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       xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
  42                                     xlog_in_core_t **, xfs_lsn_t *);
  43 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t     *mp,
  44                                 xfs_buftarg_t   *log_target,
  45                                 xfs_daddr_t     blk_offset,
  46                                 int             num_bblks);
  47 STATIC int       xlog_space_left(struct log *log, atomic64_t *head);
  48 STATIC int       xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
  49 STATIC void      xlog_dealloc_log(xlog_t *log);
  50
  51 /* local state machine functions */
  52 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  53 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
  54 STATIC int  xlog_state_get_iclog_space(xlog_t           *log,
  55                                        int              len,
  56                                        xlog_in_core_t   **iclog,
  57                                        xlog_ticket_t    *ticket,
  58                                        int              *continued_write,
  59                                        int              *logoffsetp);
  60 STATIC int  xlog_state_release_iclog(xlog_t             *log,
  61                                      xlog_in_core_t     *iclog);
  62 STATIC void xlog_state_switch_iclogs(xlog_t             *log,
  63                                      xlog_in_core_t *iclog,
  64                                      int                eventual_size);
  65 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
  66
  67 STATIC void xlog_grant_push_ail(struct log      *log,
  68                                 int             need_bytes);
  69 STATIC void xlog_regrant_reserve_log_space(xlog_t        *log,
  70                                            xlog_ticket_t *ticket);
  71 STATIC void xlog_ungrant_log_space(xlog_t        *log,
  72                                    xlog_ticket_t *ticket);
  73
  74 #if defined(DEBUG)
  75 STATIC void     xlog_verify_dest_ptr(xlog_t *log, char *ptr);
  76 STATIC void     xlog_verify_grant_tail(struct log *log);
  77 STATIC void     xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
  78                                   int count, boolean_t syncing);
  79 STATIC void     xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
  80                                      xfs_lsn_t tail_lsn);
  81 #else
  82 #define xlog_verify_dest_ptr(a,b)
  83 #define xlog_verify_grant_tail(a)
  84 #define xlog_verify_iclog(a,b,c,d)
  85 #define xlog_verify_tail_lsn(a,b,c)
  86 #endif
  87
  88 STATIC int      xlog_iclogs_empty(xlog_t *log);
  89
  90 static void
  91 xlog_grant_sub_space(
  92         struct log      *log,
  93         atomic64_t      *head,
  94         int             bytes)
  95 {
  96         int64_t head_val = atomic64_read(head);
  97         int64_t new, old;
  98
  99         do {
 100                 int     cycle, space;
 101
 102                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 103
 104                 space -= bytes;
 105                 if (space < 0) {
 106                         space += log->l_logsize;
 107                         cycle--;
 108                 }
 109
 110                 old = head_val;
 111                 new = xlog_assign_grant_head_val(cycle, space);
 112                 head_val = atomic64_cmpxchg(head, old, new);
 113         } while (head_val != old);
 114 }
 115
 116 static void
 117 xlog_grant_add_space(
 118         struct log      *log,
 119         atomic64_t      *head,
 120         int             bytes)
 121 {
 122         int64_t head_val = atomic64_read(head);
 123         int64_t new, old;
 124
 125         do {
 126                 int             tmp;
 127                 int             cycle, space;
 128
 129                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 130
 131                 tmp = log->l_logsize - space;
 132                 if (tmp > bytes)
 133                         space += bytes;
 134                 else {
 135                         space = bytes - tmp;
 136                         cycle++;
 137                 }
 138
 139                 old = head_val;
 140                 new = xlog_assign_grant_head_val(cycle, space);
 141                 head_val = atomic64_cmpxchg(head, old, new);
 142         } while (head_val != old);
 143 }
 144
 145 STATIC void
 146 xlog_grant_head_init(
 147         struct xlog_grant_head  *head)
 148 {
 149         xlog_assign_grant_head(&head->grant, 1, 0);
 150         INIT_LIST_HEAD(&head->waiters);
 151         spin_lock_init(&head->lock);
 152 }
 153
 154 STATIC void
 155 xlog_grant_head_wake_all(
 156         struct xlog_grant_head  *head)
 157 {
 158         struct xlog_ticket      *tic;
 159
 160         spin_lock(&head->lock);
 161         list_for_each_entry(tic, &head->waiters, t_queue)
 162                 wake_up_process(tic->t_task);
 163         spin_unlock(&head->lock);
 164 }
 165
 166 static inline int
 167 xlog_ticket_reservation(
 168         struct log              *log,
 169         struct xlog_grant_head  *head,
 170         struct xlog_ticket      *tic)
 171 {
 172         if (head == &log->l_write_head) {
 173                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 174                 return tic->t_unit_res;
 175         } else {
 176                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 177                         return tic->t_unit_res * tic->t_cnt;
 178                 else
 179                         return tic->t_unit_res;
 180         }
 181 }
 182
 183 STATIC bool
 184 xlog_grant_head_wake(
 185         struct log              *log,
 186         struct xlog_grant_head  *head,
 187         int                     *free_bytes)
 188 {
 189         struct xlog_ticket      *tic;
 190         int                     need_bytes;
 191
 192         list_for_each_entry(tic, &head->waiters, t_queue) {
 193                 need_bytes = xlog_ticket_reservation(log, head, tic);
 194                 if (*free_bytes < need_bytes)
 195                         return false;
 196
 197                 *free_bytes -= need_bytes;
 198                 trace_xfs_log_grant_wake_up(log, tic);
 199                 wake_up_process(tic->t_task);
 200         }
 201
 202         return true;
 203 }
 204
 205 STATIC int
 206 xlog_grant_head_wait(
 207         struct log              *log,
 208         struct xlog_grant_head  *head,
 209         struct xlog_ticket      *tic,
 210         int                     need_bytes)
 211 {
 212         list_add_tail(&tic->t_queue, &head->waiters);
 213
 214         do {
 215                 if (XLOG_FORCED_SHUTDOWN(log))
 216                         goto shutdown;
 217                 xlog_grant_push_ail(log, need_bytes);
 218
 219                 __set_current_state(TASK_UNINTERRUPTIBLE);
 220                 spin_unlock(&head->lock);
 221
 222                 XFS_STATS_INC(xs_sleep_logspace);
 223
 224                 trace_xfs_log_grant_sleep(log, tic);
 225                 schedule();
 226                 trace_xfs_log_grant_wake(log, tic);
 227
 228                 spin_lock(&head->lock);
 229                 if (XLOG_FORCED_SHUTDOWN(log))
 230                         goto shutdown;
 231         } while (xlog_space_left(log, &head->grant) < need_bytes);
 232
 233         list_del_init(&tic->t_queue);
 234         return 0;
 235 shutdown:
 236         list_del_init(&tic->t_queue);
 237         return XFS_ERROR(EIO);
 238 }
 239
 240 /*
 241  * Atomically get the log space required for a log ticket.
 242  *
 243  * Once a ticket gets put onto head->waiters, it will only return after the
 244  * needed reservation is satisfied.
 245  *
 246  * This function is structured so that it has a lock free fast path. This is
 247  * necessary because every new transaction reservation will come through this
 248  * path. Hence any lock will be globally hot if we take it unconditionally on
 249  * every pass.
 250  *
 251  * As tickets are only ever moved on and off head->waiters under head->lock, we
 252  * only need to take that lock if we are going to add the ticket to the queue
 253  * and sleep. We can avoid taking the lock if the ticket was never added to
 254  * head->waiters because the t_queue list head will be empty and we hold the
 255  * only reference to it so it can safely be checked unlocked.
 256  */
 257 STATIC int
 258 xlog_grant_head_check(
 259         struct log              *log,
 260         struct xlog_grant_head  *head,
 261         struct xlog_ticket      *tic,
 262         int                     *need_bytes)
 263 {
 264         int                     free_bytes;
 265         int                     error = 0;
 266
 267         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 268
 269         /*
 270          * If there are other waiters on the queue then give them a chance at
 271          * logspace before us.  Wake up the first waiters, if we do not wake
 272          * up all the waiters then go to sleep waiting for more free space,
 273          * otherwise try to get some space for this transaction.
 274          */
 275         *need_bytes = xlog_ticket_reservation(log, head, tic);
 276         free_bytes = xlog_space_left(log, &head->grant);
 277         if (!list_empty_careful(&head->waiters)) {
 278                 spin_lock(&head->lock);
 279                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 280                     free_bytes < *need_bytes) {
 281                         error = xlog_grant_head_wait(log, head, tic,
 282                                                      *need_bytes);
 283                 }
 284                 spin_unlock(&head->lock);
 285         } else if (free_bytes < *need_bytes) {
 286                 spin_lock(&head->lock);
 287                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 288                 spin_unlock(&head->lock);
 289         }
 290
 291         return error;
 292 }
 293
 294 static void
 295 xlog_tic_reset_res(xlog_ticket_t *tic)
 296 {
 297         tic->t_res_num = 0;
 298         tic->t_res_arr_sum = 0;
 299         tic->t_res_num_ophdrs = 0;
 300 }
 301
 302 static void
 303 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 304 {
 305         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 306                 /* add to overflow and start again */
 307                 tic->t_res_o_flow += tic->t_res_arr_sum;
 308                 tic->t_res_num = 0;
 309                 tic->t_res_arr_sum = 0;
 310         }
 311
 312         tic->t_res_arr[tic->t_res_num].r_len = len;
 313         tic->t_res_arr[tic->t_res_num].r_type = type;
 314         tic->t_res_arr_sum += len;
 315         tic->t_res_num++;
 316 }
 317
 318 /*
 319  * Replenish the byte reservation required by moving the grant write head.
 320  */
 321 int
 322 xfs_log_regrant(
 323         struct xfs_mount        *mp,
 324         struct xlog_ticket      *tic)
 325 {
 326         struct log              *log = mp->m_log;
 327         int                     need_bytes;
 328         int                     error = 0;
 329
 330         if (XLOG_FORCED_SHUTDOWN(log))
 331                 return XFS_ERROR(EIO);
 332
 333         XFS_STATS_INC(xs_try_logspace);
 334
 335         /*
 336          * This is a new transaction on the ticket, so we need to change the
 337          * transaction ID so that the next transaction has a different TID in
 338          * the log. Just add one to the existing tid so that we can see chains
 339          * of rolling transactions in the log easily.
 340          */
 341         tic->t_tid++;
 342
 343         xlog_grant_push_ail(log, tic->t_unit_res);
 344
 345         tic->t_curr_res = tic->t_unit_res;
 346         xlog_tic_reset_res(tic);
 347
 348         if (tic->t_cnt > 0)
 349                 return 0;
 350
 351         trace_xfs_log_regrant(log, tic);
 352
 353         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 354                                       &need_bytes);
 355         if (error)
 356                 goto out_error;
 357
 358         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 359         trace_xfs_log_regrant_exit(log, tic);
 360         xlog_verify_grant_tail(log);
 361         return 0;
 362
 363 out_error:
 364         /*
 365          * If we are failing, make sure the ticket doesn't have any current
 366          * reservations.  We don't want to add this back when the ticket/
 367          * transaction gets cancelled.
 368          */
 369         tic->t_curr_res = 0;
 370         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 371         return error;
 372 }
 373
 374 /*
 375  * Reserve log space and return a ticket corresponding the reservation.
 376  *
 377  * Each reservation is going to reserve extra space for a log record header.
 378  * When writes happen to the on-disk log, we don't subtract the length of the
 379  * log record header from any reservation.  By wasting space in each
 380  * reservation, we prevent over allocation problems.
 381  */
 382 int
 383 xfs_log_reserve(
 384         struct xfs_mount        *mp,
 385         int                     unit_bytes,
 386         int                     cnt,
 387         struct xlog_ticket      **ticp,
 388         __uint8_t               client,
 389         bool                    permanent,
 390         uint                    t_type)
 391 {
 392         struct log              *log = mp->m_log;
 393         struct xlog_ticket      *tic;
 394         int                     need_bytes;
 395         int                     error = 0;
 396
 397         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 398
 399         if (XLOG_FORCED_SHUTDOWN(log))
 400                 return XFS_ERROR(EIO);
 401
 402         XFS_STATS_INC(xs_try_logspace);
 403
 404         ASSERT(*ticp == NULL);
 405         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 406                                 KM_SLEEP | KM_MAYFAIL);
 407         if (!tic)
 408                 return XFS_ERROR(ENOMEM);
 409
 410         tic->t_trans_type = t_type;
 411         *ticp = tic;
 412
 413         xlog_grant_push_ail(log, tic->t_unit_res * tic->t_cnt);
 414
 415         trace_xfs_log_reserve(log, tic);
 416
 417         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 418                                       &need_bytes);
 419         if (error)
 420                 goto out_error;
 421
 422         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 423         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 424         trace_xfs_log_reserve_exit(log, tic);
 425         xlog_verify_grant_tail(log);
 426         return 0;
 427
 428 out_error:
 429         /*
 430          * If we are failing, make sure the ticket doesn't have any current
 431          * reservations.  We don't want to add this back when the ticket/
 432          * transaction gets cancelled.
 433          */
 434         tic->t_curr_res = 0;
 435         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 436         return error;
 437 }
 438
 439
 440 /*
 441  * NOTES:
 442  *
 443  *      1. currblock field gets updated at startup and after in-core logs
 444  *              marked as with WANT_SYNC.
 445  */
 446
 447 /*
 448  * This routine is called when a user of a log manager ticket is done with
 449  * the reservation.  If the ticket was ever used, then a commit record for
 450  * the associated transaction is written out as a log operation header with
 451  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 452  * a given ticket.  If the ticket was one with a permanent reservation, then
 453  * a few operations are done differently.  Permanent reservation tickets by
 454  * default don't release the reservation.  They just commit the current
 455  * transaction with the belief that the reservation is still needed.  A flag
 456  * must be passed in before permanent reservations are actually released.
 457  * When these type of tickets are not released, they need to be set into
 458  * the inited state again.  By doing this, a start record will be written
 459  * out when the next write occurs.
 460  */
 461 xfs_lsn_t
 462 xfs_log_done(
 463         struct xfs_mount        *mp,
 464         struct xlog_ticket      *ticket,
 465         struct xlog_in_core     **iclog,
 466         uint                    flags)
 467 {
 468         struct log              *log = mp->m_log;
 469         xfs_lsn_t               lsn = 0;
 470
 471         if (XLOG_FORCED_SHUTDOWN(log) ||
 472             /*
 473              * If nothing was ever written, don't write out commit record.
 474              * If we get an error, just continue and give back the log ticket.
 475              */
 476             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 477              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 478                 lsn = (xfs_lsn_t) -1;
 479                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
 480                         flags |= XFS_LOG_REL_PERM_RESERV;
 481                 }
 482         }
 483
 484
 485         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
 486             (flags & XFS_LOG_REL_PERM_RESERV)) {
 487                 trace_xfs_log_done_nonperm(log, ticket);
 488
 489                 /*
 490                  * Release ticket if not permanent reservation or a specific
 491                  * request has been made to release a permanent reservation.
 492                  */
 493                 xlog_ungrant_log_space(log, ticket);
 494                 xfs_log_ticket_put(ticket);
 495         } else {
 496                 trace_xfs_log_done_perm(log, ticket);
 497
 498                 xlog_regrant_reserve_log_space(log, ticket);
 499                 /* If this ticket was a permanent reservation and we aren't
 500                  * trying to release it, reset the inited flags; so next time
 501                  * we write, a start record will be written out.
 502                  */
 503                 ticket->t_flags |= XLOG_TIC_INITED;
 504         }
 505
 506         return lsn;
 507 }
 508
 509 /*
 510  * Attaches a new iclog I/O completion callback routine during
 511  * transaction commit.  If the log is in error state, a non-zero
 512  * return code is handed back and the caller is responsible for
 513  * executing the callback at an appropriate time.
 514  */
 515 int
 516 xfs_log_notify(
 517         struct xfs_mount        *mp,
 518         struct xlog_in_core     *iclog,
 519         xfs_log_callback_t      *cb)
 520 {
 521         int     abortflg;
 522
 523         spin_lock(&iclog->ic_callback_lock);
 524         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 525         if (!abortflg) {
 526                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 527                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 528                 cb->cb_next = NULL;
 529                 *(iclog->ic_callback_tail) = cb;
 530                 iclog->ic_callback_tail = &(cb->cb_next);
 531         }
 532         spin_unlock(&iclog->ic_callback_lock);
 533         return abortflg;
 534 }
 535
 536 int
 537 xfs_log_release_iclog(
 538         struct xfs_mount        *mp,
 539         struct xlog_in_core     *iclog)
 540 {
 541         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 542                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 543                 return EIO;
 544         }
 545
 546         return 0;
 547 }
 548
 549 /*
 550  * Mount a log filesystem
 551  *
 552  * mp           - ubiquitous xfs mount point structure
 553  * log_target   - buftarg of on-disk log device
 554  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 555  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 556  *
 557  * Return error or zero.
 558  */
 559 int
 560 xfs_log_mount(
 561         xfs_mount_t     *mp,
 562         xfs_buftarg_t   *log_target,
 563         xfs_daddr_t     blk_offset,
 564         int             num_bblks)
 565 {
 566         int             error;
 567
 568         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
 569                 xfs_notice(mp, "Mounting Filesystem");
 570         else {
 571                 xfs_notice(mp,
 572 "Mounting filesystem in no-recovery mode.  Filesystem will be inconsistent.");
 573                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 574         }
 575
 576         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 577         if (IS_ERR(mp->m_log)) {
 578                 error = -PTR_ERR(mp->m_log);
 579                 goto out;
 580         }
 581
 582         /*
 583          * Initialize the AIL now we have a log.
 584          */
 585         error = xfs_trans_ail_init(mp);
 586         if (error) {
 587                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 588                 goto out_free_log;
 589         }
 590         mp->m_log->l_ailp = mp->m_ail;
 591
 592         /*
 593          * skip log recovery on a norecovery mount.  pretend it all
 594          * just worked.
 595          */
 596         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 597                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 598
 599                 if (readonly)
 600                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 601
 602                 error = xlog_recover(mp->m_log);
 603
 604                 if (readonly)
 605                         mp->m_flags |= XFS_MOUNT_RDONLY;
 606                 if (error) {
 607                         xfs_warn(mp, "log mount/recovery failed: error %d",
 608                                 error);
 609                         goto out_destroy_ail;
 610                 }
 611         }
 612
 613         /* Normal transactions can now occur */
 614         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 615
 616         /*
 617          * Now the log has been fully initialised and we know were our
 618          * space grant counters are, we can initialise the permanent ticket
 619          * needed for delayed logging to work.
 620          */
 621         xlog_cil_init_post_recovery(mp->m_log);
 622
 623         return 0;
 624
 625 out_destroy_ail:
 626         xfs_trans_ail_destroy(mp);
 627 out_free_log:
 628         xlog_dealloc_log(mp->m_log);
 629 out:
 630         return error;
 631 }
 632
 633 /*
 634  * Finish the recovery of the file system.  This is separate from
 635  * the xfs_log_mount() call, because it depends on the code in
 636  * xfs_mountfs() to read in the root and real-time bitmap inodes
 637  * between calling xfs_log_mount() and here.
 638  *
 639  * mp           - ubiquitous xfs mount point structure
 640  */
 641 int
 642 xfs_log_mount_finish(xfs_mount_t *mp)
 643 {
 644         int     error;
 645
 646         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
 647                 error = xlog_recover_finish(mp->m_log);
 648         else {
 649                 error = 0;
 650                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 651         }
 652
 653         return error;
 654 }
 655
 656 /*
 657  * Final log writes as part of unmount.
 658  *
 659  * Mark the filesystem clean as unmount happens.  Note that during relocation
 660  * this routine needs to be executed as part of source-bag while the
 661  * deallocation must not be done until source-end.
 662  */
 663
 664 /*
 665  * Unmount record used to have a string "Unmount filesystem--" in the
 666  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 667  * We just write the magic number now since that particular field isn't
 668  * currently architecture converted and "nUmount" is a bit foo.
 669  * As far as I know, there weren't any dependencies on the old behaviour.
 670  */
 671
 672 int
 673 xfs_log_unmount_write(xfs_mount_t *mp)
 674 {
 675         xlog_t           *log = mp->m_log;
 676         xlog_in_core_t   *iclog;
 677 #ifdef DEBUG
 678         xlog_in_core_t   *first_iclog;
 679 #endif
 680         xlog_ticket_t   *tic = NULL;
 681         xfs_lsn_t        lsn;
 682         int              error;
 683
 684         /*
 685          * Don't write out unmount record on read-only mounts.
 686          * Or, if we are doing a forced umount (typically because of IO errors).
 687          */
 688         if (mp->m_flags & XFS_MOUNT_RDONLY)
 689                 return 0;
 690
 691         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
 692         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 693
 694 #ifdef DEBUG
 695         first_iclog = iclog = log->l_iclog;
 696         do {
 697                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 698                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 699                         ASSERT(iclog->ic_offset == 0);
 700                 }
 701                 iclog = iclog->ic_next;
 702         } while (iclog != first_iclog);
 703 #endif
 704         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 705                 error = xfs_log_reserve(mp, 600, 1, &tic,
 706                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
 707                 if (!error) {
 708                         /* the data section must be 32 bit size aligned */
 709                         struct {
 710                             __uint16_t magic;
 711                             __uint16_t pad1;
 712                             __uint32_t pad2; /* may as well make it 64 bits */
 713                         } magic = {
 714                                 .magic = XLOG_UNMOUNT_TYPE,
 715                         };
 716                         struct xfs_log_iovec reg = {
 717                                 .i_addr = &magic,
 718                                 .i_len = sizeof(magic),
 719                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 720                         };
 721                         struct xfs_log_vec vec = {
 722                                 .lv_niovecs = 1,
 723                                 .lv_iovecp = &reg,
 724                         };
 725
 726                         /* remove inited flag, and account for space used */
 727                         tic->t_flags = 0;
 728                         tic->t_curr_res -= sizeof(magic);
 729                         error = xlog_write(log, &vec, tic, &lsn,
 730                                            NULL, XLOG_UNMOUNT_TRANS);
 731                         /*
 732                          * At this point, we're umounting anyway,
 733                          * so there's no point in transitioning log state
 734                          * to IOERROR. Just continue...
 735                          */
 736                 }
 737
 738                 if (error)
 739                         xfs_alert(mp, "%s: unmount record failed", __func__);
 740
 741
 742                 spin_lock(&log->l_icloglock);
 743                 iclog = log->l_iclog;
 744                 atomic_inc(&iclog->ic_refcnt);
 745                 xlog_state_want_sync(log, iclog);
 746                 spin_unlock(&log->l_icloglock);
 747                 error = xlog_state_release_iclog(log, iclog);
 748
 749                 spin_lock(&log->l_icloglock);
 750                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 751                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 752                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 753                                 xlog_wait(&iclog->ic_force_wait,
 754                                                         &log->l_icloglock);
 755                         } else {
 756                                 spin_unlock(&log->l_icloglock);
 757                         }
 758                 } else {
 759                         spin_unlock(&log->l_icloglock);
 760                 }
 761                 if (tic) {
 762                         trace_xfs_log_umount_write(log, tic);
 763                         xlog_ungrant_log_space(log, tic);
 764                         xfs_log_ticket_put(tic);
 765                 }
 766         } else {
 767                 /*
 768                  * We're already in forced_shutdown mode, couldn't
 769                  * even attempt to write out the unmount transaction.
 770                  *
 771                  * Go through the motions of sync'ing and releasing
 772                  * the iclog, even though no I/O will actually happen,
 773                  * we need to wait for other log I/Os that may already
 774                  * be in progress.  Do this as a separate section of
 775                  * code so we'll know if we ever get stuck here that
 776                  * we're in this odd situation of trying to unmount
 777                  * a file system that went into forced_shutdown as
 778                  * the result of an unmount..
 779                  */
 780                 spin_lock(&log->l_icloglock);
 781                 iclog = log->l_iclog;
 782                 atomic_inc(&iclog->ic_refcnt);
 783
 784                 xlog_state_want_sync(log, iclog);
 785                 spin_unlock(&log->l_icloglock);
 786                 error =  xlog_state_release_iclog(log, iclog);
 787
 788                 spin_lock(&log->l_icloglock);
 789
 790                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 791                         || iclog->ic_state == XLOG_STATE_DIRTY
 792                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 793
 794                                 xlog_wait(&iclog->ic_force_wait,
 795                                                         &log->l_icloglock);
 796                 } else {
 797                         spin_unlock(&log->l_icloglock);
 798                 }
 799         }
 800
 801         return error;
 802 }       /* xfs_log_unmount_write */
 803
 804 /*
 805  * Deallocate log structures for unmount/relocation.
 806  *
 807  * We need to stop the aild from running before we destroy
 808  * and deallocate the log as the aild references the log.
 809  */
 810 void
 811 xfs_log_unmount(xfs_mount_t *mp)
 812 {
 813         xfs_trans_ail_destroy(mp);
 814         xlog_dealloc_log(mp->m_log);
 815 }
 816
 817 void
 818 xfs_log_item_init(
 819         struct xfs_mount        *mp,
 820         struct xfs_log_item     *item,
 821         int                     type,
 822         const struct xfs_item_ops *ops)
 823 {
 824         item->li_mountp = mp;
 825         item->li_ailp = mp->m_ail;
 826         item->li_type = type;
 827         item->li_ops = ops;
 828         item->li_lv = NULL;
 829
 830         INIT_LIST_HEAD(&item->li_ail);
 831         INIT_LIST_HEAD(&item->li_cil);
 832 }
 833
 834 /*
 835  * Wake up processes waiting for log space after we have moved the log tail.
 836  */
 837 void
 838 xfs_log_space_wake(
 839         struct xfs_mount        *mp)
 840 {
 841         struct log              *log = mp->m_log;
 842         int                     free_bytes;
 843
 844         if (XLOG_FORCED_SHUTDOWN(log))
 845                 return;
 846
 847         if (!list_empty_careful(&log->l_write_head.waiters)) {
 848                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 849
 850                 spin_lock(&log->l_write_head.lock);
 851                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
 852                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
 853                 spin_unlock(&log->l_write_head.lock);
 854         }
 855
 856         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
 857                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 858
 859                 spin_lock(&log->l_reserve_head.lock);
 860                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
 861                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
 862                 spin_unlock(&log->l_reserve_head.lock);
 863         }
 864 }
 865
 866 /*
 867  * Determine if we have a transaction that has gone to disk
 868  * that needs to be covered. To begin the transition to the idle state
 869  * firstly the log needs to be idle (no AIL and nothing in the iclogs).
 870  * If we are then in a state where covering is needed, the caller is informed
 871  * that dummy transactions are required to move the log into the idle state.
 872  *
 873  * Because this is called as part of the sync process, we should also indicate
 874  * that dummy transactions should be issued in anything but the covered or
 875  * idle states. This ensures that the log tail is accurately reflected in
 876  * the log at the end of the sync, hence if a crash occurrs avoids replay
 877  * of transactions where the metadata is already on disk.
 878  */
 879 int
 880 xfs_log_need_covered(xfs_mount_t *mp)
 881 {
 882         int             needed = 0;
 883         xlog_t          *log = mp->m_log;
 884
 885         if (!xfs_fs_writable(mp))
 886                 return 0;
 887
 888         spin_lock(&log->l_icloglock);
 889         switch (log->l_covered_state) {
 890         case XLOG_STATE_COVER_DONE:
 891         case XLOG_STATE_COVER_DONE2:
 892         case XLOG_STATE_COVER_IDLE:
 893                 break;
 894         case XLOG_STATE_COVER_NEED:
 895         case XLOG_STATE_COVER_NEED2:
 896                 if (!xfs_ail_min_lsn(log->l_ailp) &&
 897                     xlog_iclogs_empty(log)) {
 898                         if (log->l_covered_state == XLOG_STATE_COVER_NEED)
 899                                 log->l_covered_state = XLOG_STATE_COVER_DONE;
 900                         else
 901                                 log->l_covered_state = XLOG_STATE_COVER_DONE2;
 902                 }
 903                 /* FALLTHRU */
 904         default:
 905                 needed = 1;
 906                 break;
 907         }
 908         spin_unlock(&log->l_icloglock);
 909         return needed;
 910 }
 911
 912 /*
 913  * We may be holding the log iclog lock upon entering this routine.
 914  */
 915 xfs_lsn_t
 916 xlog_assign_tail_lsn_locked(
 917         struct xfs_mount        *mp)
 918 {
 919         struct log              *log = mp->m_log;
 920         struct xfs_log_item     *lip;
 921         xfs_lsn_t               tail_lsn;
 922
 923         assert_spin_locked(&mp->m_ail->xa_lock);
 924
 925         /*
 926          * To make sure we always have a valid LSN for the log tail we keep
 927          * track of the last LSN which was committed in log->l_last_sync_lsn,
 928          * and use that when the AIL was empty.
 929          */
 930         lip = xfs_ail_min(mp->m_ail);
 931         if (lip)
 932                 tail_lsn = lip->li_lsn;
 933         else
 934                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
 935         atomic64_set(&log->l_tail_lsn, tail_lsn);
 936         return tail_lsn;
 937 }
 938
 939 xfs_lsn_t
 940 xlog_assign_tail_lsn(
 941         struct xfs_mount        *mp)
 942 {
 943         xfs_lsn_t               tail_lsn;
 944
 945         spin_lock(&mp->m_ail->xa_lock);
 946         tail_lsn = xlog_assign_tail_lsn_locked(mp);
 947         spin_unlock(&mp->m_ail->xa_lock);
 948
 949         return tail_lsn;
 950 }
 951
 952 /*
 953  * Return the space in the log between the tail and the head.  The head
 954  * is passed in the cycle/bytes formal parms.  In the special case where
 955  * the reserve head has wrapped passed the tail, this calculation is no
 956  * longer valid.  In this case, just return 0 which means there is no space
 957  * in the log.  This works for all places where this function is called
 958  * with the reserve head.  Of course, if the write head were to ever
 959  * wrap the tail, we should blow up.  Rather than catch this case here,
 960  * we depend on other ASSERTions in other parts of the code.   XXXmiken
 961  *
 962  * This code also handles the case where the reservation head is behind
 963  * the tail.  The details of this case are described below, but the end
 964  * result is that we return the size of the log as the amount of space left.
 965  */
 966 STATIC int
 967 xlog_space_left(
 968         struct log      *log,
 969         atomic64_t      *head)
 970 {
 971         int             free_bytes;
 972         int             tail_bytes;
 973         int             tail_cycle;
 974         int             head_cycle;
 975         int             head_bytes;
 976
 977         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
 978         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
 979         tail_bytes = BBTOB(tail_bytes);
 980         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
 981                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
 982         else if (tail_cycle + 1 < head_cycle)
 983                 return 0;
 984         else if (tail_cycle < head_cycle) {
 985                 ASSERT(tail_cycle == (head_cycle - 1));
 986                 free_bytes = tail_bytes - head_bytes;
 987         } else {
 988                 /*
 989                  * The reservation head is behind the tail.
 990                  * In this case we just want to return the size of the
 991                  * log as the amount of space left.
 992                  */
 993                 xfs_alert(log->l_mp,
 994                         "xlog_space_left: head behind tail\n"
 995                         "  tail_cycle = %d, tail_bytes = %d\n"
 996                         "  GH   cycle = %d, GH   bytes = %d",
 997                         tail_cycle, tail_bytes, head_cycle, head_bytes);
 998                 ASSERT(0);
 999                 free_bytes = log->l_logsize;
1000         }
1001         return free_bytes;
1002 }
1003
1004
1005 /*
1006  * Log function which is called when an io completes.
1007  *
1008  * The log manager needs its own routine, in order to control what
1009  * happens with the buffer after the write completes.
1010  */
1011 void
1012 xlog_iodone(xfs_buf_t *bp)
1013 {
1014         xlog_in_core_t  *iclog = bp->b_fspriv;
1015         xlog_t          *l = iclog->ic_log;
1016         int             aborted = 0;
1017
1018         /*
1019          * Race to shutdown the filesystem if we see an error.
1020          */
1021         if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1022                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1023                 xfs_buf_ioerror_alert(bp, __func__);
1024                 xfs_buf_stale(bp);
1025                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1026                 /*
1027                  * This flag will be propagated to the trans-committed
1028                  * callback routines to let them know that the log-commit
1029                  * didn't succeed.
1030                  */
1031                 aborted = XFS_LI_ABORTED;
1032         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1033                 aborted = XFS_LI_ABORTED;
1034         }
1035
1036         /* log I/O is always issued ASYNC */
1037         ASSERT(XFS_BUF_ISASYNC(bp));
1038         xlog_state_done_syncing(iclog, aborted);
1039         /*
1040          * do not reference the buffer (bp) here as we could race
1041          * with it being freed after writing the unmount record to the
1042          * log.
1043          */
1044
1045 }       /* xlog_iodone */
1046
1047 /*
1048  * Return size of each in-core log record buffer.
1049  *
1050  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1051  *
1052  * If the filesystem blocksize is too large, we may need to choose a
1053  * larger size since the directory code currently logs entire blocks.
1054  */
1055
1056 STATIC void
1057 xlog_get_iclog_buffer_size(xfs_mount_t  *mp,
1058                            xlog_t       *log)
1059 {
1060         int size;
1061         int xhdrs;
1062
1063         if (mp->m_logbufs <= 0)
1064                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1065         else
1066                 log->l_iclog_bufs = mp->m_logbufs;
1067
1068         /*
1069          * Buffer size passed in from mount system call.
1070          */
1071         if (mp->m_logbsize > 0) {
1072                 size = log->l_iclog_size = mp->m_logbsize;
1073                 log->l_iclog_size_log = 0;
1074                 while (size != 1) {
1075                         log->l_iclog_size_log++;
1076                         size >>= 1;
1077                 }
1078
1079                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1080                         /* # headers = size / 32k
1081                          * one header holds cycles from 32k of data
1082                          */
1083
1084                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1085                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1086                                 xhdrs++;
1087                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1088                         log->l_iclog_heads = xhdrs;
1089                 } else {
1090                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1091                         log->l_iclog_hsize = BBSIZE;
1092                         log->l_iclog_heads = 1;
1093                 }
1094                 goto done;
1095         }
1096
1097         /* All machines use 32kB buffers by default. */
1098         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1099         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1100
1101         /* the default log size is 16k or 32k which is one header sector */
1102         log->l_iclog_hsize = BBSIZE;
1103         log->l_iclog_heads = 1;
1104
1105 done:
1106         /* are we being asked to make the sizes selected above visible? */
1107         if (mp->m_logbufs == 0)
1108                 mp->m_logbufs = log->l_iclog_bufs;
1109         if (mp->m_logbsize == 0)
1110                 mp->m_logbsize = log->l_iclog_size;
1111 }       /* xlog_get_iclog_buffer_size */
1112
1113
1114 /*
1115  * This routine initializes some of the log structure for a given mount point.
1116  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1117  * some other stuff may be filled in too.
1118  */
1119 STATIC xlog_t *
1120 xlog_alloc_log(xfs_mount_t      *mp,
1121                xfs_buftarg_t    *log_target,
1122                xfs_daddr_t      blk_offset,
1123                int              num_bblks)
1124 {
1125         xlog_t                  *log;
1126         xlog_rec_header_t       *head;
1127         xlog_in_core_t          **iclogp;
1128         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1129         xfs_buf_t               *bp;
1130         int                     i;
1131         int                     error = ENOMEM;
1132         uint                    log2_size = 0;
1133
1134         log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1135         if (!log) {
1136                 xfs_warn(mp, "Log allocation failed: No memory!");
1137                 goto out;
1138         }
1139
1140         log->l_mp          = mp;
1141         log->l_targ        = log_target;
1142         log->l_logsize     = BBTOB(num_bblks);
1143         log->l_logBBstart  = blk_offset;
1144         log->l_logBBsize   = num_bblks;
1145         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1146         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1147
1148         log->l_prev_block  = -1;
1149         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1150         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1151         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1152         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1153
1154         xlog_grant_head_init(&log->l_reserve_head);
1155         xlog_grant_head_init(&log->l_write_head);
1156
1157         error = EFSCORRUPTED;
1158         if (xfs_sb_version_hassector(&mp->m_sb)) {
1159                 log2_size = mp->m_sb.sb_logsectlog;
1160                 if (log2_size < BBSHIFT) {
1161                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1162                                 log2_size, BBSHIFT);
1163                         goto out_free_log;
1164                 }
1165
1166                 log2_size -= BBSHIFT;
1167                 if (log2_size > mp->m_sectbb_log) {
1168                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1169                                 log2_size, mp->m_sectbb_log);
1170                         goto out_free_log;
1171                 }
1172
1173                 /* for larger sector sizes, must have v2 or external log */
1174                 if (log2_size && log->l_logBBstart > 0 &&
1175                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1176                         xfs_warn(mp,
1177                 "log sector size (0x%x) invalid for configuration.",
1178                                 log2_size);
1179                         goto out_free_log;
1180                 }
1181         }
1182         log->l_sectBBsize = 1 << log2_size;
1183
1184         xlog_get_iclog_buffer_size(mp, log);
1185
1186         error = ENOMEM;
1187         bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1188         if (!bp)
1189                 goto out_free_log;
1190         bp->b_iodone = xlog_iodone;
1191         ASSERT(xfs_buf_islocked(bp));
1192         log->l_xbuf = bp;
1193
1194         spin_lock_init(&log->l_icloglock);
1195         init_waitqueue_head(&log->l_flush_wait);
1196
1197         iclogp = &log->l_iclog;
1198         /*
1199          * The amount of memory to allocate for the iclog structure is
1200          * rather funky due to the way the structure is defined.  It is
1201          * done this way so that we can use different sizes for machines
1202          * with different amounts of memory.  See the definition of
1203          * xlog_in_core_t in xfs_log_priv.h for details.
1204          */
1205         ASSERT(log->l_iclog_size >= 4096);
1206         for (i=0; i < log->l_iclog_bufs; i++) {
1207                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1208                 if (!*iclogp)
1209                         goto out_free_iclog;
1210
1211                 iclog = *iclogp;
1212                 iclog->ic_prev = prev_iclog;
1213                 prev_iclog = iclog;
1214
1215                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1216                                                 BTOBB(log->l_iclog_size), 0);
1217                 if (!bp)
1218                         goto out_free_iclog;
1219
1220                 bp->b_iodone = xlog_iodone;
1221                 iclog->ic_bp = bp;
1222                 iclog->ic_data = bp->b_addr;
1223 #ifdef DEBUG
1224                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1225 #endif
1226                 head = &iclog->ic_header;
1227                 memset(head, 0, sizeof(xlog_rec_header_t));
1228                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1229                 head->h_version = cpu_to_be32(
1230                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1231                 head->h_size = cpu_to_be32(log->l_iclog_size);
1232                 /* new fields */
1233                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1234                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1235
1236                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1237                 iclog->ic_state = XLOG_STATE_ACTIVE;
1238                 iclog->ic_log = log;
1239                 atomic_set(&iclog->ic_refcnt, 0);
1240                 spin_lock_init(&iclog->ic_callback_lock);
1241                 iclog->ic_callback_tail = &(iclog->ic_callback);
1242                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1243
1244                 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1245                 init_waitqueue_head(&iclog->ic_force_wait);
1246                 init_waitqueue_head(&iclog->ic_write_wait);
1247
1248                 iclogp = &iclog->ic_next;
1249         }
1250         *iclogp = log->l_iclog;                 /* complete ring */
1251         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1252
1253         error = xlog_cil_init(log);
1254         if (error)
1255                 goto out_free_iclog;
1256         return log;
1257
1258 out_free_iclog:
1259         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1260                 prev_iclog = iclog->ic_next;
1261                 if (iclog->ic_bp)
1262                         xfs_buf_free(iclog->ic_bp);
1263                 kmem_free(iclog);
1264         }
1265         spinlock_destroy(&log->l_icloglock);
1266         xfs_buf_free(log->l_xbuf);
1267 out_free_log:
1268         kmem_free(log);
1269 out:
1270         return ERR_PTR(-error);
1271 }       /* xlog_alloc_log */
1272
1273
1274 /*
1275  * Write out the commit record of a transaction associated with the given
1276  * ticket.  Return the lsn of the commit record.
1277  */
1278 STATIC int
1279 xlog_commit_record(
1280         struct log              *log,
1281         struct xlog_ticket      *ticket,
1282         struct xlog_in_core     **iclog,
1283         xfs_lsn_t               *commitlsnp)
1284 {
1285         struct xfs_mount *mp = log->l_mp;
1286         int     error;
1287         struct xfs_log_iovec reg = {
1288                 .i_addr = NULL,
1289                 .i_len = 0,
1290                 .i_type = XLOG_REG_TYPE_COMMIT,
1291         };
1292         struct xfs_log_vec vec = {
1293                 .lv_niovecs = 1,
1294                 .lv_iovecp = &reg,
1295         };
1296
1297         ASSERT_ALWAYS(iclog);
1298         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1299                                         XLOG_COMMIT_TRANS);
1300         if (error)
1301                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1302         return error;
1303 }
1304
1305 /*
1306  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1307  * log space.  This code pushes on the lsn which would supposedly free up
1308  * the 25% which we want to leave free.  We may need to adopt a policy which
1309  * pushes on an lsn which is further along in the log once we reach the high
1310  * water mark.  In this manner, we would be creating a low water mark.
1311  */
1312 STATIC void
1313 xlog_grant_push_ail(
1314         struct log      *log,
1315         int             need_bytes)
1316 {
1317         xfs_lsn_t       threshold_lsn = 0;
1318         xfs_lsn_t       last_sync_lsn;
1319         int             free_blocks;
1320         int             free_bytes;
1321         int             threshold_block;
1322         int             threshold_cycle;
1323         int             free_threshold;
1324
1325         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1326
1327         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1328         free_blocks = BTOBBT(free_bytes);
1329
1330         /*
1331          * Set the threshold for the minimum number of free blocks in the
1332          * log to the maximum of what the caller needs, one quarter of the
1333          * log, and 256 blocks.
1334          */
1335         free_threshold = BTOBB(need_bytes);
1336         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1337         free_threshold = MAX(free_threshold, 256);
1338         if (free_blocks >= free_threshold)
1339                 return;
1340
1341         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1342                                                 &threshold_block);
1343         threshold_block += free_threshold;
1344         if (threshold_block >= log->l_logBBsize) {
1345                 threshold_block -= log->l_logBBsize;
1346                 threshold_cycle += 1;
1347         }
1348         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1349                                         threshold_block);
1350         /*
1351          * Don't pass in an lsn greater than the lsn of the last
1352          * log record known to be on disk. Use a snapshot of the last sync lsn
1353          * so that it doesn't change between the compare and the set.
1354          */
1355         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1356         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1357                 threshold_lsn = last_sync_lsn;
1358
1359         /*
1360          * Get the transaction layer to kick the dirty buffers out to
1361          * disk asynchronously. No point in trying to do this if
1362          * the filesystem is shutting down.
1363          */
1364         if (!XLOG_FORCED_SHUTDOWN(log))
1365                 xfs_ail_push(log->l_ailp, threshold_lsn);
1366 }
1367
1368 /*
1369  * The bdstrat callback function for log bufs. This gives us a central
1370  * place to trap bufs in case we get hit by a log I/O error and need to
1371  * shutdown. Actually, in practice, even when we didn't get a log error,
1372  * we transition the iclogs to IOERROR state *after* flushing all existing
1373  * iclogs to disk. This is because we don't want anymore new transactions to be
1374  * started or completed afterwards.
1375  */
1376 STATIC int
1377 xlog_bdstrat(
1378         struct xfs_buf          *bp)
1379 {
1380         struct xlog_in_core     *iclog = bp->b_fspriv;
1381
1382         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1383                 xfs_buf_ioerror(bp, EIO);
1384                 xfs_buf_stale(bp);
1385                 xfs_buf_ioend(bp, 0);
1386                 /*
1387                  * It would seem logical to return EIO here, but we rely on
1388                  * the log state machine to propagate I/O errors instead of
1389                  * doing it here.
1390                  */
1391                 return 0;
1392         }
1393
1394         xfs_buf_iorequest(bp);
1395         return 0;
1396 }
1397
1398 /*
1399  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1400  * fashion.  Previously, we should have moved the current iclog
1401  * ptr in the log to point to the next available iclog.  This allows further
1402  * write to continue while this code syncs out an iclog ready to go.
1403  * Before an in-core log can be written out, the data section must be scanned
1404  * to save away the 1st word of each BBSIZE block into the header.  We replace
1405  * it with the current cycle count.  Each BBSIZE block is tagged with the
1406  * cycle count because there in an implicit assumption that drives will
1407  * guarantee that entire 512 byte blocks get written at once.  In other words,
1408  * we can't have part of a 512 byte block written and part not written.  By
1409  * tagging each block, we will know which blocks are valid when recovering
1410  * after an unclean shutdown.
1411  *
1412  * This routine is single threaded on the iclog.  No other thread can be in
1413  * this routine with the same iclog.  Changing contents of iclog can there-
1414  * fore be done without grabbing the state machine lock.  Updating the global
1415  * log will require grabbing the lock though.
1416  *
1417  * The entire log manager uses a logical block numbering scheme.  Only
1418  * log_sync (and then only bwrite()) know about the fact that the log may
1419  * not start with block zero on a given device.  The log block start offset
1420  * is added immediately before calling bwrite().
1421  */
1422
1423 STATIC int
1424 xlog_sync(xlog_t                *log,
1425           xlog_in_core_t        *iclog)
1426 {
1427         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1428         xfs_buf_t       *bp;
1429         int             i;
1430         uint            count;          /* byte count of bwrite */
1431         uint            count_init;     /* initial count before roundup */
1432         int             roundoff;       /* roundoff to BB or stripe */
1433         int             split = 0;      /* split write into two regions */
1434         int             error;
1435         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1436
1437         XFS_STATS_INC(xs_log_writes);
1438         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1439
1440         /* Add for LR header */
1441         count_init = log->l_iclog_hsize + iclog->ic_offset;
1442
1443         /* Round out the log write size */
1444         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1445                 /* we have a v2 stripe unit to use */
1446                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1447         } else {
1448                 count = BBTOB(BTOBB(count_init));
1449         }
1450         roundoff = count - count_init;
1451         ASSERT(roundoff >= 0);
1452         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1453                 roundoff < log->l_mp->m_sb.sb_logsunit)
1454                 ||
1455                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1456                  roundoff < BBTOB(1)));
1457
1458         /* move grant heads by roundoff in sync */
1459         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1460         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1461
1462         /* put cycle number in every block */
1463         xlog_pack_data(log, iclog, roundoff);
1464
1465         /* real byte length */
1466         if (v2) {
1467                 iclog->ic_header.h_len =
1468                         cpu_to_be32(iclog->ic_offset + roundoff);
1469         } else {
1470                 iclog->ic_header.h_len =
1471                         cpu_to_be32(iclog->ic_offset);
1472         }
1473
1474         bp = iclog->ic_bp;
1475         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1476
1477         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1478
1479         /* Do we need to split this write into 2 parts? */
1480         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1481                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1482                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1483                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1484         } else {
1485                 iclog->ic_bwritecnt = 1;
1486         }
1487         bp->b_io_length = BTOBB(count);
1488         bp->b_fspriv = iclog;
1489         XFS_BUF_ZEROFLAGS(bp);
1490         XFS_BUF_ASYNC(bp);
1491         bp->b_flags |= XBF_SYNCIO;
1492
1493         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1494                 bp->b_flags |= XBF_FUA;
1495
1496                 /*
1497                  * Flush the data device before flushing the log to make
1498                  * sure all meta data written back from the AIL actually made
1499                  * it to disk before stamping the new log tail LSN into the
1500                  * log buffer.  For an external log we need to issue the
1501                  * flush explicitly, and unfortunately synchronously here;
1502                  * for an internal log we can simply use the block layer
1503                  * state machine for preflushes.
1504                  */
1505                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1506                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1507                 else
1508                         bp->b_flags |= XBF_FLUSH;
1509         }
1510
1511         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1512         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1513
1514         xlog_verify_iclog(log, iclog, count, B_TRUE);
1515
1516         /* account for log which doesn't start at block #0 */
1517         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1518         /*
1519          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1520          * is shutting down.
1521          */
1522         XFS_BUF_WRITE(bp);
1523
1524         error = xlog_bdstrat(bp);
1525         if (error) {
1526                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1527                 return error;
1528         }
1529         if (split) {
1530                 bp = iclog->ic_log->l_xbuf;
1531                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1532                 xfs_buf_associate_memory(bp,
1533                                 (char *)&iclog->ic_header + count, split);
1534                 bp->b_fspriv = iclog;
1535                 XFS_BUF_ZEROFLAGS(bp);
1536                 XFS_BUF_ASYNC(bp);
1537                 bp->b_flags |= XBF_SYNCIO;
1538                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1539                         bp->b_flags |= XBF_FUA;
1540                 dptr = bp->b_addr;
1541                 /*
1542                  * Bump the cycle numbers at the start of each block
1543                  * since this part of the buffer is at the start of
1544                  * a new cycle.  Watch out for the header magic number
1545                  * case, though.
1546                  */
1547                 for (i = 0; i < split; i += BBSIZE) {
1548                         be32_add_cpu((__be32 *)dptr, 1);
1549                         if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1550                                 be32_add_cpu((__be32 *)dptr, 1);
1551                         dptr += BBSIZE;
1552                 }
1553
1554                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1555                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1556
1557                 /* account for internal log which doesn't start at block #0 */
1558                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1559                 XFS_BUF_WRITE(bp);
1560                 error = xlog_bdstrat(bp);
1561                 if (error) {
1562                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1563                         return error;
1564                 }
1565         }
1566         return 0;
1567 }       /* xlog_sync */
1568
1569
1570 /*
1571  * Deallocate a log structure
1572  */
1573 STATIC void
1574 xlog_dealloc_log(xlog_t *log)
1575 {
1576         xlog_in_core_t  *iclog, *next_iclog;
1577         int             i;
1578
1579         xlog_cil_destroy(log);
1580
1581         /*
1582          * always need to ensure that the extra buffer does not point to memory
1583          * owned by another log buffer before we free it.
1584          */
1585         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1586         xfs_buf_free(log->l_xbuf);
1587
1588         iclog = log->l_iclog;
1589         for (i=0; i<log->l_iclog_bufs; i++) {
1590                 xfs_buf_free(iclog->ic_bp);
1591                 next_iclog = iclog->ic_next;
1592                 kmem_free(iclog);
1593                 iclog = next_iclog;
1594         }
1595         spinlock_destroy(&log->l_icloglock);
1596
1597         log->l_mp->m_log = NULL;
1598         kmem_free(log);
1599 }       /* xlog_dealloc_log */
1600
1601 /*
1602  * Update counters atomically now that memcpy is done.
1603  */
1604 /* ARGSUSED */
1605 static inline void
1606 xlog_state_finish_copy(xlog_t           *log,
1607                        xlog_in_core_t   *iclog,
1608                        int              record_cnt,
1609                        int              copy_bytes)
1610 {
1611         spin_lock(&log->l_icloglock);
1612
1613         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1614         iclog->ic_offset += copy_bytes;
1615
1616         spin_unlock(&log->l_icloglock);
1617 }       /* xlog_state_finish_copy */
1618
1619
1620
1621
1622 /*
1623  * print out info relating to regions written which consume
1624  * the reservation
1625  */
1626 void
1627 xlog_print_tic_res(
1628         struct xfs_mount        *mp,
1629         struct xlog_ticket      *ticket)
1630 {
1631         uint i;
1632         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1633
1634         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1635         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1636             "bformat",
1637             "bchunk",
1638             "efi_format",
1639             "efd_format",
1640             "iformat",
1641             "icore",
1642             "iext",
1643             "ibroot",
1644             "ilocal",
1645             "iattr_ext",
1646             "iattr_broot",
1647             "iattr_local",
1648             "qformat",
1649             "dquot",
1650             "quotaoff",
1651             "LR header",
1652             "unmount",
1653             "commit",
1654             "trans header"
1655         };
1656         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1657             "SETATTR_NOT_SIZE",
1658             "SETATTR_SIZE",
1659             "INACTIVE",
1660             "CREATE",
1661             "CREATE_TRUNC",
1662             "TRUNCATE_FILE",
1663             "REMOVE",
1664             "LINK",
1665             "RENAME",
1666             "MKDIR",
1667             "RMDIR",
1668             "SYMLINK",
1669             "SET_DMATTRS",
1670             "GROWFS",
1671             "STRAT_WRITE",
1672             "DIOSTRAT",
1673             "WRITE_SYNC",
1674             "WRITEID",
1675             "ADDAFORK",
1676             "ATTRINVAL",
1677             "ATRUNCATE",
1678             "ATTR_SET",
1679             "ATTR_RM",
1680             "ATTR_FLAG",
1681             "CLEAR_AGI_BUCKET",
1682             "QM_SBCHANGE",
1683             "DUMMY1",
1684             "DUMMY2",
1685             "QM_QUOTAOFF",
1686             "QM_DQALLOC",
1687             "QM_SETQLIM",
1688             "QM_DQCLUSTER",
1689             "QM_QINOCREATE",
1690             "QM_QUOTAOFF_END",
1691             "SB_UNIT",
1692             "FSYNC_TS",
1693             "GROWFSRT_ALLOC",
1694             "GROWFSRT_ZERO",
1695             "GROWFSRT_FREE",
1696             "SWAPEXT"
1697         };
1698
1699         xfs_warn(mp,
1700                 "xlog_write: reservation summary:\n"
1701                 "  trans type  = %s (%u)\n"
1702                 "  unit res    = %d bytes\n"
1703                 "  current res = %d bytes\n"
1704                 "  total reg   = %u bytes (o/flow = %u bytes)\n"
1705                 "  ophdrs      = %u (ophdr space = %u bytes)\n"
1706                 "  ophdr + reg = %u bytes\n"
1707                 "  num regions = %u\n",
1708                 ((ticket->t_trans_type <= 0 ||
1709                   ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1710                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1711                 ticket->t_trans_type,
1712                 ticket->t_unit_res,
1713                 ticket->t_curr_res,
1714                 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1715                 ticket->t_res_num_ophdrs, ophdr_spc,
1716                 ticket->t_res_arr_sum +
1717                 ticket->t_res_o_flow + ophdr_spc,
1718                 ticket->t_res_num);
1719
1720         for (i = 0; i < ticket->t_res_num; i++) {
1721                 uint r_type = ticket->t_res_arr[i].r_type;
1722                 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1723                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1724                             "bad-rtype" : res_type_str[r_type-1]),
1725                             ticket->t_res_arr[i].r_len);
1726         }
1727
1728         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1729                 "xlog_write: reservation ran out. Need to up reservation");
1730         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1731 }
1732
1733 /*
1734  * Calculate the potential space needed by the log vector.  Each region gets
1735  * its own xlog_op_header_t and may need to be double word aligned.
1736  */
1737 static int
1738 xlog_write_calc_vec_length(
1739         struct xlog_ticket      *ticket,
1740         struct xfs_log_vec      *log_vector)
1741 {
1742         struct xfs_log_vec      *lv;
1743         int                     headers = 0;
1744         int                     len = 0;
1745         int                     i;
1746
1747         /* acct for start rec of xact */
1748         if (ticket->t_flags & XLOG_TIC_INITED)
1749                 headers++;
1750
1751         for (lv = log_vector; lv; lv = lv->lv_next) {
1752                 headers += lv->lv_niovecs;
1753
1754                 for (i = 0; i < lv->lv_niovecs; i++) {
1755                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
1756
1757                         len += vecp->i_len;
1758                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1759                 }
1760         }
1761
1762         ticket->t_res_num_ophdrs += headers;
1763         len += headers * sizeof(struct xlog_op_header);
1764
1765         return len;
1766 }
1767
1768 /*
1769  * If first write for transaction, insert start record  We can't be trying to
1770  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
1771  */
1772 static int
1773 xlog_write_start_rec(
1774         struct xlog_op_header   *ophdr,
1775         struct xlog_ticket      *ticket)
1776 {
1777         if (!(ticket->t_flags & XLOG_TIC_INITED))
1778                 return 0;
1779
1780         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
1781         ophdr->oh_clientid = ticket->t_clientid;
1782         ophdr->oh_len = 0;
1783         ophdr->oh_flags = XLOG_START_TRANS;
1784         ophdr->oh_res2 = 0;
1785
1786         ticket->t_flags &= ~XLOG_TIC_INITED;
1787
1788         return sizeof(struct xlog_op_header);
1789 }
1790
1791 static xlog_op_header_t *
1792 xlog_write_setup_ophdr(
1793         struct log              *log,
1794         struct xlog_op_header   *ophdr,
1795         struct xlog_ticket      *ticket,
1796         uint                    flags)
1797 {
1798         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1799         ophdr->oh_clientid = ticket->t_clientid;
1800         ophdr->oh_res2 = 0;
1801
1802         /* are we copying a commit or unmount record? */
1803         ophdr->oh_flags = flags;
1804
1805         /*
1806          * We've seen logs corrupted with bad transaction client ids.  This
1807          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
1808          * and shut down the filesystem.
1809          */
1810         switch (ophdr->oh_clientid)  {
1811         case XFS_TRANSACTION:
1812         case XFS_VOLUME:
1813         case XFS_LOG:
1814                 break;
1815         default:
1816                 xfs_warn(log->l_mp,
1817                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1818                         ophdr->oh_clientid, ticket);
1819                 return NULL;
1820         }
1821
1822         return ophdr;
1823 }
1824
1825 /*
1826  * Set up the parameters of the region copy into the log. This has
1827  * to handle region write split across multiple log buffers - this
1828  * state is kept external to this function so that this code can
1829  * can be written in an obvious, self documenting manner.
1830  */
1831 static int
1832 xlog_write_setup_copy(
1833         struct xlog_ticket      *ticket,
1834         struct xlog_op_header   *ophdr,
1835         int                     space_available,
1836         int                     space_required,
1837         int                     *copy_off,
1838         int                     *copy_len,
1839         int                     *last_was_partial_copy,
1840         int                     *bytes_consumed)
1841 {
1842         int                     still_to_copy;
1843
1844         still_to_copy = space_required - *bytes_consumed;
1845         *copy_off = *bytes_consumed;
1846
1847         if (still_to_copy <= space_available) {
1848                 /* write of region completes here */
1849                 *copy_len = still_to_copy;
1850                 ophdr->oh_len = cpu_to_be32(*copy_len);
1851                 if (*last_was_partial_copy)
1852                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1853                 *last_was_partial_copy = 0;
1854                 *bytes_consumed = 0;
1855                 return 0;
1856         }
1857
1858         /* partial write of region, needs extra log op header reservation */
1859         *copy_len = space_available;
1860         ophdr->oh_len = cpu_to_be32(*copy_len);
1861         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1862         if (*last_was_partial_copy)
1863                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1864         *bytes_consumed += *copy_len;
1865         (*last_was_partial_copy)++;
1866
1867         /* account for new log op header */
1868         ticket->t_curr_res -= sizeof(struct xlog_op_header);
1869         ticket->t_res_num_ophdrs++;
1870
1871         return sizeof(struct xlog_op_header);
1872 }
1873
1874 static int
1875 xlog_write_copy_finish(
1876         struct log              *log,
1877         struct xlog_in_core     *iclog,
1878         uint                    flags,
1879         int                     *record_cnt,
1880         int                     *data_cnt,
1881         int                     *partial_copy,
1882         int                     *partial_copy_len,
1883         int                     log_offset,
1884         struct xlog_in_core     **commit_iclog)
1885 {
1886         if (*partial_copy) {
1887                 /*
1888                  * This iclog has already been marked WANT_SYNC by
1889                  * xlog_state_get_iclog_space.
1890                  */
1891                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1892                 *record_cnt = 0;
1893                 *data_cnt = 0;
1894                 return xlog_state_release_iclog(log, iclog);
1895         }
1896
1897         *partial_copy = 0;
1898         *partial_copy_len = 0;
1899
1900         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1901                 /* no more space in this iclog - push it. */
1902                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1903                 *record_cnt = 0;
1904                 *data_cnt = 0;
1905
1906                 spin_lock(&log->l_icloglock);
1907                 xlog_state_want_sync(log, iclog);
1908                 spin_unlock(&log->l_icloglock);
1909
1910                 if (!commit_iclog)
1911                         return xlog_state_release_iclog(log, iclog);
1912                 ASSERT(flags & XLOG_COMMIT_TRANS);
1913                 *commit_iclog = iclog;
1914         }
1915
1916         return 0;
1917 }
1918
1919 /*
1920  * Write some region out to in-core log
1921  *
1922  * This will be called when writing externally provided regions or when
1923  * writing out a commit record for a given transaction.
1924  *
1925  * General algorithm:
1926  *      1. Find total length of this write.  This may include adding to the
1927  *              lengths passed in.
1928  *      2. Check whether we violate the tickets reservation.
1929  *      3. While writing to this iclog
1930  *          A. Reserve as much space in this iclog as can get
1931  *          B. If this is first write, save away start lsn
1932  *          C. While writing this region:
1933  *              1. If first write of transaction, write start record
1934  *              2. Write log operation header (header per region)
1935  *              3. Find out if we can fit entire region into this iclog
1936  *              4. Potentially, verify destination memcpy ptr
1937  *              5. Memcpy (partial) region
1938  *              6. If partial copy, release iclog; otherwise, continue
1939  *                      copying more regions into current iclog
1940  *      4. Mark want sync bit (in simulation mode)
1941  *      5. Release iclog for potential flush to on-disk log.
1942  *
1943  * ERRORS:
1944  * 1.   Panic if reservation is overrun.  This should never happen since
1945  *      reservation amounts are generated internal to the filesystem.
1946  * NOTES:
1947  * 1. Tickets are single threaded data structures.
1948  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1949  *      syncing routine.  When a single log_write region needs to span
1950  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1951  *      on all log operation writes which don't contain the end of the
1952  *      region.  The XLOG_END_TRANS bit is used for the in-core log
1953  *      operation which contains the end of the continued log_write region.
1954  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1955  *      we don't really know exactly how much space will be used.  As a result,
1956  *      we don't update ic_offset until the end when we know exactly how many
1957  *      bytes have been written out.
1958  */
1959 int
1960 xlog_write(
1961         struct log              *log,
1962         struct xfs_log_vec      *log_vector,
1963         struct xlog_ticket      *ticket,
1964         xfs_lsn_t               *start_lsn,
1965         struct xlog_in_core     **commit_iclog,
1966         uint                    flags)
1967 {
1968         struct xlog_in_core     *iclog = NULL;
1969         struct xfs_log_iovec    *vecp;
1970         struct xfs_log_vec      *lv;
1971         int                     len;
1972         int                     index;
1973         int                     partial_copy = 0;
1974         int                     partial_copy_len = 0;
1975         int                     contwr = 0;
1976         int                     record_cnt = 0;
1977         int                     data_cnt = 0;
1978         int                     error;
1979
1980         *start_lsn = 0;
1981
1982         len = xlog_write_calc_vec_length(ticket, log_vector);
1983
1984         /*
1985          * Region headers and bytes are already accounted for.
1986          * We only need to take into account start records and
1987          * split regions in this function.
1988          */
1989         if (ticket->t_flags & XLOG_TIC_INITED)
1990                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1991
1992         /*
1993          * Commit record headers need to be accounted for. These
1994          * come in as separate writes so are easy to detect.
1995          */
1996         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
1997                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1998
1999         if (ticket->t_curr_res < 0)
2000                 xlog_print_tic_res(log->l_mp, ticket);
2001
2002         index = 0;
2003         lv = log_vector;
2004         vecp = lv->lv_iovecp;
2005         while (lv && index < lv->lv_niovecs) {
2006                 void            *ptr;
2007                 int             log_offset;
2008
2009                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2010                                                    &contwr, &log_offset);
2011                 if (error)
2012                         return error;
2013
2014                 ASSERT(log_offset <= iclog->ic_size - 1);
2015                 ptr = iclog->ic_datap + log_offset;
2016
2017                 /* start_lsn is the first lsn written to. That's all we need. */
2018                 if (!*start_lsn)
2019                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2020
2021                 /*
2022                  * This loop writes out as many regions as can fit in the amount
2023                  * of space which was allocated by xlog_state_get_iclog_space().
2024                  */
2025                 while (lv && index < lv->lv_niovecs) {
2026                         struct xfs_log_iovec    *reg = &vecp[index];
2027                         struct xlog_op_header   *ophdr;
2028                         int                     start_rec_copy;
2029                         int                     copy_len;
2030                         int                     copy_off;
2031
2032                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2033                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2034
2035                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2036                         if (start_rec_copy) {
2037                                 record_cnt++;
2038                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2039                                                    start_rec_copy);
2040                         }
2041
2042                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2043                         if (!ophdr)
2044                                 return XFS_ERROR(EIO);
2045
2046                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2047                                            sizeof(struct xlog_op_header));
2048
2049                         len += xlog_write_setup_copy(ticket, ophdr,
2050                                                      iclog->ic_size-log_offset,
2051                                                      reg->i_len,
2052                                                      &copy_off, &copy_len,
2053                                                      &partial_copy,
2054                                                      &partial_copy_len);
2055                         xlog_verify_dest_ptr(log, ptr);
2056
2057                         /* copy region */
2058                         ASSERT(copy_len >= 0);
2059                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
2060                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2061
2062                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2063                         record_cnt++;
2064                         data_cnt += contwr ? copy_len : 0;
2065
2066                         error = xlog_write_copy_finish(log, iclog, flags,
2067                                                        &record_cnt, &data_cnt,
2068                                                        &partial_copy,
2069                                                        &partial_copy_len,
2070                                                        log_offset,
2071                                                        commit_iclog);
2072                         if (error)
2073                                 return error;
2074
2075                         /*
2076                          * if we had a partial copy, we need to get more iclog
2077                          * space but we don't want to increment the region
2078                          * index because there is still more is this region to
2079                          * write.
2080                          *
2081                          * If we completed writing this region, and we flushed
2082                          * the iclog (indicated by resetting of the record
2083                          * count), then we also need to get more log space. If
2084                          * this was the last record, though, we are done and
2085                          * can just return.
2086                          */
2087                         if (partial_copy)
2088                                 break;
2089
2090                         if (++index == lv->lv_niovecs) {
2091                                 lv = lv->lv_next;
2092                                 index = 0;
2093                                 if (lv)
2094                                         vecp = lv->lv_iovecp;
2095                         }
2096                         if (record_cnt == 0) {
2097                                 if (!lv)
2098                                         return 0;
2099                                 break;
2100                         }
2101                 }
2102         }
2103
2104         ASSERT(len == 0);
2105
2106         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2107         if (!commit_iclog)
2108                 return xlog_state_release_iclog(log, iclog);
2109
2110         ASSERT(flags & XLOG_COMMIT_TRANS);
2111         *commit_iclog = iclog;
2112         return 0;
2113 }
2114
2115
2116 /*****************************************************************************
2117  *
2118  *              State Machine functions
2119  *
2120  *****************************************************************************
2121  */
2122
2123 /* Clean iclogs starting from the head.  This ordering must be
2124  * maintained, so an iclog doesn't become ACTIVE beyond one that
2125  * is SYNCING.  This is also required to maintain the notion that we use
2126  * a ordered wait queue to hold off would be writers to the log when every
2127  * iclog is trying to sync to disk.
2128  *
2129  * State Change: DIRTY -> ACTIVE
2130  */
2131 STATIC void
2132 xlog_state_clean_log(xlog_t *log)
2133 {
2134         xlog_in_core_t  *iclog;
2135         int changed = 0;
2136
2137         iclog = log->l_iclog;
2138         do {
2139                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2140                         iclog->ic_state = XLOG_STATE_ACTIVE;
2141                         iclog->ic_offset       = 0;
2142                         ASSERT(iclog->ic_callback == NULL);
2143                         /*
2144                          * If the number of ops in this iclog indicate it just
2145                          * contains the dummy transaction, we can
2146                          * change state into IDLE (the second time around).
2147                          * Otherwise we should change the state into
2148                          * NEED a dummy.
2149                          * We don't need to cover the dummy.
2150                          */
2151                         if (!changed &&
2152                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2153                                         XLOG_COVER_OPS)) {
2154                                 changed = 1;
2155                         } else {
2156                                 /*
2157                                  * We have two dirty iclogs so start over
2158                                  * This could also be num of ops indicates
2159                                  * this is not the dummy going out.
2160                                  */
2161                                 changed = 2;
2162                         }
2163                         iclog->ic_header.h_num_logops = 0;
2164                         memset(iclog->ic_header.h_cycle_data, 0,
2165                               sizeof(iclog->ic_header.h_cycle_data));
2166                         iclog->ic_header.h_lsn = 0;
2167                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2168                         /* do nothing */;
2169                 else
2170                         break;  /* stop cleaning */
2171                 iclog = iclog->ic_next;
2172         } while (iclog != log->l_iclog);
2173
2174         /* log is locked when we are called */
2175         /*
2176          * Change state for the dummy log recording.
2177          * We usually go to NEED. But we go to NEED2 if the changed indicates
2178          * we are done writing the dummy record.
2179          * If we are done with the second dummy recored (DONE2), then
2180          * we go to IDLE.
2181          */
2182         if (changed) {
2183                 switch (log->l_covered_state) {
2184                 case XLOG_STATE_COVER_IDLE:
2185                 case XLOG_STATE_COVER_NEED:
2186                 case XLOG_STATE_COVER_NEED2:
2187                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2188                         break;
2189
2190                 case XLOG_STATE_COVER_DONE:
2191                         if (changed == 1)
2192                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2193                         else
2194                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2195                         break;
2196
2197                 case XLOG_STATE_COVER_DONE2:
2198                         if (changed == 1)
2199                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2200                         else
2201                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2202                         break;
2203
2204                 default:
2205                         ASSERT(0);
2206                 }
2207         }
2208 }       /* xlog_state_clean_log */
2209
2210 STATIC xfs_lsn_t
2211 xlog_get_lowest_lsn(
2212         xlog_t          *log)
2213 {
2214         xlog_in_core_t  *lsn_log;
2215         xfs_lsn_t       lowest_lsn, lsn;
2216
2217         lsn_log = log->l_iclog;
2218         lowest_lsn = 0;
2219         do {
2220             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2221                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2222                 if ((lsn && !lowest_lsn) ||
2223                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2224                         lowest_lsn = lsn;
2225                 }
2226             }
2227             lsn_log = lsn_log->ic_next;
2228         } while (lsn_log != log->l_iclog);
2229         return lowest_lsn;
2230 }
2231
2232
2233 STATIC void
2234 xlog_state_do_callback(
2235         xlog_t          *log,
2236         int             aborted,
2237         xlog_in_core_t  *ciclog)
2238 {
2239         xlog_in_core_t     *iclog;
2240         xlog_in_core_t     *first_iclog;        /* used to know when we've
2241                                                  * processed all iclogs once */
2242         xfs_log_callback_t *cb, *cb_next;
2243         int                flushcnt = 0;
2244         xfs_lsn_t          lowest_lsn;
2245         int                ioerrors;    /* counter: iclogs with errors */
2246         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2247         int                funcdidcallbacks; /* flag: function did callbacks */
2248         int                repeats;     /* for issuing console warnings if
2249                                          * looping too many times */
2250         int                wake = 0;
2251
2252         spin_lock(&log->l_icloglock);
2253         first_iclog = iclog = log->l_iclog;
2254         ioerrors = 0;
2255         funcdidcallbacks = 0;
2256         repeats = 0;
2257
2258         do {
2259                 /*
2260                  * Scan all iclogs starting with the one pointed to by the
2261                  * log.  Reset this starting point each time the log is
2262                  * unlocked (during callbacks).
2263                  *
2264                  * Keep looping through iclogs until one full pass is made
2265                  * without running any callbacks.
2266                  */
2267                 first_iclog = log->l_iclog;
2268                 iclog = log->l_iclog;
2269                 loopdidcallbacks = 0;
2270                 repeats++;
2271
2272                 do {
2273
2274                         /* skip all iclogs in the ACTIVE & DIRTY states */
2275                         if (iclog->ic_state &
2276                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2277                                 iclog = iclog->ic_next;
2278                                 continue;
2279                         }
2280
2281                         /*
2282                          * Between marking a filesystem SHUTDOWN and stopping
2283                          * the log, we do flush all iclogs to disk (if there
2284                          * wasn't a log I/O error). So, we do want things to
2285                          * go smoothly in case of just a SHUTDOWN  w/o a
2286                          * LOG_IO_ERROR.
2287                          */
2288                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2289                                 /*
2290                                  * Can only perform callbacks in order.  Since
2291                                  * this iclog is not in the DONE_SYNC/
2292                                  * DO_CALLBACK state, we skip the rest and
2293                                  * just try to clean up.  If we set our iclog
2294                                  * to DO_CALLBACK, we will not process it when
2295                                  * we retry since a previous iclog is in the
2296                                  * CALLBACK and the state cannot change since
2297                                  * we are holding the l_icloglock.
2298                                  */
2299                                 if (!(iclog->ic_state &
2300                                         (XLOG_STATE_DONE_SYNC |
2301                                                  XLOG_STATE_DO_CALLBACK))) {
2302                                         if (ciclog && (ciclog->ic_state ==
2303                                                         XLOG_STATE_DONE_SYNC)) {
2304                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2305                                         }
2306                                         break;
2307                                 }
2308                                 /*
2309                                  * We now have an iclog that is in either the
2310                                  * DO_CALLBACK or DONE_SYNC states. The other
2311                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2312                                  * caught by the above if and are going to
2313                                  * clean (i.e. we aren't doing their callbacks)
2314                                  * see the above if.
2315                                  */
2316
2317                                 /*
2318                                  * We will do one more check here to see if we
2319                                  * have chased our tail around.
2320                                  */
2321
2322                                 lowest_lsn = xlog_get_lowest_lsn(log);
2323                                 if (lowest_lsn &&
2324                                     XFS_LSN_CMP(lowest_lsn,
2325                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2326                                         iclog = iclog->ic_next;
2327                                         continue; /* Leave this iclog for
2328                                                    * another thread */
2329                                 }
2330
2331                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2332
2333
2334                                 /*
2335                                  * update the last_sync_lsn before we drop the
2336                                  * icloglock to ensure we are the only one that
2337                                  * can update it.
2338                                  */
2339                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2340                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2341                                 atomic64_set(&log->l_last_sync_lsn,
2342                                         be64_to_cpu(iclog->ic_header.h_lsn));
2343
2344                         } else
2345                                 ioerrors++;
2346
2347                         spin_unlock(&log->l_icloglock);
2348
2349                         /*
2350                          * Keep processing entries in the callback list until
2351                          * we come around and it is empty.  We need to
2352                          * atomically see that the list is empty and change the
2353                          * state to DIRTY so that we don't miss any more
2354                          * callbacks being added.
2355                          */
2356                         spin_lock(&iclog->ic_callback_lock);
2357                         cb = iclog->ic_callback;
2358                         while (cb) {
2359                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2360                                 iclog->ic_callback = NULL;
2361                                 spin_unlock(&iclog->ic_callback_lock);
2362
2363                                 /* perform callbacks in the order given */
2364                                 for (; cb; cb = cb_next) {
2365                                         cb_next = cb->cb_next;
2366                                         cb->cb_func(cb->cb_arg, aborted);
2367                                 }
2368                                 spin_lock(&iclog->ic_callback_lock);
2369                                 cb = iclog->ic_callback;
2370                         }
2371
2372                         loopdidcallbacks++;
2373                         funcdidcallbacks++;
2374
2375                         spin_lock(&log->l_icloglock);
2376                         ASSERT(iclog->ic_callback == NULL);
2377                         spin_unlock(&iclog->ic_callback_lock);
2378                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2379                                 iclog->ic_state = XLOG_STATE_DIRTY;
2380
2381                         /*
2382                          * Transition from DIRTY to ACTIVE if applicable.
2383                          * NOP if STATE_IOERROR.
2384                          */
2385                         xlog_state_clean_log(log);
2386
2387                         /* wake up threads waiting in xfs_log_force() */
2388                         wake_up_all(&iclog->ic_force_wait);
2389
2390                         iclog = iclog->ic_next;
2391                 } while (first_iclog != iclog);
2392
2393                 if (repeats > 5000) {
2394                         flushcnt += repeats;
2395                         repeats = 0;
2396                         xfs_warn(log->l_mp,
2397                                 "%s: possible infinite loop (%d iterations)",
2398                                 __func__, flushcnt);
2399                 }
2400         } while (!ioerrors && loopdidcallbacks);
2401
2402         /*
2403          * make one last gasp attempt to see if iclogs are being left in
2404          * limbo..
2405          */
2406 #ifdef DEBUG
2407         if (funcdidcallbacks) {
2408                 first_iclog = iclog = log->l_iclog;
2409                 do {
2410                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2411                         /*
2412                          * Terminate the loop if iclogs are found in states
2413                          * which will cause other threads to clean up iclogs.
2414                          *
2415                          * SYNCING - i/o completion will go through logs
2416                          * DONE_SYNC - interrupt thread should be waiting for
2417                          *              l_icloglock
2418                          * IOERROR - give up hope all ye who enter here
2419                          */
2420                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2421                             iclog->ic_state == XLOG_STATE_SYNCING ||
2422                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2423                             iclog->ic_state == XLOG_STATE_IOERROR )
2424                                 break;
2425                         iclog = iclog->ic_next;
2426                 } while (first_iclog != iclog);
2427         }
2428 #endif
2429
2430         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2431                 wake = 1;
2432         spin_unlock(&log->l_icloglock);
2433
2434         if (wake)
2435                 wake_up_all(&log->l_flush_wait);
2436 }
2437
2438
2439 /*
2440  * Finish transitioning this iclog to the dirty state.
2441  *
2442  * Make sure that we completely execute this routine only when this is
2443  * the last call to the iclog.  There is a good chance that iclog flushes,
2444  * when we reach the end of the physical log, get turned into 2 separate
2445  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2446  * routine.  By using the reference count bwritecnt, we guarantee that only
2447  * the second completion goes through.
2448  *
2449  * Callbacks could take time, so they are done outside the scope of the
2450  * global state machine log lock.
2451  */
2452 STATIC void
2453 xlog_state_done_syncing(
2454         xlog_in_core_t  *iclog,
2455         int             aborted)
2456 {
2457         xlog_t             *log = iclog->ic_log;
2458
2459         spin_lock(&log->l_icloglock);
2460
2461         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2462                iclog->ic_state == XLOG_STATE_IOERROR);
2463         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2464         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2465
2466
2467         /*
2468          * If we got an error, either on the first buffer, or in the case of
2469          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2470          * and none should ever be attempted to be written to disk
2471          * again.
2472          */
2473         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2474                 if (--iclog->ic_bwritecnt == 1) {
2475                         spin_unlock(&log->l_icloglock);
2476                         return;
2477                 }
2478                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2479         }
2480
2481         /*
2482          * Someone could be sleeping prior to writing out the next
2483          * iclog buffer, we wake them all, one will get to do the
2484          * I/O, the others get to wait for the result.
2485          */
2486         wake_up_all(&iclog->ic_write_wait);
2487         spin_unlock(&log->l_icloglock);
2488         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2489 }       /* xlog_state_done_syncing */
2490
2491
2492 /*
2493  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2494  * sleep.  We wait on the flush queue on the head iclog as that should be
2495  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2496  * we will wait here and all new writes will sleep until a sync completes.
2497  *
2498  * The in-core logs are used in a circular fashion. They are not used
2499  * out-of-order even when an iclog past the head is free.
2500  *
2501  * return:
2502  *      * log_offset where xlog_write() can start writing into the in-core
2503  *              log's data space.
2504  *      * in-core log pointer to which xlog_write() should write.
2505  *      * boolean indicating this is a continued write to an in-core log.
2506  *              If this is the last write, then the in-core log's offset field
2507  *              needs to be incremented, depending on the amount of data which
2508  *              is copied.
2509  */
2510 STATIC int
2511 xlog_state_get_iclog_space(xlog_t         *log,
2512                            int            len,
2513                            xlog_in_core_t **iclogp,
2514                            xlog_ticket_t  *ticket,
2515                            int            *continued_write,
2516                            int            *logoffsetp)
2517 {
2518         int               log_offset;
2519         xlog_rec_header_t *head;
2520         xlog_in_core_t    *iclog;
2521         int               error;
2522
2523 restart:
2524         spin_lock(&log->l_icloglock);
2525         if (XLOG_FORCED_SHUTDOWN(log)) {
2526                 spin_unlock(&log->l_icloglock);
2527                 return XFS_ERROR(EIO);
2528         }
2529
2530         iclog = log->l_iclog;
2531         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2532                 XFS_STATS_INC(xs_log_noiclogs);
2533
2534                 /* Wait for log writes to have flushed */
2535                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2536                 goto restart;
2537         }
2538
2539         head = &iclog->ic_header;
2540
2541         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2542         log_offset = iclog->ic_offset;
2543
2544         /* On the 1st write to an iclog, figure out lsn.  This works
2545          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2546          * committing to.  If the offset is set, that's how many blocks
2547          * must be written.
2548          */
2549         if (log_offset == 0) {
2550                 ticket->t_curr_res -= log->l_iclog_hsize;
2551                 xlog_tic_add_region(ticket,
2552                                     log->l_iclog_hsize,
2553                                     XLOG_REG_TYPE_LRHEADER);
2554                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2555                 head->h_lsn = cpu_to_be64(
2556                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2557                 ASSERT(log->l_curr_block >= 0);
2558         }
2559
2560         /* If there is enough room to write everything, then do it.  Otherwise,
2561          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2562          * bit is on, so this will get flushed out.  Don't update ic_offset
2563          * until you know exactly how many bytes get copied.  Therefore, wait
2564          * until later to update ic_offset.
2565          *
2566          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2567          * can fit into remaining data section.
2568          */
2569         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2570                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2571
2572                 /*
2573                  * If I'm the only one writing to this iclog, sync it to disk.
2574                  * We need to do an atomic compare and decrement here to avoid
2575                  * racing with concurrent atomic_dec_and_lock() calls in
2576                  * xlog_state_release_iclog() when there is more than one
2577                  * reference to the iclog.
2578                  */
2579                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2580                         /* we are the only one */
2581                         spin_unlock(&log->l_icloglock);
2582                         error = xlog_state_release_iclog(log, iclog);
2583                         if (error)
2584                                 return error;
2585                 } else {
2586                         spin_unlock(&log->l_icloglock);
2587                 }
2588                 goto restart;
2589         }
2590
2591         /* Do we have enough room to write the full amount in the remainder
2592          * of this iclog?  Or must we continue a write on the next iclog and
2593          * mark this iclog as completely taken?  In the case where we switch
2594          * iclogs (to mark it taken), this particular iclog will release/sync
2595          * to disk in xlog_write().
2596          */
2597         if (len <= iclog->ic_size - iclog->ic_offset) {
2598                 *continued_write = 0;
2599                 iclog->ic_offset += len;
2600         } else {
2601                 *continued_write = 1;
2602                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2603         }
2604         *iclogp = iclog;
2605
2606         ASSERT(iclog->ic_offset <= iclog->ic_size);
2607         spin_unlock(&log->l_icloglock);
2608
2609         *logoffsetp = log_offset;
2610         return 0;
2611 }       /* xlog_state_get_iclog_space */
2612
2613 /* The first cnt-1 times through here we don't need to
2614  * move the grant write head because the permanent
2615  * reservation has reserved cnt times the unit amount.
2616  * Release part of current permanent unit reservation and
2617  * reset current reservation to be one units worth.  Also
2618  * move grant reservation head forward.
2619  */
2620 STATIC void
2621 xlog_regrant_reserve_log_space(xlog_t        *log,
2622                                xlog_ticket_t *ticket)
2623 {
2624         trace_xfs_log_regrant_reserve_enter(log, ticket);
2625
2626         if (ticket->t_cnt > 0)
2627                 ticket->t_cnt--;
2628
2629         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2630                                         ticket->t_curr_res);
2631         xlog_grant_sub_space(log, &log->l_write_head.grant,
2632                                         ticket->t_curr_res);
2633         ticket->t_curr_res = ticket->t_unit_res;
2634         xlog_tic_reset_res(ticket);
2635
2636         trace_xfs_log_regrant_reserve_sub(log, ticket);
2637
2638         /* just return if we still have some of the pre-reserved space */
2639         if (ticket->t_cnt > 0)
2640                 return;
2641
2642         xlog_grant_add_space(log, &log->l_reserve_head.grant,
2643                                         ticket->t_unit_res);
2644
2645         trace_xfs_log_regrant_reserve_exit(log, ticket);
2646
2647         ticket->t_curr_res = ticket->t_unit_res;
2648         xlog_tic_reset_res(ticket);
2649 }       /* xlog_regrant_reserve_log_space */
2650
2651
2652 /*
2653  * Give back the space left from a reservation.
2654  *
2655  * All the information we need to make a correct determination of space left
2656  * is present.  For non-permanent reservations, things are quite easy.  The
2657  * count should have been decremented to zero.  We only need to deal with the
2658  * space remaining in the current reservation part of the ticket.  If the
2659  * ticket contains a permanent reservation, there may be left over space which
2660  * needs to be released.  A count of N means that N-1 refills of the current
2661  * reservation can be done before we need to ask for more space.  The first
2662  * one goes to fill up the first current reservation.  Once we run out of
2663  * space, the count will stay at zero and the only space remaining will be
2664  * in the current reservation field.
2665  */
2666 STATIC void
2667 xlog_ungrant_log_space(xlog_t        *log,
2668                        xlog_ticket_t *ticket)
2669 {
2670         int     bytes;
2671
2672         if (ticket->t_cnt > 0)
2673                 ticket->t_cnt--;
2674
2675         trace_xfs_log_ungrant_enter(log, ticket);
2676         trace_xfs_log_ungrant_sub(log, ticket);
2677
2678         /*
2679          * If this is a permanent reservation ticket, we may be able to free
2680          * up more space based on the remaining count.
2681          */
2682         bytes = ticket->t_curr_res;
2683         if (ticket->t_cnt > 0) {
2684                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2685                 bytes += ticket->t_unit_res*ticket->t_cnt;
2686         }
2687
2688         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2689         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2690
2691         trace_xfs_log_ungrant_exit(log, ticket);
2692
2693         xfs_log_space_wake(log->l_mp);
2694 }
2695
2696 /*
2697  * Flush iclog to disk if this is the last reference to the given iclog and
2698  * the WANT_SYNC bit is set.
2699  *
2700  * When this function is entered, the iclog is not necessarily in the
2701  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2702  *
2703  *
2704  */
2705 STATIC int
2706 xlog_state_release_iclog(
2707         xlog_t          *log,
2708         xlog_in_core_t  *iclog)
2709 {
2710         int             sync = 0;       /* do we sync? */
2711
2712         if (iclog->ic_state & XLOG_STATE_IOERROR)
2713                 return XFS_ERROR(EIO);
2714
2715         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2716         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2717                 return 0;
2718
2719         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2720                 spin_unlock(&log->l_icloglock);
2721                 return XFS_ERROR(EIO);
2722         }
2723         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2724                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2725
2726         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2727                 /* update tail before writing to iclog */
2728                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2729                 sync++;
2730                 iclog->ic_state = XLOG_STATE_SYNCING;
2731                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2732                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2733                 /* cycle incremented when incrementing curr_block */
2734         }
2735         spin_unlock(&log->l_icloglock);
2736
2737         /*
2738          * We let the log lock go, so it's possible that we hit a log I/O
2739          * error or some other SHUTDOWN condition that marks the iclog
2740          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2741          * this iclog has consistent data, so we ignore IOERROR
2742          * flags after this point.
2743          */
2744         if (sync)
2745                 return xlog_sync(log, iclog);
2746         return 0;
2747 }       /* xlog_state_release_iclog */
2748
2749
2750 /*
2751  * This routine will mark the current iclog in the ring as WANT_SYNC
2752  * and move the current iclog pointer to the next iclog in the ring.
2753  * When this routine is called from xlog_state_get_iclog_space(), the
2754  * exact size of the iclog has not yet been determined.  All we know is
2755  * that every data block.  We have run out of space in this log record.
2756  */
2757 STATIC void
2758 xlog_state_switch_iclogs(xlog_t         *log,
2759                          xlog_in_core_t *iclog,
2760                          int            eventual_size)
2761 {
2762         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2763         if (!eventual_size)
2764                 eventual_size = iclog->ic_offset;
2765         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2766         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2767         log->l_prev_block = log->l_curr_block;
2768         log->l_prev_cycle = log->l_curr_cycle;
2769
2770         /* roll log?: ic_offset changed later */
2771         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2772
2773         /* Round up to next log-sunit */
2774         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2775             log->l_mp->m_sb.sb_logsunit > 1) {
2776                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2777                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2778         }
2779
2780         if (log->l_curr_block >= log->l_logBBsize) {
2781                 log->l_curr_cycle++;
2782                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2783                         log->l_curr_cycle++;
2784                 log->l_curr_block -= log->l_logBBsize;
2785                 ASSERT(log->l_curr_block >= 0);
2786         }
2787         ASSERT(iclog == log->l_iclog);
2788         log->l_iclog = iclog->ic_next;
2789 }       /* xlog_state_switch_iclogs */
2790
2791 /*
2792  * Write out all data in the in-core log as of this exact moment in time.
2793  *
2794  * Data may be written to the in-core log during this call.  However,
2795  * we don't guarantee this data will be written out.  A change from past
2796  * implementation means this routine will *not* write out zero length LRs.
2797  *
2798  * Basically, we try and perform an intelligent scan of the in-core logs.
2799  * If we determine there is no flushable data, we just return.  There is no
2800  * flushable data if:
2801  *
2802  *      1. the current iclog is active and has no data; the previous iclog
2803  *              is in the active or dirty state.
2804  *      2. the current iclog is drity, and the previous iclog is in the
2805  *              active or dirty state.
2806  *
2807  * We may sleep if:
2808  *
2809  *      1. the current iclog is not in the active nor dirty state.
2810  *      2. the current iclog dirty, and the previous iclog is not in the
2811  *              active nor dirty state.
2812  *      3. the current iclog is active, and there is another thread writing
2813  *              to this particular iclog.
2814  *      4. a) the current iclog is active and has no other writers
2815  *         b) when we return from flushing out this iclog, it is still
2816  *              not in the active nor dirty state.
2817  */
2818 int
2819 _xfs_log_force(
2820         struct xfs_mount        *mp,
2821         uint                    flags,
2822         int                     *log_flushed)
2823 {
2824         struct log              *log = mp->m_log;
2825         struct xlog_in_core     *iclog;
2826         xfs_lsn_t               lsn;
2827
2828         XFS_STATS_INC(xs_log_force);
2829
2830         xlog_cil_force(log);
2831
2832         spin_lock(&log->l_icloglock);
2833
2834         iclog = log->l_iclog;
2835         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2836                 spin_unlock(&log->l_icloglock);
2837                 return XFS_ERROR(EIO);
2838         }
2839
2840         /* If the head iclog is not active nor dirty, we just attach
2841          * ourselves to the head and go to sleep.
2842          */
2843         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2844             iclog->ic_state == XLOG_STATE_DIRTY) {
2845                 /*
2846                  * If the head is dirty or (active and empty), then
2847                  * we need to look at the previous iclog.  If the previous
2848                  * iclog is active or dirty we are done.  There is nothing
2849                  * to sync out.  Otherwise, we attach ourselves to the
2850                  * previous iclog and go to sleep.
2851                  */
2852                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2853                     (atomic_read(&iclog->ic_refcnt) == 0
2854                      && iclog->ic_offset == 0)) {
2855                         iclog = iclog->ic_prev;
2856                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2857                             iclog->ic_state == XLOG_STATE_DIRTY)
2858                                 goto no_sleep;
2859                         else
2860                                 goto maybe_sleep;
2861                 } else {
2862                         if (atomic_read(&iclog->ic_refcnt) == 0) {
2863                                 /* We are the only one with access to this
2864                                  * iclog.  Flush it out now.  There should
2865                                  * be a roundoff of zero to show that someone
2866                                  * has already taken care of the roundoff from
2867                                  * the previous sync.
2868                                  */
2869                                 atomic_inc(&iclog->ic_refcnt);
2870                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2871                                 xlog_state_switch_iclogs(log, iclog, 0);
2872                                 spin_unlock(&log->l_icloglock);
2873
2874                                 if (xlog_state_release_iclog(log, iclog))
2875                                         return XFS_ERROR(EIO);
2876
2877                                 if (log_flushed)
2878                                         *log_flushed = 1;
2879                                 spin_lock(&log->l_icloglock);
2880                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2881                                     iclog->ic_state != XLOG_STATE_DIRTY)
2882                                         goto maybe_sleep;
2883                                 else
2884                                         goto no_sleep;
2885                         } else {
2886                                 /* Someone else is writing to this iclog.
2887                                  * Use its call to flush out the data.  However,
2888                                  * the other thread may not force out this LR,
2889                                  * so we mark it WANT_SYNC.
2890                                  */
2891                                 xlog_state_switch_iclogs(log, iclog, 0);
2892                                 goto maybe_sleep;
2893                         }
2894                 }
2895         }
2896
2897         /* By the time we come around again, the iclog could've been filled
2898          * which would give it another lsn.  If we have a new lsn, just
2899          * return because the relevant data has been flushed.
2900          */
2901 maybe_sleep:
2902         if (flags & XFS_LOG_SYNC) {
2903                 /*
2904                  * We must check if we're shutting down here, before
2905                  * we wait, while we're holding the l_icloglock.
2906                  * Then we check again after waking up, in case our
2907                  * sleep was disturbed by a bad news.
2908                  */
2909                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2910                         spin_unlock(&log->l_icloglock);
2911                         return XFS_ERROR(EIO);
2912                 }
2913                 XFS_STATS_INC(xs_log_force_sleep);
2914                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
2915                 /*
2916                  * No need to grab the log lock here since we're
2917                  * only deciding whether or not to return EIO
2918                  * and the memory read should be atomic.
2919                  */
2920                 if (iclog->ic_state & XLOG_STATE_IOERROR)
2921                         return XFS_ERROR(EIO);
2922                 if (log_flushed)
2923                         *log_flushed = 1;
2924         } else {
2925
2926 no_sleep:
2927                 spin_unlock(&log->l_icloglock);
2928         }
2929         return 0;
2930 }
2931
2932 /*
2933  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
2934  * about errors or whether the log was flushed or not. This is the normal
2935  * interface to use when trying to unpin items or move the log forward.
2936  */
2937 void
2938 xfs_log_force(
2939         xfs_mount_t     *mp,
2940         uint            flags)
2941 {
2942         int     error;
2943
2944         trace_xfs_log_force(mp, 0);
2945         error = _xfs_log_force(mp, flags, NULL);
2946         if (error)
2947                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
2948 }
2949
2950 /*
2951  * Force the in-core log to disk for a specific LSN.
2952  *
2953  * Find in-core log with lsn.
2954  *      If it is in the DIRTY state, just return.
2955  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2956  *              state and go to sleep or return.
2957  *      If it is in any other state, go to sleep or return.
2958  *
2959  * Synchronous forces are implemented with a signal variable. All callers
2960  * to force a given lsn to disk will wait on a the sv attached to the
2961  * specific in-core log.  When given in-core log finally completes its
2962  * write to disk, that thread will wake up all threads waiting on the
2963  * sv.
2964  */
2965 int
2966 _xfs_log_force_lsn(
2967         struct xfs_mount        *mp,
2968         xfs_lsn_t               lsn,
2969         uint                    flags,
2970         int                     *log_flushed)
2971 {
2972         struct log              *log = mp->m_log;
2973         struct xlog_in_core     *iclog;
2974         int                     already_slept = 0;
2975
2976         ASSERT(lsn != 0);
2977
2978         XFS_STATS_INC(xs_log_force);
2979
2980         lsn = xlog_cil_force_lsn(log, lsn);
2981         if (lsn == NULLCOMMITLSN)
2982                 return 0;
2983
2984 try_again:
2985         spin_lock(&log->l_icloglock);
2986         iclog = log->l_iclog;
2987         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2988                 spin_unlock(&log->l_icloglock);
2989                 return XFS_ERROR(EIO);
2990         }
2991
2992         do {
2993                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
2994                         iclog = iclog->ic_next;
2995                         continue;
2996                 }
2997
2998                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2999                         spin_unlock(&log->l_icloglock);
3000                         return 0;
3001                 }
3002
3003                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3004                         /*
3005                          * We sleep here if we haven't already slept (e.g.
3006                          * this is the first time we've looked at the correct
3007                          * iclog buf) and the buffer before us is going to
3008                          * be sync'ed. The reason for this is that if we
3009                          * are doing sync transactions here, by waiting for
3010                          * the previous I/O to complete, we can allow a few
3011                          * more transactions into this iclog before we close
3012                          * it down.
3013                          *
3014                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3015                          * up the refcnt so we can release the log (which
3016                          * drops the ref count).  The state switch keeps new
3017                          * transaction commits from using this buffer.  When
3018                          * the current commits finish writing into the buffer,
3019                          * the refcount will drop to zero and the buffer will
3020                          * go out then.
3021                          */
3022                         if (!already_slept &&
3023                             (iclog->ic_prev->ic_state &
3024                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3025                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3026
3027                                 XFS_STATS_INC(xs_log_force_sleep);
3028
3029                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3030                                                         &log->l_icloglock);
3031                                 if (log_flushed)
3032                                         *log_flushed = 1;
3033                                 already_slept = 1;
3034                                 goto try_again;
3035                         }
3036                         atomic_inc(&iclog->ic_refcnt);
3037                         xlog_state_switch_iclogs(log, iclog, 0);
3038                         spin_unlock(&log->l_icloglock);
3039                         if (xlog_state_release_iclog(log, iclog))
3040                                 return XFS_ERROR(EIO);
3041                         if (log_flushed)
3042                                 *log_flushed = 1;
3043                         spin_lock(&log->l_icloglock);
3044                 }
3045
3046                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3047                     !(iclog->ic_state &
3048                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3049                         /*
3050                          * Don't wait on completion if we know that we've
3051                          * gotten a log write error.
3052                          */
3053                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3054                                 spin_unlock(&log->l_icloglock);
3055                                 return XFS_ERROR(EIO);
3056                         }
3057                         XFS_STATS_INC(xs_log_force_sleep);
3058                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3059                         /*
3060                          * No need to grab the log lock here since we're
3061                          * only deciding whether or not to return EIO
3062                          * and the memory read should be atomic.
3063                          */
3064                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3065                                 return XFS_ERROR(EIO);
3066
3067                         if (log_flushed)
3068                                 *log_flushed = 1;
3069                 } else {                /* just return */
3070                         spin_unlock(&log->l_icloglock);
3071                 }
3072
3073                 return 0;
3074         } while (iclog != log->l_iclog);
3075
3076         spin_unlock(&log->l_icloglock);
3077         return 0;
3078 }
3079
3080 /*
3081  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3082  * about errors or whether the log was flushed or not. This is the normal
3083  * interface to use when trying to unpin items or move the log forward.
3084  */
3085 void
3086 xfs_log_force_lsn(
3087         xfs_mount_t     *mp,
3088         xfs_lsn_t       lsn,
3089         uint            flags)
3090 {
3091         int     error;
3092
3093         trace_xfs_log_force(mp, lsn);
3094         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3095         if (error)
3096                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3097 }
3098
3099 /*
3100  * Called when we want to mark the current iclog as being ready to sync to
3101  * disk.
3102  */
3103 STATIC void
3104 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3105 {
3106         assert_spin_locked(&log->l_icloglock);
3107
3108         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3109                 xlog_state_switch_iclogs(log, iclog, 0);
3110         } else {
3111                 ASSERT(iclog->ic_state &
3112                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3113         }
3114 }
3115
3116
3117 /*****************************************************************************
3118  *
3119  *              TICKET functions
3120  *
3121  *****************************************************************************
3122  */
3123
3124 /*
3125  * Free a used ticket when its refcount falls to zero.
3126  */
3127 void
3128 xfs_log_ticket_put(
3129         xlog_ticket_t   *ticket)
3130 {
3131         ASSERT(atomic_read(&ticket->t_ref) > 0);
3132         if (atomic_dec_and_test(&ticket->t_ref))
3133                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3134 }
3135
3136 xlog_ticket_t *
3137 xfs_log_ticket_get(
3138         xlog_ticket_t   *ticket)
3139 {
3140         ASSERT(atomic_read(&ticket->t_ref) > 0);
3141         atomic_inc(&ticket->t_ref);
3142         return ticket;
3143 }
3144
3145 /*
3146  * Allocate and initialise a new log ticket.
3147  */
3148 xlog_ticket_t *
3149 xlog_ticket_alloc(
3150         struct log      *log,
3151         int             unit_bytes,
3152         int             cnt,
3153         char            client,
3154         bool            permanent,
3155         xfs_km_flags_t  alloc_flags)
3156 {
3157         struct xlog_ticket *tic;
3158         uint            num_headers;
3159         int             iclog_space;
3160
3161         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3162         if (!tic)
3163                 return NULL;
3164
3165         /*
3166          * Permanent reservations have up to 'cnt'-1 active log operations
3167          * in the log.  A unit in this case is the amount of space for one
3168          * of these log operations.  Normal reservations have a cnt of 1
3169          * and their unit amount is the total amount of space required.
3170          *
3171          * The following lines of code account for non-transaction data
3172          * which occupy space in the on-disk log.
3173          *
3174          * Normal form of a transaction is:
3175          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3176          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3177          *
3178          * We need to account for all the leadup data and trailer data
3179          * around the transaction data.
3180          * And then we need to account for the worst case in terms of using
3181          * more space.
3182          * The worst case will happen if:
3183          * - the placement of the transaction happens to be such that the
3184          *   roundoff is at its maximum
3185          * - the transaction data is synced before the commit record is synced
3186          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3187          *   Therefore the commit record is in its own Log Record.
3188          *   This can happen as the commit record is called with its
3189          *   own region to xlog_write().
3190          *   This then means that in the worst case, roundoff can happen for
3191          *   the commit-rec as well.
3192          *   The commit-rec is smaller than padding in this scenario and so it is
3193          *   not added separately.
3194          */
3195
3196         /* for trans header */
3197         unit_bytes += sizeof(xlog_op_header_t);
3198         unit_bytes += sizeof(xfs_trans_header_t);
3199
3200         /* for start-rec */
3201         unit_bytes += sizeof(xlog_op_header_t);
3202
3203         /*
3204          * for LR headers - the space for data in an iclog is the size minus
3205          * the space used for the headers. If we use the iclog size, then we
3206          * undercalculate the number of headers required.
3207          *
3208          * Furthermore - the addition of op headers for split-recs might
3209          * increase the space required enough to require more log and op
3210          * headers, so take that into account too.
3211          *
3212          * IMPORTANT: This reservation makes the assumption that if this
3213          * transaction is the first in an iclog and hence has the LR headers
3214          * accounted to it, then the remaining space in the iclog is
3215          * exclusively for this transaction.  i.e. if the transaction is larger
3216          * than the iclog, it will be the only thing in that iclog.
3217          * Fundamentally, this means we must pass the entire log vector to
3218          * xlog_write to guarantee this.
3219          */
3220         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3221         num_headers = howmany(unit_bytes, iclog_space);
3222
3223         /* for split-recs - ophdrs added when data split over LRs */
3224         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3225
3226         /* add extra header reservations if we overrun */
3227         while (!num_headers ||
3228                howmany(unit_bytes, iclog_space) > num_headers) {
3229                 unit_bytes += sizeof(xlog_op_header_t);
3230                 num_headers++;
3231         }
3232         unit_bytes += log->l_iclog_hsize * num_headers;
3233
3234         /* for commit-rec LR header - note: padding will subsume the ophdr */
3235         unit_bytes += log->l_iclog_hsize;
3236
3237         /* for roundoff padding for transaction data and one for commit record */
3238         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3239             log->l_mp->m_sb.sb_logsunit > 1) {
3240                 /* log su roundoff */
3241                 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3242         } else {
3243                 /* BB roundoff */
3244                 unit_bytes += 2*BBSIZE;
3245         }
3246
3247         atomic_set(&tic->t_ref, 1);
3248         tic->t_task             = current;
3249         INIT_LIST_HEAD(&tic->t_queue);
3250         tic->t_unit_res         = unit_bytes;
3251         tic->t_curr_res         = unit_bytes;
3252         tic->t_cnt              = cnt;
3253         tic->t_ocnt             = cnt;
3254         tic->t_tid              = random32();
3255         tic->t_clientid         = client;
3256         tic->t_flags            = XLOG_TIC_INITED;
3257         tic->t_trans_type       = 0;
3258         if (permanent)
3259                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3260
3261         xlog_tic_reset_res(tic);
3262
3263         return tic;
3264 }
3265
3266
3267 /******************************************************************************
3268  *
3269  *              Log debug routines
3270  *
3271  ******************************************************************************
3272  */
3273 #if defined(DEBUG)
3274 /*
3275  * Make sure that the destination ptr is within the valid data region of
3276  * one of the iclogs.  This uses backup pointers stored in a different
3277  * part of the log in case we trash the log structure.
3278  */
3279 void
3280 xlog_verify_dest_ptr(
3281         struct log      *log,
3282         char            *ptr)
3283 {
3284         int i;
3285         int good_ptr = 0;
3286
3287         for (i = 0; i < log->l_iclog_bufs; i++) {
3288                 if (ptr >= log->l_iclog_bak[i] &&
3289                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3290                         good_ptr++;
3291         }
3292
3293         if (!good_ptr)
3294                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3295 }
3296
3297 /*
3298  * Check to make sure the grant write head didn't just over lap the tail.  If
3299  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3300  * the cycles differ by exactly one and check the byte count.
3301  *
3302  * This check is run unlocked, so can give false positives. Rather than assert
3303  * on failures, use a warn-once flag and a panic tag to allow the admin to
3304  * determine if they want to panic the machine when such an error occurs. For
3305  * debug kernels this will have the same effect as using an assert but, unlinke
3306  * an assert, it can be turned off at runtime.
3307  */
3308 STATIC void
3309 xlog_verify_grant_tail(
3310         struct log      *log)
3311 {
3312         int             tail_cycle, tail_blocks;
3313         int             cycle, space;
3314
3315         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3316         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3317         if (tail_cycle != cycle) {
3318                 if (cycle - 1 != tail_cycle &&
3319                     !(log->l_flags & XLOG_TAIL_WARN)) {
3320                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3321                                 "%s: cycle - 1 != tail_cycle", __func__);
3322                         log->l_flags |= XLOG_TAIL_WARN;
3323                 }
3324
3325                 if (space > BBTOB(tail_blocks) &&
3326                     !(log->l_flags & XLOG_TAIL_WARN)) {
3327                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3328                                 "%s: space > BBTOB(tail_blocks)", __func__);
3329                         log->l_flags |= XLOG_TAIL_WARN;
3330                 }
3331         }
3332 }
3333
3334 /* check if it will fit */
3335 STATIC void
3336 xlog_verify_tail_lsn(xlog_t         *log,
3337                      xlog_in_core_t *iclog,
3338                      xfs_lsn_t      tail_lsn)
3339 {
3340     int blocks;
3341
3342     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3343         blocks =
3344             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3345         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3346                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3347     } else {
3348         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3349
3350         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3351                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3352
3353         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3354         if (blocks < BTOBB(iclog->ic_offset) + 1)
3355                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3356     }
3357 }       /* xlog_verify_tail_lsn */
3358
3359 /*
3360  * Perform a number of checks on the iclog before writing to disk.
3361  *
3362  * 1. Make sure the iclogs are still circular
3363  * 2. Make sure we have a good magic number
3364  * 3. Make sure we don't have magic numbers in the data
3365  * 4. Check fields of each log operation header for:
3366  *      A. Valid client identifier
3367  *      B. tid ptr value falls in valid ptr space (user space code)
3368  *      C. Length in log record header is correct according to the
3369  *              individual operation headers within record.
3370  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3371  *      log, check the preceding blocks of the physical log to make sure all
3372  *      the cycle numbers agree with the current cycle number.
3373  */
3374 STATIC void
3375 xlog_verify_iclog(xlog_t         *log,
3376                   xlog_in_core_t *iclog,
3377                   int            count,
3378                   boolean_t      syncing)
3379 {
3380         xlog_op_header_t        *ophead;
3381         xlog_in_core_t          *icptr;
3382         xlog_in_core_2_t        *xhdr;
3383         xfs_caddr_t             ptr;
3384         xfs_caddr_t             base_ptr;
3385         __psint_t               field_offset;
3386         __uint8_t               clientid;
3387         int                     len, i, j, k, op_len;
3388         int                     idx;
3389
3390         /* check validity of iclog pointers */
3391         spin_lock(&log->l_icloglock);
3392         icptr = log->l_iclog;
3393         for (i=0; i < log->l_iclog_bufs; i++) {
3394                 if (icptr == NULL)
3395                         xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3396                 icptr = icptr->ic_next;
3397         }
3398         if (icptr != log->l_iclog)
3399                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3400         spin_unlock(&log->l_icloglock);
3401
3402         /* check log magic numbers */
3403         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3404                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3405
3406         ptr = (xfs_caddr_t) &iclog->ic_header;
3407         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3408              ptr += BBSIZE) {
3409                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3410                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3411                                 __func__);
3412         }
3413
3414         /* check fields */
3415         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3416         ptr = iclog->ic_datap;
3417         base_ptr = ptr;
3418         ophead = (xlog_op_header_t *)ptr;
3419         xhdr = iclog->ic_data;
3420         for (i = 0; i < len; i++) {
3421                 ophead = (xlog_op_header_t *)ptr;
3422
3423                 /* clientid is only 1 byte */
3424                 field_offset = (__psint_t)
3425                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3426                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3427                         clientid = ophead->oh_clientid;
3428                 } else {
3429                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3430                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3431                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3432                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3433                                 clientid = xlog_get_client_id(
3434                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3435                         } else {
3436                                 clientid = xlog_get_client_id(
3437                                         iclog->ic_header.h_cycle_data[idx]);
3438                         }
3439                 }
3440                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3441                         xfs_warn(log->l_mp,
3442                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3443                                 __func__, clientid, ophead,
3444                                 (unsigned long)field_offset);
3445
3446                 /* check length */
3447                 field_offset = (__psint_t)
3448                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3449                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3450                         op_len = be32_to_cpu(ophead->oh_len);
3451                 } else {
3452                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3453                                     (__psint_t)iclog->ic_datap);
3454                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3455                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3456                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3457                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3458                         } else {
3459                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3460                         }
3461                 }
3462                 ptr += sizeof(xlog_op_header_t) + op_len;
3463         }
3464 }       /* xlog_verify_iclog */
3465 #endif
3466
3467 /*
3468  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3469  */
3470 STATIC int
3471 xlog_state_ioerror(
3472         xlog_t  *log)
3473 {
3474         xlog_in_core_t  *iclog, *ic;
3475
3476         iclog = log->l_iclog;
3477         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3478                 /*
3479                  * Mark all the incore logs IOERROR.
3480                  * From now on, no log flushes will result.
3481                  */
3482                 ic = iclog;
3483                 do {
3484                         ic->ic_state = XLOG_STATE_IOERROR;
3485                         ic = ic->ic_next;
3486                 } while (ic != iclog);
3487                 return 0;
3488         }
3489         /*
3490          * Return non-zero, if state transition has already happened.
3491          */
3492         return 1;
3493 }
3494
3495 /*
3496  * This is called from xfs_force_shutdown, when we're forcibly
3497  * shutting down the filesystem, typically because of an IO error.
3498  * Our main objectives here are to make sure that:
3499  *      a. the filesystem gets marked 'SHUTDOWN' for all interested
3500  *         parties to find out, 'atomically'.
3501  *      b. those who're sleeping on log reservations, pinned objects and
3502  *          other resources get woken up, and be told the bad news.
3503  *      c. nothing new gets queued up after (a) and (b) are done.
3504  *      d. if !logerror, flush the iclogs to disk, then seal them off
3505  *         for business.
3506  *
3507  * Note: for delayed logging the !logerror case needs to flush the regions
3508  * held in memory out to the iclogs before flushing them to disk. This needs
3509  * to be done before the log is marked as shutdown, otherwise the flush to the
3510  * iclogs will fail.
3511  */
3512 int
3513 xfs_log_force_umount(
3514         struct xfs_mount        *mp,
3515         int                     logerror)
3516 {
3517         xlog_t          *log;
3518         int             retval;
3519
3520         log = mp->m_log;
3521
3522         /*
3523          * If this happens during log recovery, don't worry about
3524          * locking; the log isn't open for business yet.
3525          */
3526         if (!log ||
3527             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3528                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3529                 if (mp->m_sb_bp)
3530                         XFS_BUF_DONE(mp->m_sb_bp);
3531                 return 0;
3532         }
3533
3534         /*
3535          * Somebody could've already done the hard work for us.
3536          * No need to get locks for this.
3537          */
3538         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3539                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3540                 return 1;
3541         }
3542         retval = 0;
3543
3544         /*
3545          * Flush the in memory commit item list before marking the log as
3546          * being shut down. We need to do it in this order to ensure all the
3547          * completed transactions are flushed to disk with the xfs_log_force()
3548          * call below.
3549          */
3550         if (!logerror)
3551                 xlog_cil_force(log);
3552
3553         /*
3554          * mark the filesystem and the as in a shutdown state and wake
3555          * everybody up to tell them the bad news.
3556          */
3557         spin_lock(&log->l_icloglock);
3558         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3559         if (mp->m_sb_bp)
3560                 XFS_BUF_DONE(mp->m_sb_bp);
3561
3562         /*
3563          * This flag is sort of redundant because of the mount flag, but
3564          * it's good to maintain the separation between the log and the rest
3565          * of XFS.
3566          */
3567         log->l_flags |= XLOG_IO_ERROR;
3568
3569         /*
3570          * If we hit a log error, we want to mark all the iclogs IOERROR
3571          * while we're still holding the loglock.
3572          */
3573         if (logerror)
3574                 retval = xlog_state_ioerror(log);
3575         spin_unlock(&log->l_icloglock);
3576
3577         /*
3578          * We don't want anybody waiting for log reservations after this. That
3579          * means we have to wake up everybody queued up on reserveq as well as
3580          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3581          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3582          * action is protected by the grant locks.
3583          */
3584         xlog_grant_head_wake_all(&log->l_reserve_head);
3585         xlog_grant_head_wake_all(&log->l_write_head);
3586
3587         if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3588                 ASSERT(!logerror);
3589                 /*
3590                  * Force the incore logs to disk before shutting the
3591                  * log down completely.
3592                  */
3593                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3594
3595                 spin_lock(&log->l_icloglock);
3596                 retval = xlog_state_ioerror(log);
3597                 spin_unlock(&log->l_icloglock);
3598         }
3599         /*
3600          * Wake up everybody waiting on xfs_log_force.
3601          * Callback all log item committed functions as if the
3602          * log writes were completed.
3603          */
3604         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3605
3606 #ifdef XFSERRORDEBUG
3607         {
3608                 xlog_in_core_t  *iclog;
3609
3610                 spin_lock(&log->l_icloglock);
3611                 iclog = log->l_iclog;
3612                 do {
3613                         ASSERT(iclog->ic_callback == 0);
3614                         iclog = iclog->ic_next;
3615                 } while (iclog != log->l_iclog);
3616                 spin_unlock(&log->l_icloglock);
3617         }
3618 #endif
3619         /* return non-zero if log IOERROR transition had already happened */
3620         return retval;
3621 }
3622
3623 STATIC int
3624 xlog_iclogs_empty(xlog_t *log)
3625 {
3626         xlog_in_core_t  *iclog;
3627
3628         iclog = log->l_iclog;
3629         do {
3630                 /* endianness does not matter here, zero is zero in
3631                  * any language.
3632                  */
3633                 if (iclog->ic_header.h_num_logops)
3634                         return 0;
3635                 iclog = iclog->ic_next;
3636         } while (iclog != log->l_iclog);
3637         return 1;
3638 }