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