kernel/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(log->l_mp, 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(mp, 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(mp, xs_try_logspace);
 452
 453         ASSERT(*ticp == NULL);
 454         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 455                                 KM_SLEEP | KM_MAYFAIL);
 456         if (!tic)
 457                 return -ENOMEM;
 458
 459         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, "Log size out of supported range.");
 672                 xfs_crit(mp,
 673 "Continuing onwards, but if log hangs are 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                         xlog_recover_cancel(mp->m_log);
 704                         goto out_destroy_ail;
 705                 }
 706         }
 707
 708         error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 709                                "log");
 710         if (error)
 711                 goto out_destroy_ail;
 712
 713         /* Normal transactions can now occur */
 714         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 715
 716         /*
 717          * Now the log has been fully initialised and we know were our
 718          * space grant counters are, we can initialise the permanent ticket
 719          * needed for delayed logging to work.
 720          */
 721         xlog_cil_init_post_recovery(mp->m_log);
 722
 723         return 0;
 724
 725 out_destroy_ail:
 726         xfs_trans_ail_destroy(mp);
 727 out_free_log:
 728         xlog_dealloc_log(mp->m_log);
 729 out:
 730         return error;
 731 }
 732
 733 /*
 734  * Finish the recovery of the file system.  This is separate from the
 735  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 736  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 737  * here.
 738  *
 739  * If we finish recovery successfully, start the background log work. If we are
 740  * not doing recovery, then we have a RO filesystem and we don't need to start
 741  * it.
 742  */
 743 int
 744 xfs_log_mount_finish(
 745         struct xfs_mount        *mp)
 746 {
 747         int     error = 0;
 748
 749         if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
 750                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 751                 return 0;
 752         }
 753
 754         error = xlog_recover_finish(mp->m_log);
 755         if (!error)
 756                 xfs_log_work_queue(mp);
 757
 758         return error;
 759 }
 760
 761 /*
 762  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
 763  * the log.
 764  */
 765 int
 766 xfs_log_mount_cancel(
 767         struct xfs_mount        *mp)
 768 {
 769         int                     error;
 770
 771         error = xlog_recover_cancel(mp->m_log);
 772         xfs_log_unmount(mp);
 773
 774         return error;
 775 }
 776
 777 /*
 778  * Final log writes as part of unmount.
 779  *
 780  * Mark the filesystem clean as unmount happens.  Note that during relocation
 781  * this routine needs to be executed as part of source-bag while the
 782  * deallocation must not be done until source-end.
 783  */
 784
 785 /*
 786  * Unmount record used to have a string "Unmount filesystem--" in the
 787  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 788  * We just write the magic number now since that particular field isn't
 789  * currently architecture converted and "Unmount" is a bit foo.
 790  * As far as I know, there weren't any dependencies on the old behaviour.
 791  */
 792
 793 int
 794 xfs_log_unmount_write(xfs_mount_t *mp)
 795 {
 796         struct xlog      *log = mp->m_log;
 797         xlog_in_core_t   *iclog;
 798 #ifdef DEBUG
 799         xlog_in_core_t   *first_iclog;
 800 #endif
 801         xlog_ticket_t   *tic = NULL;
 802         xfs_lsn_t        lsn;
 803         int              error;
 804
 805         /*
 806          * Don't write out unmount record on read-only mounts.
 807          * Or, if we are doing a forced umount (typically because of IO errors).
 808          */
 809         if (mp->m_flags & XFS_MOUNT_RDONLY)
 810                 return 0;
 811
 812         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
 813         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 814
 815 #ifdef DEBUG
 816         first_iclog = iclog = log->l_iclog;
 817         do {
 818                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 819                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 820                         ASSERT(iclog->ic_offset == 0);
 821                 }
 822                 iclog = iclog->ic_next;
 823         } while (iclog != first_iclog);
 824 #endif
 825         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 826                 error = xfs_log_reserve(mp, 600, 1, &tic,
 827                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
 828                 if (!error) {
 829                         /* the data section must be 32 bit size aligned */
 830                         struct {
 831                             __uint16_t magic;
 832                             __uint16_t pad1;
 833                             __uint32_t pad2; /* may as well make it 64 bits */
 834                         } magic = {
 835                                 .magic = XLOG_UNMOUNT_TYPE,
 836                         };
 837                         struct xfs_log_iovec reg = {
 838                                 .i_addr = &magic,
 839                                 .i_len = sizeof(magic),
 840                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 841                         };
 842                         struct xfs_log_vec vec = {
 843                                 .lv_niovecs = 1,
 844                                 .lv_iovecp = &reg,
 845                         };
 846
 847                         /* remove inited flag, and account for space used */
 848                         tic->t_flags = 0;
 849                         tic->t_curr_res -= sizeof(magic);
 850                         error = xlog_write(log, &vec, tic, &lsn,
 851                                            NULL, XLOG_UNMOUNT_TRANS);
 852                         /*
 853                          * At this point, we're umounting anyway,
 854                          * so there's no point in transitioning log state
 855                          * to IOERROR. Just continue...
 856                          */
 857                 }
 858
 859                 if (error)
 860                         xfs_alert(mp, "%s: unmount record failed", __func__);
 861
 862
 863                 spin_lock(&log->l_icloglock);
 864                 iclog = log->l_iclog;
 865                 atomic_inc(&iclog->ic_refcnt);
 866                 xlog_state_want_sync(log, iclog);
 867                 spin_unlock(&log->l_icloglock);
 868                 error = xlog_state_release_iclog(log, iclog);
 869
 870                 spin_lock(&log->l_icloglock);
 871                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 872                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 873                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 874                                 xlog_wait(&iclog->ic_force_wait,
 875                                                         &log->l_icloglock);
 876                         } else {
 877                                 spin_unlock(&log->l_icloglock);
 878                         }
 879                 } else {
 880                         spin_unlock(&log->l_icloglock);
 881                 }
 882                 if (tic) {
 883                         trace_xfs_log_umount_write(log, tic);
 884                         xlog_ungrant_log_space(log, tic);
 885                         xfs_log_ticket_put(tic);
 886                 }
 887         } else {
 888                 /*
 889                  * We're already in forced_shutdown mode, couldn't
 890                  * even attempt to write out the unmount transaction.
 891                  *
 892                  * Go through the motions of sync'ing and releasing
 893                  * the iclog, even though no I/O will actually happen,
 894                  * we need to wait for other log I/Os that may already
 895                  * be in progress.  Do this as a separate section of
 896                  * code so we'll know if we ever get stuck here that
 897                  * we're in this odd situation of trying to unmount
 898                  * a file system that went into forced_shutdown as
 899                  * the result of an unmount..
 900                  */
 901                 spin_lock(&log->l_icloglock);
 902                 iclog = log->l_iclog;
 903                 atomic_inc(&iclog->ic_refcnt);
 904
 905                 xlog_state_want_sync(log, iclog);
 906                 spin_unlock(&log->l_icloglock);
 907                 error =  xlog_state_release_iclog(log, iclog);
 908
 909                 spin_lock(&log->l_icloglock);
 910
 911                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 912                         || iclog->ic_state == XLOG_STATE_DIRTY
 913                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 914
 915                                 xlog_wait(&iclog->ic_force_wait,
 916                                                         &log->l_icloglock);
 917                 } else {
 918                         spin_unlock(&log->l_icloglock);
 919                 }
 920         }
 921
 922         return error;
 923 }       /* xfs_log_unmount_write */
 924
 925 /*
 926  * Empty the log for unmount/freeze.
 927  *
 928  * To do this, we first need to shut down the background log work so it is not
 929  * trying to cover the log as we clean up. We then need to unpin all objects in
 930  * the log so we can then flush them out. Once they have completed their IO and
 931  * run the callbacks removing themselves from the AIL, we can write the unmount
 932  * record.
 933  */
 934 void
 935 xfs_log_quiesce(
 936         struct xfs_mount        *mp)
 937 {
 938         cancel_delayed_work_sync(&mp->m_log->l_work);
 939         xfs_log_force(mp, XFS_LOG_SYNC);
 940
 941         /*
 942          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
 943          * will push it, xfs_wait_buftarg() will not wait for it. Further,
 944          * xfs_buf_iowait() cannot be used because it was pushed with the
 945          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
 946          * the IO to complete.
 947          */
 948         xfs_ail_push_all_sync(mp->m_ail);
 949         xfs_wait_buftarg(mp->m_ddev_targp);
 950         xfs_buf_lock(mp->m_sb_bp);
 951         xfs_buf_unlock(mp->m_sb_bp);
 952
 953         xfs_log_unmount_write(mp);
 954 }
 955
 956 /*
 957  * Shut down and release the AIL and Log.
 958  *
 959  * During unmount, we need to ensure we flush all the dirty metadata objects
 960  * from the AIL so that the log is empty before we write the unmount record to
 961  * the log. Once this is done, we can tear down the AIL and the log.
 962  */
 963 void
 964 xfs_log_unmount(
 965         struct xfs_mount        *mp)
 966 {
 967         xfs_log_quiesce(mp);
 968
 969         xfs_trans_ail_destroy(mp);
 970
 971         xfs_sysfs_del(&mp->m_log->l_kobj);
 972
 973         xlog_dealloc_log(mp->m_log);
 974 }
 975
 976 void
 977 xfs_log_item_init(
 978         struct xfs_mount        *mp,
 979         struct xfs_log_item     *item,
 980         int                     type,
 981         const struct xfs_item_ops *ops)
 982 {
 983         item->li_mountp = mp;
 984         item->li_ailp = mp->m_ail;
 985         item->li_type = type;
 986         item->li_ops = ops;
 987         item->li_lv = NULL;
 988
 989         INIT_LIST_HEAD(&item->li_ail);
 990         INIT_LIST_HEAD(&item->li_cil);
 991 }
 992
 993 /*
 994  * Wake up processes waiting for log space after we have moved the log tail.
 995  */
 996 void
 997 xfs_log_space_wake(
 998         struct xfs_mount        *mp)
 999 {
1000         struct xlog             *log = mp->m_log;
1001         int                     free_bytes;
1002
1003         if (XLOG_FORCED_SHUTDOWN(log))
1004                 return;
1005
1006         if (!list_empty_careful(&log->l_write_head.waiters)) {
1007                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1008
1009                 spin_lock(&log->l_write_head.lock);
1010                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1011                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1012                 spin_unlock(&log->l_write_head.lock);
1013         }
1014
1015         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1016                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1017
1018                 spin_lock(&log->l_reserve_head.lock);
1019                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1020                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1021                 spin_unlock(&log->l_reserve_head.lock);
1022         }
1023 }
1024
1025 /*
1026  * Determine if we have a transaction that has gone to disk that needs to be
1027  * covered. To begin the transition to the idle state firstly the log needs to
1028  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1029  * we start attempting to cover the log.
1030  *
1031  * Only if we are then in a state where covering is needed, the caller is
1032  * informed that dummy transactions are required to move the log into the idle
1033  * state.
1034  *
1035  * If there are any items in the AIl or CIL, then we do not want to attempt to
1036  * cover the log as we may be in a situation where there isn't log space
1037  * available to run a dummy transaction and this can lead to deadlocks when the
1038  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1039  * there's no point in running a dummy transaction at this point because we
1040  * can't start trying to idle the log until both the CIL and AIL are empty.
1041  */
1042 int
1043 xfs_log_need_covered(xfs_mount_t *mp)
1044 {
1045         struct xlog     *log = mp->m_log;
1046         int             needed = 0;
1047
1048         if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1049                 return 0;
1050
1051         if (!xlog_cil_empty(log))
1052                 return 0;
1053
1054         spin_lock(&log->l_icloglock);
1055         switch (log->l_covered_state) {
1056         case XLOG_STATE_COVER_DONE:
1057         case XLOG_STATE_COVER_DONE2:
1058         case XLOG_STATE_COVER_IDLE:
1059                 break;
1060         case XLOG_STATE_COVER_NEED:
1061         case XLOG_STATE_COVER_NEED2:
1062                 if (xfs_ail_min_lsn(log->l_ailp))
1063                         break;
1064                 if (!xlog_iclogs_empty(log))
1065                         break;
1066
1067                 needed = 1;
1068                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1069                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1070                 else
1071                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1072                 break;
1073         default:
1074                 needed = 1;
1075                 break;
1076         }
1077         spin_unlock(&log->l_icloglock);
1078         return needed;
1079 }
1080
1081 /*
1082  * We may be holding the log iclog lock upon entering this routine.
1083  */
1084 xfs_lsn_t
1085 xlog_assign_tail_lsn_locked(
1086         struct xfs_mount        *mp)
1087 {
1088         struct xlog             *log = mp->m_log;
1089         struct xfs_log_item     *lip;
1090         xfs_lsn_t               tail_lsn;
1091
1092         assert_spin_locked(&mp->m_ail->xa_lock);
1093
1094         /*
1095          * To make sure we always have a valid LSN for the log tail we keep
1096          * track of the last LSN which was committed in log->l_last_sync_lsn,
1097          * and use that when the AIL was empty.
1098          */
1099         lip = xfs_ail_min(mp->m_ail);
1100         if (lip)
1101                 tail_lsn = lip->li_lsn;
1102         else
1103                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1104         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1105         atomic64_set(&log->l_tail_lsn, tail_lsn);
1106         return tail_lsn;
1107 }
1108
1109 xfs_lsn_t
1110 xlog_assign_tail_lsn(
1111         struct xfs_mount        *mp)
1112 {
1113         xfs_lsn_t               tail_lsn;
1114
1115         spin_lock(&mp->m_ail->xa_lock);
1116         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1117         spin_unlock(&mp->m_ail->xa_lock);
1118
1119         return tail_lsn;
1120 }
1121
1122 /*
1123  * Return the space in the log between the tail and the head.  The head
1124  * is passed in the cycle/bytes formal parms.  In the special case where
1125  * the reserve head has wrapped passed the tail, this calculation is no
1126  * longer valid.  In this case, just return 0 which means there is no space
1127  * in the log.  This works for all places where this function is called
1128  * with the reserve head.  Of course, if the write head were to ever
1129  * wrap the tail, we should blow up.  Rather than catch this case here,
1130  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1131  *
1132  * This code also handles the case where the reservation head is behind
1133  * the tail.  The details of this case are described below, but the end
1134  * result is that we return the size of the log as the amount of space left.
1135  */
1136 STATIC int
1137 xlog_space_left(
1138         struct xlog     *log,
1139         atomic64_t      *head)
1140 {
1141         int             free_bytes;
1142         int             tail_bytes;
1143         int             tail_cycle;
1144         int             head_cycle;
1145         int             head_bytes;
1146
1147         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1148         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1149         tail_bytes = BBTOB(tail_bytes);
1150         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1151                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1152         else if (tail_cycle + 1 < head_cycle)
1153                 return 0;
1154         else if (tail_cycle < head_cycle) {
1155                 ASSERT(tail_cycle == (head_cycle - 1));
1156                 free_bytes = tail_bytes - head_bytes;
1157         } else {
1158                 /*
1159                  * The reservation head is behind the tail.
1160                  * In this case we just want to return the size of the
1161                  * log as the amount of space left.
1162                  */
1163                 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1164                 xfs_alert(log->l_mp,
1165                           "  tail_cycle = %d, tail_bytes = %d",
1166                           tail_cycle, tail_bytes);
1167                 xfs_alert(log->l_mp,
1168                           "  GH   cycle = %d, GH   bytes = %d",
1169                           head_cycle, head_bytes);
1170                 ASSERT(0);
1171                 free_bytes = log->l_logsize;
1172         }
1173         return free_bytes;
1174 }
1175
1176
1177 /*
1178  * Log function which is called when an io completes.
1179  *
1180  * The log manager needs its own routine, in order to control what
1181  * happens with the buffer after the write completes.
1182  */
1183 void
1184 xlog_iodone(xfs_buf_t *bp)
1185 {
1186         struct xlog_in_core     *iclog = bp->b_fspriv;
1187         struct xlog             *l = iclog->ic_log;
1188         int                     aborted = 0;
1189
1190         /*
1191          * Race to shutdown the filesystem if we see an error.
1192          */
1193         if (XFS_TEST_ERROR(bp->b_error, l->l_mp,
1194                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1195                 xfs_buf_ioerror_alert(bp, __func__);
1196                 xfs_buf_stale(bp);
1197                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1198                 /*
1199                  * This flag will be propagated to the trans-committed
1200                  * callback routines to let them know that the log-commit
1201                  * didn't succeed.
1202                  */
1203                 aborted = XFS_LI_ABORTED;
1204         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1205                 aborted = XFS_LI_ABORTED;
1206         }
1207
1208         /* log I/O is always issued ASYNC */
1209         ASSERT(XFS_BUF_ISASYNC(bp));
1210         xlog_state_done_syncing(iclog, aborted);
1211
1212         /*
1213          * drop the buffer lock now that we are done. Nothing references
1214          * the buffer after this, so an unmount waiting on this lock can now
1215          * tear it down safely. As such, it is unsafe to reference the buffer
1216          * (bp) after the unlock as we could race with it being freed.
1217          */
1218         xfs_buf_unlock(bp);
1219 }
1220
1221 /*
1222  * Return size of each in-core log record buffer.
1223  *
1224  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1225  *
1226  * If the filesystem blocksize is too large, we may need to choose a
1227  * larger size since the directory code currently logs entire blocks.
1228  */
1229
1230 STATIC void
1231 xlog_get_iclog_buffer_size(
1232         struct xfs_mount        *mp,
1233         struct xlog             *log)
1234 {
1235         int size;
1236         int xhdrs;
1237
1238         if (mp->m_logbufs <= 0)
1239                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1240         else
1241                 log->l_iclog_bufs = mp->m_logbufs;
1242
1243         /*
1244          * Buffer size passed in from mount system call.
1245          */
1246         if (mp->m_logbsize > 0) {
1247                 size = log->l_iclog_size = mp->m_logbsize;
1248                 log->l_iclog_size_log = 0;
1249                 while (size != 1) {
1250                         log->l_iclog_size_log++;
1251                         size >>= 1;
1252                 }
1253
1254                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1255                         /* # headers = size / 32k
1256                          * one header holds cycles from 32k of data
1257                          */
1258
1259                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1260                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1261                                 xhdrs++;
1262                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1263                         log->l_iclog_heads = xhdrs;
1264                 } else {
1265                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1266                         log->l_iclog_hsize = BBSIZE;
1267                         log->l_iclog_heads = 1;
1268                 }
1269                 goto done;
1270         }
1271
1272         /* All machines use 32kB buffers by default. */
1273         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1274         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1275
1276         /* the default log size is 16k or 32k which is one header sector */
1277         log->l_iclog_hsize = BBSIZE;
1278         log->l_iclog_heads = 1;
1279
1280 done:
1281         /* are we being asked to make the sizes selected above visible? */
1282         if (mp->m_logbufs == 0)
1283                 mp->m_logbufs = log->l_iclog_bufs;
1284         if (mp->m_logbsize == 0)
1285                 mp->m_logbsize = log->l_iclog_size;
1286 }       /* xlog_get_iclog_buffer_size */
1287
1288
1289 void
1290 xfs_log_work_queue(
1291         struct xfs_mount        *mp)
1292 {
1293         queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1294                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1295 }
1296
1297 /*
1298  * Every sync period we need to unpin all items in the AIL and push them to
1299  * disk. If there is nothing dirty, then we might need to cover the log to
1300  * indicate that the filesystem is idle.
1301  */
1302 void
1303 xfs_log_worker(
1304         struct work_struct      *work)
1305 {
1306         struct xlog             *log = container_of(to_delayed_work(work),
1307                                                 struct xlog, l_work);
1308         struct xfs_mount        *mp = log->l_mp;
1309
1310         /* dgc: errors ignored - not fatal and nowhere to report them */
1311         if (xfs_log_need_covered(mp)) {
1312                 /*
1313                  * Dump a transaction into the log that contains no real change.
1314                  * This is needed to stamp the current tail LSN into the log
1315                  * during the covering operation.
1316                  *
1317                  * We cannot use an inode here for this - that will push dirty
1318                  * state back up into the VFS and then periodic inode flushing
1319                  * will prevent log covering from making progress. Hence we
1320                  * synchronously log the superblock instead to ensure the
1321                  * superblock is immediately unpinned and can be written back.
1322                  */
1323                 xfs_sync_sb(mp, true);
1324         } else
1325                 xfs_log_force(mp, 0);
1326
1327         /* start pushing all the metadata that is currently dirty */
1328         xfs_ail_push_all(mp->m_ail);
1329
1330         /* queue us up again */
1331         xfs_log_work_queue(mp);
1332 }
1333
1334 /*
1335  * This routine initializes some of the log structure for a given mount point.
1336  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1337  * some other stuff may be filled in too.
1338  */
1339 STATIC struct xlog *
1340 xlog_alloc_log(
1341         struct xfs_mount        *mp,
1342         struct xfs_buftarg      *log_target,
1343         xfs_daddr_t             blk_offset,
1344         int                     num_bblks)
1345 {
1346         struct xlog             *log;
1347         xlog_rec_header_t       *head;
1348         xlog_in_core_t          **iclogp;
1349         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1350         xfs_buf_t               *bp;
1351         int                     i;
1352         int                     error = -ENOMEM;
1353         uint                    log2_size = 0;
1354
1355         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1356         if (!log) {
1357                 xfs_warn(mp, "Log allocation failed: No memory!");
1358                 goto out;
1359         }
1360
1361         log->l_mp          = mp;
1362         log->l_targ        = log_target;
1363         log->l_logsize     = BBTOB(num_bblks);
1364         log->l_logBBstart  = blk_offset;
1365         log->l_logBBsize   = num_bblks;
1366         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1367         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1368         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1369
1370         log->l_prev_block  = -1;
1371         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1372         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1373         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1374         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1375
1376         xlog_grant_head_init(&log->l_reserve_head);
1377         xlog_grant_head_init(&log->l_write_head);
1378
1379         error = -EFSCORRUPTED;
1380         if (xfs_sb_version_hassector(&mp->m_sb)) {
1381                 log2_size = mp->m_sb.sb_logsectlog;
1382                 if (log2_size < BBSHIFT) {
1383                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1384                                 log2_size, BBSHIFT);
1385                         goto out_free_log;
1386                 }
1387
1388                 log2_size -= BBSHIFT;
1389                 if (log2_size > mp->m_sectbb_log) {
1390                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1391                                 log2_size, mp->m_sectbb_log);
1392                         goto out_free_log;
1393                 }
1394
1395                 /* for larger sector sizes, must have v2 or external log */
1396                 if (log2_size && log->l_logBBstart > 0 &&
1397                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1398                         xfs_warn(mp,
1399                 "log sector size (0x%x) invalid for configuration.",
1400                                 log2_size);
1401                         goto out_free_log;
1402                 }
1403         }
1404         log->l_sectBBsize = 1 << log2_size;
1405
1406         xlog_get_iclog_buffer_size(mp, log);
1407
1408         /*
1409          * Use a NULL block for the extra log buffer used during splits so that
1410          * it will trigger errors if we ever try to do IO on it without first
1411          * having set it up properly.
1412          */
1413         error = -ENOMEM;
1414         bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1415                            BTOBB(log->l_iclog_size), 0);
1416         if (!bp)
1417                 goto out_free_log;
1418
1419         /*
1420          * The iclogbuf buffer locks are held over IO but we are not going to do
1421          * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1422          * when appropriately.
1423          */
1424         ASSERT(xfs_buf_islocked(bp));
1425         xfs_buf_unlock(bp);
1426
1427         /* use high priority wq for log I/O completion */
1428         bp->b_ioend_wq = mp->m_log_workqueue;
1429         bp->b_iodone = xlog_iodone;
1430         log->l_xbuf = bp;
1431
1432         spin_lock_init(&log->l_icloglock);
1433         init_waitqueue_head(&log->l_flush_wait);
1434
1435         iclogp = &log->l_iclog;
1436         /*
1437          * The amount of memory to allocate for the iclog structure is
1438          * rather funky due to the way the structure is defined.  It is
1439          * done this way so that we can use different sizes for machines
1440          * with different amounts of memory.  See the definition of
1441          * xlog_in_core_t in xfs_log_priv.h for details.
1442          */
1443         ASSERT(log->l_iclog_size >= 4096);
1444         for (i=0; i < log->l_iclog_bufs; i++) {
1445                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1446                 if (!*iclogp)
1447                         goto out_free_iclog;
1448
1449                 iclog = *iclogp;
1450                 iclog->ic_prev = prev_iclog;
1451                 prev_iclog = iclog;
1452
1453                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1454                                                 BTOBB(log->l_iclog_size), 0);
1455                 if (!bp)
1456                         goto out_free_iclog;
1457
1458                 ASSERT(xfs_buf_islocked(bp));
1459                 xfs_buf_unlock(bp);
1460
1461                 /* use high priority wq for log I/O completion */
1462                 bp->b_ioend_wq = mp->m_log_workqueue;
1463                 bp->b_iodone = xlog_iodone;
1464                 iclog->ic_bp = bp;
1465                 iclog->ic_data = bp->b_addr;
1466 #ifdef DEBUG
1467                 log->l_iclog_bak[i] = &iclog->ic_header;
1468 #endif
1469                 head = &iclog->ic_header;
1470                 memset(head, 0, sizeof(xlog_rec_header_t));
1471                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1472                 head->h_version = cpu_to_be32(
1473                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1474                 head->h_size = cpu_to_be32(log->l_iclog_size);
1475                 /* new fields */
1476                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1477                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1478
1479                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1480                 iclog->ic_state = XLOG_STATE_ACTIVE;
1481                 iclog->ic_log = log;
1482                 atomic_set(&iclog->ic_refcnt, 0);
1483                 spin_lock_init(&iclog->ic_callback_lock);
1484                 iclog->ic_callback_tail = &(iclog->ic_callback);
1485                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1486
1487                 init_waitqueue_head(&iclog->ic_force_wait);
1488                 init_waitqueue_head(&iclog->ic_write_wait);
1489
1490                 iclogp = &iclog->ic_next;
1491         }
1492         *iclogp = log->l_iclog;                 /* complete ring */
1493         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1494
1495         error = xlog_cil_init(log);
1496         if (error)
1497                 goto out_free_iclog;
1498         return log;
1499
1500 out_free_iclog:
1501         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1502                 prev_iclog = iclog->ic_next;
1503                 if (iclog->ic_bp)
1504                         xfs_buf_free(iclog->ic_bp);
1505                 kmem_free(iclog);
1506         }
1507         spinlock_destroy(&log->l_icloglock);
1508         xfs_buf_free(log->l_xbuf);
1509 out_free_log:
1510         kmem_free(log);
1511 out:
1512         return ERR_PTR(error);
1513 }       /* xlog_alloc_log */
1514
1515
1516 /*
1517  * Write out the commit record of a transaction associated with the given
1518  * ticket.  Return the lsn of the commit record.
1519  */
1520 STATIC int
1521 xlog_commit_record(
1522         struct xlog             *log,
1523         struct xlog_ticket      *ticket,
1524         struct xlog_in_core     **iclog,
1525         xfs_lsn_t               *commitlsnp)
1526 {
1527         struct xfs_mount *mp = log->l_mp;
1528         int     error;
1529         struct xfs_log_iovec reg = {
1530                 .i_addr = NULL,
1531                 .i_len = 0,
1532                 .i_type = XLOG_REG_TYPE_COMMIT,
1533         };
1534         struct xfs_log_vec vec = {
1535                 .lv_niovecs = 1,
1536                 .lv_iovecp = &reg,
1537         };
1538
1539         ASSERT_ALWAYS(iclog);
1540         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1541                                         XLOG_COMMIT_TRANS);
1542         if (error)
1543                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1544         return error;
1545 }
1546
1547 /*
1548  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1549  * log space.  This code pushes on the lsn which would supposedly free up
1550  * the 25% which we want to leave free.  We may need to adopt a policy which
1551  * pushes on an lsn which is further along in the log once we reach the high
1552  * water mark.  In this manner, we would be creating a low water mark.
1553  */
1554 STATIC void
1555 xlog_grant_push_ail(
1556         struct xlog     *log,
1557         int             need_bytes)
1558 {
1559         xfs_lsn_t       threshold_lsn = 0;
1560         xfs_lsn_t       last_sync_lsn;
1561         int             free_blocks;
1562         int             free_bytes;
1563         int             threshold_block;
1564         int             threshold_cycle;
1565         int             free_threshold;
1566
1567         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1568
1569         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1570         free_blocks = BTOBBT(free_bytes);
1571
1572         /*
1573          * Set the threshold for the minimum number of free blocks in the
1574          * log to the maximum of what the caller needs, one quarter of the
1575          * log, and 256 blocks.
1576          */
1577         free_threshold = BTOBB(need_bytes);
1578         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1579         free_threshold = MAX(free_threshold, 256);
1580         if (free_blocks >= free_threshold)
1581                 return;
1582
1583         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1584                                                 &threshold_block);
1585         threshold_block += free_threshold;
1586         if (threshold_block >= log->l_logBBsize) {
1587                 threshold_block -= log->l_logBBsize;
1588                 threshold_cycle += 1;
1589         }
1590         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1591                                         threshold_block);
1592         /*
1593          * Don't pass in an lsn greater than the lsn of the last
1594          * log record known to be on disk. Use a snapshot of the last sync lsn
1595          * so that it doesn't change between the compare and the set.
1596          */
1597         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1598         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1599                 threshold_lsn = last_sync_lsn;
1600
1601         /*
1602          * Get the transaction layer to kick the dirty buffers out to
1603          * disk asynchronously. No point in trying to do this if
1604          * the filesystem is shutting down.
1605          */
1606         if (!XLOG_FORCED_SHUTDOWN(log))
1607                 xfs_ail_push(log->l_ailp, threshold_lsn);
1608 }
1609
1610 /*
1611  * Stamp cycle number in every block
1612  */
1613 STATIC void
1614 xlog_pack_data(
1615         struct xlog             *log,
1616         struct xlog_in_core     *iclog,
1617         int                     roundoff)
1618 {
1619         int                     i, j, k;
1620         int                     size = iclog->ic_offset + roundoff;
1621         __be32                  cycle_lsn;
1622         char                    *dp;
1623
1624         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1625
1626         dp = iclog->ic_datap;
1627         for (i = 0; i < BTOBB(size); i++) {
1628                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1629                         break;
1630                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1631                 *(__be32 *)dp = cycle_lsn;
1632                 dp += BBSIZE;
1633         }
1634
1635         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1636                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1637
1638                 for ( ; i < BTOBB(size); i++) {
1639                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1640                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1641                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1642                         *(__be32 *)dp = cycle_lsn;
1643                         dp += BBSIZE;
1644                 }
1645
1646                 for (i = 1; i < log->l_iclog_heads; i++)
1647                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1648         }
1649 }
1650
1651 /*
1652  * Calculate the checksum for a log buffer.
1653  *
1654  * This is a little more complicated than it should be because the various
1655  * headers and the actual data are non-contiguous.
1656  */
1657 __le32
1658 xlog_cksum(
1659         struct xlog             *log,
1660         struct xlog_rec_header  *rhead,
1661         char                    *dp,
1662         int                     size)
1663 {
1664         __uint32_t              crc;
1665
1666         /* first generate the crc for the record header ... */
1667         crc = xfs_start_cksum((char *)rhead,
1668                               sizeof(struct xlog_rec_header),
1669                               offsetof(struct xlog_rec_header, h_crc));
1670
1671         /* ... then for additional cycle data for v2 logs ... */
1672         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1673                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1674                 int             i;
1675                 int             xheads;
1676
1677                 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1678                 if (size % XLOG_HEADER_CYCLE_SIZE)
1679                         xheads++;
1680
1681                 for (i = 1; i < xheads; i++) {
1682                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1683                                      sizeof(struct xlog_rec_ext_header));
1684                 }
1685         }
1686
1687         /* ... and finally for the payload */
1688         crc = crc32c(crc, dp, size);
1689
1690         return xfs_end_cksum(crc);
1691 }
1692
1693 /*
1694  * The bdstrat callback function for log bufs. This gives us a central
1695  * place to trap bufs in case we get hit by a log I/O error and need to
1696  * shutdown. Actually, in practice, even when we didn't get a log error,
1697  * we transition the iclogs to IOERROR state *after* flushing all existing
1698  * iclogs to disk. This is because we don't want anymore new transactions to be
1699  * started or completed afterwards.
1700  *
1701  * We lock the iclogbufs here so that we can serialise against IO completion
1702  * during unmount. We might be processing a shutdown triggered during unmount,
1703  * and that can occur asynchronously to the unmount thread, and hence we need to
1704  * ensure that completes before tearing down the iclogbufs. Hence we need to
1705  * hold the buffer lock across the log IO to acheive that.
1706  */
1707 STATIC int
1708 xlog_bdstrat(
1709         struct xfs_buf          *bp)
1710 {
1711         struct xlog_in_core     *iclog = bp->b_fspriv;
1712
1713         xfs_buf_lock(bp);
1714         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1715                 xfs_buf_ioerror(bp, -EIO);
1716                 xfs_buf_stale(bp);
1717                 xfs_buf_ioend(bp);
1718                 /*
1719                  * It would seem logical to return EIO here, but we rely on
1720                  * the log state machine to propagate I/O errors instead of
1721                  * doing it here. Similarly, IO completion will unlock the
1722                  * buffer, so we don't do it here.
1723                  */
1724                 return 0;
1725         }
1726
1727         xfs_buf_submit(bp);
1728         return 0;
1729 }
1730
1731 /*
1732  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1733  * fashion.  Previously, we should have moved the current iclog
1734  * ptr in the log to point to the next available iclog.  This allows further
1735  * write to continue while this code syncs out an iclog ready to go.
1736  * Before an in-core log can be written out, the data section must be scanned
1737  * to save away the 1st word of each BBSIZE block into the header.  We replace
1738  * it with the current cycle count.  Each BBSIZE block is tagged with the
1739  * cycle count because there in an implicit assumption that drives will
1740  * guarantee that entire 512 byte blocks get written at once.  In other words,
1741  * we can't have part of a 512 byte block written and part not written.  By
1742  * tagging each block, we will know which blocks are valid when recovering
1743  * after an unclean shutdown.
1744  *
1745  * This routine is single threaded on the iclog.  No other thread can be in
1746  * this routine with the same iclog.  Changing contents of iclog can there-
1747  * fore be done without grabbing the state machine lock.  Updating the global
1748  * log will require grabbing the lock though.
1749  *
1750  * The entire log manager uses a logical block numbering scheme.  Only
1751  * log_sync (and then only bwrite()) know about the fact that the log may
1752  * not start with block zero on a given device.  The log block start offset
1753  * is added immediately before calling bwrite().
1754  */
1755
1756 STATIC int
1757 xlog_sync(
1758         struct xlog             *log,
1759         struct xlog_in_core     *iclog)
1760 {
1761         xfs_buf_t       *bp;
1762         int             i;
1763         uint            count;          /* byte count of bwrite */
1764         uint            count_init;     /* initial count before roundup */
1765         int             roundoff;       /* roundoff to BB or stripe */
1766         int             split = 0;      /* split write into two regions */
1767         int             error;
1768         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1769         int             size;
1770
1771         XFS_STATS_INC(log->l_mp, xs_log_writes);
1772         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1773
1774         /* Add for LR header */
1775         count_init = log->l_iclog_hsize + iclog->ic_offset;
1776
1777         /* Round out the log write size */
1778         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1779                 /* we have a v2 stripe unit to use */
1780                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1781         } else {
1782                 count = BBTOB(BTOBB(count_init));
1783         }
1784         roundoff = count - count_init;
1785         ASSERT(roundoff >= 0);
1786         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1787                 roundoff < log->l_mp->m_sb.sb_logsunit)
1788                 ||
1789                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1790                  roundoff < BBTOB(1)));
1791
1792         /* move grant heads by roundoff in sync */
1793         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1794         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1795
1796         /* put cycle number in every block */
1797         xlog_pack_data(log, iclog, roundoff);
1798
1799         /* real byte length */
1800         size = iclog->ic_offset;
1801         if (v2)
1802                 size += roundoff;
1803         iclog->ic_header.h_len = cpu_to_be32(size);
1804
1805         bp = iclog->ic_bp;
1806         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1807
1808         XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1809
1810         /* Do we need to split this write into 2 parts? */
1811         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1812                 char            *dptr;
1813
1814                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1815                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1816                 iclog->ic_bwritecnt = 2;
1817
1818                 /*
1819                  * Bump the cycle numbers at the start of each block in the
1820                  * part of the iclog that ends up in the buffer that gets
1821                  * written to the start of the log.
1822                  *
1823                  * Watch out for the header magic number case, though.
1824                  */
1825                 dptr = (char *)&iclog->ic_header + count;
1826                 for (i = 0; i < split; i += BBSIZE) {
1827                         __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1828                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1829                                 cycle++;
1830                         *(__be32 *)dptr = cpu_to_be32(cycle);
1831
1832                         dptr += BBSIZE;
1833                 }
1834         } else {
1835                 iclog->ic_bwritecnt = 1;
1836         }
1837
1838         /* calculcate the checksum */
1839         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1840                                             iclog->ic_datap, size);
1841
1842         bp->b_io_length = BTOBB(count);
1843         bp->b_fspriv = iclog;
1844         XFS_BUF_ZEROFLAGS(bp);
1845         XFS_BUF_ASYNC(bp);
1846         bp->b_flags |= XBF_SYNCIO;
1847
1848         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1849                 bp->b_flags |= XBF_FUA;
1850
1851                 /*
1852                  * Flush the data device before flushing the log to make
1853                  * sure all meta data written back from the AIL actually made
1854                  * it to disk before stamping the new log tail LSN into the
1855                  * log buffer.  For an external log we need to issue the
1856                  * flush explicitly, and unfortunately synchronously here;
1857                  * for an internal log we can simply use the block layer
1858                  * state machine for preflushes.
1859                  */
1860                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1861                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1862                 else
1863                         bp->b_flags |= XBF_FLUSH;
1864         }
1865
1866         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1867         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1868
1869         xlog_verify_iclog(log, iclog, count, true);
1870
1871         /* account for log which doesn't start at block #0 */
1872         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1873         /*
1874          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1875          * is shutting down.
1876          */
1877         XFS_BUF_WRITE(bp);
1878
1879         error = xlog_bdstrat(bp);
1880         if (error) {
1881                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1882                 return error;
1883         }
1884         if (split) {
1885                 bp = iclog->ic_log->l_xbuf;
1886                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1887                 xfs_buf_associate_memory(bp,
1888                                 (char *)&iclog->ic_header + count, split);
1889                 bp->b_fspriv = iclog;
1890                 XFS_BUF_ZEROFLAGS(bp);
1891                 XFS_BUF_ASYNC(bp);
1892                 bp->b_flags |= XBF_SYNCIO;
1893                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1894                         bp->b_flags |= XBF_FUA;
1895
1896                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1897                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1898
1899                 /* account for internal log which doesn't start at block #0 */
1900                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1901                 XFS_BUF_WRITE(bp);
1902                 error = xlog_bdstrat(bp);
1903                 if (error) {
1904                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1905                         return error;
1906                 }
1907         }
1908         return 0;
1909 }       /* xlog_sync */
1910
1911 /*
1912  * Deallocate a log structure
1913  */
1914 STATIC void
1915 xlog_dealloc_log(
1916         struct xlog     *log)
1917 {
1918         xlog_in_core_t  *iclog, *next_iclog;
1919         int             i;
1920
1921         xlog_cil_destroy(log);
1922
1923         /*
1924          * Cycle all the iclogbuf locks to make sure all log IO completion
1925          * is done before we tear down these buffers.
1926          */
1927         iclog = log->l_iclog;
1928         for (i = 0; i < log->l_iclog_bufs; i++) {
1929                 xfs_buf_lock(iclog->ic_bp);
1930                 xfs_buf_unlock(iclog->ic_bp);
1931                 iclog = iclog->ic_next;
1932         }
1933
1934         /*
1935          * Always need to ensure that the extra buffer does not point to memory
1936          * owned by another log buffer before we free it. Also, cycle the lock
1937          * first to ensure we've completed IO on it.
1938          */
1939         xfs_buf_lock(log->l_xbuf);
1940         xfs_buf_unlock(log->l_xbuf);
1941         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1942         xfs_buf_free(log->l_xbuf);
1943
1944         iclog = log->l_iclog;
1945         for (i = 0; i < log->l_iclog_bufs; i++) {
1946                 xfs_buf_free(iclog->ic_bp);
1947                 next_iclog = iclog->ic_next;
1948                 kmem_free(iclog);
1949                 iclog = next_iclog;
1950         }
1951         spinlock_destroy(&log->l_icloglock);
1952
1953         log->l_mp->m_log = NULL;
1954         kmem_free(log);
1955 }       /* xlog_dealloc_log */
1956
1957 /*
1958  * Update counters atomically now that memcpy is done.
1959  */
1960 /* ARGSUSED */
1961 static inline void
1962 xlog_state_finish_copy(
1963         struct xlog             *log,
1964         struct xlog_in_core     *iclog,
1965         int                     record_cnt,
1966         int                     copy_bytes)
1967 {
1968         spin_lock(&log->l_icloglock);
1969
1970         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1971         iclog->ic_offset += copy_bytes;
1972
1973         spin_unlock(&log->l_icloglock);
1974 }       /* xlog_state_finish_copy */
1975
1976
1977
1978
1979 /*
1980  * print out info relating to regions written which consume
1981  * the reservation
1982  */
1983 void
1984 xlog_print_tic_res(
1985         struct xfs_mount        *mp,
1986         struct xlog_ticket      *ticket)
1987 {
1988         uint i;
1989         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1990
1991         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1992         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1993             "bformat",
1994             "bchunk",
1995             "efi_format",
1996             "efd_format",
1997             "iformat",
1998             "icore",
1999             "iext",
2000             "ibroot",
2001             "ilocal",
2002             "iattr_ext",
2003             "iattr_broot",
2004             "iattr_local",
2005             "qformat",
2006             "dquot",
2007             "quotaoff",
2008             "LR header",
2009             "unmount",
2010             "commit",
2011             "trans header"
2012         };
2013         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
2014             "SETATTR_NOT_SIZE",
2015             "SETATTR_SIZE",
2016             "INACTIVE",
2017             "CREATE",
2018             "CREATE_TRUNC",
2019             "TRUNCATE_FILE",
2020             "REMOVE",
2021             "LINK",
2022             "RENAME",
2023             "MKDIR",
2024             "RMDIR",
2025             "SYMLINK",
2026             "SET_DMATTRS",
2027             "GROWFS",
2028             "STRAT_WRITE",
2029             "DIOSTRAT",
2030             "WRITE_SYNC",
2031             "WRITEID",
2032             "ADDAFORK",
2033             "ATTRINVAL",
2034             "ATRUNCATE",
2035             "ATTR_SET",
2036             "ATTR_RM",
2037             "ATTR_FLAG",
2038             "CLEAR_AGI_BUCKET",
2039             "QM_SBCHANGE",
2040             "DUMMY1",
2041             "DUMMY2",
2042             "QM_QUOTAOFF",
2043             "QM_DQALLOC",
2044             "QM_SETQLIM",
2045             "QM_DQCLUSTER",
2046             "QM_QINOCREATE",
2047             "QM_QUOTAOFF_END",
2048             "SB_UNIT",
2049             "FSYNC_TS",
2050             "GROWFSRT_ALLOC",
2051             "GROWFSRT_ZERO",
2052             "GROWFSRT_FREE",
2053             "SWAPEXT"
2054         };
2055
2056         xfs_warn(mp, "xlog_write: reservation summary:");
2057         xfs_warn(mp, "  trans type  = %s (%u)",
2058                  ((ticket->t_trans_type <= 0 ||
2059                    ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2060                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
2061                  ticket->t_trans_type);
2062         xfs_warn(mp, "  unit res    = %d bytes",
2063                  ticket->t_unit_res);
2064         xfs_warn(mp, "  current res = %d bytes",
2065                  ticket->t_curr_res);
2066         xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2067                  ticket->t_res_arr_sum, ticket->t_res_o_flow);
2068         xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2069                  ticket->t_res_num_ophdrs, ophdr_spc);
2070         xfs_warn(mp, "  ophdr + reg = %u bytes",
2071                  ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2072         xfs_warn(mp, "  num regions = %u",
2073                  ticket->t_res_num);
2074
2075         for (i = 0; i < ticket->t_res_num; i++) {
2076                 uint r_type = ticket->t_res_arr[i].r_type;
2077                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2078                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2079                             "bad-rtype" : res_type_str[r_type-1]),
2080                             ticket->t_res_arr[i].r_len);
2081         }
2082
2083         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2084                 "xlog_write: reservation ran out. Need to up reservation");
2085         xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2086 }
2087
2088 /*
2089  * Calculate the potential space needed by the log vector.  Each region gets
2090  * its own xlog_op_header_t and may need to be double word aligned.
2091  */
2092 static int
2093 xlog_write_calc_vec_length(
2094         struct xlog_ticket      *ticket,
2095         struct xfs_log_vec      *log_vector)
2096 {
2097         struct xfs_log_vec      *lv;
2098         int                     headers = 0;
2099         int                     len = 0;
2100         int                     i;
2101
2102         /* acct for start rec of xact */
2103         if (ticket->t_flags & XLOG_TIC_INITED)
2104                 headers++;
2105
2106         for (lv = log_vector; lv; lv = lv->lv_next) {
2107                 /* we don't write ordered log vectors */
2108                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2109                         continue;
2110
2111                 headers += lv->lv_niovecs;
2112
2113                 for (i = 0; i < lv->lv_niovecs; i++) {
2114                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2115
2116                         len += vecp->i_len;
2117                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2118                 }
2119         }
2120
2121         ticket->t_res_num_ophdrs += headers;
2122         len += headers * sizeof(struct xlog_op_header);
2123
2124         return len;
2125 }
2126
2127 /*
2128  * If first write for transaction, insert start record  We can't be trying to
2129  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2130  */
2131 static int
2132 xlog_write_start_rec(
2133         struct xlog_op_header   *ophdr,
2134         struct xlog_ticket      *ticket)
2135 {
2136         if (!(ticket->t_flags & XLOG_TIC_INITED))
2137                 return 0;
2138
2139         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2140         ophdr->oh_clientid = ticket->t_clientid;
2141         ophdr->oh_len = 0;
2142         ophdr->oh_flags = XLOG_START_TRANS;
2143         ophdr->oh_res2 = 0;
2144
2145         ticket->t_flags &= ~XLOG_TIC_INITED;
2146
2147         return sizeof(struct xlog_op_header);
2148 }
2149
2150 static xlog_op_header_t *
2151 xlog_write_setup_ophdr(
2152         struct xlog             *log,
2153         struct xlog_op_header   *ophdr,
2154         struct xlog_ticket      *ticket,
2155         uint                    flags)
2156 {
2157         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2158         ophdr->oh_clientid = ticket->t_clientid;
2159         ophdr->oh_res2 = 0;
2160
2161         /* are we copying a commit or unmount record? */
2162         ophdr->oh_flags = flags;
2163
2164         /*
2165          * We've seen logs corrupted with bad transaction client ids.  This
2166          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2167          * and shut down the filesystem.
2168          */
2169         switch (ophdr->oh_clientid)  {
2170         case XFS_TRANSACTION:
2171         case XFS_VOLUME:
2172         case XFS_LOG:
2173                 break;
2174         default:
2175                 xfs_warn(log->l_mp,
2176                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2177                         ophdr->oh_clientid, ticket);
2178                 return NULL;
2179         }
2180
2181         return ophdr;
2182 }
2183
2184 /*
2185  * Set up the parameters of the region copy into the log. This has
2186  * to handle region write split across multiple log buffers - this
2187  * state is kept external to this function so that this code can
2188  * be written in an obvious, self documenting manner.
2189  */
2190 static int
2191 xlog_write_setup_copy(
2192         struct xlog_ticket      *ticket,
2193         struct xlog_op_header   *ophdr,
2194         int                     space_available,
2195         int                     space_required,
2196         int                     *copy_off,
2197         int                     *copy_len,
2198         int                     *last_was_partial_copy,
2199         int                     *bytes_consumed)
2200 {
2201         int                     still_to_copy;
2202
2203         still_to_copy = space_required - *bytes_consumed;
2204         *copy_off = *bytes_consumed;
2205
2206         if (still_to_copy <= space_available) {
2207                 /* write of region completes here */
2208                 *copy_len = still_to_copy;
2209                 ophdr->oh_len = cpu_to_be32(*copy_len);
2210                 if (*last_was_partial_copy)
2211                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2212                 *last_was_partial_copy = 0;
2213                 *bytes_consumed = 0;
2214                 return 0;
2215         }
2216
2217         /* partial write of region, needs extra log op header reservation */
2218         *copy_len = space_available;
2219         ophdr->oh_len = cpu_to_be32(*copy_len);
2220         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2221         if (*last_was_partial_copy)
2222                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2223         *bytes_consumed += *copy_len;
2224         (*last_was_partial_copy)++;
2225
2226         /* account for new log op header */
2227         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2228         ticket->t_res_num_ophdrs++;
2229
2230         return sizeof(struct xlog_op_header);
2231 }
2232
2233 static int
2234 xlog_write_copy_finish(
2235         struct xlog             *log,
2236         struct xlog_in_core     *iclog,
2237         uint                    flags,
2238         int                     *record_cnt,
2239         int                     *data_cnt,
2240         int                     *partial_copy,
2241         int                     *partial_copy_len,
2242         int                     log_offset,
2243         struct xlog_in_core     **commit_iclog)
2244 {
2245         if (*partial_copy) {
2246                 /*
2247                  * This iclog has already been marked WANT_SYNC by
2248                  * xlog_state_get_iclog_space.
2249                  */
2250                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2251                 *record_cnt = 0;
2252                 *data_cnt = 0;
2253                 return xlog_state_release_iclog(log, iclog);
2254         }
2255
2256         *partial_copy = 0;
2257         *partial_copy_len = 0;
2258
2259         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2260                 /* no more space in this iclog - push it. */
2261                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2262                 *record_cnt = 0;
2263                 *data_cnt = 0;
2264
2265                 spin_lock(&log->l_icloglock);
2266                 xlog_state_want_sync(log, iclog);
2267                 spin_unlock(&log->l_icloglock);
2268
2269                 if (!commit_iclog)
2270                         return xlog_state_release_iclog(log, iclog);
2271                 ASSERT(flags & XLOG_COMMIT_TRANS);
2272                 *commit_iclog = iclog;
2273         }
2274
2275         return 0;
2276 }
2277
2278 /*
2279  * Write some region out to in-core log
2280  *
2281  * This will be called when writing externally provided regions or when
2282  * writing out a commit record for a given transaction.
2283  *
2284  * General algorithm:
2285  *      1. Find total length of this write.  This may include adding to the
2286  *              lengths passed in.
2287  *      2. Check whether we violate the tickets reservation.
2288  *      3. While writing to this iclog
2289  *          A. Reserve as much space in this iclog as can get
2290  *          B. If this is first write, save away start lsn
2291  *          C. While writing this region:
2292  *              1. If first write of transaction, write start record
2293  *              2. Write log operation header (header per region)
2294  *              3. Find out if we can fit entire region into this iclog
2295  *              4. Potentially, verify destination memcpy ptr
2296  *              5. Memcpy (partial) region
2297  *              6. If partial copy, release iclog; otherwise, continue
2298  *                      copying more regions into current iclog
2299  *      4. Mark want sync bit (in simulation mode)
2300  *      5. Release iclog for potential flush to on-disk log.
2301  *
2302  * ERRORS:
2303  * 1.   Panic if reservation is overrun.  This should never happen since
2304  *      reservation amounts are generated internal to the filesystem.
2305  * NOTES:
2306  * 1. Tickets are single threaded data structures.
2307  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2308  *      syncing routine.  When a single log_write region needs to span
2309  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2310  *      on all log operation writes which don't contain the end of the
2311  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2312  *      operation which contains the end of the continued log_write region.
2313  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2314  *      we don't really know exactly how much space will be used.  As a result,
2315  *      we don't update ic_offset until the end when we know exactly how many
2316  *      bytes have been written out.
2317  */
2318 int
2319 xlog_write(
2320         struct xlog             *log,
2321         struct xfs_log_vec      *log_vector,
2322         struct xlog_ticket      *ticket,
2323         xfs_lsn_t               *start_lsn,
2324         struct xlog_in_core     **commit_iclog,
2325         uint                    flags)
2326 {
2327         struct xlog_in_core     *iclog = NULL;
2328         struct xfs_log_iovec    *vecp;
2329         struct xfs_log_vec      *lv;
2330         int                     len;
2331         int                     index;
2332         int                     partial_copy = 0;
2333         int                     partial_copy_len = 0;
2334         int                     contwr = 0;
2335         int                     record_cnt = 0;
2336         int                     data_cnt = 0;
2337         int                     error;
2338
2339         *start_lsn = 0;
2340
2341         len = xlog_write_calc_vec_length(ticket, log_vector);
2342
2343         /*
2344          * Region headers and bytes are already accounted for.
2345          * We only need to take into account start records and
2346          * split regions in this function.
2347          */
2348         if (ticket->t_flags & XLOG_TIC_INITED)
2349                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2350
2351         /*
2352          * Commit record headers need to be accounted for. These
2353          * come in as separate writes so are easy to detect.
2354          */
2355         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2356                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2357
2358         if (ticket->t_curr_res < 0)
2359                 xlog_print_tic_res(log->l_mp, ticket);
2360
2361         index = 0;
2362         lv = log_vector;
2363         vecp = lv->lv_iovecp;
2364         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2365                 void            *ptr;
2366                 int             log_offset;
2367
2368                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2369                                                    &contwr, &log_offset);
2370                 if (error)
2371                         return error;
2372
2373                 ASSERT(log_offset <= iclog->ic_size - 1);
2374                 ptr = iclog->ic_datap + log_offset;
2375
2376                 /* start_lsn is the first lsn written to. That's all we need. */
2377                 if (!*start_lsn)
2378                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2379
2380                 /*
2381                  * This loop writes out as many regions as can fit in the amount
2382                  * of space which was allocated by xlog_state_get_iclog_space().
2383                  */
2384                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2385                         struct xfs_log_iovec    *reg;
2386                         struct xlog_op_header   *ophdr;
2387                         int                     start_rec_copy;
2388                         int                     copy_len;
2389                         int                     copy_off;
2390                         bool                    ordered = false;
2391
2392                         /* ordered log vectors have no regions to write */
2393                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2394                                 ASSERT(lv->lv_niovecs == 0);
2395                                 ordered = true;
2396                                 goto next_lv;
2397                         }
2398
2399                         reg = &vecp[index];
2400                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2401                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2402
2403                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2404                         if (start_rec_copy) {
2405                                 record_cnt++;
2406                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2407                                                    start_rec_copy);
2408                         }
2409
2410                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2411                         if (!ophdr)
2412                                 return -EIO;
2413
2414                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2415                                            sizeof(struct xlog_op_header));
2416
2417                         len += xlog_write_setup_copy(ticket, ophdr,
2418                                                      iclog->ic_size-log_offset,
2419                                                      reg->i_len,
2420                                                      &copy_off, &copy_len,
2421                                                      &partial_copy,
2422                                                      &partial_copy_len);
2423                         xlog_verify_dest_ptr(log, ptr);
2424
2425                         /*
2426                          * Copy region.
2427                          *
2428                          * Unmount records just log an opheader, so can have
2429                          * empty payloads with no data region to copy. Hence we
2430                          * only copy the payload if the vector says it has data
2431                          * to copy.
2432                          */
2433                         ASSERT(copy_len >= 0);
2434                         if (copy_len > 0) {
2435                                 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2436                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2437                                                    copy_len);
2438                         }
2439                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2440                         record_cnt++;
2441                         data_cnt += contwr ? copy_len : 0;
2442
2443                         error = xlog_write_copy_finish(log, iclog, flags,
2444                                                        &record_cnt, &data_cnt,
2445                                                        &partial_copy,
2446                                                        &partial_copy_len,
2447                                                        log_offset,
2448                                                        commit_iclog);
2449                         if (error)
2450                                 return error;
2451
2452                         /*
2453                          * if we had a partial copy, we need to get more iclog
2454                          * space but we don't want to increment the region
2455                          * index because there is still more is this region to
2456                          * write.
2457                          *
2458                          * If we completed writing this region, and we flushed
2459                          * the iclog (indicated by resetting of the record
2460                          * count), then we also need to get more log space. If
2461                          * this was the last record, though, we are done and
2462                          * can just return.
2463                          */
2464                         if (partial_copy)
2465                                 break;
2466
2467                         if (++index == lv->lv_niovecs) {
2468 next_lv:
2469                                 lv = lv->lv_next;
2470                                 index = 0;
2471                                 if (lv)
2472                                         vecp = lv->lv_iovecp;
2473                         }
2474                         if (record_cnt == 0 && ordered == false) {
2475                                 if (!lv)
2476                                         return 0;
2477                                 break;
2478                         }
2479                 }
2480         }
2481
2482         ASSERT(len == 0);
2483
2484         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2485         if (!commit_iclog)
2486                 return xlog_state_release_iclog(log, iclog);
2487
2488         ASSERT(flags & XLOG_COMMIT_TRANS);
2489         *commit_iclog = iclog;
2490         return 0;
2491 }
2492
2493
2494 /*****************************************************************************
2495  *
2496  *              State Machine functions
2497  *
2498  *****************************************************************************
2499  */
2500
2501 /* Clean iclogs starting from the head.  This ordering must be
2502  * maintained, so an iclog doesn't become ACTIVE beyond one that
2503  * is SYNCING.  This is also required to maintain the notion that we use
2504  * a ordered wait queue to hold off would be writers to the log when every
2505  * iclog is trying to sync to disk.
2506  *
2507  * State Change: DIRTY -> ACTIVE
2508  */
2509 STATIC void
2510 xlog_state_clean_log(
2511         struct xlog *log)
2512 {
2513         xlog_in_core_t  *iclog;
2514         int changed = 0;
2515
2516         iclog = log->l_iclog;
2517         do {
2518                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2519                         iclog->ic_state = XLOG_STATE_ACTIVE;
2520                         iclog->ic_offset       = 0;
2521                         ASSERT(iclog->ic_callback == NULL);
2522                         /*
2523                          * If the number of ops in this iclog indicate it just
2524                          * contains the dummy transaction, we can
2525                          * change state into IDLE (the second time around).
2526                          * Otherwise we should change the state into
2527                          * NEED a dummy.
2528                          * We don't need to cover the dummy.
2529                          */
2530                         if (!changed &&
2531                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2532                                         XLOG_COVER_OPS)) {
2533                                 changed = 1;
2534                         } else {
2535                                 /*
2536                                  * We have two dirty iclogs so start over
2537                                  * This could also be num of ops indicates
2538                                  * this is not the dummy going out.
2539                                  */
2540                                 changed = 2;
2541                         }
2542                         iclog->ic_header.h_num_logops = 0;
2543                         memset(iclog->ic_header.h_cycle_data, 0,
2544                               sizeof(iclog->ic_header.h_cycle_data));
2545                         iclog->ic_header.h_lsn = 0;
2546                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2547                         /* do nothing */;
2548                 else
2549                         break;  /* stop cleaning */
2550                 iclog = iclog->ic_next;
2551         } while (iclog != log->l_iclog);
2552
2553         /* log is locked when we are called */
2554         /*
2555          * Change state for the dummy log recording.
2556          * We usually go to NEED. But we go to NEED2 if the changed indicates
2557          * we are done writing the dummy record.
2558          * If we are done with the second dummy recored (DONE2), then
2559          * we go to IDLE.
2560          */
2561         if (changed) {
2562                 switch (log->l_covered_state) {
2563                 case XLOG_STATE_COVER_IDLE:
2564                 case XLOG_STATE_COVER_NEED:
2565                 case XLOG_STATE_COVER_NEED2:
2566                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2567                         break;
2568
2569                 case XLOG_STATE_COVER_DONE:
2570                         if (changed == 1)
2571                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2572                         else
2573                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2574                         break;
2575
2576                 case XLOG_STATE_COVER_DONE2:
2577                         if (changed == 1)
2578                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2579                         else
2580                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2581                         break;
2582
2583                 default:
2584                         ASSERT(0);
2585                 }
2586         }
2587 }       /* xlog_state_clean_log */
2588
2589 STATIC xfs_lsn_t
2590 xlog_get_lowest_lsn(
2591         struct xlog     *log)
2592 {
2593         xlog_in_core_t  *lsn_log;
2594         xfs_lsn_t       lowest_lsn, lsn;
2595
2596         lsn_log = log->l_iclog;
2597         lowest_lsn = 0;
2598         do {
2599             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2600                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2601                 if ((lsn && !lowest_lsn) ||
2602                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2603                         lowest_lsn = lsn;
2604                 }
2605             }
2606             lsn_log = lsn_log->ic_next;
2607         } while (lsn_log != log->l_iclog);
2608         return lowest_lsn;
2609 }
2610
2611
2612 STATIC void
2613 xlog_state_do_callback(
2614         struct xlog             *log,
2615         int                     aborted,
2616         struct xlog_in_core     *ciclog)
2617 {
2618         xlog_in_core_t     *iclog;
2619         xlog_in_core_t     *first_iclog;        /* used to know when we've
2620                                                  * processed all iclogs once */
2621         xfs_log_callback_t *cb, *cb_next;
2622         int                flushcnt = 0;
2623         xfs_lsn_t          lowest_lsn;
2624         int                ioerrors;    /* counter: iclogs with errors */
2625         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2626         int                funcdidcallbacks; /* flag: function did callbacks */
2627         int                repeats;     /* for issuing console warnings if
2628                                          * looping too many times */
2629         int                wake = 0;
2630
2631         spin_lock(&log->l_icloglock);
2632         first_iclog = iclog = log->l_iclog;
2633         ioerrors = 0;
2634         funcdidcallbacks = 0;
2635         repeats = 0;
2636
2637         do {
2638                 /*
2639                  * Scan all iclogs starting with the one pointed to by the
2640                  * log.  Reset this starting point each time the log is
2641                  * unlocked (during callbacks).
2642                  *
2643                  * Keep looping through iclogs until one full pass is made
2644                  * without running any callbacks.
2645                  */
2646                 first_iclog = log->l_iclog;
2647                 iclog = log->l_iclog;
2648                 loopdidcallbacks = 0;
2649                 repeats++;
2650
2651                 do {
2652
2653                         /* skip all iclogs in the ACTIVE & DIRTY states */
2654                         if (iclog->ic_state &
2655                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2656                                 iclog = iclog->ic_next;
2657                                 continue;
2658                         }
2659
2660                         /*
2661                          * Between marking a filesystem SHUTDOWN and stopping
2662                          * the log, we do flush all iclogs to disk (if there
2663                          * wasn't a log I/O error). So, we do want things to
2664                          * go smoothly in case of just a SHUTDOWN  w/o a
2665                          * LOG_IO_ERROR.
2666                          */
2667                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2668                                 /*
2669                                  * Can only perform callbacks in order.  Since
2670                                  * this iclog is not in the DONE_SYNC/
2671                                  * DO_CALLBACK state, we skip the rest and
2672                                  * just try to clean up.  If we set our iclog
2673                                  * to DO_CALLBACK, we will not process it when
2674                                  * we retry since a previous iclog is in the
2675                                  * CALLBACK and the state cannot change since
2676                                  * we are holding the l_icloglock.
2677                                  */
2678                                 if (!(iclog->ic_state &
2679                                         (XLOG_STATE_DONE_SYNC |
2680                                                  XLOG_STATE_DO_CALLBACK))) {
2681                                         if (ciclog && (ciclog->ic_state ==
2682                                                         XLOG_STATE_DONE_SYNC)) {
2683                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2684                                         }
2685                                         break;
2686                                 }
2687                                 /*
2688                                  * We now have an iclog that is in either the
2689                                  * DO_CALLBACK or DONE_SYNC states. The other
2690                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2691                                  * caught by the above if and are going to
2692                                  * clean (i.e. we aren't doing their callbacks)
2693                                  * see the above if.
2694                                  */
2695
2696                                 /*
2697                                  * We will do one more check here to see if we
2698                                  * have chased our tail around.
2699                                  */
2700
2701                                 lowest_lsn = xlog_get_lowest_lsn(log);
2702                                 if (lowest_lsn &&
2703                                     XFS_LSN_CMP(lowest_lsn,
2704                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2705                                         iclog = iclog->ic_next;
2706                                         continue; /* Leave this iclog for
2707                                                    * another thread */
2708                                 }
2709
2710                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2711
2712
2713                                 /*
2714                                  * Completion of a iclog IO does not imply that
2715                                  * a transaction has completed, as transactions
2716                                  * can be large enough to span many iclogs. We
2717                                  * cannot change the tail of the log half way
2718                                  * through a transaction as this may be the only
2719                                  * transaction in the log and moving th etail to
2720                                  * point to the middle of it will prevent
2721                                  * recovery from finding the start of the
2722                                  * transaction. Hence we should only update the
2723                                  * last_sync_lsn if this iclog contains
2724                                  * transaction completion callbacks on it.
2725                                  *
2726                                  * We have to do this before we drop the
2727                                  * icloglock to ensure we are the only one that
2728                                  * can update it.
2729                                  */
2730                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2731                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2732                                 if (iclog->ic_callback)
2733                                         atomic64_set(&log->l_last_sync_lsn,
2734                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2735
2736                         } else
2737                                 ioerrors++;
2738
2739                         spin_unlock(&log->l_icloglock);
2740
2741                         /*
2742                          * Keep processing entries in the callback list until
2743                          * we come around and it is empty.  We need to
2744                          * atomically see that the list is empty and change the
2745                          * state to DIRTY so that we don't miss any more
2746                          * callbacks being added.
2747                          */
2748                         spin_lock(&iclog->ic_callback_lock);
2749                         cb = iclog->ic_callback;
2750                         while (cb) {
2751                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2752                                 iclog->ic_callback = NULL;
2753                                 spin_unlock(&iclog->ic_callback_lock);
2754
2755                                 /* perform callbacks in the order given */
2756                                 for (; cb; cb = cb_next) {
2757                                         cb_next = cb->cb_next;
2758                                         cb->cb_func(cb->cb_arg, aborted);
2759                                 }
2760                                 spin_lock(&iclog->ic_callback_lock);
2761                                 cb = iclog->ic_callback;
2762                         }
2763
2764                         loopdidcallbacks++;
2765                         funcdidcallbacks++;
2766
2767                         spin_lock(&log->l_icloglock);
2768                         ASSERT(iclog->ic_callback == NULL);
2769                         spin_unlock(&iclog->ic_callback_lock);
2770                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2771                                 iclog->ic_state = XLOG_STATE_DIRTY;
2772
2773                         /*
2774                          * Transition from DIRTY to ACTIVE if applicable.
2775                          * NOP if STATE_IOERROR.
2776                          */
2777                         xlog_state_clean_log(log);
2778
2779                         /* wake up threads waiting in xfs_log_force() */
2780                         wake_up_all(&iclog->ic_force_wait);
2781
2782                         iclog = iclog->ic_next;
2783                 } while (first_iclog != iclog);
2784
2785                 if (repeats > 5000) {
2786                         flushcnt += repeats;
2787                         repeats = 0;
2788                         xfs_warn(log->l_mp,
2789                                 "%s: possible infinite loop (%d iterations)",
2790                                 __func__, flushcnt);
2791                 }
2792         } while (!ioerrors && loopdidcallbacks);
2793
2794         /*
2795          * make one last gasp attempt to see if iclogs are being left in
2796          * limbo..
2797          */
2798 #ifdef DEBUG
2799         if (funcdidcallbacks) {
2800                 first_iclog = iclog = log->l_iclog;
2801                 do {
2802                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2803                         /*
2804                          * Terminate the loop if iclogs are found in states
2805                          * which will cause other threads to clean up iclogs.
2806                          *
2807                          * SYNCING - i/o completion will go through logs
2808                          * DONE_SYNC - interrupt thread should be waiting for
2809                          *              l_icloglock
2810                          * IOERROR - give up hope all ye who enter here
2811                          */
2812                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2813                             iclog->ic_state == XLOG_STATE_SYNCING ||
2814                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2815                             iclog->ic_state == XLOG_STATE_IOERROR )
2816                                 break;
2817                         iclog = iclog->ic_next;
2818                 } while (first_iclog != iclog);
2819         }
2820 #endif
2821
2822         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2823                 wake = 1;
2824         spin_unlock(&log->l_icloglock);
2825
2826         if (wake)
2827                 wake_up_all(&log->l_flush_wait);
2828 }
2829
2830
2831 /*
2832  * Finish transitioning this iclog to the dirty state.
2833  *
2834  * Make sure that we completely execute this routine only when this is
2835  * the last call to the iclog.  There is a good chance that iclog flushes,
2836  * when we reach the end of the physical log, get turned into 2 separate
2837  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2838  * routine.  By using the reference count bwritecnt, we guarantee that only
2839  * the second completion goes through.
2840  *
2841  * Callbacks could take time, so they are done outside the scope of the
2842  * global state machine log lock.
2843  */
2844 STATIC void
2845 xlog_state_done_syncing(
2846         xlog_in_core_t  *iclog,
2847         int             aborted)
2848 {
2849         struct xlog        *log = iclog->ic_log;
2850
2851         spin_lock(&log->l_icloglock);
2852
2853         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2854                iclog->ic_state == XLOG_STATE_IOERROR);
2855         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2856         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2857
2858
2859         /*
2860          * If we got an error, either on the first buffer, or in the case of
2861          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2862          * and none should ever be attempted to be written to disk
2863          * again.
2864          */
2865         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2866                 if (--iclog->ic_bwritecnt == 1) {
2867                         spin_unlock(&log->l_icloglock);
2868                         return;
2869                 }
2870                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2871         }
2872
2873         /*
2874          * Someone could be sleeping prior to writing out the next
2875          * iclog buffer, we wake them all, one will get to do the
2876          * I/O, the others get to wait for the result.
2877          */
2878         wake_up_all(&iclog->ic_write_wait);
2879         spin_unlock(&log->l_icloglock);
2880         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2881 }       /* xlog_state_done_syncing */
2882
2883
2884 /*
2885  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2886  * sleep.  We wait on the flush queue on the head iclog as that should be
2887  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2888  * we will wait here and all new writes will sleep until a sync completes.
2889  *
2890  * The in-core logs are used in a circular fashion. They are not used
2891  * out-of-order even when an iclog past the head is free.
2892  *
2893  * return:
2894  *      * log_offset where xlog_write() can start writing into the in-core
2895  *              log's data space.
2896  *      * in-core log pointer to which xlog_write() should write.
2897  *      * boolean indicating this is a continued write to an in-core log.
2898  *              If this is the last write, then the in-core log's offset field
2899  *              needs to be incremented, depending on the amount of data which
2900  *              is copied.
2901  */
2902 STATIC int
2903 xlog_state_get_iclog_space(
2904         struct xlog             *log,
2905         int                     len,
2906         struct xlog_in_core     **iclogp,
2907         struct xlog_ticket      *ticket,
2908         int                     *continued_write,
2909         int                     *logoffsetp)
2910 {
2911         int               log_offset;
2912         xlog_rec_header_t *head;
2913         xlog_in_core_t    *iclog;
2914         int               error;
2915
2916 restart:
2917         spin_lock(&log->l_icloglock);
2918         if (XLOG_FORCED_SHUTDOWN(log)) {
2919                 spin_unlock(&log->l_icloglock);
2920                 return -EIO;
2921         }
2922
2923         iclog = log->l_iclog;
2924         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2925                 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2926
2927                 /* Wait for log writes to have flushed */
2928                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2929                 goto restart;
2930         }
2931
2932         head = &iclog->ic_header;
2933
2934         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2935         log_offset = iclog->ic_offset;
2936
2937         /* On the 1st write to an iclog, figure out lsn.  This works
2938          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2939          * committing to.  If the offset is set, that's how many blocks
2940          * must be written.
2941          */
2942         if (log_offset == 0) {
2943                 ticket->t_curr_res -= log->l_iclog_hsize;
2944                 xlog_tic_add_region(ticket,
2945                                     log->l_iclog_hsize,
2946                                     XLOG_REG_TYPE_LRHEADER);
2947                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2948                 head->h_lsn = cpu_to_be64(
2949                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2950                 ASSERT(log->l_curr_block >= 0);
2951         }
2952
2953         /* If there is enough room to write everything, then do it.  Otherwise,
2954          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2955          * bit is on, so this will get flushed out.  Don't update ic_offset
2956          * until you know exactly how many bytes get copied.  Therefore, wait
2957          * until later to update ic_offset.
2958          *
2959          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2960          * can fit into remaining data section.
2961          */
2962         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2963                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2964
2965                 /*
2966                  * If I'm the only one writing to this iclog, sync it to disk.
2967                  * We need to do an atomic compare and decrement here to avoid
2968                  * racing with concurrent atomic_dec_and_lock() calls in
2969                  * xlog_state_release_iclog() when there is more than one
2970                  * reference to the iclog.
2971                  */
2972                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2973                         /* we are the only one */
2974                         spin_unlock(&log->l_icloglock);
2975                         error = xlog_state_release_iclog(log, iclog);
2976                         if (error)
2977                                 return error;
2978                 } else {
2979                         spin_unlock(&log->l_icloglock);
2980                 }
2981                 goto restart;
2982         }
2983
2984         /* Do we have enough room to write the full amount in the remainder
2985          * of this iclog?  Or must we continue a write on the next iclog and
2986          * mark this iclog as completely taken?  In the case where we switch
2987          * iclogs (to mark it taken), this particular iclog will release/sync
2988          * to disk in xlog_write().
2989          */
2990         if (len <= iclog->ic_size - iclog->ic_offset) {
2991                 *continued_write = 0;
2992                 iclog->ic_offset += len;
2993         } else {
2994                 *continued_write = 1;
2995                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2996         }
2997         *iclogp = iclog;
2998
2999         ASSERT(iclog->ic_offset <= iclog->ic_size);
3000         spin_unlock(&log->l_icloglock);
3001
3002         *logoffsetp = log_offset;
3003         return 0;
3004 }       /* xlog_state_get_iclog_space */
3005
3006 /* The first cnt-1 times through here we don't need to
3007  * move the grant write head because the permanent
3008  * reservation has reserved cnt times the unit amount.
3009  * Release part of current permanent unit reservation and
3010  * reset current reservation to be one units worth.  Also
3011  * move grant reservation head forward.
3012  */
3013 STATIC void
3014 xlog_regrant_reserve_log_space(
3015         struct xlog             *log,
3016         struct xlog_ticket      *ticket)
3017 {
3018         trace_xfs_log_regrant_reserve_enter(log, ticket);
3019
3020         if (ticket->t_cnt > 0)
3021                 ticket->t_cnt--;
3022
3023         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3024                                         ticket->t_curr_res);
3025         xlog_grant_sub_space(log, &log->l_write_head.grant,
3026                                         ticket->t_curr_res);
3027         ticket->t_curr_res = ticket->t_unit_res;
3028         xlog_tic_reset_res(ticket);
3029
3030         trace_xfs_log_regrant_reserve_sub(log, ticket);
3031
3032         /* just return if we still have some of the pre-reserved space */
3033         if (ticket->t_cnt > 0)
3034                 return;
3035
3036         xlog_grant_add_space(log, &log->l_reserve_head.grant,
3037                                         ticket->t_unit_res);
3038
3039         trace_xfs_log_regrant_reserve_exit(log, ticket);
3040
3041         ticket->t_curr_res = ticket->t_unit_res;
3042         xlog_tic_reset_res(ticket);
3043 }       /* xlog_regrant_reserve_log_space */
3044
3045
3046 /*
3047  * Give back the space left from a reservation.
3048  *
3049  * All the information we need to make a correct determination of space left
3050  * is present.  For non-permanent reservations, things are quite easy.  The
3051  * count should have been decremented to zero.  We only need to deal with the
3052  * space remaining in the current reservation part of the ticket.  If the
3053  * ticket contains a permanent reservation, there may be left over space which
3054  * needs to be released.  A count of N means that N-1 refills of the current
3055  * reservation can be done before we need to ask for more space.  The first
3056  * one goes to fill up the first current reservation.  Once we run out of
3057  * space, the count will stay at zero and the only space remaining will be
3058  * in the current reservation field.
3059  */
3060 STATIC void
3061 xlog_ungrant_log_space(
3062         struct xlog             *log,
3063         struct xlog_ticket      *ticket)
3064 {
3065         int     bytes;
3066
3067         if (ticket->t_cnt > 0)
3068                 ticket->t_cnt--;
3069
3070         trace_xfs_log_ungrant_enter(log, ticket);
3071         trace_xfs_log_ungrant_sub(log, ticket);
3072
3073         /*
3074          * If this is a permanent reservation ticket, we may be able to free
3075          * up more space based on the remaining count.
3076          */
3077         bytes = ticket->t_curr_res;
3078         if (ticket->t_cnt > 0) {
3079                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3080                 bytes += ticket->t_unit_res*ticket->t_cnt;
3081         }
3082
3083         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3084         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3085
3086         trace_xfs_log_ungrant_exit(log, ticket);
3087
3088         xfs_log_space_wake(log->l_mp);
3089 }
3090
3091 /*
3092  * Flush iclog to disk if this is the last reference to the given iclog and
3093  * the WANT_SYNC bit is set.
3094  *
3095  * When this function is entered, the iclog is not necessarily in the
3096  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3097  *
3098  *
3099  */
3100 STATIC int
3101 xlog_state_release_iclog(
3102         struct xlog             *log,
3103         struct xlog_in_core     *iclog)
3104 {
3105         int             sync = 0;       /* do we sync? */
3106
3107         if (iclog->ic_state & XLOG_STATE_IOERROR)
3108                 return -EIO;
3109
3110         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3111         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3112                 return 0;
3113
3114         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3115                 spin_unlock(&log->l_icloglock);
3116                 return -EIO;
3117         }
3118         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3119                iclog->ic_state == XLOG_STATE_WANT_SYNC);
3120
3121         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3122                 /* update tail before writing to iclog */
3123                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3124                 sync++;
3125                 iclog->ic_state = XLOG_STATE_SYNCING;
3126                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3127                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3128                 /* cycle incremented when incrementing curr_block */
3129         }
3130         spin_unlock(&log->l_icloglock);
3131
3132         /*
3133          * We let the log lock go, so it's possible that we hit a log I/O
3134          * error or some other SHUTDOWN condition that marks the iclog
3135          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3136          * this iclog has consistent data, so we ignore IOERROR
3137          * flags after this point.
3138          */
3139         if (sync)
3140                 return xlog_sync(log, iclog);
3141         return 0;
3142 }       /* xlog_state_release_iclog */
3143
3144
3145 /*
3146  * This routine will mark the current iclog in the ring as WANT_SYNC
3147  * and move the current iclog pointer to the next iclog in the ring.
3148  * When this routine is called from xlog_state_get_iclog_space(), the
3149  * exact size of the iclog has not yet been determined.  All we know is
3150  * that every data block.  We have run out of space in this log record.
3151  */
3152 STATIC void
3153 xlog_state_switch_iclogs(
3154         struct xlog             *log,
3155         struct xlog_in_core     *iclog,
3156         int                     eventual_size)
3157 {
3158         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3159         if (!eventual_size)
3160                 eventual_size = iclog->ic_offset;
3161         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3162         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3163         log->l_prev_block = log->l_curr_block;
3164         log->l_prev_cycle = log->l_curr_cycle;
3165
3166         /* roll log?: ic_offset changed later */
3167         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3168
3169         /* Round up to next log-sunit */
3170         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3171             log->l_mp->m_sb.sb_logsunit > 1) {
3172                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3173                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3174         }
3175
3176         if (log->l_curr_block >= log->l_logBBsize) {
3177                 /*
3178                  * Rewind the current block before the cycle is bumped to make
3179                  * sure that the combined LSN never transiently moves forward
3180                  * when the log wraps to the next cycle. This is to support the
3181                  * unlocked sample of these fields from xlog_valid_lsn(). Most
3182                  * other cases should acquire l_icloglock.
3183                  */
3184                 log->l_curr_block -= log->l_logBBsize;
3185                 ASSERT(log->l_curr_block >= 0);
3186                 smp_wmb();
3187                 log->l_curr_cycle++;
3188                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3189                         log->l_curr_cycle++;
3190         }
3191         ASSERT(iclog == log->l_iclog);
3192         log->l_iclog = iclog->ic_next;
3193 }       /* xlog_state_switch_iclogs */
3194
3195 /*
3196  * Write out all data in the in-core log as of this exact moment in time.
3197  *
3198  * Data may be written to the in-core log during this call.  However,
3199  * we don't guarantee this data will be written out.  A change from past
3200  * implementation means this routine will *not* write out zero length LRs.
3201  *
3202  * Basically, we try and perform an intelligent scan of the in-core logs.
3203  * If we determine there is no flushable data, we just return.  There is no
3204  * flushable data if:
3205  *
3206  *      1. the current iclog is active and has no data; the previous iclog
3207  *              is in the active or dirty state.
3208  *      2. the current iclog is drity, and the previous iclog is in the
3209  *              active or dirty state.
3210  *
3211  * We may sleep if:
3212  *
3213  *      1. the current iclog is not in the active nor dirty state.
3214  *      2. the current iclog dirty, and the previous iclog is not in the
3215  *              active nor dirty state.
3216  *      3. the current iclog is active, and there is another thread writing
3217  *              to this particular iclog.
3218  *      4. a) the current iclog is active and has no other writers
3219  *         b) when we return from flushing out this iclog, it is still
3220  *              not in the active nor dirty state.
3221  */
3222 int
3223 _xfs_log_force(
3224         struct xfs_mount        *mp,
3225         uint                    flags,
3226         int                     *log_flushed)
3227 {
3228         struct xlog             *log = mp->m_log;
3229         struct xlog_in_core     *iclog;
3230         xfs_lsn_t               lsn;
3231
3232         XFS_STATS_INC(mp, xs_log_force);
3233
3234         xlog_cil_force(log);
3235
3236         spin_lock(&log->l_icloglock);
3237
3238         iclog = log->l_iclog;
3239         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3240                 spin_unlock(&log->l_icloglock);
3241                 return -EIO;
3242         }
3243
3244         /* If the head iclog is not active nor dirty, we just attach
3245          * ourselves to the head and go to sleep.
3246          */
3247         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3248             iclog->ic_state == XLOG_STATE_DIRTY) {
3249                 /*
3250                  * If the head is dirty or (active and empty), then
3251                  * we need to look at the previous iclog.  If the previous
3252                  * iclog is active or dirty we are done.  There is nothing
3253                  * to sync out.  Otherwise, we attach ourselves to the
3254                  * previous iclog and go to sleep.
3255                  */
3256                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3257                     (atomic_read(&iclog->ic_refcnt) == 0
3258                      && iclog->ic_offset == 0)) {
3259                         iclog = iclog->ic_prev;
3260                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3261                             iclog->ic_state == XLOG_STATE_DIRTY)
3262                                 goto no_sleep;
3263                         else
3264                                 goto maybe_sleep;
3265                 } else {
3266                         if (atomic_read(&iclog->ic_refcnt) == 0) {
3267                                 /* We are the only one with access to this
3268                                  * iclog.  Flush it out now.  There should
3269                                  * be a roundoff of zero to show that someone
3270                                  * has already taken care of the roundoff from
3271                                  * the previous sync.
3272                                  */
3273                                 atomic_inc(&iclog->ic_refcnt);
3274                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3275                                 xlog_state_switch_iclogs(log, iclog, 0);
3276                                 spin_unlock(&log->l_icloglock);
3277
3278                                 if (xlog_state_release_iclog(log, iclog))
3279                                         return -EIO;
3280
3281                                 if (log_flushed)
3282                                         *log_flushed = 1;
3283                                 spin_lock(&log->l_icloglock);
3284                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3285                                     iclog->ic_state != XLOG_STATE_DIRTY)
3286                                         goto maybe_sleep;
3287                                 else
3288                                         goto no_sleep;
3289                         } else {
3290                                 /* Someone else is writing to this iclog.
3291                                  * Use its call to flush out the data.  However,
3292                                  * the other thread may not force out this LR,
3293                                  * so we mark it WANT_SYNC.
3294                                  */
3295                                 xlog_state_switch_iclogs(log, iclog, 0);
3296                                 goto maybe_sleep;
3297                         }
3298                 }
3299         }
3300
3301         /* By the time we come around again, the iclog could've been filled
3302          * which would give it another lsn.  If we have a new lsn, just
3303          * return because the relevant data has been flushed.
3304          */
3305 maybe_sleep:
3306         if (flags & XFS_LOG_SYNC) {
3307                 /*
3308                  * We must check if we're shutting down here, before
3309                  * we wait, while we're holding the l_icloglock.
3310                  * Then we check again after waking up, in case our
3311                  * sleep was disturbed by a bad news.
3312                  */
3313                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3314                         spin_unlock(&log->l_icloglock);
3315                         return -EIO;
3316                 }
3317                 XFS_STATS_INC(mp, xs_log_force_sleep);
3318                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3319                 /*
3320                  * No need to grab the log lock here since we're
3321                  * only deciding whether or not to return EIO
3322                  * and the memory read should be atomic.
3323                  */
3324                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3325                         return -EIO;
3326                 if (log_flushed)
3327                         *log_flushed = 1;
3328         } else {
3329
3330 no_sleep:
3331                 spin_unlock(&log->l_icloglock);
3332         }
3333         return 0;
3334 }
3335
3336 /*
3337  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3338  * about errors or whether the log was flushed or not. This is the normal
3339  * interface to use when trying to unpin items or move the log forward.
3340  */
3341 void
3342 xfs_log_force(
3343         xfs_mount_t     *mp,
3344         uint            flags)
3345 {
3346         int     error;
3347
3348         trace_xfs_log_force(mp, 0);
3349         error = _xfs_log_force(mp, flags, NULL);
3350         if (error)
3351                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3352 }
3353
3354 /*
3355  * Force the in-core log to disk for a specific LSN.
3356  *
3357  * Find in-core log with lsn.
3358  *      If it is in the DIRTY state, just return.
3359  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3360  *              state and go to sleep or return.
3361  *      If it is in any other state, go to sleep or return.
3362  *
3363  * Synchronous forces are implemented with a signal variable. All callers
3364  * to force a given lsn to disk will wait on a the sv attached to the
3365  * specific in-core log.  When given in-core log finally completes its
3366  * write to disk, that thread will wake up all threads waiting on the
3367  * sv.
3368  */
3369 int
3370 _xfs_log_force_lsn(
3371         struct xfs_mount        *mp,
3372         xfs_lsn_t               lsn,
3373         uint                    flags,
3374         int                     *log_flushed)
3375 {
3376         struct xlog             *log = mp->m_log;
3377         struct xlog_in_core     *iclog;
3378         int                     already_slept = 0;
3379
3380         ASSERT(lsn != 0);
3381
3382         XFS_STATS_INC(mp, xs_log_force);
3383
3384         lsn = xlog_cil_force_lsn(log, lsn);
3385         if (lsn == NULLCOMMITLSN)
3386                 return 0;
3387
3388 try_again:
3389         spin_lock(&log->l_icloglock);
3390         iclog = log->l_iclog;
3391         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3392                 spin_unlock(&log->l_icloglock);
3393                 return -EIO;
3394         }
3395
3396         do {
3397                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3398                         iclog = iclog->ic_next;
3399                         continue;
3400                 }
3401
3402                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3403                         spin_unlock(&log->l_icloglock);
3404                         return 0;
3405                 }
3406
3407                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3408                         /*
3409                          * We sleep here if we haven't already slept (e.g.
3410                          * this is the first time we've looked at the correct
3411                          * iclog buf) and the buffer before us is going to
3412                          * be sync'ed. The reason for this is that if we
3413                          * are doing sync transactions here, by waiting for
3414                          * the previous I/O to complete, we can allow a few
3415                          * more transactions into this iclog before we close
3416                          * it down.
3417                          *
3418                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3419                          * up the refcnt so we can release the log (which
3420                          * drops the ref count).  The state switch keeps new
3421                          * transaction commits from using this buffer.  When
3422                          * the current commits finish writing into the buffer,
3423                          * the refcount will drop to zero and the buffer will
3424                          * go out then.
3425                          */
3426                         if (!already_slept &&
3427                             (iclog->ic_prev->ic_state &
3428                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3429                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3430
3431                                 XFS_STATS_INC(mp, xs_log_force_sleep);
3432
3433                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3434                                                         &log->l_icloglock);
3435                                 if (log_flushed)
3436                                         *log_flushed = 1;
3437                                 already_slept = 1;
3438                                 goto try_again;
3439                         }
3440                         atomic_inc(&iclog->ic_refcnt);
3441                         xlog_state_switch_iclogs(log, iclog, 0);
3442                         spin_unlock(&log->l_icloglock);
3443                         if (xlog_state_release_iclog(log, iclog))
3444                                 return -EIO;
3445                         if (log_flushed)
3446                                 *log_flushed = 1;
3447                         spin_lock(&log->l_icloglock);
3448                 }
3449
3450                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3451                     !(iclog->ic_state &
3452                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3453                         /*
3454                          * Don't wait on completion if we know that we've
3455                          * gotten a log write error.
3456                          */
3457                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3458                                 spin_unlock(&log->l_icloglock);
3459                                 return -EIO;
3460                         }
3461                         XFS_STATS_INC(mp, xs_log_force_sleep);
3462                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3463                         /*
3464                          * No need to grab the log lock here since we're
3465                          * only deciding whether or not to return EIO
3466                          * and the memory read should be atomic.
3467                          */
3468                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3469                                 return -EIO;
3470
3471                         if (log_flushed)
3472                                 *log_flushed = 1;
3473                 } else {                /* just return */
3474                         spin_unlock(&log->l_icloglock);
3475                 }
3476
3477                 return 0;
3478         } while (iclog != log->l_iclog);
3479
3480         spin_unlock(&log->l_icloglock);
3481         return 0;
3482 }
3483
3484 /*
3485  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3486  * about errors or whether the log was flushed or not. This is the normal
3487  * interface to use when trying to unpin items or move the log forward.
3488  */
3489 void
3490 xfs_log_force_lsn(
3491         xfs_mount_t     *mp,
3492         xfs_lsn_t       lsn,
3493         uint            flags)
3494 {
3495         int     error;
3496
3497         trace_xfs_log_force(mp, lsn);
3498         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3499         if (error)
3500                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3501 }
3502
3503 /*
3504  * Called when we want to mark the current iclog as being ready to sync to
3505  * disk.
3506  */
3507 STATIC void
3508 xlog_state_want_sync(
3509         struct xlog             *log,
3510         struct xlog_in_core     *iclog)
3511 {
3512         assert_spin_locked(&log->l_icloglock);
3513
3514         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3515                 xlog_state_switch_iclogs(log, iclog, 0);
3516         } else {
3517                 ASSERT(iclog->ic_state &
3518                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3519         }
3520 }
3521
3522
3523 /*****************************************************************************
3524  *
3525  *              TICKET functions
3526  *
3527  *****************************************************************************
3528  */
3529
3530 /*
3531  * Free a used ticket when its refcount falls to zero.
3532  */
3533 void
3534 xfs_log_ticket_put(
3535         xlog_ticket_t   *ticket)
3536 {
3537         ASSERT(atomic_read(&ticket->t_ref) > 0);
3538         if (atomic_dec_and_test(&ticket->t_ref))
3539                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3540 }
3541
3542 xlog_ticket_t *
3543 xfs_log_ticket_get(
3544         xlog_ticket_t   *ticket)
3545 {
3546         ASSERT(atomic_read(&ticket->t_ref) > 0);
3547         atomic_inc(&ticket->t_ref);
3548         return ticket;
3549 }
3550
3551 /*
3552  * Figure out the total log space unit (in bytes) that would be
3553  * required for a log ticket.
3554  */
3555 int
3556 xfs_log_calc_unit_res(
3557         struct xfs_mount        *mp,
3558         int                     unit_bytes)
3559 {
3560         struct xlog             *log = mp->m_log;
3561         int                     iclog_space;
3562         uint                    num_headers;
3563
3564         /*
3565          * Permanent reservations have up to 'cnt'-1 active log operations
3566          * in the log.  A unit in this case is the amount of space for one
3567          * of these log operations.  Normal reservations have a cnt of 1
3568          * and their unit amount is the total amount of space required.
3569          *
3570          * The following lines of code account for non-transaction data
3571          * which occupy space in the on-disk log.
3572          *
3573          * Normal form of a transaction is:
3574          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3575          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3576          *
3577          * We need to account for all the leadup data and trailer data
3578          * around the transaction data.
3579          * And then we need to account for the worst case in terms of using
3580          * more space.
3581          * The worst case will happen if:
3582          * - the placement of the transaction happens to be such that the
3583          *   roundoff is at its maximum
3584          * - the transaction data is synced before the commit record is synced
3585          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3586          *   Therefore the commit record is in its own Log Record.
3587          *   This can happen as the commit record is called with its
3588          *   own region to xlog_write().
3589          *   This then means that in the worst case, roundoff can happen for
3590          *   the commit-rec as well.
3591          *   The commit-rec is smaller than padding in this scenario and so it is
3592          *   not added separately.
3593          */
3594
3595         /* for trans header */
3596         unit_bytes += sizeof(xlog_op_header_t);
3597         unit_bytes += sizeof(xfs_trans_header_t);
3598
3599         /* for start-rec */
3600         unit_bytes += sizeof(xlog_op_header_t);
3601
3602         /*
3603          * for LR headers - the space for data in an iclog is the size minus
3604          * the space used for the headers. If we use the iclog size, then we
3605          * undercalculate the number of headers required.
3606          *
3607          * Furthermore - the addition of op headers for split-recs might
3608          * increase the space required enough to require more log and op
3609          * headers, so take that into account too.
3610          *
3611          * IMPORTANT: This reservation makes the assumption that if this
3612          * transaction is the first in an iclog and hence has the LR headers
3613          * accounted to it, then the remaining space in the iclog is
3614          * exclusively for this transaction.  i.e. if the transaction is larger
3615          * than the iclog, it will be the only thing in that iclog.
3616          * Fundamentally, this means we must pass the entire log vector to
3617          * xlog_write to guarantee this.
3618          */
3619         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3620         num_headers = howmany(unit_bytes, iclog_space);
3621
3622         /* for split-recs - ophdrs added when data split over LRs */
3623         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3624
3625         /* add extra header reservations if we overrun */
3626         while (!num_headers ||
3627                howmany(unit_bytes, iclog_space) > num_headers) {
3628                 unit_bytes += sizeof(xlog_op_header_t);
3629                 num_headers++;
3630         }
3631         unit_bytes += log->l_iclog_hsize * num_headers;
3632
3633         /* for commit-rec LR header - note: padding will subsume the ophdr */
3634         unit_bytes += log->l_iclog_hsize;
3635
3636         /* for roundoff padding for transaction data and one for commit record */
3637         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3638                 /* log su roundoff */
3639                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3640         } else {
3641                 /* BB roundoff */
3642                 unit_bytes += 2 * BBSIZE;
3643         }
3644
3645         return unit_bytes;
3646 }
3647
3648 /*
3649  * Allocate and initialise a new log ticket.
3650  */
3651 struct xlog_ticket *
3652 xlog_ticket_alloc(
3653         struct xlog             *log,
3654         int                     unit_bytes,
3655         int                     cnt,
3656         char                    client,
3657         bool                    permanent,
3658         xfs_km_flags_t          alloc_flags)
3659 {
3660         struct xlog_ticket      *tic;
3661         int                     unit_res;
3662
3663         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3664         if (!tic)
3665                 return NULL;
3666
3667         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3668
3669         atomic_set(&tic->t_ref, 1);
3670         tic->t_task             = current;
3671         INIT_LIST_HEAD(&tic->t_queue);
3672         tic->t_unit_res         = unit_res;
3673         tic->t_curr_res         = unit_res;
3674         tic->t_cnt              = cnt;
3675         tic->t_ocnt             = cnt;
3676         tic->t_tid              = prandom_u32();
3677         tic->t_clientid         = client;
3678         tic->t_flags            = XLOG_TIC_INITED;
3679         tic->t_trans_type       = 0;
3680         if (permanent)
3681                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3682
3683         xlog_tic_reset_res(tic);
3684
3685         return tic;
3686 }
3687
3688
3689 /******************************************************************************
3690  *
3691  *              Log debug routines
3692  *
3693  ******************************************************************************
3694  */
3695 #if defined(DEBUG)
3696 /*
3697  * Make sure that the destination ptr is within the valid data region of
3698  * one of the iclogs.  This uses backup pointers stored in a different
3699  * part of the log in case we trash the log structure.
3700  */
3701 void
3702 xlog_verify_dest_ptr(
3703         struct xlog     *log,
3704         void            *ptr)
3705 {
3706         int i;
3707         int good_ptr = 0;
3708
3709         for (i = 0; i < log->l_iclog_bufs; i++) {
3710                 if (ptr >= log->l_iclog_bak[i] &&
3711                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3712                         good_ptr++;
3713         }
3714
3715         if (!good_ptr)
3716                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3717 }
3718
3719 /*
3720  * Check to make sure the grant write head didn't just over lap the tail.  If
3721  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3722  * the cycles differ by exactly one and check the byte count.
3723  *
3724  * This check is run unlocked, so can give false positives. Rather than assert
3725  * on failures, use a warn-once flag and a panic tag to allow the admin to
3726  * determine if they want to panic the machine when such an error occurs. For
3727  * debug kernels this will have the same effect as using an assert but, unlinke
3728  * an assert, it can be turned off at runtime.
3729  */
3730 STATIC void
3731 xlog_verify_grant_tail(
3732         struct xlog     *log)
3733 {
3734         int             tail_cycle, tail_blocks;
3735         int             cycle, space;
3736
3737         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3738         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3739         if (tail_cycle != cycle) {
3740                 if (cycle - 1 != tail_cycle &&
3741                     !(log->l_flags & XLOG_TAIL_WARN)) {
3742                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3743                                 "%s: cycle - 1 != tail_cycle", __func__);
3744                         log->l_flags |= XLOG_TAIL_WARN;
3745                 }
3746
3747                 if (space > BBTOB(tail_blocks) &&
3748                     !(log->l_flags & XLOG_TAIL_WARN)) {
3749                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3750                                 "%s: space > BBTOB(tail_blocks)", __func__);
3751                         log->l_flags |= XLOG_TAIL_WARN;
3752                 }
3753         }
3754 }
3755
3756 /* check if it will fit */
3757 STATIC void
3758 xlog_verify_tail_lsn(
3759         struct xlog             *log,
3760         struct xlog_in_core     *iclog,
3761         xfs_lsn_t               tail_lsn)
3762 {
3763     int blocks;
3764
3765     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3766         blocks =
3767             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3768         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3769                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3770     } else {
3771         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3772
3773         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3774                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3775
3776         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3777         if (blocks < BTOBB(iclog->ic_offset) + 1)
3778                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3779     }
3780 }       /* xlog_verify_tail_lsn */
3781
3782 /*
3783  * Perform a number of checks on the iclog before writing to disk.
3784  *
3785  * 1. Make sure the iclogs are still circular
3786  * 2. Make sure we have a good magic number
3787  * 3. Make sure we don't have magic numbers in the data
3788  * 4. Check fields of each log operation header for:
3789  *      A. Valid client identifier
3790  *      B. tid ptr value falls in valid ptr space (user space code)
3791  *      C. Length in log record header is correct according to the
3792  *              individual operation headers within record.
3793  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3794  *      log, check the preceding blocks of the physical log to make sure all
3795  *      the cycle numbers agree with the current cycle number.
3796  */
3797 STATIC void
3798 xlog_verify_iclog(
3799         struct xlog             *log,
3800         struct xlog_in_core     *iclog,
3801         int                     count,
3802         bool                    syncing)
3803 {
3804         xlog_op_header_t        *ophead;
3805         xlog_in_core_t          *icptr;
3806         xlog_in_core_2_t        *xhdr;
3807         void                    *base_ptr, *ptr, *p;
3808         ptrdiff_t               field_offset;
3809         __uint8_t               clientid;
3810         int                     len, i, j, k, op_len;
3811         int                     idx;
3812
3813         /* check validity of iclog pointers */
3814         spin_lock(&log->l_icloglock);
3815         icptr = log->l_iclog;
3816         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3817                 ASSERT(icptr);
3818
3819         if (icptr != log->l_iclog)
3820                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3821         spin_unlock(&log->l_icloglock);
3822
3823         /* check log magic numbers */
3824         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3825                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3826
3827         base_ptr = ptr = &iclog->ic_header;
3828         p = &iclog->ic_header;
3829         for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3830                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3831                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3832                                 __func__);
3833         }
3834
3835         /* check fields */
3836         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3837         base_ptr = ptr = iclog->ic_datap;
3838         ophead = ptr;
3839         xhdr = iclog->ic_data;
3840         for (i = 0; i < len; i++) {
3841                 ophead = ptr;
3842
3843                 /* clientid is only 1 byte */
3844                 p = &ophead->oh_clientid;
3845                 field_offset = p - base_ptr;
3846                 if (!syncing || (field_offset & 0x1ff)) {
3847                         clientid = ophead->oh_clientid;
3848                 } else {
3849                         idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3850                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3851                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3852                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3853                                 clientid = xlog_get_client_id(
3854                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3855                         } else {
3856                                 clientid = xlog_get_client_id(
3857                                         iclog->ic_header.h_cycle_data[idx]);
3858                         }
3859                 }
3860                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3861                         xfs_warn(log->l_mp,
3862                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3863                                 __func__, clientid, ophead,
3864                                 (unsigned long)field_offset);
3865
3866                 /* check length */
3867                 p = &ophead->oh_len;
3868                 field_offset = p - base_ptr;
3869                 if (!syncing || (field_offset & 0x1ff)) {
3870                         op_len = be32_to_cpu(ophead->oh_len);
3871                 } else {
3872                         idx = BTOBBT((uintptr_t)&ophead->oh_len -
3873                                     (uintptr_t)iclog->ic_datap);
3874                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3875                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3876                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3877                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3878                         } else {
3879                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3880                         }
3881                 }
3882                 ptr += sizeof(xlog_op_header_t) + op_len;
3883         }
3884 }       /* xlog_verify_iclog */
3885 #endif
3886
3887 /*
3888  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3889  */
3890 STATIC int
3891 xlog_state_ioerror(
3892         struct xlog     *log)
3893 {
3894         xlog_in_core_t  *iclog, *ic;
3895
3896         iclog = log->l_iclog;
3897         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3898                 /*
3899                  * Mark all the incore logs IOERROR.
3900                  * From now on, no log flushes will result.
3901                  */
3902                 ic = iclog;
3903                 do {
3904                         ic->ic_state = XLOG_STATE_IOERROR;
3905                         ic = ic->ic_next;
3906                 } while (ic != iclog);
3907                 return 0;
3908         }
3909         /*
3910          * Return non-zero, if state transition has already happened.
3911          */
3912         return 1;
3913 }
3914
3915 /*
3916  * This is called from xfs_force_shutdown, when we're forcibly
3917  * shutting down the filesystem, typically because of an IO error.
3918  * Our main objectives here are to make sure that:
3919  *      a. if !logerror, flush the logs to disk. Anything modified
3920  *         after this is ignored.
3921  *      b. the filesystem gets marked 'SHUTDOWN' for all interested
3922  *         parties to find out, 'atomically'.
3923  *      c. those who're sleeping on log reservations, pinned objects and
3924  *          other resources get woken up, and be told the bad news.
3925  *      d. nothing new gets queued up after (b) and (c) are done.
3926  *
3927  * Note: for the !logerror case we need to flush the regions held in memory out
3928  * to disk first. This needs to be done before the log is marked as shutdown,
3929  * otherwise the iclog writes will fail.
3930  */
3931 int
3932 xfs_log_force_umount(
3933         struct xfs_mount        *mp,
3934         int                     logerror)
3935 {
3936         struct xlog     *log;
3937         int             retval;
3938
3939         log = mp->m_log;
3940
3941         /*
3942          * If this happens during log recovery, don't worry about
3943          * locking; the log isn't open for business yet.
3944          */
3945         if (!log ||
3946             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3947                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3948                 if (mp->m_sb_bp)
3949                         XFS_BUF_DONE(mp->m_sb_bp);
3950                 return 0;
3951         }
3952
3953         /*
3954          * Somebody could've already done the hard work for us.
3955          * No need to get locks for this.
3956          */
3957         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3958                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3959                 return 1;
3960         }
3961
3962         /*
3963          * Flush all the completed transactions to disk before marking the log
3964          * being shut down. We need to do it in this order to ensure that
3965          * completed operations are safely on disk before we shut down, and that
3966          * we don't have to issue any buffer IO after the shutdown flags are set
3967          * to guarantee this.
3968          */
3969         if (!logerror)
3970                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3971
3972         /*
3973          * mark the filesystem and the as in a shutdown state and wake
3974          * everybody up to tell them the bad news.
3975          */
3976         spin_lock(&log->l_icloglock);
3977         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3978         if (mp->m_sb_bp)
3979                 XFS_BUF_DONE(mp->m_sb_bp);
3980
3981         /*
3982          * Mark the log and the iclogs with IO error flags to prevent any
3983          * further log IO from being issued or completed.
3984          */
3985         log->l_flags |= XLOG_IO_ERROR;
3986         retval = xlog_state_ioerror(log);
3987         spin_unlock(&log->l_icloglock);
3988
3989         /*
3990          * We don't want anybody waiting for log reservations after this. That
3991          * means we have to wake up everybody queued up on reserveq as well as
3992          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3993          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3994          * action is protected by the grant locks.
3995          */
3996         xlog_grant_head_wake_all(&log->l_reserve_head);
3997         xlog_grant_head_wake_all(&log->l_write_head);
3998
3999         /*
4000          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
4001          * as if the log writes were completed. The abort handling in the log
4002          * item committed callback functions will do this again under lock to
4003          * avoid races.
4004          */
4005         wake_up_all(&log->l_cilp->xc_commit_wait);
4006         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
4007
4008 #ifdef XFSERRORDEBUG
4009         {
4010                 xlog_in_core_t  *iclog;
4011
4012                 spin_lock(&log->l_icloglock);
4013                 iclog = log->l_iclog;
4014                 do {
4015                         ASSERT(iclog->ic_callback == 0);
4016                         iclog = iclog->ic_next;
4017                 } while (iclog != log->l_iclog);
4018                 spin_unlock(&log->l_icloglock);
4019         }
4020 #endif
4021         /* return non-zero if log IOERROR transition had already happened */
4022         return retval;
4023 }
4024
4025 STATIC int
4026 xlog_iclogs_empty(
4027         struct xlog     *log)
4028 {
4029         xlog_in_core_t  *iclog;
4030
4031         iclog = log->l_iclog;
4032         do {
4033                 /* endianness does not matter here, zero is zero in
4034                  * any language.
4035                  */
4036                 if (iclog->ic_header.h_num_logops)
4037                         return 0;
4038                 iclog = iclog->ic_next;
4039         } while (iclog != log->l_iclog);
4040         return 1;
4041 }
4042
4043 /*
4044  * Verify that an LSN stamped into a piece of metadata is valid. This is
4045  * intended for use in read verifiers on v5 superblocks.
4046  */
4047 bool
4048 xfs_log_check_lsn(
4049         struct xfs_mount        *mp,
4050         xfs_lsn_t               lsn)
4051 {
4052         struct xlog             *log = mp->m_log;
4053         bool                    valid;
4054
4055         /*
4056          * norecovery mode skips mount-time log processing and unconditionally
4057          * resets the in-core LSN. We can't validate in this mode, but
4058          * modifications are not allowed anyways so just return true.
4059          */
4060         if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4061                 return true;
4062
4063         /*
4064          * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4065          * handled by recovery and thus safe to ignore here.
4066          */
4067         if (lsn == NULLCOMMITLSN)
4068                 return true;
4069
4070         valid = xlog_valid_lsn(mp->m_log, lsn);
4071
4072         /* warn the user about what's gone wrong before verifier failure */
4073         if (!valid) {
4074                 spin_lock(&log->l_icloglock);
4075                 xfs_warn(mp,
4076 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4077 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4078                          CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4079                          log->l_curr_cycle, log->l_curr_block);
4080                 spin_unlock(&log->l_icloglock);
4081         }
4082
4083         return valid;
4084 }