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
  37 kmem_zone_t     *xfs_log_ticket_zone;
  38
  39 /* Local miscellaneous function prototypes */
  40 STATIC int
  41 xlog_commit_record(
  42         struct xlog             *log,
  43         struct xlog_ticket      *ticket,
  44         struct xlog_in_core     **iclog,
  45         xfs_lsn_t               *commitlsnp);
  46
  47 STATIC struct xlog *
  48 xlog_alloc_log(
  49         struct xfs_mount        *mp,
  50         struct xfs_buftarg      *log_target,
  51         xfs_daddr_t             blk_offset,
  52         int                     num_bblks);
  53 STATIC int
  54 xlog_space_left(
  55         struct xlog             *log,
  56         atomic64_t              *head);
  57 STATIC int
  58 xlog_sync(
  59         struct xlog             *log,
  60         struct xlog_in_core     *iclog);
  61 STATIC void
  62 xlog_dealloc_log(
  63         struct xlog             *log);
  64
  65 /* local state machine functions */
  66 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  67 STATIC void
  68 xlog_state_do_callback(
  69         struct xlog             *log,
  70         int                     aborted,
  71         struct xlog_in_core     *iclog);
  72 STATIC int
  73 xlog_state_get_iclog_space(
  74         struct xlog             *log,
  75         int                     len,
  76         struct xlog_in_core     **iclog,
  77         struct xlog_ticket      *ticket,
  78         int                     *continued_write,
  79         int                     *logoffsetp);
  80 STATIC int
  81 xlog_state_release_iclog(
  82         struct xlog             *log,
  83         struct xlog_in_core     *iclog);
  84 STATIC void
  85 xlog_state_switch_iclogs(
  86         struct xlog             *log,
  87         struct xlog_in_core     *iclog,
  88         int                     eventual_size);
  89 STATIC void
  90 xlog_state_want_sync(
  91         struct xlog             *log,
  92         struct xlog_in_core     *iclog);
  93
  94 STATIC void
  95 xlog_grant_push_ail(
  96         struct xlog             *log,
  97         int                     need_bytes);
  98 STATIC void
  99 xlog_regrant_reserve_log_space(
 100         struct xlog             *log,
 101         struct xlog_ticket      *ticket);
 102 STATIC void
 103 xlog_ungrant_log_space(
 104         struct xlog             *log,
 105         struct xlog_ticket      *ticket);
 106
 107 #if defined(DEBUG)
 108 STATIC void
 109 xlog_verify_dest_ptr(
 110         struct xlog             *log,
 111         char                    *ptr);
 112 STATIC void
 113 xlog_verify_grant_tail(
 114         struct xlog *log);
 115 STATIC void
 116 xlog_verify_iclog(
 117         struct xlog             *log,
 118         struct xlog_in_core     *iclog,
 119         int                     count,
 120         bool                    syncing);
 121 STATIC void
 122 xlog_verify_tail_lsn(
 123         struct xlog             *log,
 124         struct xlog_in_core     *iclog,
 125         xfs_lsn_t               tail_lsn);
 126 #else
 127 #define xlog_verify_dest_ptr(a,b)
 128 #define xlog_verify_grant_tail(a)
 129 #define xlog_verify_iclog(a,b,c,d)
 130 #define xlog_verify_tail_lsn(a,b,c)
 131 #endif
 132
 133 STATIC int
 134 xlog_iclogs_empty(
 135         struct xlog             *log);
 136
 137 static void
 138 xlog_grant_sub_space(
 139         struct xlog             *log,
 140         atomic64_t              *head,
 141         int                     bytes)
 142 {
 143         int64_t head_val = atomic64_read(head);
 144         int64_t new, old;
 145
 146         do {
 147                 int     cycle, space;
 148
 149                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 150
 151                 space -= bytes;
 152                 if (space < 0) {
 153                         space += log->l_logsize;
 154                         cycle--;
 155                 }
 156
 157                 old = head_val;
 158                 new = xlog_assign_grant_head_val(cycle, space);
 159                 head_val = atomic64_cmpxchg(head, old, new);
 160         } while (head_val != old);
 161 }
 162
 163 static void
 164 xlog_grant_add_space(
 165         struct xlog             *log,
 166         atomic64_t              *head,
 167         int                     bytes)
 168 {
 169         int64_t head_val = atomic64_read(head);
 170         int64_t new, old;
 171
 172         do {
 173                 int             tmp;
 174                 int             cycle, space;
 175
 176                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 177
 178                 tmp = log->l_logsize - space;
 179                 if (tmp > bytes)
 180                         space += bytes;
 181                 else {
 182                         space = bytes - tmp;
 183                         cycle++;
 184                 }
 185
 186                 old = head_val;
 187                 new = xlog_assign_grant_head_val(cycle, space);
 188                 head_val = atomic64_cmpxchg(head, old, new);
 189         } while (head_val != old);
 190 }
 191
 192 STATIC void
 193 xlog_grant_head_init(
 194         struct xlog_grant_head  *head)
 195 {
 196         xlog_assign_grant_head(&head->grant, 1, 0);
 197         INIT_LIST_HEAD(&head->waiters);
 198         spin_lock_init(&head->lock);
 199 }
 200
 201 STATIC void
 202 xlog_grant_head_wake_all(
 203         struct xlog_grant_head  *head)
 204 {
 205         struct xlog_ticket      *tic;
 206
 207         spin_lock(&head->lock);
 208         list_for_each_entry(tic, &head->waiters, t_queue)
 209                 wake_up_process(tic->t_task);
 210         spin_unlock(&head->lock);
 211 }
 212
 213 static inline int
 214 xlog_ticket_reservation(
 215         struct xlog             *log,
 216         struct xlog_grant_head  *head,
 217         struct xlog_ticket      *tic)
 218 {
 219         if (head == &log->l_write_head) {
 220                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 221                 return tic->t_unit_res;
 222         } else {
 223                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 224                         return tic->t_unit_res * tic->t_cnt;
 225                 else
 226                         return tic->t_unit_res;
 227         }
 228 }
 229
 230 STATIC bool
 231 xlog_grant_head_wake(
 232         struct xlog             *log,
 233         struct xlog_grant_head  *head,
 234         int                     *free_bytes)
 235 {
 236         struct xlog_ticket      *tic;
 237         int                     need_bytes;
 238
 239         list_for_each_entry(tic, &head->waiters, t_queue) {
 240                 need_bytes = xlog_ticket_reservation(log, head, tic);
 241                 if (*free_bytes < need_bytes)
 242                         return false;
 243
 244                 *free_bytes -= need_bytes;
 245                 trace_xfs_log_grant_wake_up(log, tic);
 246                 wake_up_process(tic->t_task);
 247         }
 248
 249         return true;
 250 }
 251
 252 STATIC int
 253 xlog_grant_head_wait(
 254         struct xlog             *log,
 255         struct xlog_grant_head  *head,
 256         struct xlog_ticket      *tic,
 257         int                     need_bytes) __releases(&head->lock)
 258                                             __acquires(&head->lock)
 259 {
 260         list_add_tail(&tic->t_queue, &head->waiters);
 261
 262         do {
 263                 if (XLOG_FORCED_SHUTDOWN(log))
 264                         goto shutdown;
 265                 xlog_grant_push_ail(log, need_bytes);
 266
 267                 __set_current_state(TASK_UNINTERRUPTIBLE);
 268                 spin_unlock(&head->lock);
 269
 270                 XFS_STATS_INC(xs_sleep_logspace);
 271
 272                 trace_xfs_log_grant_sleep(log, tic);
 273                 schedule();
 274                 trace_xfs_log_grant_wake(log, tic);
 275
 276                 spin_lock(&head->lock);
 277                 if (XLOG_FORCED_SHUTDOWN(log))
 278                         goto shutdown;
 279         } while (xlog_space_left(log, &head->grant) < need_bytes);
 280
 281         list_del_init(&tic->t_queue);
 282         return 0;
 283 shutdown:
 284         list_del_init(&tic->t_queue);
 285         return -EIO;
 286 }
 287
 288 /*
 289  * Atomically get the log space required for a log ticket.
 290  *
 291  * Once a ticket gets put onto head->waiters, it will only return after the
 292  * needed reservation is satisfied.
 293  *
 294  * This function is structured so that it has a lock free fast path. This is
 295  * necessary because every new transaction reservation will come through this
 296  * path. Hence any lock will be globally hot if we take it unconditionally on
 297  * every pass.
 298  *
 299  * As tickets are only ever moved on and off head->waiters under head->lock, we
 300  * only need to take that lock if we are going to add the ticket to the queue
 301  * and sleep. We can avoid taking the lock if the ticket was never added to
 302  * head->waiters because the t_queue list head will be empty and we hold the
 303  * only reference to it so it can safely be checked unlocked.
 304  */
 305 STATIC int
 306 xlog_grant_head_check(
 307         struct xlog             *log,
 308         struct xlog_grant_head  *head,
 309         struct xlog_ticket      *tic,
 310         int                     *need_bytes)
 311 {
 312         int                     free_bytes;
 313         int                     error = 0;
 314
 315         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 316
 317         /*
 318          * If there are other waiters on the queue then give them a chance at
 319          * logspace before us.  Wake up the first waiters, if we do not wake
 320          * up all the waiters then go to sleep waiting for more free space,
 321          * otherwise try to get some space for this transaction.
 322          */
 323         *need_bytes = xlog_ticket_reservation(log, head, tic);
 324         free_bytes = xlog_space_left(log, &head->grant);
 325         if (!list_empty_careful(&head->waiters)) {
 326                 spin_lock(&head->lock);
 327                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 328                     free_bytes < *need_bytes) {
 329                         error = xlog_grant_head_wait(log, head, tic,
 330                                                      *need_bytes);
 331                 }
 332                 spin_unlock(&head->lock);
 333         } else if (free_bytes < *need_bytes) {
 334                 spin_lock(&head->lock);
 335                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 336                 spin_unlock(&head->lock);
 337         }
 338
 339         return error;
 340 }
 341
 342 static void
 343 xlog_tic_reset_res(xlog_ticket_t *tic)
 344 {
 345         tic->t_res_num = 0;
 346         tic->t_res_arr_sum = 0;
 347         tic->t_res_num_ophdrs = 0;
 348 }
 349
 350 static void
 351 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 352 {
 353         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 354                 /* add to overflow and start again */
 355                 tic->t_res_o_flow += tic->t_res_arr_sum;
 356                 tic->t_res_num = 0;
 357                 tic->t_res_arr_sum = 0;
 358         }
 359
 360         tic->t_res_arr[tic->t_res_num].r_len = len;
 361         tic->t_res_arr[tic->t_res_num].r_type = type;
 362         tic->t_res_arr_sum += len;
 363         tic->t_res_num++;
 364 }
 365
 366 /*
 367  * Replenish the byte reservation required by moving the grant write head.
 368  */
 369 int
 370 xfs_log_regrant(
 371         struct xfs_mount        *mp,
 372         struct xlog_ticket      *tic)
 373 {
 374         struct xlog             *log = mp->m_log;
 375         int                     need_bytes;
 376         int                     error = 0;
 377
 378         if (XLOG_FORCED_SHUTDOWN(log))
 379                 return -EIO;
 380
 381         XFS_STATS_INC(xs_try_logspace);
 382
 383         /*
 384          * This is a new transaction on the ticket, so we need to change the
 385          * transaction ID so that the next transaction has a different TID in
 386          * the log. Just add one to the existing tid so that we can see chains
 387          * of rolling transactions in the log easily.
 388          */
 389         tic->t_tid++;
 390
 391         xlog_grant_push_ail(log, tic->t_unit_res);
 392
 393         tic->t_curr_res = tic->t_unit_res;
 394         xlog_tic_reset_res(tic);
 395
 396         if (tic->t_cnt > 0)
 397                 return 0;
 398
 399         trace_xfs_log_regrant(log, tic);
 400
 401         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 402                                       &need_bytes);
 403         if (error)
 404                 goto out_error;
 405
 406         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 407         trace_xfs_log_regrant_exit(log, tic);
 408         xlog_verify_grant_tail(log);
 409         return 0;
 410
 411 out_error:
 412         /*
 413          * If we are failing, make sure the ticket doesn't have any current
 414          * reservations.  We don't want to add this back when the ticket/
 415          * transaction gets cancelled.
 416          */
 417         tic->t_curr_res = 0;
 418         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 419         return error;
 420 }
 421
 422 /*
 423  * Reserve log space and return a ticket corresponding the reservation.
 424  *
 425  * Each reservation is going to reserve extra space for a log record header.
 426  * When writes happen to the on-disk log, we don't subtract the length of the
 427  * log record header from any reservation.  By wasting space in each
 428  * reservation, we prevent over allocation problems.
 429  */
 430 int
 431 xfs_log_reserve(
 432         struct xfs_mount        *mp,
 433         int                     unit_bytes,
 434         int                     cnt,
 435         struct xlog_ticket      **ticp,
 436         __uint8_t               client,
 437         bool                    permanent,
 438         uint                    t_type)
 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(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         tic->t_trans_type = t_type;
 459         *ticp = tic;
 460
 461         xlog_grant_push_ail(log, tic->t_cnt ? tic->t_unit_res * tic->t_cnt
 462                                             : tic->t_unit_res);
 463
 464         trace_xfs_log_reserve(log, tic);
 465
 466         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 467                                       &need_bytes);
 468         if (error)
 469                 goto out_error;
 470
 471         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 472         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 473         trace_xfs_log_reserve_exit(log, tic);
 474         xlog_verify_grant_tail(log);
 475         return 0;
 476
 477 out_error:
 478         /*
 479          * If we are failing, make sure the ticket doesn't have any current
 480          * reservations.  We don't want to add this back when the ticket/
 481          * transaction gets cancelled.
 482          */
 483         tic->t_curr_res = 0;
 484         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 485         return error;
 486 }
 487
 488
 489 /*
 490  * NOTES:
 491  *
 492  *      1. currblock field gets updated at startup and after in-core logs
 493  *              marked as with WANT_SYNC.
 494  */
 495
 496 /*
 497  * This routine is called when a user of a log manager ticket is done with
 498  * the reservation.  If the ticket was ever used, then a commit record for
 499  * the associated transaction is written out as a log operation header with
 500  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 501  * a given ticket.  If the ticket was one with a permanent reservation, then
 502  * a few operations are done differently.  Permanent reservation tickets by
 503  * default don't release the reservation.  They just commit the current
 504  * transaction with the belief that the reservation is still needed.  A flag
 505  * must be passed in before permanent reservations are actually released.
 506  * When these type of tickets are not released, they need to be set into
 507  * the inited state again.  By doing this, a start record will be written
 508  * out when the next write occurs.
 509  */
 510 xfs_lsn_t
 511 xfs_log_done(
 512         struct xfs_mount        *mp,
 513         struct xlog_ticket      *ticket,
 514         struct xlog_in_core     **iclog,
 515         uint                    flags)
 516 {
 517         struct xlog             *log = mp->m_log;
 518         xfs_lsn_t               lsn = 0;
 519
 520         if (XLOG_FORCED_SHUTDOWN(log) ||
 521             /*
 522              * If nothing was ever written, don't write out commit record.
 523              * If we get an error, just continue and give back the log ticket.
 524              */
 525             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 526              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 527                 lsn = (xfs_lsn_t) -1;
 528                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
 529                         flags |= XFS_LOG_REL_PERM_RESERV;
 530                 }
 531         }
 532
 533
 534         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
 535             (flags & XFS_LOG_REL_PERM_RESERV)) {
 536                 trace_xfs_log_done_nonperm(log, ticket);
 537
 538                 /*
 539                  * Release ticket if not permanent reservation or a specific
 540                  * request has been made to release a permanent reservation.
 541                  */
 542                 xlog_ungrant_log_space(log, ticket);
 543                 xfs_log_ticket_put(ticket);
 544         } else {
 545                 trace_xfs_log_done_perm(log, ticket);
 546
 547                 xlog_regrant_reserve_log_space(log, ticket);
 548                 /* If this ticket was a permanent reservation and we aren't
 549                  * trying to release it, reset the inited flags; so next time
 550                  * we write, a start record will be written out.
 551                  */
 552                 ticket->t_flags |= XLOG_TIC_INITED;
 553         }
 554
 555         return lsn;
 556 }
 557
 558 /*
 559  * Attaches a new iclog I/O completion callback routine during
 560  * transaction commit.  If the log is in error state, a non-zero
 561  * return code is handed back and the caller is responsible for
 562  * executing the callback at an appropriate time.
 563  */
 564 int
 565 xfs_log_notify(
 566         struct xfs_mount        *mp,
 567         struct xlog_in_core     *iclog,
 568         xfs_log_callback_t      *cb)
 569 {
 570         int     abortflg;
 571
 572         spin_lock(&iclog->ic_callback_lock);
 573         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 574         if (!abortflg) {
 575                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 576                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 577                 cb->cb_next = NULL;
 578                 *(iclog->ic_callback_tail) = cb;
 579                 iclog->ic_callback_tail = &(cb->cb_next);
 580         }
 581         spin_unlock(&iclog->ic_callback_lock);
 582         return abortflg;
 583 }
 584
 585 int
 586 xfs_log_release_iclog(
 587         struct xfs_mount        *mp,
 588         struct xlog_in_core     *iclog)
 589 {
 590         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 591                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 592                 return -EIO;
 593         }
 594
 595         return 0;
 596 }
 597
 598 /*
 599  * Mount a log filesystem
 600  *
 601  * mp           - ubiquitous xfs mount point structure
 602  * log_target   - buftarg of on-disk log device
 603  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 604  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 605  *
 606  * Return error or zero.
 607  */
 608 int
 609 xfs_log_mount(
 610         xfs_mount_t     *mp,
 611         xfs_buftarg_t   *log_target,
 612         xfs_daddr_t     blk_offset,
 613         int             num_bblks)
 614 {
 615         int             error = 0;
 616         int             min_logfsbs;
 617
 618         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 619                 xfs_notice(mp, "Mounting V%d Filesystem",
 620                            XFS_SB_VERSION_NUM(&mp->m_sb));
 621         } else {
 622                 xfs_notice(mp,
 623 "Mounting V%d filesystem in no-recovery mode. Filesystem will be inconsistent.",
 624                            XFS_SB_VERSION_NUM(&mp->m_sb));
 625                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 626         }
 627
 628         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 629         if (IS_ERR(mp->m_log)) {
 630                 error = PTR_ERR(mp->m_log);
 631                 goto out;
 632         }
 633
 634         /*
 635          * Validate the given log space and drop a critical message via syslog
 636          * if the log size is too small that would lead to some unexpected
 637          * situations in transaction log space reservation stage.
 638          *
 639          * Note: we can't just reject the mount if the validation fails.  This
 640          * would mean that people would have to downgrade their kernel just to
 641          * remedy the situation as there is no way to grow the log (short of
 642          * black magic surgery with xfs_db).
 643          *
 644          * We can, however, reject mounts for CRC format filesystems, as the
 645          * mkfs binary being used to make the filesystem should never create a
 646          * filesystem with a log that is too small.
 647          */
 648         min_logfsbs = xfs_log_calc_minimum_size(mp);
 649
 650         if (mp->m_sb.sb_logblocks < min_logfsbs) {
 651                 xfs_warn(mp,
 652                 "Log size %d blocks too small, minimum size is %d blocks",
 653                          mp->m_sb.sb_logblocks, min_logfsbs);
 654                 error = -EINVAL;
 655         } else if (mp->m_sb.sb_logblocks > XFS_MAX_LOG_BLOCKS) {
 656                 xfs_warn(mp,
 657                 "Log size %d blocks too large, maximum size is %lld blocks",
 658                          mp->m_sb.sb_logblocks, XFS_MAX_LOG_BLOCKS);
 659                 error = -EINVAL;
 660         } else if (XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks) > XFS_MAX_LOG_BYTES) {
 661                 xfs_warn(mp,
 662                 "log size %lld bytes too large, maximum size is %lld bytes",
 663                          XFS_FSB_TO_B(mp, mp->m_sb.sb_logblocks),
 664                          XFS_MAX_LOG_BYTES);
 665                 error = -EINVAL;
 666         }
 667         if (error) {
 668                 if (xfs_sb_version_hascrc(&mp->m_sb)) {
 669                         xfs_crit(mp, "AAIEEE! Log failed size checks. Abort!");
 670                         ASSERT(0);
 671                         goto out_free_log;
 672                 }
 673                 xfs_crit(mp,
 674 "Log size out of supported range. Continuing onwards, but if log hangs are\n"
 675 "experienced then please report this message in the bug report.");
 676         }
 677
 678         /*
 679          * Initialize the AIL now we have a log.
 680          */
 681         error = xfs_trans_ail_init(mp);
 682         if (error) {
 683                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 684                 goto out_free_log;
 685         }
 686         mp->m_log->l_ailp = mp->m_ail;
 687
 688         /*
 689          * skip log recovery on a norecovery mount.  pretend it all
 690          * just worked.
 691          */
 692         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 693                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 694
 695                 if (readonly)
 696                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 697
 698                 error = xlog_recover(mp->m_log);
 699
 700                 if (readonly)
 701                         mp->m_flags |= XFS_MOUNT_RDONLY;
 702                 if (error) {
 703                         xfs_warn(mp, "log mount/recovery failed: error %d",
 704                                 error);
 705                         goto out_destroy_ail;
 706                 }
 707         }
 708
 709         error = xfs_sysfs_init(&mp->m_log->l_kobj, &xfs_log_ktype, &mp->m_kobj,
 710                                "log");
 711         if (error)
 712                 goto out_destroy_ail;
 713
 714         /* Normal transactions can now occur */
 715         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 716
 717         /*
 718          * Now the log has been fully initialised and we know were our
 719          * space grant counters are, we can initialise the permanent ticket
 720          * needed for delayed logging to work.
 721          */
 722         xlog_cil_init_post_recovery(mp->m_log);
 723
 724         return 0;
 725
 726 out_destroy_ail:
 727         xfs_trans_ail_destroy(mp);
 728 out_free_log:
 729         xlog_dealloc_log(mp->m_log);
 730 out:
 731         return error;
 732 }
 733
 734 /*
 735  * Finish the recovery of the file system.  This is separate from the
 736  * xfs_log_mount() call, because it depends on the code in xfs_mountfs() to read
 737  * in the root and real-time bitmap inodes between calling xfs_log_mount() and
 738  * here.
 739  *
 740  * If we finish recovery successfully, start the background log work. If we are
 741  * not doing recovery, then we have a RO filesystem and we don't need to start
 742  * it.
 743  */
 744 int
 745 xfs_log_mount_finish(xfs_mount_t *mp)
 746 {
 747         int     error = 0;
 748
 749         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 750                 error = xlog_recover_finish(mp->m_log);
 751                 if (!error)
 752                         xfs_log_work_queue(mp);
 753         } else {
 754                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 755         }
 756
 757
 758         return error;
 759 }
 760
 761 /*
 762  * Final log writes as part of unmount.
 763  *
 764  * Mark the filesystem clean as unmount happens.  Note that during relocation
 765  * this routine needs to be executed as part of source-bag while the
 766  * deallocation must not be done until source-end.
 767  */
 768
 769 /*
 770  * Unmount record used to have a string "Unmount filesystem--" in the
 771  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 772  * We just write the magic number now since that particular field isn't
 773  * currently architecture converted and "Unmount" is a bit foo.
 774  * As far as I know, there weren't any dependencies on the old behaviour.
 775  */
 776
 777 int
 778 xfs_log_unmount_write(xfs_mount_t *mp)
 779 {
 780         struct xlog      *log = mp->m_log;
 781         xlog_in_core_t   *iclog;
 782 #ifdef DEBUG
 783         xlog_in_core_t   *first_iclog;
 784 #endif
 785         xlog_ticket_t   *tic = NULL;
 786         xfs_lsn_t        lsn;
 787         int              error;
 788
 789         /*
 790          * Don't write out unmount record on read-only mounts.
 791          * Or, if we are doing a forced umount (typically because of IO errors).
 792          */
 793         if (mp->m_flags & XFS_MOUNT_RDONLY)
 794                 return 0;
 795
 796         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
 797         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 798
 799 #ifdef DEBUG
 800         first_iclog = iclog = log->l_iclog;
 801         do {
 802                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 803                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 804                         ASSERT(iclog->ic_offset == 0);
 805                 }
 806                 iclog = iclog->ic_next;
 807         } while (iclog != first_iclog);
 808 #endif
 809         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 810                 error = xfs_log_reserve(mp, 600, 1, &tic,
 811                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
 812                 if (!error) {
 813                         /* the data section must be 32 bit size aligned */
 814                         struct {
 815                             __uint16_t magic;
 816                             __uint16_t pad1;
 817                             __uint32_t pad2; /* may as well make it 64 bits */
 818                         } magic = {
 819                                 .magic = XLOG_UNMOUNT_TYPE,
 820                         };
 821                         struct xfs_log_iovec reg = {
 822                                 .i_addr = &magic,
 823                                 .i_len = sizeof(magic),
 824                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 825                         };
 826                         struct xfs_log_vec vec = {
 827                                 .lv_niovecs = 1,
 828                                 .lv_iovecp = &reg,
 829                         };
 830
 831                         /* remove inited flag, and account for space used */
 832                         tic->t_flags = 0;
 833                         tic->t_curr_res -= sizeof(magic);
 834                         error = xlog_write(log, &vec, tic, &lsn,
 835                                            NULL, XLOG_UNMOUNT_TRANS);
 836                         /*
 837                          * At this point, we're umounting anyway,
 838                          * so there's no point in transitioning log state
 839                          * to IOERROR. Just continue...
 840                          */
 841                 }
 842
 843                 if (error)
 844                         xfs_alert(mp, "%s: unmount record failed", __func__);
 845
 846
 847                 spin_lock(&log->l_icloglock);
 848                 iclog = log->l_iclog;
 849                 atomic_inc(&iclog->ic_refcnt);
 850                 xlog_state_want_sync(log, iclog);
 851                 spin_unlock(&log->l_icloglock);
 852                 error = xlog_state_release_iclog(log, iclog);
 853
 854                 spin_lock(&log->l_icloglock);
 855                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 856                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 857                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 858                                 xlog_wait(&iclog->ic_force_wait,
 859                                                         &log->l_icloglock);
 860                         } else {
 861                                 spin_unlock(&log->l_icloglock);
 862                         }
 863                 } else {
 864                         spin_unlock(&log->l_icloglock);
 865                 }
 866                 if (tic) {
 867                         trace_xfs_log_umount_write(log, tic);
 868                         xlog_ungrant_log_space(log, tic);
 869                         xfs_log_ticket_put(tic);
 870                 }
 871         } else {
 872                 /*
 873                  * We're already in forced_shutdown mode, couldn't
 874                  * even attempt to write out the unmount transaction.
 875                  *
 876                  * Go through the motions of sync'ing and releasing
 877                  * the iclog, even though no I/O will actually happen,
 878                  * we need to wait for other log I/Os that may already
 879                  * be in progress.  Do this as a separate section of
 880                  * code so we'll know if we ever get stuck here that
 881                  * we're in this odd situation of trying to unmount
 882                  * a file system that went into forced_shutdown as
 883                  * the result of an unmount..
 884                  */
 885                 spin_lock(&log->l_icloglock);
 886                 iclog = log->l_iclog;
 887                 atomic_inc(&iclog->ic_refcnt);
 888
 889                 xlog_state_want_sync(log, iclog);
 890                 spin_unlock(&log->l_icloglock);
 891                 error =  xlog_state_release_iclog(log, iclog);
 892
 893                 spin_lock(&log->l_icloglock);
 894
 895                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 896                         || iclog->ic_state == XLOG_STATE_DIRTY
 897                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 898
 899                                 xlog_wait(&iclog->ic_force_wait,
 900                                                         &log->l_icloglock);
 901                 } else {
 902                         spin_unlock(&log->l_icloglock);
 903                 }
 904         }
 905
 906         return error;
 907 }       /* xfs_log_unmount_write */
 908
 909 /*
 910  * Empty the log for unmount/freeze.
 911  *
 912  * To do this, we first need to shut down the background log work so it is not
 913  * trying to cover the log as we clean up. We then need to unpin all objects in
 914  * the log so we can then flush them out. Once they have completed their IO and
 915  * run the callbacks removing themselves from the AIL, we can write the unmount
 916  * record.
 917  */
 918 void
 919 xfs_log_quiesce(
 920         struct xfs_mount        *mp)
 921 {
 922         cancel_delayed_work_sync(&mp->m_log->l_work);
 923         xfs_log_force(mp, XFS_LOG_SYNC);
 924
 925         /*
 926          * The superblock buffer is uncached and while xfs_ail_push_all_sync()
 927          * will push it, xfs_wait_buftarg() will not wait for it. Further,
 928          * xfs_buf_iowait() cannot be used because it was pushed with the
 929          * XBF_ASYNC flag set, so we need to use a lock/unlock pair to wait for
 930          * the IO to complete.
 931          */
 932         xfs_ail_push_all_sync(mp->m_ail);
 933         xfs_wait_buftarg(mp->m_ddev_targp);
 934         xfs_buf_lock(mp->m_sb_bp);
 935         xfs_buf_unlock(mp->m_sb_bp);
 936
 937         xfs_log_unmount_write(mp);
 938 }
 939
 940 /*
 941  * Shut down and release the AIL and Log.
 942  *
 943  * During unmount, we need to ensure we flush all the dirty metadata objects
 944  * from the AIL so that the log is empty before we write the unmount record to
 945  * the log. Once this is done, we can tear down the AIL and the log.
 946  */
 947 void
 948 xfs_log_unmount(
 949         struct xfs_mount        *mp)
 950 {
 951         xfs_log_quiesce(mp);
 952
 953         xfs_trans_ail_destroy(mp);
 954
 955         xfs_sysfs_del(&mp->m_log->l_kobj);
 956
 957         xlog_dealloc_log(mp->m_log);
 958 }
 959
 960 void
 961 xfs_log_item_init(
 962         struct xfs_mount        *mp,
 963         struct xfs_log_item     *item,
 964         int                     type,
 965         const struct xfs_item_ops *ops)
 966 {
 967         item->li_mountp = mp;
 968         item->li_ailp = mp->m_ail;
 969         item->li_type = type;
 970         item->li_ops = ops;
 971         item->li_lv = NULL;
 972
 973         INIT_LIST_HEAD(&item->li_ail);
 974         INIT_LIST_HEAD(&item->li_cil);
 975 }
 976
 977 /*
 978  * Wake up processes waiting for log space after we have moved the log tail.
 979  */
 980 void
 981 xfs_log_space_wake(
 982         struct xfs_mount        *mp)
 983 {
 984         struct xlog             *log = mp->m_log;
 985         int                     free_bytes;
 986
 987         if (XLOG_FORCED_SHUTDOWN(log))
 988                 return;
 989
 990         if (!list_empty_careful(&log->l_write_head.waiters)) {
 991                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 992
 993                 spin_lock(&log->l_write_head.lock);
 994                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
 995                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
 996                 spin_unlock(&log->l_write_head.lock);
 997         }
 998
 999         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
1000                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
1001
1002                 spin_lock(&log->l_reserve_head.lock);
1003                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1004                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
1005                 spin_unlock(&log->l_reserve_head.lock);
1006         }
1007 }
1008
1009 /*
1010  * Determine if we have a transaction that has gone to disk that needs to be
1011  * covered. To begin the transition to the idle state firstly the log needs to
1012  * be idle. That means the CIL, the AIL and the iclogs needs to be empty before
1013  * we start attempting to cover the log.
1014  *
1015  * Only if we are then in a state where covering is needed, the caller is
1016  * informed that dummy transactions are required to move the log into the idle
1017  * state.
1018  *
1019  * If there are any items in the AIl or CIL, then we do not want to attempt to
1020  * cover the log as we may be in a situation where there isn't log space
1021  * available to run a dummy transaction and this can lead to deadlocks when the
1022  * tail of the log is pinned by an item that is modified in the CIL.  Hence
1023  * there's no point in running a dummy transaction at this point because we
1024  * can't start trying to idle the log until both the CIL and AIL are empty.
1025  */
1026 int
1027 xfs_log_need_covered(xfs_mount_t *mp)
1028 {
1029         struct xlog     *log = mp->m_log;
1030         int             needed = 0;
1031
1032         if (!xfs_fs_writable(mp, SB_FREEZE_WRITE))
1033                 return 0;
1034
1035         if (!xlog_cil_empty(log))
1036                 return 0;
1037
1038         spin_lock(&log->l_icloglock);
1039         switch (log->l_covered_state) {
1040         case XLOG_STATE_COVER_DONE:
1041         case XLOG_STATE_COVER_DONE2:
1042         case XLOG_STATE_COVER_IDLE:
1043                 break;
1044         case XLOG_STATE_COVER_NEED:
1045         case XLOG_STATE_COVER_NEED2:
1046                 if (xfs_ail_min_lsn(log->l_ailp))
1047                         break;
1048                 if (!xlog_iclogs_empty(log))
1049                         break;
1050
1051                 needed = 1;
1052                 if (log->l_covered_state == XLOG_STATE_COVER_NEED)
1053                         log->l_covered_state = XLOG_STATE_COVER_DONE;
1054                 else
1055                         log->l_covered_state = XLOG_STATE_COVER_DONE2;
1056                 break;
1057         default:
1058                 needed = 1;
1059                 break;
1060         }
1061         spin_unlock(&log->l_icloglock);
1062         return needed;
1063 }
1064
1065 /*
1066  * We may be holding the log iclog lock upon entering this routine.
1067  */
1068 xfs_lsn_t
1069 xlog_assign_tail_lsn_locked(
1070         struct xfs_mount        *mp)
1071 {
1072         struct xlog             *log = mp->m_log;
1073         struct xfs_log_item     *lip;
1074         xfs_lsn_t               tail_lsn;
1075
1076         assert_spin_locked(&mp->m_ail->xa_lock);
1077
1078         /*
1079          * To make sure we always have a valid LSN for the log tail we keep
1080          * track of the last LSN which was committed in log->l_last_sync_lsn,
1081          * and use that when the AIL was empty.
1082          */
1083         lip = xfs_ail_min(mp->m_ail);
1084         if (lip)
1085                 tail_lsn = lip->li_lsn;
1086         else
1087                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
1088         trace_xfs_log_assign_tail_lsn(log, tail_lsn);
1089         atomic64_set(&log->l_tail_lsn, tail_lsn);
1090         return tail_lsn;
1091 }
1092
1093 xfs_lsn_t
1094 xlog_assign_tail_lsn(
1095         struct xfs_mount        *mp)
1096 {
1097         xfs_lsn_t               tail_lsn;
1098
1099         spin_lock(&mp->m_ail->xa_lock);
1100         tail_lsn = xlog_assign_tail_lsn_locked(mp);
1101         spin_unlock(&mp->m_ail->xa_lock);
1102
1103         return tail_lsn;
1104 }
1105
1106 /*
1107  * Return the space in the log between the tail and the head.  The head
1108  * is passed in the cycle/bytes formal parms.  In the special case where
1109  * the reserve head has wrapped passed the tail, this calculation is no
1110  * longer valid.  In this case, just return 0 which means there is no space
1111  * in the log.  This works for all places where this function is called
1112  * with the reserve head.  Of course, if the write head were to ever
1113  * wrap the tail, we should blow up.  Rather than catch this case here,
1114  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1115  *
1116  * This code also handles the case where the reservation head is behind
1117  * the tail.  The details of this case are described below, but the end
1118  * result is that we return the size of the log as the amount of space left.
1119  */
1120 STATIC int
1121 xlog_space_left(
1122         struct xlog     *log,
1123         atomic64_t      *head)
1124 {
1125         int             free_bytes;
1126         int             tail_bytes;
1127         int             tail_cycle;
1128         int             head_cycle;
1129         int             head_bytes;
1130
1131         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1132         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1133         tail_bytes = BBTOB(tail_bytes);
1134         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1135                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1136         else if (tail_cycle + 1 < head_cycle)
1137                 return 0;
1138         else if (tail_cycle < head_cycle) {
1139                 ASSERT(tail_cycle == (head_cycle - 1));
1140                 free_bytes = tail_bytes - head_bytes;
1141         } else {
1142                 /*
1143                  * The reservation head is behind the tail.
1144                  * In this case we just want to return the size of the
1145                  * log as the amount of space left.
1146                  */
1147                 xfs_alert(log->l_mp,
1148                         "xlog_space_left: head behind tail\n"
1149                         "  tail_cycle = %d, tail_bytes = %d\n"
1150                         "  GH   cycle = %d, GH   bytes = %d",
1151                         tail_cycle, tail_bytes, head_cycle, head_bytes);
1152                 ASSERT(0);
1153                 free_bytes = log->l_logsize;
1154         }
1155         return free_bytes;
1156 }
1157
1158
1159 /*
1160  * Log function which is called when an io completes.
1161  *
1162  * The log manager needs its own routine, in order to control what
1163  * happens with the buffer after the write completes.
1164  */
1165 void
1166 xlog_iodone(xfs_buf_t *bp)
1167 {
1168         struct xlog_in_core     *iclog = bp->b_fspriv;
1169         struct xlog             *l = iclog->ic_log;
1170         int                     aborted = 0;
1171
1172         /*
1173          * Race to shutdown the filesystem if we see an error.
1174          */
1175         if (XFS_TEST_ERROR(bp->b_error, l->l_mp,
1176                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1177                 xfs_buf_ioerror_alert(bp, __func__);
1178                 xfs_buf_stale(bp);
1179                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1180                 /*
1181                  * This flag will be propagated to the trans-committed
1182                  * callback routines to let them know that the log-commit
1183                  * didn't succeed.
1184                  */
1185                 aborted = XFS_LI_ABORTED;
1186         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1187                 aborted = XFS_LI_ABORTED;
1188         }
1189
1190         /* log I/O is always issued ASYNC */
1191         ASSERT(XFS_BUF_ISASYNC(bp));
1192         xlog_state_done_syncing(iclog, aborted);
1193
1194         /*
1195          * drop the buffer lock now that we are done. Nothing references
1196          * the buffer after this, so an unmount waiting on this lock can now
1197          * tear it down safely. As such, it is unsafe to reference the buffer
1198          * (bp) after the unlock as we could race with it being freed.
1199          */
1200         xfs_buf_unlock(bp);
1201 }
1202
1203 /*
1204  * Return size of each in-core log record buffer.
1205  *
1206  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1207  *
1208  * If the filesystem blocksize is too large, we may need to choose a
1209  * larger size since the directory code currently logs entire blocks.
1210  */
1211
1212 STATIC void
1213 xlog_get_iclog_buffer_size(
1214         struct xfs_mount        *mp,
1215         struct xlog             *log)
1216 {
1217         int size;
1218         int xhdrs;
1219
1220         if (mp->m_logbufs <= 0)
1221                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1222         else
1223                 log->l_iclog_bufs = mp->m_logbufs;
1224
1225         /*
1226          * Buffer size passed in from mount system call.
1227          */
1228         if (mp->m_logbsize > 0) {
1229                 size = log->l_iclog_size = mp->m_logbsize;
1230                 log->l_iclog_size_log = 0;
1231                 while (size != 1) {
1232                         log->l_iclog_size_log++;
1233                         size >>= 1;
1234                 }
1235
1236                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1237                         /* # headers = size / 32k
1238                          * one header holds cycles from 32k of data
1239                          */
1240
1241                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1242                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1243                                 xhdrs++;
1244                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1245                         log->l_iclog_heads = xhdrs;
1246                 } else {
1247                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1248                         log->l_iclog_hsize = BBSIZE;
1249                         log->l_iclog_heads = 1;
1250                 }
1251                 goto done;
1252         }
1253
1254         /* All machines use 32kB buffers by default. */
1255         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1256         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1257
1258         /* the default log size is 16k or 32k which is one header sector */
1259         log->l_iclog_hsize = BBSIZE;
1260         log->l_iclog_heads = 1;
1261
1262 done:
1263         /* are we being asked to make the sizes selected above visible? */
1264         if (mp->m_logbufs == 0)
1265                 mp->m_logbufs = log->l_iclog_bufs;
1266         if (mp->m_logbsize == 0)
1267                 mp->m_logbsize = log->l_iclog_size;
1268 }       /* xlog_get_iclog_buffer_size */
1269
1270
1271 void
1272 xfs_log_work_queue(
1273         struct xfs_mount        *mp)
1274 {
1275         queue_delayed_work(mp->m_log_workqueue, &mp->m_log->l_work,
1276                                 msecs_to_jiffies(xfs_syncd_centisecs * 10));
1277 }
1278
1279 /*
1280  * Every sync period we need to unpin all items in the AIL and push them to
1281  * disk. If there is nothing dirty, then we might need to cover the log to
1282  * indicate that the filesystem is idle.
1283  */
1284 void
1285 xfs_log_worker(
1286         struct work_struct      *work)
1287 {
1288         struct xlog             *log = container_of(to_delayed_work(work),
1289                                                 struct xlog, l_work);
1290         struct xfs_mount        *mp = log->l_mp;
1291
1292         /* dgc: errors ignored - not fatal and nowhere to report them */
1293         if (xfs_log_need_covered(mp))
1294                 xfs_fs_log_dummy(mp);
1295         else
1296                 xfs_log_force(mp, 0);
1297
1298         /* start pushing all the metadata that is currently dirty */
1299         xfs_ail_push_all(mp->m_ail);
1300
1301         /* queue us up again */
1302         xfs_log_work_queue(mp);
1303 }
1304
1305 /*
1306  * This routine initializes some of the log structure for a given mount point.
1307  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1308  * some other stuff may be filled in too.
1309  */
1310 STATIC struct xlog *
1311 xlog_alloc_log(
1312         struct xfs_mount        *mp,
1313         struct xfs_buftarg      *log_target,
1314         xfs_daddr_t             blk_offset,
1315         int                     num_bblks)
1316 {
1317         struct xlog             *log;
1318         xlog_rec_header_t       *head;
1319         xlog_in_core_t          **iclogp;
1320         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1321         xfs_buf_t               *bp;
1322         int                     i;
1323         int                     error = -ENOMEM;
1324         uint                    log2_size = 0;
1325
1326         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1327         if (!log) {
1328                 xfs_warn(mp, "Log allocation failed: No memory!");
1329                 goto out;
1330         }
1331
1332         log->l_mp          = mp;
1333         log->l_targ        = log_target;
1334         log->l_logsize     = BBTOB(num_bblks);
1335         log->l_logBBstart  = blk_offset;
1336         log->l_logBBsize   = num_bblks;
1337         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1338         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1339         INIT_DELAYED_WORK(&log->l_work, xfs_log_worker);
1340
1341         log->l_prev_block  = -1;
1342         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1343         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1344         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1345         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1346
1347         xlog_grant_head_init(&log->l_reserve_head);
1348         xlog_grant_head_init(&log->l_write_head);
1349
1350         error = -EFSCORRUPTED;
1351         if (xfs_sb_version_hassector(&mp->m_sb)) {
1352                 log2_size = mp->m_sb.sb_logsectlog;
1353                 if (log2_size < BBSHIFT) {
1354                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1355                                 log2_size, BBSHIFT);
1356                         goto out_free_log;
1357                 }
1358
1359                 log2_size -= BBSHIFT;
1360                 if (log2_size > mp->m_sectbb_log) {
1361                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1362                                 log2_size, mp->m_sectbb_log);
1363                         goto out_free_log;
1364                 }
1365
1366                 /* for larger sector sizes, must have v2 or external log */
1367                 if (log2_size && log->l_logBBstart > 0 &&
1368                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1369                         xfs_warn(mp,
1370                 "log sector size (0x%x) invalid for configuration.",
1371                                 log2_size);
1372                         goto out_free_log;
1373                 }
1374         }
1375         log->l_sectBBsize = 1 << log2_size;
1376
1377         xlog_get_iclog_buffer_size(mp, log);
1378
1379         /*
1380          * Use a NULL block for the extra log buffer used during splits so that
1381          * it will trigger errors if we ever try to do IO on it without first
1382          * having set it up properly.
1383          */
1384         error = -ENOMEM;
1385         bp = xfs_buf_alloc(mp->m_logdev_targp, XFS_BUF_DADDR_NULL,
1386                            BTOBB(log->l_iclog_size), 0);
1387         if (!bp)
1388                 goto out_free_log;
1389
1390         /*
1391          * The iclogbuf buffer locks are held over IO but we are not going to do
1392          * IO yet.  Hence unlock the buffer so that the log IO path can grab it
1393          * when appropriately.
1394          */
1395         ASSERT(xfs_buf_islocked(bp));
1396         xfs_buf_unlock(bp);
1397
1398         bp->b_iodone = xlog_iodone;
1399         log->l_xbuf = bp;
1400
1401         spin_lock_init(&log->l_icloglock);
1402         init_waitqueue_head(&log->l_flush_wait);
1403
1404         iclogp = &log->l_iclog;
1405         /*
1406          * The amount of memory to allocate for the iclog structure is
1407          * rather funky due to the way the structure is defined.  It is
1408          * done this way so that we can use different sizes for machines
1409          * with different amounts of memory.  See the definition of
1410          * xlog_in_core_t in xfs_log_priv.h for details.
1411          */
1412         ASSERT(log->l_iclog_size >= 4096);
1413         for (i=0; i < log->l_iclog_bufs; i++) {
1414                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1415                 if (!*iclogp)
1416                         goto out_free_iclog;
1417
1418                 iclog = *iclogp;
1419                 iclog->ic_prev = prev_iclog;
1420                 prev_iclog = iclog;
1421
1422                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1423                                                 BTOBB(log->l_iclog_size), 0);
1424                 if (!bp)
1425                         goto out_free_iclog;
1426
1427                 ASSERT(xfs_buf_islocked(bp));
1428                 xfs_buf_unlock(bp);
1429
1430                 bp->b_iodone = xlog_iodone;
1431                 iclog->ic_bp = bp;
1432                 iclog->ic_data = bp->b_addr;
1433 #ifdef DEBUG
1434                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1435 #endif
1436                 head = &iclog->ic_header;
1437                 memset(head, 0, sizeof(xlog_rec_header_t));
1438                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1439                 head->h_version = cpu_to_be32(
1440                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1441                 head->h_size = cpu_to_be32(log->l_iclog_size);
1442                 /* new fields */
1443                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1444                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1445
1446                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1447                 iclog->ic_state = XLOG_STATE_ACTIVE;
1448                 iclog->ic_log = log;
1449                 atomic_set(&iclog->ic_refcnt, 0);
1450                 spin_lock_init(&iclog->ic_callback_lock);
1451                 iclog->ic_callback_tail = &(iclog->ic_callback);
1452                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1453
1454                 init_waitqueue_head(&iclog->ic_force_wait);
1455                 init_waitqueue_head(&iclog->ic_write_wait);
1456
1457                 iclogp = &iclog->ic_next;
1458         }
1459         *iclogp = log->l_iclog;                 /* complete ring */
1460         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1461
1462         error = xlog_cil_init(log);
1463         if (error)
1464                 goto out_free_iclog;
1465         return log;
1466
1467 out_free_iclog:
1468         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1469                 prev_iclog = iclog->ic_next;
1470                 if (iclog->ic_bp)
1471                         xfs_buf_free(iclog->ic_bp);
1472                 kmem_free(iclog);
1473         }
1474         spinlock_destroy(&log->l_icloglock);
1475         xfs_buf_free(log->l_xbuf);
1476 out_free_log:
1477         kmem_free(log);
1478 out:
1479         return ERR_PTR(error);
1480 }       /* xlog_alloc_log */
1481
1482
1483 /*
1484  * Write out the commit record of a transaction associated with the given
1485  * ticket.  Return the lsn of the commit record.
1486  */
1487 STATIC int
1488 xlog_commit_record(
1489         struct xlog             *log,
1490         struct xlog_ticket      *ticket,
1491         struct xlog_in_core     **iclog,
1492         xfs_lsn_t               *commitlsnp)
1493 {
1494         struct xfs_mount *mp = log->l_mp;
1495         int     error;
1496         struct xfs_log_iovec reg = {
1497                 .i_addr = NULL,
1498                 .i_len = 0,
1499                 .i_type = XLOG_REG_TYPE_COMMIT,
1500         };
1501         struct xfs_log_vec vec = {
1502                 .lv_niovecs = 1,
1503                 .lv_iovecp = &reg,
1504         };
1505
1506         ASSERT_ALWAYS(iclog);
1507         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1508                                         XLOG_COMMIT_TRANS);
1509         if (error)
1510                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1511         return error;
1512 }
1513
1514 /*
1515  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1516  * log space.  This code pushes on the lsn which would supposedly free up
1517  * the 25% which we want to leave free.  We may need to adopt a policy which
1518  * pushes on an lsn which is further along in the log once we reach the high
1519  * water mark.  In this manner, we would be creating a low water mark.
1520  */
1521 STATIC void
1522 xlog_grant_push_ail(
1523         struct xlog     *log,
1524         int             need_bytes)
1525 {
1526         xfs_lsn_t       threshold_lsn = 0;
1527         xfs_lsn_t       last_sync_lsn;
1528         int             free_blocks;
1529         int             free_bytes;
1530         int             threshold_block;
1531         int             threshold_cycle;
1532         int             free_threshold;
1533
1534         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1535
1536         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1537         free_blocks = BTOBBT(free_bytes);
1538
1539         /*
1540          * Set the threshold for the minimum number of free blocks in the
1541          * log to the maximum of what the caller needs, one quarter of the
1542          * log, and 256 blocks.
1543          */
1544         free_threshold = BTOBB(need_bytes);
1545         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1546         free_threshold = MAX(free_threshold, 256);
1547         if (free_blocks >= free_threshold)
1548                 return;
1549
1550         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1551                                                 &threshold_block);
1552         threshold_block += free_threshold;
1553         if (threshold_block >= log->l_logBBsize) {
1554                 threshold_block -= log->l_logBBsize;
1555                 threshold_cycle += 1;
1556         }
1557         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1558                                         threshold_block);
1559         /*
1560          * Don't pass in an lsn greater than the lsn of the last
1561          * log record known to be on disk. Use a snapshot of the last sync lsn
1562          * so that it doesn't change between the compare and the set.
1563          */
1564         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1565         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1566                 threshold_lsn = last_sync_lsn;
1567
1568         /*
1569          * Get the transaction layer to kick the dirty buffers out to
1570          * disk asynchronously. No point in trying to do this if
1571          * the filesystem is shutting down.
1572          */
1573         if (!XLOG_FORCED_SHUTDOWN(log))
1574                 xfs_ail_push(log->l_ailp, threshold_lsn);
1575 }
1576
1577 /*
1578  * Stamp cycle number in every block
1579  */
1580 STATIC void
1581 xlog_pack_data(
1582         struct xlog             *log,
1583         struct xlog_in_core     *iclog,
1584         int                     roundoff)
1585 {
1586         int                     i, j, k;
1587         int                     size = iclog->ic_offset + roundoff;
1588         __be32                  cycle_lsn;
1589         xfs_caddr_t             dp;
1590
1591         cycle_lsn = CYCLE_LSN_DISK(iclog->ic_header.h_lsn);
1592
1593         dp = iclog->ic_datap;
1594         for (i = 0; i < BTOBB(size); i++) {
1595                 if (i >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE))
1596                         break;
1597                 iclog->ic_header.h_cycle_data[i] = *(__be32 *)dp;
1598                 *(__be32 *)dp = cycle_lsn;
1599                 dp += BBSIZE;
1600         }
1601
1602         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1603                 xlog_in_core_2_t *xhdr = iclog->ic_data;
1604
1605                 for ( ; i < BTOBB(size); i++) {
1606                         j = i / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1607                         k = i % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
1608                         xhdr[j].hic_xheader.xh_cycle_data[k] = *(__be32 *)dp;
1609                         *(__be32 *)dp = cycle_lsn;
1610                         dp += BBSIZE;
1611                 }
1612
1613                 for (i = 1; i < log->l_iclog_heads; i++)
1614                         xhdr[i].hic_xheader.xh_cycle = cycle_lsn;
1615         }
1616 }
1617
1618 /*
1619  * Calculate the checksum for a log buffer.
1620  *
1621  * This is a little more complicated than it should be because the various
1622  * headers and the actual data are non-contiguous.
1623  */
1624 __le32
1625 xlog_cksum(
1626         struct xlog             *log,
1627         struct xlog_rec_header  *rhead,
1628         char                    *dp,
1629         int                     size)
1630 {
1631         __uint32_t              crc;
1632
1633         /* first generate the crc for the record header ... */
1634         crc = xfs_start_cksum((char *)rhead,
1635                               sizeof(struct xlog_rec_header),
1636                               offsetof(struct xlog_rec_header, h_crc));
1637
1638         /* ... then for additional cycle data for v2 logs ... */
1639         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb)) {
1640                 union xlog_in_core2 *xhdr = (union xlog_in_core2 *)rhead;
1641                 int             i;
1642
1643                 for (i = 1; i < log->l_iclog_heads; i++) {
1644                         crc = crc32c(crc, &xhdr[i].hic_xheader,
1645                                      sizeof(struct xlog_rec_ext_header));
1646                 }
1647         }
1648
1649         /* ... and finally for the payload */
1650         crc = crc32c(crc, dp, size);
1651
1652         return xfs_end_cksum(crc);
1653 }
1654
1655 /*
1656  * The bdstrat callback function for log bufs. This gives us a central
1657  * place to trap bufs in case we get hit by a log I/O error and need to
1658  * shutdown. Actually, in practice, even when we didn't get a log error,
1659  * we transition the iclogs to IOERROR state *after* flushing all existing
1660  * iclogs to disk. This is because we don't want anymore new transactions to be
1661  * started or completed afterwards.
1662  *
1663  * We lock the iclogbufs here so that we can serialise against IO completion
1664  * during unmount. We might be processing a shutdown triggered during unmount,
1665  * and that can occur asynchronously to the unmount thread, and hence we need to
1666  * ensure that completes before tearing down the iclogbufs. Hence we need to
1667  * hold the buffer lock across the log IO to acheive that.
1668  */
1669 STATIC int
1670 xlog_bdstrat(
1671         struct xfs_buf          *bp)
1672 {
1673         struct xlog_in_core     *iclog = bp->b_fspriv;
1674
1675         xfs_buf_lock(bp);
1676         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1677                 xfs_buf_ioerror(bp, -EIO);
1678                 xfs_buf_stale(bp);
1679                 xfs_buf_ioend(bp);
1680                 /*
1681                  * It would seem logical to return EIO here, but we rely on
1682                  * the log state machine to propagate I/O errors instead of
1683                  * doing it here. Similarly, IO completion will unlock the
1684                  * buffer, so we don't do it here.
1685                  */
1686                 return 0;
1687         }
1688
1689         xfs_buf_submit(bp);
1690         return 0;
1691 }
1692
1693 /*
1694  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1695  * fashion.  Previously, we should have moved the current iclog
1696  * ptr in the log to point to the next available iclog.  This allows further
1697  * write to continue while this code syncs out an iclog ready to go.
1698  * Before an in-core log can be written out, the data section must be scanned
1699  * to save away the 1st word of each BBSIZE block into the header.  We replace
1700  * it with the current cycle count.  Each BBSIZE block is tagged with the
1701  * cycle count because there in an implicit assumption that drives will
1702  * guarantee that entire 512 byte blocks get written at once.  In other words,
1703  * we can't have part of a 512 byte block written and part not written.  By
1704  * tagging each block, we will know which blocks are valid when recovering
1705  * after an unclean shutdown.
1706  *
1707  * This routine is single threaded on the iclog.  No other thread can be in
1708  * this routine with the same iclog.  Changing contents of iclog can there-
1709  * fore be done without grabbing the state machine lock.  Updating the global
1710  * log will require grabbing the lock though.
1711  *
1712  * The entire log manager uses a logical block numbering scheme.  Only
1713  * log_sync (and then only bwrite()) know about the fact that the log may
1714  * not start with block zero on a given device.  The log block start offset
1715  * is added immediately before calling bwrite().
1716  */
1717
1718 STATIC int
1719 xlog_sync(
1720         struct xlog             *log,
1721         struct xlog_in_core     *iclog)
1722 {
1723         xfs_buf_t       *bp;
1724         int             i;
1725         uint            count;          /* byte count of bwrite */
1726         uint            count_init;     /* initial count before roundup */
1727         int             roundoff;       /* roundoff to BB or stripe */
1728         int             split = 0;      /* split write into two regions */
1729         int             error;
1730         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1731         int             size;
1732
1733         XFS_STATS_INC(xs_log_writes);
1734         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1735
1736         /* Add for LR header */
1737         count_init = log->l_iclog_hsize + iclog->ic_offset;
1738
1739         /* Round out the log write size */
1740         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1741                 /* we have a v2 stripe unit to use */
1742                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1743         } else {
1744                 count = BBTOB(BTOBB(count_init));
1745         }
1746         roundoff = count - count_init;
1747         ASSERT(roundoff >= 0);
1748         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1749                 roundoff < log->l_mp->m_sb.sb_logsunit)
1750                 ||
1751                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1752                  roundoff < BBTOB(1)));
1753
1754         /* move grant heads by roundoff in sync */
1755         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1756         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1757
1758         /* put cycle number in every block */
1759         xlog_pack_data(log, iclog, roundoff);
1760
1761         /* real byte length */
1762         size = iclog->ic_offset;
1763         if (v2)
1764                 size += roundoff;
1765         iclog->ic_header.h_len = cpu_to_be32(size);
1766
1767         bp = iclog->ic_bp;
1768         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1769
1770         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1771
1772         /* Do we need to split this write into 2 parts? */
1773         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1774                 char            *dptr;
1775
1776                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1777                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1778                 iclog->ic_bwritecnt = 2;
1779
1780                 /*
1781                  * Bump the cycle numbers at the start of each block in the
1782                  * part of the iclog that ends up in the buffer that gets
1783                  * written to the start of the log.
1784                  *
1785                  * Watch out for the header magic number case, though.
1786                  */
1787                 dptr = (char *)&iclog->ic_header + count;
1788                 for (i = 0; i < split; i += BBSIZE) {
1789                         __uint32_t cycle = be32_to_cpu(*(__be32 *)dptr);
1790                         if (++cycle == XLOG_HEADER_MAGIC_NUM)
1791                                 cycle++;
1792                         *(__be32 *)dptr = cpu_to_be32(cycle);
1793
1794                         dptr += BBSIZE;
1795                 }
1796         } else {
1797                 iclog->ic_bwritecnt = 1;
1798         }
1799
1800         /* calculcate the checksum */
1801         iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
1802                                             iclog->ic_datap, size);
1803
1804         bp->b_io_length = BTOBB(count);
1805         bp->b_fspriv = iclog;
1806         XFS_BUF_ZEROFLAGS(bp);
1807         XFS_BUF_ASYNC(bp);
1808         bp->b_flags |= XBF_SYNCIO;
1809         /* use high priority completion wq */
1810         bp->b_ioend_wq = log->l_mp->m_log_workqueue;
1811
1812         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1813                 bp->b_flags |= XBF_FUA;
1814
1815                 /*
1816                  * Flush the data device before flushing the log to make
1817                  * sure all meta data written back from the AIL actually made
1818                  * it to disk before stamping the new log tail LSN into the
1819                  * log buffer.  For an external log we need to issue the
1820                  * flush explicitly, and unfortunately synchronously here;
1821                  * for an internal log we can simply use the block layer
1822                  * state machine for preflushes.
1823                  */
1824                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1825                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1826                 else
1827                         bp->b_flags |= XBF_FLUSH;
1828         }
1829
1830         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1831         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1832
1833         xlog_verify_iclog(log, iclog, count, true);
1834
1835         /* account for log which doesn't start at block #0 */
1836         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1837         /*
1838          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1839          * is shutting down.
1840          */
1841         XFS_BUF_WRITE(bp);
1842
1843         error = xlog_bdstrat(bp);
1844         if (error) {
1845                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1846                 return error;
1847         }
1848         if (split) {
1849                 bp = iclog->ic_log->l_xbuf;
1850                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1851                 xfs_buf_associate_memory(bp,
1852                                 (char *)&iclog->ic_header + count, split);
1853                 bp->b_fspriv = iclog;
1854                 XFS_BUF_ZEROFLAGS(bp);
1855                 XFS_BUF_ASYNC(bp);
1856                 bp->b_flags |= XBF_SYNCIO;
1857                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1858                         bp->b_flags |= XBF_FUA;
1859                 /* use high priority completion wq */
1860                 bp->b_ioend_wq = log->l_mp->m_log_workqueue;
1861
1862                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1863                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1864
1865                 /* account for internal log which doesn't start at block #0 */
1866                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1867                 XFS_BUF_WRITE(bp);
1868                 error = xlog_bdstrat(bp);
1869                 if (error) {
1870                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1871                         return error;
1872                 }
1873         }
1874         return 0;
1875 }       /* xlog_sync */
1876
1877 /*
1878  * Deallocate a log structure
1879  */
1880 STATIC void
1881 xlog_dealloc_log(
1882         struct xlog     *log)
1883 {
1884         xlog_in_core_t  *iclog, *next_iclog;
1885         int             i;
1886
1887         xlog_cil_destroy(log);
1888
1889         /*
1890          * Cycle all the iclogbuf locks to make sure all log IO completion
1891          * is done before we tear down these buffers.
1892          */
1893         iclog = log->l_iclog;
1894         for (i = 0; i < log->l_iclog_bufs; i++) {
1895                 xfs_buf_lock(iclog->ic_bp);
1896                 xfs_buf_unlock(iclog->ic_bp);
1897                 iclog = iclog->ic_next;
1898         }
1899
1900         /*
1901          * Always need to ensure that the extra buffer does not point to memory
1902          * owned by another log buffer before we free it. Also, cycle the lock
1903          * first to ensure we've completed IO on it.
1904          */
1905         xfs_buf_lock(log->l_xbuf);
1906         xfs_buf_unlock(log->l_xbuf);
1907         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1908         xfs_buf_free(log->l_xbuf);
1909
1910         iclog = log->l_iclog;
1911         for (i = 0; i < log->l_iclog_bufs; i++) {
1912                 xfs_buf_free(iclog->ic_bp);
1913                 next_iclog = iclog->ic_next;
1914                 kmem_free(iclog);
1915                 iclog = next_iclog;
1916         }
1917         spinlock_destroy(&log->l_icloglock);
1918
1919         log->l_mp->m_log = NULL;
1920         kmem_free(log);
1921 }       /* xlog_dealloc_log */
1922
1923 /*
1924  * Update counters atomically now that memcpy is done.
1925  */
1926 /* ARGSUSED */
1927 static inline void
1928 xlog_state_finish_copy(
1929         struct xlog             *log,
1930         struct xlog_in_core     *iclog,
1931         int                     record_cnt,
1932         int                     copy_bytes)
1933 {
1934         spin_lock(&log->l_icloglock);
1935
1936         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1937         iclog->ic_offset += copy_bytes;
1938
1939         spin_unlock(&log->l_icloglock);
1940 }       /* xlog_state_finish_copy */
1941
1942
1943
1944
1945 /*
1946  * print out info relating to regions written which consume
1947  * the reservation
1948  */
1949 void
1950 xlog_print_tic_res(
1951         struct xfs_mount        *mp,
1952         struct xlog_ticket      *ticket)
1953 {
1954         uint i;
1955         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1956
1957         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1958         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1959             "bformat",
1960             "bchunk",
1961             "efi_format",
1962             "efd_format",
1963             "iformat",
1964             "icore",
1965             "iext",
1966             "ibroot",
1967             "ilocal",
1968             "iattr_ext",
1969             "iattr_broot",
1970             "iattr_local",
1971             "qformat",
1972             "dquot",
1973             "quotaoff",
1974             "LR header",
1975             "unmount",
1976             "commit",
1977             "trans header"
1978         };
1979         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1980             "SETATTR_NOT_SIZE",
1981             "SETATTR_SIZE",
1982             "INACTIVE",
1983             "CREATE",
1984             "CREATE_TRUNC",
1985             "TRUNCATE_FILE",
1986             "REMOVE",
1987             "LINK",
1988             "RENAME",
1989             "MKDIR",
1990             "RMDIR",
1991             "SYMLINK",
1992             "SET_DMATTRS",
1993             "GROWFS",
1994             "STRAT_WRITE",
1995             "DIOSTRAT",
1996             "WRITE_SYNC",
1997             "WRITEID",
1998             "ADDAFORK",
1999             "ATTRINVAL",
2000             "ATRUNCATE",
2001             "ATTR_SET",
2002             "ATTR_RM",
2003             "ATTR_FLAG",
2004             "CLEAR_AGI_BUCKET",
2005             "QM_SBCHANGE",
2006             "DUMMY1",
2007             "DUMMY2",
2008             "QM_QUOTAOFF",
2009             "QM_DQALLOC",
2010             "QM_SETQLIM",
2011             "QM_DQCLUSTER",
2012             "QM_QINOCREATE",
2013             "QM_QUOTAOFF_END",
2014             "SB_UNIT",
2015             "FSYNC_TS",
2016             "GROWFSRT_ALLOC",
2017             "GROWFSRT_ZERO",
2018             "GROWFSRT_FREE",
2019             "SWAPEXT"
2020         };
2021
2022         xfs_warn(mp,
2023                 "xlog_write: reservation summary:\n"
2024                 "  trans type  = %s (%u)\n"
2025                 "  unit res    = %d bytes\n"
2026                 "  current res = %d bytes\n"
2027                 "  total reg   = %u bytes (o/flow = %u bytes)\n"
2028                 "  ophdrs      = %u (ophdr space = %u bytes)\n"
2029                 "  ophdr + reg = %u bytes\n"
2030                 "  num regions = %u\n",
2031                 ((ticket->t_trans_type <= 0 ||
2032                   ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
2033                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
2034                 ticket->t_trans_type,
2035                 ticket->t_unit_res,
2036                 ticket->t_curr_res,
2037                 ticket->t_res_arr_sum, ticket->t_res_o_flow,
2038                 ticket->t_res_num_ophdrs, ophdr_spc,
2039                 ticket->t_res_arr_sum +
2040                 ticket->t_res_o_flow + ophdr_spc,
2041                 ticket->t_res_num);
2042
2043         for (i = 0; i < ticket->t_res_num; i++) {
2044                 uint r_type = ticket->t_res_arr[i].r_type;
2045                 xfs_warn(mp, "region[%u]: %s - %u bytes", i,
2046                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
2047                             "bad-rtype" : res_type_str[r_type-1]),
2048                             ticket->t_res_arr[i].r_len);
2049         }
2050
2051         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
2052                 "xlog_write: reservation ran out. Need to up reservation");
2053         xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
2054 }
2055
2056 /*
2057  * Calculate the potential space needed by the log vector.  Each region gets
2058  * its own xlog_op_header_t and may need to be double word aligned.
2059  */
2060 static int
2061 xlog_write_calc_vec_length(
2062         struct xlog_ticket      *ticket,
2063         struct xfs_log_vec      *log_vector)
2064 {
2065         struct xfs_log_vec      *lv;
2066         int                     headers = 0;
2067         int                     len = 0;
2068         int                     i;
2069
2070         /* acct for start rec of xact */
2071         if (ticket->t_flags & XLOG_TIC_INITED)
2072                 headers++;
2073
2074         for (lv = log_vector; lv; lv = lv->lv_next) {
2075                 /* we don't write ordered log vectors */
2076                 if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED)
2077                         continue;
2078
2079                 headers += lv->lv_niovecs;
2080
2081                 for (i = 0; i < lv->lv_niovecs; i++) {
2082                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
2083
2084                         len += vecp->i_len;
2085                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
2086                 }
2087         }
2088
2089         ticket->t_res_num_ophdrs += headers;
2090         len += headers * sizeof(struct xlog_op_header);
2091
2092         return len;
2093 }
2094
2095 /*
2096  * If first write for transaction, insert start record  We can't be trying to
2097  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
2098  */
2099 static int
2100 xlog_write_start_rec(
2101         struct xlog_op_header   *ophdr,
2102         struct xlog_ticket      *ticket)
2103 {
2104         if (!(ticket->t_flags & XLOG_TIC_INITED))
2105                 return 0;
2106
2107         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
2108         ophdr->oh_clientid = ticket->t_clientid;
2109         ophdr->oh_len = 0;
2110         ophdr->oh_flags = XLOG_START_TRANS;
2111         ophdr->oh_res2 = 0;
2112
2113         ticket->t_flags &= ~XLOG_TIC_INITED;
2114
2115         return sizeof(struct xlog_op_header);
2116 }
2117
2118 static xlog_op_header_t *
2119 xlog_write_setup_ophdr(
2120         struct xlog             *log,
2121         struct xlog_op_header   *ophdr,
2122         struct xlog_ticket      *ticket,
2123         uint                    flags)
2124 {
2125         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
2126         ophdr->oh_clientid = ticket->t_clientid;
2127         ophdr->oh_res2 = 0;
2128
2129         /* are we copying a commit or unmount record? */
2130         ophdr->oh_flags = flags;
2131
2132         /*
2133          * We've seen logs corrupted with bad transaction client ids.  This
2134          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
2135          * and shut down the filesystem.
2136          */
2137         switch (ophdr->oh_clientid)  {
2138         case XFS_TRANSACTION:
2139         case XFS_VOLUME:
2140         case XFS_LOG:
2141                 break;
2142         default:
2143                 xfs_warn(log->l_mp,
2144                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
2145                         ophdr->oh_clientid, ticket);
2146                 return NULL;
2147         }
2148
2149         return ophdr;
2150 }
2151
2152 /*
2153  * Set up the parameters of the region copy into the log. This has
2154  * to handle region write split across multiple log buffers - this
2155  * state is kept external to this function so that this code can
2156  * be written in an obvious, self documenting manner.
2157  */
2158 static int
2159 xlog_write_setup_copy(
2160         struct xlog_ticket      *ticket,
2161         struct xlog_op_header   *ophdr,
2162         int                     space_available,
2163         int                     space_required,
2164         int                     *copy_off,
2165         int                     *copy_len,
2166         int                     *last_was_partial_copy,
2167         int                     *bytes_consumed)
2168 {
2169         int                     still_to_copy;
2170
2171         still_to_copy = space_required - *bytes_consumed;
2172         *copy_off = *bytes_consumed;
2173
2174         if (still_to_copy <= space_available) {
2175                 /* write of region completes here */
2176                 *copy_len = still_to_copy;
2177                 ophdr->oh_len = cpu_to_be32(*copy_len);
2178                 if (*last_was_partial_copy)
2179                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
2180                 *last_was_partial_copy = 0;
2181                 *bytes_consumed = 0;
2182                 return 0;
2183         }
2184
2185         /* partial write of region, needs extra log op header reservation */
2186         *copy_len = space_available;
2187         ophdr->oh_len = cpu_to_be32(*copy_len);
2188         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
2189         if (*last_was_partial_copy)
2190                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
2191         *bytes_consumed += *copy_len;
2192         (*last_was_partial_copy)++;
2193
2194         /* account for new log op header */
2195         ticket->t_curr_res -= sizeof(struct xlog_op_header);
2196         ticket->t_res_num_ophdrs++;
2197
2198         return sizeof(struct xlog_op_header);
2199 }
2200
2201 static int
2202 xlog_write_copy_finish(
2203         struct xlog             *log,
2204         struct xlog_in_core     *iclog,
2205         uint                    flags,
2206         int                     *record_cnt,
2207         int                     *data_cnt,
2208         int                     *partial_copy,
2209         int                     *partial_copy_len,
2210         int                     log_offset,
2211         struct xlog_in_core     **commit_iclog)
2212 {
2213         if (*partial_copy) {
2214                 /*
2215                  * This iclog has already been marked WANT_SYNC by
2216                  * xlog_state_get_iclog_space.
2217                  */
2218                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2219                 *record_cnt = 0;
2220                 *data_cnt = 0;
2221                 return xlog_state_release_iclog(log, iclog);
2222         }
2223
2224         *partial_copy = 0;
2225         *partial_copy_len = 0;
2226
2227         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
2228                 /* no more space in this iclog - push it. */
2229                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
2230                 *record_cnt = 0;
2231                 *data_cnt = 0;
2232
2233                 spin_lock(&log->l_icloglock);
2234                 xlog_state_want_sync(log, iclog);
2235                 spin_unlock(&log->l_icloglock);
2236
2237                 if (!commit_iclog)
2238                         return xlog_state_release_iclog(log, iclog);
2239                 ASSERT(flags & XLOG_COMMIT_TRANS);
2240                 *commit_iclog = iclog;
2241         }
2242
2243         return 0;
2244 }
2245
2246 /*
2247  * Write some region out to in-core log
2248  *
2249  * This will be called when writing externally provided regions or when
2250  * writing out a commit record for a given transaction.
2251  *
2252  * General algorithm:
2253  *      1. Find total length of this write.  This may include adding to the
2254  *              lengths passed in.
2255  *      2. Check whether we violate the tickets reservation.
2256  *      3. While writing to this iclog
2257  *          A. Reserve as much space in this iclog as can get
2258  *          B. If this is first write, save away start lsn
2259  *          C. While writing this region:
2260  *              1. If first write of transaction, write start record
2261  *              2. Write log operation header (header per region)
2262  *              3. Find out if we can fit entire region into this iclog
2263  *              4. Potentially, verify destination memcpy ptr
2264  *              5. Memcpy (partial) region
2265  *              6. If partial copy, release iclog; otherwise, continue
2266  *                      copying more regions into current iclog
2267  *      4. Mark want sync bit (in simulation mode)
2268  *      5. Release iclog for potential flush to on-disk log.
2269  *
2270  * ERRORS:
2271  * 1.   Panic if reservation is overrun.  This should never happen since
2272  *      reservation amounts are generated internal to the filesystem.
2273  * NOTES:
2274  * 1. Tickets are single threaded data structures.
2275  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2276  *      syncing routine.  When a single log_write region needs to span
2277  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2278  *      on all log operation writes which don't contain the end of the
2279  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2280  *      operation which contains the end of the continued log_write region.
2281  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2282  *      we don't really know exactly how much space will be used.  As a result,
2283  *      we don't update ic_offset until the end when we know exactly how many
2284  *      bytes have been written out.
2285  */
2286 int
2287 xlog_write(
2288         struct xlog             *log,
2289         struct xfs_log_vec      *log_vector,
2290         struct xlog_ticket      *ticket,
2291         xfs_lsn_t               *start_lsn,
2292         struct xlog_in_core     **commit_iclog,
2293         uint                    flags)
2294 {
2295         struct xlog_in_core     *iclog = NULL;
2296         struct xfs_log_iovec    *vecp;
2297         struct xfs_log_vec      *lv;
2298         int                     len;
2299         int                     index;
2300         int                     partial_copy = 0;
2301         int                     partial_copy_len = 0;
2302         int                     contwr = 0;
2303         int                     record_cnt = 0;
2304         int                     data_cnt = 0;
2305         int                     error;
2306
2307         *start_lsn = 0;
2308
2309         len = xlog_write_calc_vec_length(ticket, log_vector);
2310
2311         /*
2312          * Region headers and bytes are already accounted for.
2313          * We only need to take into account start records and
2314          * split regions in this function.
2315          */
2316         if (ticket->t_flags & XLOG_TIC_INITED)
2317                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2318
2319         /*
2320          * Commit record headers need to be accounted for. These
2321          * come in as separate writes so are easy to detect.
2322          */
2323         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2324                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2325
2326         if (ticket->t_curr_res < 0)
2327                 xlog_print_tic_res(log->l_mp, ticket);
2328
2329         index = 0;
2330         lv = log_vector;
2331         vecp = lv->lv_iovecp;
2332         while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2333                 void            *ptr;
2334                 int             log_offset;
2335
2336                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2337                                                    &contwr, &log_offset);
2338                 if (error)
2339                         return error;
2340
2341                 ASSERT(log_offset <= iclog->ic_size - 1);
2342                 ptr = iclog->ic_datap + log_offset;
2343
2344                 /* start_lsn is the first lsn written to. That's all we need. */
2345                 if (!*start_lsn)
2346                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2347
2348                 /*
2349                  * This loop writes out as many regions as can fit in the amount
2350                  * of space which was allocated by xlog_state_get_iclog_space().
2351                  */
2352                 while (lv && (!lv->lv_niovecs || index < lv->lv_niovecs)) {
2353                         struct xfs_log_iovec    *reg;
2354                         struct xlog_op_header   *ophdr;
2355                         int                     start_rec_copy;
2356                         int                     copy_len;
2357                         int                     copy_off;
2358                         bool                    ordered = false;
2359
2360                         /* ordered log vectors have no regions to write */
2361                         if (lv->lv_buf_len == XFS_LOG_VEC_ORDERED) {
2362                                 ASSERT(lv->lv_niovecs == 0);
2363                                 ordered = true;
2364                                 goto next_lv;
2365                         }
2366
2367                         reg = &vecp[index];
2368                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2369                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2370
2371                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2372                         if (start_rec_copy) {
2373                                 record_cnt++;
2374                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2375                                                    start_rec_copy);
2376                         }
2377
2378                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2379                         if (!ophdr)
2380                                 return -EIO;
2381
2382                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2383                                            sizeof(struct xlog_op_header));
2384
2385                         len += xlog_write_setup_copy(ticket, ophdr,
2386                                                      iclog->ic_size-log_offset,
2387                                                      reg->i_len,
2388                                                      &copy_off, &copy_len,
2389                                                      &partial_copy,
2390                                                      &partial_copy_len);
2391                         xlog_verify_dest_ptr(log, ptr);
2392
2393                         /* copy region */
2394                         ASSERT(copy_len >= 0);
2395                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
2396                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2397
2398                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2399                         record_cnt++;
2400                         data_cnt += contwr ? copy_len : 0;
2401
2402                         error = xlog_write_copy_finish(log, iclog, flags,
2403                                                        &record_cnt, &data_cnt,
2404                                                        &partial_copy,
2405                                                        &partial_copy_len,
2406                                                        log_offset,
2407                                                        commit_iclog);
2408                         if (error)
2409                                 return error;
2410
2411                         /*
2412                          * if we had a partial copy, we need to get more iclog
2413                          * space but we don't want to increment the region
2414                          * index because there is still more is this region to
2415                          * write.
2416                          *
2417                          * If we completed writing this region, and we flushed
2418                          * the iclog (indicated by resetting of the record
2419                          * count), then we also need to get more log space. If
2420                          * this was the last record, though, we are done and
2421                          * can just return.
2422                          */
2423                         if (partial_copy)
2424                                 break;
2425
2426                         if (++index == lv->lv_niovecs) {
2427 next_lv:
2428                                 lv = lv->lv_next;
2429                                 index = 0;
2430                                 if (lv)
2431                                         vecp = lv->lv_iovecp;
2432                         }
2433                         if (record_cnt == 0 && ordered == false) {
2434                                 if (!lv)
2435                                         return 0;
2436                                 break;
2437                         }
2438                 }
2439         }
2440
2441         ASSERT(len == 0);
2442
2443         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2444         if (!commit_iclog)
2445                 return xlog_state_release_iclog(log, iclog);
2446
2447         ASSERT(flags & XLOG_COMMIT_TRANS);
2448         *commit_iclog = iclog;
2449         return 0;
2450 }
2451
2452
2453 /*****************************************************************************
2454  *
2455  *              State Machine functions
2456  *
2457  *****************************************************************************
2458  */
2459
2460 /* Clean iclogs starting from the head.  This ordering must be
2461  * maintained, so an iclog doesn't become ACTIVE beyond one that
2462  * is SYNCING.  This is also required to maintain the notion that we use
2463  * a ordered wait queue to hold off would be writers to the log when every
2464  * iclog is trying to sync to disk.
2465  *
2466  * State Change: DIRTY -> ACTIVE
2467  */
2468 STATIC void
2469 xlog_state_clean_log(
2470         struct xlog *log)
2471 {
2472         xlog_in_core_t  *iclog;
2473         int changed = 0;
2474
2475         iclog = log->l_iclog;
2476         do {
2477                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2478                         iclog->ic_state = XLOG_STATE_ACTIVE;
2479                         iclog->ic_offset       = 0;
2480                         ASSERT(iclog->ic_callback == NULL);
2481                         /*
2482                          * If the number of ops in this iclog indicate it just
2483                          * contains the dummy transaction, we can
2484                          * change state into IDLE (the second time around).
2485                          * Otherwise we should change the state into
2486                          * NEED a dummy.
2487                          * We don't need to cover the dummy.
2488                          */
2489                         if (!changed &&
2490                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2491                                         XLOG_COVER_OPS)) {
2492                                 changed = 1;
2493                         } else {
2494                                 /*
2495                                  * We have two dirty iclogs so start over
2496                                  * This could also be num of ops indicates
2497                                  * this is not the dummy going out.
2498                                  */
2499                                 changed = 2;
2500                         }
2501                         iclog->ic_header.h_num_logops = 0;
2502                         memset(iclog->ic_header.h_cycle_data, 0,
2503                               sizeof(iclog->ic_header.h_cycle_data));
2504                         iclog->ic_header.h_lsn = 0;
2505                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2506                         /* do nothing */;
2507                 else
2508                         break;  /* stop cleaning */
2509                 iclog = iclog->ic_next;
2510         } while (iclog != log->l_iclog);
2511
2512         /* log is locked when we are called */
2513         /*
2514          * Change state for the dummy log recording.
2515          * We usually go to NEED. But we go to NEED2 if the changed indicates
2516          * we are done writing the dummy record.
2517          * If we are done with the second dummy recored (DONE2), then
2518          * we go to IDLE.
2519          */
2520         if (changed) {
2521                 switch (log->l_covered_state) {
2522                 case XLOG_STATE_COVER_IDLE:
2523                 case XLOG_STATE_COVER_NEED:
2524                 case XLOG_STATE_COVER_NEED2:
2525                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2526                         break;
2527
2528                 case XLOG_STATE_COVER_DONE:
2529                         if (changed == 1)
2530                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2531                         else
2532                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2533                         break;
2534
2535                 case XLOG_STATE_COVER_DONE2:
2536                         if (changed == 1)
2537                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2538                         else
2539                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2540                         break;
2541
2542                 default:
2543                         ASSERT(0);
2544                 }
2545         }
2546 }       /* xlog_state_clean_log */
2547
2548 STATIC xfs_lsn_t
2549 xlog_get_lowest_lsn(
2550         struct xlog     *log)
2551 {
2552         xlog_in_core_t  *lsn_log;
2553         xfs_lsn_t       lowest_lsn, lsn;
2554
2555         lsn_log = log->l_iclog;
2556         lowest_lsn = 0;
2557         do {
2558             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2559                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2560                 if ((lsn && !lowest_lsn) ||
2561                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2562                         lowest_lsn = lsn;
2563                 }
2564             }
2565             lsn_log = lsn_log->ic_next;
2566         } while (lsn_log != log->l_iclog);
2567         return lowest_lsn;
2568 }
2569
2570
2571 STATIC void
2572 xlog_state_do_callback(
2573         struct xlog             *log,
2574         int                     aborted,
2575         struct xlog_in_core     *ciclog)
2576 {
2577         xlog_in_core_t     *iclog;
2578         xlog_in_core_t     *first_iclog;        /* used to know when we've
2579                                                  * processed all iclogs once */
2580         xfs_log_callback_t *cb, *cb_next;
2581         int                flushcnt = 0;
2582         xfs_lsn_t          lowest_lsn;
2583         int                ioerrors;    /* counter: iclogs with errors */
2584         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2585         int                funcdidcallbacks; /* flag: function did callbacks */
2586         int                repeats;     /* for issuing console warnings if
2587                                          * looping too many times */
2588         int                wake = 0;
2589
2590         spin_lock(&log->l_icloglock);
2591         first_iclog = iclog = log->l_iclog;
2592         ioerrors = 0;
2593         funcdidcallbacks = 0;
2594         repeats = 0;
2595
2596         do {
2597                 /*
2598                  * Scan all iclogs starting with the one pointed to by the
2599                  * log.  Reset this starting point each time the log is
2600                  * unlocked (during callbacks).
2601                  *
2602                  * Keep looping through iclogs until one full pass is made
2603                  * without running any callbacks.
2604                  */
2605                 first_iclog = log->l_iclog;
2606                 iclog = log->l_iclog;
2607                 loopdidcallbacks = 0;
2608                 repeats++;
2609
2610                 do {
2611
2612                         /* skip all iclogs in the ACTIVE & DIRTY states */
2613                         if (iclog->ic_state &
2614                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2615                                 iclog = iclog->ic_next;
2616                                 continue;
2617                         }
2618
2619                         /*
2620                          * Between marking a filesystem SHUTDOWN and stopping
2621                          * the log, we do flush all iclogs to disk (if there
2622                          * wasn't a log I/O error). So, we do want things to
2623                          * go smoothly in case of just a SHUTDOWN  w/o a
2624                          * LOG_IO_ERROR.
2625                          */
2626                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2627                                 /*
2628                                  * Can only perform callbacks in order.  Since
2629                                  * this iclog is not in the DONE_SYNC/
2630                                  * DO_CALLBACK state, we skip the rest and
2631                                  * just try to clean up.  If we set our iclog
2632                                  * to DO_CALLBACK, we will not process it when
2633                                  * we retry since a previous iclog is in the
2634                                  * CALLBACK and the state cannot change since
2635                                  * we are holding the l_icloglock.
2636                                  */
2637                                 if (!(iclog->ic_state &
2638                                         (XLOG_STATE_DONE_SYNC |
2639                                                  XLOG_STATE_DO_CALLBACK))) {
2640                                         if (ciclog && (ciclog->ic_state ==
2641                                                         XLOG_STATE_DONE_SYNC)) {
2642                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2643                                         }
2644                                         break;
2645                                 }
2646                                 /*
2647                                  * We now have an iclog that is in either the
2648                                  * DO_CALLBACK or DONE_SYNC states. The other
2649                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2650                                  * caught by the above if and are going to
2651                                  * clean (i.e. we aren't doing their callbacks)
2652                                  * see the above if.
2653                                  */
2654
2655                                 /*
2656                                  * We will do one more check here to see if we
2657                                  * have chased our tail around.
2658                                  */
2659
2660                                 lowest_lsn = xlog_get_lowest_lsn(log);
2661                                 if (lowest_lsn &&
2662                                     XFS_LSN_CMP(lowest_lsn,
2663                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2664                                         iclog = iclog->ic_next;
2665                                         continue; /* Leave this iclog for
2666                                                    * another thread */
2667                                 }
2668
2669                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2670
2671
2672                                 /*
2673                                  * Completion of a iclog IO does not imply that
2674                                  * a transaction has completed, as transactions
2675                                  * can be large enough to span many iclogs. We
2676                                  * cannot change the tail of the log half way
2677                                  * through a transaction as this may be the only
2678                                  * transaction in the log and moving th etail to
2679                                  * point to the middle of it will prevent
2680                                  * recovery from finding the start of the
2681                                  * transaction. Hence we should only update the
2682                                  * last_sync_lsn if this iclog contains
2683                                  * transaction completion callbacks on it.
2684                                  *
2685                                  * We have to do this before we drop the
2686                                  * icloglock to ensure we are the only one that
2687                                  * can update it.
2688                                  */
2689                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2690                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2691                                 if (iclog->ic_callback)
2692                                         atomic64_set(&log->l_last_sync_lsn,
2693                                                 be64_to_cpu(iclog->ic_header.h_lsn));
2694
2695                         } else
2696                                 ioerrors++;
2697
2698                         spin_unlock(&log->l_icloglock);
2699
2700                         /*
2701                          * Keep processing entries in the callback list until
2702                          * we come around and it is empty.  We need to
2703                          * atomically see that the list is empty and change the
2704                          * state to DIRTY so that we don't miss any more
2705                          * callbacks being added.
2706                          */
2707                         spin_lock(&iclog->ic_callback_lock);
2708                         cb = iclog->ic_callback;
2709                         while (cb) {
2710                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2711                                 iclog->ic_callback = NULL;
2712                                 spin_unlock(&iclog->ic_callback_lock);
2713
2714                                 /* perform callbacks in the order given */
2715                                 for (; cb; cb = cb_next) {
2716                                         cb_next = cb->cb_next;
2717                                         cb->cb_func(cb->cb_arg, aborted);
2718                                 }
2719                                 spin_lock(&iclog->ic_callback_lock);
2720                                 cb = iclog->ic_callback;
2721                         }
2722
2723                         loopdidcallbacks++;
2724                         funcdidcallbacks++;
2725
2726                         spin_lock(&log->l_icloglock);
2727                         ASSERT(iclog->ic_callback == NULL);
2728                         spin_unlock(&iclog->ic_callback_lock);
2729                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2730                                 iclog->ic_state = XLOG_STATE_DIRTY;
2731
2732                         /*
2733                          * Transition from DIRTY to ACTIVE if applicable.
2734                          * NOP if STATE_IOERROR.
2735                          */
2736                         xlog_state_clean_log(log);
2737
2738                         /* wake up threads waiting in xfs_log_force() */
2739                         wake_up_all(&iclog->ic_force_wait);
2740
2741                         iclog = iclog->ic_next;
2742                 } while (first_iclog != iclog);
2743
2744                 if (repeats > 5000) {
2745                         flushcnt += repeats;
2746                         repeats = 0;
2747                         xfs_warn(log->l_mp,
2748                                 "%s: possible infinite loop (%d iterations)",
2749                                 __func__, flushcnt);
2750                 }
2751         } while (!ioerrors && loopdidcallbacks);
2752
2753         /*
2754          * make one last gasp attempt to see if iclogs are being left in
2755          * limbo..
2756          */
2757 #ifdef DEBUG
2758         if (funcdidcallbacks) {
2759                 first_iclog = iclog = log->l_iclog;
2760                 do {
2761                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2762                         /*
2763                          * Terminate the loop if iclogs are found in states
2764                          * which will cause other threads to clean up iclogs.
2765                          *
2766                          * SYNCING - i/o completion will go through logs
2767                          * DONE_SYNC - interrupt thread should be waiting for
2768                          *              l_icloglock
2769                          * IOERROR - give up hope all ye who enter here
2770                          */
2771                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2772                             iclog->ic_state == XLOG_STATE_SYNCING ||
2773                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2774                             iclog->ic_state == XLOG_STATE_IOERROR )
2775                                 break;
2776                         iclog = iclog->ic_next;
2777                 } while (first_iclog != iclog);
2778         }
2779 #endif
2780
2781         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2782                 wake = 1;
2783         spin_unlock(&log->l_icloglock);
2784
2785         if (wake)
2786                 wake_up_all(&log->l_flush_wait);
2787 }
2788
2789
2790 /*
2791  * Finish transitioning this iclog to the dirty state.
2792  *
2793  * Make sure that we completely execute this routine only when this is
2794  * the last call to the iclog.  There is a good chance that iclog flushes,
2795  * when we reach the end of the physical log, get turned into 2 separate
2796  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2797  * routine.  By using the reference count bwritecnt, we guarantee that only
2798  * the second completion goes through.
2799  *
2800  * Callbacks could take time, so they are done outside the scope of the
2801  * global state machine log lock.
2802  */
2803 STATIC void
2804 xlog_state_done_syncing(
2805         xlog_in_core_t  *iclog,
2806         int             aborted)
2807 {
2808         struct xlog        *log = iclog->ic_log;
2809
2810         spin_lock(&log->l_icloglock);
2811
2812         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2813                iclog->ic_state == XLOG_STATE_IOERROR);
2814         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2815         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2816
2817
2818         /*
2819          * If we got an error, either on the first buffer, or in the case of
2820          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2821          * and none should ever be attempted to be written to disk
2822          * again.
2823          */
2824         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2825                 if (--iclog->ic_bwritecnt == 1) {
2826                         spin_unlock(&log->l_icloglock);
2827                         return;
2828                 }
2829                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2830         }
2831
2832         /*
2833          * Someone could be sleeping prior to writing out the next
2834          * iclog buffer, we wake them all, one will get to do the
2835          * I/O, the others get to wait for the result.
2836          */
2837         wake_up_all(&iclog->ic_write_wait);
2838         spin_unlock(&log->l_icloglock);
2839         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2840 }       /* xlog_state_done_syncing */
2841
2842
2843 /*
2844  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2845  * sleep.  We wait on the flush queue on the head iclog as that should be
2846  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2847  * we will wait here and all new writes will sleep until a sync completes.
2848  *
2849  * The in-core logs are used in a circular fashion. They are not used
2850  * out-of-order even when an iclog past the head is free.
2851  *
2852  * return:
2853  *      * log_offset where xlog_write() can start writing into the in-core
2854  *              log's data space.
2855  *      * in-core log pointer to which xlog_write() should write.
2856  *      * boolean indicating this is a continued write to an in-core log.
2857  *              If this is the last write, then the in-core log's offset field
2858  *              needs to be incremented, depending on the amount of data which
2859  *              is copied.
2860  */
2861 STATIC int
2862 xlog_state_get_iclog_space(
2863         struct xlog             *log,
2864         int                     len,
2865         struct xlog_in_core     **iclogp,
2866         struct xlog_ticket      *ticket,
2867         int                     *continued_write,
2868         int                     *logoffsetp)
2869 {
2870         int               log_offset;
2871         xlog_rec_header_t *head;
2872         xlog_in_core_t    *iclog;
2873         int               error;
2874
2875 restart:
2876         spin_lock(&log->l_icloglock);
2877         if (XLOG_FORCED_SHUTDOWN(log)) {
2878                 spin_unlock(&log->l_icloglock);
2879                 return -EIO;
2880         }
2881
2882         iclog = log->l_iclog;
2883         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2884                 XFS_STATS_INC(xs_log_noiclogs);
2885
2886                 /* Wait for log writes to have flushed */
2887                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2888                 goto restart;
2889         }
2890
2891         head = &iclog->ic_header;
2892
2893         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2894         log_offset = iclog->ic_offset;
2895
2896         /* On the 1st write to an iclog, figure out lsn.  This works
2897          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2898          * committing to.  If the offset is set, that's how many blocks
2899          * must be written.
2900          */
2901         if (log_offset == 0) {
2902                 ticket->t_curr_res -= log->l_iclog_hsize;
2903                 xlog_tic_add_region(ticket,
2904                                     log->l_iclog_hsize,
2905                                     XLOG_REG_TYPE_LRHEADER);
2906                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2907                 head->h_lsn = cpu_to_be64(
2908                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2909                 ASSERT(log->l_curr_block >= 0);
2910         }
2911
2912         /* If there is enough room to write everything, then do it.  Otherwise,
2913          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2914          * bit is on, so this will get flushed out.  Don't update ic_offset
2915          * until you know exactly how many bytes get copied.  Therefore, wait
2916          * until later to update ic_offset.
2917          *
2918          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2919          * can fit into remaining data section.
2920          */
2921         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2922                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2923
2924                 /*
2925                  * If I'm the only one writing to this iclog, sync it to disk.
2926                  * We need to do an atomic compare and decrement here to avoid
2927                  * racing with concurrent atomic_dec_and_lock() calls in
2928                  * xlog_state_release_iclog() when there is more than one
2929                  * reference to the iclog.
2930                  */
2931                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2932                         /* we are the only one */
2933                         spin_unlock(&log->l_icloglock);
2934                         error = xlog_state_release_iclog(log, iclog);
2935                         if (error)
2936                                 return error;
2937                 } else {
2938                         spin_unlock(&log->l_icloglock);
2939                 }
2940                 goto restart;
2941         }
2942
2943         /* Do we have enough room to write the full amount in the remainder
2944          * of this iclog?  Or must we continue a write on the next iclog and
2945          * mark this iclog as completely taken?  In the case where we switch
2946          * iclogs (to mark it taken), this particular iclog will release/sync
2947          * to disk in xlog_write().
2948          */
2949         if (len <= iclog->ic_size - iclog->ic_offset) {
2950                 *continued_write = 0;
2951                 iclog->ic_offset += len;
2952         } else {
2953                 *continued_write = 1;
2954                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2955         }
2956         *iclogp = iclog;
2957
2958         ASSERT(iclog->ic_offset <= iclog->ic_size);
2959         spin_unlock(&log->l_icloglock);
2960
2961         *logoffsetp = log_offset;
2962         return 0;
2963 }       /* xlog_state_get_iclog_space */
2964
2965 /* The first cnt-1 times through here we don't need to
2966  * move the grant write head because the permanent
2967  * reservation has reserved cnt times the unit amount.
2968  * Release part of current permanent unit reservation and
2969  * reset current reservation to be one units worth.  Also
2970  * move grant reservation head forward.
2971  */
2972 STATIC void
2973 xlog_regrant_reserve_log_space(
2974         struct xlog             *log,
2975         struct xlog_ticket      *ticket)
2976 {
2977         trace_xfs_log_regrant_reserve_enter(log, ticket);
2978
2979         if (ticket->t_cnt > 0)
2980                 ticket->t_cnt--;
2981
2982         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2983                                         ticket->t_curr_res);
2984         xlog_grant_sub_space(log, &log->l_write_head.grant,
2985                                         ticket->t_curr_res);
2986         ticket->t_curr_res = ticket->t_unit_res;
2987         xlog_tic_reset_res(ticket);
2988
2989         trace_xfs_log_regrant_reserve_sub(log, ticket);
2990
2991         /* just return if we still have some of the pre-reserved space */
2992         if (ticket->t_cnt > 0)
2993                 return;
2994
2995         xlog_grant_add_space(log, &log->l_reserve_head.grant,
2996                                         ticket->t_unit_res);
2997
2998         trace_xfs_log_regrant_reserve_exit(log, ticket);
2999
3000         ticket->t_curr_res = ticket->t_unit_res;
3001         xlog_tic_reset_res(ticket);
3002 }       /* xlog_regrant_reserve_log_space */
3003
3004
3005 /*
3006  * Give back the space left from a reservation.
3007  *
3008  * All the information we need to make a correct determination of space left
3009  * is present.  For non-permanent reservations, things are quite easy.  The
3010  * count should have been decremented to zero.  We only need to deal with the
3011  * space remaining in the current reservation part of the ticket.  If the
3012  * ticket contains a permanent reservation, there may be left over space which
3013  * needs to be released.  A count of N means that N-1 refills of the current
3014  * reservation can be done before we need to ask for more space.  The first
3015  * one goes to fill up the first current reservation.  Once we run out of
3016  * space, the count will stay at zero and the only space remaining will be
3017  * in the current reservation field.
3018  */
3019 STATIC void
3020 xlog_ungrant_log_space(
3021         struct xlog             *log,
3022         struct xlog_ticket      *ticket)
3023 {
3024         int     bytes;
3025
3026         if (ticket->t_cnt > 0)
3027                 ticket->t_cnt--;
3028
3029         trace_xfs_log_ungrant_enter(log, ticket);
3030         trace_xfs_log_ungrant_sub(log, ticket);
3031
3032         /*
3033          * If this is a permanent reservation ticket, we may be able to free
3034          * up more space based on the remaining count.
3035          */
3036         bytes = ticket->t_curr_res;
3037         if (ticket->t_cnt > 0) {
3038                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
3039                 bytes += ticket->t_unit_res*ticket->t_cnt;
3040         }
3041
3042         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
3043         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
3044
3045         trace_xfs_log_ungrant_exit(log, ticket);
3046
3047         xfs_log_space_wake(log->l_mp);
3048 }
3049
3050 /*
3051  * Flush iclog to disk if this is the last reference to the given iclog and
3052  * the WANT_SYNC bit is set.
3053  *
3054  * When this function is entered, the iclog is not necessarily in the
3055  * WANT_SYNC state.  It may be sitting around waiting to get filled.
3056  *
3057  *
3058  */
3059 STATIC int
3060 xlog_state_release_iclog(
3061         struct xlog             *log,
3062         struct xlog_in_core     *iclog)
3063 {
3064         int             sync = 0;       /* do we sync? */
3065
3066         if (iclog->ic_state & XLOG_STATE_IOERROR)
3067                 return -EIO;
3068
3069         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
3070         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
3071                 return 0;
3072
3073         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3074                 spin_unlock(&log->l_icloglock);
3075                 return -EIO;
3076         }
3077         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
3078                iclog->ic_state == XLOG_STATE_WANT_SYNC);
3079
3080         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
3081                 /* update tail before writing to iclog */
3082                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
3083                 sync++;
3084                 iclog->ic_state = XLOG_STATE_SYNCING;
3085                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
3086                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
3087                 /* cycle incremented when incrementing curr_block */
3088         }
3089         spin_unlock(&log->l_icloglock);
3090
3091         /*
3092          * We let the log lock go, so it's possible that we hit a log I/O
3093          * error or some other SHUTDOWN condition that marks the iclog
3094          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
3095          * this iclog has consistent data, so we ignore IOERROR
3096          * flags after this point.
3097          */
3098         if (sync)
3099                 return xlog_sync(log, iclog);
3100         return 0;
3101 }       /* xlog_state_release_iclog */
3102
3103
3104 /*
3105  * This routine will mark the current iclog in the ring as WANT_SYNC
3106  * and move the current iclog pointer to the next iclog in the ring.
3107  * When this routine is called from xlog_state_get_iclog_space(), the
3108  * exact size of the iclog has not yet been determined.  All we know is
3109  * that every data block.  We have run out of space in this log record.
3110  */
3111 STATIC void
3112 xlog_state_switch_iclogs(
3113         struct xlog             *log,
3114         struct xlog_in_core     *iclog,
3115         int                     eventual_size)
3116 {
3117         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
3118         if (!eventual_size)
3119                 eventual_size = iclog->ic_offset;
3120         iclog->ic_state = XLOG_STATE_WANT_SYNC;
3121         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
3122         log->l_prev_block = log->l_curr_block;
3123         log->l_prev_cycle = log->l_curr_cycle;
3124
3125         /* roll log?: ic_offset changed later */
3126         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
3127
3128         /* Round up to next log-sunit */
3129         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3130             log->l_mp->m_sb.sb_logsunit > 1) {
3131                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
3132                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
3133         }
3134
3135         if (log->l_curr_block >= log->l_logBBsize) {
3136                 log->l_curr_cycle++;
3137                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
3138                         log->l_curr_cycle++;
3139                 log->l_curr_block -= log->l_logBBsize;
3140                 ASSERT(log->l_curr_block >= 0);
3141         }
3142         ASSERT(iclog == log->l_iclog);
3143         log->l_iclog = iclog->ic_next;
3144 }       /* xlog_state_switch_iclogs */
3145
3146 /*
3147  * Write out all data in the in-core log as of this exact moment in time.
3148  *
3149  * Data may be written to the in-core log during this call.  However,
3150  * we don't guarantee this data will be written out.  A change from past
3151  * implementation means this routine will *not* write out zero length LRs.
3152  *
3153  * Basically, we try and perform an intelligent scan of the in-core logs.
3154  * If we determine there is no flushable data, we just return.  There is no
3155  * flushable data if:
3156  *
3157  *      1. the current iclog is active and has no data; the previous iclog
3158  *              is in the active or dirty state.
3159  *      2. the current iclog is drity, and the previous iclog is in the
3160  *              active or dirty state.
3161  *
3162  * We may sleep if:
3163  *
3164  *      1. the current iclog is not in the active nor dirty state.
3165  *      2. the current iclog dirty, and the previous iclog is not in the
3166  *              active nor dirty state.
3167  *      3. the current iclog is active, and there is another thread writing
3168  *              to this particular iclog.
3169  *      4. a) the current iclog is active and has no other writers
3170  *         b) when we return from flushing out this iclog, it is still
3171  *              not in the active nor dirty state.
3172  */
3173 int
3174 _xfs_log_force(
3175         struct xfs_mount        *mp,
3176         uint                    flags,
3177         int                     *log_flushed)
3178 {
3179         struct xlog             *log = mp->m_log;
3180         struct xlog_in_core     *iclog;
3181         xfs_lsn_t               lsn;
3182
3183         XFS_STATS_INC(xs_log_force);
3184
3185         xlog_cil_force(log);
3186
3187         spin_lock(&log->l_icloglock);
3188
3189         iclog = log->l_iclog;
3190         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3191                 spin_unlock(&log->l_icloglock);
3192                 return -EIO;
3193         }
3194
3195         /* If the head iclog is not active nor dirty, we just attach
3196          * ourselves to the head and go to sleep.
3197          */
3198         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3199             iclog->ic_state == XLOG_STATE_DIRTY) {
3200                 /*
3201                  * If the head is dirty or (active and empty), then
3202                  * we need to look at the previous iclog.  If the previous
3203                  * iclog is active or dirty we are done.  There is nothing
3204                  * to sync out.  Otherwise, we attach ourselves to the
3205                  * previous iclog and go to sleep.
3206                  */
3207                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
3208                     (atomic_read(&iclog->ic_refcnt) == 0
3209                      && iclog->ic_offset == 0)) {
3210                         iclog = iclog->ic_prev;
3211                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
3212                             iclog->ic_state == XLOG_STATE_DIRTY)
3213                                 goto no_sleep;
3214                         else
3215                                 goto maybe_sleep;
3216                 } else {
3217                         if (atomic_read(&iclog->ic_refcnt) == 0) {
3218                                 /* We are the only one with access to this
3219                                  * iclog.  Flush it out now.  There should
3220                                  * be a roundoff of zero to show that someone
3221                                  * has already taken care of the roundoff from
3222                                  * the previous sync.
3223                                  */
3224                                 atomic_inc(&iclog->ic_refcnt);
3225                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
3226                                 xlog_state_switch_iclogs(log, iclog, 0);
3227                                 spin_unlock(&log->l_icloglock);
3228
3229                                 if (xlog_state_release_iclog(log, iclog))
3230                                         return -EIO;
3231
3232                                 if (log_flushed)
3233                                         *log_flushed = 1;
3234                                 spin_lock(&log->l_icloglock);
3235                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
3236                                     iclog->ic_state != XLOG_STATE_DIRTY)
3237                                         goto maybe_sleep;
3238                                 else
3239                                         goto no_sleep;
3240                         } else {
3241                                 /* Someone else is writing to this iclog.
3242                                  * Use its call to flush out the data.  However,
3243                                  * the other thread may not force out this LR,
3244                                  * so we mark it WANT_SYNC.
3245                                  */
3246                                 xlog_state_switch_iclogs(log, iclog, 0);
3247                                 goto maybe_sleep;
3248                         }
3249                 }
3250         }
3251
3252         /* By the time we come around again, the iclog could've been filled
3253          * which would give it another lsn.  If we have a new lsn, just
3254          * return because the relevant data has been flushed.
3255          */
3256 maybe_sleep:
3257         if (flags & XFS_LOG_SYNC) {
3258                 /*
3259                  * We must check if we're shutting down here, before
3260                  * we wait, while we're holding the l_icloglock.
3261                  * Then we check again after waking up, in case our
3262                  * sleep was disturbed by a bad news.
3263                  */
3264                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
3265                         spin_unlock(&log->l_icloglock);
3266                         return -EIO;
3267                 }
3268                 XFS_STATS_INC(xs_log_force_sleep);
3269                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3270                 /*
3271                  * No need to grab the log lock here since we're
3272                  * only deciding whether or not to return EIO
3273                  * and the memory read should be atomic.
3274                  */
3275                 if (iclog->ic_state & XLOG_STATE_IOERROR)
3276                         return -EIO;
3277                 if (log_flushed)
3278                         *log_flushed = 1;
3279         } else {
3280
3281 no_sleep:
3282                 spin_unlock(&log->l_icloglock);
3283         }
3284         return 0;
3285 }
3286
3287 /*
3288  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
3289  * about errors or whether the log was flushed or not. This is the normal
3290  * interface to use when trying to unpin items or move the log forward.
3291  */
3292 void
3293 xfs_log_force(
3294         xfs_mount_t     *mp,
3295         uint            flags)
3296 {
3297         int     error;
3298
3299         trace_xfs_log_force(mp, 0);
3300         error = _xfs_log_force(mp, flags, NULL);
3301         if (error)
3302                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3303 }
3304
3305 /*
3306  * Force the in-core log to disk for a specific LSN.
3307  *
3308  * Find in-core log with lsn.
3309  *      If it is in the DIRTY state, just return.
3310  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3311  *              state and go to sleep or return.
3312  *      If it is in any other state, go to sleep or return.
3313  *
3314  * Synchronous forces are implemented with a signal variable. All callers
3315  * to force a given lsn to disk will wait on a the sv attached to the
3316  * specific in-core log.  When given in-core log finally completes its
3317  * write to disk, that thread will wake up all threads waiting on the
3318  * sv.
3319  */
3320 int
3321 _xfs_log_force_lsn(
3322         struct xfs_mount        *mp,
3323         xfs_lsn_t               lsn,
3324         uint                    flags,
3325         int                     *log_flushed)
3326 {
3327         struct xlog             *log = mp->m_log;
3328         struct xlog_in_core     *iclog;
3329         int                     already_slept = 0;
3330
3331         ASSERT(lsn != 0);
3332
3333         XFS_STATS_INC(xs_log_force);
3334
3335         lsn = xlog_cil_force_lsn(log, lsn);
3336         if (lsn == NULLCOMMITLSN)
3337                 return 0;
3338
3339 try_again:
3340         spin_lock(&log->l_icloglock);
3341         iclog = log->l_iclog;
3342         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3343                 spin_unlock(&log->l_icloglock);
3344                 return -EIO;
3345         }
3346
3347         do {
3348                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3349                         iclog = iclog->ic_next;
3350                         continue;
3351                 }
3352
3353                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3354                         spin_unlock(&log->l_icloglock);
3355                         return 0;
3356                 }
3357
3358                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3359                         /*
3360                          * We sleep here if we haven't already slept (e.g.
3361                          * this is the first time we've looked at the correct
3362                          * iclog buf) and the buffer before us is going to
3363                          * be sync'ed. The reason for this is that if we
3364                          * are doing sync transactions here, by waiting for
3365                          * the previous I/O to complete, we can allow a few
3366                          * more transactions into this iclog before we close
3367                          * it down.
3368                          *
3369                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3370                          * up the refcnt so we can release the log (which
3371                          * drops the ref count).  The state switch keeps new
3372                          * transaction commits from using this buffer.  When
3373                          * the current commits finish writing into the buffer,
3374                          * the refcount will drop to zero and the buffer will
3375                          * go out then.
3376                          */
3377                         if (!already_slept &&
3378                             (iclog->ic_prev->ic_state &
3379                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3380                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3381
3382                                 XFS_STATS_INC(xs_log_force_sleep);
3383
3384                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3385                                                         &log->l_icloglock);
3386                                 if (log_flushed)
3387                                         *log_flushed = 1;
3388                                 already_slept = 1;
3389                                 goto try_again;
3390                         }
3391                         atomic_inc(&iclog->ic_refcnt);
3392                         xlog_state_switch_iclogs(log, iclog, 0);
3393                         spin_unlock(&log->l_icloglock);
3394                         if (xlog_state_release_iclog(log, iclog))
3395                                 return -EIO;
3396                         if (log_flushed)
3397                                 *log_flushed = 1;
3398                         spin_lock(&log->l_icloglock);
3399                 }
3400
3401                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3402                     !(iclog->ic_state &
3403                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3404                         /*
3405                          * Don't wait on completion if we know that we've
3406                          * gotten a log write error.
3407                          */
3408                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3409                                 spin_unlock(&log->l_icloglock);
3410                                 return -EIO;
3411                         }
3412                         XFS_STATS_INC(xs_log_force_sleep);
3413                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3414                         /*
3415                          * No need to grab the log lock here since we're
3416                          * only deciding whether or not to return EIO
3417                          * and the memory read should be atomic.
3418                          */
3419                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3420                                 return -EIO;
3421
3422                         if (log_flushed)
3423                                 *log_flushed = 1;
3424                 } else {                /* just return */
3425                         spin_unlock(&log->l_icloglock);
3426                 }
3427
3428                 return 0;
3429         } while (iclog != log->l_iclog);
3430
3431         spin_unlock(&log->l_icloglock);
3432         return 0;
3433 }
3434
3435 /*
3436  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3437  * about errors or whether the log was flushed or not. This is the normal
3438  * interface to use when trying to unpin items or move the log forward.
3439  */
3440 void
3441 xfs_log_force_lsn(
3442         xfs_mount_t     *mp,
3443         xfs_lsn_t       lsn,
3444         uint            flags)
3445 {
3446         int     error;
3447
3448         trace_xfs_log_force(mp, lsn);
3449         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3450         if (error)
3451                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3452 }
3453
3454 /*
3455  * Called when we want to mark the current iclog as being ready to sync to
3456  * disk.
3457  */
3458 STATIC void
3459 xlog_state_want_sync(
3460         struct xlog             *log,
3461         struct xlog_in_core     *iclog)
3462 {
3463         assert_spin_locked(&log->l_icloglock);
3464
3465         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3466                 xlog_state_switch_iclogs(log, iclog, 0);
3467         } else {
3468                 ASSERT(iclog->ic_state &
3469                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3470         }
3471 }
3472
3473
3474 /*****************************************************************************
3475  *
3476  *              TICKET functions
3477  *
3478  *****************************************************************************
3479  */
3480
3481 /*
3482  * Free a used ticket when its refcount falls to zero.
3483  */
3484 void
3485 xfs_log_ticket_put(
3486         xlog_ticket_t   *ticket)
3487 {
3488         ASSERT(atomic_read(&ticket->t_ref) > 0);
3489         if (atomic_dec_and_test(&ticket->t_ref))
3490                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3491 }
3492
3493 xlog_ticket_t *
3494 xfs_log_ticket_get(
3495         xlog_ticket_t   *ticket)
3496 {
3497         ASSERT(atomic_read(&ticket->t_ref) > 0);
3498         atomic_inc(&ticket->t_ref);
3499         return ticket;
3500 }
3501
3502 /*
3503  * Figure out the total log space unit (in bytes) that would be
3504  * required for a log ticket.
3505  */
3506 int
3507 xfs_log_calc_unit_res(
3508         struct xfs_mount        *mp,
3509         int                     unit_bytes)
3510 {
3511         struct xlog             *log = mp->m_log;
3512         int                     iclog_space;
3513         uint                    num_headers;
3514
3515         /*
3516          * Permanent reservations have up to 'cnt'-1 active log operations
3517          * in the log.  A unit in this case is the amount of space for one
3518          * of these log operations.  Normal reservations have a cnt of 1
3519          * and their unit amount is the total amount of space required.
3520          *
3521          * The following lines of code account for non-transaction data
3522          * which occupy space in the on-disk log.
3523          *
3524          * Normal form of a transaction is:
3525          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3526          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3527          *
3528          * We need to account for all the leadup data and trailer data
3529          * around the transaction data.
3530          * And then we need to account for the worst case in terms of using
3531          * more space.
3532          * The worst case will happen if:
3533          * - the placement of the transaction happens to be such that the
3534          *   roundoff is at its maximum
3535          * - the transaction data is synced before the commit record is synced
3536          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3537          *   Therefore the commit record is in its own Log Record.
3538          *   This can happen as the commit record is called with its
3539          *   own region to xlog_write().
3540          *   This then means that in the worst case, roundoff can happen for
3541          *   the commit-rec as well.
3542          *   The commit-rec is smaller than padding in this scenario and so it is
3543          *   not added separately.
3544          */
3545
3546         /* for trans header */
3547         unit_bytes += sizeof(xlog_op_header_t);
3548         unit_bytes += sizeof(xfs_trans_header_t);
3549
3550         /* for start-rec */
3551         unit_bytes += sizeof(xlog_op_header_t);
3552
3553         /*
3554          * for LR headers - the space for data in an iclog is the size minus
3555          * the space used for the headers. If we use the iclog size, then we
3556          * undercalculate the number of headers required.
3557          *
3558          * Furthermore - the addition of op headers for split-recs might
3559          * increase the space required enough to require more log and op
3560          * headers, so take that into account too.
3561          *
3562          * IMPORTANT: This reservation makes the assumption that if this
3563          * transaction is the first in an iclog and hence has the LR headers
3564          * accounted to it, then the remaining space in the iclog is
3565          * exclusively for this transaction.  i.e. if the transaction is larger
3566          * than the iclog, it will be the only thing in that iclog.
3567          * Fundamentally, this means we must pass the entire log vector to
3568          * xlog_write to guarantee this.
3569          */
3570         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3571         num_headers = howmany(unit_bytes, iclog_space);
3572
3573         /* for split-recs - ophdrs added when data split over LRs */
3574         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3575
3576         /* add extra header reservations if we overrun */
3577         while (!num_headers ||
3578                howmany(unit_bytes, iclog_space) > num_headers) {
3579                 unit_bytes += sizeof(xlog_op_header_t);
3580                 num_headers++;
3581         }
3582         unit_bytes += log->l_iclog_hsize * num_headers;
3583
3584         /* for commit-rec LR header - note: padding will subsume the ophdr */
3585         unit_bytes += log->l_iclog_hsize;
3586
3587         /* for roundoff padding for transaction data and one for commit record */
3588         if (xfs_sb_version_haslogv2(&mp->m_sb) && mp->m_sb.sb_logsunit > 1) {
3589                 /* log su roundoff */
3590                 unit_bytes += 2 * mp->m_sb.sb_logsunit;
3591         } else {
3592                 /* BB roundoff */
3593                 unit_bytes += 2 * BBSIZE;
3594         }
3595
3596         return unit_bytes;
3597 }
3598
3599 /*
3600  * Allocate and initialise a new log ticket.
3601  */
3602 struct xlog_ticket *
3603 xlog_ticket_alloc(
3604         struct xlog             *log,
3605         int                     unit_bytes,
3606         int                     cnt,
3607         char                    client,
3608         bool                    permanent,
3609         xfs_km_flags_t          alloc_flags)
3610 {
3611         struct xlog_ticket      *tic;
3612         int                     unit_res;
3613
3614         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3615         if (!tic)
3616                 return NULL;
3617
3618         unit_res = xfs_log_calc_unit_res(log->l_mp, unit_bytes);
3619
3620         atomic_set(&tic->t_ref, 1);
3621         tic->t_task             = current;
3622         INIT_LIST_HEAD(&tic->t_queue);
3623         tic->t_unit_res         = unit_res;
3624         tic->t_curr_res         = unit_res;
3625         tic->t_cnt              = cnt;
3626         tic->t_ocnt             = cnt;
3627         tic->t_tid              = prandom_u32();
3628         tic->t_clientid         = client;
3629         tic->t_flags            = XLOG_TIC_INITED;
3630         tic->t_trans_type       = 0;
3631         if (permanent)
3632                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3633
3634         xlog_tic_reset_res(tic);
3635
3636         return tic;
3637 }
3638
3639
3640 /******************************************************************************
3641  *
3642  *              Log debug routines
3643  *
3644  ******************************************************************************
3645  */
3646 #if defined(DEBUG)
3647 /*
3648  * Make sure that the destination ptr is within the valid data region of
3649  * one of the iclogs.  This uses backup pointers stored in a different
3650  * part of the log in case we trash the log structure.
3651  */
3652 void
3653 xlog_verify_dest_ptr(
3654         struct xlog     *log,
3655         char            *ptr)
3656 {
3657         int i;
3658         int good_ptr = 0;
3659
3660         for (i = 0; i < log->l_iclog_bufs; i++) {
3661                 if (ptr >= log->l_iclog_bak[i] &&
3662                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3663                         good_ptr++;
3664         }
3665
3666         if (!good_ptr)
3667                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3668 }
3669
3670 /*
3671  * Check to make sure the grant write head didn't just over lap the tail.  If
3672  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3673  * the cycles differ by exactly one and check the byte count.
3674  *
3675  * This check is run unlocked, so can give false positives. Rather than assert
3676  * on failures, use a warn-once flag and a panic tag to allow the admin to
3677  * determine if they want to panic the machine when such an error occurs. For
3678  * debug kernels this will have the same effect as using an assert but, unlinke
3679  * an assert, it can be turned off at runtime.
3680  */
3681 STATIC void
3682 xlog_verify_grant_tail(
3683         struct xlog     *log)
3684 {
3685         int             tail_cycle, tail_blocks;
3686         int             cycle, space;
3687
3688         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3689         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3690         if (tail_cycle != cycle) {
3691                 if (cycle - 1 != tail_cycle &&
3692                     !(log->l_flags & XLOG_TAIL_WARN)) {
3693                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3694                                 "%s: cycle - 1 != tail_cycle", __func__);
3695                         log->l_flags |= XLOG_TAIL_WARN;
3696                 }
3697
3698                 if (space > BBTOB(tail_blocks) &&
3699                     !(log->l_flags & XLOG_TAIL_WARN)) {
3700                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3701                                 "%s: space > BBTOB(tail_blocks)", __func__);
3702                         log->l_flags |= XLOG_TAIL_WARN;
3703                 }
3704         }
3705 }
3706
3707 /* check if it will fit */
3708 STATIC void
3709 xlog_verify_tail_lsn(
3710         struct xlog             *log,
3711         struct xlog_in_core     *iclog,
3712         xfs_lsn_t               tail_lsn)
3713 {
3714     int blocks;
3715
3716     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3717         blocks =
3718             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3719         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3720                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3721     } else {
3722         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3723
3724         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3725                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3726
3727         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3728         if (blocks < BTOBB(iclog->ic_offset) + 1)
3729                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3730     }
3731 }       /* xlog_verify_tail_lsn */
3732
3733 /*
3734  * Perform a number of checks on the iclog before writing to disk.
3735  *
3736  * 1. Make sure the iclogs are still circular
3737  * 2. Make sure we have a good magic number
3738  * 3. Make sure we don't have magic numbers in the data
3739  * 4. Check fields of each log operation header for:
3740  *      A. Valid client identifier
3741  *      B. tid ptr value falls in valid ptr space (user space code)
3742  *      C. Length in log record header is correct according to the
3743  *              individual operation headers within record.
3744  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3745  *      log, check the preceding blocks of the physical log to make sure all
3746  *      the cycle numbers agree with the current cycle number.
3747  */
3748 STATIC void
3749 xlog_verify_iclog(
3750         struct xlog             *log,
3751         struct xlog_in_core     *iclog,
3752         int                     count,
3753         bool                    syncing)
3754 {
3755         xlog_op_header_t        *ophead;
3756         xlog_in_core_t          *icptr;
3757         xlog_in_core_2_t        *xhdr;
3758         xfs_caddr_t             ptr;
3759         xfs_caddr_t             base_ptr;
3760         __psint_t               field_offset;
3761         __uint8_t               clientid;
3762         int                     len, i, j, k, op_len;
3763         int                     idx;
3764
3765         /* check validity of iclog pointers */
3766         spin_lock(&log->l_icloglock);
3767         icptr = log->l_iclog;
3768         for (i = 0; i < log->l_iclog_bufs; i++, icptr = icptr->ic_next)
3769                 ASSERT(icptr);
3770
3771         if (icptr != log->l_iclog)
3772                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3773         spin_unlock(&log->l_icloglock);
3774
3775         /* check log magic numbers */
3776         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3777                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3778
3779         ptr = (xfs_caddr_t) &iclog->ic_header;
3780         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3781              ptr += BBSIZE) {
3782                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3783                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3784                                 __func__);
3785         }
3786
3787         /* check fields */
3788         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3789         ptr = iclog->ic_datap;
3790         base_ptr = ptr;
3791         ophead = (xlog_op_header_t *)ptr;
3792         xhdr = iclog->ic_data;
3793         for (i = 0; i < len; i++) {
3794                 ophead = (xlog_op_header_t *)ptr;
3795
3796                 /* clientid is only 1 byte */
3797                 field_offset = (__psint_t)
3798                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3799                 if (!syncing || (field_offset & 0x1ff)) {
3800                         clientid = ophead->oh_clientid;
3801                 } else {
3802                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3803                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3804                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3805                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3806                                 clientid = xlog_get_client_id(
3807                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3808                         } else {
3809                                 clientid = xlog_get_client_id(
3810                                         iclog->ic_header.h_cycle_data[idx]);
3811                         }
3812                 }
3813                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3814                         xfs_warn(log->l_mp,
3815                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3816                                 __func__, clientid, ophead,
3817                                 (unsigned long)field_offset);
3818
3819                 /* check length */
3820                 field_offset = (__psint_t)
3821                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3822                 if (!syncing || (field_offset & 0x1ff)) {
3823                         op_len = be32_to_cpu(ophead->oh_len);
3824                 } else {
3825                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3826                                     (__psint_t)iclog->ic_datap);
3827                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3828                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3829                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3830                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3831                         } else {
3832                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3833                         }
3834                 }
3835                 ptr += sizeof(xlog_op_header_t) + op_len;
3836         }
3837 }       /* xlog_verify_iclog */
3838 #endif
3839
3840 /*
3841  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3842  */
3843 STATIC int
3844 xlog_state_ioerror(
3845         struct xlog     *log)
3846 {
3847         xlog_in_core_t  *iclog, *ic;
3848
3849         iclog = log->l_iclog;
3850         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3851                 /*
3852                  * Mark all the incore logs IOERROR.
3853                  * From now on, no log flushes will result.
3854                  */
3855                 ic = iclog;
3856                 do {
3857                         ic->ic_state = XLOG_STATE_IOERROR;
3858                         ic = ic->ic_next;
3859                 } while (ic != iclog);
3860                 return 0;
3861         }
3862         /*
3863          * Return non-zero, if state transition has already happened.
3864          */
3865         return 1;
3866 }
3867
3868 /*
3869  * This is called from xfs_force_shutdown, when we're forcibly
3870  * shutting down the filesystem, typically because of an IO error.
3871  * Our main objectives here are to make sure that:
3872  *      a. if !logerror, flush the logs to disk. Anything modified
3873  *         after this is ignored.
3874  *      b. the filesystem gets marked 'SHUTDOWN' for all interested
3875  *         parties to find out, 'atomically'.
3876  *      c. those who're sleeping on log reservations, pinned objects and
3877  *          other resources get woken up, and be told the bad news.
3878  *      d. nothing new gets queued up after (b) and (c) are done.
3879  *
3880  * Note: for the !logerror case we need to flush the regions held in memory out
3881  * to disk first. This needs to be done before the log is marked as shutdown,
3882  * otherwise the iclog writes will fail.
3883  */
3884 int
3885 xfs_log_force_umount(
3886         struct xfs_mount        *mp,
3887         int                     logerror)
3888 {
3889         struct xlog     *log;
3890         int             retval;
3891
3892         log = mp->m_log;
3893
3894         /*
3895          * If this happens during log recovery, don't worry about
3896          * locking; the log isn't open for business yet.
3897          */
3898         if (!log ||
3899             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3900                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3901                 if (mp->m_sb_bp)
3902                         XFS_BUF_DONE(mp->m_sb_bp);
3903                 return 0;
3904         }
3905
3906         /*
3907          * Somebody could've already done the hard work for us.
3908          * No need to get locks for this.
3909          */
3910         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3911                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3912                 return 1;
3913         }
3914
3915         /*
3916          * Flush all the completed transactions to disk before marking the log
3917          * being shut down. We need to do it in this order to ensure that
3918          * completed operations are safely on disk before we shut down, and that
3919          * we don't have to issue any buffer IO after the shutdown flags are set
3920          * to guarantee this.
3921          */
3922         if (!logerror)
3923                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3924
3925         /*
3926          * mark the filesystem and the as in a shutdown state and wake
3927          * everybody up to tell them the bad news.
3928          */
3929         spin_lock(&log->l_icloglock);
3930         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3931         if (mp->m_sb_bp)
3932                 XFS_BUF_DONE(mp->m_sb_bp);
3933
3934         /*
3935          * Mark the log and the iclogs with IO error flags to prevent any
3936          * further log IO from being issued or completed.
3937          */
3938         log->l_flags |= XLOG_IO_ERROR;
3939         retval = xlog_state_ioerror(log);
3940         spin_unlock(&log->l_icloglock);
3941
3942         /*
3943          * We don't want anybody waiting for log reservations after this. That
3944          * means we have to wake up everybody queued up on reserveq as well as
3945          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3946          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3947          * action is protected by the grant locks.
3948          */
3949         xlog_grant_head_wake_all(&log->l_reserve_head);
3950         xlog_grant_head_wake_all(&log->l_write_head);
3951
3952         /*
3953          * Wake up everybody waiting on xfs_log_force. Wake the CIL push first
3954          * as if the log writes were completed. The abort handling in the log
3955          * item committed callback functions will do this again under lock to
3956          * avoid races.
3957          */
3958         wake_up_all(&log->l_cilp->xc_commit_wait);
3959         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3960
3961 #ifdef XFSERRORDEBUG
3962         {
3963                 xlog_in_core_t  *iclog;
3964
3965                 spin_lock(&log->l_icloglock);
3966                 iclog = log->l_iclog;
3967                 do {
3968                         ASSERT(iclog->ic_callback == 0);
3969                         iclog = iclog->ic_next;
3970                 } while (iclog != log->l_iclog);
3971                 spin_unlock(&log->l_icloglock);
3972         }
3973 #endif
3974         /* return non-zero if log IOERROR transition had already happened */
3975         return retval;
3976 }
3977
3978 STATIC int
3979 xlog_iclogs_empty(
3980         struct xlog     *log)
3981 {
3982         xlog_in_core_t  *iclog;
3983
3984         iclog = log->l_iclog;
3985         do {
3986                 /* endianness does not matter here, zero is zero in
3987                  * any language.
3988                  */
3989                 if (iclog->ic_header.h_num_logops)
3990                         return 0;
3991                 iclog = iclog->ic_next;
3992         } while (iclog != log->l_iclog);
3993         return 1;
3994 }
3995