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