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_types.h"
  21 #include "xfs_log.h"
  22 #include "xfs_trans.h"
  23 #include "xfs_sb.h"
  24 #include "xfs_ag.h"
  25 #include "xfs_mount.h"
  26 #include "xfs_error.h"
  27 #include "xfs_log_priv.h"
  28 #include "xfs_buf_item.h"
  29 #include "xfs_bmap_btree.h"
  30 #include "xfs_alloc_btree.h"
  31 #include "xfs_ialloc_btree.h"
  32 #include "xfs_log_recover.h"
  33 #include "xfs_trans_priv.h"
  34 #include "xfs_dinode.h"
  35 #include "xfs_inode.h"
  36 #include "xfs_trace.h"
  37
  38 kmem_zone_t     *xfs_log_ticket_zone;
  39
  40 /* Local miscellaneous function prototypes */
  41 STATIC int
  42 xlog_commit_record(
  43         struct xlog             *log,
  44         struct xlog_ticket      *ticket,
  45         struct xlog_in_core     **iclog,
  46         xfs_lsn_t               *commitlsnp);
  47
  48 STATIC struct xlog *
  49 xlog_alloc_log(
  50         struct xfs_mount        *mp,
  51         struct xfs_buftarg      *log_target,
  52         xfs_daddr_t             blk_offset,
  53         int                     num_bblks);
  54 STATIC int
  55 xlog_space_left(
  56         struct xlog             *log,
  57         atomic64_t              *head);
  58 STATIC int
  59 xlog_sync(
  60         struct xlog             *log,
  61         struct xlog_in_core     *iclog);
  62 STATIC void
  63 xlog_dealloc_log(
  64         struct xlog             *log);
  65
  66 /* local state machine functions */
  67 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  68 STATIC void
  69 xlog_state_do_callback(
  70         struct xlog             *log,
  71         int                     aborted,
  72         struct xlog_in_core     *iclog);
  73 STATIC int
  74 xlog_state_get_iclog_space(
  75         struct xlog             *log,
  76         int                     len,
  77         struct xlog_in_core     **iclog,
  78         struct xlog_ticket      *ticket,
  79         int                     *continued_write,
  80         int                     *logoffsetp);
  81 STATIC int
  82 xlog_state_release_iclog(
  83         struct xlog             *log,
  84         struct xlog_in_core     *iclog);
  85 STATIC void
  86 xlog_state_switch_iclogs(
  87         struct xlog             *log,
  88         struct xlog_in_core     *iclog,
  89         int                     eventual_size);
  90 STATIC void
  91 xlog_state_want_sync(
  92         struct xlog             *log,
  93         struct xlog_in_core     *iclog);
  94
  95 STATIC void
  96 xlog_grant_push_ail(
  97         struct xlog             *log,
  98         int                     need_bytes);
  99 STATIC void
 100 xlog_regrant_reserve_log_space(
 101         struct xlog             *log,
 102         struct xlog_ticket      *ticket);
 103 STATIC void
 104 xlog_ungrant_log_space(
 105         struct xlog             *log,
 106         struct xlog_ticket      *ticket);
 107
 108 #if defined(DEBUG)
 109 STATIC void
 110 xlog_verify_dest_ptr(
 111         struct xlog             *log,
 112         char                    *ptr);
 113 STATIC void
 114 xlog_verify_grant_tail(
 115         struct xlog *log);
 116 STATIC void
 117 xlog_verify_iclog(
 118         struct xlog             *log,
 119         struct xlog_in_core     *iclog,
 120         int                     count,
 121         boolean_t               syncing);
 122 STATIC void
 123 xlog_verify_tail_lsn(
 124         struct xlog             *log,
 125         struct xlog_in_core     *iclog,
 126         xfs_lsn_t               tail_lsn);
 127 #else
 128 #define xlog_verify_dest_ptr(a,b)
 129 #define xlog_verify_grant_tail(a)
 130 #define xlog_verify_iclog(a,b,c,d)
 131 #define xlog_verify_tail_lsn(a,b,c)
 132 #endif
 133
 134 STATIC int
 135 xlog_iclogs_empty(
 136         struct xlog             *log);
 137
 138 static void
 139 xlog_grant_sub_space(
 140         struct xlog             *log,
 141         atomic64_t              *head,
 142         int                     bytes)
 143 {
 144         int64_t head_val = atomic64_read(head);
 145         int64_t new, old;
 146
 147         do {
 148                 int     cycle, space;
 149
 150                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 151
 152                 space -= bytes;
 153                 if (space < 0) {
 154                         space += log->l_logsize;
 155                         cycle--;
 156                 }
 157
 158                 old = head_val;
 159                 new = xlog_assign_grant_head_val(cycle, space);
 160                 head_val = atomic64_cmpxchg(head, old, new);
 161         } while (head_val != old);
 162 }
 163
 164 static void
 165 xlog_grant_add_space(
 166         struct xlog             *log,
 167         atomic64_t              *head,
 168         int                     bytes)
 169 {
 170         int64_t head_val = atomic64_read(head);
 171         int64_t new, old;
 172
 173         do {
 174                 int             tmp;
 175                 int             cycle, space;
 176
 177                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 178
 179                 tmp = log->l_logsize - space;
 180                 if (tmp > bytes)
 181                         space += bytes;
 182                 else {
 183                         space = bytes - tmp;
 184                         cycle++;
 185                 }
 186
 187                 old = head_val;
 188                 new = xlog_assign_grant_head_val(cycle, space);
 189                 head_val = atomic64_cmpxchg(head, old, new);
 190         } while (head_val != old);
 191 }
 192
 193 STATIC void
 194 xlog_grant_head_init(
 195         struct xlog_grant_head  *head)
 196 {
 197         xlog_assign_grant_head(&head->grant, 1, 0);
 198         INIT_LIST_HEAD(&head->waiters);
 199         spin_lock_init(&head->lock);
 200 }
 201
 202 STATIC void
 203 xlog_grant_head_wake_all(
 204         struct xlog_grant_head  *head)
 205 {
 206         struct xlog_ticket      *tic;
 207
 208         spin_lock(&head->lock);
 209         list_for_each_entry(tic, &head->waiters, t_queue)
 210                 wake_up_process(tic->t_task);
 211         spin_unlock(&head->lock);
 212 }
 213
 214 static inline int
 215 xlog_ticket_reservation(
 216         struct xlog             *log,
 217         struct xlog_grant_head  *head,
 218         struct xlog_ticket      *tic)
 219 {
 220         if (head == &log->l_write_head) {
 221                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 222                 return tic->t_unit_res;
 223         } else {
 224                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 225                         return tic->t_unit_res * tic->t_cnt;
 226                 else
 227                         return tic->t_unit_res;
 228         }
 229 }
 230
 231 STATIC bool
 232 xlog_grant_head_wake(
 233         struct xlog             *log,
 234         struct xlog_grant_head  *head,
 235         int                     *free_bytes)
 236 {
 237         struct xlog_ticket      *tic;
 238         int                     need_bytes;
 239
 240         list_for_each_entry(tic, &head->waiters, t_queue) {
 241                 need_bytes = xlog_ticket_reservation(log, head, tic);
 242                 if (*free_bytes < need_bytes)
 243                         return false;
 244
 245                 *free_bytes -= need_bytes;
 246                 trace_xfs_log_grant_wake_up(log, tic);
 247                 wake_up_process(tic->t_task);
 248         }
 249
 250         return true;
 251 }
 252
 253 STATIC int
 254 xlog_grant_head_wait(
 255         struct xlog             *log,
 256         struct xlog_grant_head  *head,
 257         struct xlog_ticket      *tic,
 258         int                     need_bytes)
 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 XFS_ERROR(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 XFS_ERROR(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 XFS_ERROR(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 XFS_ERROR(ENOMEM);
 457
 458         tic->t_trans_type = t_type;
 459         *ticp = tic;
 460
 461         xlog_grant_push_ail(log, tic->t_unit_res * tic->t_cnt);
 462
 463         trace_xfs_log_reserve(log, tic);
 464
 465         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 466                                       &need_bytes);
 467         if (error)
 468                 goto out_error;
 469
 470         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 471         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 472         trace_xfs_log_reserve_exit(log, tic);
 473         xlog_verify_grant_tail(log);
 474         return 0;
 475
 476 out_error:
 477         /*
 478          * If we are failing, make sure the ticket doesn't have any current
 479          * reservations.  We don't want to add this back when the ticket/
 480          * transaction gets cancelled.
 481          */
 482         tic->t_curr_res = 0;
 483         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 484         return error;
 485 }
 486
 487
 488 /*
 489  * NOTES:
 490  *
 491  *      1. currblock field gets updated at startup and after in-core logs
 492  *              marked as with WANT_SYNC.
 493  */
 494
 495 /*
 496  * This routine is called when a user of a log manager ticket is done with
 497  * the reservation.  If the ticket was ever used, then a commit record for
 498  * the associated transaction is written out as a log operation header with
 499  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 500  * a given ticket.  If the ticket was one with a permanent reservation, then
 501  * a few operations are done differently.  Permanent reservation tickets by
 502  * default don't release the reservation.  They just commit the current
 503  * transaction with the belief that the reservation is still needed.  A flag
 504  * must be passed in before permanent reservations are actually released.
 505  * When these type of tickets are not released, they need to be set into
 506  * the inited state again.  By doing this, a start record will be written
 507  * out when the next write occurs.
 508  */
 509 xfs_lsn_t
 510 xfs_log_done(
 511         struct xfs_mount        *mp,
 512         struct xlog_ticket      *ticket,
 513         struct xlog_in_core     **iclog,
 514         uint                    flags)
 515 {
 516         struct xlog             *log = mp->m_log;
 517         xfs_lsn_t               lsn = 0;
 518
 519         if (XLOG_FORCED_SHUTDOWN(log) ||
 520             /*
 521              * If nothing was ever written, don't write out commit record.
 522              * If we get an error, just continue and give back the log ticket.
 523              */
 524             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 525              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 526                 lsn = (xfs_lsn_t) -1;
 527                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
 528                         flags |= XFS_LOG_REL_PERM_RESERV;
 529                 }
 530         }
 531
 532
 533         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
 534             (flags & XFS_LOG_REL_PERM_RESERV)) {
 535                 trace_xfs_log_done_nonperm(log, ticket);
 536
 537                 /*
 538                  * Release ticket if not permanent reservation or a specific
 539                  * request has been made to release a permanent reservation.
 540                  */
 541                 xlog_ungrant_log_space(log, ticket);
 542                 xfs_log_ticket_put(ticket);
 543         } else {
 544                 trace_xfs_log_done_perm(log, ticket);
 545
 546                 xlog_regrant_reserve_log_space(log, ticket);
 547                 /* If this ticket was a permanent reservation and we aren't
 548                  * trying to release it, reset the inited flags; so next time
 549                  * we write, a start record will be written out.
 550                  */
 551                 ticket->t_flags |= XLOG_TIC_INITED;
 552         }
 553
 554         return lsn;
 555 }
 556
 557 /*
 558  * Attaches a new iclog I/O completion callback routine during
 559  * transaction commit.  If the log is in error state, a non-zero
 560  * return code is handed back and the caller is responsible for
 561  * executing the callback at an appropriate time.
 562  */
 563 int
 564 xfs_log_notify(
 565         struct xfs_mount        *mp,
 566         struct xlog_in_core     *iclog,
 567         xfs_log_callback_t      *cb)
 568 {
 569         int     abortflg;
 570
 571         spin_lock(&iclog->ic_callback_lock);
 572         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 573         if (!abortflg) {
 574                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 575                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 576                 cb->cb_next = NULL;
 577                 *(iclog->ic_callback_tail) = cb;
 578                 iclog->ic_callback_tail = &(cb->cb_next);
 579         }
 580         spin_unlock(&iclog->ic_callback_lock);
 581         return abortflg;
 582 }
 583
 584 int
 585 xfs_log_release_iclog(
 586         struct xfs_mount        *mp,
 587         struct xlog_in_core     *iclog)
 588 {
 589         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 590                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 591                 return EIO;
 592         }
 593
 594         return 0;
 595 }
 596
 597 /*
 598  * Mount a log filesystem
 599  *
 600  * mp           - ubiquitous xfs mount point structure
 601  * log_target   - buftarg of on-disk log device
 602  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 603  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 604  *
 605  * Return error or zero.
 606  */
 607 int
 608 xfs_log_mount(
 609         xfs_mount_t     *mp,
 610         xfs_buftarg_t   *log_target,
 611         xfs_daddr_t     blk_offset,
 612         int             num_bblks)
 613 {
 614         int             error;
 615
 616         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
 617                 xfs_notice(mp, "Mounting Filesystem");
 618         else {
 619                 xfs_notice(mp,
 620 "Mounting filesystem in no-recovery mode.  Filesystem will be inconsistent.");
 621                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 622         }
 623
 624         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 625         if (IS_ERR(mp->m_log)) {
 626                 error = -PTR_ERR(mp->m_log);
 627                 goto out;
 628         }
 629
 630         /*
 631          * Initialize the AIL now we have a log.
 632          */
 633         error = xfs_trans_ail_init(mp);
 634         if (error) {
 635                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 636                 goto out_free_log;
 637         }
 638         mp->m_log->l_ailp = mp->m_ail;
 639
 640         /*
 641          * skip log recovery on a norecovery mount.  pretend it all
 642          * just worked.
 643          */
 644         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 645                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 646
 647                 if (readonly)
 648                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 649
 650                 error = xlog_recover(mp->m_log);
 651
 652                 if (readonly)
 653                         mp->m_flags |= XFS_MOUNT_RDONLY;
 654                 if (error) {
 655                         xfs_warn(mp, "log mount/recovery failed: error %d",
 656                                 error);
 657                         goto out_destroy_ail;
 658                 }
 659         }
 660
 661         /* Normal transactions can now occur */
 662         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 663
 664         /*
 665          * Now the log has been fully initialised and we know were our
 666          * space grant counters are, we can initialise the permanent ticket
 667          * needed for delayed logging to work.
 668          */
 669         xlog_cil_init_post_recovery(mp->m_log);
 670
 671         return 0;
 672
 673 out_destroy_ail:
 674         xfs_trans_ail_destroy(mp);
 675 out_free_log:
 676         xlog_dealloc_log(mp->m_log);
 677 out:
 678         return error;
 679 }
 680
 681 /*
 682  * Finish the recovery of the file system.  This is separate from
 683  * the xfs_log_mount() call, because it depends on the code in
 684  * xfs_mountfs() to read in the root and real-time bitmap inodes
 685  * between calling xfs_log_mount() and here.
 686  *
 687  * mp           - ubiquitous xfs mount point structure
 688  */
 689 int
 690 xfs_log_mount_finish(xfs_mount_t *mp)
 691 {
 692         int     error;
 693
 694         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
 695                 error = xlog_recover_finish(mp->m_log);
 696         else {
 697                 error = 0;
 698                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 699         }
 700
 701         return error;
 702 }
 703
 704 /*
 705  * Final log writes as part of unmount.
 706  *
 707  * Mark the filesystem clean as unmount happens.  Note that during relocation
 708  * this routine needs to be executed as part of source-bag while the
 709  * deallocation must not be done until source-end.
 710  */
 711
 712 /*
 713  * Unmount record used to have a string "Unmount filesystem--" in the
 714  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 715  * We just write the magic number now since that particular field isn't
 716  * currently architecture converted and "nUmount" is a bit foo.
 717  * As far as I know, there weren't any dependencies on the old behaviour.
 718  */
 719
 720 int
 721 xfs_log_unmount_write(xfs_mount_t *mp)
 722 {
 723         struct xlog      *log = mp->m_log;
 724         xlog_in_core_t   *iclog;
 725 #ifdef DEBUG
 726         xlog_in_core_t   *first_iclog;
 727 #endif
 728         xlog_ticket_t   *tic = NULL;
 729         xfs_lsn_t        lsn;
 730         int              error;
 731
 732         /*
 733          * Don't write out unmount record on read-only mounts.
 734          * Or, if we are doing a forced umount (typically because of IO errors).
 735          */
 736         if (mp->m_flags & XFS_MOUNT_RDONLY)
 737                 return 0;
 738
 739         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
 740         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 741
 742 #ifdef DEBUG
 743         first_iclog = iclog = log->l_iclog;
 744         do {
 745                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 746                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 747                         ASSERT(iclog->ic_offset == 0);
 748                 }
 749                 iclog = iclog->ic_next;
 750         } while (iclog != first_iclog);
 751 #endif
 752         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 753                 error = xfs_log_reserve(mp, 600, 1, &tic,
 754                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
 755                 if (!error) {
 756                         /* the data section must be 32 bit size aligned */
 757                         struct {
 758                             __uint16_t magic;
 759                             __uint16_t pad1;
 760                             __uint32_t pad2; /* may as well make it 64 bits */
 761                         } magic = {
 762                                 .magic = XLOG_UNMOUNT_TYPE,
 763                         };
 764                         struct xfs_log_iovec reg = {
 765                                 .i_addr = &magic,
 766                                 .i_len = sizeof(magic),
 767                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 768                         };
 769                         struct xfs_log_vec vec = {
 770                                 .lv_niovecs = 1,
 771                                 .lv_iovecp = &reg,
 772                         };
 773
 774                         /* remove inited flag, and account for space used */
 775                         tic->t_flags = 0;
 776                         tic->t_curr_res -= sizeof(magic);
 777                         error = xlog_write(log, &vec, tic, &lsn,
 778                                            NULL, XLOG_UNMOUNT_TRANS);
 779                         /*
 780                          * At this point, we're umounting anyway,
 781                          * so there's no point in transitioning log state
 782                          * to IOERROR. Just continue...
 783                          */
 784                 }
 785
 786                 if (error)
 787                         xfs_alert(mp, "%s: unmount record failed", __func__);
 788
 789
 790                 spin_lock(&log->l_icloglock);
 791                 iclog = log->l_iclog;
 792                 atomic_inc(&iclog->ic_refcnt);
 793                 xlog_state_want_sync(log, iclog);
 794                 spin_unlock(&log->l_icloglock);
 795                 error = xlog_state_release_iclog(log, iclog);
 796
 797                 spin_lock(&log->l_icloglock);
 798                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 799                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 800                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 801                                 xlog_wait(&iclog->ic_force_wait,
 802                                                         &log->l_icloglock);
 803                         } else {
 804                                 spin_unlock(&log->l_icloglock);
 805                         }
 806                 } else {
 807                         spin_unlock(&log->l_icloglock);
 808                 }
 809                 if (tic) {
 810                         trace_xfs_log_umount_write(log, tic);
 811                         xlog_ungrant_log_space(log, tic);
 812                         xfs_log_ticket_put(tic);
 813                 }
 814         } else {
 815                 /*
 816                  * We're already in forced_shutdown mode, couldn't
 817                  * even attempt to write out the unmount transaction.
 818                  *
 819                  * Go through the motions of sync'ing and releasing
 820                  * the iclog, even though no I/O will actually happen,
 821                  * we need to wait for other log I/Os that may already
 822                  * be in progress.  Do this as a separate section of
 823                  * code so we'll know if we ever get stuck here that
 824                  * we're in this odd situation of trying to unmount
 825                  * a file system that went into forced_shutdown as
 826                  * the result of an unmount..
 827                  */
 828                 spin_lock(&log->l_icloglock);
 829                 iclog = log->l_iclog;
 830                 atomic_inc(&iclog->ic_refcnt);
 831
 832                 xlog_state_want_sync(log, iclog);
 833                 spin_unlock(&log->l_icloglock);
 834                 error =  xlog_state_release_iclog(log, iclog);
 835
 836                 spin_lock(&log->l_icloglock);
 837
 838                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 839                         || iclog->ic_state == XLOG_STATE_DIRTY
 840                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 841
 842                                 xlog_wait(&iclog->ic_force_wait,
 843                                                         &log->l_icloglock);
 844                 } else {
 845                         spin_unlock(&log->l_icloglock);
 846                 }
 847         }
 848
 849         return error;
 850 }       /* xfs_log_unmount_write */
 851
 852 /*
 853  * Deallocate log structures for unmount/relocation.
 854  *
 855  * We need to stop the aild from running before we destroy
 856  * and deallocate the log as the aild references the log.
 857  */
 858 void
 859 xfs_log_unmount(xfs_mount_t *mp)
 860 {
 861         cancel_delayed_work_sync(&mp->m_sync_work);
 862         xfs_trans_ail_destroy(mp);
 863         xlog_dealloc_log(mp->m_log);
 864 }
 865
 866 void
 867 xfs_log_item_init(
 868         struct xfs_mount        *mp,
 869         struct xfs_log_item     *item,
 870         int                     type,
 871         const struct xfs_item_ops *ops)
 872 {
 873         item->li_mountp = mp;
 874         item->li_ailp = mp->m_ail;
 875         item->li_type = type;
 876         item->li_ops = ops;
 877         item->li_lv = NULL;
 878
 879         INIT_LIST_HEAD(&item->li_ail);
 880         INIT_LIST_HEAD(&item->li_cil);
 881 }
 882
 883 /*
 884  * Wake up processes waiting for log space after we have moved the log tail.
 885  */
 886 void
 887 xfs_log_space_wake(
 888         struct xfs_mount        *mp)
 889 {
 890         struct xlog             *log = mp->m_log;
 891         int                     free_bytes;
 892
 893         if (XLOG_FORCED_SHUTDOWN(log))
 894                 return;
 895
 896         if (!list_empty_careful(&log->l_write_head.waiters)) {
 897                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 898
 899                 spin_lock(&log->l_write_head.lock);
 900                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
 901                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
 902                 spin_unlock(&log->l_write_head.lock);
 903         }
 904
 905         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
 906                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 907
 908                 spin_lock(&log->l_reserve_head.lock);
 909                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
 910                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
 911                 spin_unlock(&log->l_reserve_head.lock);
 912         }
 913 }
 914
 915 /*
 916  * Determine if we have a transaction that has gone to disk
 917  * that needs to be covered. To begin the transition to the idle state
 918  * firstly the log needs to be idle (no AIL and nothing in the iclogs).
 919  * If we are then in a state where covering is needed, the caller is informed
 920  * that dummy transactions are required to move the log into the idle state.
 921  *
 922  * Because this is called as part of the sync process, we should also indicate
 923  * that dummy transactions should be issued in anything but the covered or
 924  * idle states. This ensures that the log tail is accurately reflected in
 925  * the log at the end of the sync, hence if a crash occurrs avoids replay
 926  * of transactions where the metadata is already on disk.
 927  */
 928 int
 929 xfs_log_need_covered(xfs_mount_t *mp)
 930 {
 931         int             needed = 0;
 932         struct xlog     *log = mp->m_log;
 933
 934         if (!xfs_fs_writable(mp))
 935                 return 0;
 936
 937         spin_lock(&log->l_icloglock);
 938         switch (log->l_covered_state) {
 939         case XLOG_STATE_COVER_DONE:
 940         case XLOG_STATE_COVER_DONE2:
 941         case XLOG_STATE_COVER_IDLE:
 942                 break;
 943         case XLOG_STATE_COVER_NEED:
 944         case XLOG_STATE_COVER_NEED2:
 945                 if (!xfs_ail_min_lsn(log->l_ailp) &&
 946                     xlog_iclogs_empty(log)) {
 947                         if (log->l_covered_state == XLOG_STATE_COVER_NEED)
 948                                 log->l_covered_state = XLOG_STATE_COVER_DONE;
 949                         else
 950                                 log->l_covered_state = XLOG_STATE_COVER_DONE2;
 951                 }
 952                 /* FALLTHRU */
 953         default:
 954                 needed = 1;
 955                 break;
 956         }
 957         spin_unlock(&log->l_icloglock);
 958         return needed;
 959 }
 960
 961 /*
 962  * We may be holding the log iclog lock upon entering this routine.
 963  */
 964 xfs_lsn_t
 965 xlog_assign_tail_lsn_locked(
 966         struct xfs_mount        *mp)
 967 {
 968         struct xlog             *log = mp->m_log;
 969         struct xfs_log_item     *lip;
 970         xfs_lsn_t               tail_lsn;
 971
 972         assert_spin_locked(&mp->m_ail->xa_lock);
 973
 974         /*
 975          * To make sure we always have a valid LSN for the log tail we keep
 976          * track of the last LSN which was committed in log->l_last_sync_lsn,
 977          * and use that when the AIL was empty.
 978          */
 979         lip = xfs_ail_min(mp->m_ail);
 980         if (lip)
 981                 tail_lsn = lip->li_lsn;
 982         else
 983                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
 984         atomic64_set(&log->l_tail_lsn, tail_lsn);
 985         return tail_lsn;
 986 }
 987
 988 xfs_lsn_t
 989 xlog_assign_tail_lsn(
 990         struct xfs_mount        *mp)
 991 {
 992         xfs_lsn_t               tail_lsn;
 993
 994         spin_lock(&mp->m_ail->xa_lock);
 995         tail_lsn = xlog_assign_tail_lsn_locked(mp);
 996         spin_unlock(&mp->m_ail->xa_lock);
 997
 998         return tail_lsn;
 999 }
1000
1001 /*
1002  * Return the space in the log between the tail and the head.  The head
1003  * is passed in the cycle/bytes formal parms.  In the special case where
1004  * the reserve head has wrapped passed the tail, this calculation is no
1005  * longer valid.  In this case, just return 0 which means there is no space
1006  * in the log.  This works for all places where this function is called
1007  * with the reserve head.  Of course, if the write head were to ever
1008  * wrap the tail, we should blow up.  Rather than catch this case here,
1009  * we depend on other ASSERTions in other parts of the code.   XXXmiken
1010  *
1011  * This code also handles the case where the reservation head is behind
1012  * the tail.  The details of this case are described below, but the end
1013  * result is that we return the size of the log as the amount of space left.
1014  */
1015 STATIC int
1016 xlog_space_left(
1017         struct xlog     *log,
1018         atomic64_t      *head)
1019 {
1020         int             free_bytes;
1021         int             tail_bytes;
1022         int             tail_cycle;
1023         int             head_cycle;
1024         int             head_bytes;
1025
1026         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
1027         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
1028         tail_bytes = BBTOB(tail_bytes);
1029         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
1030                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
1031         else if (tail_cycle + 1 < head_cycle)
1032                 return 0;
1033         else if (tail_cycle < head_cycle) {
1034                 ASSERT(tail_cycle == (head_cycle - 1));
1035                 free_bytes = tail_bytes - head_bytes;
1036         } else {
1037                 /*
1038                  * The reservation head is behind the tail.
1039                  * In this case we just want to return the size of the
1040                  * log as the amount of space left.
1041                  */
1042                 xfs_alert(log->l_mp,
1043                         "xlog_space_left: head behind tail\n"
1044                         "  tail_cycle = %d, tail_bytes = %d\n"
1045                         "  GH   cycle = %d, GH   bytes = %d",
1046                         tail_cycle, tail_bytes, head_cycle, head_bytes);
1047                 ASSERT(0);
1048                 free_bytes = log->l_logsize;
1049         }
1050         return free_bytes;
1051 }
1052
1053
1054 /*
1055  * Log function which is called when an io completes.
1056  *
1057  * The log manager needs its own routine, in order to control what
1058  * happens with the buffer after the write completes.
1059  */
1060 void
1061 xlog_iodone(xfs_buf_t *bp)
1062 {
1063         struct xlog_in_core     *iclog = bp->b_fspriv;
1064         struct xlog             *l = iclog->ic_log;
1065         int                     aborted = 0;
1066
1067         /*
1068          * Race to shutdown the filesystem if we see an error.
1069          */
1070         if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1071                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1072                 xfs_buf_ioerror_alert(bp, __func__);
1073                 xfs_buf_stale(bp);
1074                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1075                 /*
1076                  * This flag will be propagated to the trans-committed
1077                  * callback routines to let them know that the log-commit
1078                  * didn't succeed.
1079                  */
1080                 aborted = XFS_LI_ABORTED;
1081         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1082                 aborted = XFS_LI_ABORTED;
1083         }
1084
1085         /* log I/O is always issued ASYNC */
1086         ASSERT(XFS_BUF_ISASYNC(bp));
1087         xlog_state_done_syncing(iclog, aborted);
1088         /*
1089          * do not reference the buffer (bp) here as we could race
1090          * with it being freed after writing the unmount record to the
1091          * log.
1092          */
1093
1094 }       /* xlog_iodone */
1095
1096 /*
1097  * Return size of each in-core log record buffer.
1098  *
1099  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1100  *
1101  * If the filesystem blocksize is too large, we may need to choose a
1102  * larger size since the directory code currently logs entire blocks.
1103  */
1104
1105 STATIC void
1106 xlog_get_iclog_buffer_size(
1107         struct xfs_mount        *mp,
1108         struct xlog             *log)
1109 {
1110         int size;
1111         int xhdrs;
1112
1113         if (mp->m_logbufs <= 0)
1114                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1115         else
1116                 log->l_iclog_bufs = mp->m_logbufs;
1117
1118         /*
1119          * Buffer size passed in from mount system call.
1120          */
1121         if (mp->m_logbsize > 0) {
1122                 size = log->l_iclog_size = mp->m_logbsize;
1123                 log->l_iclog_size_log = 0;
1124                 while (size != 1) {
1125                         log->l_iclog_size_log++;
1126                         size >>= 1;
1127                 }
1128
1129                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1130                         /* # headers = size / 32k
1131                          * one header holds cycles from 32k of data
1132                          */
1133
1134                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1135                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1136                                 xhdrs++;
1137                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1138                         log->l_iclog_heads = xhdrs;
1139                 } else {
1140                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1141                         log->l_iclog_hsize = BBSIZE;
1142                         log->l_iclog_heads = 1;
1143                 }
1144                 goto done;
1145         }
1146
1147         /* All machines use 32kB buffers by default. */
1148         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1149         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1150
1151         /* the default log size is 16k or 32k which is one header sector */
1152         log->l_iclog_hsize = BBSIZE;
1153         log->l_iclog_heads = 1;
1154
1155 done:
1156         /* are we being asked to make the sizes selected above visible? */
1157         if (mp->m_logbufs == 0)
1158                 mp->m_logbufs = log->l_iclog_bufs;
1159         if (mp->m_logbsize == 0)
1160                 mp->m_logbsize = log->l_iclog_size;
1161 }       /* xlog_get_iclog_buffer_size */
1162
1163
1164 /*
1165  * This routine initializes some of the log structure for a given mount point.
1166  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1167  * some other stuff may be filled in too.
1168  */
1169 STATIC struct xlog *
1170 xlog_alloc_log(
1171         struct xfs_mount        *mp,
1172         struct xfs_buftarg      *log_target,
1173         xfs_daddr_t             blk_offset,
1174         int                     num_bblks)
1175 {
1176         struct xlog             *log;
1177         xlog_rec_header_t       *head;
1178         xlog_in_core_t          **iclogp;
1179         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1180         xfs_buf_t               *bp;
1181         int                     i;
1182         int                     error = ENOMEM;
1183         uint                    log2_size = 0;
1184
1185         log = kmem_zalloc(sizeof(struct xlog), KM_MAYFAIL);
1186         if (!log) {
1187                 xfs_warn(mp, "Log allocation failed: No memory!");
1188                 goto out;
1189         }
1190
1191         log->l_mp          = mp;
1192         log->l_targ        = log_target;
1193         log->l_logsize     = BBTOB(num_bblks);
1194         log->l_logBBstart  = blk_offset;
1195         log->l_logBBsize   = num_bblks;
1196         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1197         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1198
1199         log->l_prev_block  = -1;
1200         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1201         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1202         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1203         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1204
1205         xlog_grant_head_init(&log->l_reserve_head);
1206         xlog_grant_head_init(&log->l_write_head);
1207
1208         error = EFSCORRUPTED;
1209         if (xfs_sb_version_hassector(&mp->m_sb)) {
1210                 log2_size = mp->m_sb.sb_logsectlog;
1211                 if (log2_size < BBSHIFT) {
1212                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1213                                 log2_size, BBSHIFT);
1214                         goto out_free_log;
1215                 }
1216
1217                 log2_size -= BBSHIFT;
1218                 if (log2_size > mp->m_sectbb_log) {
1219                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1220                                 log2_size, mp->m_sectbb_log);
1221                         goto out_free_log;
1222                 }
1223
1224                 /* for larger sector sizes, must have v2 or external log */
1225                 if (log2_size && log->l_logBBstart > 0 &&
1226                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1227                         xfs_warn(mp,
1228                 "log sector size (0x%x) invalid for configuration.",
1229                                 log2_size);
1230                         goto out_free_log;
1231                 }
1232         }
1233         log->l_sectBBsize = 1 << log2_size;
1234
1235         xlog_get_iclog_buffer_size(mp, log);
1236
1237         error = ENOMEM;
1238         bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1239         if (!bp)
1240                 goto out_free_log;
1241         bp->b_iodone = xlog_iodone;
1242         ASSERT(xfs_buf_islocked(bp));
1243         log->l_xbuf = bp;
1244
1245         spin_lock_init(&log->l_icloglock);
1246         init_waitqueue_head(&log->l_flush_wait);
1247
1248         iclogp = &log->l_iclog;
1249         /*
1250          * The amount of memory to allocate for the iclog structure is
1251          * rather funky due to the way the structure is defined.  It is
1252          * done this way so that we can use different sizes for machines
1253          * with different amounts of memory.  See the definition of
1254          * xlog_in_core_t in xfs_log_priv.h for details.
1255          */
1256         ASSERT(log->l_iclog_size >= 4096);
1257         for (i=0; i < log->l_iclog_bufs; i++) {
1258                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1259                 if (!*iclogp)
1260                         goto out_free_iclog;
1261
1262                 iclog = *iclogp;
1263                 iclog->ic_prev = prev_iclog;
1264                 prev_iclog = iclog;
1265
1266                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1267                                                 BTOBB(log->l_iclog_size), 0);
1268                 if (!bp)
1269                         goto out_free_iclog;
1270
1271                 bp->b_iodone = xlog_iodone;
1272                 iclog->ic_bp = bp;
1273                 iclog->ic_data = bp->b_addr;
1274 #ifdef DEBUG
1275                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1276 #endif
1277                 head = &iclog->ic_header;
1278                 memset(head, 0, sizeof(xlog_rec_header_t));
1279                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1280                 head->h_version = cpu_to_be32(
1281                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1282                 head->h_size = cpu_to_be32(log->l_iclog_size);
1283                 /* new fields */
1284                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1285                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1286
1287                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1288                 iclog->ic_state = XLOG_STATE_ACTIVE;
1289                 iclog->ic_log = log;
1290                 atomic_set(&iclog->ic_refcnt, 0);
1291                 spin_lock_init(&iclog->ic_callback_lock);
1292                 iclog->ic_callback_tail = &(iclog->ic_callback);
1293                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1294
1295                 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1296                 init_waitqueue_head(&iclog->ic_force_wait);
1297                 init_waitqueue_head(&iclog->ic_write_wait);
1298
1299                 iclogp = &iclog->ic_next;
1300         }
1301         *iclogp = log->l_iclog;                 /* complete ring */
1302         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1303
1304         error = xlog_cil_init(log);
1305         if (error)
1306                 goto out_free_iclog;
1307         return log;
1308
1309 out_free_iclog:
1310         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1311                 prev_iclog = iclog->ic_next;
1312                 if (iclog->ic_bp)
1313                         xfs_buf_free(iclog->ic_bp);
1314                 kmem_free(iclog);
1315         }
1316         spinlock_destroy(&log->l_icloglock);
1317         xfs_buf_free(log->l_xbuf);
1318 out_free_log:
1319         kmem_free(log);
1320 out:
1321         return ERR_PTR(-error);
1322 }       /* xlog_alloc_log */
1323
1324
1325 /*
1326  * Write out the commit record of a transaction associated with the given
1327  * ticket.  Return the lsn of the commit record.
1328  */
1329 STATIC int
1330 xlog_commit_record(
1331         struct xlog             *log,
1332         struct xlog_ticket      *ticket,
1333         struct xlog_in_core     **iclog,
1334         xfs_lsn_t               *commitlsnp)
1335 {
1336         struct xfs_mount *mp = log->l_mp;
1337         int     error;
1338         struct xfs_log_iovec reg = {
1339                 .i_addr = NULL,
1340                 .i_len = 0,
1341                 .i_type = XLOG_REG_TYPE_COMMIT,
1342         };
1343         struct xfs_log_vec vec = {
1344                 .lv_niovecs = 1,
1345                 .lv_iovecp = &reg,
1346         };
1347
1348         ASSERT_ALWAYS(iclog);
1349         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1350                                         XLOG_COMMIT_TRANS);
1351         if (error)
1352                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1353         return error;
1354 }
1355
1356 /*
1357  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1358  * log space.  This code pushes on the lsn which would supposedly free up
1359  * the 25% which we want to leave free.  We may need to adopt a policy which
1360  * pushes on an lsn which is further along in the log once we reach the high
1361  * water mark.  In this manner, we would be creating a low water mark.
1362  */
1363 STATIC void
1364 xlog_grant_push_ail(
1365         struct xlog     *log,
1366         int             need_bytes)
1367 {
1368         xfs_lsn_t       threshold_lsn = 0;
1369         xfs_lsn_t       last_sync_lsn;
1370         int             free_blocks;
1371         int             free_bytes;
1372         int             threshold_block;
1373         int             threshold_cycle;
1374         int             free_threshold;
1375
1376         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1377
1378         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1379         free_blocks = BTOBBT(free_bytes);
1380
1381         /*
1382          * Set the threshold for the minimum number of free blocks in the
1383          * log to the maximum of what the caller needs, one quarter of the
1384          * log, and 256 blocks.
1385          */
1386         free_threshold = BTOBB(need_bytes);
1387         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1388         free_threshold = MAX(free_threshold, 256);
1389         if (free_blocks >= free_threshold)
1390                 return;
1391
1392         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1393                                                 &threshold_block);
1394         threshold_block += free_threshold;
1395         if (threshold_block >= log->l_logBBsize) {
1396                 threshold_block -= log->l_logBBsize;
1397                 threshold_cycle += 1;
1398         }
1399         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1400                                         threshold_block);
1401         /*
1402          * Don't pass in an lsn greater than the lsn of the last
1403          * log record known to be on disk. Use a snapshot of the last sync lsn
1404          * so that it doesn't change between the compare and the set.
1405          */
1406         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1407         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1408                 threshold_lsn = last_sync_lsn;
1409
1410         /*
1411          * Get the transaction layer to kick the dirty buffers out to
1412          * disk asynchronously. No point in trying to do this if
1413          * the filesystem is shutting down.
1414          */
1415         if (!XLOG_FORCED_SHUTDOWN(log))
1416                 xfs_ail_push(log->l_ailp, threshold_lsn);
1417 }
1418
1419 /*
1420  * The bdstrat callback function for log bufs. This gives us a central
1421  * place to trap bufs in case we get hit by a log I/O error and need to
1422  * shutdown. Actually, in practice, even when we didn't get a log error,
1423  * we transition the iclogs to IOERROR state *after* flushing all existing
1424  * iclogs to disk. This is because we don't want anymore new transactions to be
1425  * started or completed afterwards.
1426  */
1427 STATIC int
1428 xlog_bdstrat(
1429         struct xfs_buf          *bp)
1430 {
1431         struct xlog_in_core     *iclog = bp->b_fspriv;
1432
1433         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1434                 xfs_buf_ioerror(bp, EIO);
1435                 xfs_buf_stale(bp);
1436                 xfs_buf_ioend(bp, 0);
1437                 /*
1438                  * It would seem logical to return EIO here, but we rely on
1439                  * the log state machine to propagate I/O errors instead of
1440                  * doing it here.
1441                  */
1442                 return 0;
1443         }
1444
1445         xfs_buf_iorequest(bp);
1446         return 0;
1447 }
1448
1449 /*
1450  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1451  * fashion.  Previously, we should have moved the current iclog
1452  * ptr in the log to point to the next available iclog.  This allows further
1453  * write to continue while this code syncs out an iclog ready to go.
1454  * Before an in-core log can be written out, the data section must be scanned
1455  * to save away the 1st word of each BBSIZE block into the header.  We replace
1456  * it with the current cycle count.  Each BBSIZE block is tagged with the
1457  * cycle count because there in an implicit assumption that drives will
1458  * guarantee that entire 512 byte blocks get written at once.  In other words,
1459  * we can't have part of a 512 byte block written and part not written.  By
1460  * tagging each block, we will know which blocks are valid when recovering
1461  * after an unclean shutdown.
1462  *
1463  * This routine is single threaded on the iclog.  No other thread can be in
1464  * this routine with the same iclog.  Changing contents of iclog can there-
1465  * fore be done without grabbing the state machine lock.  Updating the global
1466  * log will require grabbing the lock though.
1467  *
1468  * The entire log manager uses a logical block numbering scheme.  Only
1469  * log_sync (and then only bwrite()) know about the fact that the log may
1470  * not start with block zero on a given device.  The log block start offset
1471  * is added immediately before calling bwrite().
1472  */
1473
1474 STATIC int
1475 xlog_sync(
1476         struct xlog             *log,
1477         struct xlog_in_core     *iclog)
1478 {
1479         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1480         xfs_buf_t       *bp;
1481         int             i;
1482         uint            count;          /* byte count of bwrite */
1483         uint            count_init;     /* initial count before roundup */
1484         int             roundoff;       /* roundoff to BB or stripe */
1485         int             split = 0;      /* split write into two regions */
1486         int             error;
1487         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1488
1489         XFS_STATS_INC(xs_log_writes);
1490         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1491
1492         /* Add for LR header */
1493         count_init = log->l_iclog_hsize + iclog->ic_offset;
1494
1495         /* Round out the log write size */
1496         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1497                 /* we have a v2 stripe unit to use */
1498                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1499         } else {
1500                 count = BBTOB(BTOBB(count_init));
1501         }
1502         roundoff = count - count_init;
1503         ASSERT(roundoff >= 0);
1504         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1505                 roundoff < log->l_mp->m_sb.sb_logsunit)
1506                 ||
1507                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1508                  roundoff < BBTOB(1)));
1509
1510         /* move grant heads by roundoff in sync */
1511         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1512         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1513
1514         /* put cycle number in every block */
1515         xlog_pack_data(log, iclog, roundoff);
1516
1517         /* real byte length */
1518         if (v2) {
1519                 iclog->ic_header.h_len =
1520                         cpu_to_be32(iclog->ic_offset + roundoff);
1521         } else {
1522                 iclog->ic_header.h_len =
1523                         cpu_to_be32(iclog->ic_offset);
1524         }
1525
1526         bp = iclog->ic_bp;
1527         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1528
1529         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1530
1531         /* Do we need to split this write into 2 parts? */
1532         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1533                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1534                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1535                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1536         } else {
1537                 iclog->ic_bwritecnt = 1;
1538         }
1539         bp->b_io_length = BTOBB(count);
1540         bp->b_fspriv = iclog;
1541         XFS_BUF_ZEROFLAGS(bp);
1542         XFS_BUF_ASYNC(bp);
1543         bp->b_flags |= XBF_SYNCIO;
1544
1545         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1546                 bp->b_flags |= XBF_FUA;
1547
1548                 /*
1549                  * Flush the data device before flushing the log to make
1550                  * sure all meta data written back from the AIL actually made
1551                  * it to disk before stamping the new log tail LSN into the
1552                  * log buffer.  For an external log we need to issue the
1553                  * flush explicitly, and unfortunately synchronously here;
1554                  * for an internal log we can simply use the block layer
1555                  * state machine for preflushes.
1556                  */
1557                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1558                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1559                 else
1560                         bp->b_flags |= XBF_FLUSH;
1561         }
1562
1563         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1564         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1565
1566         xlog_verify_iclog(log, iclog, count, B_TRUE);
1567
1568         /* account for log which doesn't start at block #0 */
1569         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1570         /*
1571          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1572          * is shutting down.
1573          */
1574         XFS_BUF_WRITE(bp);
1575
1576         error = xlog_bdstrat(bp);
1577         if (error) {
1578                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1579                 return error;
1580         }
1581         if (split) {
1582                 bp = iclog->ic_log->l_xbuf;
1583                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1584                 xfs_buf_associate_memory(bp,
1585                                 (char *)&iclog->ic_header + count, split);
1586                 bp->b_fspriv = iclog;
1587                 XFS_BUF_ZEROFLAGS(bp);
1588                 XFS_BUF_ASYNC(bp);
1589                 bp->b_flags |= XBF_SYNCIO;
1590                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1591                         bp->b_flags |= XBF_FUA;
1592                 dptr = bp->b_addr;
1593                 /*
1594                  * Bump the cycle numbers at the start of each block
1595                  * since this part of the buffer is at the start of
1596                  * a new cycle.  Watch out for the header magic number
1597                  * case, though.
1598                  */
1599                 for (i = 0; i < split; i += BBSIZE) {
1600                         be32_add_cpu((__be32 *)dptr, 1);
1601                         if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1602                                 be32_add_cpu((__be32 *)dptr, 1);
1603                         dptr += BBSIZE;
1604                 }
1605
1606                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1607                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1608
1609                 /* account for internal log which doesn't start at block #0 */
1610                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1611                 XFS_BUF_WRITE(bp);
1612                 error = xlog_bdstrat(bp);
1613                 if (error) {
1614                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1615                         return error;
1616                 }
1617         }
1618         return 0;
1619 }       /* xlog_sync */
1620
1621
1622 /*
1623  * Deallocate a log structure
1624  */
1625 STATIC void
1626 xlog_dealloc_log(
1627         struct xlog     *log)
1628 {
1629         xlog_in_core_t  *iclog, *next_iclog;
1630         int             i;
1631
1632         xlog_cil_destroy(log);
1633
1634         /*
1635          * always need to ensure that the extra buffer does not point to memory
1636          * owned by another log buffer before we free it.
1637          */
1638         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1639         xfs_buf_free(log->l_xbuf);
1640
1641         iclog = log->l_iclog;
1642         for (i=0; i<log->l_iclog_bufs; i++) {
1643                 xfs_buf_free(iclog->ic_bp);
1644                 next_iclog = iclog->ic_next;
1645                 kmem_free(iclog);
1646                 iclog = next_iclog;
1647         }
1648         spinlock_destroy(&log->l_icloglock);
1649
1650         log->l_mp->m_log = NULL;
1651         kmem_free(log);
1652 }       /* xlog_dealloc_log */
1653
1654 /*
1655  * Update counters atomically now that memcpy is done.
1656  */
1657 /* ARGSUSED */
1658 static inline void
1659 xlog_state_finish_copy(
1660         struct xlog             *log,
1661         struct xlog_in_core     *iclog,
1662         int                     record_cnt,
1663         int                     copy_bytes)
1664 {
1665         spin_lock(&log->l_icloglock);
1666
1667         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1668         iclog->ic_offset += copy_bytes;
1669
1670         spin_unlock(&log->l_icloglock);
1671 }       /* xlog_state_finish_copy */
1672
1673
1674
1675
1676 /*
1677  * print out info relating to regions written which consume
1678  * the reservation
1679  */
1680 void
1681 xlog_print_tic_res(
1682         struct xfs_mount        *mp,
1683         struct xlog_ticket      *ticket)
1684 {
1685         uint i;
1686         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1687
1688         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1689         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1690             "bformat",
1691             "bchunk",
1692             "efi_format",
1693             "efd_format",
1694             "iformat",
1695             "icore",
1696             "iext",
1697             "ibroot",
1698             "ilocal",
1699             "iattr_ext",
1700             "iattr_broot",
1701             "iattr_local",
1702             "qformat",
1703             "dquot",
1704             "quotaoff",
1705             "LR header",
1706             "unmount",
1707             "commit",
1708             "trans header"
1709         };
1710         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1711             "SETATTR_NOT_SIZE",
1712             "SETATTR_SIZE",
1713             "INACTIVE",
1714             "CREATE",
1715             "CREATE_TRUNC",
1716             "TRUNCATE_FILE",
1717             "REMOVE",
1718             "LINK",
1719             "RENAME",
1720             "MKDIR",
1721             "RMDIR",
1722             "SYMLINK",
1723             "SET_DMATTRS",
1724             "GROWFS",
1725             "STRAT_WRITE",
1726             "DIOSTRAT",
1727             "WRITE_SYNC",
1728             "WRITEID",
1729             "ADDAFORK",
1730             "ATTRINVAL",
1731             "ATRUNCATE",
1732             "ATTR_SET",
1733             "ATTR_RM",
1734             "ATTR_FLAG",
1735             "CLEAR_AGI_BUCKET",
1736             "QM_SBCHANGE",
1737             "DUMMY1",
1738             "DUMMY2",
1739             "QM_QUOTAOFF",
1740             "QM_DQALLOC",
1741             "QM_SETQLIM",
1742             "QM_DQCLUSTER",
1743             "QM_QINOCREATE",
1744             "QM_QUOTAOFF_END",
1745             "SB_UNIT",
1746             "FSYNC_TS",
1747             "GROWFSRT_ALLOC",
1748             "GROWFSRT_ZERO",
1749             "GROWFSRT_FREE",
1750             "SWAPEXT"
1751         };
1752
1753         xfs_warn(mp,
1754                 "xlog_write: reservation summary:\n"
1755                 "  trans type  = %s (%u)\n"
1756                 "  unit res    = %d bytes\n"
1757                 "  current res = %d bytes\n"
1758                 "  total reg   = %u bytes (o/flow = %u bytes)\n"
1759                 "  ophdrs      = %u (ophdr space = %u bytes)\n"
1760                 "  ophdr + reg = %u bytes\n"
1761                 "  num regions = %u\n",
1762                 ((ticket->t_trans_type <= 0 ||
1763                   ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1764                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1765                 ticket->t_trans_type,
1766                 ticket->t_unit_res,
1767                 ticket->t_curr_res,
1768                 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1769                 ticket->t_res_num_ophdrs, ophdr_spc,
1770                 ticket->t_res_arr_sum +
1771                 ticket->t_res_o_flow + ophdr_spc,
1772                 ticket->t_res_num);
1773
1774         for (i = 0; i < ticket->t_res_num; i++) {
1775                 uint r_type = ticket->t_res_arr[i].r_type;
1776                 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1777                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1778                             "bad-rtype" : res_type_str[r_type-1]),
1779                             ticket->t_res_arr[i].r_len);
1780         }
1781
1782         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1783                 "xlog_write: reservation ran out. Need to up reservation");
1784         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1785 }
1786
1787 /*
1788  * Calculate the potential space needed by the log vector.  Each region gets
1789  * its own xlog_op_header_t and may need to be double word aligned.
1790  */
1791 static int
1792 xlog_write_calc_vec_length(
1793         struct xlog_ticket      *ticket,
1794         struct xfs_log_vec      *log_vector)
1795 {
1796         struct xfs_log_vec      *lv;
1797         int                     headers = 0;
1798         int                     len = 0;
1799         int                     i;
1800
1801         /* acct for start rec of xact */
1802         if (ticket->t_flags & XLOG_TIC_INITED)
1803                 headers++;
1804
1805         for (lv = log_vector; lv; lv = lv->lv_next) {
1806                 headers += lv->lv_niovecs;
1807
1808                 for (i = 0; i < lv->lv_niovecs; i++) {
1809                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
1810
1811                         len += vecp->i_len;
1812                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1813                 }
1814         }
1815
1816         ticket->t_res_num_ophdrs += headers;
1817         len += headers * sizeof(struct xlog_op_header);
1818
1819         return len;
1820 }
1821
1822 /*
1823  * If first write for transaction, insert start record  We can't be trying to
1824  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
1825  */
1826 static int
1827 xlog_write_start_rec(
1828         struct xlog_op_header   *ophdr,
1829         struct xlog_ticket      *ticket)
1830 {
1831         if (!(ticket->t_flags & XLOG_TIC_INITED))
1832                 return 0;
1833
1834         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
1835         ophdr->oh_clientid = ticket->t_clientid;
1836         ophdr->oh_len = 0;
1837         ophdr->oh_flags = XLOG_START_TRANS;
1838         ophdr->oh_res2 = 0;
1839
1840         ticket->t_flags &= ~XLOG_TIC_INITED;
1841
1842         return sizeof(struct xlog_op_header);
1843 }
1844
1845 static xlog_op_header_t *
1846 xlog_write_setup_ophdr(
1847         struct xlog             *log,
1848         struct xlog_op_header   *ophdr,
1849         struct xlog_ticket      *ticket,
1850         uint                    flags)
1851 {
1852         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1853         ophdr->oh_clientid = ticket->t_clientid;
1854         ophdr->oh_res2 = 0;
1855
1856         /* are we copying a commit or unmount record? */
1857         ophdr->oh_flags = flags;
1858
1859         /*
1860          * We've seen logs corrupted with bad transaction client ids.  This
1861          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
1862          * and shut down the filesystem.
1863          */
1864         switch (ophdr->oh_clientid)  {
1865         case XFS_TRANSACTION:
1866         case XFS_VOLUME:
1867         case XFS_LOG:
1868                 break;
1869         default:
1870                 xfs_warn(log->l_mp,
1871                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1872                         ophdr->oh_clientid, ticket);
1873                 return NULL;
1874         }
1875
1876         return ophdr;
1877 }
1878
1879 /*
1880  * Set up the parameters of the region copy into the log. This has
1881  * to handle region write split across multiple log buffers - this
1882  * state is kept external to this function so that this code can
1883  * can be written in an obvious, self documenting manner.
1884  */
1885 static int
1886 xlog_write_setup_copy(
1887         struct xlog_ticket      *ticket,
1888         struct xlog_op_header   *ophdr,
1889         int                     space_available,
1890         int                     space_required,
1891         int                     *copy_off,
1892         int                     *copy_len,
1893         int                     *last_was_partial_copy,
1894         int                     *bytes_consumed)
1895 {
1896         int                     still_to_copy;
1897
1898         still_to_copy = space_required - *bytes_consumed;
1899         *copy_off = *bytes_consumed;
1900
1901         if (still_to_copy <= space_available) {
1902                 /* write of region completes here */
1903                 *copy_len = still_to_copy;
1904                 ophdr->oh_len = cpu_to_be32(*copy_len);
1905                 if (*last_was_partial_copy)
1906                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1907                 *last_was_partial_copy = 0;
1908                 *bytes_consumed = 0;
1909                 return 0;
1910         }
1911
1912         /* partial write of region, needs extra log op header reservation */
1913         *copy_len = space_available;
1914         ophdr->oh_len = cpu_to_be32(*copy_len);
1915         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1916         if (*last_was_partial_copy)
1917                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1918         *bytes_consumed += *copy_len;
1919         (*last_was_partial_copy)++;
1920
1921         /* account for new log op header */
1922         ticket->t_curr_res -= sizeof(struct xlog_op_header);
1923         ticket->t_res_num_ophdrs++;
1924
1925         return sizeof(struct xlog_op_header);
1926 }
1927
1928 static int
1929 xlog_write_copy_finish(
1930         struct xlog             *log,
1931         struct xlog_in_core     *iclog,
1932         uint                    flags,
1933         int                     *record_cnt,
1934         int                     *data_cnt,
1935         int                     *partial_copy,
1936         int                     *partial_copy_len,
1937         int                     log_offset,
1938         struct xlog_in_core     **commit_iclog)
1939 {
1940         if (*partial_copy) {
1941                 /*
1942                  * This iclog has already been marked WANT_SYNC by
1943                  * xlog_state_get_iclog_space.
1944                  */
1945                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1946                 *record_cnt = 0;
1947                 *data_cnt = 0;
1948                 return xlog_state_release_iclog(log, iclog);
1949         }
1950
1951         *partial_copy = 0;
1952         *partial_copy_len = 0;
1953
1954         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1955                 /* no more space in this iclog - push it. */
1956                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1957                 *record_cnt = 0;
1958                 *data_cnt = 0;
1959
1960                 spin_lock(&log->l_icloglock);
1961                 xlog_state_want_sync(log, iclog);
1962                 spin_unlock(&log->l_icloglock);
1963
1964                 if (!commit_iclog)
1965                         return xlog_state_release_iclog(log, iclog);
1966                 ASSERT(flags & XLOG_COMMIT_TRANS);
1967                 *commit_iclog = iclog;
1968         }
1969
1970         return 0;
1971 }
1972
1973 /*
1974  * Write some region out to in-core log
1975  *
1976  * This will be called when writing externally provided regions or when
1977  * writing out a commit record for a given transaction.
1978  *
1979  * General algorithm:
1980  *      1. Find total length of this write.  This may include adding to the
1981  *              lengths passed in.
1982  *      2. Check whether we violate the tickets reservation.
1983  *      3. While writing to this iclog
1984  *          A. Reserve as much space in this iclog as can get
1985  *          B. If this is first write, save away start lsn
1986  *          C. While writing this region:
1987  *              1. If first write of transaction, write start record
1988  *              2. Write log operation header (header per region)
1989  *              3. Find out if we can fit entire region into this iclog
1990  *              4. Potentially, verify destination memcpy ptr
1991  *              5. Memcpy (partial) region
1992  *              6. If partial copy, release iclog; otherwise, continue
1993  *                      copying more regions into current iclog
1994  *      4. Mark want sync bit (in simulation mode)
1995  *      5. Release iclog for potential flush to on-disk log.
1996  *
1997  * ERRORS:
1998  * 1.   Panic if reservation is overrun.  This should never happen since
1999  *      reservation amounts are generated internal to the filesystem.
2000  * NOTES:
2001  * 1. Tickets are single threaded data structures.
2002  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
2003  *      syncing routine.  When a single log_write region needs to span
2004  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
2005  *      on all log operation writes which don't contain the end of the
2006  *      region.  The XLOG_END_TRANS bit is used for the in-core log
2007  *      operation which contains the end of the continued log_write region.
2008  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
2009  *      we don't really know exactly how much space will be used.  As a result,
2010  *      we don't update ic_offset until the end when we know exactly how many
2011  *      bytes have been written out.
2012  */
2013 int
2014 xlog_write(
2015         struct xlog             *log,
2016         struct xfs_log_vec      *log_vector,
2017         struct xlog_ticket      *ticket,
2018         xfs_lsn_t               *start_lsn,
2019         struct xlog_in_core     **commit_iclog,
2020         uint                    flags)
2021 {
2022         struct xlog_in_core     *iclog = NULL;
2023         struct xfs_log_iovec    *vecp;
2024         struct xfs_log_vec      *lv;
2025         int                     len;
2026         int                     index;
2027         int                     partial_copy = 0;
2028         int                     partial_copy_len = 0;
2029         int                     contwr = 0;
2030         int                     record_cnt = 0;
2031         int                     data_cnt = 0;
2032         int                     error;
2033
2034         *start_lsn = 0;
2035
2036         len = xlog_write_calc_vec_length(ticket, log_vector);
2037
2038         /*
2039          * Region headers and bytes are already accounted for.
2040          * We only need to take into account start records and
2041          * split regions in this function.
2042          */
2043         if (ticket->t_flags & XLOG_TIC_INITED)
2044                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2045
2046         /*
2047          * Commit record headers need to be accounted for. These
2048          * come in as separate writes so are easy to detect.
2049          */
2050         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2051                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2052
2053         if (ticket->t_curr_res < 0)
2054                 xlog_print_tic_res(log->l_mp, ticket);
2055
2056         index = 0;
2057         lv = log_vector;
2058         vecp = lv->lv_iovecp;
2059         while (lv && index < lv->lv_niovecs) {
2060                 void            *ptr;
2061                 int             log_offset;
2062
2063                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2064                                                    &contwr, &log_offset);
2065                 if (error)
2066                         return error;
2067
2068                 ASSERT(log_offset <= iclog->ic_size - 1);
2069                 ptr = iclog->ic_datap + log_offset;
2070
2071                 /* start_lsn is the first lsn written to. That's all we need. */
2072                 if (!*start_lsn)
2073                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2074
2075                 /*
2076                  * This loop writes out as many regions as can fit in the amount
2077                  * of space which was allocated by xlog_state_get_iclog_space().
2078                  */
2079                 while (lv && index < lv->lv_niovecs) {
2080                         struct xfs_log_iovec    *reg = &vecp[index];
2081                         struct xlog_op_header   *ophdr;
2082                         int                     start_rec_copy;
2083                         int                     copy_len;
2084                         int                     copy_off;
2085
2086                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2087                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2088
2089                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2090                         if (start_rec_copy) {
2091                                 record_cnt++;
2092                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2093                                                    start_rec_copy);
2094                         }
2095
2096                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2097                         if (!ophdr)
2098                                 return XFS_ERROR(EIO);
2099
2100                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2101                                            sizeof(struct xlog_op_header));
2102
2103                         len += xlog_write_setup_copy(ticket, ophdr,
2104                                                      iclog->ic_size-log_offset,
2105                                                      reg->i_len,
2106                                                      &copy_off, &copy_len,
2107                                                      &partial_copy,
2108                                                      &partial_copy_len);
2109                         xlog_verify_dest_ptr(log, ptr);
2110
2111                         /* copy region */
2112                         ASSERT(copy_len >= 0);
2113                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
2114                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2115
2116                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2117                         record_cnt++;
2118                         data_cnt += contwr ? copy_len : 0;
2119
2120                         error = xlog_write_copy_finish(log, iclog, flags,
2121                                                        &record_cnt, &data_cnt,
2122                                                        &partial_copy,
2123                                                        &partial_copy_len,
2124                                                        log_offset,
2125                                                        commit_iclog);
2126                         if (error)
2127                                 return error;
2128
2129                         /*
2130                          * if we had a partial copy, we need to get more iclog
2131                          * space but we don't want to increment the region
2132                          * index because there is still more is this region to
2133                          * write.
2134                          *
2135                          * If we completed writing this region, and we flushed
2136                          * the iclog (indicated by resetting of the record
2137                          * count), then we also need to get more log space. If
2138                          * this was the last record, though, we are done and
2139                          * can just return.
2140                          */
2141                         if (partial_copy)
2142                                 break;
2143
2144                         if (++index == lv->lv_niovecs) {
2145                                 lv = lv->lv_next;
2146                                 index = 0;
2147                                 if (lv)
2148                                         vecp = lv->lv_iovecp;
2149                         }
2150                         if (record_cnt == 0) {
2151                                 if (!lv)
2152                                         return 0;
2153                                 break;
2154                         }
2155                 }
2156         }
2157
2158         ASSERT(len == 0);
2159
2160         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2161         if (!commit_iclog)
2162                 return xlog_state_release_iclog(log, iclog);
2163
2164         ASSERT(flags & XLOG_COMMIT_TRANS);
2165         *commit_iclog = iclog;
2166         return 0;
2167 }
2168
2169
2170 /*****************************************************************************
2171  *
2172  *              State Machine functions
2173  *
2174  *****************************************************************************
2175  */
2176
2177 /* Clean iclogs starting from the head.  This ordering must be
2178  * maintained, so an iclog doesn't become ACTIVE beyond one that
2179  * is SYNCING.  This is also required to maintain the notion that we use
2180  * a ordered wait queue to hold off would be writers to the log when every
2181  * iclog is trying to sync to disk.
2182  *
2183  * State Change: DIRTY -> ACTIVE
2184  */
2185 STATIC void
2186 xlog_state_clean_log(
2187         struct xlog *log)
2188 {
2189         xlog_in_core_t  *iclog;
2190         int changed = 0;
2191
2192         iclog = log->l_iclog;
2193         do {
2194                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2195                         iclog->ic_state = XLOG_STATE_ACTIVE;
2196                         iclog->ic_offset       = 0;
2197                         ASSERT(iclog->ic_callback == NULL);
2198                         /*
2199                          * If the number of ops in this iclog indicate it just
2200                          * contains the dummy transaction, we can
2201                          * change state into IDLE (the second time around).
2202                          * Otherwise we should change the state into
2203                          * NEED a dummy.
2204                          * We don't need to cover the dummy.
2205                          */
2206                         if (!changed &&
2207                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2208                                         XLOG_COVER_OPS)) {
2209                                 changed = 1;
2210                         } else {
2211                                 /*
2212                                  * We have two dirty iclogs so start over
2213                                  * This could also be num of ops indicates
2214                                  * this is not the dummy going out.
2215                                  */
2216                                 changed = 2;
2217                         }
2218                         iclog->ic_header.h_num_logops = 0;
2219                         memset(iclog->ic_header.h_cycle_data, 0,
2220                               sizeof(iclog->ic_header.h_cycle_data));
2221                         iclog->ic_header.h_lsn = 0;
2222                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2223                         /* do nothing */;
2224                 else
2225                         break;  /* stop cleaning */
2226                 iclog = iclog->ic_next;
2227         } while (iclog != log->l_iclog);
2228
2229         /* log is locked when we are called */
2230         /*
2231          * Change state for the dummy log recording.
2232          * We usually go to NEED. But we go to NEED2 if the changed indicates
2233          * we are done writing the dummy record.
2234          * If we are done with the second dummy recored (DONE2), then
2235          * we go to IDLE.
2236          */
2237         if (changed) {
2238                 switch (log->l_covered_state) {
2239                 case XLOG_STATE_COVER_IDLE:
2240                 case XLOG_STATE_COVER_NEED:
2241                 case XLOG_STATE_COVER_NEED2:
2242                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2243                         break;
2244
2245                 case XLOG_STATE_COVER_DONE:
2246                         if (changed == 1)
2247                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2248                         else
2249                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2250                         break;
2251
2252                 case XLOG_STATE_COVER_DONE2:
2253                         if (changed == 1)
2254                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2255                         else
2256                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2257                         break;
2258
2259                 default:
2260                         ASSERT(0);
2261                 }
2262         }
2263 }       /* xlog_state_clean_log */
2264
2265 STATIC xfs_lsn_t
2266 xlog_get_lowest_lsn(
2267         struct xlog     *log)
2268 {
2269         xlog_in_core_t  *lsn_log;
2270         xfs_lsn_t       lowest_lsn, lsn;
2271
2272         lsn_log = log->l_iclog;
2273         lowest_lsn = 0;
2274         do {
2275             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2276                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2277                 if ((lsn && !lowest_lsn) ||
2278                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2279                         lowest_lsn = lsn;
2280                 }
2281             }
2282             lsn_log = lsn_log->ic_next;
2283         } while (lsn_log != log->l_iclog);
2284         return lowest_lsn;
2285 }
2286
2287
2288 STATIC void
2289 xlog_state_do_callback(
2290         struct xlog             *log,
2291         int                     aborted,
2292         struct xlog_in_core     *ciclog)
2293 {
2294         xlog_in_core_t     *iclog;
2295         xlog_in_core_t     *first_iclog;        /* used to know when we've
2296                                                  * processed all iclogs once */
2297         xfs_log_callback_t *cb, *cb_next;
2298         int                flushcnt = 0;
2299         xfs_lsn_t          lowest_lsn;
2300         int                ioerrors;    /* counter: iclogs with errors */
2301         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2302         int                funcdidcallbacks; /* flag: function did callbacks */
2303         int                repeats;     /* for issuing console warnings if
2304                                          * looping too many times */
2305         int                wake = 0;
2306
2307         spin_lock(&log->l_icloglock);
2308         first_iclog = iclog = log->l_iclog;
2309         ioerrors = 0;
2310         funcdidcallbacks = 0;
2311         repeats = 0;
2312
2313         do {
2314                 /*
2315                  * Scan all iclogs starting with the one pointed to by the
2316                  * log.  Reset this starting point each time the log is
2317                  * unlocked (during callbacks).
2318                  *
2319                  * Keep looping through iclogs until one full pass is made
2320                  * without running any callbacks.
2321                  */
2322                 first_iclog = log->l_iclog;
2323                 iclog = log->l_iclog;
2324                 loopdidcallbacks = 0;
2325                 repeats++;
2326
2327                 do {
2328
2329                         /* skip all iclogs in the ACTIVE & DIRTY states */
2330                         if (iclog->ic_state &
2331                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2332                                 iclog = iclog->ic_next;
2333                                 continue;
2334                         }
2335
2336                         /*
2337                          * Between marking a filesystem SHUTDOWN and stopping
2338                          * the log, we do flush all iclogs to disk (if there
2339                          * wasn't a log I/O error). So, we do want things to
2340                          * go smoothly in case of just a SHUTDOWN  w/o a
2341                          * LOG_IO_ERROR.
2342                          */
2343                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2344                                 /*
2345                                  * Can only perform callbacks in order.  Since
2346                                  * this iclog is not in the DONE_SYNC/
2347                                  * DO_CALLBACK state, we skip the rest and
2348                                  * just try to clean up.  If we set our iclog
2349                                  * to DO_CALLBACK, we will not process it when
2350                                  * we retry since a previous iclog is in the
2351                                  * CALLBACK and the state cannot change since
2352                                  * we are holding the l_icloglock.
2353                                  */
2354                                 if (!(iclog->ic_state &
2355                                         (XLOG_STATE_DONE_SYNC |
2356                                                  XLOG_STATE_DO_CALLBACK))) {
2357                                         if (ciclog && (ciclog->ic_state ==
2358                                                         XLOG_STATE_DONE_SYNC)) {
2359                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2360                                         }
2361                                         break;
2362                                 }
2363                                 /*
2364                                  * We now have an iclog that is in either the
2365                                  * DO_CALLBACK or DONE_SYNC states. The other
2366                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2367                                  * caught by the above if and are going to
2368                                  * clean (i.e. we aren't doing their callbacks)
2369                                  * see the above if.
2370                                  */
2371
2372                                 /*
2373                                  * We will do one more check here to see if we
2374                                  * have chased our tail around.
2375                                  */
2376
2377                                 lowest_lsn = xlog_get_lowest_lsn(log);
2378                                 if (lowest_lsn &&
2379                                     XFS_LSN_CMP(lowest_lsn,
2380                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2381                                         iclog = iclog->ic_next;
2382                                         continue; /* Leave this iclog for
2383                                                    * another thread */
2384                                 }
2385
2386                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2387
2388
2389                                 /*
2390                                  * update the last_sync_lsn before we drop the
2391                                  * icloglock to ensure we are the only one that
2392                                  * can update it.
2393                                  */
2394                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2395                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2396                                 atomic64_set(&log->l_last_sync_lsn,
2397                                         be64_to_cpu(iclog->ic_header.h_lsn));
2398
2399                         } else
2400                                 ioerrors++;
2401
2402                         spin_unlock(&log->l_icloglock);
2403
2404                         /*
2405                          * Keep processing entries in the callback list until
2406                          * we come around and it is empty.  We need to
2407                          * atomically see that the list is empty and change the
2408                          * state to DIRTY so that we don't miss any more
2409                          * callbacks being added.
2410                          */
2411                         spin_lock(&iclog->ic_callback_lock);
2412                         cb = iclog->ic_callback;
2413                         while (cb) {
2414                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2415                                 iclog->ic_callback = NULL;
2416                                 spin_unlock(&iclog->ic_callback_lock);
2417
2418                                 /* perform callbacks in the order given */
2419                                 for (; cb; cb = cb_next) {
2420                                         cb_next = cb->cb_next;
2421                                         cb->cb_func(cb->cb_arg, aborted);
2422                                 }
2423                                 spin_lock(&iclog->ic_callback_lock);
2424                                 cb = iclog->ic_callback;
2425                         }
2426
2427                         loopdidcallbacks++;
2428                         funcdidcallbacks++;
2429
2430                         spin_lock(&log->l_icloglock);
2431                         ASSERT(iclog->ic_callback == NULL);
2432                         spin_unlock(&iclog->ic_callback_lock);
2433                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2434                                 iclog->ic_state = XLOG_STATE_DIRTY;
2435
2436                         /*
2437                          * Transition from DIRTY to ACTIVE if applicable.
2438                          * NOP if STATE_IOERROR.
2439                          */
2440                         xlog_state_clean_log(log);
2441
2442                         /* wake up threads waiting in xfs_log_force() */
2443                         wake_up_all(&iclog->ic_force_wait);
2444
2445                         iclog = iclog->ic_next;
2446                 } while (first_iclog != iclog);
2447
2448                 if (repeats > 5000) {
2449                         flushcnt += repeats;
2450                         repeats = 0;
2451                         xfs_warn(log->l_mp,
2452                                 "%s: possible infinite loop (%d iterations)",
2453                                 __func__, flushcnt);
2454                 }
2455         } while (!ioerrors && loopdidcallbacks);
2456
2457         /*
2458          * make one last gasp attempt to see if iclogs are being left in
2459          * limbo..
2460          */
2461 #ifdef DEBUG
2462         if (funcdidcallbacks) {
2463                 first_iclog = iclog = log->l_iclog;
2464                 do {
2465                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2466                         /*
2467                          * Terminate the loop if iclogs are found in states
2468                          * which will cause other threads to clean up iclogs.
2469                          *
2470                          * SYNCING - i/o completion will go through logs
2471                          * DONE_SYNC - interrupt thread should be waiting for
2472                          *              l_icloglock
2473                          * IOERROR - give up hope all ye who enter here
2474                          */
2475                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2476                             iclog->ic_state == XLOG_STATE_SYNCING ||
2477                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2478                             iclog->ic_state == XLOG_STATE_IOERROR )
2479                                 break;
2480                         iclog = iclog->ic_next;
2481                 } while (first_iclog != iclog);
2482         }
2483 #endif
2484
2485         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2486                 wake = 1;
2487         spin_unlock(&log->l_icloglock);
2488
2489         if (wake)
2490                 wake_up_all(&log->l_flush_wait);
2491 }
2492
2493
2494 /*
2495  * Finish transitioning this iclog to the dirty state.
2496  *
2497  * Make sure that we completely execute this routine only when this is
2498  * the last call to the iclog.  There is a good chance that iclog flushes,
2499  * when we reach the end of the physical log, get turned into 2 separate
2500  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2501  * routine.  By using the reference count bwritecnt, we guarantee that only
2502  * the second completion goes through.
2503  *
2504  * Callbacks could take time, so they are done outside the scope of the
2505  * global state machine log lock.
2506  */
2507 STATIC void
2508 xlog_state_done_syncing(
2509         xlog_in_core_t  *iclog,
2510         int             aborted)
2511 {
2512         struct xlog        *log = iclog->ic_log;
2513
2514         spin_lock(&log->l_icloglock);
2515
2516         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2517                iclog->ic_state == XLOG_STATE_IOERROR);
2518         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2519         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2520
2521
2522         /*
2523          * If we got an error, either on the first buffer, or in the case of
2524          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2525          * and none should ever be attempted to be written to disk
2526          * again.
2527          */
2528         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2529                 if (--iclog->ic_bwritecnt == 1) {
2530                         spin_unlock(&log->l_icloglock);
2531                         return;
2532                 }
2533                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2534         }
2535
2536         /*
2537          * Someone could be sleeping prior to writing out the next
2538          * iclog buffer, we wake them all, one will get to do the
2539          * I/O, the others get to wait for the result.
2540          */
2541         wake_up_all(&iclog->ic_write_wait);
2542         spin_unlock(&log->l_icloglock);
2543         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2544 }       /* xlog_state_done_syncing */
2545
2546
2547 /*
2548  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2549  * sleep.  We wait on the flush queue on the head iclog as that should be
2550  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2551  * we will wait here and all new writes will sleep until a sync completes.
2552  *
2553  * The in-core logs are used in a circular fashion. They are not used
2554  * out-of-order even when an iclog past the head is free.
2555  *
2556  * return:
2557  *      * log_offset where xlog_write() can start writing into the in-core
2558  *              log's data space.
2559  *      * in-core log pointer to which xlog_write() should write.
2560  *      * boolean indicating this is a continued write to an in-core log.
2561  *              If this is the last write, then the in-core log's offset field
2562  *              needs to be incremented, depending on the amount of data which
2563  *              is copied.
2564  */
2565 STATIC int
2566 xlog_state_get_iclog_space(
2567         struct xlog             *log,
2568         int                     len,
2569         struct xlog_in_core     **iclogp,
2570         struct xlog_ticket      *ticket,
2571         int                     *continued_write,
2572         int                     *logoffsetp)
2573 {
2574         int               log_offset;
2575         xlog_rec_header_t *head;
2576         xlog_in_core_t    *iclog;
2577         int               error;
2578
2579 restart:
2580         spin_lock(&log->l_icloglock);
2581         if (XLOG_FORCED_SHUTDOWN(log)) {
2582                 spin_unlock(&log->l_icloglock);
2583                 return XFS_ERROR(EIO);
2584         }
2585
2586         iclog = log->l_iclog;
2587         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2588                 XFS_STATS_INC(xs_log_noiclogs);
2589
2590                 /* Wait for log writes to have flushed */
2591                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2592                 goto restart;
2593         }
2594
2595         head = &iclog->ic_header;
2596
2597         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2598         log_offset = iclog->ic_offset;
2599
2600         /* On the 1st write to an iclog, figure out lsn.  This works
2601          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2602          * committing to.  If the offset is set, that's how many blocks
2603          * must be written.
2604          */
2605         if (log_offset == 0) {
2606                 ticket->t_curr_res -= log->l_iclog_hsize;
2607                 xlog_tic_add_region(ticket,
2608                                     log->l_iclog_hsize,
2609                                     XLOG_REG_TYPE_LRHEADER);
2610                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2611                 head->h_lsn = cpu_to_be64(
2612                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2613                 ASSERT(log->l_curr_block >= 0);
2614         }
2615
2616         /* If there is enough room to write everything, then do it.  Otherwise,
2617          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2618          * bit is on, so this will get flushed out.  Don't update ic_offset
2619          * until you know exactly how many bytes get copied.  Therefore, wait
2620          * until later to update ic_offset.
2621          *
2622          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2623          * can fit into remaining data section.
2624          */
2625         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2626                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2627
2628                 /*
2629                  * If I'm the only one writing to this iclog, sync it to disk.
2630                  * We need to do an atomic compare and decrement here to avoid
2631                  * racing with concurrent atomic_dec_and_lock() calls in
2632                  * xlog_state_release_iclog() when there is more than one
2633                  * reference to the iclog.
2634                  */
2635                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2636                         /* we are the only one */
2637                         spin_unlock(&log->l_icloglock);
2638                         error = xlog_state_release_iclog(log, iclog);
2639                         if (error)
2640                                 return error;
2641                 } else {
2642                         spin_unlock(&log->l_icloglock);
2643                 }
2644                 goto restart;
2645         }
2646
2647         /* Do we have enough room to write the full amount in the remainder
2648          * of this iclog?  Or must we continue a write on the next iclog and
2649          * mark this iclog as completely taken?  In the case where we switch
2650          * iclogs (to mark it taken), this particular iclog will release/sync
2651          * to disk in xlog_write().
2652          */
2653         if (len <= iclog->ic_size - iclog->ic_offset) {
2654                 *continued_write = 0;
2655                 iclog->ic_offset += len;
2656         } else {
2657                 *continued_write = 1;
2658                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2659         }
2660         *iclogp = iclog;
2661
2662         ASSERT(iclog->ic_offset <= iclog->ic_size);
2663         spin_unlock(&log->l_icloglock);
2664
2665         *logoffsetp = log_offset;
2666         return 0;
2667 }       /* xlog_state_get_iclog_space */
2668
2669 /* The first cnt-1 times through here we don't need to
2670  * move the grant write head because the permanent
2671  * reservation has reserved cnt times the unit amount.
2672  * Release part of current permanent unit reservation and
2673  * reset current reservation to be one units worth.  Also
2674  * move grant reservation head forward.
2675  */
2676 STATIC void
2677 xlog_regrant_reserve_log_space(
2678         struct xlog             *log,
2679         struct xlog_ticket      *ticket)
2680 {
2681         trace_xfs_log_regrant_reserve_enter(log, ticket);
2682
2683         if (ticket->t_cnt > 0)
2684                 ticket->t_cnt--;
2685
2686         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2687                                         ticket->t_curr_res);
2688         xlog_grant_sub_space(log, &log->l_write_head.grant,
2689                                         ticket->t_curr_res);
2690         ticket->t_curr_res = ticket->t_unit_res;
2691         xlog_tic_reset_res(ticket);
2692
2693         trace_xfs_log_regrant_reserve_sub(log, ticket);
2694
2695         /* just return if we still have some of the pre-reserved space */
2696         if (ticket->t_cnt > 0)
2697                 return;
2698
2699         xlog_grant_add_space(log, &log->l_reserve_head.grant,
2700                                         ticket->t_unit_res);
2701
2702         trace_xfs_log_regrant_reserve_exit(log, ticket);
2703
2704         ticket->t_curr_res = ticket->t_unit_res;
2705         xlog_tic_reset_res(ticket);
2706 }       /* xlog_regrant_reserve_log_space */
2707
2708
2709 /*
2710  * Give back the space left from a reservation.
2711  *
2712  * All the information we need to make a correct determination of space left
2713  * is present.  For non-permanent reservations, things are quite easy.  The
2714  * count should have been decremented to zero.  We only need to deal with the
2715  * space remaining in the current reservation part of the ticket.  If the
2716  * ticket contains a permanent reservation, there may be left over space which
2717  * needs to be released.  A count of N means that N-1 refills of the current
2718  * reservation can be done before we need to ask for more space.  The first
2719  * one goes to fill up the first current reservation.  Once we run out of
2720  * space, the count will stay at zero and the only space remaining will be
2721  * in the current reservation field.
2722  */
2723 STATIC void
2724 xlog_ungrant_log_space(
2725         struct xlog             *log,
2726         struct xlog_ticket      *ticket)
2727 {
2728         int     bytes;
2729
2730         if (ticket->t_cnt > 0)
2731                 ticket->t_cnt--;
2732
2733         trace_xfs_log_ungrant_enter(log, ticket);
2734         trace_xfs_log_ungrant_sub(log, ticket);
2735
2736         /*
2737          * If this is a permanent reservation ticket, we may be able to free
2738          * up more space based on the remaining count.
2739          */
2740         bytes = ticket->t_curr_res;
2741         if (ticket->t_cnt > 0) {
2742                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2743                 bytes += ticket->t_unit_res*ticket->t_cnt;
2744         }
2745
2746         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2747         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2748
2749         trace_xfs_log_ungrant_exit(log, ticket);
2750
2751         xfs_log_space_wake(log->l_mp);
2752 }
2753
2754 /*
2755  * Flush iclog to disk if this is the last reference to the given iclog and
2756  * the WANT_SYNC bit is set.
2757  *
2758  * When this function is entered, the iclog is not necessarily in the
2759  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2760  *
2761  *
2762  */
2763 STATIC int
2764 xlog_state_release_iclog(
2765         struct xlog             *log,
2766         struct xlog_in_core     *iclog)
2767 {
2768         int             sync = 0;       /* do we sync? */
2769
2770         if (iclog->ic_state & XLOG_STATE_IOERROR)
2771                 return XFS_ERROR(EIO);
2772
2773         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2774         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2775                 return 0;
2776
2777         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2778                 spin_unlock(&log->l_icloglock);
2779                 return XFS_ERROR(EIO);
2780         }
2781         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2782                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2783
2784         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2785                 /* update tail before writing to iclog */
2786                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2787                 sync++;
2788                 iclog->ic_state = XLOG_STATE_SYNCING;
2789                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2790                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2791                 /* cycle incremented when incrementing curr_block */
2792         }
2793         spin_unlock(&log->l_icloglock);
2794
2795         /*
2796          * We let the log lock go, so it's possible that we hit a log I/O
2797          * error or some other SHUTDOWN condition that marks the iclog
2798          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2799          * this iclog has consistent data, so we ignore IOERROR
2800          * flags after this point.
2801          */
2802         if (sync)
2803                 return xlog_sync(log, iclog);
2804         return 0;
2805 }       /* xlog_state_release_iclog */
2806
2807
2808 /*
2809  * This routine will mark the current iclog in the ring as WANT_SYNC
2810  * and move the current iclog pointer to the next iclog in the ring.
2811  * When this routine is called from xlog_state_get_iclog_space(), the
2812  * exact size of the iclog has not yet been determined.  All we know is
2813  * that every data block.  We have run out of space in this log record.
2814  */
2815 STATIC void
2816 xlog_state_switch_iclogs(
2817         struct xlog             *log,
2818         struct xlog_in_core     *iclog,
2819         int                     eventual_size)
2820 {
2821         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2822         if (!eventual_size)
2823                 eventual_size = iclog->ic_offset;
2824         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2825         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2826         log->l_prev_block = log->l_curr_block;
2827         log->l_prev_cycle = log->l_curr_cycle;
2828
2829         /* roll log?: ic_offset changed later */
2830         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2831
2832         /* Round up to next log-sunit */
2833         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2834             log->l_mp->m_sb.sb_logsunit > 1) {
2835                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2836                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2837         }
2838
2839         if (log->l_curr_block >= log->l_logBBsize) {
2840                 log->l_curr_cycle++;
2841                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2842                         log->l_curr_cycle++;
2843                 log->l_curr_block -= log->l_logBBsize;
2844                 ASSERT(log->l_curr_block >= 0);
2845         }
2846         ASSERT(iclog == log->l_iclog);
2847         log->l_iclog = iclog->ic_next;
2848 }       /* xlog_state_switch_iclogs */
2849
2850 /*
2851  * Write out all data in the in-core log as of this exact moment in time.
2852  *
2853  * Data may be written to the in-core log during this call.  However,
2854  * we don't guarantee this data will be written out.  A change from past
2855  * implementation means this routine will *not* write out zero length LRs.
2856  *
2857  * Basically, we try and perform an intelligent scan of the in-core logs.
2858  * If we determine there is no flushable data, we just return.  There is no
2859  * flushable data if:
2860  *
2861  *      1. the current iclog is active and has no data; the previous iclog
2862  *              is in the active or dirty state.
2863  *      2. the current iclog is drity, and the previous iclog is in the
2864  *              active or dirty state.
2865  *
2866  * We may sleep if:
2867  *
2868  *      1. the current iclog is not in the active nor dirty state.
2869  *      2. the current iclog dirty, and the previous iclog is not in the
2870  *              active nor dirty state.
2871  *      3. the current iclog is active, and there is another thread writing
2872  *              to this particular iclog.
2873  *      4. a) the current iclog is active and has no other writers
2874  *         b) when we return from flushing out this iclog, it is still
2875  *              not in the active nor dirty state.
2876  */
2877 int
2878 _xfs_log_force(
2879         struct xfs_mount        *mp,
2880         uint                    flags,
2881         int                     *log_flushed)
2882 {
2883         struct xlog             *log = mp->m_log;
2884         struct xlog_in_core     *iclog;
2885         xfs_lsn_t               lsn;
2886
2887         XFS_STATS_INC(xs_log_force);
2888
2889         xlog_cil_force(log);
2890
2891         spin_lock(&log->l_icloglock);
2892
2893         iclog = log->l_iclog;
2894         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2895                 spin_unlock(&log->l_icloglock);
2896                 return XFS_ERROR(EIO);
2897         }
2898
2899         /* If the head iclog is not active nor dirty, we just attach
2900          * ourselves to the head and go to sleep.
2901          */
2902         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2903             iclog->ic_state == XLOG_STATE_DIRTY) {
2904                 /*
2905                  * If the head is dirty or (active and empty), then
2906                  * we need to look at the previous iclog.  If the previous
2907                  * iclog is active or dirty we are done.  There is nothing
2908                  * to sync out.  Otherwise, we attach ourselves to the
2909                  * previous iclog and go to sleep.
2910                  */
2911                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2912                     (atomic_read(&iclog->ic_refcnt) == 0
2913                      && iclog->ic_offset == 0)) {
2914                         iclog = iclog->ic_prev;
2915                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2916                             iclog->ic_state == XLOG_STATE_DIRTY)
2917                                 goto no_sleep;
2918                         else
2919                                 goto maybe_sleep;
2920                 } else {
2921                         if (atomic_read(&iclog->ic_refcnt) == 0) {
2922                                 /* We are the only one with access to this
2923                                  * iclog.  Flush it out now.  There should
2924                                  * be a roundoff of zero to show that someone
2925                                  * has already taken care of the roundoff from
2926                                  * the previous sync.
2927                                  */
2928                                 atomic_inc(&iclog->ic_refcnt);
2929                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2930                                 xlog_state_switch_iclogs(log, iclog, 0);
2931                                 spin_unlock(&log->l_icloglock);
2932
2933                                 if (xlog_state_release_iclog(log, iclog))
2934                                         return XFS_ERROR(EIO);
2935
2936                                 if (log_flushed)
2937                                         *log_flushed = 1;
2938                                 spin_lock(&log->l_icloglock);
2939                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2940                                     iclog->ic_state != XLOG_STATE_DIRTY)
2941                                         goto maybe_sleep;
2942                                 else
2943                                         goto no_sleep;
2944                         } else {
2945                                 /* Someone else is writing to this iclog.
2946                                  * Use its call to flush out the data.  However,
2947                                  * the other thread may not force out this LR,
2948                                  * so we mark it WANT_SYNC.
2949                                  */
2950                                 xlog_state_switch_iclogs(log, iclog, 0);
2951                                 goto maybe_sleep;
2952                         }
2953                 }
2954         }
2955
2956         /* By the time we come around again, the iclog could've been filled
2957          * which would give it another lsn.  If we have a new lsn, just
2958          * return because the relevant data has been flushed.
2959          */
2960 maybe_sleep:
2961         if (flags & XFS_LOG_SYNC) {
2962                 /*
2963                  * We must check if we're shutting down here, before
2964                  * we wait, while we're holding the l_icloglock.
2965                  * Then we check again after waking up, in case our
2966                  * sleep was disturbed by a bad news.
2967                  */
2968                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2969                         spin_unlock(&log->l_icloglock);
2970                         return XFS_ERROR(EIO);
2971                 }
2972                 XFS_STATS_INC(xs_log_force_sleep);
2973                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
2974                 /*
2975                  * No need to grab the log lock here since we're
2976                  * only deciding whether or not to return EIO
2977                  * and the memory read should be atomic.
2978                  */
2979                 if (iclog->ic_state & XLOG_STATE_IOERROR)
2980                         return XFS_ERROR(EIO);
2981                 if (log_flushed)
2982                         *log_flushed = 1;
2983         } else {
2984
2985 no_sleep:
2986                 spin_unlock(&log->l_icloglock);
2987         }
2988         return 0;
2989 }
2990
2991 /*
2992  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
2993  * about errors or whether the log was flushed or not. This is the normal
2994  * interface to use when trying to unpin items or move the log forward.
2995  */
2996 void
2997 xfs_log_force(
2998         xfs_mount_t     *mp,
2999         uint            flags)
3000 {
3001         int     error;
3002
3003         trace_xfs_log_force(mp, 0);
3004         error = _xfs_log_force(mp, flags, NULL);
3005         if (error)
3006                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3007 }
3008
3009 /*
3010  * Force the in-core log to disk for a specific LSN.
3011  *
3012  * Find in-core log with lsn.
3013  *      If it is in the DIRTY state, just return.
3014  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
3015  *              state and go to sleep or return.
3016  *      If it is in any other state, go to sleep or return.
3017  *
3018  * Synchronous forces are implemented with a signal variable. All callers
3019  * to force a given lsn to disk will wait on a the sv attached to the
3020  * specific in-core log.  When given in-core log finally completes its
3021  * write to disk, that thread will wake up all threads waiting on the
3022  * sv.
3023  */
3024 int
3025 _xfs_log_force_lsn(
3026         struct xfs_mount        *mp,
3027         xfs_lsn_t               lsn,
3028         uint                    flags,
3029         int                     *log_flushed)
3030 {
3031         struct xlog             *log = mp->m_log;
3032         struct xlog_in_core     *iclog;
3033         int                     already_slept = 0;
3034
3035         ASSERT(lsn != 0);
3036
3037         XFS_STATS_INC(xs_log_force);
3038
3039         lsn = xlog_cil_force_lsn(log, lsn);
3040         if (lsn == NULLCOMMITLSN)
3041                 return 0;
3042
3043 try_again:
3044         spin_lock(&log->l_icloglock);
3045         iclog = log->l_iclog;
3046         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3047                 spin_unlock(&log->l_icloglock);
3048                 return XFS_ERROR(EIO);
3049         }
3050
3051         do {
3052                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
3053                         iclog = iclog->ic_next;
3054                         continue;
3055                 }
3056
3057                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3058                         spin_unlock(&log->l_icloglock);
3059                         return 0;
3060                 }
3061
3062                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3063                         /*
3064                          * We sleep here if we haven't already slept (e.g.
3065                          * this is the first time we've looked at the correct
3066                          * iclog buf) and the buffer before us is going to
3067                          * be sync'ed. The reason for this is that if we
3068                          * are doing sync transactions here, by waiting for
3069                          * the previous I/O to complete, we can allow a few
3070                          * more transactions into this iclog before we close
3071                          * it down.
3072                          *
3073                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3074                          * up the refcnt so we can release the log (which
3075                          * drops the ref count).  The state switch keeps new
3076                          * transaction commits from using this buffer.  When
3077                          * the current commits finish writing into the buffer,
3078                          * the refcount will drop to zero and the buffer will
3079                          * go out then.
3080                          */
3081                         if (!already_slept &&
3082                             (iclog->ic_prev->ic_state &
3083                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3084                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3085
3086                                 XFS_STATS_INC(xs_log_force_sleep);
3087
3088                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3089                                                         &log->l_icloglock);
3090                                 if (log_flushed)
3091                                         *log_flushed = 1;
3092                                 already_slept = 1;
3093                                 goto try_again;
3094                         }
3095                         atomic_inc(&iclog->ic_refcnt);
3096                         xlog_state_switch_iclogs(log, iclog, 0);
3097                         spin_unlock(&log->l_icloglock);
3098                         if (xlog_state_release_iclog(log, iclog))
3099                                 return XFS_ERROR(EIO);
3100                         if (log_flushed)
3101                                 *log_flushed = 1;
3102                         spin_lock(&log->l_icloglock);
3103                 }
3104
3105                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3106                     !(iclog->ic_state &
3107                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3108                         /*
3109                          * Don't wait on completion if we know that we've
3110                          * gotten a log write error.
3111                          */
3112                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3113                                 spin_unlock(&log->l_icloglock);
3114                                 return XFS_ERROR(EIO);
3115                         }
3116                         XFS_STATS_INC(xs_log_force_sleep);
3117                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3118                         /*
3119                          * No need to grab the log lock here since we're
3120                          * only deciding whether or not to return EIO
3121                          * and the memory read should be atomic.
3122                          */
3123                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3124                                 return XFS_ERROR(EIO);
3125
3126                         if (log_flushed)
3127                                 *log_flushed = 1;
3128                 } else {                /* just return */
3129                         spin_unlock(&log->l_icloglock);
3130                 }
3131
3132                 return 0;
3133         } while (iclog != log->l_iclog);
3134
3135         spin_unlock(&log->l_icloglock);
3136         return 0;
3137 }
3138
3139 /*
3140  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3141  * about errors or whether the log was flushed or not. This is the normal
3142  * interface to use when trying to unpin items or move the log forward.
3143  */
3144 void
3145 xfs_log_force_lsn(
3146         xfs_mount_t     *mp,
3147         xfs_lsn_t       lsn,
3148         uint            flags)
3149 {
3150         int     error;
3151
3152         trace_xfs_log_force(mp, lsn);
3153         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3154         if (error)
3155                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3156 }
3157
3158 /*
3159  * Called when we want to mark the current iclog as being ready to sync to
3160  * disk.
3161  */
3162 STATIC void
3163 xlog_state_want_sync(
3164         struct xlog             *log,
3165         struct xlog_in_core     *iclog)
3166 {
3167         assert_spin_locked(&log->l_icloglock);
3168
3169         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3170                 xlog_state_switch_iclogs(log, iclog, 0);
3171         } else {
3172                 ASSERT(iclog->ic_state &
3173                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3174         }
3175 }
3176
3177
3178 /*****************************************************************************
3179  *
3180  *              TICKET functions
3181  *
3182  *****************************************************************************
3183  */
3184
3185 /*
3186  * Free a used ticket when its refcount falls to zero.
3187  */
3188 void
3189 xfs_log_ticket_put(
3190         xlog_ticket_t   *ticket)
3191 {
3192         ASSERT(atomic_read(&ticket->t_ref) > 0);
3193         if (atomic_dec_and_test(&ticket->t_ref))
3194                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3195 }
3196
3197 xlog_ticket_t *
3198 xfs_log_ticket_get(
3199         xlog_ticket_t   *ticket)
3200 {
3201         ASSERT(atomic_read(&ticket->t_ref) > 0);
3202         atomic_inc(&ticket->t_ref);
3203         return ticket;
3204 }
3205
3206 /*
3207  * Allocate and initialise a new log ticket.
3208  */
3209 struct xlog_ticket *
3210 xlog_ticket_alloc(
3211         struct xlog     *log,
3212         int             unit_bytes,
3213         int             cnt,
3214         char            client,
3215         bool            permanent,
3216         xfs_km_flags_t  alloc_flags)
3217 {
3218         struct xlog_ticket *tic;
3219         uint            num_headers;
3220         int             iclog_space;
3221
3222         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3223         if (!tic)
3224                 return NULL;
3225
3226         /*
3227          * Permanent reservations have up to 'cnt'-1 active log operations
3228          * in the log.  A unit in this case is the amount of space for one
3229          * of these log operations.  Normal reservations have a cnt of 1
3230          * and their unit amount is the total amount of space required.
3231          *
3232          * The following lines of code account for non-transaction data
3233          * which occupy space in the on-disk log.
3234          *
3235          * Normal form of a transaction is:
3236          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3237          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3238          *
3239          * We need to account for all the leadup data and trailer data
3240          * around the transaction data.
3241          * And then we need to account for the worst case in terms of using
3242          * more space.
3243          * The worst case will happen if:
3244          * - the placement of the transaction happens to be such that the
3245          *   roundoff is at its maximum
3246          * - the transaction data is synced before the commit record is synced
3247          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3248          *   Therefore the commit record is in its own Log Record.
3249          *   This can happen as the commit record is called with its
3250          *   own region to xlog_write().
3251          *   This then means that in the worst case, roundoff can happen for
3252          *   the commit-rec as well.
3253          *   The commit-rec is smaller than padding in this scenario and so it is
3254          *   not added separately.
3255          */
3256
3257         /* for trans header */
3258         unit_bytes += sizeof(xlog_op_header_t);
3259         unit_bytes += sizeof(xfs_trans_header_t);
3260
3261         /* for start-rec */
3262         unit_bytes += sizeof(xlog_op_header_t);
3263
3264         /*
3265          * for LR headers - the space for data in an iclog is the size minus
3266          * the space used for the headers. If we use the iclog size, then we
3267          * undercalculate the number of headers required.
3268          *
3269          * Furthermore - the addition of op headers for split-recs might
3270          * increase the space required enough to require more log and op
3271          * headers, so take that into account too.
3272          *
3273          * IMPORTANT: This reservation makes the assumption that if this
3274          * transaction is the first in an iclog and hence has the LR headers
3275          * accounted to it, then the remaining space in the iclog is
3276          * exclusively for this transaction.  i.e. if the transaction is larger
3277          * than the iclog, it will be the only thing in that iclog.
3278          * Fundamentally, this means we must pass the entire log vector to
3279          * xlog_write to guarantee this.
3280          */
3281         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3282         num_headers = howmany(unit_bytes, iclog_space);
3283
3284         /* for split-recs - ophdrs added when data split over LRs */
3285         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3286
3287         /* add extra header reservations if we overrun */
3288         while (!num_headers ||
3289                howmany(unit_bytes, iclog_space) > num_headers) {
3290                 unit_bytes += sizeof(xlog_op_header_t);
3291                 num_headers++;
3292         }
3293         unit_bytes += log->l_iclog_hsize * num_headers;
3294
3295         /* for commit-rec LR header - note: padding will subsume the ophdr */
3296         unit_bytes += log->l_iclog_hsize;
3297
3298         /* for roundoff padding for transaction data and one for commit record */
3299         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3300             log->l_mp->m_sb.sb_logsunit > 1) {
3301                 /* log su roundoff */
3302                 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3303         } else {
3304                 /* BB roundoff */
3305                 unit_bytes += 2*BBSIZE;
3306         }
3307
3308         atomic_set(&tic->t_ref, 1);
3309         tic->t_task             = current;
3310         INIT_LIST_HEAD(&tic->t_queue);
3311         tic->t_unit_res         = unit_bytes;
3312         tic->t_curr_res         = unit_bytes;
3313         tic->t_cnt              = cnt;
3314         tic->t_ocnt             = cnt;
3315         tic->t_tid              = random32();
3316         tic->t_clientid         = client;
3317         tic->t_flags            = XLOG_TIC_INITED;
3318         tic->t_trans_type       = 0;
3319         if (permanent)
3320                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3321
3322         xlog_tic_reset_res(tic);
3323
3324         return tic;
3325 }
3326
3327
3328 /******************************************************************************
3329  *
3330  *              Log debug routines
3331  *
3332  ******************************************************************************
3333  */
3334 #if defined(DEBUG)
3335 /*
3336  * Make sure that the destination ptr is within the valid data region of
3337  * one of the iclogs.  This uses backup pointers stored in a different
3338  * part of the log in case we trash the log structure.
3339  */
3340 void
3341 xlog_verify_dest_ptr(
3342         struct xlog     *log,
3343         char            *ptr)
3344 {
3345         int i;
3346         int good_ptr = 0;
3347
3348         for (i = 0; i < log->l_iclog_bufs; i++) {
3349                 if (ptr >= log->l_iclog_bak[i] &&
3350                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3351                         good_ptr++;
3352         }
3353
3354         if (!good_ptr)
3355                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3356 }
3357
3358 /*
3359  * Check to make sure the grant write head didn't just over lap the tail.  If
3360  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3361  * the cycles differ by exactly one and check the byte count.
3362  *
3363  * This check is run unlocked, so can give false positives. Rather than assert
3364  * on failures, use a warn-once flag and a panic tag to allow the admin to
3365  * determine if they want to panic the machine when such an error occurs. For
3366  * debug kernels this will have the same effect as using an assert but, unlinke
3367  * an assert, it can be turned off at runtime.
3368  */
3369 STATIC void
3370 xlog_verify_grant_tail(
3371         struct xlog     *log)
3372 {
3373         int             tail_cycle, tail_blocks;
3374         int             cycle, space;
3375
3376         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3377         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3378         if (tail_cycle != cycle) {
3379                 if (cycle - 1 != tail_cycle &&
3380                     !(log->l_flags & XLOG_TAIL_WARN)) {
3381                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3382                                 "%s: cycle - 1 != tail_cycle", __func__);
3383                         log->l_flags |= XLOG_TAIL_WARN;
3384                 }
3385
3386                 if (space > BBTOB(tail_blocks) &&
3387                     !(log->l_flags & XLOG_TAIL_WARN)) {
3388                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3389                                 "%s: space > BBTOB(tail_blocks)", __func__);
3390                         log->l_flags |= XLOG_TAIL_WARN;
3391                 }
3392         }
3393 }
3394
3395 /* check if it will fit */
3396 STATIC void
3397 xlog_verify_tail_lsn(
3398         struct xlog             *log,
3399         struct xlog_in_core     *iclog,
3400         xfs_lsn_t               tail_lsn)
3401 {
3402     int blocks;
3403
3404     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3405         blocks =
3406             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3407         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3408                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3409     } else {
3410         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3411
3412         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3413                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3414
3415         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3416         if (blocks < BTOBB(iclog->ic_offset) + 1)
3417                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3418     }
3419 }       /* xlog_verify_tail_lsn */
3420
3421 /*
3422  * Perform a number of checks on the iclog before writing to disk.
3423  *
3424  * 1. Make sure the iclogs are still circular
3425  * 2. Make sure we have a good magic number
3426  * 3. Make sure we don't have magic numbers in the data
3427  * 4. Check fields of each log operation header for:
3428  *      A. Valid client identifier
3429  *      B. tid ptr value falls in valid ptr space (user space code)
3430  *      C. Length in log record header is correct according to the
3431  *              individual operation headers within record.
3432  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3433  *      log, check the preceding blocks of the physical log to make sure all
3434  *      the cycle numbers agree with the current cycle number.
3435  */
3436 STATIC void
3437 xlog_verify_iclog(
3438         struct xlog             *log,
3439         struct xlog_in_core     *iclog,
3440         int                     count,
3441         boolean_t               syncing)
3442 {
3443         xlog_op_header_t        *ophead;
3444         xlog_in_core_t          *icptr;
3445         xlog_in_core_2_t        *xhdr;
3446         xfs_caddr_t             ptr;
3447         xfs_caddr_t             base_ptr;
3448         __psint_t               field_offset;
3449         __uint8_t               clientid;
3450         int                     len, i, j, k, op_len;
3451         int                     idx;
3452
3453         /* check validity of iclog pointers */
3454         spin_lock(&log->l_icloglock);
3455         icptr = log->l_iclog;
3456         for (i=0; i < log->l_iclog_bufs; i++) {
3457                 if (icptr == NULL)
3458                         xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3459                 icptr = icptr->ic_next;
3460         }
3461         if (icptr != log->l_iclog)
3462                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3463         spin_unlock(&log->l_icloglock);
3464
3465         /* check log magic numbers */
3466         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3467                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3468
3469         ptr = (xfs_caddr_t) &iclog->ic_header;
3470         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3471              ptr += BBSIZE) {
3472                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3473                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3474                                 __func__);
3475         }
3476
3477         /* check fields */
3478         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3479         ptr = iclog->ic_datap;
3480         base_ptr = ptr;
3481         ophead = (xlog_op_header_t *)ptr;
3482         xhdr = iclog->ic_data;
3483         for (i = 0; i < len; i++) {
3484                 ophead = (xlog_op_header_t *)ptr;
3485
3486                 /* clientid is only 1 byte */
3487                 field_offset = (__psint_t)
3488                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3489                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3490                         clientid = ophead->oh_clientid;
3491                 } else {
3492                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3493                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3494                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3495                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3496                                 clientid = xlog_get_client_id(
3497                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3498                         } else {
3499                                 clientid = xlog_get_client_id(
3500                                         iclog->ic_header.h_cycle_data[idx]);
3501                         }
3502                 }
3503                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3504                         xfs_warn(log->l_mp,
3505                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3506                                 __func__, clientid, ophead,
3507                                 (unsigned long)field_offset);
3508
3509                 /* check length */
3510                 field_offset = (__psint_t)
3511                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3512                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3513                         op_len = be32_to_cpu(ophead->oh_len);
3514                 } else {
3515                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3516                                     (__psint_t)iclog->ic_datap);
3517                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3518                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3519                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3520                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3521                         } else {
3522                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3523                         }
3524                 }
3525                 ptr += sizeof(xlog_op_header_t) + op_len;
3526         }
3527 }       /* xlog_verify_iclog */
3528 #endif
3529
3530 /*
3531  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3532  */
3533 STATIC int
3534 xlog_state_ioerror(
3535         struct xlog     *log)
3536 {
3537         xlog_in_core_t  *iclog, *ic;
3538
3539         iclog = log->l_iclog;
3540         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3541                 /*
3542                  * Mark all the incore logs IOERROR.
3543                  * From now on, no log flushes will result.
3544                  */
3545                 ic = iclog;
3546                 do {
3547                         ic->ic_state = XLOG_STATE_IOERROR;
3548                         ic = ic->ic_next;
3549                 } while (ic != iclog);
3550                 return 0;
3551         }
3552         /*
3553          * Return non-zero, if state transition has already happened.
3554          */
3555         return 1;
3556 }
3557
3558 /*
3559  * This is called from xfs_force_shutdown, when we're forcibly
3560  * shutting down the filesystem, typically because of an IO error.
3561  * Our main objectives here are to make sure that:
3562  *      a. the filesystem gets marked 'SHUTDOWN' for all interested
3563  *         parties to find out, 'atomically'.
3564  *      b. those who're sleeping on log reservations, pinned objects and
3565  *          other resources get woken up, and be told the bad news.
3566  *      c. nothing new gets queued up after (a) and (b) are done.
3567  *      d. if !logerror, flush the iclogs to disk, then seal them off
3568  *         for business.
3569  *
3570  * Note: for delayed logging the !logerror case needs to flush the regions
3571  * held in memory out to the iclogs before flushing them to disk. This needs
3572  * to be done before the log is marked as shutdown, otherwise the flush to the
3573  * iclogs will fail.
3574  */
3575 int
3576 xfs_log_force_umount(
3577         struct xfs_mount        *mp,
3578         int                     logerror)
3579 {
3580         struct xlog     *log;
3581         int             retval;
3582
3583         log = mp->m_log;
3584
3585         /*
3586          * If this happens during log recovery, don't worry about
3587          * locking; the log isn't open for business yet.
3588          */
3589         if (!log ||
3590             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3591                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3592                 if (mp->m_sb_bp)
3593                         XFS_BUF_DONE(mp->m_sb_bp);
3594                 return 0;
3595         }
3596
3597         /*
3598          * Somebody could've already done the hard work for us.
3599          * No need to get locks for this.
3600          */
3601         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3602                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3603                 return 1;
3604         }
3605         retval = 0;
3606
3607         /*
3608          * Flush the in memory commit item list before marking the log as
3609          * being shut down. We need to do it in this order to ensure all the
3610          * completed transactions are flushed to disk with the xfs_log_force()
3611          * call below.
3612          */
3613         if (!logerror)
3614                 xlog_cil_force(log);
3615
3616         /*
3617          * mark the filesystem and the as in a shutdown state and wake
3618          * everybody up to tell them the bad news.
3619          */
3620         spin_lock(&log->l_icloglock);
3621         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3622         if (mp->m_sb_bp)
3623                 XFS_BUF_DONE(mp->m_sb_bp);
3624
3625         /*
3626          * This flag is sort of redundant because of the mount flag, but
3627          * it's good to maintain the separation between the log and the rest
3628          * of XFS.
3629          */
3630         log->l_flags |= XLOG_IO_ERROR;
3631
3632         /*
3633          * If we hit a log error, we want to mark all the iclogs IOERROR
3634          * while we're still holding the loglock.
3635          */
3636         if (logerror)
3637                 retval = xlog_state_ioerror(log);
3638         spin_unlock(&log->l_icloglock);
3639
3640         /*
3641          * We don't want anybody waiting for log reservations after this. That
3642          * means we have to wake up everybody queued up on reserveq as well as
3643          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3644          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3645          * action is protected by the grant locks.
3646          */
3647         xlog_grant_head_wake_all(&log->l_reserve_head);
3648         xlog_grant_head_wake_all(&log->l_write_head);
3649
3650         if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3651                 ASSERT(!logerror);
3652                 /*
3653                  * Force the incore logs to disk before shutting the
3654                  * log down completely.
3655                  */
3656                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3657
3658                 spin_lock(&log->l_icloglock);
3659                 retval = xlog_state_ioerror(log);
3660                 spin_unlock(&log->l_icloglock);
3661         }
3662         /*
3663          * Wake up everybody waiting on xfs_log_force.
3664          * Callback all log item committed functions as if the
3665          * log writes were completed.
3666          */
3667         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3668
3669 #ifdef XFSERRORDEBUG
3670         {
3671                 xlog_in_core_t  *iclog;
3672
3673                 spin_lock(&log->l_icloglock);
3674                 iclog = log->l_iclog;
3675                 do {
3676                         ASSERT(iclog->ic_callback == 0);
3677                         iclog = iclog->ic_next;
3678                 } while (iclog != log->l_iclog);
3679                 spin_unlock(&log->l_icloglock);
3680         }
3681 #endif
3682         /* return non-zero if log IOERROR transition had already happened */
3683         return retval;
3684 }
3685
3686 STATIC int
3687 xlog_iclogs_empty(
3688         struct xlog     *log)
3689 {
3690         xlog_in_core_t  *iclog;
3691
3692         iclog = log->l_iclog;
3693         do {
3694                 /* endianness does not matter here, zero is zero in
3695                  * any language.
3696                  */
3697                 if (iclog->ic_header.h_num_logops)
3698                         return 0;
3699                 iclog = iclog->ic_next;
3700         } while (iclog != log->l_iclog);
3701         return 1;
3702 }