fs/xfs/xfs_log.c

   1 /*
   2  * Copyright (c) 2000-2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_shared.h"
  21 #include "xfs_format.h"
  22 #include "xfs_log_format.h"
  23 #include "xfs_trans_resv.h"
  24 #include "xfs_mount.h"
  25 #include "xfs_error.h"
  26 #include "xfs_trans.h"
  27 #include "xfs_trans_priv.h"
  28 #include "xfs_log.h"
  29 #include "xfs_log_priv.h"
  30 #include "xfs_log_recover.h"
  31 #include "xfs_inode.h"
  32 #include "xfs_trace.h"
  33 #include "xfs_fsops.h"
  34 #include "xfs_cksum.h"
  35 #include "xfs_sysfs.h"
  36 #include "xfs_sb.h"
  37
  38 kmem_zone_t     *xfs_log_ticket_zone;
  39
  40 /* Local miscellaneous function prototypes */
  41 STATIC int
  42 xlog_commit_record(
  43         struct xlog             *log,
  44         struct xlog_ticket      *ticket,
  45         struct xlog_in_core     **iclog,
  46         xfs_lsn_t               *commitlsnp);
  47
  48 STATIC struct xlog *
  49 xlog_alloc_log(
  50         struct xfs_mount        *mp,
  51         struct xfs_buftarg      *log_target,
  52         xfs_daddr_t             blk_offset,
  53         int                     num_bblks);
  54 STATIC int
  55 xlog_space_left(
  56         struct xlog             *log,
  57         atomic64_t              *head);
  58 STATIC int
  59 xlog_sync(
  60         struct xlog             *log,
  61         struct xlog_in_core     *iclog);
  62 STATIC void
  63 xlog_dealloc_log(
  64         struct xlog             *log);
  65
  66 /* local state machine functions */
  67 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  68 STATIC void
  69 xlog_state_do_callback(
  70         struct xlog             *log,
  71         int                     aborted,
  72         struct xlog_in_core     *iclog);
  73 STATIC int
  74 xlog_state_get_iclog_space(
  75         struct xlog             *log,
  76         int                     len,
  77         struct xlog_in_core     **iclog,
  78         struct xlog_ticket      *ticket,
  79         int                     *continued_write,
  80         int                     *logoffsetp);
  81 STATIC int
  82 xlog_state_release_iclog(
  83         struct xlog             *log,
  84         struct xlog_in_core     *iclog);
  85 STATIC void
  86 xlog_state_switch_iclogs(
  87         struct xlog             *log,
  88         struct xlog_in_core     *iclog,
  89         int                     eventual_size);
  90 STATIC void
  91 xlog_state_want_sync(
  92         struct xlog             *log,
  93         struct xlog_in_core     *iclog);
  94
  95 STATIC void
  96 xlog_grant_push_ail(
  97         struct xlog             *log,
  98         int                     need_bytes);
  99 STATIC void
 100 xlog_regrant_reserve_log_space(
 101         struct xlog             *log,
 102         struct xlog_ticket      *ticket);
 103 STATIC void
 104 xlog_ungrant_log_space(
 105         struct xlog             *log,
 106         struct xlog_ticket      *ticket);
 107
 108 #if defined(DEBUG)
 109 STATIC void
 110 xlog_verify_dest_ptr(
 111         struct xlog             *log,
 112         void                    *ptr);
 113 STATIC void
 114 xlog_verify_grant_tail(
 115         struct xlog *log);
 116 STATIC void
 117 xlog_verify_iclog(
 118         struct xlog             *log,
 119         struct xlog_in_core     *iclog,
 120         int                     count,
 121         bool                    syncing);
 122 STATIC void
 123 xlog_verify_tail_lsn(
 124         struct xlog             *log,
 125         struct xlog_in_core     *iclog,
 126         xfs_lsn_t               tail_lsn);
 127 #else
 128 #define xlog_verify_dest_ptr(a,b)
 129 #define xlog_verify_grant_tail(a)
 130 #define xlog_verify_iclog(a,b,c,d)
 131 #define xlog_verify_tail_lsn(a,b,c)
 132 #endif
 133
 134 STATIC int
 135 xlog_iclogs_empty(
 136         struct xlog             *log);
 137
 138 static void
 139 xlog_grant_sub_space(
 140         struct xlog             *log,
 141         atomic64_t              *head,
 142         int                     bytes)
 143 {
 144         int64_t head_val = atomic64_read(head);
 145         int64_t new, old;
 146
 147         do {
 148                 int     cycle, space;
 149
 150                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 151
 152                 space -= bytes;
 153                 if (space < 0) {
 154                         space += log->l_logsize;
 155                         cycle--;
 156                 }
 157
 158                 old = head_val;
 159                 new = xlog_assign_grant_head_val(cycle, space);
 160                 head_val = atomic64_cmpxchg(head, old, new);
 161         } while (head_val != old);
 162 }
 163
 164 static void
 165 xlog_grant_add_space(
 166         struct xlog             *log,
 167         atomic64_t              *head,
 168         int                     bytes)
 169 {
 170         int64_t head_val = atomic64_read(head);
 171         int64_t new, old;
 172
 173         do {
 174                 int             tmp;
 175                 int             cycle, space;
 176
 177                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 178
 179                 tmp = log->l_logsize - space;
 180                 if (tmp > bytes)
 181                         space += bytes;
 182                 else {
 183                         space = bytes - tmp;
 184                         cycle++;
 185                 }
 186
 187                 old = head_val;
 188                 new = xlog_assign_grant_head_val(cycle, space);
 189                 head_val = atomic64_cmpxchg(head, old, new);
 190         } while (head_val != old);
 191 }
 192
 193 STATIC void
 194 xlog_grant_head_init(
 195         struct xlog_grant_head  *head)
 196 {
 197         xlog_assign_grant_head(&head->grant, 1, 0);
 198         INIT_LIST_HEAD(&head->waiters);
 199         spin_lock_init(&head->lock);
 200 }
 201
 202 STATIC void
 203 xlog_grant_head_wake_all(
 204         struct xlog_grant_head  *head)
 205 {
 206         struct xlog_ticket      *tic;
 207
 208         spin_lock(&head->lock);
 209         list_for_each_entry(tic, &head->waiters, t_queue)
 210                 wake_up_process(tic->t_task);
 211         spin_unlock(&head->lock);
 212 }
 213
 214 static inline int
 215 xlog_ticket_reservation(
 216         struct xlog             *log,
 217         struct xlog_grant_head  *head,
 218         struct xlog_ticket      *tic)
 219 {
 220         if (head == &log->l_write_head) {
 221                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 222                 return tic->t_unit_res;
 223         } else {
 224                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 225                         return tic->t_unit_res * tic->t_cnt;
 226                 else
 227                         return tic->t_unit_res;
 228         }
 229 }
 230
 231 STATIC bool
 232 xlog_grant_head_wake(
 233         struct xlog             *log,
 234         struct xlog_grant_head  *head,
 235         int                     *free_bytes)
 236 {
 237         struct xlog_ticket      *tic;
 238         int                     need_bytes;
 239
 240         list_for_each_entry(tic, &head->waiters, t_queue) {
 241                 need_bytes = xlog_ticket_reservation(log, head, tic);
 242                 if (*free_bytes < need_bytes)
 243                         return false;
 244
 245                 *free_bytes -= need_bytes;
 246                 trace_xfs_log_grant_wake_up(log, tic);
 247                 wake_up_process(tic->t_task);
 248         }
 249
 250         return true;
 251 }
 252
 253 STATIC int
 254 xlog_grant_head_wait(
 255         struct xlog             *log,
 256         struct xlog_grant_head  *head,
 257         struct xlog_ticket      *tic,
 258         int                     need_bytes) __releases(&head->lock)
 259                                             __acquires(&head->lock)
 260 {
 261         list_add_tail(&tic->t_queue, &head->waiters);
 262
 263         do {
 264                 if (XLOG_FORCED_SHUTDOWN(log))
 265                         goto shutdown;
 266                 xlog_grant_push_ail(log, need_bytes);
 267
 268                 __set_current_state(TASK_UNINTERRUPTIBLE);
 269                 spin_unlock(&head->lock);
 270
 271                 XFS_STATS_INC(log->l_mp, xs_sleep_logspace);
 272
 273                 trace_xfs_log_grant_sleep(log, tic);
 274                 schedule();
 275                 trace_xfs_log_grant_wake(log, tic);
 276
 277                 spin_lock(&head->lock);
 278                 if (XLOG_FORCED_SHUTDOWN(log))
 279                         goto shutdown;
 280         } while (xlog_space_left(log, &head->grant) < need_bytes);
 281
 282         list_del_init(&tic->t_queue);
 283         return 0;
 284 shutdown:
 285         list_del_init(&tic->t_queue);
 286         return -EIO;
 287 }
 288
 289 /*
 290  * Atomically get the log space required for a log ticket.
 291  *
 292  * Once a ticket gets put onto head->waiters, it will only return after the
 293  * needed reservation is satisfied.
 294  *
 295  * This function is structured so that it has a lock free fast path. This is
 296  * necessary because every new transaction reservation will come through this
 297  * path. Hence any lock will be globally hot if we take it unconditionally on
 298  * every pass.
 299  *
 300  * As tickets are only ever moved on and off head->waiters under head->lock, we
 301  * only need to take that lock if we are going to add the ticket to the queue
 302  * and sleep. We can avoid taking the lock if the ticket was never added to
 303  * head->waiters because the t_queue list head will be empty and we hold the
 304  * only reference to it so it can safely be checked unlocked.
 305  */
 306 STATIC int
 307 xlog_grant_head_check(
 308         struct xlog             *log,
 309         struct xlog_grant_head  *head,
 310         struct xlog_ticket      *tic,
 311         int                     *need_bytes)
 312 {
 313         int                     free_bytes;
 314         int                     error = 0;
 315
 316         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 317
 318         /*
 319          * If there are other waiters on the queue then give them a chance at
 320          * logspace before us.  Wake up the first waiters, if we do not wake
 321          * up all the waiters then go to sleep waiting for more free space,
 322          * otherwise try to get some space for this transaction.
 323          */
 324         *need_bytes = xlog_ticket_reservation(log, head, tic);
 325         free_bytes = xlog_space_left(log, &head->grant);
 326         if (!list_empty_careful(&head->waiters)) {
 327                 spin_lock(&head->lock);
 328                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 329                     free_bytes < *need_bytes) {
 330                         error = xlog_grant_head_wait(log, head, tic,
 331                                                      *need_bytes);
 332                 }
 333                 spin_unlock(&head->lock);
 334         } else if (free_bytes < *need_bytes) {
 335                 spin_lock(&head->lock);
 336                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 337                 spin_unlock(&head->lock);
 338         }
 339
 340         return error;
 341 }
 342
 343 static void
 344 xlog_tic_reset_res(xlog_ticket_t *tic)
 345 {
 346         tic->t_res_num = 0;
 347         tic->t_res_arr_sum = 0;
 348         tic->t_res_num_ophdrs = 0;
 349 }
 350
 351 static void
 352 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 353 {
 354         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 355                 /* add to overflow and start again */
 356                 tic->t_res_o_flow += tic->t_res_arr_sum;
 357                 tic->t_res_num = 0;
 358                 tic->t_res_arr_sum = 0;
 359         }
 360
 361         tic->t_res_arr[tic->t_res_num].r_len = len;
 362         tic->t_res_arr[tic->t_res_num].r_type = type;
 363         tic->t_res_arr_sum += len;
 364         tic->t_res_num++;
 365 }
 366
 367 /*
 368  * Replenish the byte reservation required by moving the grant write head.
 369  */
 370 int
 371 xfs_log_regrant(
 372         struct xfs_mount        *mp,
 373         struct xlog_ticket      *tic)
 374 {
 375         struct xlog             *log = mp->m_log;
 376         int                     need_bytes;
 377         int                     error = 0;
 378
 379         if (XLOG_FORCED_SHUTDOWN(log))
 380                 return -EIO;
 381
 382         XFS_STATS_INC(mp, xs_try_logspace);
 383
 384         /*
 385          * This is a new transaction on the ticket, so we need to change the
 386          * transaction ID so that the next transaction has a different TID in
 387          * the log. Just add one to the existing tid so that we can see chains
 388          * of rolling transactions in the log easily.
 389          */
 390         tic->t_tid++;
 391
 392         xlog_grant_push_ail(log, tic->t_unit_res);
 393
 394         tic->t_curr_res = tic->t_unit_res;
 395         xlog_tic_reset_res(tic);
 396
 397         if (tic->t_cnt > 0)
 398                 return 0;
 399
 400         trace_xfs_log_regrant(log, tic);
 401
 402         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 403                                       &need_bytes);
 404         if (error)
 405                 goto out_error;
 406
 407         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 408         trace_xfs_log_regrant_exit(log, tic);
 409         xlog_verify_grant_tail(log);
 410         return 0;
 411
 412 out_error:
 413         /*
 414          * If we are failing, make sure the ticket doesn't have any current
 415          * reservations.  We don't want to add this back when the ticket/
 416          * transaction gets cancelled.
 417          */
 418         tic->t_curr_res = 0;
 419         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 420         return error;
 421 }
 422
 423 /*
 424  * Reserve log space and return a ticket corresponding the reservation.
 425  *
 426  * Each reservation is going to reserve extra space for a log record header.
 427  * When writes happen to the on-disk log, we don't subtract the length of the
 428  * log record header from any reservation.  By wasting space in each
 429  * reservation, we prevent over allocation problems.
 430  */
 431 int
 432 xfs_log_reserve(
 433         struct xfs_mount        *mp,
 434         int                     unit_bytes,
 435         int                     cnt,
 436         struct xlog_ticket      **ticp,
 437         __uint8_t               client,
 438         bool                    permanent)
 439 {
 440         struct xlog             *log = mp->m_log;
 441         struct xlog_ticket      *tic;
 442         int                     need_bytes;
 443         int                     error = 0;
 444
 445         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 446
 447         if (XLOG_FORCED_SHUTDOWN(log))
 448                 return -EIO;
 449
 450         XFS_STATS_INC(mp, xs_try_logspace);
 451
 452         ASSERT(*ticp == NULL);
 453         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 454                                 KM_SLEEP | KM_MAYFAIL);
 455         if (!tic)
 456                 return -ENOMEM;
 457
 458         *ticp = tic;
 459
 460         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 461                                             : tic->t_unit_res);
 462
 463         trace_xfs_log_reserve(log, tic);
 464
 465         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 466                                       &need_bytes);
 467         if (error)
 468                 goto out_error;
 469
 470         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 471         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 472         trace_xfs_log_reserve_exit(log, tic);
 473         xlog_verify_grant_tail(log);
 474         return 0;
 475
 476 out_error:
 477         /*
 478          * If we are failing, make sure the ticket doesn't have any current
 479          * reservations.  We don't want to add this back when the ticket/
 480          * transaction gets cancelled.
 481          */
 482         tic->t_curr_res = 0;
 483         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 484         return error;
 485 }
 486
 487
 488 /*
 489  * NOTES:
 490  *
 491  *      1. currblock field gets updated at startup and after in-core logs
 492  *              marked as with WANT_SYNC.
 493  */
 494
 495 /*
 496  * This routine is called when a user of a log manager ticket is done with
 497  * the reservation.  If the ticket was ever used, then a commit record for
 498  * the associated transaction is written out as a log operation header with
 499  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 500  * a given ticket.  If the ticket was one with a permanent reservation, then
 501  * a few operations are done differently.  Permanent reservation tickets by
 502  * default don't release the reservation.  They just commit the current
 503  * transaction with the belief that the reservation is still needed.  A flag
 504  * must be passed in before permanent reservations are actually released.
 505  * When these type of tickets are not released, they need to be set into
 506  * the inited state again.  By doing this, a start record will be written
 507  * out when the next write occurs.
 508  */
 509 xfs_lsn_t
 510 xfs_log_done(
 511         struct xfs_mount        *mp,
 512         struct xlog_ticket      *ticket,
 513         struct xlog_in_core     **iclog,
 514         bool                    regrant)
 515 {
 516         struct xlog             *log = mp->m_log;
 517         xfs_lsn_t               lsn = 0;
 518
 519         if (XLOG_FORCED_SHUTDOWN(log) ||
 520             /*
 521              * If nothing was ever written, don't write out commit record.
 522              * If we get an error, just continue and give back the log ticket.
 523              */
 524             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 525              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 526                 lsn = (xfs_lsn_t) -1;
 527                 regrant = false;
 528         }
 529
 530
 531         if (!regrant) {
 532                 trace_xfs_log_done_nonperm(log, ticket);
 533
 534                 /*
 535                  * Release ticket if not permanent reservation or a specific
 536                  * request has been made to release a permanent reservation.
 537                  */
 538                 xlog_ungrant_log_space(log, ticket);
 539         } else {
 540                 trace_xfs_log_done_perm(log, ticket);
 541
 542                 xlog_regrant_reserve_log_space(log, ticket);
 543                 /* If this ticket was a permanent reservation and we aren't
 544                  * trying to release it, reset the inited flags; so next time
 545                  * we write, a start record will be written out.
 546                  */
 547                 ticket->t_flags |= XLOG_TIC_INITED;
 548         }
 549
 550         xfs_log_ticket_put(ticket);
 551         return lsn;
 552 }
 553
 554 /*
 555  * Attaches a new iclog I/O completion callback routine during
 556  * transaction commit.  If the log is in error state, a non-zero
 557  * return code is handed back and the caller is responsible for
 558  * executing the callback at an appropriate time.
 559  */
 560 int
 561 xfs_log_notify(
 562         struct xfs_mount        *mp,
 563         struct xlog_in_core     *iclog,
 564         xfs_log_callback_t      *cb)
 565 {
 566         int     abortflg;
 567
 568         spin_lock(&iclog->ic_callback_lock);
 569         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 570         if (!abortflg) {
 571                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 572                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 573                 cb->cb_next = NULL;
 574                 *(iclog->ic_callback_tail) = cb;
 575                 iclog->ic_callback_tail = &(cb->cb_next);
 576         }
 577         spin_unlock(&iclog->ic_callback_lock);
 578         return abortflg;
 579 }
 580
 581 int
 582 xfs_log_release_iclog(
 583         struct xfs_mount        *mp,
 584         struct xlog_in_core     *iclog)
 585 {
 586         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 587                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 588                 return -EIO;
 589         }
 590
 591         return 0;
 592 }
 593
 594 /*
 595  * Mount a log filesystem
 596  *
 597  * mp           - ubiquitous xfs mount point structure
 598  * log_target   - buftarg of on-disk log device
 599  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 600  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 601  *
 602  * Return error or zero.
 603  */
 604 int
 605 xfs_log_mount(
 606         xfs_mount_t     *mp,
 607         xfs_buftarg_t   *log_target,
 608         xfs_daddr_t     blk_offset,
 609         int             num_bblks)
 610 {
 611         int             error = 0;
 612         int             min_logfsbs;
 613
 614         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 615                 xfs_notice(mp, "Mounting V%d Filesystem",
 616                            XFS_SB_VERSION_NUM(&mp->m_sb));
 617         } else {
 618                 xfs_notice(mp,
 619 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
 620                            XFS_SB_VERSION_NUM(&mp->m_sb));
 621                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 622         }
 623
 624         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 625         if (IS_ERR(mp->m_log)) {
 626                 error = PTR_ERR(mp->m_log);
 627                 goto out;
 628         }
 629
 630         /*
 631          * Validate the given log space and drop a critical message via syslog
 632          * if the log size is too small that would lead to some unexpected
 633          * situations in transaction log space reservation stage.
 634          *
 635          * Note: we can't just reject the mount if the validation fails.  This
 636          * would mean that people would have to downgrade their kernel just to
 637          * remedy the situation as there is no way to grow the log (short of
 638          * black magic surgery with xfs_db).
 639          *
 640          * We can, however, reject mounts for CRC format filesystems, as the
 641          * mkfs binary being used to make the filesystem should never create a
 642          * filesystem with a log that is too small.
 643          */
 644         min_logfsbs = xfs_log_calc_minimum_size(mp);
 645
 646         if (mp->m_sb.sb_logblocks < min_logfsbs) {
 647                 xfs_warn(mp,
 648                 "Log size %d blocks too small, minimum size is %d blocks",
 649                          mp->m_sb.sb_logblocks, min_logfsbs);
 650                 error = -EINVAL;
 651         } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
 652                 xfs_warn(mp,
 653                 "Log size %d blocks too large, maximum size is %lld blocks",
 654                          mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
 655                 error = -EINVAL;
 656         } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
 657                 xfs_warn(mp,
 658                 "log size %lld bytes too large, maximum size is %lld bytes",
 659                          XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
 660                          XFS_MAX_LOG_BYTES);
 661                 error = -EINVAL;
 662         }
 663         if (error) {
 664                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
 665                         xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 666                         ASSERT(0);
 667                         goto out_free_log;
 668                 }
 669                 xfs_crit(mp, "Log size out of supported range.");
 670                 xfs_crit(mp,
 671 "Continuing onwards, but if log hangs are experienced then please report this message in the bug report.");
 672         }
 673
 674         /*
 675          * Initialize the AIL now we have a log.
 676          */
 677         error = xfs_trans_ail_init(mp);
 678         if (error) {
 679                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 680                 goto out_free_log;
 681         }
 682         mp->m_log->l_ailp = mp->m_ail;
 683
 684         /*
 685          * skip log recovery on a norecovery mount.  pretend it all
 686          * just worked.
 687          */
 688         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 689                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 690
 691                 if (readonly)
 692                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 693
 694                 error = xlog_recover(mp->m_log);
 695
 696                 if (readonly)
 697                         mp->m_flags |= XFS_MOUNT_RDONLY;
 698                 if (error) {
 699                         xfs_warn(mp, "log mount/recovery failed: error %d",
 700                                 error);
 701                         xlog_recover_cancel(mp->m_log);
 702                         goto out_destroy_ail;
 703                 }
 704         }
 705
 706         error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 707                                "log");
 708         if (error)
 709                 goto out_destroy_ail;
 710
 711         /* Normal transactions can now occur */
 712         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 713
 714         /*
 715          * Now the log has been fully initialised and we know were our
 716          * space grant counters are, we can initialise the permanent ticket
 717          * needed for delayed logging to work.
 718          */
 719         xlog_cil_init_post_recovery(mp->m_log);
 720
 721         return 0;
 722
 723 out_destroy_ail:
 724         xfs_trans_ail_destroy(mp);
 725 out_free_log:
 726         xlog_dealloc_log(mp->m_log);
 727 out:
 728         return error;
 729 }
 730
 731 /*
 732  * Finish the recovery of the file system.  This is separate from the
 733  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 734  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 735  * here.
 736  *
 737  * If we finish recovery successfully, start the background log work. If we are
 738  * not doing recovery, then we have a RO filesystem and we don't need to start
 739  * it.
 740  */
 741 int
 742 xfs_log_mount_finish(
 743         struct xfs_mount        *mp)
 744 {
 745         int     error = 0;
 746
 747         if (mp->m_flags & XFS_MOUNT_NORECOVERY) {
 748                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 749                 return 0;
 750         }
 751
 752         error = xlog_recover_finish(mp->m_log);
 753         if (!error)
 754                 xfs_log_work_queue(mp);
 755
 756         return error;
 757 }
 758
 759 /*
 760  * The mount has failed. Cancel the recovery if it hasn't completed and destroy
 761  * the log.
 762  */
 763 int
 764 xfs_log_mount_cancel(
 765         struct xfs_mount        *mp)
 766 {
 767         int                     error;
 768
 769         error = xlog_recover_cancel(mp->m_log);
 770         xfs_log_unmount(mp);
 771
 772         return error;
 773 }
 774
 775 /*
 776  * Final log writes as part of unmount.
 777  *
 778  * Mark the filesystem clean as unmount happens.  Note that during relocation
 779  * this routine needs to be executed as part of source-bag while the
 780  * deallocation must not be done until source-end.
 781  */
 782
 783 /*
 784  * Unmount record used to have a string "Unmount filesystem--" in the
 785  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 786  * We just write the magic number now since that particular field isn't
 787  * currently architecture converted and "Unmount" is a bit foo.
 788  * As far as I know, there weren't any dependencies on the old behaviour.
 789  */
 790
 791 int
 792 xfs_log_unmount_write(xfs_mount_t *mp)
 793 {
 794         struct xlog      *log = mp->m_log;
 795         xlog_in_core_t   *iclog;
 796 #ifdef DEBUG
 797         xlog_in_core_t   *first_iclog;
 798 #endif
 799         xlog_ticket_t   *tic = NULL;
 800         xfs_lsn_t        lsn;
 801         int              error;
 802
 803         /*
 804          * Don't write out unmount record on read-only mounts.
 805          * Or, if we are doing a forced umount (typically because of IO errors).
 806          */
 807         if (mp->m_flags & XFS_MOUNT_RDONLY)
 808                 return 0;
 809
 810         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
 811         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 812
 813 #ifdef DEBUG
 814         first_iclog = iclog = log->l_iclog;
 815         do {
 816                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 817                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 818                         ASSERT(iclog->ic_offset == 0);
 819                 }
 820                 iclog = iclog->ic_next;
 821         } while (iclog != first_iclog);
 822 #endif
 823         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 824                 error = xfs_log_reserve(mp, 600, 1, &tic, XFS_LOG, 0);
 825                 if (!error) {
 826                         /* the data section must be 32 bit size aligned */
 827                         struct {
 828                             __uint16_t magic;
 829                             __uint16_t pad1;
 830                             __uint32_t pad2; /* may as well make it 64 bits */
 831                         } magic = {
 832                                 .magic = XLOG_UNMOUNT_TYPE,
 833                         };
 834                         struct xfs_log_iovec reg = {
 835                                 .i_addr = &magic,
 836                                 .i_len = sizeof(magic),
 837                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 838                         };
 839                         struct xfs_log_vec vec = {
 840                                 .lv_niovecs = 1,
 841                                 .lv_iovecp = &reg,
 842                         };
 843
 844                         /* remove inited flag, and account for space used */
 845                         tic->t_flags = 0;
 846                         tic->t_curr_res -= sizeof(magic);
 847                         error = xlog_write(log, &vec, tic, &lsn,
 848                                            NULL, XLOG_UNMOUNT_TRANS);
 849                         /*
 850                          * At this point, we're umounting anyway,
 851                          * so there's no point in transitioning log state
 852                          * to IOERROR. Just continue...
 853                          */
 854                 }
 855
 856                 if (error)
 857                         xfs_alert(mp, "%s: unmount record failed", __func__);
 858
 859
 860                 spin_lock(&log->l_icloglock);
 861                 iclog = log->l_iclog;
 862                 atomic_inc(&iclog->ic_refcnt);
 863                 xlog_state_want_sync(log, iclog);
 864                 spin_unlock(&log->l_icloglock);
 865                 error = xlog_state_release_iclog(log, iclog);
 866
 867                 spin_lock(&log->l_icloglock);
 868                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 869                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 870                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 871                                 xlog_wait(&iclog->ic_force_wait,
 872                                                         &log->l_icloglock);
 873                         } else {
 874                                 spin_unlock(&log->l_icloglock);
 875                         }
 876                 } else {
 877                         spin_unlock(&log->l_icloglock);
 878                 }
 879                 if (tic) {
 880                         trace_xfs_log_umount_write(log, tic);
 881                         xlog_ungrant_log_space(log, tic);
 882                         xfs_log_ticket_put(tic);
 883                 }
 884         } else {
 885                 /*
 886                  * We're already in forced_shutdown mode, couldn't
 887                  * even attempt to write out the unmount transaction.
 888                  *
 889                  * Go through the motions of sync'ing and releasing
 890                  * the iclog, even though no I/O will actually happen,
 891                  * we need to wait for other log I/Os that may already
 892                  * be in progress.  Do this as a separate section of
 893                  * code so we'll know if we ever get stuck here that
 894                  * we're in this odd situation of trying to unmount
 895                  * a file system that went into forced_shutdown as
 896                  * the result of an unmount..
 897                  */
 898                 spin_lock(&log->l_icloglock);
 899                 iclog = log->l_iclog;
 900                 atomic_inc(&iclog->ic_refcnt);
 901
 902                 xlog_state_want_sync(log, iclog);
 903                 spin_unlock(&log->l_icloglock);
 904                 error =  xlog_state_release_iclog(log, iclog);
 905
 906                 spin_lock(&log->l_icloglock);
 907
 908                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 909                         || iclog->ic_state == XLOG_STATE_DIRTY
 910                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 911
 912                                 xlog_wait(&iclog->ic_force_wait,
 913                                                         &log->l_icloglock);
 914                 } else {
 915                         spin_unlock(&log->l_icloglock);
 916                 }
 917         }
 918
 919         return error;
 920 }       /* xfs_log_unmount_write */
 921
 922 /*
 923  * Empty the log for unmount/freeze.
 924  *
 925  * To do this, we first need to shut down the background log work so it is not
 926  * trying to cover the log as we clean up. We then need to unpin all objects in
 927  * the log so we can then flush them out. Once they have completed their IO and
 928  * run the callbacks removing themselves from the AIL, we can write the unmount
 929  * record.
 930  */
 931 void
 932 xfs_log_quiesce(
 933         struct xfs_mount        *mp)
 934 {
 935         cancel_delayed_work_sync(&mp->m_log->l_work);
 936         xfs_log_force(mp, XFS_LOG_SYNC);
 937
 938         /*
 939          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
 940          * will push it, xfs_wait_buftarg() will not wait for it. Further,
 941          * xfs_buf_iowait() cannot be used because it was pushed with the
 942          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
 943          * the IO to complete.
 944          */
 945         xfs_ail_push_all_sync(mp->m_ail);
 946         xfs_wait_buftarg(mp->m_ddev_targp);
 947         xfs_buf_lock(mp->m_sb_bp);
 948         xfs_buf_unlock(mp->m_sb_bp);
 949
 950         xfs_log_unmount_write(mp);
 951 }
 952
 953 /*
 954  * Shut down and release the AIL and Log.
 955  *
 956  * During unmount, we need to ensure we flush all the dirty metadata objects
 957  * from the AIL so that the log is empty before we write the unmount record to
 958  * the log. Once this is done, we can tear down the AIL and the log.
 959  */
 960 void
 961 xfs_log_unmount(
 962         struct xfs_mount        *mp)
 963 {
 964         xfs_log_quiesce(mp);
 965
 966         xfs_trans_ail_destroy(mp);
 967
 968         xfs_sysfs_del(&mp->m_log->l_kobj);
 969
 970         xlog_dealloc_log(mp->m_log);
 971 }
 972
 973 void
 974 xfs_log_item_init(
 975         struct xfs_mount        *mp,
 976         struct xfs_log_item     *item,
 977         int                     type,
 978         const struct xfs_item_ops *ops)
 979 {
 980         item->li_mountp = mp;
 981         item->li_ailp = mp->m_ail;
 982         item->li_type = type;
 983         item->li_ops = ops;
 984         item->li_lv = NULL;
 985
 986         INIT_LIST_HEAD(&item->li_ail);
 987         INIT_LIST_HEAD(&item->li_cil);
 988 }
 989
 990 /*
 991  * Wake up processes waiting for log space after we have moved the log tail.
 992  */
 993 void
 994 xfs_log_space_wake(
 995         struct xfs_mount        *mp)
 996 {
 997         struct xlog             *log = mp->m_log;
 998         int                     free_bytes;
 999
1000         if (XLOG_FORCED_SHUTDOWN(log))
1001                 return;
1002
1003         if (!list_empty_careful(&log->l_write_head.waiters)) {
1004                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1005
1006                 spin_lock(&log->l_write_head.lock);
1007                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
1008                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
1009                 spin_unlock(&log->l_write_head.lock);
1010         }
1011
1012         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1013                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1014
1015                 spin_lock(&log->l_reserve_head.lock);
1016                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1017                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1018                 spin_unlock(&log->l_reserve_head.lock);
1019         }
1020 }
1021
1022 /*
1023  * Determine if we have a transaction that has gone to disk that needs to be
1024  * covered. To begin the transition to the idle state firstly the log needs to
1025  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1026  * we start attempting to cover the log.
1027  *
1028  * Only if we are then in a state where covering is needed, the caller is
1029  * informed that dummy transactions are required to move the log into the idle
1030  * state.
1031  *
1032  * If there are any items in the AIl or CIL, then we do not want to attempt to
1033  * cover the log as we may be in a situation where there isn't log space
1034  * available to run a dummy transaction and this can lead to deadlocks when the
1035  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1036  * there's no point in running a dummy transaction at this point because we
1037  * can't start trying to idle the log until both the CIL and AIL are empty.
1038  */
1039 int
1040 xfs_log_need_covered(xfs_mount_t *mp)
1041 {
1042         struct xlog     *log = mp->m_log;
1043         int             needed = 0;
1044
1045         if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1046                 return 0;
1047
1048         if (!xlog_cil_empty(log))
1049                 return 0;
1050
1051         spin_lock(&log->l_icloglock);
1052         switch (log->l_covered_state) {
1053         case XLOG_STATE_COVER_DONE:
1054         case XLOG_STATE_COVER_DONE2:
1055         case XLOG_STATE_COVER_IDLE:
1056                 break;
1057         case XLOG_STATE_COVER_NEED:
1058         case XLOG_STATE_COVER_NEED2:
1059                 if (xfs_ail_min_lsn(log->l_ailp))
1060                         break;
1061                 if (!xlog_iclogs_empty(log))
1062                         break;
1063
1064                 needed = 1;
1065                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1066                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1067                 else
1068                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1069                 break;
1070         default:
1071                 needed = 1;
1072                 break;
1073         }
1074         spin_unlock(&log->l_icloglock);
1075         return needed;
1076 }
1077
1078 /*
1079  * We may be holding the log iclog lock upon entering this routine.
1080  */
1081 xfs_lsn_t
1082 xlog_assign_tail_lsn_locked(
1083         struct xfs_mount        *mp)
1084 {
1085         struct xlog             *log = mp->m_log;
1086         struct xfs_log_item     *lip;
1087         xfs_lsn_t               tail_lsn;
1088
1089         assert_spin_locked(&mp->m_ail->xa_lock);
1090
1091         /*
1092          * To make sure we always have a valid LSN for the log tail we keep
1093          * track of the last LSN which was committed in log->l_last_sync_lsn,
1094          * and use that when the AIL was empty.
1095          */
1096         lip = xfs_ail_min(mp->m_ail);
1097         if (lip)
1098                 tail_lsn = lip->li_lsn;
1099         else
1100                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1101         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1102         atomic64_set(&log->l_tail_lsn, tail_lsn);
1103         return tail_lsn;
1104 }
1105
1106 xfs_lsn_t
1107 xlog_assign_tail_lsn(
1108         struct xfs_mount        *mp)
1109 {
1110         xfs_lsn_t               tail_lsn;
1111
1112         spin_lock(&mp->m_ail->xa_lock);
1113         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1114         spin_unlock(&mp->m_ail->xa_lock);
1115
1116         return tail_lsn;
1117 }
1118
1119 /*
1120  * Return the space in the log between the tail and the head.  The head
1121  * is passed in the cycle/bytes formal parms.  In the special case where
1122  * the reserve head has wrapped passed the tail, this calculation is no
1123  * longer valid.  In this case, just return 0 which means there is no space
1124  * in the log.  This works for all places where this function is called
1125  * with the reserve head.  Of course, if the write head were to ever
1126  * wrap the tail, we should blow up.  Rather than catch this case here,
1127  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1128  *
1129  * This code also handles the case where the reservation head is behind
1130  * the tail.  The details of this case are described below, but the end
1131  * result is that we return the size of the log as the amount of space left.
1132  */
1133 STATIC int
1134 xlog_space_left(
1135         struct xlog     *log,
1136         atomic64_t      *head)
1137 {
1138         int             free_bytes;
1139         int             tail_bytes;
1140         int             tail_cycle;
1141         int             head_cycle;
1142         int             head_bytes;
1143
1144         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1145         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1146         tail_bytes = BBTOB(tail_bytes);
1147         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1148                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1149         else if (tail_cycle + 1 < head_cycle)
1150                 return 0;
1151         else if (tail_cycle < head_cycle) {
1152                 ASSERT(tail_cycle == (head_cycle - 1));
1153                 free_bytes = tail_bytes - head_bytes;
1154         } else {
1155                 /*
1156                  * The reservation head is behind the tail.
1157                  * In this case we just want to return the size of the
1158                  * log as the amount of space left.
1159                  */
1160                 xfs_alert(log->l_mp, "xlog_space_left: head behind tail");
1161                 xfs_alert(log->l_mp,
1162                           "  tail_cycle = %d, tail_bytes = %d",
1163                           tail_cycle, tail_bytes);
1164                 xfs_alert(log->l_mp,
1165                           "  GH   cycle = %d, GH   bytes = %d",
1166                           head_cycle, head_bytes);
1167                 ASSERT(0);
1168                 free_bytes = log->l_logsize;
1169         }
1170         return free_bytes;
1171 }
1172
1173
1174 /*
1175  * Log function which is called when an io completes.
1176  *
1177  * The log manager needs its own routine, in order to control what
1178  * happens with the buffer after the write completes.
1179  */
1180 void
1181 xlog_iodone(xfs_buf_t *bp)
1182 {
1183         struct xlog_in_core     *iclog = bp->b_fspriv;
1184         struct xlog             *l = iclog->ic_log;
1185         int                     aborted = 0;
1186
1187         /*
1188          * Race to shutdown the filesystem if we see an error or the iclog is in
1189          * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
1190          * CRC errors into log recovery.
1191          */
1192         if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR,
1193                            XFS_RANDOM_IODONE_IOERR) ||
1194             iclog->ic_state & XLOG_STATE_IOABORT) {
1195                 if (iclog->ic_state & XLOG_STATE_IOABORT)
1196                         iclog->ic_state &= ~XLOG_STATE_IOABORT;
1197
1198                 xfs_buf_ioerror_alert(bp, __func__);
1199                 xfs_buf_stale(bp);
1200                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1201                 /*
1202                  * This flag will be propagated to the trans-committed
1203                  * callback routines to let them know that the log-commit
1204                  * didn't succeed.
1205                  */
1206                 aborted = XFS_LI_ABORTED;
1207         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1208                 aborted = XFS_LI_ABORTED;
1209         }
1210
1211         /* log I/O is always issued ASYNC */
1212         ASSERT(bp->b_flags & XBF_ASYNC);
1213         xlog_state_done_syncing(iclog, aborted);
1214
1215         /*
1216          * drop the buffer lock now that we are done. Nothing references
1217          * the buffer after this, so an unmount waiting on this lock can now
1218          * tear it down safely. As such, it is unsafe to reference the buffer
1219          * (bp) after the unlock as we could race with it being freed.
1220          */
1221         xfs_buf_unlock(bp);
1222 }
1223
1224 /*
1225  * Return size of each in-core log record buffer.
1226  *
1227  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1228  *
1229  * If the filesystem blocksize is too large, we may need to choose a
1230  * larger size since the directory code currently logs entire blocks.
1231  */
1232
1233 STATIC void
1234 xlog_get_iclog_buffer_size(
1235         struct xfs_mount        *mp,
1236         struct xlog             *log)
1237 {
1238         int size;
1239         int xhdrs;
1240
1241         if (mp->m_logbufs <= 0)
1242                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1243         else
1244                 log->l_iclog_bufs = mp->m_logbufs;
1245
1246         /*
1247          * Buffer size passed in from mount system call.
1248          */
1249         if (mp->m_logbsize > 0) {
1250                 size = log->l_iclog_size = mp->m_logbsize;
1251                 log->l_iclog_size_log = 0;
1252                 while (size != 1) {
1253                         log->l_iclog_size_log++;
1254                         size >>= 1;
1255                 }
1256
1257                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1258                         /* # headers = size / 32k
1259                          * one header holds cycles from 32k of data
1260                          */
1261
1262                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1263                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1264                                 xhdrs++;
1265                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1266                         log->l_iclog_heads = xhdrs;
1267                 } else {
1268                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1269                         log->l_iclog_hsize = BBSIZE;
1270                         log->l_iclog_heads = 1;
1271                 }
1272                 goto done;
1273         }
1274
1275         /* All machines use 32kB buffers by default. */
1276         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1277         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1278
1279         /* the default log size is 16k or 32k which is one header sector */
1280         log->l_iclog_hsize = BBSIZE;
1281         log->l_iclog_heads = 1;
1282
1283 done:
1284         /* are we being asked to make the sizes selected above visible? */
1285         if (mp->m_logbufs == 0)
1286                 mp->m_logbufs = log->l_iclog_bufs;
1287         if (mp->m_logbsize == 0)
1288                 mp->m_logbsize = log->l_iclog_size;
1289 }       /* xlog_get_iclog_buffer_size */
1290
1291
1292 void
1293 xfs_log_work_queue(
1294         struct xfs_mount        *mp)
1295 {
1296         queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1297                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1298 }
1299
1300 /*
1301  * Every sync period we need to unpin all items in the AIL and push them to
1302  * disk. If there is nothing dirty, then we might need to cover the log to
1303  * indicate that the filesystem is idle.
1304  */
1305 void
1306 xfs_log_worker(
1307         struct work_struct      *work)
1308 {
1309         struct xlog             *log = container_of(to_delayed_work(work),
1310                                                 struct xlog, l_work);
1311         struct xfs_mount        *mp = log->l_mp;
1312
1313         /* dgc: errors ignored - not fatal and nowhere to report them */
1314         if (xfs_log_need_covered(mp)) {
1315                 /*
1316                  * Dump a transaction into the log that contains no real change.
1317                  * This is needed to stamp the current tail LSN into the log
1318                  * during the covering operation.
1319                  *
1320                  * We cannot use an inode here for this - that will push dirty
1321                  * state back up into the VFS and then periodic inode flushing
1322                  * will prevent log covering from making progress. Hence we
1323                  * synchronously log the superblock instead to ensure the
1324                  * superblock is immediately unpinned and can be written back.
1325                  */
1326                 xfs_sync_sb(mp, true);
1327         } else
1328                 xfs_log_force(mp, 0);
1329
1330         /* start pushing all the metadata that is currently dirty */
1331         xfs_ail_push_all(mp->m_ail);
1332
1333         /* queue us up again */
1334         xfs_log_work_queue(mp);
1335 }
1336
1337 /*
1338  * This routine initializes some of the log structure for a given mount point.
1339  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1340  * some other stuff may be filled in too.
1341  */
1342 STATIC struct xlog *
1343 xlog_alloc_log(
1344         struct xfs_mount        *mp,
1345         struct xfs_buftarg      *log_target,
1346         xfs_daddr_t             blk_offset,
1347         int                     num_bblks)
1348 {
1349         struct xlog             *log;
1350         xlog_rec_header_t       *head;
1351         xlog_in_core_t          **iclogp;
1352         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1353         xfs_buf_t               *bp;
1354         int                     i;
1355         int                     error = -ENOMEM;
1356         uint                    log2_size = 0;
1357
1358         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1359         if (!log) {
1360                 xfs_warn(mp, "Log allocation failed: No memory!");
1361                 goto out;
1362         }
1363
1364         log->l_mp          = mp;
1365         log->l_targ        = log_target;
1366         log->l_logsize     = BBTOB(num_bblks);
1367         log->l_logBBstart  = blk_offset;
1368         log->l_logBBsize   = num_bblks;
1369         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1370         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1371         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1372
1373         log->l_prev_block  = -1;
1374         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1375         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1376         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1377         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1378
1379         xlog_grant_head_init(&log->l_reserve_head);
1380         xlog_grant_head_init(&log->l_write_head);
1381
1382         error = -EFSCORRUPTED;
1383         if (xfs_sb_version_hassector(&mp->m_sb)) {
1384                 log2_size = mp->m_sb.sb_logsectlog;
1385                 if (log2_size < BBSHIFT) {
1386                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1387                                 log2_size, BBSHIFT);
1388                         goto out_free_log;
1389                 }
1390
1391                 log2_size -= BBSHIFT;
1392                 if (log2_size > mp->m_sectbb_log) {
1393                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1394                                 log2_size, mp->m_sectbb_log);
1395                         goto out_free_log;
1396                 }
1397
1398                 /* for larger sector sizes, must have v2 or external log */
1399                 if (log2_size && log->l_logBBstart > 0 &&
1400                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1401                         xfs_warn(mp,
1402                 "log sector size (0x%x) invalid for configuration.",
1403                                 log2_size);
1404                         goto out_free_log;
1405                 }
1406         }
1407         log->l_sectBBsize = 1 << log2_size;
1408
1409         xlog_get_iclog_buffer_size(mp, log);
1410
1411         /*
1412          * Use a NULL block for the extra log buffer used during splits so that
1413          * it will trigger errors if we ever try to do IO on it without first
1414          * having set it up properly.
1415          */
1416         error = -ENOMEM;
1417         bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1418                            BTOBB(log->l_iclog_size), 0);
1419         if (!bp)
1420                 goto out_free_log;
1421
1422         /*
1423          * The iclogbuf buffer locks are held over IO but we are not going to do
1424          * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1425          * when appropriately.
1426          */
1427         ASSERT(xfs_buf_islocked(bp));
1428         xfs_buf_unlock(bp);
1429
1430         /* use high priority wq for log I/O completion */
1431         bp->b_ioend_wq = mp->m_log_workqueue;
1432         bp->b_iodone = xlog_iodone;
1433         log->l_xbuf = bp;
1434
1435         spin_lock_init(&log->l_icloglock);
1436         init_waitqueue_head(&log->l_flush_wait);
1437
1438         iclogp = &log->l_iclog;
1439         /*
1440          * The amount of memory to allocate for the iclog structure is
1441          * rather funky due to the way the structure is defined.  It is
1442          * done this way so that we can use different sizes for machines
1443          * with different amounts of memory.  See the definition of
1444          * xlog_in_core_t in xfs_log_priv.h for details.
1445          */
1446         ASSERT(log->l_iclog_size >= 4096);
1447         for (i=0; i < log->l_iclog_bufs; i++) {
1448                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1449                 if (!*iclogp)
1450                         goto out_free_iclog;
1451
1452                 iclog = *iclogp;
1453                 iclog->ic_prev = prev_iclog;
1454                 prev_iclog = iclog;
1455
1456                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1457                                                 BTOBB(log->l_iclog_size), 0);
1458                 if (!bp)
1459                         goto out_free_iclog;
1460
1461                 ASSERT(xfs_buf_islocked(bp));
1462                 xfs_buf_unlock(bp);
1463
1464                 /* use high priority wq for log I/O completion */
1465                 bp->b_ioend_wq = mp->m_log_workqueue;
1466                 bp->b_iodone = xlog_iodone;
1467                 iclog->ic_bp = bp;
1468                 iclog->ic_data = bp->b_addr;
1469 #ifdef DEBUG
1470                 log->l_iclog_bak[i] = &iclog->ic_header;
1471 #endif
1472                 head = &iclog->ic_header;
1473                 memset(head, 0, sizeof(xlog_rec_header_t));
1474                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1475                 head->h_version = cpu_to_be32(
1476                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1477                 head->h_size = cpu_to_be32(log->l_iclog_size);
1478                 /* new fields */
1479                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1480                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1481
1482                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1483                 iclog->ic_state = XLOG_STATE_ACTIVE;
1484                 iclog->ic_log = log;
1485                 atomic_set(&iclog->ic_refcnt, 0);
1486                 spin_lock_init(&iclog->ic_callback_lock);
1487                 iclog->ic_callback_tail = &(iclog->ic_callback);
1488                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1489
1490                 init_waitqueue_head(&iclog->ic_force_wait);
1491                 init_waitqueue_head(&iclog->ic_write_wait);
1492
1493                 iclogp = &iclog->ic_next;
1494         }
1495         *iclogp = log->l_iclog;                 /* complete ring */
1496         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1497
1498         error = xlog_cil_init(log);
1499         if (error)
1500                 goto out_free_iclog;
1501         return log;
1502
1503 out_free_iclog:
1504         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1505                 prev_iclog = iclog->ic_next;
1506                 if (iclog->ic_bp)
1507                         xfs_buf_free(iclog->ic_bp);
1508                 kmem_free(iclog);
1509         }
1510         spinlock_destroy(&log->l_icloglock);
1511         xfs_buf_free(log->l_xbuf);
1512 out_free_log:
1513         kmem_free(log);
1514 out:
1515         return ERR_PTR(error);
1516 }       /* xlog_alloc_log */
1517
1518
1519 /*
1520  * Write out the commit record of a transaction associated with the given
1521  * ticket.  Return the lsn of the commit record.
1522  */
1523 STATIC int
1524 xlog_commit_record(
1525         struct xlog             *log,
1526         struct xlog_ticket      *ticket,
1527         struct xlog_in_core     **iclog,
1528         xfs_lsn_t               *commitlsnp)
1529 {
1530         struct xfs_mount *mp = log->l_mp;
1531         int     error;
1532         struct xfs_log_iovec reg = {
1533                 .i_addr = NULL,
1534                 .i_len = 0,
1535                 .i_type = XLOG_REG_TYPE_COMMIT,
1536         };
1537         struct xfs_log_vec vec = {
1538                 .lv_niovecs = 1,
1539                 .lv_iovecp = &reg,
1540         };
1541
1542         ASSERT_ALWAYS(iclog);
1543         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1544                                         XLOG_COMMIT_TRANS);
1545         if (error)
1546                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1547         return error;
1548 }
1549
1550 /*
1551  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1552  * log space.  This code pushes on the lsn which would supposedly free up
1553  * the 25% which we want to leave free.  We may need to adopt a policy which
1554  * pushes on an lsn which is further along in the log once we reach the high
1555  * water mark.  In this manner, we would be creating a low water mark.
1556  */
1557 STATIC void
1558 xlog_grant_push_ail(
1559         struct xlog     *log,
1560         int             need_bytes)
1561 {
1562         xfs_lsn_t       threshold_lsn = 0;
1563         xfs_lsn_t       last_sync_lsn;
1564         int             free_blocks;
1565         int             free_bytes;
1566         int             threshold_block;
1567         int             threshold_cycle;
1568         int             free_threshold;
1569
1570         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1571
1572         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1573         free_blocks = BTOBBT(free_bytes);
1574
1575         /*
1576          * Set the threshold for the minimum number of free blocks in the
1577          * log to the maximum of what the caller needs, one quarter of the
1578          * log, and 256 blocks.
1579          */
1580         free_threshold = BTOBB(need_bytes);
1581         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1582         free_threshold = MAX(free_threshold, 256);
1583         if (free_blocks >= free_threshold)
1584                 return;
1585
1586         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1587                                                 &threshold_block);
1588         threshold_block += free_threshold;
1589         if (threshold_block >= log->l_logBBsize) {
1590                 threshold_block -= log->l_logBBsize;
1591                 threshold_cycle += 1;
1592         }
1593         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1594                                         threshold_block);
1595         /*
1596          * Don't pass in an lsn greater than the lsn of the last
1597          * log record known to be on disk. Use a snapshot of the last sync lsn
1598          * so that it doesn't change between the compare and the set.
1599          */
1600         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1601         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1602                 threshold_lsn = last_sync_lsn;
1603
1604         /*
1605          * Get the transaction layer to kick the dirty buffers out to
1606          * disk asynchronously. No point in trying to do this if
1607          * the filesystem is shutting down.
1608          */
1609         if (!XLOG_FORCED_SHUTDOWN(log))
1610                 xfs_ail_push(log->l_ailp, threshold_lsn);
1611 }
1612
1613 /*
1614  * Stamp cycle number in every block
1615  */
1616 STATIC void
1617 xlog_pack_data(
1618         struct xlog             *log,
1619         struct xlog_in_core     *iclog,
1620         int                     roundoff)
1621 {
1622         int                     i, j, k;
1623         int                     size = iclog->ic_offset + roundoff;
1624         __be32                  cycle_lsn;
1625         char                    *dp;
1626
1627         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1628
1629         dp = iclog->ic_datap;
1630         for (i = 0; i < BTOBB(size); i++) {
1631                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1632                         break;
1633                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1634                 *(__be32 *)dp = cycle_lsn;
1635                 dp += BBSIZE;
1636         }
1637
1638         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1639                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1640
1641                 for ( ; i < BTOBB(size); i++) {
1642                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1643                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1644                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1645                         *(__be32 *)dp = cycle_lsn;
1646                         dp += BBSIZE;
1647                 }
1648
1649                 for (i = 1; i < log->l_iclog_heads; i++)
1650                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1651         }
1652 }
1653
1654 /*
1655  * Calculate the checksum for a log buffer.
1656  *
1657  * This is a little more complicated than it should be because the various
1658  * headers and the actual data are non-contiguous.
1659  */
1660 __le32
1661 xlog_cksum(
1662         struct xlog             *log,
1663         struct xlog_rec_header  *rhead,
1664         char                    *dp,
1665         int                     size)
1666 {
1667         __uint32_t              crc;
1668
1669         /* first generate the crc for the record header ... */
1670         crc = xfs_start_cksum((char *)rhead,
1671                               sizeof(struct xlog_rec_header),
1672                               offsetof(struct xlog_rec_header, h_crc));
1673
1674         /* ... then for additional cycle data for v2 logs ... */
1675         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1676                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1677                 int             i;
1678                 int             xheads;
1679
1680                 xheads = size / XLOG_HEADER_CYCLE_SIZE;
1681                 if (size % XLOG_HEADER_CYCLE_SIZE)
1682                         xheads++;
1683
1684                 for (i = 1; i < xheads; i++) {
1685                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1686                                      sizeof(struct xlog_rec_ext_header));
1687                 }
1688         }
1689
1690         /* ... and finally for the payload */
1691         crc = crc32c(crc, dp, size);
1692
1693         return xfs_end_cksum(crc);
1694 }
1695
1696 /*
1697  * The bdstrat callback function for log bufs. This gives us a central
1698  * place to trap bufs in case we get hit by a log I/O error and need to
1699  * shutdown. Actually, in practice, even when we didn't get a log error,
1700  * we transition the iclogs to IOERROR state *after* flushing all existing
1701  * iclogs to disk. This is because we don't want anymore new transactions to be
1702  * started or completed afterwards.
1703  *
1704  * We lock the iclogbufs here so that we can serialise against IO completion
1705  * during unmount. We might be processing a shutdown triggered during unmount,
1706  * and that can occur asynchronously to the unmount thread, and hence we need to
1707  * ensure that completes before tearing down the iclogbufs. Hence we need to
1708  * hold the buffer lock across the log IO to acheive that.
1709  */
1710 STATIC int
1711 xlog_bdstrat(
1712         struct xfs_buf          *bp)
1713 {
1714         struct xlog_in_core     *iclog = bp->b_fspriv;
1715
1716         xfs_buf_lock(bp);
1717         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1718                 xfs_buf_ioerror(bp, -EIO);
1719                 xfs_buf_stale(bp);
1720                 xfs_buf_ioend(bp);
1721                 /*
1722                  * It would seem logical to return EIO here, but we rely on
1723                  * the log state machine to propagate I/O errors instead of
1724                  * doing it here. Similarly, IO completion will unlock the
1725                  * buffer, so we don't do it here.
1726                  */
1727                 return 0;
1728         }
1729
1730         xfs_buf_submit(bp);
1731         return 0;
1732 }
1733
1734 /*
1735  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1736  * fashion.  Previously, we should have moved the current iclog
1737  * ptr in the log to point to the next available iclog.  This allows further
1738  * write to continue while this code syncs out an iclog ready to go.
1739  * Before an in-core log can be written out, the data section must be scanned
1740  * to save away the 1st word of each BBSIZE block into the header.  We replace
1741  * it with the current cycle count.  Each BBSIZE block is tagged with the
1742  * cycle count because there in an implicit assumption that drives will
1743  * guarantee that entire 512 byte blocks get written at once.  In other words,
1744  * we can't have part of a 512 byte block written and part not written.  By
1745  * tagging each block, we will know which blocks are valid when recovering
1746  * after an unclean shutdown.
1747  *
1748  * This routine is single threaded on the iclog.  No other thread can be in
1749  * this routine with the same iclog.  Changing contents of iclog can there-
1750  * fore be done without grabbing the state machine lock.  Updating the global
1751  * log will require grabbing the lock though.
1752  *
1753  * The entire log manager uses a logical block numbering scheme.  Only
1754  * log_sync (and then only bwrite()) know about the fact that the log may
1755  * not start with block zero on a given device.  The log block start offset
1756  * is added immediately before calling bwrite().
1757  */
1758
1759 STATIC int
1760 xlog_sync(
1761         struct xlog             *log,
1762         struct xlog_in_core     *iclog)
1763 {
1764         xfs_buf_t       *bp;
1765         int             i;
1766         uint            count;          /* byte count of bwrite */
1767         uint            count_init;     /* initial count before roundup */
1768         int             roundoff;       /* roundoff to BB or stripe */
1769         int             split = 0;      /* split write into two regions */
1770         int             error;
1771         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1772         int             size;
1773
1774         XFS_STATS_INC(log->l_mp, xs_log_writes);
1775         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1776
1777         /* Add for LR header */
1778         count_init = log->l_iclog_hsize + iclog->ic_offset;
1779
1780         /* Round out the log write size */
1781         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1782                 /* we have a v2 stripe unit to use */
1783                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1784         } else {
1785                 count = BBTOB(BTOBB(count_init));
1786         }
1787         roundoff = count - count_init;
1788         ASSERT(roundoff >= 0);
1789         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1790                 roundoff < log->l_mp->m_sb.sb_logsunit)
1791                 ||
1792                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1793                  roundoff < BBTOB(1)));
1794
1795         /* move grant heads by roundoff in sync */
1796         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1797         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1798
1799         /* put cycle number in every block */
1800         xlog_pack_data(log, iclog, roundoff);
1801
1802         /* real byte length */
1803         size = iclog->ic_offset;
1804         if (v2)
1805                 size += roundoff;
1806         iclog->ic_header.h_len = cpu_to_be32(size);
1807
1808         bp = iclog->ic_bp;
1809         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1810
1811         XFS_STATS_ADD(log->l_mp, xs_log_blocks, BTOBB(count));
1812
1813         /* Do we need to split this write into 2 parts? */
1814         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1815                 char            *dptr;
1816
1817                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1818                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1819                 iclog->ic_bwritecnt = 2;
1820
1821                 /*
1822                  * Bump the cycle numbers at the start of each block in the
1823                  * part of the iclog that ends up in the buffer that gets
1824                  * written to the start of the log.
1825                  *
1826                  * Watch out for the header magic number case, though.
1827                  */
1828                 dptr = (char *)&iclog->ic_header + count;
1829                 for (i = 0; i < split; i += BBSIZE) {
1830                         __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1831                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1832                                 cycle++;
1833                         *(__be32 *)dptr = cpu_to_be32(cycle);
1834
1835                         dptr += BBSIZE;
1836                 }
1837         } else {
1838                 iclog->ic_bwritecnt = 1;
1839         }
1840
1841         /* calculcate the checksum */
1842         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1843                                             iclog->ic_datap, size);
1844 #ifdef DEBUG
1845         /*
1846          * Intentionally corrupt the log record CRC based on the error injection
1847          * frequency, if defined. This facilitates testing log recovery in the
1848          * event of torn writes. Hence, set the IOABORT state to abort the log
1849          * write on I/O completion and shutdown the fs. The subsequent mount
1850          * detects the bad CRC and attempts to recover.
1851          */
1852         if (log->l_badcrc_factor &&
1853             (prandom_u32() % log->l_badcrc_factor == 0)) {
1854                 iclog->ic_header.h_crc &= 0xAAAAAAAA;
1855                 iclog->ic_state |= XLOG_STATE_IOABORT;
1856                 xfs_warn(log->l_mp,
1857         "Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
1858                          be64_to_cpu(iclog->ic_header.h_lsn));
1859         }
1860 #endif
1861
1862         bp->b_io_length = BTOBB(count);
1863         bp->b_fspriv = iclog;
1864         bp->b_flags &= ~(XBF_FUA | XBF_FLUSH);
1865         bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE);
1866
1867         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1868                 bp->b_flags |= XBF_FUA;
1869
1870                 /*
1871                  * Flush the data device before flushing the log to make
1872                  * sure all meta data written back from the AIL actually made
1873                  * it to disk before stamping the new log tail LSN into the
1874                  * log buffer.  For an external log we need to issue the
1875                  * flush explicitly, and unfortunately synchronously here;
1876                  * for an internal log we can simply use the block layer
1877                  * state machine for preflushes.
1878                  */
1879                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1880                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1881                 else
1882                         bp->b_flags |= XBF_FLUSH;
1883         }
1884
1885         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1886         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1887
1888         xlog_verify_iclog(log, iclog, count, true);
1889
1890         /* account for log which doesn't start at block #0 */
1891         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1892
1893         /*
1894          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1895          * is shutting down.
1896          */
1897         error = xlog_bdstrat(bp);
1898         if (error) {
1899                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1900                 return error;
1901         }
1902         if (split) {
1903                 bp = iclog->ic_log->l_xbuf;
1904                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1905                 xfs_buf_associate_memory(bp,
1906                                 (char *)&iclog->ic_header + count, split);
1907                 bp->b_fspriv = iclog;
1908                 bp->b_flags &= ~(XBF_FUA | XBF_FLUSH);
1909                 bp->b_flags |= (XBF_ASYNC | XBF_SYNCIO | XBF_WRITE);
1910                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1911                         bp->b_flags |= XBF_FUA;
1912
1913                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1914                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1915
1916                 /* account for internal log which doesn't start at block #0 */
1917                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1918                 error = xlog_bdstrat(bp);
1919                 if (error) {
1920                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1921                         return error;
1922                 }
1923         }
1924         return 0;
1925 }       /* xlog_sync */
1926
1927 /*
1928  * Deallocate a log structure
1929  */
1930 STATIC void
1931 xlog_dealloc_log(
1932         struct xlog     *log)
1933 {
1934         xlog_in_core_t  *iclog, *next_iclog;
1935         int             i;
1936
1937         xlog_cil_destroy(log);
1938
1939         /*
1940          * Cycle all the iclogbuf locks to make sure all log IO completion
1941          * is done before we tear down these buffers.
1942          */
1943         iclog = log->l_iclog;
1944         for (i = 0; i < log->l_iclog_bufs; i++) {
1945                 xfs_buf_lock(iclog->ic_bp);
1946                 xfs_buf_unlock(iclog->ic_bp);
1947                 iclog = iclog->ic_next;
1948         }
1949
1950         /*
1951          * Always need to ensure that the extra buffer does not point to memory
1952          * owned by another log buffer before we free it. Also, cycle the lock
1953          * first to ensure we've completed IO on it.
1954          */
1955         xfs_buf_lock(log->l_xbuf);
1956         xfs_buf_unlock(log->l_xbuf);
1957         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1958         xfs_buf_free(log->l_xbuf);
1959
1960         iclog = log->l_iclog;
1961         for (i = 0; i < log->l_iclog_bufs; i++) {
1962                 xfs_buf_free(iclog->ic_bp);
1963                 next_iclog = iclog->ic_next;
1964                 kmem_free(iclog);
1965                 iclog = next_iclog;
1966         }
1967         spinlock_destroy(&log->l_icloglock);
1968
1969         log->l_mp->m_log = NULL;
1970         kmem_free(log);
1971 }       /* xlog_dealloc_log */
1972
1973 /*
1974  * Update counters atomically now that memcpy is done.
1975  */
1976 /* ARGSUSED */
1977 static inline void
1978 xlog_state_finish_copy(
1979         struct xlog             *log,
1980         struct xlog_in_core     *iclog,
1981         int                     record_cnt,
1982         int                     copy_bytes)
1983 {
1984         spin_lock(&log->l_icloglock);
1985
1986         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1987         iclog->ic_offset += copy_bytes;
1988
1989         spin_unlock(&log->l_icloglock);
1990 }       /* xlog_state_finish_copy */
1991
1992
1993
1994
1995 /*
1996  * print out info relating to regions written which consume
1997  * the reservation
1998  */
1999 void
2000 xlog_print_tic_res(
2001         struct xfs_mount        *mp,
2002         struct xlog_ticket      *ticket)
2003 {
2004         uint i;
2005         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
2006
2007         /* match with XLOG_REG_TYPE_* in xfs_log.h */
2008 #define REG_TYPE_STR(type, str) [XLOG_REG_TYPE_##type] = str
2009         static char *res_type_str[XLOG_REG_TYPE_MAX + 1] = {
2010             REG_TYPE_STR(BFORMAT, "bformat"),
2011             REG_TYPE_STR(BCHUNK, "bchunk"),
2012             REG_TYPE_STR(EFI_FORMAT, "efi_format"),
2013             REG_TYPE_STR(EFD_FORMAT, "efd_format"),
2014             REG_TYPE_STR(IFORMAT, "iformat"),
2015             REG_TYPE_STR(ICORE, "icore"),
2016             REG_TYPE_STR(IEXT, "iext"),
2017             REG_TYPE_STR(IBROOT, "ibroot"),
2018             REG_TYPE_STR(ILOCAL, "ilocal"),
2019             REG_TYPE_STR(IATTR_EXT, "iattr_ext"),
2020             REG_TYPE_STR(IATTR_BROOT, "iattr_broot"),
2021             REG_TYPE_STR(IATTR_LOCAL, "iattr_local"),
2022             REG_TYPE_STR(QFORMAT, "qformat"),
2023             REG_TYPE_STR(DQUOT, "dquot"),
2024             REG_TYPE_STR(QUOTAOFF, "quotaoff"),
2025             REG_TYPE_STR(LRHEADER, "LR header"),
2026             REG_TYPE_STR(UNMOUNT, "unmount"),
2027             REG_TYPE_STR(COMMIT, "commit"),
2028             REG_TYPE_STR(TRANSHDR, "trans header"),
2029             REG_TYPE_STR(ICREATE, "inode create")
2030         };
2031 #undef REG_TYPE_STR
2032
2033         xfs_warn(mp, "xlog_write: reservation summary:");
2034         xfs_warn(mp, "  unit res    = %d bytes",
2035                  ticket->t_unit_res);
2036         xfs_warn(mp, "  current res = %d bytes",
2037                  ticket->t_curr_res);
2038         xfs_warn(mp, "  total reg   = %u bytes (o/flow = %u bytes)",
2039                  ticket->t_res_arr_sum, ticket->t_res_o_flow);
2040         xfs_warn(mp, "  ophdrs      = %u (ophdr space = %u bytes)",
2041                  ticket->t_res_num_ophdrs, ophdr_spc);
2042         xfs_warn(mp, "  ophdr + reg = %u bytes",
2043                  ticket->t_res_arr_sum + ticket->t_res_o_flow + ophdr_spc);
2044         xfs_warn(mp, "  num regions = %u",
2045                  ticket->t_res_num);
2046
2047         for (i = 0; i < ticket->t_res_num; i++) {
2048                 uint r_type = ticket->t_res_arr[i].r_type;
2049                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2050                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2051                             "bad-rtype" : res_type_str[r_type]),
2052                             ticket->t_res_arr[i].r_len);
2053         }
2054
2055         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2056                 "xlog_write: reservation ran out. Need to up reservation");
2057         xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2058 }
2059
2060 /*
2061  * Calculate the potential space needed by the log vector.  Each region gets
2062  * its own xlog_op_header_t and may need to be double word aligned.
2063  */
2064 static int
2065 xlog_write_calc_vec_length(
2066         struct xlog_ticket      *ticket,
2067         struct xfs_log_vec      *log_vector)
2068 {
2069         struct xfs_log_vec      *lv;
2070         int                     headers = 0;
2071         int                     len = 0;
2072         int                     i;
2073
2074         /* acct for start rec of xact */
2075         if (ticket->t_flags & XLOG_TIC_INITED)
2076                 headers++;
2077
2078         for (lv = log_vector; lv; lv = lv->lv_next) {
2079                 /* we don't write ordered log vectors */
2080                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2081                         continue;
2082
2083                 headers += lv->lv_niovecs;
2084
2085                 for (i = 0; i < lv->lv_niovecs; i++) {
2086                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2087
2088                         len += vecp->i_len;
2089                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2090                 }
2091         }
2092
2093         ticket->t_res_num_ophdrs += headers;
2094         len += headers * sizeof(struct xlog_op_header);
2095
2096         return len;
2097 }
2098
2099 /*
2100  * If first write for transaction, insert start record  We can't be trying to
2101  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2102  */
2103 static int
2104 xlog_write_start_rec(
2105         struct xlog_op_header   *ophdr,
2106         struct xlog_ticket      *ticket)
2107 {
2108         if (!(ticket->t_flags & XLOG_TIC_INITED))
2109                 return 0;
2110
2111         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2112         ophdr->oh_clientid = ticket->t_clientid;
2113         ophdr->oh_len = 0;
2114         ophdr->oh_flags = XLOG_START_TRANS;
2115         ophdr->oh_res2 = 0;
2116
2117         ticket->t_flags &= ~XLOG_TIC_INITED;
2118
2119         return sizeof(struct xlog_op_header);
2120 }
2121
2122 static xlog_op_header_t *
2123 xlog_write_setup_ophdr(
2124         struct xlog             *log,
2125         struct xlog_op_header   *ophdr,
2126         struct xlog_ticket      *ticket,
2127         uint                    flags)
2128 {
2129         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2130         ophdr->oh_clientid = ticket->t_clientid;
2131         ophdr->oh_res2 = 0;
2132
2133         /* are we copying a commit or unmount record? */
2134         ophdr->oh_flags = flags;
2135
2136         /*
2137          * We've seen logs corrupted with bad transaction client ids.  This
2138          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2139          * and shut down the filesystem.
2140          */
2141         switch (ophdr->oh_clientid)  {
2142         case XFS_TRANSACTION:
2143         case XFS_VOLUME:
2144         case XFS_LOG:
2145                 break;
2146         default:
2147                 xfs_warn(log->l_mp,
2148                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2149                         ophdr->oh_clientid, ticket);
2150                 return NULL;
2151         }
2152
2153         return ophdr;
2154 }
2155
2156 /*
2157  * Set up the parameters of the region copy into the log. This has
2158  * to handle region write split across multiple log buffers - this
2159  * state is kept external to this function so that this code can
2160  * be written in an obvious, self documenting manner.
2161  */
2162 static int
2163 xlog_write_setup_copy(
2164         struct xlog_ticket      *ticket,
2165         struct xlog_op_header   *ophdr,
2166         int                     space_available,
2167         int                     space_required,
2168         int                     *copy_off,
2169         int                     *copy_len,
2170         int                     *last_was_partial_copy,
2171         int                     *bytes_consumed)
2172 {
2173         int                     still_to_copy;
2174
2175         still_to_copy = space_required - *bytes_consumed;
2176         *copy_off = *bytes_consumed;
2177
2178         if (still_to_copy <= space_available) {
2179                 /* write of region completes here */
2180                 *copy_len = still_to_copy;
2181                 ophdr->oh_len = cpu_to_be32(*copy_len);
2182                 if (*last_was_partial_copy)
2183                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2184                 *last_was_partial_copy = 0;
2185                 *bytes_consumed = 0;
2186                 return 0;
2187         }
2188
2189         /* partial write of region, needs extra log op header reservation */
2190         *copy_len = space_available;
2191         ophdr->oh_len = cpu_to_be32(*copy_len);
2192         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2193         if (*last_was_partial_copy)
2194                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2195         *bytes_consumed += *copy_len;
2196         (*last_was_partial_copy)++;
2197
2198         /* account for new log op header */
2199         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2200         ticket->t_res_num_ophdrs++;
2201
2202         return sizeof(struct xlog_op_header);
2203 }
2204
2205 static int
2206 xlog_write_copy_finish(
2207         struct xlog             *log,
2208         struct xlog_in_core     *iclog,
2209         uint                    flags,
2210         int                     *record_cnt,
2211         int                     *data_cnt,
2212         int                     *partial_copy,
2213         int                     *partial_copy_len,
2214         int                     log_offset,
2215         struct xlog_in_core     **commit_iclog)
2216 {
2217         if (*partial_copy) {
2218                 /*
2219                  * This iclog has already been marked WANT_SYNC by
2220                  * xlog_state_get_iclog_space.
2221                  */
2222                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2223                 *record_cnt = 0;
2224                 *data_cnt = 0;
2225                 return xlog_state_release_iclog(log, iclog);
2226         }
2227
2228         *partial_copy = 0;
2229         *partial_copy_len = 0;
2230
2231         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2232                 /* no more space in this iclog - push it. */
2233                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2234                 *record_cnt = 0;
2235                 *data_cnt = 0;
2236
2237                 spin_lock(&log->l_icloglock);
2238                 xlog_state_want_sync(log, iclog);
2239                 spin_unlock(&log->l_icloglock);
2240
2241                 if (!commit_iclog)
2242                         return xlog_state_release_iclog(log, iclog);
2243                 ASSERT(flags & XLOG_COMMIT_TRANS);
2244                 *commit_iclog = iclog;
2245         }
2246
2247         return 0;
2248 }
2249
2250 /*
2251  * Write some region out to in-core log
2252  *
2253  * This will be called when writing externally provided regions or when
2254  * writing out a commit record for a given transaction.
2255  *
2256  * General algorithm:
2257  *      1. Find total length of this write.  This may include adding to the
2258  *              lengths passed in.
2259  *      2. Check whether we violate the tickets reservation.
2260  *      3. While writing to this iclog
2261  *          A. Reserve as much space in this iclog as can get
2262  *          B. If this is first write, save away start lsn
2263  *          C. While writing this region:
2264  *              1. If first write of transaction, write start record
2265  *              2. Write log operation header (header per region)
2266  *              3. Find out if we can fit entire region into this iclog
2267  *              4. Potentially, verify destination memcpy ptr
2268  *              5. Memcpy (partial) region
2269  *              6. If partial copy, release iclog; otherwise, continue
2270  *                      copying more regions into current iclog
2271  *      4. Mark want sync bit (in simulation mode)
2272  *      5. Release iclog for potential flush to on-disk log.
2273  *
2274  * ERRORS:
2275  * 1.   Panic if reservation is overrun.  This should never happen since
2276  *      reservation amounts are generated internal to the filesystem.
2277  * NOTES:
2278  * 1. Tickets are single threaded data structures.
2279  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2280  *      syncing routine.  When a single log_write region needs to span
2281  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2282  *      on all log operation writes which don't contain the end of the
2283  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2284  *      operation which contains the end of the continued log_write region.
2285  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2286  *      we don't really know exactly how much space will be used.  As a result,
2287  *      we don't update ic_offset until the end when we know exactly how many
2288  *      bytes have been written out.
2289  */
2290 int
2291 xlog_write(
2292         struct xlog             *log,
2293         struct xfs_log_vec      *log_vector,
2294         struct xlog_ticket      *ticket,
2295         xfs_lsn_t               *start_lsn,
2296         struct xlog_in_core     **commit_iclog,
2297         uint                    flags)
2298 {
2299         struct xlog_in_core     *iclog = NULL;
2300         struct xfs_log_iovec    *vecp;
2301         struct xfs_log_vec      *lv;
2302         int                     len;
2303         int                     index;
2304         int                     partial_copy = 0;
2305         int                     partial_copy_len = 0;
2306         int                     contwr = 0;
2307         int                     record_cnt = 0;
2308         int                     data_cnt = 0;
2309         int                     error;
2310
2311         *start_lsn = 0;
2312
2313         len = xlog_write_calc_vec_length(ticket, log_vector);
2314
2315         /*
2316          * Region headers and bytes are already accounted for.
2317          * We only need to take into account start records and
2318          * split regions in this function.
2319          */
2320         if (ticket->t_flags & XLOG_TIC_INITED)
2321                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2322
2323         /*
2324          * Commit record headers need to be accounted for. These
2325          * come in as separate writes so are easy to detect.
2326          */
2327         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2328                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2329
2330         if (ticket->t_curr_res < 0)
2331                 xlog_print_tic_res(log->l_mp, ticket);
2332
2333         index = 0;
2334         lv = log_vector;
2335         vecp = lv->lv_iovecp;
2336         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2337                 void            *ptr;
2338                 int             log_offset;
2339
2340                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2341                                                    &contwr, &log_offset);
2342                 if (error)
2343                         return error;
2344
2345                 ASSERT(log_offset <= iclog->ic_size - 1);
2346                 ptr = iclog->ic_datap + log_offset;
2347
2348                 /* start_lsn is the first lsn written to. That's all we need. */
2349                 if (!*start_lsn)
2350                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2351
2352                 /*
2353                  * This loop writes out as many regions as can fit in the amount
2354                  * of space which was allocated by xlog_state_get_iclog_space().
2355                  */
2356                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2357                         struct xfs_log_iovec    *reg;
2358                         struct xlog_op_header   *ophdr;
2359                         int                     start_rec_copy;
2360                         int                     copy_len;
2361                         int                     copy_off;
2362                         bool                    ordered = false;
2363
2364                         /* ordered log vectors have no regions to write */
2365                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2366                                 ASSERT(lv->lv_niovecs == 0);
2367                                 ordered = true;
2368                                 goto next_lv;
2369                         }
2370
2371                         reg = &vecp[index];
2372                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2373                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2374
2375                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2376                         if (start_rec_copy) {
2377                                 record_cnt++;
2378                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2379                                                    start_rec_copy);
2380                         }
2381
2382                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2383                         if (!ophdr)
2384                                 return -EIO;
2385
2386                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2387                                            sizeof(struct xlog_op_header));
2388
2389                         len += xlog_write_setup_copy(ticket, ophdr,
2390                                                      iclog->ic_size-log_offset,
2391                                                      reg->i_len,
2392                                                      &copy_off, &copy_len,
2393                                                      &partial_copy,
2394                                                      &partial_copy_len);
2395                         xlog_verify_dest_ptr(log, ptr);
2396
2397                         /*
2398                          * Copy region.
2399                          *
2400                          * Unmount records just log an opheader, so can have
2401                          * empty payloads with no data region to copy. Hence we
2402                          * only copy the payload if the vector says it has data
2403                          * to copy.
2404                          */
2405                         ASSERT(copy_len >= 0);
2406                         if (copy_len > 0) {
2407                                 memcpy(ptr, reg->i_addr + copy_off, copy_len);
2408                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2409                                                    copy_len);
2410                         }
2411                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2412                         record_cnt++;
2413                         data_cnt += contwr ? copy_len : 0;
2414
2415                         error = xlog_write_copy_finish(log, iclog, flags,
2416                                                        &record_cnt, &data_cnt,
2417                                                        &partial_copy,
2418                                                        &partial_copy_len,
2419                                                        log_offset,
2420                                                        commit_iclog);
2421                         if (error)
2422                                 return error;
2423
2424                         /*
2425                          * if we had a partial copy, we need to get more iclog
2426                          * space but we don't want to increment the region
2427                          * index because there is still more is this region to
2428                          * write.
2429                          *
2430                          * If we completed writing this region, and we flushed
2431                          * the iclog (indicated by resetting of the record
2432                          * count), then we also need to get more log space. If
2433                          * this was the last record, though, we are done and
2434                          * can just return.
2435                          */
2436                         if (partial_copy)
2437                                 break;
2438
2439                         if (++index == lv->lv_niovecs) {
2440 next_lv:
2441                                 lv = lv->lv_next;
2442                                 index = 0;
2443                                 if (lv)
2444                                         vecp = lv->lv_iovecp;
2445                         }
2446                         if (record_cnt == 0 && ordered == false) {
2447                                 if (!lv)
2448                                         return 0;
2449                                 break;
2450                         }
2451                 }
2452         }
2453
2454         ASSERT(len == 0);
2455
2456         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2457         if (!commit_iclog)
2458                 return xlog_state_release_iclog(log, iclog);
2459
2460         ASSERT(flags & XLOG_COMMIT_TRANS);
2461         *commit_iclog = iclog;
2462         return 0;
2463 }
2464
2465
2466 /*****************************************************************************
2467  *
2468  *              State Machine functions
2469  *
2470  *****************************************************************************
2471  */
2472
2473 /* Clean iclogs starting from the head.  This ordering must be
2474  * maintained, so an iclog doesn't become ACTIVE beyond one that
2475  * is SYNCING.  This is also required to maintain the notion that we use
2476  * a ordered wait queue to hold off would be writers to the log when every
2477  * iclog is trying to sync to disk.
2478  *
2479  * State Change: DIRTY -> ACTIVE
2480  */
2481 STATIC void
2482 xlog_state_clean_log(
2483         struct xlog *log)
2484 {
2485         xlog_in_core_t  *iclog;
2486         int changed = 0;
2487
2488         iclog = log->l_iclog;
2489         do {
2490                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2491                         iclog->ic_state = XLOG_STATE_ACTIVE;
2492                         iclog->ic_offset       = 0;
2493                         ASSERT(iclog->ic_callback == NULL);
2494                         /*
2495                          * If the number of ops in this iclog indicate it just
2496                          * contains the dummy transaction, we can
2497                          * change state into IDLE (the second time around).
2498                          * Otherwise we should change the state into
2499                          * NEED a dummy.
2500                          * We don't need to cover the dummy.
2501                          */
2502                         if (!changed &&
2503                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2504                                         XLOG_COVER_OPS)) {
2505                                 changed = 1;
2506                         } else {
2507                                 /*
2508                                  * We have two dirty iclogs so start over
2509                                  * This could also be num of ops indicates
2510                                  * this is not the dummy going out.
2511                                  */
2512                                 changed = 2;
2513                         }
2514                         iclog->ic_header.h_num_logops = 0;
2515                         memset(iclog->ic_header.h_cycle_data, 0,
2516                               sizeof(iclog->ic_header.h_cycle_data));
2517                         iclog->ic_header.h_lsn = 0;
2518                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2519                         /* do nothing */;
2520                 else
2521                         break;  /* stop cleaning */
2522                 iclog = iclog->ic_next;
2523         } while (iclog != log->l_iclog);
2524
2525         /* log is locked when we are called */
2526         /*
2527          * Change state for the dummy log recording.
2528          * We usually go to NEED. But we go to NEED2 if the changed indicates
2529          * we are done writing the dummy record.
2530          * If we are done with the second dummy recored (DONE2), then
2531          * we go to IDLE.
2532          */
2533         if (changed) {
2534                 switch (log->l_covered_state) {
2535                 case XLOG_STATE_COVER_IDLE:
2536                 case XLOG_STATE_COVER_NEED:
2537                 case XLOG_STATE_COVER_NEED2:
2538                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2539                         break;
2540
2541                 case XLOG_STATE_COVER_DONE:
2542                         if (changed == 1)
2543                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2544                         else
2545                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2546                         break;
2547
2548                 case XLOG_STATE_COVER_DONE2:
2549                         if (changed == 1)
2550                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2551                         else
2552                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2553                         break;
2554
2555                 default:
2556                         ASSERT(0);
2557                 }
2558         }
2559 }       /* xlog_state_clean_log */
2560
2561 STATIC xfs_lsn_t
2562 xlog_get_lowest_lsn(
2563         struct xlog     *log)
2564 {
2565         xlog_in_core_t  *lsn_log;
2566         xfs_lsn_t       lowest_lsn, lsn;
2567
2568         lsn_log = log->l_iclog;
2569         lowest_lsn = 0;
2570         do {
2571             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2572                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2573                 if ((lsn && !lowest_lsn) ||
2574                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2575                         lowest_lsn = lsn;
2576                 }
2577             }
2578             lsn_log = lsn_log->ic_next;
2579         } while (lsn_log != log->l_iclog);
2580         return lowest_lsn;
2581 }
2582
2583
2584 STATIC void
2585 xlog_state_do_callback(
2586         struct xlog             *log,
2587         int                     aborted,
2588         struct xlog_in_core     *ciclog)
2589 {
2590         xlog_in_core_t     *iclog;
2591         xlog_in_core_t     *first_iclog;        /* used to know when we've
2592                                                  * processed all iclogs once */
2593         xfs_log_callback_t *cb, *cb_next;
2594         int                flushcnt = 0;
2595         xfs_lsn_t          lowest_lsn;
2596         int                ioerrors;    /* counter: iclogs with errors */
2597         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2598         int                funcdidcallbacks; /* flag: function did callbacks */
2599         int                repeats;     /* for issuing console warnings if
2600                                          * looping too many times */
2601         int                wake = 0;
2602
2603         spin_lock(&log->l_icloglock);
2604         first_iclog = iclog = log->l_iclog;
2605         ioerrors = 0;
2606         funcdidcallbacks = 0;
2607         repeats = 0;
2608
2609         do {
2610                 /*
2611                  * Scan all iclogs starting with the one pointed to by the
2612                  * log.  Reset this starting point each time the log is
2613                  * unlocked (during callbacks).
2614                  *
2615                  * Keep looping through iclogs until one full pass is made
2616                  * without running any callbacks.
2617                  */
2618                 first_iclog = log->l_iclog;
2619                 iclog = log->l_iclog;
2620                 loopdidcallbacks = 0;
2621                 repeats++;
2622
2623                 do {
2624
2625                         /* skip all iclogs in the ACTIVE & DIRTY states */
2626                         if (iclog->ic_state &
2627                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2628                                 iclog = iclog->ic_next;
2629                                 continue;
2630                         }
2631
2632                         /*
2633                          * Between marking a filesystem SHUTDOWN and stopping
2634                          * the log, we do flush all iclogs to disk (if there
2635                          * wasn't a log I/O error). So, we do want things to
2636                          * go smoothly in case of just a SHUTDOWN  w/o a
2637                          * LOG_IO_ERROR.
2638                          */
2639                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2640                                 /*
2641                                  * Can only perform callbacks in order.  Since
2642                                  * this iclog is not in the DONE_SYNC/
2643                                  * DO_CALLBACK state, we skip the rest and
2644                                  * just try to clean up.  If we set our iclog
2645                                  * to DO_CALLBACK, we will not process it when
2646                                  * we retry since a previous iclog is in the
2647                                  * CALLBACK and the state cannot change since
2648                                  * we are holding the l_icloglock.
2649                                  */
2650                                 if (!(iclog->ic_state &
2651                                         (XLOG_STATE_DONE_SYNC |
2652                                                  XLOG_STATE_DO_CALLBACK))) {
2653                                         if (ciclog && (ciclog->ic_state ==
2654                                                         XLOG_STATE_DONE_SYNC)) {
2655                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2656                                         }
2657                                         break;
2658                                 }
2659                                 /*
2660                                  * We now have an iclog that is in either the
2661                                  * DO_CALLBACK or DONE_SYNC states. The other
2662                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2663                                  * caught by the above if and are going to
2664                                  * clean (i.e. we aren't doing their callbacks)
2665                                  * see the above if.
2666                                  */
2667
2668                                 /*
2669                                  * We will do one more check here to see if we
2670                                  * have chased our tail around.
2671                                  */
2672
2673                                 lowest_lsn = xlog_get_lowest_lsn(log);
2674                                 if (lowest_lsn &&
2675                                     XFS_LSN_CMP(lowest_lsn,
2676                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2677                                         iclog = iclog->ic_next;
2678                                         continue; /* Leave this iclog for
2679                                                    * another thread */
2680                                 }
2681
2682                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2683
2684
2685                                 /*
2686                                  * Completion of a iclog IO does not imply that
2687                                  * a transaction has completed, as transactions
2688                                  * can be large enough to span many iclogs. We
2689                                  * cannot change the tail of the log half way
2690                                  * through a transaction as this may be the only
2691                                  * transaction in the log and moving th etail to
2692                                  * point to the middle of it will prevent
2693                                  * recovery from finding the start of the
2694                                  * transaction. Hence we should only update the
2695                                  * last_sync_lsn if this iclog contains
2696                                  * transaction completion callbacks on it.
2697                                  *
2698                                  * We have to do this before we drop the
2699                                  * icloglock to ensure we are the only one that
2700                                  * can update it.
2701                                  */
2702                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2703                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2704                                 if (iclog->ic_callback)
2705                                         atomic64_set(&log->l_last_sync_lsn,
2706                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2707
2708                         } else
2709                                 ioerrors++;
2710
2711                         spin_unlock(&log->l_icloglock);
2712
2713                         /*
2714                          * Keep processing entries in the callback list until
2715                          * we come around and it is empty.  We need to
2716                          * atomically see that the list is empty and change the
2717                          * state to DIRTY so that we don't miss any more
2718                          * callbacks being added.
2719                          */
2720                         spin_lock(&iclog->ic_callback_lock);
2721                         cb = iclog->ic_callback;
2722                         while (cb) {
2723                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2724                                 iclog->ic_callback = NULL;
2725                                 spin_unlock(&iclog->ic_callback_lock);
2726
2727                                 /* perform callbacks in the order given */
2728                                 for (; cb; cb = cb_next) {
2729                                         cb_next = cb->cb_next;
2730                                         cb->cb_func(cb->cb_arg, aborted);
2731                                 }
2732                                 spin_lock(&iclog->ic_callback_lock);
2733                                 cb = iclog->ic_callback;
2734                         }
2735
2736                         loopdidcallbacks++;
2737                         funcdidcallbacks++;
2738
2739                         spin_lock(&log->l_icloglock);
2740                         ASSERT(iclog->ic_callback == NULL);
2741                         spin_unlock(&iclog->ic_callback_lock);
2742                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2743                                 iclog->ic_state = XLOG_STATE_DIRTY;
2744
2745                         /*
2746                          * Transition from DIRTY to ACTIVE if applicable.
2747                          * NOP if STATE_IOERROR.
2748                          */
2749                         xlog_state_clean_log(log);
2750
2751                         /* wake up threads waiting in xfs_log_force() */
2752                         wake_up_all(&iclog->ic_force_wait);
2753
2754                         iclog = iclog->ic_next;
2755                 } while (first_iclog != iclog);
2756
2757                 if (repeats > 5000) {
2758                         flushcnt += repeats;
2759                         repeats = 0;
2760                         xfs_warn(log->l_mp,
2761                                 "%s: possible infinite loop (%d iterations)",
2762                                 __func__, flushcnt);
2763                 }
2764         } while (!ioerrors && loopdidcallbacks);
2765
2766 #ifdef DEBUG
2767         /*
2768          * Make one last gasp attempt to see if iclogs are being left in limbo.
2769          * If the above loop finds an iclog earlier than the current iclog and
2770          * in one of the syncing states, the current iclog is put into
2771          * DO_CALLBACK and the callbacks are deferred to the completion of the
2772          * earlier iclog. Walk the iclogs in order and make sure that no iclog
2773          * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
2774          * states.
2775          *
2776          * Note that SYNCING|IOABORT is a valid state so we cannot just check
2777          * for ic_state == SYNCING.
2778          */
2779         if (funcdidcallbacks) {
2780                 first_iclog = iclog = log->l_iclog;
2781                 do {
2782                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2783                         /*
2784                          * Terminate the loop if iclogs are found in states
2785                          * which will cause other threads to clean up iclogs.
2786                          *
2787                          * SYNCING - i/o completion will go through logs
2788                          * DONE_SYNC - interrupt thread should be waiting for
2789                          *              l_icloglock
2790                          * IOERROR - give up hope all ye who enter here
2791                          */
2792                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2793                             iclog->ic_state & XLOG_STATE_SYNCING ||
2794                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2795                             iclog->ic_state == XLOG_STATE_IOERROR )
2796                                 break;
2797                         iclog = iclog->ic_next;
2798                 } while (first_iclog != iclog);
2799         }
2800 #endif
2801
2802         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2803                 wake = 1;
2804         spin_unlock(&log->l_icloglock);
2805
2806         if (wake)
2807                 wake_up_all(&log->l_flush_wait);
2808 }
2809
2810
2811 /*
2812  * Finish transitioning this iclog to the dirty state.
2813  *
2814  * Make sure that we completely execute this routine only when this is
2815  * the last call to the iclog.  There is a good chance that iclog flushes,
2816  * when we reach the end of the physical log, get turned into 2 separate
2817  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2818  * routine.  By using the reference count bwritecnt, we guarantee that only
2819  * the second completion goes through.
2820  *
2821  * Callbacks could take time, so they are done outside the scope of the
2822  * global state machine log lock.
2823  */
2824 STATIC void
2825 xlog_state_done_syncing(
2826         xlog_in_core_t  *iclog,
2827         int             aborted)
2828 {
2829         struct xlog        *log = iclog->ic_log;
2830
2831         spin_lock(&log->l_icloglock);
2832
2833         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2834                iclog->ic_state == XLOG_STATE_IOERROR);
2835         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2836         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2837
2838
2839         /*
2840          * If we got an error, either on the first buffer, or in the case of
2841          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2842          * and none should ever be attempted to be written to disk
2843          * again.
2844          */
2845         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2846                 if (--iclog->ic_bwritecnt == 1) {
2847                         spin_unlock(&log->l_icloglock);
2848                         return;
2849                 }
2850                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2851         }
2852
2853         /*
2854          * Someone could be sleeping prior to writing out the next
2855          * iclog buffer, we wake them all, one will get to do the
2856          * I/O, the others get to wait for the result.
2857          */
2858         wake_up_all(&iclog->ic_write_wait);
2859         spin_unlock(&log->l_icloglock);
2860         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2861 }       /* xlog_state_done_syncing */
2862
2863
2864 /*
2865  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2866  * sleep.  We wait on the flush queue on the head iclog as that should be
2867  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2868  * we will wait here and all new writes will sleep until a sync completes.
2869  *
2870  * The in-core logs are used in a circular fashion. They are not used
2871  * out-of-order even when an iclog past the head is free.
2872  *
2873  * return:
2874  *      * log_offset where xlog_write() can start writing into the in-core
2875  *              log's data space.
2876  *      * in-core log pointer to which xlog_write() should write.
2877  *      * boolean indicating this is a continued write to an in-core log.
2878  *              If this is the last write, then the in-core log's offset field
2879  *              needs to be incremented, depending on the amount of data which
2880  *              is copied.
2881  */
2882 STATIC int
2883 xlog_state_get_iclog_space(
2884         struct xlog             *log,
2885         int                     len,
2886         struct xlog_in_core     **iclogp,
2887         struct xlog_ticket      *ticket,
2888         int                     *continued_write,
2889         int                     *logoffsetp)
2890 {
2891         int               log_offset;
2892         xlog_rec_header_t *head;
2893         xlog_in_core_t    *iclog;
2894         int               error;
2895
2896 restart:
2897         spin_lock(&log->l_icloglock);
2898         if (XLOG_FORCED_SHUTDOWN(log)) {
2899                 spin_unlock(&log->l_icloglock);
2900                 return -EIO;
2901         }
2902
2903         iclog = log->l_iclog;
2904         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2905                 XFS_STATS_INC(log->l_mp, xs_log_noiclogs);
2906
2907                 /* Wait for log writes to have flushed */
2908                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2909                 goto restart;
2910         }
2911
2912         head = &iclog->ic_header;
2913
2914         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2915         log_offset = iclog->ic_offset;
2916
2917         /* On the 1st write to an iclog, figure out lsn.  This works
2918          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2919          * committing to.  If the offset is set, that's how many blocks
2920          * must be written.
2921          */
2922         if (log_offset == 0) {
2923                 ticket->t_curr_res -= log->l_iclog_hsize;
2924                 xlog_tic_add_region(ticket,
2925                                     log->l_iclog_hsize,
2926                                     XLOG_REG_TYPE_LRHEADER);
2927                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2928                 head->h_lsn = cpu_to_be64(
2929                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2930                 ASSERT(log->l_curr_block >= 0);
2931         }
2932
2933         /* If there is enough room to write everything, then do it.  Otherwise,
2934          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2935          * bit is on, so this will get flushed out.  Don't update ic_offset
2936          * until you know exactly how many bytes get copied.  Therefore, wait
2937          * until later to update ic_offset.
2938          *
2939          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2940          * can fit into remaining data section.
2941          */
2942         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2943                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2944
2945                 /*
2946                  * If I'm the only one writing to this iclog, sync it to disk.
2947                  * We need to do an atomic compare and decrement here to avoid
2948                  * racing with concurrent atomic_dec_and_lock() calls in
2949                  * xlog_state_release_iclog() when there is more than one
2950                  * reference to the iclog.
2951                  */
2952                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2953                         /* we are the only one */
2954                         spin_unlock(&log->l_icloglock);
2955                         error = xlog_state_release_iclog(log, iclog);
2956                         if (error)
2957                                 return error;
2958                 } else {
2959                         spin_unlock(&log->l_icloglock);
2960                 }
2961                 goto restart;
2962         }
2963
2964         /* Do we have enough room to write the full amount in the remainder
2965          * of this iclog?  Or must we continue a write on the next iclog and
2966          * mark this iclog as completely taken?  In the case where we switch
2967          * iclogs (to mark it taken), this particular iclog will release/sync
2968          * to disk in xlog_write().
2969          */
2970         if (len <= iclog->ic_size - iclog->ic_offset) {
2971                 *continued_write = 0;
2972                 iclog->ic_offset += len;
2973         } else {
2974                 *continued_write = 1;
2975                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2976         }
2977         *iclogp = iclog;
2978
2979         ASSERT(iclog->ic_offset <= iclog->ic_size);
2980         spin_unlock(&log->l_icloglock);
2981
2982         *logoffsetp = log_offset;
2983         return 0;
2984 }       /* xlog_state_get_iclog_space */
2985
2986 /* The first cnt-1 times through here we don't need to
2987  * move the grant write head because the permanent
2988  * reservation has reserved cnt times the unit amount.
2989  * Release part of current permanent unit reservation and
2990  * reset current reservation to be one units worth.  Also
2991  * move grant reservation head forward.
2992  */
2993 STATIC void
2994 xlog_regrant_reserve_log_space(
2995         struct xlog             *log,
2996         struct xlog_ticket      *ticket)
2997 {
2998         trace_xfs_log_regrant_reserve_enter(log, ticket);
2999
3000         if (ticket->t_cnt > 0)
3001                 ticket->t_cnt--;
3002
3003         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
3004                                         ticket->t_curr_res);
3005         xlog_grant_sub_space(log, &log->l_write_head.grant,
3006                                         ticket->t_curr_res);
3007         ticket->t_curr_res = ticket->t_unit_res;
3008         xlog_tic_reset_res(ticket);
3009
3010         trace_xfs_log_regrant_reserve_sub(log, ticket);
3011
3012         /* just return if we still have some of the pre-reserved space */
3013         if (ticket->t_cnt > 0)
3014                 return;
3015
3016         xlog_grant_add_space(log, &log->l_reserve_head.grant,
3017                                         ticket->t_unit_res);
3018
3019         trace_xfs_log_regrant_reserve_exit(log, ticket);
3020
3021         ticket->t_curr_res = ticket->t_unit_res;
3022         xlog_tic_reset_res(ticket);
3023 }       /* xlog_regrant_reserve_log_space */
3024
3025
3026 /*
3027  * Give back the space left from a reservation.
3028  *
3029  * All the information we need to make a correct determination of space left
3030  * is present.  For non-permanent reservations, things are quite easy.  The
3031  * count should have been decremented to zero.  We only need to deal with the
3032  * space remaining in the current reservation part of the ticket.  If the
3033  * ticket contains a permanent reservation, there may be left over space which
3034  * needs to be released.  A count of N means that N-1 refills of the current
3035  * reservation can be done before we need to ask for more space.  The first
3036  * one goes to fill up the first current reservation.  Once we run out of
3037  * space, the count will stay at zero and the only space remaining will be
3038  * in the current reservation field.
3039  */
3040 STATIC void
3041 xlog_ungrant_log_space(
3042         struct xlog             *log,
3043         struct xlog_ticket      *ticket)
3044 {
3045         int     bytes;
3046
3047         if (ticket->t_cnt > 0)
3048                 ticket->t_cnt--;
3049
3050         trace_xfs_log_ungrant_enter(log, ticket);
3051         trace_xfs_log_ungrant_sub(log, ticket);
3052
3053         /*
3054          * If this is a permanent reservation ticket, we may be able to free
3055          * up more space based on the remaining count.
3056          */
3057         bytes = ticket->t_curr_res;
3058         if (ticket->t_cnt > 0) {
3059                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3060                 bytes += ticket->t_unit_res*ticket->t_cnt;
3061         }
3062
3063         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3064         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3065
3066         trace_xfs_log_ungrant_exit(log, ticket);
3067
3068         xfs_log_space_wake(log->l_mp);
3069 }
3070
3071 /*
3072  * Flush iclog to disk if this is the last reference to the given iclog and
3073  * the WANT_SYNC bit is set.
3074  *
3075  * When this function is entered, the iclog is not necessarily in the
3076  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3077  *
3078  *
3079  */
3080 STATIC int
3081 xlog_state_release_iclog(
3082         struct xlog             *log,
3083         struct xlog_in_core     *iclog)
3084 {
3085         int             sync = 0;       /* do we sync? */
3086
3087         if (iclog->ic_state & XLOG_STATE_IOERROR)
3088                 return -EIO;
3089
3090         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3091         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3092                 return 0;
3093
3094         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3095                 spin_unlock(&log->l_icloglock);
3096                 return -EIO;
3097         }
3098         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3099                iclog->ic_state == XLOG_STATE_WANT_SYNC);
3100
3101         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3102                 /* update tail before writing to iclog */
3103                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3104                 sync++;
3105                 iclog->ic_state = XLOG_STATE_SYNCING;
3106                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3107                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3108                 /* cycle incremented when incrementing curr_block */
3109         }
3110         spin_unlock(&log->l_icloglock);
3111
3112         /*
3113          * We let the log lock go, so it's possible that we hit a log I/O
3114          * error or some other SHUTDOWN condition that marks the iclog
3115          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3116          * this iclog has consistent data, so we ignore IOERROR
3117          * flags after this point.
3118          */
3119         if (sync)
3120                 return xlog_sync(log, iclog);
3121         return 0;
3122 }       /* xlog_state_release_iclog */
3123
3124
3125 /*
3126  * This routine will mark the current iclog in the ring as WANT_SYNC
3127  * and move the current iclog pointer to the next iclog in the ring.
3128  * When this routine is called from xlog_state_get_iclog_space(), the
3129  * exact size of the iclog has not yet been determined.  All we know is
3130  * that every data block.  We have run out of space in this log record.
3131  */
3132 STATIC void
3133 xlog_state_switch_iclogs(
3134         struct xlog             *log,
3135         struct xlog_in_core     *iclog,
3136         int                     eventual_size)
3137 {
3138         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3139         if (!eventual_size)
3140                 eventual_size = iclog->ic_offset;
3141         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3142         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3143         log->l_prev_block = log->l_curr_block;
3144         log->l_prev_cycle = log->l_curr_cycle;
3145
3146         /* roll log?: ic_offset changed later */
3147         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3148
3149         /* Round up to next log-sunit */
3150         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3151             log->l_mp->m_sb.sb_logsunit > 1) {
3152                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3153                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3154         }
3155
3156         if (log->l_curr_block >= log->l_logBBsize) {
3157                 /*
3158                  * Rewind the current block before the cycle is bumped to make
3159                  * sure that the combined LSN never transiently moves forward
3160                  * when the log wraps to the next cycle. This is to support the
3161                  * unlocked sample of these fields from xlog_valid_lsn(). Most
3162                  * other cases should acquire l_icloglock.
3163                  */
3164                 log->l_curr_block -= log->l_logBBsize;
3165                 ASSERT(log->l_curr_block >= 0);
3166                 smp_wmb();
3167                 log->l_curr_cycle++;
3168                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3169                         log->l_curr_cycle++;
3170         }
3171         ASSERT(iclog == log->l_iclog);
3172         log->l_iclog = iclog->ic_next;
3173 }       /* xlog_state_switch_iclogs */
3174
3175 /*
3176  * Write out all data in the in-core log as of this exact moment in time.
3177  *
3178  * Data may be written to the in-core log during this call.  However,
3179  * we don't guarantee this data will be written out.  A change from past
3180  * implementation means this routine will *not* write out zero length LRs.
3181  *
3182  * Basically, we try and perform an intelligent scan of the in-core logs.
3183  * If we determine there is no flushable data, we just return.  There is no
3184  * flushable data if:
3185  *
3186  *      1. the current iclog is active and has no data; the previous iclog
3187  *              is in the active or dirty state.
3188  *      2. the current iclog is drity, and the previous iclog is in the
3189  *              active or dirty state.
3190  *
3191  * We may sleep if:
3192  *
3193  *      1. the current iclog is not in the active nor dirty state.
3194  *      2. the current iclog dirty, and the previous iclog is not in the
3195  *              active nor dirty state.
3196  *      3. the current iclog is active, and there is another thread writing
3197  *              to this particular iclog.
3198  *      4. a) the current iclog is active and has no other writers
3199  *         b) when we return from flushing out this iclog, it is still
3200  *              not in the active nor dirty state.
3201  */
3202 int
3203 _xfs_log_force(
3204         struct xfs_mount        *mp,
3205         uint                    flags,
3206         int                     *log_flushed)
3207 {
3208         struct xlog             *log = mp->m_log;
3209         struct xlog_in_core     *iclog;
3210         xfs_lsn_t               lsn;
3211
3212         XFS_STATS_INC(mp, xs_log_force);
3213
3214         xlog_cil_force(log);
3215
3216         spin_lock(&log->l_icloglock);
3217
3218         iclog = log->l_iclog;
3219         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3220                 spin_unlock(&log->l_icloglock);
3221                 return -EIO;
3222         }
3223
3224         /* If the head iclog is not active nor dirty, we just attach
3225          * ourselves to the head and go to sleep.
3226          */
3227         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3228             iclog->ic_state == XLOG_STATE_DIRTY) {
3229                 /*
3230                  * If the head is dirty or (active and empty), then
3231                  * we need to look at the previous iclog.  If the previous
3232                  * iclog is active or dirty we are done.  There is nothing
3233                  * to sync out.  Otherwise, we attach ourselves to the
3234                  * previous iclog and go to sleep.
3235                  */
3236                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3237                     (atomic_read(&iclog->ic_refcnt) == 0
3238                      && iclog->ic_offset == 0)) {
3239                         iclog = iclog->ic_prev;
3240                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3241                             iclog->ic_state == XLOG_STATE_DIRTY)
3242                                 goto no_sleep;
3243                         else
3244                                 goto maybe_sleep;
3245                 } else {
3246                         if (atomic_read(&iclog->ic_refcnt) == 0) {
3247                                 /* We are the only one with access to this
3248                                  * iclog.  Flush it out now.  There should
3249                                  * be a roundoff of zero to show that someone
3250                                  * has already taken care of the roundoff from
3251                                  * the previous sync.
3252                                  */
3253                                 atomic_inc(&iclog->ic_refcnt);
3254                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3255                                 xlog_state_switch_iclogs(log, iclog, 0);
3256                                 spin_unlock(&log->l_icloglock);
3257
3258                                 if (xlog_state_release_iclog(log, iclog))
3259                                         return -EIO;
3260
3261                                 if (log_flushed)
3262                                         *log_flushed = 1;
3263                                 spin_lock(&log->l_icloglock);
3264                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3265                                     iclog->ic_state != XLOG_STATE_DIRTY)
3266                                         goto maybe_sleep;
3267                                 else
3268                                         goto no_sleep;
3269                         } else {
3270                                 /* Someone else is writing to this iclog.
3271                                  * Use its call to flush out the data.  However,
3272                                  * the other thread may not force out this LR,
3273                                  * so we mark it WANT_SYNC.
3274                                  */
3275                                 xlog_state_switch_iclogs(log, iclog, 0);
3276                                 goto maybe_sleep;
3277                         }
3278                 }
3279         }
3280
3281         /* By the time we come around again, the iclog could've been filled
3282          * which would give it another lsn.  If we have a new lsn, just
3283          * return because the relevant data has been flushed.
3284          */
3285 maybe_sleep:
3286         if (flags & XFS_LOG_SYNC) {
3287                 /*
3288                  * We must check if we're shutting down here, before
3289                  * we wait, while we're holding the l_icloglock.
3290                  * Then we check again after waking up, in case our
3291                  * sleep was disturbed by a bad news.
3292                  */
3293                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3294                         spin_unlock(&log->l_icloglock);
3295                         return -EIO;
3296                 }
3297                 XFS_STATS_INC(mp, xs_log_force_sleep);
3298                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3299                 /*
3300                  * No need to grab the log lock here since we're
3301                  * only deciding whether or not to return EIO
3302                  * and the memory read should be atomic.
3303                  */
3304                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3305                         return -EIO;
3306                 if (log_flushed)
3307                         *log_flushed = 1;
3308         } else {
3309
3310 no_sleep:
3311                 spin_unlock(&log->l_icloglock);
3312         }
3313         return 0;
3314 }
3315
3316 /*
3317  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3318  * about errors or whether the log was flushed or not. This is the normal
3319  * interface to use when trying to unpin items or move the log forward.
3320  */
3321 void
3322 xfs_log_force(
3323         xfs_mount_t     *mp,
3324         uint            flags)
3325 {
3326         int     error;
3327
3328         trace_xfs_log_force(mp, 0, _RET_IP_);
3329         error = _xfs_log_force(mp, flags, NULL);
3330         if (error)
3331                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3332 }
3333
3334 /*
3335  * Force the in-core log to disk for a specific LSN.
3336  *
3337  * Find in-core log with lsn.
3338  *      If it is in the DIRTY state, just return.
3339  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3340  *              state and go to sleep or return.
3341  *      If it is in any other state, go to sleep or return.
3342  *
3343  * Synchronous forces are implemented with a signal variable. All callers
3344  * to force a given lsn to disk will wait on a the sv attached to the
3345  * specific in-core log.  When given in-core log finally completes its
3346  * write to disk, that thread will wake up all threads waiting on the
3347  * sv.
3348  */
3349 int
3350 _xfs_log_force_lsn(
3351         struct xfs_mount        *mp,
3352         xfs_lsn_t               lsn,
3353         uint                    flags,
3354         int                     *log_flushed)
3355 {
3356         struct xlog             *log = mp->m_log;
3357         struct xlog_in_core     *iclog;
3358         int                     already_slept = 0;
3359
3360         ASSERT(lsn != 0);
3361
3362         XFS_STATS_INC(mp, xs_log_force);
3363
3364         lsn = xlog_cil_force_lsn(log, lsn);
3365         if (lsn == NULLCOMMITLSN)
3366                 return 0;
3367
3368 try_again:
3369         spin_lock(&log->l_icloglock);
3370         iclog = log->l_iclog;
3371         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3372                 spin_unlock(&log->l_icloglock);
3373                 return -EIO;
3374         }
3375
3376         do {
3377                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3378                         iclog = iclog->ic_next;
3379                         continue;
3380                 }
3381
3382                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3383                         spin_unlock(&log->l_icloglock);
3384                         return 0;
3385                 }
3386
3387                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3388                         /*
3389                          * We sleep here if we haven't already slept (e.g.
3390                          * this is the first time we've looked at the correct
3391                          * iclog buf) and the buffer before us is going to
3392                          * be sync'ed. The reason for this is that if we
3393                          * are doing sync transactions here, by waiting for
3394                          * the previous I/O to complete, we can allow a few
3395                          * more transactions into this iclog before we close
3396                          * it down.
3397                          *
3398                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3399                          * up the refcnt so we can release the log (which
3400                          * drops the ref count).  The state switch keeps new
3401                          * transaction commits from using this buffer.  When
3402                          * the current commits finish writing into the buffer,
3403                          * the refcount will drop to zero and the buffer will
3404                          * go out then.
3405                          */
3406                         if (!already_slept &&
3407                             (iclog->ic_prev->ic_state &
3408                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3409                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3410
3411                                 XFS_STATS_INC(mp, xs_log_force_sleep);
3412
3413                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3414                                                         &log->l_icloglock);
3415                                 if (log_flushed)
3416                                         *log_flushed = 1;
3417                                 already_slept = 1;
3418                                 goto try_again;
3419                         }
3420                         atomic_inc(&iclog->ic_refcnt);
3421                         xlog_state_switch_iclogs(log, iclog, 0);
3422                         spin_unlock(&log->l_icloglock);
3423                         if (xlog_state_release_iclog(log, iclog))
3424                                 return -EIO;
3425                         if (log_flushed)
3426                                 *log_flushed = 1;
3427                         spin_lock(&log->l_icloglock);
3428                 }
3429
3430                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3431                     !(iclog->ic_state &
3432                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3433                         /*
3434                          * Don't wait on completion if we know that we've
3435                          * gotten a log write error.
3436                          */
3437                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3438                                 spin_unlock(&log->l_icloglock);
3439                                 return -EIO;
3440                         }
3441                         XFS_STATS_INC(mp, xs_log_force_sleep);
3442                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3443                         /*
3444                          * No need to grab the log lock here since we're
3445                          * only deciding whether or not to return EIO
3446                          * and the memory read should be atomic.
3447                          */
3448                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3449                                 return -EIO;
3450
3451                         if (log_flushed)
3452                                 *log_flushed = 1;
3453                 } else {                /* just return */
3454                         spin_unlock(&log->l_icloglock);
3455                 }
3456
3457                 return 0;
3458         } while (iclog != log->l_iclog);
3459
3460         spin_unlock(&log->l_icloglock);
3461         return 0;
3462 }
3463
3464 /*
3465  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3466  * about errors or whether the log was flushed or not. This is the normal
3467  * interface to use when trying to unpin items or move the log forward.
3468  */
3469 void
3470 xfs_log_force_lsn(
3471         xfs_mount_t     *mp,
3472         xfs_lsn_t       lsn,
3473         uint            flags)
3474 {
3475         int     error;
3476
3477         trace_xfs_log_force(mp, lsn, _RET_IP_);
3478         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3479         if (error)
3480                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3481 }
3482
3483 /*
3484  * Called when we want to mark the current iclog as being ready to sync to
3485  * disk.
3486  */
3487 STATIC void
3488 xlog_state_want_sync(
3489         struct xlog             *log,
3490         struct xlog_in_core     *iclog)
3491 {
3492         assert_spin_locked(&log->l_icloglock);
3493
3494         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3495                 xlog_state_switch_iclogs(log, iclog, 0);
3496         } else {
3497                 ASSERT(iclog->ic_state &
3498                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3499         }
3500 }
3501
3502
3503 /*****************************************************************************
3504  *
3505  *              TICKET functions
3506  *
3507  *****************************************************************************
3508  */
3509
3510 /*
3511  * Free a used ticket when its refcount falls to zero.
3512  */
3513 void
3514 xfs_log_ticket_put(
3515         xlog_ticket_t   *ticket)
3516 {
3517         ASSERT(atomic_read(&ticket->t_ref) > 0);
3518         if (atomic_dec_and_test(&ticket->t_ref))
3519                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3520 }
3521
3522 xlog_ticket_t *
3523 xfs_log_ticket_get(
3524         xlog_ticket_t   *ticket)
3525 {
3526         ASSERT(atomic_read(&ticket->t_ref) > 0);
3527         atomic_inc(&ticket->t_ref);
3528         return ticket;
3529 }
3530
3531 /*
3532  * Figure out the total log space unit (in bytes) that would be
3533  * required for a log ticket.
3534  */
3535 int
3536 xfs_log_calc_unit_res(
3537         struct xfs_mount        *mp,
3538         int                     unit_bytes)
3539 {
3540         struct xlog             *log = mp->m_log;
3541         int                     iclog_space;
3542         uint                    num_headers;
3543
3544         /*
3545          * Permanent reservations have up to 'cnt'-1 active log operations
3546          * in the log.  A unit in this case is the amount of space for one
3547          * of these log operations.  Normal reservations have a cnt of 1
3548          * and their unit amount is the total amount of space required.
3549          *
3550          * The following lines of code account for non-transaction data
3551          * which occupy space in the on-disk log.
3552          *
3553          * Normal form of a transaction is:
3554          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3555          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3556          *
3557          * We need to account for all the leadup data and trailer data
3558          * around the transaction data.
3559          * And then we need to account for the worst case in terms of using
3560          * more space.
3561          * The worst case will happen if:
3562          * - the placement of the transaction happens to be such that the
3563          *   roundoff is at its maximum
3564          * - the transaction data is synced before the commit record is synced
3565          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3566          *   Therefore the commit record is in its own Log Record.
3567          *   This can happen as the commit record is called with its
3568          *   own region to xlog_write().
3569          *   This then means that in the worst case, roundoff can happen for
3570          *   the commit-rec as well.
3571          *   The commit-rec is smaller than padding in this scenario and so it is
3572          *   not added separately.
3573          */
3574
3575         /* for trans header */
3576         unit_bytes += sizeof(xlog_op_header_t);
3577         unit_bytes += sizeof(xfs_trans_header_t);
3578
3579         /* for start-rec */
3580         unit_bytes += sizeof(xlog_op_header_t);
3581
3582         /*
3583          * for LR headers - the space for data in an iclog is the size minus
3584          * the space used for the headers. If we use the iclog size, then we
3585          * undercalculate the number of headers required.
3586          *
3587          * Furthermore - the addition of op headers for split-recs might
3588          * increase the space required enough to require more log and op
3589          * headers, so take that into account too.
3590          *
3591          * IMPORTANT: This reservation makes the assumption that if this
3592          * transaction is the first in an iclog and hence has the LR headers
3593          * accounted to it, then the remaining space in the iclog is
3594          * exclusively for this transaction.  i.e. if the transaction is larger
3595          * than the iclog, it will be the only thing in that iclog.
3596          * Fundamentally, this means we must pass the entire log vector to
3597          * xlog_write to guarantee this.
3598          */
3599         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3600         num_headers = howmany(unit_bytes, iclog_space);
3601
3602         /* for split-recs - ophdrs added when data split over LRs */
3603         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3604
3605         /* add extra header reservations if we overrun */
3606         while (!num_headers ||
3607                howmany(unit_bytes, iclog_space) > num_headers) {
3608                 unit_bytes += sizeof(xlog_op_header_t);
3609                 num_headers++;
3610         }
3611         unit_bytes += log->l_iclog_hsize * num_headers;
3612
3613         /* for commit-rec LR header - note: padding will subsume the ophdr */
3614         unit_bytes += log->l_iclog_hsize;
3615
3616         /* for roundoff padding for transaction data and one for commit record */
3617         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3618                 /* log su roundoff */
3619                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3620         } else {
3621                 /* BB roundoff */
3622                 unit_bytes += 2 * BBSIZE;
3623         }
3624
3625         return unit_bytes;
3626 }
3627
3628 /*
3629  * Allocate and initialise a new log ticket.
3630  */
3631 struct xlog_ticket *
3632 xlog_ticket_alloc(
3633         struct xlog             *log,
3634         int                     unit_bytes,
3635         int                     cnt,
3636         char                    client,
3637         bool                    permanent,
3638         xfs_km_flags_t          alloc_flags)
3639 {
3640         struct xlog_ticket      *tic;
3641         int                     unit_res;
3642
3643         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3644         if (!tic)
3645                 return NULL;
3646
3647         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3648
3649         atomic_set(&tic->t_ref, 1);
3650         tic->t_task             = current;
3651         INIT_LIST_HEAD(&tic->t_queue);
3652         tic->t_unit_res         = unit_res;
3653         tic->t_curr_res         = unit_res;
3654         tic->t_cnt              = cnt;
3655         tic->t_ocnt             = cnt;
3656         tic->t_tid              = prandom_u32();
3657         tic->t_clientid         = client;
3658         tic->t_flags            = XLOG_TIC_INITED;
3659         if (permanent)
3660                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3661
3662         xlog_tic_reset_res(tic);
3663
3664         return tic;
3665 }
3666
3667
3668 /******************************************************************************
3669  *
3670  *              Log debug routines
3671  *
3672  ******************************************************************************
3673  */
3674 #if defined(DEBUG)
3675 /*
3676  * Make sure that the destination ptr is within the valid data region of
3677  * one of the iclogs.  This uses backup pointers stored in a different
3678  * part of the log in case we trash the log structure.
3679  */
3680 void
3681 xlog_verify_dest_ptr(
3682         struct xlog     *log,
3683         void            *ptr)
3684 {
3685         int i;
3686         int good_ptr = 0;
3687
3688         for (i = 0; i < log->l_iclog_bufs; i++) {
3689                 if (ptr >= log->l_iclog_bak[i] &&
3690                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3691                         good_ptr++;
3692         }
3693
3694         if (!good_ptr)
3695                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3696 }
3697
3698 /*
3699  * Check to make sure the grant write head didn't just over lap the tail.  If
3700  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3701  * the cycles differ by exactly one and check the byte count.
3702  *
3703  * This check is run unlocked, so can give false positives. Rather than assert
3704  * on failures, use a warn-once flag and a panic tag to allow the admin to
3705  * determine if they want to panic the machine when such an error occurs. For
3706  * debug kernels this will have the same effect as using an assert but, unlinke
3707  * an assert, it can be turned off at runtime.
3708  */
3709 STATIC void
3710 xlog_verify_grant_tail(
3711         struct xlog     *log)
3712 {
3713         int             tail_cycle, tail_blocks;
3714         int             cycle, space;
3715
3716         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3717         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3718         if (tail_cycle != cycle) {
3719                 if (cycle - 1 != tail_cycle &&
3720                     !(log->l_flags & XLOG_TAIL_WARN)) {
3721                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3722                                 "%s: cycle - 1 != tail_cycle", __func__);
3723                         log->l_flags |= XLOG_TAIL_WARN;
3724                 }
3725
3726                 if (space > BBTOB(tail_blocks) &&
3727                     !(log->l_flags & XLOG_TAIL_WARN)) {
3728                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3729                                 "%s: space > BBTOB(tail_blocks)", __func__);
3730                         log->l_flags |= XLOG_TAIL_WARN;
3731                 }
3732         }
3733 }
3734
3735 /* check if it will fit */
3736 STATIC void
3737 xlog_verify_tail_lsn(
3738         struct xlog             *log,
3739         struct xlog_in_core     *iclog,
3740         xfs_lsn_t               tail_lsn)
3741 {
3742     int blocks;
3743
3744     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3745         blocks =
3746             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3747         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3748                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3749     } else {
3750         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3751
3752         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3753                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3754
3755         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3756         if (blocks < BTOBB(iclog->ic_offset) + 1)
3757                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3758     }
3759 }       /* xlog_verify_tail_lsn */
3760
3761 /*
3762  * Perform a number of checks on the iclog before writing to disk.
3763  *
3764  * 1. Make sure the iclogs are still circular
3765  * 2. Make sure we have a good magic number
3766  * 3. Make sure we don't have magic numbers in the data
3767  * 4. Check fields of each log operation header for:
3768  *      A. Valid client identifier
3769  *      B. tid ptr value falls in valid ptr space (user space code)
3770  *      C. Length in log record header is correct according to the
3771  *              individual operation headers within record.
3772  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3773  *      log, check the preceding blocks of the physical log to make sure all
3774  *      the cycle numbers agree with the current cycle number.
3775  */
3776 STATIC void
3777 xlog_verify_iclog(
3778         struct xlog             *log,
3779         struct xlog_in_core     *iclog,
3780         int                     count,
3781         bool                    syncing)
3782 {
3783         xlog_op_header_t        *ophead;
3784         xlog_in_core_t          *icptr;
3785         xlog_in_core_2_t        *xhdr;
3786         void                    *base_ptr, *ptr, *p;
3787         ptrdiff_t               field_offset;
3788         __uint8_t               clientid;
3789         int                     len, i, j, k, op_len;
3790         int                     idx;
3791
3792         /* check validity of iclog pointers */
3793         spin_lock(&log->l_icloglock);
3794         icptr = log->l_iclog;
3795         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3796                 ASSERT(icptr);
3797
3798         if (icptr != log->l_iclog)
3799                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3800         spin_unlock(&log->l_icloglock);
3801
3802         /* check log magic numbers */
3803         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3804                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3805
3806         base_ptr = ptr = &iclog->ic_header;
3807         p = &iclog->ic_header;
3808         for (ptr += BBSIZE; ptr < base_ptr + count; ptr += BBSIZE) {
3809                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3810                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3811                                 __func__);
3812         }
3813
3814         /* check fields */
3815         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3816         base_ptr = ptr = iclog->ic_datap;
3817         ophead = ptr;
3818         xhdr = iclog->ic_data;
3819         for (i = 0; i < len; i++) {
3820                 ophead = ptr;
3821
3822                 /* clientid is only 1 byte */
3823                 p = &ophead->oh_clientid;
3824                 field_offset = p - base_ptr;
3825                 if (!syncing || (field_offset & 0x1ff)) {
3826                         clientid = ophead->oh_clientid;
3827                 } else {
3828                         idx = BTOBBT((char *)&ophead->oh_clientid - iclog->ic_datap);
3829                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3830                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3831                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3832                                 clientid = xlog_get_client_id(
3833                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3834                         } else {
3835                                 clientid = xlog_get_client_id(
3836                                         iclog->ic_header.h_cycle_data[idx]);
3837                         }
3838                 }
3839                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3840                         xfs_warn(log->l_mp,
3841                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3842                                 __func__, clientid, ophead,
3843                                 (unsigned long)field_offset);
3844
3845                 /* check length */
3846                 p = &ophead->oh_len;
3847                 field_offset = p - base_ptr;
3848                 if (!syncing || (field_offset & 0x1ff)) {
3849                         op_len = be32_to_cpu(ophead->oh_len);
3850                 } else {
3851                         idx = BTOBBT((uintptr_t)&ophead->oh_len -
3852                                     (uintptr_t)iclog->ic_datap);
3853                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3854                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3855                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3856                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3857                         } else {
3858                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3859                         }
3860                 }
3861                 ptr += sizeof(xlog_op_header_t) + op_len;
3862         }
3863 }       /* xlog_verify_iclog */
3864 #endif
3865
3866 /*
3867  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3868  */
3869 STATIC int
3870 xlog_state_ioerror(
3871         struct xlog     *log)
3872 {
3873         xlog_in_core_t  *iclog, *ic;
3874
3875         iclog = log->l_iclog;
3876         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3877                 /*
3878                  * Mark all the incore logs IOERROR.
3879                  * From now on, no log flushes will result.
3880                  */
3881                 ic = iclog;
3882                 do {
3883                         ic->ic_state = XLOG_STATE_IOERROR;
3884                         ic = ic->ic_next;
3885                 } while (ic != iclog);
3886                 return 0;
3887         }
3888         /*
3889          * Return non-zero, if state transition has already happened.
3890          */
3891         return 1;
3892 }
3893
3894 /*
3895  * This is called from xfs_force_shutdown, when we're forcibly
3896  * shutting down the filesystem, typically because of an IO error.
3897  * Our main objectives here are to make sure that:
3898  *      a. if !logerror, flush the logs to disk. Anything modified
3899  *         after this is ignored.
3900  *      b. the filesystem gets marked 'SHUTDOWN' for all interested
3901  *         parties to find out, 'atomically'.
3902  *      c. those who're sleeping on log reservations, pinned objects and
3903  *          other resources get woken up, and be told the bad news.
3904  *      d. nothing new gets queued up after (b) and (c) are done.
3905  *
3906  * Note: for the !logerror case we need to flush the regions held in memory out
3907  * to disk first. This needs to be done before the log is marked as shutdown,
3908  * otherwise the iclog writes will fail.
3909  */
3910 int
3911 xfs_log_force_umount(
3912         struct xfs_mount        *mp,
3913         int                     logerror)
3914 {
3915         struct xlog     *log;
3916         int             retval;
3917
3918         log = mp->m_log;
3919
3920         /*
3921          * If this happens during log recovery, don't worry about
3922          * locking; the log isn't open for business yet.
3923          */
3924         if (!log ||
3925             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3926                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3927                 if (mp->m_sb_bp)
3928                         mp->m_sb_bp->b_flags |= XBF_DONE;
3929                 return 0;
3930         }
3931
3932         /*
3933          * Somebody could've already done the hard work for us.
3934          * No need to get locks for this.
3935          */
3936         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3937                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3938                 return 1;
3939         }
3940
3941         /*
3942          * Flush all the completed transactions to disk before marking the log
3943          * being shut down. We need to do it in this order to ensure that
3944          * completed operations are safely on disk before we shut down, and that
3945          * we don't have to issue any buffer IO after the shutdown flags are set
3946          * to guarantee this.
3947          */
3948         if (!logerror)
3949                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3950
3951         /*
3952          * mark the filesystem and the as in a shutdown state and wake
3953          * everybody up to tell them the bad news.
3954          */
3955         spin_lock(&log->l_icloglock);
3956         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3957         if (mp->m_sb_bp)
3958                 mp->m_sb_bp->b_flags |= XBF_DONE;
3959
3960         /*
3961          * Mark the log and the iclogs with IO error flags to prevent any
3962          * further log IO from being issued or completed.
3963          */
3964         log->l_flags |= XLOG_IO_ERROR;
3965         retval = xlog_state_ioerror(log);
3966         spin_unlock(&log->l_icloglock);
3967
3968         /*
3969          * We don't want anybody waiting for log reservations after this. That
3970          * means we have to wake up everybody queued up on reserveq as well as
3971          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3972          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3973          * action is protected by the grant locks.
3974          */
3975         xlog_grant_head_wake_all(&log->l_reserve_head);
3976         xlog_grant_head_wake_all(&log->l_write_head);
3977
3978         /*
3979          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3980          * as if the log writes were completed. The abort handling in the log
3981          * item committed callback functions will do this again under lock to
3982          * avoid races.
3983          */
3984         wake_up_all(&log->l_cilp->xc_commit_wait);
3985         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3986
3987 #ifdef XFSERRORDEBUG
3988         {
3989                 xlog_in_core_t  *iclog;
3990
3991                 spin_lock(&log->l_icloglock);
3992                 iclog = log->l_iclog;
3993                 do {
3994                         ASSERT(iclog->ic_callback == 0);
3995                         iclog = iclog->ic_next;
3996                 } while (iclog != log->l_iclog);
3997                 spin_unlock(&log->l_icloglock);
3998         }
3999 #endif
4000         /* return non-zero if log IOERROR transition had already happened */
4001         return retval;
4002 }
4003
4004 STATIC int
4005 xlog_iclogs_empty(
4006         struct xlog     *log)
4007 {
4008         xlog_in_core_t  *iclog;
4009
4010         iclog = log->l_iclog;
4011         do {
4012                 /* endianness does not matter here, zero is zero in
4013                  * any language.
4014                  */
4015                 if (iclog->ic_header.h_num_logops)
4016                         return 0;
4017                 iclog = iclog->ic_next;
4018         } while (iclog != log->l_iclog);
4019         return 1;
4020 }
4021
4022 /*
4023  * Verify that an LSN stamped into a piece of metadata is valid. This is
4024  * intended for use in read verifiers on v5 superblocks.
4025  */
4026 bool
4027 xfs_log_check_lsn(
4028         struct xfs_mount        *mp,
4029         xfs_lsn_t               lsn)
4030 {
4031         struct xlog             *log = mp->m_log;
4032         bool                    valid;
4033
4034         /*
4035          * norecovery mode skips mount-time log processing and unconditionally
4036          * resets the in-core LSN. We can't validate in this mode, but
4037          * modifications are not allowed anyways so just return true.
4038          */
4039         if (mp->m_flags & XFS_MOUNT_NORECOVERY)
4040                 return true;
4041
4042         /*
4043          * Some metadata LSNs are initialized to NULL (e.g., the agfl). This is
4044          * handled by recovery and thus safe to ignore here.
4045          */
4046         if (lsn == NULLCOMMITLSN)
4047                 return true;
4048
4049         valid = xlog_valid_lsn(mp->m_log, lsn);
4050
4051         /* warn the user about what's gone wrong before verifier failure */
4052         if (!valid) {
4053                 spin_lock(&log->l_icloglock);
4054                 xfs_warn(mp,
4055 "Corruption warning: Metadata has LSN (%d:%d) ahead of current LSN (%d:%d). "
4056 "Please unmount and run xfs_repair (>= v4.3) to resolve.",
4057                          CYCLE_LSN(lsn), BLOCK_LSN(lsn),
4058                          log->l_curr_cycle, log->l_curr_block);
4059                 spin_unlock(&log->l_icloglock);
4060         }
4061
4062         return valid;
4063 }