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