fs/xfs/xfs_log.c

   1 /*
   2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_shared.h"
  21 #include "xfs_format.h"
  22 #include "xfs_log_format.h"
  23 #include "xfs_trans_resv.h"
  24 #include "xfs_mount.h"
  25 #include "xfs_error.h"
  26 #include "xfs_trans.h"
  27 #include "xfs_trans_priv.h"
  28 #include "xfs_log.h"
  29 #include "xfs_log_priv.h"
  30 #include "xfs_log_recover.h"
  31 #include "xfs_inode.h"
  32 #include "xfs_trace.h"
  33 #include "xfs_fsops.h"
  34 #include "xfs_cksum.h"
  35 #include "xfs_sysfs.h"
  36 #include "xfs_sb.h"
  37
  38 kmem_zone_t     *xfs_log_ticket_zone;
  39
  40 /* Local miscellaneous function prototypes */
  41 STATIC int
  42 xlog_commit_record(
  43         struct xlog             *log,
  44         struct xlog_ticket      *ticket,
  45         struct xlog_in_core     **iclog,
  46         xfs_lsn_t               *commitlsnp);
  47
  48 STATIC struct xlog *
  49 xlog_alloc_log(
  50         struct xfs_mount        *mp,
  51         struct xfs_buftarg      *log_target,
  52         xfs_daddr_t             blk_offset,
  53         int                     num_bblks);
  54 STATIC int
  55 xlog_space_left(
  56         struct xlog             *log,
  57         atomic64_t              *head);
  58 STATIC int
  59 xlog_sync(
  60         struct xlog             *log,
  61         struct xlog_in_core     *iclog);
  62 STATIC void
  63 xlog_dealloc_log(
  64         struct xlog             *log);
  65
  66 /* local state machine functions */
  67 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  68 STATIC void
  69 xlog_state_do_callback(
  70         struct xlog             *log,
  71         int                     aborted,
  72         struct xlog_in_core     *iclog);
  73 STATIC int
  74 xlog_state_get_iclog_space(
  75         struct xlog             *log,
  76         int                     len,
  77         struct xlog_in_core     **iclog,
  78         struct xlog_ticket      *ticket,
  79         int                     *continued_write,
  80         int                     *logoffsetp);
  81 STATIC int
  82 xlog_state_release_iclog(
  83         struct xlog             *log,
  84         struct xlog_in_core     *iclog);
  85 STATIC void
  86 xlog_state_switch_iclogs(
  87         struct xlog             *log,
  88         struct xlog_in_core     *iclog,
  89         int                     eventual_size);
  90 STATIC void
  91 xlog_state_want_sync(
  92         struct xlog             *log,
  93         struct xlog_in_core     *iclog);
  94
  95 STATIC void
  96 xlog_grant_push_ail(
  97         struct xlog             *log,
  98         int                     need_bytes);
  99 STATIC void
 100 xlog_regrant_reserve_log_space(
 101         struct xlog             *log,
 102         struct xlog_ticket      *ticket);
 103 STATIC void
 104 xlog_ungrant_log_space(
 105         struct xlog             *log,
 106         struct xlog_ticket      *ticket);
 107
 108 #if defined(DEBUG)
 109 STATIC void
 110 xlog_verify_dest_ptr(
 111         struct xlog             *log,
 112         void                    *ptr);
 113 STATIC void
 114 xlog_verify_grant_tail(
 115         struct xlog *log);
 116 STATIC void
 117 xlog_verify_iclog(
 118         struct xlog             *log,
 119         struct xlog_in_core     *iclog,
 120         int                     count,
 121         bool                    syncing);
 122 STATIC void
 123 xlog_verify_tail_lsn(
 124         struct xlog             *log,
 125         struct xlog_in_core     *iclog,
 126         xfs_lsn_t               tail_lsn);
 127 #else
 128 #define xlog_verify_dest_ptr(a,b)
 129 #define xlog_verify_grant_tail(a)
 130 #define xlog_verify_iclog(a,b,c,d)
 131 #define xlog_verify_tail_lsn(a,b,c)
 132 #endif
 133
 134 STATIC int
 135 xlog_iclogs_empty(
 136         struct xlog             *log);
 137
 138 static void
 139 xlog_grant_sub_space(
 140         struct xlog             *log,
 141         atomic64_t              *head,
 142         int                     bytes)
 143 {
 144         int64_t head_val = atomic64_read(head);
 145         int64_t new, old;
 146
 147         do {
 148                 int     cycle, space;
 149
 150                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 151
 152                 space -= bytes;
 153                 if (space < 0) {
 154                         space += log->l_logsize;
 155                         cycle--;
 156                 }
 157
 158                 old = head_val;
 159                 new = xlog_assign_grant_head_val(cycle, space);
 160                 head_val = atomic64_cmpxchg(head, old, new);
 161         } while (head_val != old);
 162 }
 163
 164 static void
 165 xlog_grant_add_space(
 166         struct xlog             *log,
 167         atomic64_t              *head,
 168         int                     bytes)
 169 {
 170         int64_t head_val = atomic64_read(head);
 171         int64_t new, old;
 172
 173         do {
 174                 int             tmp;
 175                 int             cycle, space;
 176
 177                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 178
 179                 tmp = log->l_logsize - space;
 180                 if (tmp > bytes)
 181                         space += bytes;
 182                 else {
 183                         space = bytes - tmp;
 184                         cycle++;
 185                 }
 186
 187                 old = head_val;
 188                 new = xlog_assign_grant_head_val(cycle, space);
 189                 head_val = atomic64_cmpxchg(head, old, new);
 190         } while (head_val != old);
 191 }
 192
 193 STATIC void
 194 xlog_grant_head_init(
 195         struct xlog_grant_head  *head)
 196 {
 197         xlog_assign_grant_head(&head->grant, 1, 0);
 198         INIT_LIST_HEAD(&head->waiters);
 199         spin_lock_init(&head->lock);
 200 }
 201
 202 STATIC void
 203 xlog_grant_head_wake_all(
 204         struct xlog_grant_head  *head)
 205 {
 206         struct xlog_ticket      *tic;
 207
 208         spin_lock(&head->lock);
 209         list_for_each_entry(tic, &head->waiters, t_queue)
 210                 wake_up_process(tic->t_task);
 211         spin_unlock(&head->lock);
 212 }
 213
 214 static inline int
 215 xlog_ticket_reservation(
 216         struct xlog             *log,
 217         struct xlog_grant_head  *head,
 218         struct xlog_ticket      *tic)
 219 {
 220         if (head == &log->l_write_head) {
 221                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 222                 return tic->t_unit_res;
 223         } else {
 224                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 225                         return tic->t_unit_res * tic->t_cnt;
 226                 else
 227                         return tic->t_unit_res;
 228         }
 229 }
 230
 231 STATIC bool
 232 xlog_grant_head_wake(
 233         struct xlog             *log,
 234         struct xlog_grant_head  *head,
 235         int                     *free_bytes)
 236 {
 237         struct xlog_ticket      *tic;
 238         int                     need_bytes;
 239
 240         list_for_each_entry(tic, &head->waiters, t_queue) {
 241                 need_bytes = xlog_ticket_reservation(log, head, tic);
 242                 if (*free_bytes < need_bytes)
 243                         return false;
 244
 245                 *free_bytes -= need_bytes;
 246                 trace_xfs_log_grant_wake_up(log, tic);
 247                 wake_up_process(tic->t_task);
 248         }
 249
 250         return true;
 251 }
 252
 253 STATIC int
 254 xlog_grant_head_wait(
 255         struct xlog             *log,
 256         struct xlog_grant_head  *head,
 257         struct xlog_ticket      *tic,
 258         int                     need_bytes) __releases(&head->lock)
 259                                             __acquires(&head->lock)
 260 {
 261         list_add_tail(&tic->t_queue, &head->waiters);
 262
 263         do {
 264                 if (XLOG_FORCED_SHUTDOWN(log))
 265                         goto shutdown;
 266                 xlog_grant_push_ail(log, need_bytes);
 267
 268                 __set_current_state(TASK_UNINTERRUPTIBLE);
 269                 spin_unlock(&head->lock);
 270
 271                 XFS_STATS_INC(xs_sleep_logspace);
 272
 273                 trace_xfs_log_grant_sleep(log, tic);
 274                 schedule();
 275                 trace_xfs_log_grant_wake(log, tic);
 276
 277                 spin_lock(&head->lock);
 278                 if (XLOG_FORCED_SHUTDOWN(log))
 279                         goto shutdown;
 280         } while (xlog_space_left(log, &head->grant) < need_bytes);
 281
 282         list_del_init(&tic->t_queue);
 283         return 0;
 284 shutdown:
 285         list_del_init(&tic->t_queue);
 286         return -EIO;
 287 }
 288
 289 /*
 290  * Atomically get the log space required for a log ticket.
 291  *
 292  * Once a ticket gets put onto head->waiters, it will only return after the
 293  * needed reservation is satisfied.
 294  *
 295  * This function is structured so that it has a lock free fast path. This is
 296  * necessary because every new transaction reservation will come through this
 297  * path. Hence any lock will be globally hot if we take it unconditionally on
 298  * every pass.
 299  *
 300  * As tickets are only ever moved on and off head->waiters under head->lock, we
 301  * only need to take that lock if we are going to add the ticket to the queue
 302  * and sleep. We can avoid taking the lock if the ticket was never added to
 303  * head->waiters because the t_queue list head will be empty and we hold the
 304  * only reference to it so it can safely be checked unlocked.
 305  */
 306 STATIC int
 307 xlog_grant_head_check(
 308         struct xlog             *log,
 309         struct xlog_grant_head  *head,
 310         struct xlog_ticket      *tic,
 311         int                     *need_bytes)
 312 {
 313         int                     free_bytes;
 314         int                     error = 0;
 315
 316         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 317
 318         /*
 319          * If there are other waiters on the queue then give them a chance at
 320          * logspace before us.  Wake up the first waiters, if we do not wake
 321          * up all the waiters then go to sleep waiting for more free space,
 322          * otherwise try to get some space for this transaction.
 323          */
 324         *need_bytes = xlog_ticket_reservation(log, head, tic);
 325         free_bytes = xlog_space_left(log, &head->grant);
 326         if (!list_empty_careful(&head->waiters)) {
 327                 spin_lock(&head->lock);
 328                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 329                     free_bytes < *need_bytes) {
 330                         error = xlog_grant_head_wait(log, head, tic,
 331                                                      *need_bytes);
 332                 }
 333                 spin_unlock(&head->lock);
 334         } else if (free_bytes < *need_bytes) {
 335                 spin_lock(&head->lock);
 336                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 337                 spin_unlock(&head->lock);
 338         }
 339
 340         return error;
 341 }
 342
 343 static void
 344 xlog_tic_reset_res(xlog_ticket_t *tic)
 345 {
 346         tic->t_res_num = 0;
 347         tic->t_res_arr_sum = 0;
 348         tic->t_res_num_ophdrs = 0;
 349 }
 350
 351 static void
 352 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 353 {
 354         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 355                 /* add to overflow and start again */
 356                 tic->t_res_o_flow += tic->t_res_arr_sum;
 357                 tic->t_res_num = 0;
 358                 tic->t_res_arr_sum = 0;
 359         }
 360
 361         tic->t_res_arr[tic->t_res_num].r_len = len;
 362         tic->t_res_arr[tic->t_res_num].r_type = type;
 363         tic->t_res_arr_sum += len;
 364         tic->t_res_num++;
 365 }
 366
 367 /*
 368  * Replenish the byte reservation required by moving the grant write head.
 369  */
 370 int
 371 xfs_log_regrant(
 372         struct xfs_mount        *mp,
 373         struct xlog_ticket      *tic)
 374 {
 375         struct xlog             *log = mp->m_log;
 376         int                     need_bytes;
 377         int                     error = 0;
 378
 379         if (XLOG_FORCED_SHUTDOWN(log))
 380                 return -EIO;
 381
 382         XFS_STATS_INC(xs_try_logspace);
 383
 384         /*
 385          * This is a new transaction on the ticket, so we need to change the
 386          * transaction ID so that the next transaction has a different TID in
 387          * the log. Just add one to the existing tid so that we can see chains
 388          * of rolling transactions in the log easily.
 389          */
 390         tic->t_tid++;
 391
 392         xlog_grant_push_ail(log, tic->t_unit_res);
 393
 394         tic->t_curr_res = tic->t_unit_res;
 395         xlog_tic_reset_res(tic);
 396
 397         if (tic->t_cnt > 0)
 398                 return 0;
 399
 400         trace_xfs_log_regrant(log, tic);
 401
 402         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 403                                       &need_bytes);
 404         if (error)
 405                 goto out_error;
 406
 407         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 408         trace_xfs_log_regrant_exit(log, tic);
 409         xlog_verify_grant_tail(log);
 410         return 0;
 411
 412 out_error:
 413         /*
 414          * If we are failing, make sure the ticket doesn't have any current
 415          * reservations.  We don't want to add this back when the ticket/
 416          * transaction gets cancelled.
 417          */
 418         tic->t_curr_res = 0;
 419         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 420         return error;
 421 }
 422
 423 /*
 424  * Reserve log space and return a ticket corresponding the reservation.
 425  *
 426  * Each reservation is going to reserve extra space for a log record header.
 427  * When writes happen to the on-disk log, we don't subtract the length of the
 428  * log record header from any reservation.  By wasting space in each
 429  * reservation, we prevent over allocation problems.
 430  */
 431 int
 432 xfs_log_reserve(
 433         struct xfs_mount        *mp,
 434         int                     unit_bytes,
 435         int                     cnt,
 436         struct xlog_ticket      **ticp,
 437         __uint8_t               client,
 438         bool                    permanent,
 439         uint                    t_type)
 440 {
 441         struct xlog             *log = mp->m_log;
 442         struct xlog_ticket      *tic;
 443         int                     need_bytes;
 444         int                     error = 0;
 445
 446         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 447
 448         if (XLOG_FORCED_SHUTDOWN(log))
 449                 return -EIO;
 450
 451         XFS_STATS_INC(xs_try_logspace);
 452
 453         ASSERT(*ticp == NULL);
 454         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 455                                 KM_SLEEP | KM_MAYFAIL);
 456         if (!tic)
 457                 return -ENOMEM;
 458
 459         tic->t_trans_type = t_type;
 460         *ticp = tic;
 461
 462         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 463                                             : tic->t_unit_res);
 464
 465         trace_xfs_log_reserve(log, tic);
 466
 467         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 468                                       &need_bytes);
 469         if (error)
 470                 goto out_error;
 471
 472         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 473         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 474         trace_xfs_log_reserve_exit(log, tic);
 475         xlog_verify_grant_tail(log);
 476         return 0;
 477
 478 out_error:
 479         /*
 480          * If we are failing, make sure the ticket doesn't have any current
 481          * reservations.  We don't want to add this back when the ticket/
 482          * transaction gets cancelled.
 483          */
 484         tic->t_curr_res = 0;
 485         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 486         return error;
 487 }
 488
 489
 490 /*
 491  * NOTES:
 492  *
 493  *      1. currblock field gets updated at startup and after in-core logs
 494  *              marked as with WANT_SYNC.
 495  */
 496
 497 /*
 498  * This routine is called when a user of a log manager ticket is done with
 499  * the reservation.  If the ticket was ever used, then a commit record for
 500  * the associated transaction is written out as a log operation header with
 501  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 502  * a given ticket.  If the ticket was one with a permanent reservation, then
 503  * a few operations are done differently.  Permanent reservation tickets by
 504  * default don't release the reservation.  They just commit the current
 505  * transaction with the belief that the reservation is still needed.  A flag
 506  * must be passed in before permanent reservations are actually released.
 507  * When these type of tickets are not released, they need to be set into
 508  * the inited state again.  By doing this, a start record will be written
 509  * out when the next write occurs.
 510  */
 511 xfs_lsn_t
 512 xfs_log_done(
 513         struct xfs_mount        *mp,
 514         struct xlog_ticket      *ticket,
 515         struct xlog_in_core     **iclog,
 516         bool                    regrant)
 517 {
 518         struct xlog             *log = mp->m_log;
 519         xfs_lsn_t               lsn = 0;
 520
 521         if (XLOG_FORCED_SHUTDOWN(log) ||
 522             /*
 523              * If nothing was ever written, don't write out commit record.
 524              * If we get an error, just continue and give back the log ticket.
 525              */
 526             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 527              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 528                 lsn = (xfs_lsn_t) -1;
 529                 regrant = false;
 530         }
 531
 532
 533         if (!regrant) {
 534                 trace_xfs_log_done_nonperm(log, ticket);
 535
 536                 /*
 537                  * Release ticket if not permanent reservation or a specific
 538                  * request has been made to release a permanent reservation.
 539                  */
 540                 xlog_ungrant_log_space(log, ticket);
 541         } else {
 542                 trace_xfs_log_done_perm(log, ticket);
 543
 544                 xlog_regrant_reserve_log_space(log, ticket);
 545                 /* If this ticket was a permanent reservation and we aren't
 546                  * trying to release it, reset the inited flags; so next time
 547                  * we write, a start record will be written out.
 548                  */
 549                 ticket->t_flags |= XLOG_TIC_INITED;
 550         }
 551
 552         xfs_log_ticket_put(ticket);
 553         return lsn;
 554 }
 555
 556 /*
 557  * Attaches a new iclog I/O completion callback routine during
 558  * transaction commit.  If the log is in error state, a non-zero
 559  * return code is handed back and the caller is responsible for
 560  * executing the callback at an appropriate time.
 561  */
 562 int
 563 xfs_log_notify(
 564         struct xfs_mount        *mp,
 565         struct xlog_in_core     *iclog,
 566         xfs_log_callback_t      *cb)
 567 {
 568         int     abortflg;
 569
 570         spin_lock(&iclog->ic_callback_lock);
 571         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 572         if (!abortflg) {
 573                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 574                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 575                 cb->cb_next = NULL;
 576                 *(iclog->ic_callback_tail) = cb;
 577                 iclog->ic_callback_tail = &(cb->cb_next);
 578         }
 579         spin_unlock(&iclog->ic_callback_lock);
 580         return abortflg;
 581 }
 582
 583 int
 584 xfs_log_release_iclog(
 585         struct xfs_mount        *mp,
 586         struct xlog_in_core     *iclog)
 587 {
 588         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 589                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 590                 return -EIO;
 591         }
 592
 593         return 0;
 594 }
 595
 596 /*
 597  * Mount a log filesystem
 598  *
 599  * mp           - ubiquitous xfs mount point structure
 600  * log_target   - buftarg of on-disk log device
 601  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 602  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 603  *
 604  * Return error or zero.
 605  */
 606 int
 607 xfs_log_mount(
 608         xfs_mount_t     *mp,
 609         xfs_buftarg_t   *log_target,
 610         xfs_daddr_t     blk_offset,
 611         int             num_bblks)
 612 {
 613         int             error = 0;
 614         int             min_logfsbs;
 615
 616         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 617                 xfs_notice(mp, "Mounting V%d Filesystem",
 618                            XFS_SB_VERSION_NUM(&mp->m_sb));
 619         } else {
 620                 xfs_notice(mp,
 621 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
 622                            XFS_SB_VERSION_NUM(&mp->m_sb));
 623                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 624         }
 625
 626         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 627         if (IS_ERR(mp->m_log)) {
 628                 error = PTR_ERR(mp->m_log);
 629                 goto out;
 630         }
 631
 632         /*
 633          * Validate the given log space and drop a critical message via syslog
 634          * if the log size is too small that would lead to some unexpected
 635          * situations in transaction log space reservation stage.
 636          *
 637          * Note: we can't just reject the mount if the validation fails.  This
 638          * would mean that people would have to downgrade their kernel just to
 639          * remedy the situation as there is no way to grow the log (short of
 640          * black magic surgery with xfs_db).
 641          *
 642          * We can, however, reject mounts for CRC format filesystems, as the
 643          * mkfs binary being used to make the filesystem should never create a
 644          * filesystem with a log that is too small.
 645          */
 646         min_logfsbs = xfs_log_calc_minimum_size(mp);
 647
 648         if (mp->m_sb.sb_logblocks < min_logfsbs) {
 649                 xfs_warn(mp,
 650                 "Log size %d blocks too small, minimum size is %d blocks",
 651                          mp->m_sb.sb_logblocks, min_logfsbs);
 652                 error = -EINVAL;
 653         } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
 654                 xfs_warn(mp,
 655                 "Log size %d blocks too large, maximum size is %lld blocks",
 656                          mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
 657                 error = -EINVAL;
 658         } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
 659                 xfs_warn(mp,
 660                 "log size %lld bytes too large, maximum size is %lld bytes",
 661                          XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
 662                          XFS_MAX_LOG_BYTES);
 663                 error = -EINVAL;
 664         }
 665         if (error) {
 666                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
 667                         xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 668                         ASSERT(0);
 669                         goto out_free_log;
 670                 }
 671                 xfs_crit(mp,
 672 "Log size out of supported range. Continuing onwards, but if log hangs are\n"
 673 "experienced then please report this message in the bug report.");
 674         }
 675
 676         /*
 677          * Initialize the AIL now we have a log.
 678          */
 679         error = xfs_trans_ail_init(mp);
 680         if (error) {
 681                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 682                 goto out_free_log;
 683         }
 684         mp->m_log->l_ailp = mp->m_ail;
 685
 686         /*
 687          * skip log recovery on a norecovery mount.  pretend it all
 688          * just worked.
 689          */
 690         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 691                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 692
 693                 if (readonly)
 694                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 695
 696                 error = xlog_recover(mp->m_log);
 697
 698                 if (readonly)
 699                         mp->m_flags |= XFS_MOUNT_RDONLY;
 700                 if (error) {
 701                         xfs_warn(mp, "log mount/recovery failed: error %d",
 702                                 error);
 703                         goto out_destroy_ail;
 704                 }
 705         }
 706
 707         error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 708                                "log");
 709         if (error)
 710                 goto out_destroy_ail;
 711
 712         /* Normal transactions can now occur */
 713         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 714
 715         /*
 716          * Now the log has been fully initialised and we know were our
 717          * space grant counters are, we can initialise the permanent ticket
 718          * needed for delayed logging to work.
 719          */
 720         xlog_cil_init_post_recovery(mp->m_log);
 721
 722         return 0;
 723
 724 out_destroy_ail:
 725         xfs_trans_ail_destroy(mp);
 726 out_free_log:
 727         xlog_dealloc_log(mp->m_log);
 728 out:
 729         return error;
 730 }
 731
 732 /*
 733  * Finish the recovery of the file system.  This is separate from the
 734  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 735  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 736  * here.
 737  *
 738  * If we finish recovery successfully, start the background log work. If we are
 739  * not doing recovery, then we have a RO filesystem and we don't need to start
 740  * it.
 741  */
 742 int
 743 xfs_log_mount_finish(xfs_mount_t *mp)
 744 {
 745         int     error = 0;
 746
 747         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 748                 error = xlog_recover_finish(mp->m_log);
 749                 if (!error)
 750                         xfs_log_work_queue(mp);
 751         } else {
 752                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 753         }
 754
 755
 756         return error;
 757 }
 758
 759 /*
 760  * Final log writes as part of unmount.
 761  *
 762  * Mark the filesystem clean as unmount happens.  Note that during relocation
 763  * this routine needs to be executed as part of source-bag while the
 764  * deallocation must not be done until source-end.
 765  */
 766
 767 /*
 768  * Unmount record used to have a string "Unmount filesystem--" in the
 769  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 770  * We just write the magic number now since that particular field isn't
 771  * currently architecture converted and "Unmount" is a bit foo.
 772  * As far as I know, there weren't any dependencies on the old behaviour.
 773  */
 774
 775 int
 776 xfs_log_unmount_write(xfs_mount_t *mp)
 777 {
 778         struct xlog      *log = mp->m_log;
 779         xlog_in_core_t   *iclog;
 780 #ifdef DEBUG
 781         xlog_in_core_t   *first_iclog;
 782 #endif
 783         xlog_ticket_t   *tic = NULL;
 784         xfs_lsn_t        lsn;
 785         int              error;
 786
 787         /*
 788          * Don't write out unmount record on read-only mounts.
 789          * Or, if we are doing a forced umount (typically because of IO errors).
 790          */
 791         if (mp->m_flags & XFS_MOUNT_RDONLY)
 792                 return 0;
 793
 794         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
 795         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 796
 797 #ifdef DEBUG
 798         first_iclog = iclog = log->l_iclog;
 799         do {
 800                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 801                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 802                         ASSERT(iclog->ic_offset == 0);
 803                 }
 804                 iclog = iclog->ic_next;
 805         } while (iclog != first_iclog);
 806 #endif
 807         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 808                 error = xfs_log_reserve(mp, 600, 1, &tic,
 809                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
 810                 if (!error) {
 811                         /* the data section must be 32 bit size aligned */
 812                         struct {
 813                             __uint16_t magic;
 814                             __uint16_t pad1;
 815                             __uint32_t pad2; /* may as well make it 64 bits */
 816                         } magic = {
 817                                 .magic = XLOG_UNMOUNT_TYPE,
 818                         };
 819                         struct xfs_log_iovec reg = {
 820                                 .i_addr = &magic,
 821                                 .i_len = sizeof(magic),
 822                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 823                         };
 824                         struct xfs_log_vec vec = {
 825                                 .lv_niovecs = 1,
 826                                 .lv_iovecp = &reg,
 827                         };
 828
 829                         /* remove inited flag, and account for space used */
 830                         tic->t_flags = 0;
 831                         tic->t_curr_res -= sizeof(magic);
 832                         error = xlog_write(log, &vec, tic, &lsn,
 833                                            NULL, XLOG_UNMOUNT_TRANS);
 834                         /*
 835                          * At this point, we're umounting anyway,
 836                          * so there's no point in transitioning log state
 837                          * to IOERROR. Just continue...
 838                          */
 839                 }
 840
 841                 if (error)
 842                         xfs_alert(mp, "%s: unmount record failed", __func__);
 843
 844
 845                 spin_lock(&log->l_icloglock);
 846                 iclog = log->l_iclog;
 847                 atomic_inc(&iclog->ic_refcnt);
 848                 xlog_state_want_sync(log, iclog);
 849                 spin_unlock(&log->l_icloglock);
 850                 error = xlog_state_release_iclog(log, iclog);
 851
 852                 spin_lock(&log->l_icloglock);
 853                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 854                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 855                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 856                                 xlog_wait(&iclog->ic_force_wait,
 857                                                         &log->l_icloglock);
 858                         } else {
 859                                 spin_unlock(&log->l_icloglock);
 860                         }
 861                 } else {
 862                         spin_unlock(&log->l_icloglock);
 863                 }
 864                 if (tic) {
 865                         trace_xfs_log_umount_write(log, tic);
 866                         xlog_ungrant_log_space(log, tic);
 867                         xfs_log_ticket_put(tic);
 868                 }
 869         } else {
 870                 /*
 871                  * We're already in forced_shutdown mode, couldn't
 872                  * even attempt to write out the unmount transaction.
 873                  *
 874                  * Go through the motions of sync'ing and releasing
 875                  * the iclog, even though no I/O will actually happen,
 876                  * we need to wait for other log I/Os that may already
 877                  * be in progress.  Do this as a separate section of
 878                  * code so we'll know if we ever get stuck here that
 879                  * we're in this odd situation of trying to unmount
 880                  * a file system that went into forced_shutdown as
 881                  * the result of an unmount..
 882                  */
 883                 spin_lock(&log->l_icloglock);
 884                 iclog = log->l_iclog;
 885                 atomic_inc(&iclog->ic_refcnt);
 886
 887                 xlog_state_want_sync(log, iclog);
 888                 spin_unlock(&log->l_icloglock);
 889                 error =  xlog_state_release_iclog(log, iclog);
 890
 891                 spin_lock(&log->l_icloglock);
 892
 893                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 894                         || iclog->ic_state == XLOG_STATE_DIRTY
 895                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 896
 897                                 xlog_wait(&iclog->ic_force_wait,
 898                                                         &log->l_icloglock);
 899                 } else {
 900                         spin_unlock(&log->l_icloglock);
 901                 }
 902         }
 903
 904         return error;
 905 }       /* xfs_log_unmount_write */
 906
 907 /*
 908  * Empty the log for unmount/freeze.
 909  *
 910  * To do this, we first need to shut down the background log work so it is not
 911  * trying to cover the log as we clean up. We then need to unpin all objects in
 912  * the log so we can then flush them out. Once they have completed their IO and
 913  * run the callbacks removing themselves from the AIL, we can write the unmount
 914  * record.
 915  */
 916 void
 917 xfs_log_quiesce(
 918         struct xfs_mount        *mp)
 919 {
 920         cancel_delayed_work_sync(&mp->m_log->l_work);
 921         xfs_log_force(mp, XFS_LOG_SYNC);
 922
 923         /*
 924          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
 925          * will push it, xfs_wait_buftarg() will not wait for it. Further,
 926          * xfs_buf_iowait() cannot be used because it was pushed with the
 927          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
 928          * the IO to complete.
 929          */
 930         xfs_ail_push_all_sync(mp->m_ail);
 931         xfs_wait_buftarg(mp->m_ddev_targp);
 932         xfs_buf_lock(mp->m_sb_bp);
 933         xfs_buf_unlock(mp->m_sb_bp);
 934
 935         xfs_log_unmount_write(mp);
 936 }
 937
 938 /*
 939  * Shut down and release the AIL and Log.
 940  *
 941  * During unmount, we need to ensure we flush all the dirty metadata objects
 942  * from the AIL so that the log is empty before we write the unmount record to
 943  * the log. Once this is done, we can tear down the AIL and the log.
 944  */
 945 void
 946 xfs_log_unmount(
 947         struct xfs_mount        *mp)
 948 {
 949         xfs_log_quiesce(mp);
 950
 951         xfs_trans_ail_destroy(mp);
 952
 953         xfs_sysfs_del(&mp->m_log->l_kobj);
 954
 955         xlog_dealloc_log(mp->m_log);
 956 }
 957
 958 void
 959 xfs_log_item_init(
 960         struct xfs_mount        *mp,
 961         struct xfs_log_item     *item,
 962         int                     type,
 963         const struct xfs_item_ops *ops)
 964 {
 965         item->li_mountp = mp;
 966         item->li_ailp = mp->m_ail;
 967         item->li_type = type;
 968         item->li_ops = ops;
 969         item->li_lv = NULL;
 970
 971         INIT_LIST_HEAD(&item->li_ail);
 972         INIT_LIST_HEAD(&item->li_cil);
 973 }
 974
 975 /*
 976  * Wake up processes waiting for log space after we have moved the log tail.
 977  */
 978 void
 979 xfs_log_space_wake(
 980         struct xfs_mount        *mp)
 981 {
 982         struct xlog             *log = mp->m_log;
 983         int                     free_bytes;
 984
 985         if (XLOG_FORCED_SHUTDOWN(log))
 986                 return;
 987
 988         if (!list_empty_careful(&log->l_write_head.waiters)) {
 989                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 990
 991                 spin_lock(&log->l_write_head.lock);
 992                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
 993                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
 994                 spin_unlock(&log->l_write_head.lock);
 995         }
 996
 997         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
 998                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 999
1000                 spin_lock(&log->l_reserve_head.lock);
1001                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1002                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1003                 spin_unlock(&log->l_reserve_head.lock);
1004         }
1005 }
1006
1007 /*
1008  * Determine if we have a transaction that has gone to disk that needs to be
1009  * covered. To begin the transition to the idle state firstly the log needs to
1010  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1011  * we start attempting to cover the log.
1012  *
1013  * Only if we are then in a state where covering is needed, the caller is
1014  * informed that dummy transactions are required to move the log into the idle
1015  * state.
1016  *
1017  * If there are any items in the AIl or CIL, then we do not want to attempt to
1018  * cover the log as we may be in a situation where there isn't log space
1019  * available to run a dummy transaction and this can lead to deadlocks when the
1020  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1021  * there's no point in running a dummy transaction at this point because we
1022  * can't start trying to idle the log until both the CIL and AIL are empty.
1023  */
1024 int
1025 xfs_log_need_covered(xfs_mount_t *mp)
1026 {
1027         struct xlog     *log = mp->m_log;
1028         int             needed = 0;
1029
1030         if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1031                 return 0;
1032
1033         if (!xlog_cil_empty(log))
1034                 return 0;
1035
1036         spin_lock(&log->l_icloglock);
1037         switch (log->l_covered_state) {
1038         case XLOG_STATE_COVER_DONE:
1039         case XLOG_STATE_COVER_DONE2:
1040         case XLOG_STATE_COVER_IDLE:
1041                 break;
1042         case XLOG_STATE_COVER_NEED:
1043         case XLOG_STATE_COVER_NEED2:
1044                 if (xfs_ail_min_lsn(log->l_ailp))
1045                         break;
1046                 if (!xlog_iclogs_empty(log))
1047                         break;
1048
1049                 needed = 1;
1050                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1051                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1052                 else
1053                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1054                 break;
1055         default:
1056                 needed = 1;
1057                 break;
1058         }
1059         spin_unlock(&log->l_icloglock);
1060         return needed;
1061 }
1062
1063 /*
1064  * We may be holding the log iclog lock upon entering this routine.
1065  */
1066 xfs_lsn_t
1067 xlog_assign_tail_lsn_locked(
1068         struct xfs_mount        *mp)
1069 {
1070         struct xlog             *log = mp->m_log;
1071         struct xfs_log_item     *lip;
1072         xfs_lsn_t               tail_lsn;
1073
1074         assert_spin_locked(&mp->m_ail->xa_lock);
1075
1076         /*
1077          * To make sure we always have a valid LSN for the log tail we keep
1078          * track of the last LSN which was committed in log->l_last_sync_lsn,
1079          * and use that when the AIL was empty.
1080          */
1081         lip = xfs_ail_min(mp->m_ail);
1082         if (lip)
1083                 tail_lsn = lip->li_lsn;
1084         else
1085                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1086         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1087         atomic64_set(&log->l_tail_lsn, tail_lsn);
1088         return tail_lsn;
1089 }
1090
1091 xfs_lsn_t
1092 xlog_assign_tail_lsn(
1093         struct xfs_mount        *mp)
1094 {
1095         xfs_lsn_t               tail_lsn;
1096
1097         spin_lock(&mp->m_ail->xa_lock);
1098         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1099         spin_unlock(&mp->m_ail->xa_lock);
1100
1101         return tail_lsn;
1102 }
1103
1104 /*
1105  * Return the space in the log between the tail and the head.  The head
1106  * is passed in the cycle/bytes formal parms.  In the special case where
1107  * the reserve head has wrapped passed the tail, this calculation is no
1108  * longer valid.  In this case, just return 0 which means there is no space
1109  * in the log.  This works for all places where this function is called
1110  * with the reserve head.  Of course, if the write head were to ever
1111  * wrap the tail, we should blow up.  Rather than catch this case here,
1112  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1113  *
1114  * This code also handles the case where the reservation head is behind
1115  * the tail.  The details of this case are described below, but the end
1116  * result is that we return the size of the log as the amount of space left.
1117  */
1118 STATIC int
1119 xlog_space_left(
1120         struct xlog     *log,
1121         atomic64_t      *head)
1122 {
1123         int             free_bytes;
1124         int             tail_bytes;
1125         int             tail_cycle;
1126         int             head_cycle;
1127         int             head_bytes;
1128
1129         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1130         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1131         tail_bytes = BBTOB(tail_bytes);
1132         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1133                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1134         else if (tail_cycle + 1 < head_cycle)
1135                 return 0;
1136         else if (tail_cycle < head_cycle) {
1137                 ASSERT(tail_cycle == (head_cycle - 1));
1138                 free_bytes = tail_bytes - head_bytes;
1139         } else {
1140                 /*
1141                  * The reservation head is behind the tail.
1142                  * In this case we just want to return the size of the
1143                  * log as the amount of space left.
1144                  */
1145                 xfs_alert(log->l_mp,
1146                         "xlog_space_left: head behind tail\n"
1147                         "  tail_cycle = %d, tail_bytes = %d\n"
1148                         "  GH   cycle = %d, GH   bytes = %d",
1149                         tail_cycle, tail_bytes, head_cycle, head_bytes);
1150                 ASSERT(0);
1151                 free_bytes = log->l_logsize;
1152         }
1153         return free_bytes;
1154 }
1155
1156
1157 /*
1158  * Log function which is called when an io completes.
1159  *
1160  * The log manager needs its own routine, in order to control what
1161  * happens with the buffer after the write completes.
1162  */
1163 void
1164 xlog_iodone(xfs_buf_t *bp)
1165 {
1166         struct xlog_in_core     *iclog = bp->b_fspriv;
1167         struct xlog             *l = iclog->ic_log;
1168         int                     aborted = 0;
1169
1170         /*
1171          * Race to shutdown the filesystem if we see an error.
1172          */
1173         if (XFS_TEST_ERROR(bp->b_error, l->l_mp,
1174                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1175                 xfs_buf_ioerror_alert(bp, __func__);
1176                 xfs_buf_stale(bp);
1177                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1178                 /*
1179                  * This flag will be propagated to the trans-committed
1180                  * callback routines to let them know that the log-commit
1181                  * didn't succeed.
1182                  */
1183                 aborted = XFS_LI_ABORTED;
1184         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1185                 aborted = XFS_LI_ABORTED;
1186         }
1187
1188         /* log I/O is always issued ASYNC */
1189         ASSERT(XFS_BUF_ISASYNC(bp));
1190         xlog_state_done_syncing(iclog, aborted);
1191
1192         /*
1193          * drop the buffer lock now that we are done. Nothing references
1194          * the buffer after this, so an unmount waiting on this lock can now
1195          * tear it down safely. As such, it is unsafe to reference the buffer
1196          * (bp) after the unlock as we could race with it being freed.
1197          */
1198         xfs_buf_unlock(bp);
1199 }
1200
1201 /*
1202  * Return size of each in-core log record buffer.
1203  *
1204  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1205  *
1206  * If the filesystem blocksize is too large, we may need to choose a
1207  * larger size since the directory code currently logs entire blocks.
1208  */
1209
1210 STATIC void
1211 xlog_get_iclog_buffer_size(
1212         struct xfs_mount        *mp,
1213         struct xlog             *log)
1214 {
1215         int size;
1216         int xhdrs;
1217
1218         if (mp->m_logbufs <= 0)
1219                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1220         else
1221                 log->l_iclog_bufs = mp->m_logbufs;
1222
1223         /*
1224          * Buffer size passed in from mount system call.
1225          */
1226         if (mp->m_logbsize > 0) {
1227                 size = log->l_iclog_size = mp->m_logbsize;
1228                 log->l_iclog_size_log = 0;
1229                 while (size != 1) {
1230                         log->l_iclog_size_log++;
1231                         size >>= 1;
1232                 }
1233
1234                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1235                         /* # headers = size / 32k
1236                          * one header holds cycles from 32k of data
1237                          */
1238
1239                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1240                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1241                                 xhdrs++;
1242                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1243                         log->l_iclog_heads = xhdrs;
1244                 } else {
1245                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1246                         log->l_iclog_hsize = BBSIZE;
1247                         log->l_iclog_heads = 1;
1248                 }
1249                 goto done;
1250         }
1251
1252         /* All machines use 32kB buffers by default. */
1253         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1254         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1255
1256         /* the default log size is 16k or 32k which is one header sector */
1257         log->l_iclog_hsize = BBSIZE;
1258         log->l_iclog_heads = 1;
1259
1260 done:
1261         /* are we being asked to make the sizes selected above visible? */
1262         if (mp->m_logbufs == 0)
1263                 mp->m_logbufs = log->l_iclog_bufs;
1264         if (mp->m_logbsize == 0)
1265                 mp->m_logbsize = log->l_iclog_size;
1266 }       /* xlog_get_iclog_buffer_size */
1267
1268
1269 void
1270 xfs_log_work_queue(
1271         struct xfs_mount        *mp)
1272 {
1273         queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1274                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1275 }
1276
1277 /*
1278  * Every sync period we need to unpin all items in the AIL and push them to
1279  * disk. If there is nothing dirty, then we might need to cover the log to
1280  * indicate that the filesystem is idle.
1281  */
1282 void
1283 xfs_log_worker(
1284         struct work_struct      *work)
1285 {
1286         struct xlog             *log = container_of(to_delayed_work(work),
1287                                                 struct xlog, l_work);
1288         struct xfs_mount        *mp = log->l_mp;
1289
1290         /* dgc: errors ignored - not fatal and nowhere to report them */
1291         if (xfs_log_need_covered(mp)) {
1292                 /*
1293                  * Dump a transaction into the log that contains no real change.
1294                  * This is needed to stamp the current tail LSN into the log
1295                  * during the covering operation.
1296                  *
1297                  * We cannot use an inode here for this - that will push dirty
1298                  * state back up into the VFS and then periodic inode flushing
1299                  * will prevent log covering from making progress. Hence we
1300                  * synchronously log the superblock instead to ensure the
1301                  * superblock is immediately unpinned and can be written back.
1302                  */
1303                 xfs_sync_sb(mp, true);
1304         } else
1305                 xfs_log_force(mp, 0);
1306
1307         /* start pushing all the metadata that is currently dirty */
1308         xfs_ail_push_all(mp->m_ail);
1309
1310         /* queue us up again */
1311         xfs_log_work_queue(mp);
1312 }
1313
1314 /*
1315  * This routine initializes some of the log structure for a given mount point.
1316  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1317  * some other stuff may be filled in too.
1318  */
1319 STATIC struct xlog *
1320 xlog_alloc_log(
1321         struct xfs_mount        *mp,
1322         struct xfs_buftarg      *log_target,
1323         xfs_daddr_t             blk_offset,
1324         int                     num_bblks)
1325 {
1326         struct xlog             *log;
1327         xlog_rec_header_t       *head;
1328         xlog_in_core_t          **iclogp;
1329         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1330         xfs_buf_t               *bp;
1331         int                     i;
1332         int                     error = -ENOMEM;
1333         uint                    log2_size = 0;
1334
1335         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1336         if (!log) {
1337                 xfs_warn(mp, "Log allocation failed: No memory!");
1338                 goto out;
1339         }
1340
1341         log->l_mp          = mp;
1342         log->l_targ        = log_target;
1343         log->l_logsize     = BBTOB(num_bblks);
1344         log->l_logBBstart  = blk_offset;
1345         log->l_logBBsize   = num_bblks;
1346         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1347         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1348         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1349
1350         log->l_prev_block  = -1;
1351         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1352         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1353         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1354         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1355
1356         xlog_grant_head_init(&log->l_reserve_head);
1357         xlog_grant_head_init(&log->l_write_head);
1358
1359         error = -EFSCORRUPTED;
1360         if (xfs_sb_version_hassector(&mp->m_sb)) {
1361                 log2_size = mp->m_sb.sb_logsectlog;
1362                 if (log2_size < BBSHIFT) {
1363                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1364                                 log2_size, BBSHIFT);
1365                         goto out_free_log;
1366                 }
1367
1368                 log2_size -= BBSHIFT;
1369                 if (log2_size > mp->m_sectbb_log) {
1370                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1371                                 log2_size, mp->m_sectbb_log);
1372                         goto out_free_log;
1373                 }
1374
1375                 /* for larger sector sizes, must have v2 or external log */
1376                 if (log2_size && log->l_logBBstart > 0 &&
1377                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1378                         xfs_warn(mp,
1379                 "log sector size (0x%x) invalid for configuration.",
1380                                 log2_size);
1381                         goto out_free_log;
1382                 }
1383         }
1384         log->l_sectBBsize = 1 << log2_size;
1385
1386         xlog_get_iclog_buffer_size(mp, log);
1387
1388         /*
1389          * Use a NULL block for the extra log buffer used during splits so that
1390          * it will trigger errors if we ever try to do IO on it without first
1391          * having set it up properly.
1392          */
1393         error = -ENOMEM;
1394         bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1395                            BTOBB(log->l_iclog_size), 0);
1396         if (!bp)
1397                 goto out_free_log;
1398
1399         /*
1400          * The iclogbuf buffer locks are held over IO but we are not going to do
1401          * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1402          * when appropriately.
1403          */
1404         ASSERT(xfs_buf_islocked(bp));
1405         xfs_buf_unlock(bp);
1406
1407         /* use high priority wq for log I/O completion */
1408         bp->b_ioend_wq = mp->m_log_workqueue;
1409         bp->b_iodone = xlog_iodone;
1410         log->l_xbuf = bp;
1411
1412         spin_lock_init(&log->l_icloglock);
1413         init_waitqueue_head(&log->l_flush_wait);
1414
1415         iclogp = &log->l_iclog;
1416         /*
1417          * The amount of memory to allocate for the iclog structure is
1418          * rather funky due to the way the structure is defined.  It is
1419          * done this way so that we can use different sizes for machines
1420          * with different amounts of memory.  See the definition of
1421          * xlog_in_core_t in xfs_log_priv.h for details.
1422          */
1423         ASSERT(log->l_iclog_size >= 4096);
1424         for (i=0; i < log->l_iclog_bufs; i++) {
1425                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1426                 if (!*iclogp)
1427                         goto out_free_iclog;
1428
1429                 iclog = *iclogp;
1430                 iclog->ic_prev = prev_iclog;
1431                 prev_iclog = iclog;
1432
1433                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1434                                                 BTOBB(log->l_iclog_size), 0);
1435                 if (!bp)
1436                         goto out_free_iclog;
1437
1438                 ASSERT(xfs_buf_islocked(bp));
1439                 xfs_buf_unlock(bp);
1440
1441                 /* use high priority wq for log I/O completion */
1442                 bp->b_ioend_wq = mp->m_log_workqueue;
1443                 bp->b_iodone = xlog_iodone;
1444                 iclog->ic_bp = bp;
1445                 iclog->ic_data = bp->b_addr;
1446 #ifdef DEBUG
1447                 log->l_iclog_bak[i] = &iclog->ic_header;
1448 #endif
1449                 head = &iclog->ic_header;
1450                 memset(head, 0, sizeof(xlog_rec_header_t));
1451                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1452                 head->h_version = cpu_to_be32(
1453                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1454                 head->h_size = cpu_to_be32(log->l_iclog_size);
1455                 /* new fields */
1456                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1457                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1458
1459                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1460                 iclog->ic_state = XLOG_STATE_ACTIVE;
1461                 iclog->ic_log = log;
1462                 atomic_set(&iclog->ic_refcnt, 0);
1463                 spin_lock_init(&iclog->ic_callback_lock);
1464                 iclog->ic_callback_tail = &(iclog->ic_callback);
1465                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1466
1467                 init_waitqueue_head(&iclog->ic_force_wait);
1468                 init_waitqueue_head(&iclog->ic_write_wait);
1469
1470                 iclogp = &iclog->ic_next;
1471         }
1472         *iclogp = log->l_iclog;                 /* complete ring */
1473         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1474
1475         error = xlog_cil_init(log);
1476         if (error)
1477                 goto out_free_iclog;
1478         return log;
1479
1480 out_free_iclog:
1481         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1482                 prev_iclog = iclog->ic_next;
1483                 if (iclog->ic_bp)
1484                         xfs_buf_free(iclog->ic_bp);
1485                 kmem_free(iclog);
1486         }
1487         spinlock_destroy(&log->l_icloglock);
1488         xfs_buf_free(log->l_xbuf);
1489 out_free_log:
1490         kmem_free(log);
1491 out:
1492         return ERR_PTR(error);
1493 }       /* xlog_alloc_log */
1494
1495
1496 /*
1497  * Write out the commit record of a transaction associated with the given
1498  * ticket.  Return the lsn of the commit record.
1499  */
1500 STATIC int
1501 xlog_commit_record(
1502         struct xlog             *log,
1503         struct xlog_ticket      *ticket,
1504         struct xlog_in_core     **iclog,
1505         xfs_lsn_t               *commitlsnp)
1506 {
1507         struct xfs_mount *mp = log->l_mp;
1508         int     error;
1509         struct xfs_log_iovec reg = {
1510                 .i_addr = NULL,
1511                 .i_len = 0,
1512                 .i_type = XLOG_REG_TYPE_COMMIT,
1513         };
1514         struct xfs_log_vec vec = {
1515                 .lv_niovecs = 1,
1516                 .lv_iovecp = &reg,
1517         };
1518
1519         ASSERT_ALWAYS(iclog);
1520         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1521                                         XLOG_COMMIT_TRANS);
1522         if (error)
1523                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1524         return error;
1525 }
1526
1527 /*
1528  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1529  * log space.  This code pushes on the lsn which would supposedly free up
1530  * the 25% which we want to leave free.  We may need to adopt a policy which
1531  * pushes on an lsn which is further along in the log once we reach the high
1532  * water mark.  In this manner, we would be creating a low water mark.
1533  */
1534 STATIC void
1535 xlog_grant_push_ail(
1536         struct xlog     *log,
1537         int             need_bytes)
1538 {
1539         xfs_lsn_t       threshold_lsn = 0;
1540         xfs_lsn_t       last_sync_lsn;
1541         int             free_blocks;
1542         int             free_bytes;
1543         int             threshold_block;
1544         int             threshold_cycle;
1545         int             free_threshold;
1546
1547         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1548
1549         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1550         free_blocks = BTOBBT(free_bytes);
1551
1552         /*
1553          * Set the threshold for the minimum number of free blocks in the
1554          * log to the maximum of what the caller needs, one quarter of the
1555          * log, and 256 blocks.
1556          */
1557         free_threshold = BTOBB(need_bytes);
1558         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1559         free_threshold = MAX(free_threshold, 256);
1560         if (free_blocks >= free_threshold)
1561                 return;
1562
1563         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1564                                                 &threshold_block);
1565         threshold_block += free_threshold;
1566         if (threshold_block >= log->l_logBBsize) {
1567                 threshold_block -= log->l_logBBsize;
1568                 threshold_cycle += 1;
1569         }
1570         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1571                                         threshold_block);
1572         /*
1573          * Don't pass in an lsn greater than the lsn of the last
1574          * log record known to be on disk. Use a snapshot of the last sync lsn
1575          * so that it doesn't change between the compare and the set.
1576          */
1577         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1578         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1579                 threshold_lsn = last_sync_lsn;
1580
1581         /*
1582          * Get the transaction layer to kick the dirty buffers out to
1583          * disk asynchronously. No point in trying to do this if
1584          * the filesystem is shutting down.
1585          */
1586         if (!XLOG_FORCED_SHUTDOWN(log))
1587                 xfs_ail_push(log->l_ailp, threshold_lsn);
1588 }
1589
1590 /*
1591  * Stamp cycle number in every block
1592  */
1593 STATIC void
1594 xlog_pack_data(
1595         struct xlog             *log,
1596         struct xlog_in_core     *iclog,
1597         int                     roundoff)
1598 {
1599         int                     i, j, k;
1600         int                     size = iclog->ic_offset + roundoff;
1601         __be32                  cycle_lsn;
1602         char                    *dp;
1603
1604         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1605
1606         dp = iclog->ic_datap;
1607         for (i = 0; i < BTOBB(size); i++) {
1608                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1609                         break;
1610                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1611                 *(__be32 *)dp = cycle_lsn;
1612                 dp += BBSIZE;
1613         }
1614
1615         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1616                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1617
1618                 for ( ; i < BTOBB(size); i++) {
1619                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1620                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1621                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1622                         *(__be32 *)dp = cycle_lsn;
1623                         dp += BBSIZE;
1624                 }
1625
1626                 for (i = 1; i < log->l_iclog_heads; i++)
1627                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1628         }
1629 }
1630
1631 /*
1632  * Calculate the checksum for a log buffer.
1633  *
1634  * This is a little more complicated than it should be because the various
1635  * headers and the actual data are non-contiguous.
1636  */
1637 __le32
1638 xlog_cksum(
1639         struct xlog             *log,
1640         struct xlog_rec_header  *rhead,
1641         char                    *dp,
1642         int                     size)
1643 {
1644         __uint32_t              crc;
1645
1646         /* first generate the crc for the record header ... */
1647         crc = xfs_start_cksum((char *)rhead,
1648                               sizeof(struct xlog_rec_header),
1649                               offsetof(struct xlog_rec_header, h_crc));
1650
1651         /* ... then for additional cycle data for v2 logs ... */
1652         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1653                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1654                 int             i;
1655
1656                 for (i = 1; i < log->l_iclog_heads; i++) {
1657                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1658                                      sizeof(struct xlog_rec_ext_header));
1659                 }
1660         }
1661
1662         /* ... and finally for the payload */
1663         crc = crc32c(crc, dp, size);
1664
1665         return xfs_end_cksum(crc);
1666 }
1667
1668 /*
1669  * The bdstrat callback function for log bufs. This gives us a central
1670  * place to trap bufs in case we get hit by a log I/O error and need to
1671  * shutdown. Actually, in practice, even when we didn't get a log error,
1672  * we transition the iclogs to IOERROR state *after* flushing all existing
1673  * iclogs to disk. This is because we don't want anymore new transactions to be
1674  * started or completed afterwards.
1675  *
1676  * We lock the iclogbufs here so that we can serialise against IO completion
1677  * during unmount. We might be processing a shutdown triggered during unmount,
1678  * and that can occur asynchronously to the unmount thread, and hence we need to
1679  * ensure that completes before tearing down the iclogbufs. Hence we need to
1680  * hold the buffer lock across the log IO to acheive that.
1681  */
1682 STATIC int
1683 xlog_bdstrat(
1684         struct xfs_buf          *bp)
1685 {
1686         struct xlog_in_core     *iclog = bp->b_fspriv;
1687
1688         xfs_buf_lock(bp);
1689         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1690                 xfs_buf_ioerror(bp, -EIO);
1691                 xfs_buf_stale(bp);
1692                 xfs_buf_ioend(bp);
1693                 /*
1694                  * It would seem logical to return EIO here, but we rely on
1695                  * the log state machine to propagate I/O errors instead of
1696                  * doing it here. Similarly, IO completion will unlock the
1697                  * buffer, so we don't do it here.
1698                  */
1699                 return 0;
1700         }
1701
1702         xfs_buf_submit(bp);
1703         return 0;
1704 }
1705
1706 /*
1707  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1708  * fashion.  Previously, we should have moved the current iclog
1709  * ptr in the log to point to the next available iclog.  This allows further
1710  * write to continue while this code syncs out an iclog ready to go.
1711  * Before an in-core log can be written out, the data section must be scanned
1712  * to save away the 1st word of each BBSIZE block into the header.  We replace
1713  * it with the current cycle count.  Each BBSIZE block is tagged with the
1714  * cycle count because there in an implicit assumption that drives will
1715  * guarantee that entire 512 byte blocks get written at once.  In other words,
1716  * we can't have part of a 512 byte block written and part not written.  By
1717  * tagging each block, we will know which blocks are valid when recovering
1718  * after an unclean shutdown.
1719  *
1720  * This routine is single threaded on the iclog.  No other thread can be in
1721  * this routine with the same iclog.  Changing contents of iclog can there-
1722  * fore be done without grabbing the state machine lock.  Updating the global
1723  * log will require grabbing the lock though.
1724  *
1725  * The entire log manager uses a logical block numbering scheme.  Only
1726  * log_sync (and then only bwrite()) know about the fact that the log may
1727  * not start with block zero on a given device.  The log block start offset
1728  * is added immediately before calling bwrite().
1729  */
1730
1731 STATIC int
1732 xlog_sync(
1733         struct xlog             *log,
1734         struct xlog_in_core     *iclog)
1735 {
1736         xfs_buf_t       *bp;
1737         int             i;
1738         uint            count;          /* byte count of bwrite */
1739         uint            count_init;     /* initial count before roundup */
1740         int             roundoff;       /* roundoff to BB or stripe */
1741         int             split = 0;      /* split write into two regions */
1742         int             error;
1743         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1744         int             size;
1745
1746         XFS_STATS_INC(xs_log_writes);
1747         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1748
1749         /* Add for LR header */
1750         count_init = log->l_iclog_hsize + iclog->ic_offset;
1751
1752         /* Round out the log write size */
1753         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1754                 /* we have a v2 stripe unit to use */
1755                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1756         } else {
1757                 count = BBTOB(BTOBB(count_init));
1758         }
1759         roundoff = count - count_init;
1760         ASSERT(roundoff >= 0);
1761         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1762                 roundoff < log->l_mp->m_sb.sb_logsunit)
1763                 ||
1764                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1765                  roundoff < BBTOB(1)));
1766
1767         /* move grant heads by roundoff in sync */
1768         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1769         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1770
1771         /* put cycle number in every block */
1772         xlog_pack_data(log, iclog, roundoff);
1773
1774         /* real byte length */
1775         size = iclog->ic_offset;
1776         if (v2)
1777                 size += roundoff;
1778         iclog->ic_header.h_len = cpu_to_be32(size);
1779
1780         bp = iclog->ic_bp;
1781         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1782
1783         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1784
1785         /* Do we need to split this write into 2 parts? */
1786         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1787                 char            *dptr;
1788
1789                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1790                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1791                 iclog->ic_bwritecnt = 2;
1792
1793                 /*
1794                  * Bump the cycle numbers at the start of each block in the
1795                  * part of the iclog that ends up in the buffer that gets
1796                  * written to the start of the log.
1797                  *
1798                  * Watch out for the header magic number case, though.
1799                  */
1800                 dptr = (char *)&iclog->ic_header + count;
1801                 for (i = 0; i < split; i += BBSIZE) {
1802                         __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1803                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1804                                 cycle++;
1805                         *(__be32 *)dptr = cpu_to_be32(cycle);
1806
1807                         dptr += BBSIZE;
1808                 }
1809         } else {
1810                 iclog->ic_bwritecnt = 1;
1811         }
1812
1813         /* calculcate the checksum */
1814         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1815                                             iclog->ic_datap, size);
1816
1817         bp->b_io_length = BTOBB(count);
1818         bp->b_fspriv = iclog;
1819         XFS_BUF_ZEROFLAGS(bp);
1820         XFS_BUF_ASYNC(bp);
1821         bp->b_flags |= XBF_SYNCIO;
1822
1823         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1824                 bp->b_flags |= XBF_FUA;
1825
1826                 /*
1827                  * Flush the data device before flushing the log to make
1828                  * sure all meta data written back from the AIL actually made
1829                  * it to disk before stamping the new log tail LSN into the
1830                  * log buffer.  For an external log we need to issue the
1831                  * flush explicitly, and unfortunately synchronously here;
1832                  * for an internal log we can simply use the block layer
1833                  * state machine for preflushes.
1834                  */
1835                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1836                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1837                 else
1838                         bp->b_flags |= XBF_FLUSH;
1839         }
1840
1841         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1842         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1843
1844         xlog_verify_iclog(log, iclog, count, true);
1845
1846         /* account for log which doesn't start at block #0 */
1847         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1848         /*
1849          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1850          * is shutting down.
1851          */
1852         XFS_BUF_WRITE(bp);
1853
1854         error = xlog_bdstrat(bp);
1855         if (error) {
1856                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1857                 return error;
1858         }
1859         if (split) {
1860                 bp = iclog->ic_log->l_xbuf;
1861                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1862                 xfs_buf_associate_memory(bp,
1863                                 (char *)&iclog->ic_header + count, split);
1864                 bp->b_fspriv = iclog;
1865                 XFS_BUF_ZEROFLAGS(bp);
1866                 XFS_BUF_ASYNC(bp);
1867                 bp->b_flags |= XBF_SYNCIO;
1868                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1869                         bp->b_flags |= XBF_FUA;
1870
1871                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1872                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1873
1874                 /* account for internal log which doesn't start at block #0 */
1875                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1876                 XFS_BUF_WRITE(bp);
1877                 error = xlog_bdstrat(bp);
1878                 if (error) {
1879                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1880                         return error;
1881                 }
1882         }
1883         return 0;
1884 }       /* xlog_sync */
1885
1886 /*
1887  * Deallocate a log structure
1888  */
1889 STATIC void
1890 xlog_dealloc_log(
1891         struct xlog     *log)
1892 {
1893         xlog_in_core_t  *iclog, *next_iclog;
1894         int             i;
1895
1896         xlog_cil_destroy(log);
1897
1898         /*
1899          * Cycle all the iclogbuf locks to make sure all log IO completion
1900          * is done before we tear down these buffers.
1901          */
1902         iclog = log->l_iclog;
1903         for (i = 0; i < log->l_iclog_bufs; i++) {
1904                 xfs_buf_lock(iclog->ic_bp);
1905                 xfs_buf_unlock(iclog->ic_bp);
1906                 iclog = iclog->ic_next;
1907         }
1908
1909         /*
1910          * Always need to ensure that the extra buffer does not point to memory
1911          * owned by another log buffer before we free it. Also, cycle the lock
1912          * first to ensure we've completed IO on it.
1913          */
1914         xfs_buf_lock(log->l_xbuf);
1915         xfs_buf_unlock(log->l_xbuf);
1916         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1917         xfs_buf_free(log->l_xbuf);
1918
1919         iclog = log->l_iclog;
1920         for (i = 0; i < log->l_iclog_bufs; i++) {
1921                 xfs_buf_free(iclog->ic_bp);
1922                 next_iclog = iclog->ic_next;
1923                 kmem_free(iclog);
1924                 iclog = next_iclog;
1925         }
1926         spinlock_destroy(&log->l_icloglock);
1927
1928         log->l_mp->m_log = NULL;
1929         kmem_free(log);
1930 }       /* xlog_dealloc_log */
1931
1932 /*
1933  * Update counters atomically now that memcpy is done.
1934  */
1935 /* ARGSUSED */
1936 static inline void
1937 xlog_state_finish_copy(
1938         struct xlog             *log,
1939         struct xlog_in_core     *iclog,
1940         int                     record_cnt,
1941         int                     copy_bytes)
1942 {
1943         spin_lock(&log->l_icloglock);
1944
1945         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1946         iclog->ic_offset += copy_bytes;
1947
1948         spin_unlock(&log->l_icloglock);
1949 }       /* xlog_state_finish_copy */
1950
1951
1952
1953
1954 /*
1955  * print out info relating to regions written which consume
1956  * the reservation
1957  */
1958 void
1959 xlog_print_tic_res(
1960         struct xfs_mount        *mp,
1961         struct xlog_ticket      *ticket)
1962 {
1963         uint i;
1964         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1965
1966         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1967         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1968             "bformat",
1969             "bchunk",
1970             "efi_format",
1971             "efd_format",
1972             "iformat",
1973             "icore",
1974             "iext",
1975             "ibroot",
1976             "ilocal",
1977             "iattr_ext",
1978             "iattr_broot",
1979             "iattr_local",
1980             "qformat",
1981             "dquot",
1982             "quotaoff",
1983             "LR header",
1984             "unmount",
1985             "commit",
1986             "trans header"
1987         };
1988         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1989             "SETATTR_NOT_SIZE",
1990             "SETATTR_SIZE",
1991             "INACTIVE",
1992             "CREATE",
1993             "CREATE_TRUNC",
1994             "TRUNCATE_FILE",
1995             "REMOVE",
1996             "LINK",
1997             "RENAME",
1998             "MKDIR",
1999             "RMDIR",
2000             "SYMLINK",
2001             "SET_DMATTRS",
2002             "GROWFS",
2003             "STRAT_WRITE",
2004             "DIOSTRAT",
2005             "WRITE_SYNC",
2006             "WRITEID",
2007             "ADDAFORK",
2008             "ATTRINVAL",
2009             "ATRUNCATE",
2010             "ATTR_SET",
2011             "ATTR_RM",
2012             "ATTR_FLAG",
2013             "CLEAR_AGI_BUCKET",
2014             "QM_SBCHANGE",
2015             "DUMMY1",
2016             "DUMMY2",
2017             "QM_QUOTAOFF",
2018             "QM_DQALLOC",
2019             "QM_SETQLIM",
2020             "QM_DQCLUSTER",
2021             "QM_QINOCREATE",
2022             "QM_QUOTAOFF_END",
2023             "SB_UNIT",
2024             "FSYNC_TS",
2025             "GROWFSRT_ALLOC",
2026             "GROWFSRT_ZERO",
2027             "GROWFSRT_FREE",
2028             "SWAPEXT"
2029         };
2030
2031         xfs_warn(mp,
2032                 "xlog_write: reservation summary:\n"
2033                 "  trans type  = %s (%u)\n"
2034                 "  unit res    = %d bytes\n"
2035                 "  current res = %d bytes\n"
2036                 "  total reg   = %u bytes (o/flow = %u bytes)\n"
2037                 "  ophdrs      = %u (ophdr space = %u bytes)\n"
2038                 "  ophdr + reg = %u bytes\n"
2039                 "  num regions = %u",
2040                 ((ticket->t_trans_type <= 0 ||
2041                   ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2042                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
2043                 ticket->t_trans_type,
2044                 ticket->t_unit_res,
2045                 ticket->t_curr_res,
2046                 ticket->t_res_arr_sum, ticket->t_res_o_flow,
2047                 ticket->t_res_num_ophdrs, ophdr_spc,
2048                 ticket->t_res_arr_sum +
2049                 ticket->t_res_o_flow + ophdr_spc,
2050                 ticket->t_res_num);
2051
2052         for (i = 0; i < ticket->t_res_num; i++) {
2053                 uint r_type = ticket->t_res_arr[i].r_type;
2054                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2055                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2056                             "bad-rtype" : res_type_str[r_type-1]),
2057                             ticket->t_res_arr[i].r_len);
2058         }
2059
2060         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2061                 "xlog_write: reservation ran out. Need to up reservation");
2062         xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2063 }
2064
2065 /*
2066  * Calculate the potential space needed by the log vector.  Each region gets
2067  * its own xlog_op_header_t and may need to be double word aligned.
2068  */
2069 static int
2070 xlog_write_calc_vec_length(
2071         struct xlog_ticket      *ticket,
2072         struct xfs_log_vec      *log_vector)
2073 {
2074         struct xfs_log_vec      *lv;
2075         int                     headers = 0;
2076         int                     len = 0;
2077         int                     i;
2078
2079         /* acct for start rec of xact */
2080         if (ticket->t_flags & XLOG_TIC_INITED)
2081                 headers++;
2082
2083         for (lv = log_vector; lv; lv = lv->lv_next) {
2084                 /* we don't write ordered log vectors */
2085                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2086                         continue;
2087
2088                 headers += lv->lv_niovecs;
2089
2090                 for (i = 0; i < lv->lv_niovecs; i++) {
2091                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2092
2093                         len += vecp->i_len;
2094                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2095                 }
2096         }
2097
2098         ticket->t_res_num_ophdrs += headers;
2099         len += headers * sizeof(struct xlog_op_header);
2100
2101         return len;
2102 }
2103
2104 /*
2105  * If first write for transaction, insert start record  We can't be trying to
2106  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2107  */
2108 static int
2109 xlog_write_start_rec(
2110         struct xlog_op_header   *ophdr,
2111         struct xlog_ticket      *ticket)
2112 {
2113         if (!(ticket->t_flags & XLOG_TIC_INITED))
2114                 return 0;
2115
2116         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2117         ophdr->oh_clientid = ticket->t_clientid;
2118         ophdr->oh_len = 0;
2119         ophdr->oh_flags = XLOG_START_TRANS;
2120         ophdr->oh_res2 = 0;
2121
2122         ticket->t_flags &= ~XLOG_TIC_INITED;
2123
2124         return sizeof(struct xlog_op_header);
2125 }
2126
2127 static xlog_op_header_t *
2128 xlog_write_setup_ophdr(
2129         struct xlog             *log,
2130         struct xlog_op_header   *ophdr,
2131         struct xlog_ticket      *ticket,
2132         uint                    flags)
2133 {
2134         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2135         ophdr->oh_clientid = ticket->t_clientid;
2136         ophdr->oh_res2 = 0;
2137
2138         /* are we copying a commit or unmount record? */
2139         ophdr->oh_flags = flags;
2140
2141         /*
2142          * We've seen logs corrupted with bad transaction client ids.  This
2143          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2144          * and shut down the filesystem.
2145          */
2146         switch (ophdr->oh_clientid)  {
2147         case XFS_TRANSACTION:
2148         case XFS_VOLUME:
2149         case XFS_LOG:
2150                 break;
2151         default:
2152                 xfs_warn(log->l_mp,
2153                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2154                         ophdr->oh_clientid, ticket);
2155                 return NULL;
2156         }
2157
2158         return ophdr;
2159 }
2160
2161 /*
2162  * Set up the parameters of the region copy into the log. This has
2163  * to handle region write split across multiple log buffers - this
2164  * state is kept external to this function so that this code can
2165  * be written in an obvious, self documenting manner.
2166  */
2167 static int
2168 xlog_write_setup_copy(
2169         struct xlog_ticket      *ticket,
2170         struct xlog_op_header   *ophdr,
2171         int                     space_available,
2172         int                     space_required,
2173         int                     *copy_off,
2174         int                     *copy_len,
2175         int                     *last_was_partial_copy,
2176         int                     *bytes_consumed)
2177 {
2178         int                     still_to_copy;
2179
2180         still_to_copy = space_required - *bytes_consumed;
2181         *copy_off = *bytes_consumed;
2182
2183         if (still_to_copy <= space_available) {
2184                 /* write of region completes here */
2185                 *copy_len = still_to_copy;
2186                 ophdr->oh_len = cpu_to_be32(*copy_len);
2187                 if (*last_was_partial_copy)
2188                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2189                 *last_was_partial_copy = 0;
2190                 *bytes_consumed = 0;
2191                 return 0;
2192         }
2193
2194         /* partial write of region, needs extra log op header reservation */
2195         *copy_len = space_available;
2196         ophdr->oh_len = cpu_to_be32(*copy_len);
2197         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2198         if (*last_was_partial_copy)
2199                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2200         *bytes_consumed += *copy_len;
2201         (*last_was_partial_copy)++;
2202
2203         /* account for new log op header */
2204         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2205         ticket->t_res_num_ophdrs++;
2206
2207         return sizeof(struct xlog_op_header);
2208 }
2209
2210 static int
2211 xlog_write_copy_finish(
2212         struct xlog             *log,
2213         struct xlog_in_core     *iclog,
2214         uint                    flags,
2215         int                     *record_cnt,
2216         int                     *data_cnt,
2217         int                     *partial_copy,
2218         int                     *partial_copy_len,
2219         int                     log_offset,
2220         struct xlog_in_core     **commit_iclog)
2221 {
2222         if (*partial_copy) {
2223                 /*
2224                  * This iclog has already been marked WANT_SYNC by
2225                  * xlog_state_get_iclog_space.
2226                  */
2227                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2228                 *record_cnt = 0;
2229                 *data_cnt = 0;
2230                 return xlog_state_release_iclog(log, iclog);
2231         }
2232
2233         *partial_copy = 0;
2234         *partial_copy_len = 0;
2235
2236         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2237                 /* no more space in this iclog - push it. */
2238                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2239                 *record_cnt = 0;
2240                 *data_cnt = 0;
2241
2242                 spin_lock(&log->l_icloglock);
2243                 xlog_state_want_sync(log, iclog);
2244                 spin_unlock(&log->l_icloglock);
2245
2246                 if (!commit_iclog)
2247                         return xlog_state_release_iclog(log, iclog);
2248                 ASSERT(flags & XLOG_COMMIT_TRANS);
2249                 *commit_iclog = iclog;
2250         }
2251
2252         return 0;
2253 }
2254
2255 /*
2256  * Write some region out to in-core log
2257  *
2258  * This will be called when writing externally provided regions or when
2259  * writing out a commit record for a given transaction.
2260  *
2261  * General algorithm:
2262  *      1. Find total length of this write.  This may include adding to the
2263  *              lengths passed in.
2264  *      2. Check whether we violate the tickets reservation.
2265  *      3. While writing to this iclog
2266  *          A. Reserve as much space in this iclog as can get
2267  *          B. If this is first write, save away start lsn
2268  *          C. While writing this region:
2269  *              1. If first write of transaction, write start record
2270  *              2. Write log operation header (header per region)
2271  *              3. Find out if we can fit entire region into this iclog
2272  *              4. Potentially, verify destination memcpy ptr
2273  *              5. Memcpy (partial) region
2274  *              6. If partial copy, release iclog; otherwise, continue
2275  *                      copying more regions into current iclog
2276  *      4. Mark want sync bit (in simulation mode)
2277  *      5. Release iclog for potential flush to on-disk log.
2278  *
2279  * ERRORS:
2280  * 1.   Panic if reservation is overrun.  This should never happen since
2281  *      reservation amounts are generated internal to the filesystem.
2282  * NOTES:
2283  * 1. Tickets are single threaded data structures.
2284  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2285  *      syncing routine.  When a single log_write region needs to span
2286  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2287  *      on all log operation writes which don't contain the end of the
2288  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2289  *      operation which contains the end of the continued log_write region.
2290  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2291  *      we don't really know exactly how much space will be used.  As a result,
2292  *      we don't update ic_offset until the end when we know exactly how many
2293  *      bytes have been written out.
2294  */
2295 int
2296 xlog_write(
2297         struct xlog             *log,
2298         struct xfs_log_vec      *log_vector,
2299         struct xlog_ticket      *ticket,
2300         xfs_lsn_t               *start_lsn,
2301         struct xlog_in_core     **commit_iclog,
2302         uint                    flags)
2303 {
2304         struct xlog_in_core     *iclog = NULL;
2305         struct xfs_log_iovec    *vecp;
2306         struct xfs_log_vec      *lv;
2307         int                     len;
2308         int                     index;
2309         int                     partial_copy = 0;
2310         int                     partial_copy_len = 0;
2311         int                     contwr = 0;
2312         int                     record_cnt = 0;
2313         int                     data_cnt = 0;
2314         int                     error;
2315
2316         *start_lsn = 0;
2317
2318         len = xlog_write_calc_vec_length(ticket, log_vector);
2319
2320         /*
2321          * Region headers and bytes are already accounted for.
2322          * We only need to take into account start records and
2323          * split regions in this function.
2324          */
2325         if (ticket->t_flags & XLOG_TIC_INITED)
2326                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2327
2328         /*
2329          * Commit record headers need to be accounted for. These
2330          * come in as separate writes so are easy to detect.
2331          */
2332         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2333                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2334
2335         if (ticket->t_curr_res < 0)
2336                 xlog_print_tic_res(log->l_mp, ticket);
2337
2338         index = 0;
2339         lv = log_vector;
2340         vecp = lv->lv_iovecp;
2341         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2342                 void            *ptr;
2343                 int             log_offset;
2344
2345                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2346                                                    &contwr, &log_offset);
2347                 if (error)
2348                         return error;
2349
2350                 ASSERT(log_offset <= iclog->ic_size - 1);
2351                 ptr = iclog->ic_datap + log_offset;
2352
2353                 /* start_lsn is the first lsn written to. That's all we need. */
2354                 if (!*start_lsn)
2355                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2356
2357                 /*
2358                  * This loop writes out as many regions as can fit in the amount
2359                  * of space which was allocated by xlog_state_get_iclog_space().
2360                  */
2361                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2362                         struct xfs_log_iovec    *reg;
2363                         struct xlog_op_header   *ophdr;
2364                         int                     start_rec_copy;
2365                         int                     copy_len;
2366                         int                     copy_off;
2367                         bool                    ordered = false;
2368
2369                         /* ordered log vectors have no regions to write */
2370                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2371                                 ASSERT(lv->lv_niovecs == 0);
2372                                 ordered = true;
2373                                 goto next_lv;
2374                         }
2375
2376                         reg = &vecp[index];
2377                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2378                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2379
2380                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2381                         if (start_rec_copy) {
2382                                 record_cnt++;
2383                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2384                                                    start_rec_copy);
2385                         }
2386
2387                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2388                         if (!ophdr)
2389                                 return -EIO;
2390
2391                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2392                                            sizeof(struct xlog_op_header));
2393
2394                         len += xlog_write_setup_copy(ticket, ophdr,
2395                                                      iclog->ic_size-log_offset,
2396                                                      reg->i_len,
2397                                                      &copy_off, &copy_len,
2398                                                      &partial_copy,
2399                                                      &partial_copy_len);
2400                         xlog_verify_dest_ptr(log, ptr);
2401
2402                         /* copy region */
2403                         ASSERT(copy_len >= 0);
2404                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
2405                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2406
2407                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2408                         record_cnt++;
2409                         data_cnt += contwr ? copy_len : 0;
2410
2411                         error = xlog_write_copy_finish(log, iclog, flags,
2412                                                        &record_cnt, &data_cnt,
2413                                                        &partial_copy,
2414                                                        &partial_copy_len,
2415                                                        log_offset,
2416                                                        commit_iclog);
2417                         if (error)
2418                                 return error;
2419
2420                         /*
2421                          * if we had a partial copy, we need to get more iclog
2422                          * space but we don't want to increment the region
2423                          * index because there is still more is this region to
2424                          * write.
2425                          *
2426                          * If we completed writing this region, and we flushed
2427                          * the iclog (indicated by resetting of the record
2428                          * count), then we also need to get more log space. If
2429                          * this was the last record, though, we are done and
2430                          * can just return.
2431                          */
2432                         if (partial_copy)
2433                                 break;
2434
2435                         if (++index == lv->lv_niovecs) {
2436 next_lv:
2437                                 lv = lv->lv_next;
2438                                 index = 0;
2439                                 if (lv)
2440                                         vecp = lv->lv_iovecp;
2441                         }
2442                         if (record_cnt == 0 && ordered == false) {
2443                                 if (!lv)
2444                                         return 0;
2445                                 break;
2446                         }
2447                 }
2448         }
2449
2450         ASSERT(len == 0);
2451
2452         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2453         if (!commit_iclog)
2454                 return xlog_state_release_iclog(log, iclog);
2455
2456         ASSERT(flags & XLOG_COMMIT_TRANS);
2457         *commit_iclog = iclog;
2458         return 0;
2459 }
2460
2461
2462 /*****************************************************************************
2463  *
2464  *              State Machine functions
2465  *
2466  *****************************************************************************
2467  */
2468
2469 /* Clean iclogs starting from the head.  This ordering must be
2470  * maintained, so an iclog doesn't become ACTIVE beyond one that
2471  * is SYNCING.  This is also required to maintain the notion that we use
2472  * a ordered wait queue to hold off would be writers to the log when every
2473  * iclog is trying to sync to disk.
2474  *
2475  * State Change: DIRTY -> ACTIVE
2476  */
2477 STATIC void
2478 xlog_state_clean_log(
2479         struct xlog *log)
2480 {
2481         xlog_in_core_t  *iclog;
2482         int changed = 0;
2483
2484         iclog = log->l_iclog;
2485         do {
2486                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2487                         iclog->ic_state = XLOG_STATE_ACTIVE;
2488                         iclog->ic_offset       = 0;
2489                         ASSERT(iclog->ic_callback == NULL);
2490                         /*
2491                          * If the number of ops in this iclog indicate it just
2492                          * contains the dummy transaction, we can
2493                          * change state into IDLE (the second time around).
2494                          * Otherwise we should change the state into
2495                          * NEED a dummy.
2496                          * We don't need to cover the dummy.
2497                          */
2498                         if (!changed &&
2499                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2500                                         XLOG_COVER_OPS)) {
2501                                 changed = 1;
2502                         } else {
2503                                 /*
2504                                  * We have two dirty iclogs so start over
2505                                  * This could also be num of ops indicates
2506                                  * this is not the dummy going out.
2507                                  */
2508                                 changed = 2;
2509                         }
2510                         iclog->ic_header.h_num_logops = 0;
2511                         memset(iclog->ic_header.h_cycle_data, 0,
2512                               sizeof(iclog->ic_header.h_cycle_data));
2513                         iclog->ic_header.h_lsn = 0;
2514                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2515                         /* do nothing */;
2516                 else
2517                         break;  /* stop cleaning */
2518                 iclog = iclog->ic_next;
2519         } while (iclog != log->l_iclog);
2520
2521         /* log is locked when we are called */
2522         /*
2523          * Change state for the dummy log recording.
2524          * We usually go to NEED. But we go to NEED2 if the changed indicates
2525          * we are done writing the dummy record.
2526          * If we are done with the second dummy recored (DONE2), then
2527          * we go to IDLE.
2528          */
2529         if (changed) {
2530                 switch (log->l_covered_state) {
2531                 case XLOG_STATE_COVER_IDLE:
2532                 case XLOG_STATE_COVER_NEED:
2533                 case XLOG_STATE_COVER_NEED2:
2534                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2535                         break;
2536
2537                 case XLOG_STATE_COVER_DONE:
2538                         if (changed == 1)
2539                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2540                         else
2541                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2542                         break;
2543
2544                 case XLOG_STATE_COVER_DONE2:
2545                         if (changed == 1)
2546                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2547                         else
2548                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2549                         break;
2550
2551                 default:
2552                         ASSERT(0);
2553                 }
2554         }
2555 }       /* xlog_state_clean_log */
2556
2557 STATIC xfs_lsn_t
2558 xlog_get_lowest_lsn(
2559         struct xlog     *log)
2560 {
2561         xlog_in_core_t  *lsn_log;
2562         xfs_lsn_t       lowest_lsn, lsn;
2563
2564         lsn_log = log->l_iclog;
2565         lowest_lsn = 0;
2566         do {
2567             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2568                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2569                 if ((lsn && !lowest_lsn) ||
2570                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2571                         lowest_lsn = lsn;
2572                 }
2573             }
2574             lsn_log = lsn_log->ic_next;
2575         } while (lsn_log != log->l_iclog);
2576         return lowest_lsn;
2577 }
2578
2579
2580 STATIC void
2581 xlog_state_do_callback(
2582         struct xlog             *log,
2583         int                     aborted,
2584         struct xlog_in_core     *ciclog)
2585 {
2586         xlog_in_core_t     *iclog;
2587         xlog_in_core_t     *first_iclog;        /* used to know when we've
2588                                                  * processed all iclogs once */
2589         xfs_log_callback_t *cb, *cb_next;
2590         int                flushcnt = 0;
2591         xfs_lsn_t          lowest_lsn;
2592         int                ioerrors;    /* counter: iclogs with errors */
2593         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2594         int                funcdidcallbacks; /* flag: function did callbacks */
2595         int                repeats;     /* for issuing console warnings if
2596                                          * looping too many times */
2597         int                wake = 0;
2598
2599         spin_lock(&log->l_icloglock);
2600         first_iclog = iclog = log->l_iclog;
2601         ioerrors = 0;
2602         funcdidcallbacks = 0;
2603         repeats = 0;
2604
2605         do {
2606                 /*
2607                  * Scan all iclogs starting with the one pointed to by the
2608                  * log.  Reset this starting point each time the log is
2609                  * unlocked (during callbacks).
2610                  *
2611                  * Keep looping through iclogs until one full pass is made
2612                  * without running any callbacks.
2613                  */
2614                 first_iclog = log->l_iclog;
2615                 iclog = log->l_iclog;
2616                 loopdidcallbacks = 0;
2617                 repeats++;
2618
2619                 do {
2620
2621                         /* skip all iclogs in the ACTIVE & DIRTY states */
2622                         if (iclog->ic_state &
2623                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2624                                 iclog = iclog->ic_next;
2625                                 continue;
2626                         }
2627
2628                         /*
2629                          * Between marking a filesystem SHUTDOWN and stopping
2630                          * the log, we do flush all iclogs to disk (if there
2631                          * wasn't a log I/O error). So, we do want things to
2632                          * go smoothly in case of just a SHUTDOWN  w/o a
2633                          * LOG_IO_ERROR.
2634                          */
2635                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2636                                 /*
2637                                  * Can only perform callbacks in order.  Since
2638                                  * this iclog is not in the DONE_SYNC/
2639                                  * DO_CALLBACK state, we skip the rest and
2640                                  * just try to clean up.  If we set our iclog
2641                                  * to DO_CALLBACK, we will not process it when
2642                                  * we retry since a previous iclog is in the
2643                                  * CALLBACK and the state cannot change since
2644                                  * we are holding the l_icloglock.
2645                                  */
2646                                 if (!(iclog->ic_state &
2647                                         (XLOG_STATE_DONE_SYNC |
2648                                                  XLOG_STATE_DO_CALLBACK))) {
2649                                         if (ciclog && (ciclog->ic_state ==
2650                                                         XLOG_STATE_DONE_SYNC)) {
2651                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2652                                         }
2653                                         break;
2654                                 }
2655                                 /*
2656                                  * We now have an iclog that is in either the
2657                                  * DO_CALLBACK or DONE_SYNC states. The other
2658                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2659                                  * caught by the above if and are going to
2660                                  * clean (i.e. we aren't doing their callbacks)
2661                                  * see the above if.
2662                                  */
2663
2664                                 /*
2665                                  * We will do one more check here to see if we
2666                                  * have chased our tail around.
2667                                  */
2668
2669                                 lowest_lsn = xlog_get_lowest_lsn(log);
2670                                 if (lowest_lsn &&
2671                                     XFS_LSN_CMP(lowest_lsn,
2672                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2673                                         iclog = iclog->ic_next;
2674                                         continue; /* Leave this iclog for
2675                                                    * another thread */
2676                                 }
2677
2678                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2679
2680
2681                                 /*
2682                                  * Completion of a iclog IO does not imply that
2683                                  * a transaction has completed, as transactions
2684                                  * can be large enough to span many iclogs. We
2685                                  * cannot change the tail of the log half way
2686                                  * through a transaction as this may be the only
2687                                  * transaction in the log and moving th etail to
2688                                  * point to the middle of it will prevent
2689                                  * recovery from finding the start of the
2690                                  * transaction. Hence we should only update the
2691                                  * last_sync_lsn if this iclog contains
2692                                  * transaction completion callbacks on it.
2693                                  *
2694                                  * We have to do this before we drop the
2695                                  * icloglock to ensure we are the only one that
2696                                  * can update it.
2697                                  */
2698                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2699                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2700                                 if (iclog->ic_callback)
2701                                         atomic64_set(&log->l_last_sync_lsn,
2702                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2703
2704                         } else
2705                                 ioerrors++;
2706
2707                         spin_unlock(&log->l_icloglock);
2708
2709                         /*
2710                          * Keep processing entries in the callback list until
2711                          * we come around and it is empty.  We need to
2712                          * atomically see that the list is empty and change the
2713                          * state to DIRTY so that we don't miss any more
2714                          * callbacks being added.
2715                          */
2716                         spin_lock(&iclog->ic_callback_lock);
2717                         cb = iclog->ic_callback;
2718                         while (cb) {
2719                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2720                                 iclog->ic_callback = NULL;
2721                                 spin_unlock(&iclog->ic_callback_lock);
2722
2723                                 /* perform callbacks in the order given */
2724                                 for (; cb; cb = cb_next) {
2725                                         cb_next = cb->cb_next;
2726                                         cb->cb_func(cb->cb_arg, aborted);
2727                                 }
2728                                 spin_lock(&iclog->ic_callback_lock);
2729                                 cb = iclog->ic_callback;
2730                         }
2731
2732                         loopdidcallbacks++;
2733                         funcdidcallbacks++;
2734
2735                         spin_lock(&log->l_icloglock);
2736                         ASSERT(iclog->ic_callback == NULL);
2737                         spin_unlock(&iclog->ic_callback_lock);
2738                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2739                                 iclog->ic_state = XLOG_STATE_DIRTY;
2740
2741                         /*
2742                          * Transition from DIRTY to ACTIVE if applicable.
2743                          * NOP if STATE_IOERROR.
2744                          */
2745                         xlog_state_clean_log(log);
2746
2747                         /* wake up threads waiting in xfs_log_force() */
2748                         wake_up_all(&iclog->ic_force_wait);
2749
2750                         iclog = iclog->ic_next;
2751                 } while (first_iclog != iclog);
2752
2753                 if (repeats > 5000) {
2754                         flushcnt += repeats;
2755                         repeats = 0;
2756                         xfs_warn(log->l_mp,
2757                                 "%s: possible infinite loop (%d iterations)",
2758                                 __func__, flushcnt);
2759                 }
2760         } while (!ioerrors && loopdidcallbacks);
2761
2762         /*
2763          * make one last gasp attempt to see if iclogs are being left in
2764          * limbo..
2765          */
2766 #ifdef DEBUG
2767         if (funcdidcallbacks) {
2768                 first_iclog = iclog = log->l_iclog;
2769                 do {
2770                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2771                         /*
2772                          * Terminate the loop if iclogs are found in states
2773                          * which will cause other threads to clean up iclogs.
2774                          *
2775                          * SYNCING - i/o completion will go through logs
2776                          * DONE_SYNC - interrupt thread should be waiting for
2777                          *              l_icloglock
2778                          * IOERROR - give up hope all ye who enter here
2779                          */
2780                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2781                             iclog->ic_state == XLOG_STATE_SYNCING ||
2782                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2783                             iclog->ic_state == XLOG_STATE_IOERROR )
2784                                 break;
2785                         iclog = iclog->ic_next;
2786                 } while (first_iclog != iclog);
2787         }
2788 #endif
2789
2790         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2791                 wake = 1;
2792         spin_unlock(&log->l_icloglock);
2793
2794         if (wake)
2795                 wake_up_all(&log->l_flush_wait);
2796 }
2797
2798
2799 /*
2800  * Finish transitioning this iclog to the dirty state.
2801  *
2802  * Make sure that we completely execute this routine only when this is
2803  * the last call to the iclog.  There is a good chance that iclog flushes,
2804  * when we reach the end of the physical log, get turned into 2 separate
2805  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2806  * routine.  By using the reference count bwritecnt, we guarantee that only
2807  * the second completion goes through.
2808  *
2809  * Callbacks could take time, so they are done outside the scope of the
2810  * global state machine log lock.
2811  */
2812 STATIC void
2813 xlog_state_done_syncing(
2814         xlog_in_core_t  *iclog,
2815         int             aborted)
2816 {
2817         struct xlog        *log = iclog->ic_log;
2818
2819         spin_lock(&log->l_icloglock);
2820
2821         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2822                iclog->ic_state == XLOG_STATE_IOERROR);
2823         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2824         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2825
2826
2827         /*
2828          * If we got an error, either on the first buffer, or in the case of
2829          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2830          * and none should ever be attempted to be written to disk
2831          * again.
2832          */
2833         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2834                 if (--iclog->ic_bwritecnt == 1) {
2835                         spin_unlock(&log->l_icloglock);
2836                         return;
2837                 }
2838                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2839         }
2840
2841         /*
2842          * Someone could be sleeping prior to writing out the next
2843          * iclog buffer, we wake them all, one will get to do the
2844          * I/O, the others get to wait for the result.
2845          */
2846         wake_up_all(&iclog->ic_write_wait);
2847         spin_unlock(&log->l_icloglock);
2848         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2849 }       /* xlog_state_done_syncing */
2850
2851
2852 /*
2853  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2854  * sleep.  We wait on the flush queue on the head iclog as that should be
2855  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2856  * we will wait here and all new writes will sleep until a sync completes.
2857  *
2858  * The in-core logs are used in a circular fashion. They are not used
2859  * out-of-order even when an iclog past the head is free.
2860  *
2861  * return:
2862  *      * log_offset where xlog_write() can start writing into the in-core
2863  *              log's data space.
2864  *      * in-core log pointer to which xlog_write() should write.
2865  *      * boolean indicating this is a continued write to an in-core log.
2866  *              If this is the last write, then the in-core log's offset field
2867  *              needs to be incremented, depending on the amount of data which
2868  *              is copied.
2869  */
2870 STATIC int
2871 xlog_state_get_iclog_space(
2872         struct xlog             *log,
2873         int                     len,
2874         struct xlog_in_core     **iclogp,
2875         struct xlog_ticket      *ticket,
2876         int                     *continued_write,
2877         int                     *logoffsetp)
2878 {
2879         int               log_offset;
2880         xlog_rec_header_t *head;
2881         xlog_in_core_t    *iclog;
2882         int               error;
2883
2884 restart:
2885         spin_lock(&log->l_icloglock);
2886         if (XLOG_FORCED_SHUTDOWN(log)) {
2887                 spin_unlock(&log->l_icloglock);
2888                 return -EIO;
2889         }
2890
2891         iclog = log->l_iclog;
2892         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2893                 XFS_STATS_INC(xs_log_noiclogs);
2894
2895                 /* Wait for log writes to have flushed */
2896                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2897                 goto restart;
2898         }
2899
2900         head = &iclog->ic_header;
2901
2902         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2903         log_offset = iclog->ic_offset;
2904
2905         /* On the 1st write to an iclog, figure out lsn.  This works
2906          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2907          * committing to.  If the offset is set, that's how many blocks
2908          * must be written.
2909          */
2910         if (log_offset == 0) {
2911                 ticket->t_curr_res -= log->l_iclog_hsize;
2912                 xlog_tic_add_region(ticket,
2913                                     log->l_iclog_hsize,
2914                                     XLOG_REG_TYPE_LRHEADER);
2915                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2916                 head->h_lsn = cpu_to_be64(
2917                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2918                 ASSERT(log->l_curr_block >= 0);
2919         }
2920
2921         /* If there is enough room to write everything, then do it.  Otherwise,
2922          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2923          * bit is on, so this will get flushed out.  Don't update ic_offset
2924          * until you know exactly how many bytes get copied.  Therefore, wait
2925          * until later to update ic_offset.
2926          *
2927          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2928          * can fit into remaining data section.
2929          */
2930         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2931                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2932
2933                 /*
2934                  * If I'm the only one writing to this iclog, sync it to disk.
2935                  * We need to do an atomic compare and decrement here to avoid
2936                  * racing with concurrent atomic_dec_and_lock() calls in
2937                  * xlog_state_release_iclog() when there is more than one
2938                  * reference to the iclog.
2939                  */
2940                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2941                         /* we are the only one */
2942                         spin_unlock(&log->l_icloglock);
2943                         error = xlog_state_release_iclog(log, iclog);
2944                         if (error)
2945                                 return error;
2946                 } else {
2947                         spin_unlock(&log->l_icloglock);
2948                 }
2949                 goto restart;
2950         }
2951
2952         /* Do we have enough room to write the full amount in the remainder
2953          * of this iclog?  Or must we continue a write on the next iclog and
2954          * mark this iclog as completely taken?  In the case where we switch
2955          * iclogs (to mark it taken), this particular iclog will release/sync
2956          * to disk in xlog_write().
2957          */
2958         if (len <= iclog->ic_size - iclog->ic_offset) {
2959                 *continued_write = 0;
2960                 iclog->ic_offset += len;
2961         } else {
2962                 *continued_write = 1;
2963                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2964         }
2965         *iclogp = iclog;
2966
2967         ASSERT(iclog->ic_offset <= iclog->ic_size);
2968         spin_unlock(&log->l_icloglock);
2969
2970         *logoffsetp = log_offset;
2971         return 0;
2972 }       /* xlog_state_get_iclog_space */
2973
2974 /* The first cnt-1 times through here we don't need to
2975  * move the grant write head because the permanent
2976  * reservation has reserved cnt times the unit amount.
2977  * Release part of current permanent unit reservation and
2978  * reset current reservation to be one units worth.  Also
2979  * move grant reservation head forward.
2980  */
2981 STATIC void
2982 xlog_regrant_reserve_log_space(
2983         struct xlog             *log,
2984         struct xlog_ticket      *ticket)
2985 {
2986         trace_xfs_log_regrant_reserve_enter(log, ticket);
2987
2988         if (ticket->t_cnt > 0)
2989                 ticket->t_cnt--;
2990
2991         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2992                                         ticket->t_curr_res);
2993         xlog_grant_sub_space(log, &log->l_write_head.grant,
2994                                         ticket->t_curr_res);
2995         ticket->t_curr_res = ticket->t_unit_res;
2996         xlog_tic_reset_res(ticket);
2997
2998         trace_xfs_log_regrant_reserve_sub(log, ticket);
2999
3000         /* just return if we still have some of the pre-reserved space */
3001         if (ticket->t_cnt > 0)
3002                 return;
3003
3004         xlog_grant_add_space(log, &log->l_reserve_head.grant,
3005                                         ticket->t_unit_res);
3006
3007         trace_xfs_log_regrant_reserve_exit(log, ticket);
3008
3009         ticket->t_curr_res = ticket->t_unit_res;
3010         xlog_tic_reset_res(ticket);
3011 }       /* xlog_regrant_reserve_log_space */
3012
3013
3014 /*
3015  * Give back the space left from a reservation.
3016  *
3017  * All the information we need to make a correct determination of space left
3018  * is present.  For non-permanent reservations, things are quite easy.  The
3019  * count should have been decremented to zero.  We only need to deal with the
3020  * space remaining in the current reservation part of the ticket.  If the
3021  * ticket contains a permanent reservation, there may be left over space which
3022  * needs to be released.  A count of N means that N-1 refills of the current
3023  * reservation can be done before we need to ask for more space.  The first
3024  * one goes to fill up the first current reservation.  Once we run out of
3025  * space, the count will stay at zero and the only space remaining will be
3026  * in the current reservation field.
3027  */
3028 STATIC void
3029 xlog_ungrant_log_space(
3030         struct xlog             *log,
3031         struct xlog_ticket      *ticket)
3032 {
3033         int     bytes;
3034
3035         if (ticket->t_cnt > 0)
3036                 ticket->t_cnt--;
3037
3038         trace_xfs_log_ungrant_enter(log, ticket);
3039         trace_xfs_log_ungrant_sub(log, ticket);
3040
3041         /*
3042          * If this is a permanent reservation ticket, we may be able to free
3043          * up more space based on the remaining count.
3044          */
3045         bytes = ticket->t_curr_res;
3046         if (ticket->t_cnt > 0) {
3047                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3048                 bytes += ticket->t_unit_res*ticket->t_cnt;
3049         }
3050
3051         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3052         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3053
3054         trace_xfs_log_ungrant_exit(log, ticket);
3055
3056         xfs_log_space_wake(log->l_mp);
3057 }
3058
3059 /*
3060  * Flush iclog to disk if this is the last reference to the given iclog and
3061  * the WANT_SYNC bit is set.
3062  *
3063  * When this function is entered, the iclog is not necessarily in the
3064  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3065  *
3066  *
3067  */
3068 STATIC int
3069 xlog_state_release_iclog(
3070         struct xlog             *log,
3071         struct xlog_in_core     *iclog)
3072 {
3073         int             sync = 0;       /* do we sync? */
3074
3075         if (iclog->ic_state & XLOG_STATE_IOERROR)
3076                 return -EIO;
3077
3078         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3079         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3080                 return 0;
3081
3082         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3083                 spin_unlock(&log->l_icloglock);
3084                 return -EIO;
3085         }
3086         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3087                iclog->ic_state == XLOG_STATE_WANT_SYNC);
3088
3089         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3090                 /* update tail before writing to iclog */
3091                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3092                 sync++;
3093                 iclog->ic_state = XLOG_STATE_SYNCING;
3094                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3095                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3096                 /* cycle incremented when incrementing curr_block */
3097         }
3098         spin_unlock(&log->l_icloglock);
3099
3100         /*
3101          * We let the log lock go, so it's possible that we hit a log I/O
3102          * error or some other SHUTDOWN condition that marks the iclog
3103          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3104          * this iclog has consistent data, so we ignore IOERROR
3105          * flags after this point.
3106          */
3107         if (sync)
3108                 return xlog_sync(log, iclog);
3109         return 0;
3110 }       /* xlog_state_release_iclog */
3111
3112
3113 /*
3114  * This routine will mark the current iclog in the ring as WANT_SYNC
3115  * and move the current iclog pointer to the next iclog in the ring.
3116  * When this routine is called from xlog_state_get_iclog_space(), the
3117  * exact size of the iclog has not yet been determined.  All we know is
3118  * that every data block.  We have run out of space in this log record.
3119  */
3120 STATIC void
3121 xlog_state_switch_iclogs(
3122         struct xlog             *log,
3123         struct xlog_in_core     *iclog,
3124         int                     eventual_size)
3125 {
3126         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3127         if (!eventual_size)
3128                 eventual_size = iclog->ic_offset;
3129         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3130         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3131         log->l_prev_block = log->l_curr_block;
3132         log->l_prev_cycle = log->l_curr_cycle;
3133
3134         /* roll log?: ic_offset changed later */
3135         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3136
3137         /* Round up to next log-sunit */
3138         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3139             log->l_mp->m_sb.sb_logsunit > 1) {
3140                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3141                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3142         }
3143
3144         if (log->l_curr_block >= log->l_logBBsize) {
3145                 log->l_curr_cycle++;
3146                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3147                         log->l_curr_cycle++;
3148                 log->l_curr_block -= log->l_logBBsize;
3149                 ASSERT(log->l_curr_block >= 0);
3150         }
3151         ASSERT(iclog == log->l_iclog);
3152         log->l_iclog = iclog->ic_next;
3153 }       /* xlog_state_switch_iclogs */
3154
3155 /*
3156  * Write out all data in the in-core log as of this exact moment in time.
3157  *
3158  * Data may be written to the in-core log during this call.  However,
3159  * we don't guarantee this data will be written out.  A change from past
3160  * implementation means this routine will *not* write out zero length LRs.
3161  *
3162  * Basically, we try and perform an intelligent scan of the in-core logs.
3163  * If we determine there is no flushable data, we just return.  There is no
3164  * flushable data if:
3165  *
3166  *      1. the current iclog is active and has no data; the previous iclog
3167  *              is in the active or dirty state.
3168  *      2. the current iclog is drity, and the previous iclog is in the
3169  *              active or dirty state.
3170  *
3171  * We may sleep if:
3172  *
3173  *      1. the current iclog is not in the active nor dirty state.
3174  *      2. the current iclog dirty, and the previous iclog is not in the
3175  *              active nor dirty state.
3176  *      3. the current iclog is active, and there is another thread writing
3177  *              to this particular iclog.
3178  *      4. a) the current iclog is active and has no other writers
3179  *         b) when we return from flushing out this iclog, it is still
3180  *              not in the active nor dirty state.
3181  */
3182 int
3183 _xfs_log_force(
3184         struct xfs_mount        *mp,
3185         uint                    flags,
3186         int                     *log_flushed)
3187 {
3188         struct xlog             *log = mp->m_log;
3189         struct xlog_in_core     *iclog;
3190         xfs_lsn_t               lsn;
3191
3192         XFS_STATS_INC(xs_log_force);
3193
3194         xlog_cil_force(log);
3195
3196         spin_lock(&log->l_icloglock);
3197
3198         iclog = log->l_iclog;
3199         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3200                 spin_unlock(&log->l_icloglock);
3201                 return -EIO;
3202         }
3203
3204         /* If the head iclog is not active nor dirty, we just attach
3205          * ourselves to the head and go to sleep.
3206          */
3207         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3208             iclog->ic_state == XLOG_STATE_DIRTY) {
3209                 /*
3210                  * If the head is dirty or (active and empty), then
3211                  * we need to look at the previous iclog.  If the previous
3212                  * iclog is active or dirty we are done.  There is nothing
3213                  * to sync out.  Otherwise, we attach ourselves to the
3214                  * previous iclog and go to sleep.
3215                  */
3216                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3217                     (atomic_read(&iclog->ic_refcnt) == 0
3218                      && iclog->ic_offset == 0)) {
3219                         iclog = iclog->ic_prev;
3220                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3221                             iclog->ic_state == XLOG_STATE_DIRTY)
3222                                 goto no_sleep;
3223                         else
3224                                 goto maybe_sleep;
3225                 } else {
3226                         if (atomic_read(&iclog->ic_refcnt) == 0) {
3227                                 /* We are the only one with access to this
3228                                  * iclog.  Flush it out now.  There should
3229                                  * be a roundoff of zero to show that someone
3230                                  * has already taken care of the roundoff from
3231                                  * the previous sync.
3232                                  */
3233                                 atomic_inc(&iclog->ic_refcnt);
3234                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3235                                 xlog_state_switch_iclogs(log, iclog, 0);
3236                                 spin_unlock(&log->l_icloglock);
3237
3238                                 if (xlog_state_release_iclog(log, iclog))
3239                                         return -EIO;
3240
3241                                 if (log_flushed)
3242                                         *log_flushed = 1;
3243                                 spin_lock(&log->l_icloglock);
3244                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3245                                     iclog->ic_state != XLOG_STATE_DIRTY)
3246                                         goto maybe_sleep;
3247                                 else
3248                                         goto no_sleep;
3249                         } else {
3250                                 /* Someone else is writing to this iclog.
3251                                  * Use its call to flush out the data.  However,
3252                                  * the other thread may not force out this LR,
3253                                  * so we mark it WANT_SYNC.
3254                                  */
3255                                 xlog_state_switch_iclogs(log, iclog, 0);
3256                                 goto maybe_sleep;
3257                         }
3258                 }
3259         }
3260
3261         /* By the time we come around again, the iclog could've been filled
3262          * which would give it another lsn.  If we have a new lsn, just
3263          * return because the relevant data has been flushed.
3264          */
3265 maybe_sleep:
3266         if (flags & XFS_LOG_SYNC) {
3267                 /*
3268                  * We must check if we're shutting down here, before
3269                  * we wait, while we're holding the l_icloglock.
3270                  * Then we check again after waking up, in case our
3271                  * sleep was disturbed by a bad news.
3272                  */
3273                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3274                         spin_unlock(&log->l_icloglock);
3275                         return -EIO;
3276                 }
3277                 XFS_STATS_INC(xs_log_force_sleep);
3278                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3279                 /*
3280                  * No need to grab the log lock here since we're
3281                  * only deciding whether or not to return EIO
3282                  * and the memory read should be atomic.
3283                  */
3284                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3285                         return -EIO;
3286                 if (log_flushed)
3287                         *log_flushed = 1;
3288         } else {
3289
3290 no_sleep:
3291                 spin_unlock(&log->l_icloglock);
3292         }
3293         return 0;
3294 }
3295
3296 /*
3297  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3298  * about errors or whether the log was flushed or not. This is the normal
3299  * interface to use when trying to unpin items or move the log forward.
3300  */
3301 void
3302 xfs_log_force(
3303         xfs_mount_t     *mp,
3304         uint            flags)
3305 {
3306         int     error;
3307
3308         trace_xfs_log_force(mp, 0);
3309         error = _xfs_log_force(mp, flags, NULL);
3310         if (error)
3311                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3312 }
3313
3314 /*
3315  * Force the in-core log to disk for a specific LSN.
3316  *
3317  * Find in-core log with lsn.
3318  *      If it is in the DIRTY state, just return.
3319  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3320  *              state and go to sleep or return.
3321  *      If it is in any other state, go to sleep or return.
3322  *
3323  * Synchronous forces are implemented with a signal variable. All callers
3324  * to force a given lsn to disk will wait on a the sv attached to the
3325  * specific in-core log.  When given in-core log finally completes its
3326  * write to disk, that thread will wake up all threads waiting on the
3327  * sv.
3328  */
3329 int
3330 _xfs_log_force_lsn(
3331         struct xfs_mount        *mp,
3332         xfs_lsn_t               lsn,
3333         uint                    flags,
3334         int                     *log_flushed)
3335 {
3336         struct xlog             *log = mp->m_log;
3337         struct xlog_in_core     *iclog;
3338         int                     already_slept = 0;
3339
3340         ASSERT(lsn != 0);
3341
3342         XFS_STATS_INC(xs_log_force);
3343
3344         lsn = xlog_cil_force_lsn(log, lsn);
3345         if (lsn == NULLCOMMITLSN)
3346                 return 0;
3347
3348 try_again:
3349         spin_lock(&log->l_icloglock);
3350         iclog = log->l_iclog;
3351         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3352                 spin_unlock(&log->l_icloglock);
3353                 return -EIO;
3354         }
3355
3356         do {
3357                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3358                         iclog = iclog->ic_next;
3359                         continue;
3360                 }
3361
3362                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3363                         spin_unlock(&log->l_icloglock);
3364                         return 0;
3365                 }
3366
3367                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3368                         /*
3369                          * We sleep here if we haven't already slept (e.g.
3370                          * this is the first time we've looked at the correct
3371                          * iclog buf) and the buffer before us is going to
3372                          * be sync'ed. The reason for this is that if we
3373                          * are doing sync transactions here, by waiting for
3374                          * the previous I/O to complete, we can allow a few
3375                          * more transactions into this iclog before we close
3376                          * it down.
3377                          *
3378                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3379                          * up the refcnt so we can release the log (which
3380                          * drops the ref count).  The state switch keeps new
3381                          * transaction commits from using this buffer.  When
3382                          * the current commits finish writing into the buffer,
3383                          * the refcount will drop to zero and the buffer will
3384                          * go out then.
3385                          */
3386                         if (!already_slept &&
3387                             (iclog->ic_prev->ic_state &
3388                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3389                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3390
3391                                 XFS_STATS_INC(xs_log_force_sleep);
3392
3393                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3394                                                         &log->l_icloglock);
3395                                 if (log_flushed)
3396                                         *log_flushed = 1;
3397                                 already_slept = 1;
3398                                 goto try_again;
3399                         }
3400                         atomic_inc(&iclog->ic_refcnt);
3401                         xlog_state_switch_iclogs(log, iclog, 0);
3402                         spin_unlock(&log->l_icloglock);
3403                         if (xlog_state_release_iclog(log, iclog))
3404                                 return -EIO;
3405                         if (log_flushed)
3406                                 *log_flushed = 1;
3407                         spin_lock(&log->l_icloglock);
3408                 }
3409
3410                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3411                     !(iclog->ic_state &
3412                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3413                         /*
3414                          * Don't wait on completion if we know that we've
3415                          * gotten a log write error.
3416                          */
3417                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3418                                 spin_unlock(&log->l_icloglock);
3419                                 return -EIO;
3420                         }
3421                         XFS_STATS_INC(xs_log_force_sleep);
3422                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3423                         /*
3424                          * No need to grab the log lock here since we're
3425                          * only deciding whether or not to return EIO
3426                          * and the memory read should be atomic.
3427                          */
3428                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3429                                 return -EIO;
3430
3431                         if (log_flushed)
3432                                 *log_flushed = 1;
3433                 } else {                /* just return */
3434                         spin_unlock(&log->l_icloglock);
3435                 }
3436
3437                 return 0;
3438         } while (iclog != log->l_iclog);
3439
3440         spin_unlock(&log->l_icloglock);
3441         return 0;
3442 }
3443
3444 /*
3445  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3446  * about errors or whether the log was flushed or not. This is the normal
3447  * interface to use when trying to unpin items or move the log forward.
3448  */
3449 void
3450 xfs_log_force_lsn(
3451         xfs_mount_t     *mp,
3452         xfs_lsn_t       lsn,
3453         uint            flags)
3454 {
3455         int     error;
3456
3457         trace_xfs_log_force(mp, lsn);
3458         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3459         if (error)
3460                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3461 }
3462
3463 /*
3464  * Called when we want to mark the current iclog as being ready to sync to
3465  * disk.
3466  */
3467 STATIC void
3468 xlog_state_want_sync(
3469         struct xlog             *log,
3470         struct xlog_in_core     *iclog)
3471 {
3472         assert_spin_locked(&log->l_icloglock);
3473
3474         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3475                 xlog_state_switch_iclogs(log, iclog, 0);
3476         } else {
3477                 ASSERT(iclog->ic_state &
3478                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3479         }
3480 }
3481
3482
3483 /*****************************************************************************
3484  *
3485  *              TICKET functions
3486  *
3487  *****************************************************************************
3488  */
3489
3490 /*
3491  * Free a used ticket when its refcount falls to zero.
3492  */
3493 void
3494 xfs_log_ticket_put(
3495         xlog_ticket_t   *ticket)
3496 {
3497         ASSERT(atomic_read(&ticket->t_ref) > 0);
3498         if (atomic_dec_and_test(&ticket->t_ref))
3499                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3500 }
3501
3502 xlog_ticket_t *
3503 xfs_log_ticket_get(
3504         xlog_ticket_t   *ticket)
3505 {
3506         ASSERT(atomic_read(&ticket->t_ref) > 0);
3507         atomic_inc(&ticket->t_ref);
3508         return ticket;
3509 }
3510
3511 /*
3512  * Figure out the total log space unit (in bytes) that would be
3513  * required for a log ticket.
3514  */
3515 int
3516 xfs_log_calc_unit_res(
3517         struct xfs_mount        *mp,
3518         int                     unit_bytes)
3519 {
3520         struct xlog             *log = mp->m_log;
3521         int                     iclog_space;
3522         uint                    num_headers;
3523
3524         /*
3525          * Permanent reservations have up to 'cnt'-1 active log operations
3526          * in the log.  A unit in this case is the amount of space for one
3527          * of these log operations.  Normal reservations have a cnt of 1
3528          * and their unit amount is the total amount of space required.
3529          *
3530          * The following lines of code account for non-transaction data
3531          * which occupy space in the on-disk log.
3532          *
3533          * Normal form of a transaction is:
3534          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3535          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3536          *
3537          * We need to account for all the leadup data and trailer data
3538          * around the transaction data.
3539          * And then we need to account for the worst case in terms of using
3540          * more space.
3541          * The worst case will happen if:
3542          * - the placement of the transaction happens to be such that the
3543          *   roundoff is at its maximum
3544          * - the transaction data is synced before the commit record is synced
3545          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3546          *   Therefore the commit record is in its own Log Record.
3547          *   This can happen as the commit record is called with its
3548          *   own region to xlog_write().
3549          *   This then means that in the worst case, roundoff can happen for
3550          *   the commit-rec as well.
3551          *   The commit-rec is smaller than padding in this scenario and so it is
3552          *   not added separately.
3553          */
3554
3555         /* for trans header */
3556         unit_bytes += sizeof(xlog_op_header_t);
3557         unit_bytes += sizeof(xfs_trans_header_t);
3558
3559         /* for start-rec */
3560         unit_bytes += sizeof(xlog_op_header_t);
3561
3562         /*
3563          * for LR headers - the space for data in an iclog is the size minus
3564          * the space used for the headers. If we use the iclog size, then we
3565          * undercalculate the number of headers required.
3566          *
3567          * Furthermore - the addition of op headers for split-recs might
3568          * increase the space required enough to require more log and op
3569          * headers, so take that into account too.
3570          *
3571          * IMPORTANT: This reservation makes the assumption that if this
3572          * transaction is the first in an iclog and hence has the LR headers
3573          * accounted to it, then the remaining space in the iclog is
3574          * exclusively for this transaction.  i.e. if the transaction is larger
3575          * than the iclog, it will be the only thing in that iclog.
3576          * Fundamentally, this means we must pass the entire log vector to
3577          * xlog_write to guarantee this.
3578          */
3579         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3580         num_headers = howmany(unit_bytes, iclog_space);
3581
3582         /* for split-recs - ophdrs added when data split over LRs */
3583         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3584
3585         /* add extra header reservations if we overrun */
3586         while (!num_headers ||
3587                howmany(unit_bytes, iclog_space) > num_headers) {
3588                 unit_bytes += sizeof(xlog_op_header_t);
3589                 num_headers++;
3590         }
3591         unit_bytes += log->l_iclog_hsize * num_headers;
3592
3593         /* for commit-rec LR header - note: padding will subsume the ophdr */
3594         unit_bytes += log->l_iclog_hsize;
3595
3596         /* for roundoff padding for transaction data and one for commit record */
3597         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3598                 /* log su roundoff */
3599                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3600         } else {
3601                 /* BB roundoff */
3602                 unit_bytes += 2 * BBSIZE;
3603         }
3604
3605         return unit_bytes;
3606 }
3607
3608 /*
3609  * Allocate and initialise a new log ticket.
3610  */
3611 struct xlog_ticket *
3612 xlog_ticket_alloc(
3613         struct xlog             *log,
3614         int                     unit_bytes,
3615         int                     cnt,
3616         char                    client,
3617         bool                    permanent,
3618         xfs_km_flags_t          alloc_flags)
3619 {
3620         struct xlog_ticket      *tic;
3621         int                     unit_res;
3622
3623         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3624         if (!tic)
3625                 return NULL;
3626
3627         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3628
3629         atomic_set(&tic->t_ref, 1);
3630         tic->t_task             = current;
3631         INIT_LIST_HEAD(&tic->t_queue);
3632         tic->t_unit_res         = unit_res;
3633         tic->t_curr_res         = unit_res;
3634         tic->t_cnt              = cnt;
3635         tic->t_ocnt             = cnt;
3636         tic->t_tid              = prandom_u32();
3637         tic->t_clientid         = client;
3638         tic->t_flags            = XLOG_TIC_INITED;
3639         tic->t_trans_type       = 0;
3640         if (permanent)
3641                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3642
3643         xlog_tic_reset_res(tic);
3644
3645         return tic;
3646 }
3647
3648
3649 /******************************************************************************
3650  *
3651  *              Log debug routines
3652  *
3653  ******************************************************************************
3654  */
3655 #if defined(DEBUG)
3656 /*
3657  * Make sure that the destination ptr is within the valid data region of
3658  * one of the iclogs.  This uses backup pointers stored in a different
3659  * part of the log in case we trash the log structure.
3660  */
3661 void
3662 xlog_verify_dest_ptr(
3663         struct xlog     *log,
3664         void            *ptr)
3665 {
3666         int i;
3667         int good_ptr = 0;
3668
3669         for (i = 0; i < log->l_iclog_bufs; i++) {
3670                 if (ptr >= log->l_iclog_bak[i] &&
3671                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3672                         good_ptr++;
3673         }
3674
3675         if (!good_ptr)
3676                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3677 }
3678
3679 /*
3680  * Check to make sure the grant write head didn't just over lap the tail.  If
3681  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3682  * the cycles differ by exactly one and check the byte count.
3683  *
3684  * This check is run unlocked, so can give false positives. Rather than assert
3685  * on failures, use a warn-once flag and a panic tag to allow the admin to
3686  * determine if they want to panic the machine when such an error occurs. For
3687  * debug kernels this will have the same effect as using an assert but, unlinke
3688  * an assert, it can be turned off at runtime.
3689  */
3690 STATIC void
3691 xlog_verify_grant_tail(
3692         struct xlog     *log)
3693 {
3694         int             tail_cycle, tail_blocks;
3695         int             cycle, space;
3696
3697         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3698         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3699         if (tail_cycle != cycle) {
3700                 if (cycle - 1 != tail_cycle &&
3701                     !(log->l_flags & XLOG_TAIL_WARN)) {
3702                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3703                                 "%s: cycle - 1 != tail_cycle", __func__);
3704                         log->l_flags |= XLOG_TAIL_WARN;
3705                 }
3706
3707                 if (space > BBTOB(tail_blocks) &&
3708                     !(log->l_flags & XLOG_TAIL_WARN)) {
3709                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3710                                 "%s: space > BBTOB(tail_blocks)", __func__);
3711                         log->l_flags |= XLOG_TAIL_WARN;
3712                 }
3713         }
3714 }
3715
3716 /* check if it will fit */
3717 STATIC void
3718 xlog_verify_tail_lsn(
3719         struct xlog             *log,
3720         struct xlog_in_core     *iclog,
3721         xfs_lsn_t               tail_lsn)
3722 {
3723     int blocks;
3724
3725     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3726         blocks =
3727             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3728         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3729                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3730     } else {
3731         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3732
3733         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3734                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3735
3736         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3737         if (blocks < BTOBB(iclog->ic_offset) + 1)
3738                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3739     }
3740 }       /* xlog_verify_tail_lsn */
3741
3742 /*
3743  * Perform a number of checks on the iclog before writing to disk.
3744  *
3745  * 1. Make sure the iclogs are still circular
3746  * 2. Make sure we have a good magic number
3747  * 3. Make sure we don't have magic numbers in the data
3748  * 4. Check fields of each log operation header for:
3749  *      A. Valid client identifier
3750  *      B. tid ptr value falls in valid ptr space (user space code)
3751  *      C. Length in log record header is correct according to the
3752  *              individual operation headers within record.
3753  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3754  *      log, check the preceding blocks of the physical log to make sure all
3755  *      the cycle numbers agree with the current cycle number.
3756  */
3757 STATIC void
3758 xlog_verify_iclog(
3759         struct xlog             *log,
3760         struct xlog_in_core     *iclog,
3761         int                     count,
3762         bool                    syncing)
3763 {
3764         xlog_op_header_t        *ophead;
3765         xlog_in_core_t          *icptr;
3766         xlog_in_core_2_t        *xhdr;
3767         void                    *base_ptr, *ptr, *p;
3768         ptrdiff_t               field_offset;
3769         __uint8_t               clientid;
3770         int                     len, i, j, k, op_len;
3771         int                     idx;
3772
3773         /* check validity of iclog pointers */
3774         spin_lock(&log->l_icloglock);
3775         icptr = log->l_iclog;
3776         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3777                 ASSERT(icptr);
3778
3779         if (icptr != log->l_iclog)
3780                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3781         spin_unlock(&log->l_icloglock);
3782
3783         /* check log magic numbers */
3784         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3785                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3786
3787         base_ptr = ptr = &iclog->ic_header;
3788         p = &iclog->ic_header;
3789         for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3790                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3791                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3792                                 __func__);
3793         }
3794
3795         /* check fields */
3796         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3797         base_ptr = ptr = iclog->ic_datap;
3798         ophead = ptr;
3799         xhdr = iclog->ic_data;
3800         for (i = 0; i < len; i++) {
3801                 ophead = ptr;
3802
3803                 /* clientid is only 1 byte */
3804                 p = &ophead->oh_clientid;
3805                 field_offset = p - base_ptr;
3806                 if (!syncing || (field_offset & 0x1ff)) {
3807                         clientid = ophead->oh_clientid;
3808                 } else {
3809                         idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3810                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3811                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3812                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3813                                 clientid = xlog_get_client_id(
3814                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3815                         } else {
3816                                 clientid = xlog_get_client_id(
3817                                         iclog->ic_header.h_cycle_data[idx]);
3818                         }
3819                 }
3820                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3821                         xfs_warn(log->l_mp,
3822                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3823                                 __func__, clientid, ophead,
3824                                 (unsigned long)field_offset);
3825
3826                 /* check length */
3827                 p = &ophead->oh_len;
3828                 field_offset = p - base_ptr;
3829                 if (!syncing || (field_offset & 0x1ff)) {
3830                         op_len = be32_to_cpu(ophead->oh_len);
3831                 } else {
3832                         idx = BTOBBT((uintptr_t)&ophead->oh_len -
3833                                     (uintptr_t)iclog->ic_datap);
3834                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3835                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3836                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3837                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3838                         } else {
3839                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3840                         }
3841                 }
3842                 ptr += sizeof(xlog_op_header_t) + op_len;
3843         }
3844 }       /* xlog_verify_iclog */
3845 #endif
3846
3847 /*
3848  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3849  */
3850 STATIC int
3851 xlog_state_ioerror(
3852         struct xlog     *log)
3853 {
3854         xlog_in_core_t  *iclog, *ic;
3855
3856         iclog = log->l_iclog;
3857         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3858                 /*
3859                  * Mark all the incore logs IOERROR.
3860                  * From now on, no log flushes will result.
3861                  */
3862                 ic = iclog;
3863                 do {
3864                         ic->ic_state = XLOG_STATE_IOERROR;
3865                         ic = ic->ic_next;
3866                 } while (ic != iclog);
3867                 return 0;
3868         }
3869         /*
3870          * Return non-zero, if state transition has already happened.
3871          */
3872         return 1;
3873 }
3874
3875 /*
3876  * This is called from xfs_force_shutdown, when we're forcibly
3877  * shutting down the filesystem, typically because of an IO error.
3878  * Our main objectives here are to make sure that:
3879  *      a. if !logerror, flush the logs to disk. Anything modified
3880  *         after this is ignored.
3881  *      b. the filesystem gets marked 'SHUTDOWN' for all interested
3882  *         parties to find out, 'atomically'.
3883  *      c. those who're sleeping on log reservations, pinned objects and
3884  *          other resources get woken up, and be told the bad news.
3885  *      d. nothing new gets queued up after (b) and (c) are done.
3886  *
3887  * Note: for the !logerror case we need to flush the regions held in memory out
3888  * to disk first. This needs to be done before the log is marked as shutdown,
3889  * otherwise the iclog writes will fail.
3890  */
3891 int
3892 xfs_log_force_umount(
3893         struct xfs_mount        *mp,
3894         int                     logerror)
3895 {
3896         struct xlog     *log;
3897         int             retval;
3898
3899         log = mp->m_log;
3900
3901         /*
3902          * If this happens during log recovery, don't worry about
3903          * locking; the log isn't open for business yet.
3904          */
3905         if (!log ||
3906             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3907                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3908                 if (mp->m_sb_bp)
3909                         XFS_BUF_DONE(mp->m_sb_bp);
3910                 return 0;
3911         }
3912
3913         /*
3914          * Somebody could've already done the hard work for us.
3915          * No need to get locks for this.
3916          */
3917         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3918                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3919                 return 1;
3920         }
3921
3922         /*
3923          * Flush all the completed transactions to disk before marking the log
3924          * being shut down. We need to do it in this order to ensure that
3925          * completed operations are safely on disk before we shut down, and that
3926          * we don't have to issue any buffer IO after the shutdown flags are set
3927          * to guarantee this.
3928          */
3929         if (!logerror)
3930                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3931
3932         /*
3933          * mark the filesystem and the as in a shutdown state and wake
3934          * everybody up to tell them the bad news.
3935          */
3936         spin_lock(&log->l_icloglock);
3937         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3938         if (mp->m_sb_bp)
3939                 XFS_BUF_DONE(mp->m_sb_bp);
3940
3941         /*
3942          * Mark the log and the iclogs with IO error flags to prevent any
3943          * further log IO from being issued or completed.
3944          */
3945         log->l_flags |= XLOG_IO_ERROR;
3946         retval = xlog_state_ioerror(log);
3947         spin_unlock(&log->l_icloglock);
3948
3949         /*
3950          * We don't want anybody waiting for log reservations after this. That
3951          * means we have to wake up everybody queued up on reserveq as well as
3952          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3953          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3954          * action is protected by the grant locks.
3955          */
3956         xlog_grant_head_wake_all(&log->l_reserve_head);
3957         xlog_grant_head_wake_all(&log->l_write_head);
3958
3959         /*
3960          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3961          * as if the log writes were completed. The abort handling in the log
3962          * item committed callback functions will do this again under lock to
3963          * avoid races.
3964          */
3965         wake_up_all(&log->l_cilp->xc_commit_wait);
3966         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3967
3968 #ifdef XFSERRORDEBUG
3969         {
3970                 xlog_in_core_t  *iclog;
3971
3972                 spin_lock(&log->l_icloglock);
3973                 iclog = log->l_iclog;
3974                 do {
3975                         ASSERT(iclog->ic_callback == 0);
3976                         iclog = iclog->ic_next;
3977                 } while (iclog != log->l_iclog);
3978                 spin_unlock(&log->l_icloglock);
3979         }
3980 #endif
3981         /* return non-zero if log IOERROR transition had already happened */
3982         return retval;
3983 }
3984
3985 STATIC int
3986 xlog_iclogs_empty(
3987         struct xlog     *log)
3988 {
3989         xlog_in_core_t  *iclog;
3990
3991         iclog = log->l_iclog;
3992         do {
3993                 /* endianness does not matter here, zero is zero in
3994                  * any language.
3995                  */
3996                 if (iclog->ic_header.h_num_logops)
3997                         return 0;
3998                 iclog = iclog->ic_next;
3999         } while (iclog != log->l_iclog);
4000         return 1;
4001 }
4002