to be written to disk
2814          * again.
2815          */
2816         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2817                 if (--iclog->ic_bwritecnt == 1) {
2818                         spin_unlock(&log->l_icloglock);
2819                         return;
2820                 }
2821                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2822         }
2823
2824         /*
2825          * Someone could be sleeping prior to writing out the next
2826          * iclog buffer, we wake them all, one will get to do the
2827          * I/O, the others get to wait for the result.
2828          */
2829         wake_up_all(&iclog->ic_write_wait);
2830         spin_unlock(&log->l_icloglock);
2831         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2832 }       /* xlog_state_done_syncing */
2833
2834
2835 /*
2836  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2837  * sleep.  We wait on the flush queue on the head iclog as that should be
2838  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2839  * we will wait here and all new writes will sleep until a sync completes.
2840  *
2841  * The in-core logs are used in a circular fashion. They are not used
2842  * out-of-order even when an iclog past the head is free.
2843  *
2844  * return:
2845  *      * log_offset where xlog_write() can start writing into the in-core
2846  *              log's data space.
2847  *      * in-core log pointer to which xlog_write() should write.
2848  *      * boolean indicating this is a continued write to an in-core log.
2849  *              If this is the last write, then the in-core log's offset field
2850  *              needs to be incremented, depending on the amount of data which
2851  *              is copied.
2852  */
2853 STATIC int
2854 xlog_state_get_iclog_space(
2855         struct xlog             *log,
2856         int                     len,
2857         struct xlog_in_core     **iclogp,
2858         struct xlog_ticket      *ticket,
2859         int                     *continued_write,
2860         int                     *logoffsetp)
2861 {
2862         int               log_offset;
2863         xlog_rec_header_t *head;
2864         xlog_in_core_t    *iclog;
2865         int               error;
2866
2867 restart:
2868         spin_lock(&log->l_icloglock);
2869         if (XLOG_FORCED_SHUTDOWN(log)) {
2870                 spin_unlock(&log->l_icloglock);
2871                 return -EIO;
2872         }
2873
2874         iclog = log->l_iclog;
2875         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2876                 XFS_STATS_INC(xs_log_noiclogs);
2877
2878                 /* Wait for log writes to have flushed */
2879                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2880                 goto restart;
2881         }
2882
2883         head = &iclog->ic_header;
2884
2885         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2886         log_offset = iclog->ic_offset;
2887
2888         /* On the 1st write to an iclog, figure out lsn.  This works
2889          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2890          * committing to.  If the offset is set, that's how many blocks
2891          * must be written.
2892          */
2893         if (log_offset == 0) {
2894                 ticket->t_curr_res -= log->l_iclog_hsize;
2895                 xlog_tic_add_region(ticket,
2896                                     log->l_iclog_hsize,
2897                                     XLOG_REG_TYPE_LRHEADER);
2898                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2899                 head->h_lsn = cpu_to_be64(
2900                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2901                 ASSERT(log->l_curr_block >= 0);
2902         }
2903
2904         /* If there is enough room to write everything, then do it.  Otherwise,
2905          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2906          * bit is on, so this will get flushed out.  Don't update ic_offset
2907          * until you know exactly how many bytes get copied.  Therefore, wait
2908          * until later to update ic_offset.
2909          *
2910          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2911          * can fit into remaining data section.
2912          */
2913         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2914                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2915
2916                 /*
2917                  * If I'm the only one writing to this iclog, sync it to disk.
2918                  * We need to do an atomic compare and decrement here to avoid
2919                  * racing with concurrent atomic_dec_and_lock() calls in
2920                  * xlog_state_release_iclog() when there is more than one
2921                  * reference to the iclog.
2922                  */
2923                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2924                         /* we are the only one */
2925                         spin_unlock(&log->l_icloglock);
2926                         error = xlog_state_release_iclog(log, iclog);
2927                         if (error)
2928                                 return error;
2929                 } else {
2930                         spin_unlock(&log->l_icloglock);
2931                 }
2932                 goto restart;
2933         }
2934
2935         /* Do we have enough room to write the full amount in the remainder
2936          * of this iclog?  Or must we continue a write on the next iclog and
2937          * mark this iclog as completely taken?  In the case where we switch
2938          * iclogs (to mark it taken), this particular iclog will release/sync
2939          * to disk in xlog_write().
2940          */
2941         if (len <= iclog->ic_size - iclog->ic_offset) {
2942                 *continued_write = 0;
2943                 iclog->ic_offset += len;
2944         } else {
2945                 *continued_write = 1;
2946                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2947         }
2948         *iclogp = iclog;
2949
2950         ASSERT(iclog->ic_offset <= iclog->ic_size);
2951         spin_unlock(&log->l_icloglock);
2952
2953         *logoffsetp = log_offset;
2954         return 0;
2955 }       /* xlog_state_get_iclog_space */
2956
2957 /* The first cnt-1 times through here we don't need to
2958  * move the grant write head because the permanent
2959  * reservation has reserved cnt times the unit amount.
2960  * Release part of current permanent unit reservation and
2961  * reset current reservation to be one units worth.  Also
2962  * move grant reservation head forward.
2963  */
2964 STATIC void
2965 xlog_regrant_reserve_log_space(
2966         struct xlog             *log,
2967         struct xlog_ticket      *ticket)
2968 {
2969         trace_xfs_log_regrant_reserve_enter(log, ticket);
2970
2971         if (ticket->t_cnt > 0)
2972                 ticket->t_cnt--;
2973
2974         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2975                                         ticket->t_curr_res);
2976         xlog_grant_sub_space(log, &log->l_write_head.grant,
2977                                         ticket->t_curr_res);
2978         ticket->t_curr_res = ticket->t_unit_res;
2979         xlog_tic_reset_res(ticket);
2980
2981         trace_xfs_log_regrant_reserve_sub(log, ticket);
2982
2983         /* just return if we still have some of the pre-reserved space */
2984         if (ticket->t_cnt > 0)
2985                 return;
2986
2987         xlog_grant_add_space(log, &log->l_reserve_head.grant,
2988                                         ticket->t_unit_res);
2989
2990         trace_xfs_log_regrant_reserve_exit(log, ticket);
2991
2992         ticket->t_curr_res = ticket->t_unit_res;
2993         xlog_tic_reset_res(ticket);
2994 }       /* xlog_regrant_reserve_log_space */
2995
2996
2997 /*
2998  * Give back the space left from a reservation.
2999  *
3000  * All the information we need to make a correct determination of space left
3001  * is present.  For non-permanent reservations, things are quite easy.  The
3002  * count should have been decremented to zero.  We only need to deal with the
3003  * space remaining in the current reservation part of the ticket.  If the
3004  * ticket contains a permanent reservation, there may be left over space which
3005  * needs to be released.  A count of N means that N-1 refills of the current
3006  * reservation can be done before we need to ask for more space.  The first
3007  * one goes to fill up the first current reservation.  Once we run out of
3008  * space, the count will stay at zero and the only space remaining will be
3009  * in the current reservation field.
3010  */
3011 STATIC void
3012 xlog_ungrant_log_space(
3013         struct xlog             *log,
3014         struct xlog_ticket      *ticket)
3015 {
3016         int     bytes;
3017
3018         if (ticket->t_cnt > 0)
3019                 ticket->t_cnt--;
3020
3021         trace_xfs_log_ungrant_enter(log, ticket);
3022         trace_xfs_log_ungrant_sub(log, ticket);
3023
3024         /*
3025          * If this is a permanent reservation ticket, we may be able to free
3026          * up more space based on the remaining count.
3027          */
3028         bytes = ticket->t_curr_res;
3029         if (ticket->t_cnt > 0) {
3030                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3031                 bytes += ticket->t_unit_res*ticket->t_cnt;
3032         }
3033
3034         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3035         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3036
3037         trace_xfs_log_ungrant_exit(log, ticket);
3038
3039         xfs_log_space_wake(log->l_mp);
3040 }
3041
3042 /*
3043  * Flush iclog to disk if this is the last reference to the given iclog and
3044  * the WANT_SYNC bit is set.
3045  *
3046  * When this function is entered, the iclog is not necessarily in the
3047  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3048  *
3049  *
3050  */
3051 STATIC int
3052 xlog_state_release_iclog(
3053         struct xlog             *log,
3054         struct xlog_in_core     *iclog)
3055 {
3056         int             sync = 0;       /* do we sync? */
3057
3058         if (iclog->ic_state & XLOG_STATE_IOERROR)
3059                 return -EIO;
3060
3061         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3062         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3063                 return 0;
3064
3065         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3066                 spin_unlock(&log->l_icloglock);
3067                 return -EIO;
3068         }
3069         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3070                iclog->ic_state == XLOG_STATE_WANT_SYNC);
3071
3072         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3073                 /* update tail before writing to iclog */
3074                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3075                 sync++;
3076                 iclog->ic_state = XLOG_STATE_SYNCING;
3077                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3078                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3079                 /* cycle incremented when incrementing curr_block */
3080         }
3081         spin_unlock(&log->l_icloglock);
3082
3083         /*
3084          * We let the log lock go, so it's possible that we hit a log I/O
3085          * error or some other SHUTDOWN condition that marks the iclog
3086          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3087          * this iclog has consistent data, so we ignore IOERROR
3088          * flags after this point.
3089          */
3090         if (sync)
3091                 return xlog_sync(log, iclog);
3092         return 0;
3093 }       /* xlog_state_release_iclog */
3094
3095
3096 /*
3097  * This routine will mark the current iclog in the ring as WANT_SYNC
3098  * and move the current iclog pointer to the next iclog in the ring.
3099  * When this routine is called from xlog_state_get_iclog_space(), the
3100  * exact size of the iclog has not yet been determined.  All we know is
3101  * that every data block.  We have run out of space in this log record.
3102  */
3103 STATIC void
3104 xlog_state_switch_iclogs(
3105         struct xlog             *log,
3106         struct xlog_in_core     *iclog,
3107         int                     eventual_size)
3108 {
3109         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3110         if (!eventual_size)
3111                 eventual_size = iclog->ic_offset;
3112         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3113         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3114         log->l_prev_block = log->l_curr_block;
3115         log->l_prev_cycle = log->l_curr_cycle;
3116
3117         /* roll log?: ic_offset changed later */
3118         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3119
3120         /* Round up to next log-sunit */
3121         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3122             log->l_mp->m_sb.sb_logsunit > 1) {
3123                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3124                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3125         }
3126
3127         if (log->l_curr_block >= log->l_logBBsize) {
3128                 log->l_curr_cycle++;
3129                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3130                         log->l_curr_cycle++;
3131                 log->l_curr_block -= log->l_logBBsize;
3132                 ASSERT(log->l_curr_block >= 0);
3133         }
3134         ASSERT(iclog == log->l_iclog);
3135         log->l_iclog = iclog->ic_next;
3136 }       /* xlog_state_switch_iclogs */
3137
3138 /*
3139  * Write out all data in the in-core log as of this exact moment in time.
3140  *
3141  * Data may be written to the in-core log during this call.  However,
3142  * we don't guarantee this data will be written out.  A change from past
3143  * implementation means this routine will *not* write out zero length LRs.
3144  *
3145  * Basically, we try and perform an intelligent scan of the in-core logs.
3146  * If we determine there is no flushable data, we just return.  There is no
3147  * flushable data if:
3148  *
3149  *      1. the current iclog is active and has no data; the previous iclog
3150  *              is in the active or dirty state.
3151  *      2. the current iclog is drity, and the previous iclog is in the
3152  *              active or dirty state.
3153  *
3154  * We may sleep if:
3155  *
3156  *      1. the current iclog is not in the active nor dirty state.
3157  *      2. the current iclog dirty, and the previous iclog is not in the
3158  *              active nor dirty state.
3159  *      3. the current iclog is active, and there is another thread writing
3160  *              to this particular iclog.
3161  *      4. a) the current iclog is active and has no other writers
3162  *         b) when we return from flushing out this iclog, it is still
3163  *              not in the active nor dirty state.
3164  */
3165 int
3166 _xfs_log_force(
3167         struct xfs_mount        *mp,
3168         uint                    flags,
3169         int                     *log_flushed)
3170 {
3171         struct xlog             *log = mp->m_log;
3172         struct xlog_in_core     *iclog;
3173         xfs_lsn_t               lsn;
3174
3175         XFS_STATS_INC(xs_log_force);
3176
3177         xlog_cil_force(log);
3178
3179         spin_lock(&log->l_icloglock);
3180
3181         iclog = log->l_iclog;
3182         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3183                 spin_unlock(&log->l_icloglock);
3184                 return -EIO;
3185         }
3186
3187         /* If the head iclog is not active nor dirty, we just attach
3188          * ourselves to the head and go to sleep.
3189          */
3190         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3191             iclog->ic_state == XLOG_STATE_DIRTY) {
3192                 /*
3193                  * If the head is dirty or (active and empty), then
3194                  * we need to look at the previous iclog.  If the previous
3195                  * iclog is active or dirty we are done.  There is nothing
3196                  * to sync out.  Otherwise, we attach ourselves to the
3197                  * previous iclog and go to sleep.
3198                  */
3199                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3200                     (atomic_read(&iclog->ic_refcnt) == 0
3201                      && iclog->ic_offset == 0)) {
3202                         iclog = iclog->ic_prev;
3203                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3204                             iclog->ic_state == XLOG_STATE_DIRTY)
3205                                 goto no_sleep;
3206                         else
3207                                 goto maybe_sleep;
3208                 } else {
3209                         if (atomic_read(&iclog->ic_refcnt) == 0) {
3210                                 /* We are the only one with access to this
3211                                  * iclog.  Flush it out now.  There should
3212                                  * be a roundoff of zero to show that someone
3213                                  * has already taken care of the roundoff from
3214                                  * the previous sync.
3215                                  */
3216                                 atomic_inc(&iclog->ic_refcnt);
3217                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3218                                 xlog_state_switch_iclogs(log, iclog, 0);
3219                                 spin_unlock(&log->l_icloglock);
3220
3221                                 if (xlog_state_release_iclog(log, iclog))
3222                                         return -EIO;
3223
3224                                 if (log_flushed)
3225                                         *log_flushed = 1;
3226                                 spin_lock(&log->l_icloglock);
3227                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3228                                     iclog->ic_state != XLOG_STATE_DIRTY)
3229                                         goto maybe_sleep;
3230                                 else
3231                                         goto no_sleep;
3232                         } else {
3233                                 /* Someone else is writing to this iclog.
3234                                  * Use its call to flush out the data.  However,
3235                                  * the other thread may not force out this LR,
3236                                  * so we mark it WANT_SYNC.
3237                                  */
3238                                 xlog_state_switch_iclogs(log, iclog, 0);
3239                                 goto maybe_sleep;
3240                         }
3241                 }
3242         }
3243
3244         /* By the time we come around again, the iclog could've been filled
3245          * which would give it another lsn.  If we have a new lsn, just
3246          * return because the relevant data has been flushed.
3247          */
3248 maybe_sleep:
3249         if (flags & XFS_LOG_SYNC) {
3250                 /*
3251                  * We must check if we're shutting down here, before
3252                  * we wait, while we're holding the l_icloglock.
3253                  * Then we check again after waking up, in case our
3254                  * sleep was disturbed by a bad news.
3255                  */
3256                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3257                         spin_unlock(&log->l_icloglock);
3258                         return -EIO;
3259                 }
3260                 XFS_STATS_INC(xs_log_force_sleep);
3261                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3262                 /*
3263                  * No need to grab the log lock here since we're
3264                  * only deciding whether or not to return EIO
3265                  * and the memory read should be atomic.
3266                  */
3267                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3268                         return -EIO;
3269                 if (log_flushed)
3270                         *log_flushed = 1;
3271         } else {
3272
3273 no_sleep:
3274                 spin_unlock(&log->l_icloglock);
3275         }
3276         return 0;
3277 }
3278
3279 /*
3280  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3281  * about errors or whether the log was flushed or not. This is the normal
3282  * interface to use when trying to unpin items or move the log forward.
3283  */
3284 void
3285 xfs_log_force(
3286         xfs_mount_t     *mp,
3287         uint            flags)
3288 {
3289         int     error;
3290
3291         trace_xfs_log_force(mp, 0);
3292         error = _xfs_log_force(mp, flags, NULL);
3293         if (error)
3294                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3295 }
3296
3297 /*
3298  * Force the in-core log to disk for a specific LSN.
3299  *
3300  * Find in-core log with lsn.
3301  *      If it is in the DIRTY state, just return.
3302  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3303  *              state and go to sleep or return.
3304  *      If it is in any other state, go to sleep or return.
3305  *
3306  * Synchronous forces are implemented with a signal variable. All callers
3307  * to force a given lsn to disk will wait on a the sv attached to the
3308  * specific in-core log.  When given in-core log finally completes its
3309  * write to disk, that thread will wake up all threads waiting on the
3310  * sv.
3311  */
3312 int
3313 _xfs_log_force_lsn(
3314         struct xfs_mount        *mp,
3315         xfs_lsn_t               lsn,
3316         uint                    flags,
3317         int                     *log_flushed)
3318 {
3319         struct xlog             *log = mp->m_log;
3320         struct xlog_in_core     *iclog;
3321         int                     already_slept = 0;
3322
3323         ASSERT(lsn != 0);
3324
3325         XFS_STATS_INC(xs_log_force);
3326
3327         lsn = xlog_cil_force_lsn(log, lsn);
3328         if (lsn == NULLCOMMITLSN)
3329                 return 0;
3330
3331 try_again:
3332         spin_lock(&log->l_icloglock);
3333         iclog = log->l_iclog;
3334         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3335                 spin_unlock(&log->l_icloglock);
3336                 return -EIO;
3337         }
3338
3339         do {
3340                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3341                         iclog = iclog->ic_next;
3342                         continue;
3343                 }
3344
3345                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3346                         spin_unlock(&log->l_icloglock);
3347                         return 0;
3348                 }
3349
3350                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3351                         /*
3352                          * We sleep here if we haven't already slept (e.g.
3353                          * this is the first time we've looked at the correct
3354                          * iclog buf) and the buffer before us is going to
3355                          * be sync'ed. The reason for this is that if we
3356                          * are doing sync transactions here, by waiting for
3357                          * the previous I/O to complete, we can allow a few
3358                          * more transactions into this iclog before we close
3359                          * it down.
3360                          *
3361                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3362                          * up the refcnt so we can release the log (which
3363                          * drops the ref count).  The state switch keeps new
3364                          * transaction commits from using this buffer.  When
3365                          * the current commits finish writing into the buffer,
3366                          * the refcount will drop to zero and the buffer will
3367                          * go out then.
3368                          */
3369                         if (!already_slept &&
3370                             (iclog->ic_prev->ic_state &
3371                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3372                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3373
3374                                 XFS_STATS_INC(xs_log_force_sleep);
3375
3376                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3377                                                         &log->l_icloglock);
3378                                 if (log_flushed)
3379                                         *log_flushed = 1;
3380                                 already_slept = 1;
3381                                 goto try_again;
3382                         }
3383                         atomic_inc(&iclog->ic_refcnt);
3384                         xlog_state_switch_iclogs(log, iclog, 0);
3385                         spin_unlock(&log->l_icloglock);
3386                         if (xlog_state_release_iclog(log, iclog))
3387                                 return -EIO;
3388                         if (log_flushed)
3389                                 *log_flushed = 1;
3390                         spin_lock(&log->l_icloglock);
3391                 }
3392
3393                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3394                     !(iclog->ic_state &
3395                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3396                         /*
3397                          * Don't wait on completion if we know that we've
3398                          * gotten a log write error.
3399                          */
3400                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3401                                 spin_unlock(&log->l_icloglock);
3402                                 return -EIO;
3403                         }
3404                         XFS_STATS_INC(xs_log_force_sleep);
3405                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3406                         /*
3407                          * No need to grab the log lock here since we're
3408                          * only deciding whether or not to return EIO
3409                          * and the memory read should be atomic.
3410                          */
3411                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3412                                 return -EIO;
3413
3414                         if (log_flushed)
3415                                 *log_flushed = 1;
3416                 } else {                /* just return */
3417                         spin_unlock(&log->l_icloglock);
3418                 }
3419
3420                 return 0;
3421         } while (iclog != log->l_iclog);
3422
3423         spin_unlock(&log->l_icloglock);
3424         return 0;
3425 }
3426
3427 /*
3428  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3429  * about errors or whether the log was flushed or not. This is the normal
3430  * interface to use when trying to unpin items or move the log forward.
3431  */
3432 void
3433 xfs_log_force_lsn(
3434         xfs_mount_t     *mp,
3435         xfs_lsn_t       lsn,
3436         uint            flags)
3437 {
3438         int     error;
3439
3440         trace_xfs_log_force(mp, lsn);
3441         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3442         if (error)
3443                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3444 }
3445
3446 /*
3447  * Called when we want to mark the current iclog as being ready to sync to
3448  * disk.
3449  */
3450 STATIC void
3451 xlog_state_want_sync(
3452         struct xlog             *log,
3453         struct xlog_in_core     *iclog)
3454 {
3455         assert_spin_locked(&log->l_icloglock);
3456
3457         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3458                 xlog_state_switch_iclogs(log, iclog, 0);
3459         } else {
3460                 ASSERT(iclog->ic_state &
3461                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3462         }
3463 }
3464
3465
3466 /*****************************************************************************
3467  *
3468  *              TICKET functions
3469  *
3470  *****************************************************************************
3471  */
3472
3473 /*
3474  * Free a used ticket when its refcount falls to zero.
3475  */
3476 void
3477 xfs_log_ticket_put(
3478         xlog_ticket_t   *ticket)
3479 {
3480         ASSERT(atomic_read(&ticket->t_ref) > 0);
3481         if (atomic_dec_and_test(&ticket->t_ref))
3482                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3483 }
3484
3485 xlog_ticket_t *
3486 xfs_log_ticket_get(
3487         xlog_ticket_t   *ticket)
3488 {
3489         ASSERT(atomic_read(&ticket->t_ref) > 0);
3490         atomic_inc(&ticket->t_ref);
3491         return ticket;
3492 }
3493
3494 /*
3495  * Figure out the total log space unit (in bytes) that would be
3496  * required for a log ticket.
3497  */
3498 int
3499 xfs_log_calc_unit_res(
3500         struct xfs_mount        *mp,
3501         int                     unit_bytes)
3502 {
3503         struct xlog             *log = mp->m_log;
3504         int                     iclog_space;
3505         uint                    num_headers;
3506
3507         /*
3508          * Permanent reservations have up to 'cnt'-1 active log operations
3509          * in the log.  A unit in this case is the amount of space for one
3510          * of these log operations.  Normal reservations have a cnt of 1
3511          * and their unit amount is the total amount of space required.
3512          *
3513          * The following lines of code account for non-transaction data
3514          * which occupy space in the on-disk log.
3515          *
3516          * Normal form of a transaction is:
3517          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3518          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3519          *
3520          * We need to account for all the leadup data and trailer data
3521          * around the transaction data.
3522          * And then we need to account for the worst case in terms of using
3523          * more space.
3524          * The worst case will happen if:
3525          * - the placement of the transaction happens to be such that the
3526          *   roundoff is at its maximum
3527          * - the transaction data is synced before the commit record is synced
3528          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3529          *   Therefore the commit record is in its own Log Record.
3530          *   This can happen as the commit record is called with its
3531          *   own region to xlog_write().
3532          *   This then means that in the worst case, roundoff can happen for
3533          *   the commit-rec as well.
3534          *   The commit-rec is smaller than padding in this scenario and so it is
3535          *   not added separately.
3536          */
3537
3538         /* for trans header */
3539         unit_bytes += sizeof(xlog_op_header_t);
3540         unit_bytes += sizeof(xfs_trans_header_t);
3541
3542         /* for start-rec */
3543         unit_bytes += sizeof(xlog_op_header_t);
3544
3545         /*
3546          * for LR headers - the space for data in an iclog is the size minus
3547          * the space used for the headers. If we use the iclog size, then we
3548          * undercalculate the number of headers required.
3549          *
3550          * Furthermore - the addition of op headers for split-recs might
3551          * increase the space required enough to require more log and op
3552          * headers, so take that into account too.
3553          *
3554          * IMPORTANT: This reservation makes the assumption that if this
3555          * transaction is the first in an iclog and hence has the LR headers
3556          * accounted to it, then the remaining space in the iclog is
3557          * exclusively for this transaction.  i.e. if the transaction is larger
3558          * than the iclog, it will be the only thing in that iclog.
3559          * Fundamentally, this means we must pass the entire log vector to
3560          * xlog_write to guarantee this.
3561          */
3562         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3563         num_headers = howmany(unit_bytes, iclog_space);
3564
3565         /* for split-recs - ophdrs added when data split over LRs */
3566         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3567
3568         /* add extra header reservations if we overrun */
3569         while (!num_headers ||
3570                howmany(unit_bytes, iclog_space) > num_headers) {
3571                 unit_bytes += sizeof(xlog_op_header_t);
3572                 num_headers++;
3573         }
3574         unit_bytes += log->l_iclog_hsize * num_headers;
3575
3576         /* for commit-rec LR header - note: padding will subsume the ophdr */
3577         unit_bytes += log->l_iclog_hsize;
3578
3579         /* for roundoff padding for transaction data and one for commit record */
3580         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3581                 /* log su roundoff */
3582                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3583         } else {
3584                 /* BB roundoff */
3585                 unit_bytes += 2 * BBSIZE;
3586         }
3587
3588         return unit_bytes;
3589 }
3590
3591 /*
3592  * Allocate and initialise a new log ticket.
3593  */
3594 struct xlog_ticket *
3595 xlog_ticket_alloc(
3596         struct xlog             *log,
3597         int                     unit_bytes,
3598         int                     cnt,
3599         char                    client,
3600         bool                    permanent,
3601         xfs_km_flags_t          alloc_flags)
3602 {
3603         struct xlog_ticket      *tic;
3604         int                     unit_res;
3605
3606         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3607         if (!tic)
3608                 return NULL;
3609
3610         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3611
3612         atomic_set(&tic->t_ref, 1);
3613         tic->t_task             = current;
3614         INIT_LIST_HEAD(&tic->t_queue);
3615         tic->t_unit_res         = unit_res;
3616         tic->t_curr_res         = unit_res;
3617         tic->t_cnt              = cnt;
3618         tic->t_ocnt             = cnt;
3619         tic->t_tid              = prandom_u32();
3620         tic->t_clientid         = client;
3621         tic->t_flags            = XLOG_TIC_INITED;
3622         tic->t_trans_type       = 0;
3623         if (permanent)
3624                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3625
3626         xlog_tic_reset_res(tic);
3627
3628         return tic;
3629 }
3630
3631
3632 /******************************************************************************
3633  *
3634  *              Log debug routines
3635  *
3636  ******************************************************************************
3637  */
3638 #if defined(DEBUG)
3639 /*
3640  * Make sure that the destination ptr is within the valid data region of
3641  * one of the iclogs.  This uses backup pointers stored in a different
3642  * part of the log in case we trash the log structure.
3643  */
3644 void
3645 xlog_verify_dest_ptr(
3646         struct xlog     *log,
3647         char            *ptr)
3648 {
3649         int i;
3650         int good_ptr = 0;
3651
3652         for (i = 0; i < log->l_iclog_bufs; i++) {
3653                 if (ptr >= log->l_iclog_bak[i] &&
3654                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3655                         good_ptr++;
3656         }
3657
3658         if (!good_ptr)
3659                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3660 }
3661
3662 /*
3663  * Check to make sure the grant write head didn't just over lap the tail.  If
3664  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3665  * the cycles differ by exactly one and check the byte count.
3666  *
3667  * This check is run unlocked, so can give false positives. Rather than assert
3668  * on failures, use a warn-once flag and a panic tag to allow the admin to
3669  * determine if they want to panic the machine when such an error occurs. For
3670  * debug kernels this will have the same effect as using an assert but, unlinke
3671  * an assert, it can be turned off at runtime.
3672  */
3673 STATIC void
3674 xlog_verify_grant_tail(
3675         struct xlog     *log)
3676 {
3677         int             tail_cycle, tail_blocks;
3678         int             cycle, space;
3679
3680         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3681         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3682         if (tail_cycle != cycle) {
3683                 if (cycle - 1 != tail_cycle &&
3684                     !(log->l_flags & XLOG_TAIL_WARN)) {
3685                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3686                                 "%s: cycle - 1 != tail_cycle", __func__);
3687                         log->l_flags |= XLOG_TAIL_WARN;
3688                 }
3689
3690                 if (space > BBTOB(tail_blocks) &&
3691                     !(log->l_flags & XLOG_TAIL_WARN)) {
3692                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3693                                 "%s: space > BBTOB(tail_blocks)", __func__);
3694                         log->l_flags |= XLOG_TAIL_WARN;
3695                 }
3696         }
3697 }
3698
3699 /* check if it will fit */
3700 STATIC void
3701 xlog_verify_tail_lsn(
3702         struct xlog             *log,
3703         struct xlog_in_core     *iclog,
3704         xfs_lsn_t               tail_lsn)
3705 {
3706     int blocks;
3707
3708     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3709         blocks =
3710             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3711         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3712                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3713     } else {
3714         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3715
3716         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3717                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3718
3719         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3720         if (blocks < BTOBB(iclog->ic_offset) + 1)
3721                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3722     }
3723 }       /* xlog_verify_tail_lsn */
3724
3725 /*
3726  * Perform a number of checks on the iclog before writing to disk.
3727  *
3728  * 1. Make sure the iclogs are still circular
3729  * 2. Make sure we have a good magic number
3730  * 3. Make sure we don't have magic numbers in the data
3731  * 4. Check fields of each log operation header for:
3732  *      A. Valid client identifier
3733  *      B. tid ptr value falls in valid ptr space (user space code)
3734  *      C. Length in log record header is correct according to the
3735  *              individual operation headers within record.
3736  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3737  *      log, check the preceding blocks of the physical log to make sure all
3738  *      the cycle numbers agree with the current cycle number.
3739  */
3740 STATIC void
3741 xlog_verify_iclog(
3742         struct xlog             *log,
3743         struct xlog_in_core     *iclog,
3744         int                     count,
3745         bool                    syncing)
3746 {
3747         xlog_op_header_t        *ophead;
3748         xlog_in_core_t          *icptr;
3749         xlog_in_core_2_t        *xhdr;
3750         xfs_caddr_t             ptr;
3751         xfs_caddr_t             base_ptr;
3752         __psint_t               field_offset;
3753         __uint8_t               clientid;
3754         int                     len, i, j, k, op_len;
3755         int                     idx;
3756
3757         /* check validity of iclog pointers */
3758         spin_lock(&log->l_icloglock);
3759         icptr = log->l_iclog;
3760         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3761                 ASSERT(icptr);
3762
3763         if (icptr != log->l_iclog)
3764                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3765         spin_unlock(&log->l_icloglock);
3766
3767         /* check log magic numbers */
3768         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3769                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3770
3771         ptr = (xfs_caddr_t) &iclog->ic_header;
3772         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3773              ptr += BBSIZE) {
3774                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3775                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3776                                 __func__);
3777         }
3778
3779         /* check fields */
3780         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3781         ptr = iclog->ic_datap;
3782         base_ptr = ptr;
3783         ophead = (xlog_op_header_t *)ptr;
3784         xhdr = iclog->ic_data;
3785         for (i = 0; i < len; i++) {
3786                 ophead = (xlog_op_header_t *)ptr;
3787
3788                 /* clientid is only 1 byte */
3789                 field_offset = (__psint_t)
3790                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3791                 if (!syncing || (field_offset & 0x1ff)) {
3792                         clientid = ophead->oh_clientid;
3793                 } else {
3794                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3795                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3796                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3797                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3798                                 clientid = xlog_get_client_id(
3799                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3800                         } else {
3801                                 clientid = xlog_get_client_id(
3802                                         iclog->ic_header.h_cycle_data[idx]);
3803                         }
3804                 }
3805                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3806                         xfs_warn(log->l_mp,
3807                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3808                                 __func__, clientid, ophead,
3809                                 (unsigned long)field_offset);
3810
3811                 /* check length */
3812                 field_offset = (__psint_t)
3813                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3814                 if (!syncing || (field_offset & 0x1ff)) {
3815                         op_len = be32_to_cpu(ophead->oh_len);
3816                 } else {
3817                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3818                                     (__psint_t)iclog->ic_datap);
3819                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3820                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3821                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3822                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3823                         } else {
3824                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3825                         }
3826                 }
3827                 ptr += sizeof(xlog_op_header_t) + op_len;
3828         }
3829 }       /* xlog_verify_iclog */
3830 #endif
3831
3832 /*
3833  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3834  */
3835 STATIC int
3836 xlog_state_ioerror(
3837         struct xlog     *log)
3838 {
3839         xlog_in_core_t  *iclog, *ic;
3840
3841         iclog = log->l_iclog;
3842         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3843                 /*
3844                  * Mark all the incore logs IOERROR.
3845                  * From now on, no log flushes will result.
3846                  */
3847                 ic = iclog;
3848                 do {
3849                         ic->ic_state = XLOG_STATE_IOERROR;
3850                         ic = ic->ic_next;
3851                 } while (ic != iclog);
3852                 return 0;
3853         }
3854         /*
3855          * Return non-zero, if state transition has already happened.
3856          */
3857         return 1;
3858 }
3859
3860 /*
3861  * This is called from xfs_force_shutdown, when we're forcibly
3862  * shutting down the filesystem, typically because of an IO error.
3863  * Our main objectives here are to make sure that:
3864  *      a. the filesystem gets marked 'SHUTDOWN' for all interested
3865  *         parties to find out, 'atomically'.
3866  *      b. those who're sleeping on log reservations, pinned objects and
3867  *          other resources get woken up, and be told the bad news.
3868  *      c. nothing new gets queued up after (a) and (b) are done.
3869  *      d. if !logerror, flush the iclogs to disk, then seal them off
3870  *         for business.
3871  *
3872  * Note: for delayed logging the !logerror case needs to flush the regions
3873  * held in memory out to the iclogs before flushing them to disk. This needs
3874  * to be done before the log is marked as shutdown, otherwise the flush to the
3875  * iclogs will fail.
3876  */
3877 int
3878 xfs_log_force_umount(
3879         struct xfs_mount        *mp,
3880         int                     logerror)
3881 {
3882         struct xlog     *log;
3883         int             retval;
3884
3885         log = mp->m_log;
3886
3887         /*
3888          * If this happens during log recovery, don't worry about
3889          * locking; the log isn't open for business yet.
3890          */
3891         if (!log ||
3892             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3893                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3894                 if (mp->m_sb_bp)
3895                         XFS_BUF_DONE(mp->m_sb_bp);
3896                 return 0;
3897         }
3898
3899         /*
3900          * Somebody could've already done the hard work for us.
3901          * No need to get locks for this.
3902          */
3903         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3904                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3905                 return 1;
3906         }
3907         retval = 0;
3908
3909         /*
3910          * Flush the in memory commit item list before marking the log as
3911          * being shut down. We need to do it in this order to ensure all the
3912          * completed transactions are flushed to disk with the xfs_log_force()
3913          * call below.
3914          */
3915         if (!logerror)
3916                 xlog_cil_force(log);
3917
3918         /*
3919          * mark the filesystem and the as in a shutdown state and wake
3920          * everybody up to tell them the bad news.
3921          */
3922         spin_lock(&log->l_icloglock);
3923         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3924         if (mp->m_sb_bp)
3925                 XFS_BUF_DONE(mp->m_sb_bp);
3926
3927         /*
3928          * This flag is sort of redundant because of the mount flag, but
3929          * it's good to maintain the separation between the log and the rest
3930          * of XFS.
3931          */
3932         log->l_flags |= XLOG_IO_ERROR;
3933
3934         /*
3935          * If we hit a log error, we want to mark all the iclogs IOERROR
3936          * while we're still holding the loglock.
3937          */
3938         if (logerror)
3939                 retval = xlog_state_ioerror(log);
3940         spin_unlock(&log->l_icloglock);
3941
3942         /*
3943          * We don't want anybody waiting for log reservations after this. That
3944          * means we have to wake up everybody queued up on reserveq as well as
3945          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3946          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3947          * action is protected by the grant locks.
3948          */
3949         xlog_grant_head_wake_all(&log->l_reserve_head);
3950         xlog_grant_head_wake_all(&log->l_write_head);
3951
3952         if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3953                 ASSERT(!logerror);
3954                 /*
3955                  * Force the incore logs to disk before shutting the
3956                  * log down completely.
3957                  */
3958                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3959
3960                 spin_lock(&log->l_icloglock);
3961                 retval = xlog_state_ioerror(log);
3962                 spin_unlock(&log->l_icloglock);
3963         }
3964
3965         /*
3966          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3967          * as if the log writes were completed. The abort handling in the log
3968          * item committed callback functions will do this again under lock to
3969          * avoid races.
3970          */
3971         wake_up_all(&log->l_cilp->xc_commit_wait);
3972         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3973
3974 #ifdef XFSERRORDEBUG
3975         {
3976                 xlog_in_core_t  *iclog;
3977
3978                 spin_lock(&log->l_icloglock);
3979                 iclog = log->l_iclog;
3980                 do {
3981                         ASSERT(iclog->ic_callback == 0);
3982                         iclog = iclog->ic_next;
3983                 } while (iclog != log->l_iclog);
3984                 spin_unlock(&log->l_icloglock);
3985         }
3986 #endif
3987         /* return non-zero if log IOERROR transition had already happened */
3988         return retval;
3989 }
3990
3991 STATIC int
3992 xlog_iclogs_empty(
3993         struct xlog     *log)
3994 {
3995         xlog_in_core_t  *iclog;
3996
3997         iclog = log->l_iclog;
3998         do {
3999                 /* endianness does not matter here, zero is zero in
4000                  * any language.
4001                  */
4002                 if (iclog->ic_header.h_num_logops)
4003                         return 0;
4004                 iclog = iclog->ic_next;
4005         } while (iclog != log->l_iclog);
4006         return 1;
4007 }
4008