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_bit.h"
  22 #include "xfs_log.h"
  23 #include "xfs_inum.h"
  24 #include "xfs_trans.h"
  25 #include "xfs_sb.h"
  26 #include "xfs_ag.h"
  27 #include "xfs_mount.h"
  28 #include "xfs_error.h"
  29 #include "xfs_log_priv.h"
  30 #include "xfs_buf_item.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_alloc_btree.h"
  33 #include "xfs_ialloc_btree.h"
  34 #include "xfs_log_recover.h"
  35 #include "xfs_trans_priv.h"
  36 #include "xfs_dinode.h"
  37 #include "xfs_inode.h"
  38 #include "xfs_rw.h"
  39 #include "xfs_trace.h"
  40
  41 kmem_zone_t     *xfs_log_ticket_zone;
  42
  43 /* Local miscellaneous function prototypes */
  44 STATIC int       xlog_commit_record(struct log *log, struct xlog_ticket *ticket,
  45                                     xlog_in_core_t **, xfs_lsn_t *);
  46 STATIC xlog_t *  xlog_alloc_log(xfs_mount_t     *mp,
  47                                 xfs_buftarg_t   *log_target,
  48                                 xfs_daddr_t     blk_offset,
  49                                 int             num_bblks);
  50 STATIC int       xlog_space_left(struct log *log, atomic64_t *head);
  51 STATIC int       xlog_sync(xlog_t *log, xlog_in_core_t *iclog);
  52 STATIC void      xlog_dealloc_log(xlog_t *log);
  53
  54 /* local state machine functions */
  55 STATIC void xlog_state_done_syncing(xlog_in_core_t *iclog, int);
  56 STATIC void xlog_state_do_callback(xlog_t *log,int aborted, xlog_in_core_t *iclog);
  57 STATIC int  xlog_state_get_iclog_space(xlog_t           *log,
  58                                        int              len,
  59                                        xlog_in_core_t   **iclog,
  60                                        xlog_ticket_t    *ticket,
  61                                        int              *continued_write,
  62                                        int              *logoffsetp);
  63 STATIC int  xlog_state_release_iclog(xlog_t             *log,
  64                                      xlog_in_core_t     *iclog);
  65 STATIC void xlog_state_switch_iclogs(xlog_t             *log,
  66                                      xlog_in_core_t *iclog,
  67                                      int                eventual_size);
  68 STATIC void xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog);
  69
  70 STATIC void xlog_grant_push_ail(struct log      *log,
  71                                 int             need_bytes);
  72 STATIC void xlog_regrant_reserve_log_space(xlog_t        *log,
  73                                            xlog_ticket_t *ticket);
  74 STATIC void xlog_ungrant_log_space(xlog_t        *log,
  75                                    xlog_ticket_t *ticket);
  76
  77 #if defined(DEBUG)
  78 STATIC void     xlog_verify_dest_ptr(xlog_t *log, char *ptr);
  79 STATIC void     xlog_verify_grant_tail(struct log *log);
  80 STATIC void     xlog_verify_iclog(xlog_t *log, xlog_in_core_t *iclog,
  81                                   int count, boolean_t syncing);
  82 STATIC void     xlog_verify_tail_lsn(xlog_t *log, xlog_in_core_t *iclog,
  83                                      xfs_lsn_t tail_lsn);
  84 #else
  85 #define xlog_verify_dest_ptr(a,b)
  86 #define xlog_verify_grant_tail(a)
  87 #define xlog_verify_iclog(a,b,c,d)
  88 #define xlog_verify_tail_lsn(a,b,c)
  89 #endif
  90
  91 STATIC int      xlog_iclogs_empty(xlog_t *log);
  92
  93 static void
  94 xlog_grant_sub_space(
  95         struct log      *log,
  96         atomic64_t      *head,
  97         int             bytes)
  98 {
  99         int64_t head_val = atomic64_read(head);
 100         int64_t new, old;
 101
 102         do {
 103                 int     cycle, space;
 104
 105                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 106
 107                 space -= bytes;
 108                 if (space < 0) {
 109                         space += log->l_logsize;
 110                         cycle--;
 111                 }
 112
 113                 old = head_val;
 114                 new = xlog_assign_grant_head_val(cycle, space);
 115                 head_val = atomic64_cmpxchg(head, old, new);
 116         } while (head_val != old);
 117 }
 118
 119 static void
 120 xlog_grant_add_space(
 121         struct log      *log,
 122         atomic64_t      *head,
 123         int             bytes)
 124 {
 125         int64_t head_val = atomic64_read(head);
 126         int64_t new, old;
 127
 128         do {
 129                 int             tmp;
 130                 int             cycle, space;
 131
 132                 xlog_crack_grant_head_val(head_val, &cycle, &space);
 133
 134                 tmp = log->l_logsize - space;
 135                 if (tmp > bytes)
 136                         space += bytes;
 137                 else {
 138                         space = bytes - tmp;
 139                         cycle++;
 140                 }
 141
 142                 old = head_val;
 143                 new = xlog_assign_grant_head_val(cycle, space);
 144                 head_val = atomic64_cmpxchg(head, old, new);
 145         } while (head_val != old);
 146 }
 147
 148 STATIC void
 149 xlog_grant_head_init(
 150         struct xlog_grant_head  *head)
 151 {
 152         xlog_assign_grant_head(&head->grant, 1, 0);
 153         INIT_LIST_HEAD(&head->waiters);
 154         spin_lock_init(&head->lock);
 155 }
 156
 157 STATIC void
 158 xlog_grant_head_wake_all(
 159         struct xlog_grant_head  *head)
 160 {
 161         struct xlog_ticket      *tic;
 162
 163         spin_lock(&head->lock);
 164         list_for_each_entry(tic, &head->waiters, t_queue)
 165                 wake_up_process(tic->t_task);
 166         spin_unlock(&head->lock);
 167 }
 168
 169 static inline int
 170 xlog_ticket_reservation(
 171         struct log              *log,
 172         struct xlog_grant_head  *head,
 173         struct xlog_ticket      *tic)
 174 {
 175         if (head == &log->l_write_head) {
 176                 ASSERT(tic->t_flags & XLOG_TIC_PERM_RESERV);
 177                 return tic->t_unit_res;
 178         } else {
 179                 if (tic->t_flags & XLOG_TIC_PERM_RESERV)
 180                         return tic->t_unit_res * tic->t_cnt;
 181                 else
 182                         return tic->t_unit_res;
 183         }
 184 }
 185
 186 STATIC bool
 187 xlog_grant_head_wake(
 188         struct log              *log,
 189         struct xlog_grant_head  *head,
 190         int                     *free_bytes)
 191 {
 192         struct xlog_ticket      *tic;
 193         int                     need_bytes;
 194
 195         list_for_each_entry(tic, &head->waiters, t_queue) {
 196                 need_bytes = xlog_ticket_reservation(log, head, tic);
 197                 if (*free_bytes < need_bytes)
 198                         return false;
 199
 200                 *free_bytes -= need_bytes;
 201                 trace_xfs_log_grant_wake_up(log, tic);
 202                 wake_up_process(tic->t_task);
 203         }
 204
 205         return true;
 206 }
 207
 208 STATIC int
 209 xlog_grant_head_wait(
 210         struct log              *log,
 211         struct xlog_grant_head  *head,
 212         struct xlog_ticket      *tic,
 213         int                     need_bytes)
 214 {
 215         list_add_tail(&tic->t_queue, &head->waiters);
 216
 217         do {
 218                 if (XLOG_FORCED_SHUTDOWN(log))
 219                         goto shutdown;
 220                 xlog_grant_push_ail(log, need_bytes);
 221
 222                 __set_current_state(TASK_UNINTERRUPTIBLE);
 223                 spin_unlock(&head->lock);
 224
 225                 XFS_STATS_INC(xs_sleep_logspace);
 226
 227                 trace_xfs_log_grant_sleep(log, tic);
 228                 schedule();
 229                 trace_xfs_log_grant_wake(log, tic);
 230
 231                 spin_lock(&head->lock);
 232                 if (XLOG_FORCED_SHUTDOWN(log))
 233                         goto shutdown;
 234         } while (xlog_space_left(log, &head->grant) < need_bytes);
 235
 236         list_del_init(&tic->t_queue);
 237         return 0;
 238 shutdown:
 239         list_del_init(&tic->t_queue);
 240         return XFS_ERROR(EIO);
 241 }
 242
 243 /*
 244  * Atomically get the log space required for a log ticket.
 245  *
 246  * Once a ticket gets put onto head->waiters, it will only return after the
 247  * needed reservation is satisfied.
 248  *
 249  * This function is structured so that it has a lock free fast path. This is
 250  * necessary because every new transaction reservation will come through this
 251  * path. Hence any lock will be globally hot if we take it unconditionally on
 252  * every pass.
 253  *
 254  * As tickets are only ever moved on and off head->waiters under head->lock, we
 255  * only need to take that lock if we are going to add the ticket to the queue
 256  * and sleep. We can avoid taking the lock if the ticket was never added to
 257  * head->waiters because the t_queue list head will be empty and we hold the
 258  * only reference to it so it can safely be checked unlocked.
 259  */
 260 STATIC int
 261 xlog_grant_head_check(
 262         struct log              *log,
 263         struct xlog_grant_head  *head,
 264         struct xlog_ticket      *tic,
 265         int                     *need_bytes)
 266 {
 267         int                     free_bytes;
 268         int                     error = 0;
 269
 270         ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 271
 272         /*
 273          * If there are other waiters on the queue then give them a chance at
 274          * logspace before us.  Wake up the first waiters, if we do not wake
 275          * up all the waiters then go to sleep waiting for more free space,
 276          * otherwise try to get some space for this transaction.
 277          */
 278         *need_bytes = xlog_ticket_reservation(log, head, tic);
 279         free_bytes = xlog_space_left(log, &head->grant);
 280         if (!list_empty_careful(&head->waiters)) {
 281                 spin_lock(&head->lock);
 282                 if (!xlog_grant_head_wake(log, head, &free_bytes) ||
 283                     free_bytes < *need_bytes) {
 284                         error = xlog_grant_head_wait(log, head, tic,
 285                                                      *need_bytes);
 286                 }
 287                 spin_unlock(&head->lock);
 288         } else if (free_bytes < *need_bytes) {
 289                 spin_lock(&head->lock);
 290                 error = xlog_grant_head_wait(log, head, tic, *need_bytes);
 291                 spin_unlock(&head->lock);
 292         }
 293
 294         return error;
 295 }
 296
 297 static void
 298 xlog_tic_reset_res(xlog_ticket_t *tic)
 299 {
 300         tic->t_res_num = 0;
 301         tic->t_res_arr_sum = 0;
 302         tic->t_res_num_ophdrs = 0;
 303 }
 304
 305 static void
 306 xlog_tic_add_region(xlog_ticket_t *tic, uint len, uint type)
 307 {
 308         if (tic->t_res_num == XLOG_TIC_LEN_MAX) {
 309                 /* add to overflow and start again */
 310                 tic->t_res_o_flow += tic->t_res_arr_sum;
 311                 tic->t_res_num = 0;
 312                 tic->t_res_arr_sum = 0;
 313         }
 314
 315         tic->t_res_arr[tic->t_res_num].r_len = len;
 316         tic->t_res_arr[tic->t_res_num].r_type = type;
 317         tic->t_res_arr_sum += len;
 318         tic->t_res_num++;
 319 }
 320
 321 /*
 322  * Replenish the byte reservation required by moving the grant write head.
 323  */
 324 int
 325 xfs_log_regrant(
 326         struct xfs_mount        *mp,
 327         struct xlog_ticket      *tic)
 328 {
 329         struct log              *log = mp->m_log;
 330         int                     need_bytes;
 331         int                     error = 0;
 332
 333         if (XLOG_FORCED_SHUTDOWN(log))
 334                 return XFS_ERROR(EIO);
 335
 336         XFS_STATS_INC(xs_try_logspace);
 337
 338         /*
 339          * This is a new transaction on the ticket, so we need to change the
 340          * transaction ID so that the next transaction has a different TID in
 341          * the log. Just add one to the existing tid so that we can see chains
 342          * of rolling transactions in the log easily.
 343          */
 344         tic->t_tid++;
 345
 346         xlog_grant_push_ail(log, tic->t_unit_res);
 347
 348         tic->t_curr_res = tic->t_unit_res;
 349         xlog_tic_reset_res(tic);
 350
 351         if (tic->t_cnt > 0)
 352                 return 0;
 353
 354         trace_xfs_log_regrant(log, tic);
 355
 356         error = xlog_grant_head_check(log, &log->l_write_head, tic,
 357                                       &need_bytes);
 358         if (error)
 359                 goto out_error;
 360
 361         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 362         trace_xfs_log_regrant_exit(log, tic);
 363         xlog_verify_grant_tail(log);
 364         return 0;
 365
 366 out_error:
 367         /*
 368          * If we are failing, make sure the ticket doesn't have any current
 369          * reservations.  We don't want to add this back when the ticket/
 370          * transaction gets cancelled.
 371          */
 372         tic->t_curr_res = 0;
 373         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 374         return error;
 375 }
 376
 377 /*
 378  * Reserve log space and return a ticket corresponding the reservation.
 379  *
 380  * Each reservation is going to reserve extra space for a log record header.
 381  * When writes happen to the on-disk log, we don't subtract the length of the
 382  * log record header from any reservation.  By wasting space in each
 383  * reservation, we prevent over allocation problems.
 384  */
 385 int
 386 xfs_log_reserve(
 387         struct xfs_mount        *mp,
 388         int                     unit_bytes,
 389         int                     cnt,
 390         struct xlog_ticket      **ticp,
 391         __uint8_t               client,
 392         bool                    permanent,
 393         uint                    t_type)
 394 {
 395         struct log              *log = mp->m_log;
 396         struct xlog_ticket      *tic;
 397         int                     need_bytes;
 398         int                     error = 0;
 399
 400         ASSERT(client == XFS_TRANSACTION || client == XFS_LOG);
 401
 402         if (XLOG_FORCED_SHUTDOWN(log))
 403                 return XFS_ERROR(EIO);
 404
 405         XFS_STATS_INC(xs_try_logspace);
 406
 407         ASSERT(*ticp == NULL);
 408         tic = xlog_ticket_alloc(log, unit_bytes, cnt, client, permanent,
 409                                 KM_SLEEP | KM_MAYFAIL);
 410         if (!tic)
 411                 return XFS_ERROR(ENOMEM);
 412
 413         tic->t_trans_type = t_type;
 414         *ticp = tic;
 415
 416         xlog_grant_push_ail(log, tic->t_unit_res * tic->t_cnt);
 417
 418         trace_xfs_log_reserve(log, tic);
 419
 420         error = xlog_grant_head_check(log, &log->l_reserve_head, tic,
 421                                       &need_bytes);
 422         if (error)
 423                 goto out_error;
 424
 425         xlog_grant_add_space(log, &log->l_reserve_head.grant, need_bytes);
 426         xlog_grant_add_space(log, &log->l_write_head.grant, need_bytes);
 427         trace_xfs_log_reserve_exit(log, tic);
 428         xlog_verify_grant_tail(log);
 429         return 0;
 430
 431 out_error:
 432         /*
 433          * If we are failing, make sure the ticket doesn't have any current
 434          * reservations.  We don't want to add this back when the ticket/
 435          * transaction gets cancelled.
 436          */
 437         tic->t_curr_res = 0;
 438         tic->t_cnt = 0; /* ungrant will give back unit_res * t_cnt. */
 439         return error;
 440 }
 441
 442
 443 /*
 444  * NOTES:
 445  *
 446  *      1. currblock field gets updated at startup and after in-core logs
 447  *              marked as with WANT_SYNC.
 448  */
 449
 450 /*
 451  * This routine is called when a user of a log manager ticket is done with
 452  * the reservation.  If the ticket was ever used, then a commit record for
 453  * the associated transaction is written out as a log operation header with
 454  * no data.  The flag XLOG_TIC_INITED is set when the first write occurs with
 455  * a given ticket.  If the ticket was one with a permanent reservation, then
 456  * a few operations are done differently.  Permanent reservation tickets by
 457  * default don't release the reservation.  They just commit the current
 458  * transaction with the belief that the reservation is still needed.  A flag
 459  * must be passed in before permanent reservations are actually released.
 460  * When these type of tickets are not released, they need to be set into
 461  * the inited state again.  By doing this, a start record will be written
 462  * out when the next write occurs.
 463  */
 464 xfs_lsn_t
 465 xfs_log_done(
 466         struct xfs_mount        *mp,
 467         struct xlog_ticket      *ticket,
 468         struct xlog_in_core     **iclog,
 469         uint                    flags)
 470 {
 471         struct log              *log = mp->m_log;
 472         xfs_lsn_t               lsn = 0;
 473
 474         if (XLOG_FORCED_SHUTDOWN(log) ||
 475             /*
 476              * If nothing was ever written, don't write out commit record.
 477              * If we get an error, just continue and give back the log ticket.
 478              */
 479             (((ticket->t_flags & XLOG_TIC_INITED) == 0) &&
 480              (xlog_commit_record(log, ticket, iclog, &lsn)))) {
 481                 lsn = (xfs_lsn_t) -1;
 482                 if (ticket->t_flags & XLOG_TIC_PERM_RESERV) {
 483                         flags |= XFS_LOG_REL_PERM_RESERV;
 484                 }
 485         }
 486
 487
 488         if ((ticket->t_flags & XLOG_TIC_PERM_RESERV) == 0 ||
 489             (flags & XFS_LOG_REL_PERM_RESERV)) {
 490                 trace_xfs_log_done_nonperm(log, ticket);
 491
 492                 /*
 493                  * Release ticket if not permanent reservation or a specific
 494                  * request has been made to release a permanent reservation.
 495                  */
 496                 xlog_ungrant_log_space(log, ticket);
 497                 xfs_log_ticket_put(ticket);
 498         } else {
 499                 trace_xfs_log_done_perm(log, ticket);
 500
 501                 xlog_regrant_reserve_log_space(log, ticket);
 502                 /* If this ticket was a permanent reservation and we aren't
 503                  * trying to release it, reset the inited flags; so next time
 504                  * we write, a start record will be written out.
 505                  */
 506                 ticket->t_flags |= XLOG_TIC_INITED;
 507         }
 508
 509         return lsn;
 510 }
 511
 512 /*
 513  * Attaches a new iclog I/O completion callback routine during
 514  * transaction commit.  If the log is in error state, a non-zero
 515  * return code is handed back and the caller is responsible for
 516  * executing the callback at an appropriate time.
 517  */
 518 int
 519 xfs_log_notify(
 520         struct xfs_mount        *mp,
 521         struct xlog_in_core     *iclog,
 522         xfs_log_callback_t      *cb)
 523 {
 524         int     abortflg;
 525
 526         spin_lock(&iclog->ic_callback_lock);
 527         abortflg = (iclog->ic_state & XLOG_STATE_IOERROR);
 528         if (!abortflg) {
 529                 ASSERT_ALWAYS((iclog->ic_state == XLOG_STATE_ACTIVE) ||
 530                               (iclog->ic_state == XLOG_STATE_WANT_SYNC));
 531                 cb->cb_next = NULL;
 532                 *(iclog->ic_callback_tail) = cb;
 533                 iclog->ic_callback_tail = &(cb->cb_next);
 534         }
 535         spin_unlock(&iclog->ic_callback_lock);
 536         return abortflg;
 537 }
 538
 539 int
 540 xfs_log_release_iclog(
 541         struct xfs_mount        *mp,
 542         struct xlog_in_core     *iclog)
 543 {
 544         if (xlog_state_release_iclog(mp->m_log, iclog)) {
 545                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
 546                 return EIO;
 547         }
 548
 549         return 0;
 550 }
 551
 552 /*
 553  * Mount a log filesystem
 554  *
 555  * mp           - ubiquitous xfs mount point structure
 556  * log_target   - buftarg of on-disk log device
 557  * blk_offset   - Start block # where block size is 512 bytes (BBSIZE)
 558  * num_bblocks  - Number of BBSIZE blocks in on-disk log
 559  *
 560  * Return error or zero.
 561  */
 562 int
 563 xfs_log_mount(
 564         xfs_mount_t     *mp,
 565         xfs_buftarg_t   *log_target,
 566         xfs_daddr_t     blk_offset,
 567         int             num_bblks)
 568 {
 569         int             error;
 570
 571         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
 572                 xfs_notice(mp, "Mounting Filesystem");
 573         else {
 574                 xfs_notice(mp,
 575 "Mounting filesystem in no-recovery mode.  Filesystem will be inconsistent.");
 576                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 577         }
 578
 579         mp->m_log = xlog_alloc_log(mp, log_target, blk_offset, num_bblks);
 580         if (IS_ERR(mp->m_log)) {
 581                 error = -PTR_ERR(mp->m_log);
 582                 goto out;
 583         }
 584
 585         /*
 586          * Initialize the AIL now we have a log.
 587          */
 588         error = xfs_trans_ail_init(mp);
 589         if (error) {
 590                 xfs_warn(mp, "AIL initialisation failed: error %d", error);
 591                 goto out_free_log;
 592         }
 593         mp->m_log->l_ailp = mp->m_ail;
 594
 595         /*
 596          * skip log recovery on a norecovery mount.  pretend it all
 597          * just worked.
 598          */
 599         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY)) {
 600                 int     readonly = (mp->m_flags & XFS_MOUNT_RDONLY);
 601
 602                 if (readonly)
 603                         mp->m_flags &= ~XFS_MOUNT_RDONLY;
 604
 605                 error = xlog_recover(mp->m_log);
 606
 607                 if (readonly)
 608                         mp->m_flags |= XFS_MOUNT_RDONLY;
 609                 if (error) {
 610                         xfs_warn(mp, "log mount/recovery failed: error %d",
 611                                 error);
 612                         goto out_destroy_ail;
 613                 }
 614         }
 615
 616         /* Normal transactions can now occur */
 617         mp->m_log->l_flags &= ~XLOG_ACTIVE_RECOVERY;
 618
 619         /*
 620          * Now the log has been fully initialised and we know were our
 621          * space grant counters are, we can initialise the permanent ticket
 622          * needed for delayed logging to work.
 623          */
 624         xlog_cil_init_post_recovery(mp->m_log);
 625
 626         return 0;
 627
 628 out_destroy_ail:
 629         xfs_trans_ail_destroy(mp);
 630 out_free_log:
 631         xlog_dealloc_log(mp->m_log);
 632 out:
 633         return error;
 634 }
 635
 636 /*
 637  * Finish the recovery of the file system.  This is separate from
 638  * the xfs_log_mount() call, because it depends on the code in
 639  * xfs_mountfs() to read in the root and real-time bitmap inodes
 640  * between calling xfs_log_mount() and here.
 641  *
 642  * mp           - ubiquitous xfs mount point structure
 643  */
 644 int
 645 xfs_log_mount_finish(xfs_mount_t *mp)
 646 {
 647         int     error;
 648
 649         if (!(mp->m_flags & XFS_MOUNT_NORECOVERY))
 650                 error = xlog_recover_finish(mp->m_log);
 651         else {
 652                 error = 0;
 653                 ASSERT(mp->m_flags & XFS_MOUNT_RDONLY);
 654         }
 655
 656         return error;
 657 }
 658
 659 /*
 660  * Final log writes as part of unmount.
 661  *
 662  * Mark the filesystem clean as unmount happens.  Note that during relocation
 663  * this routine needs to be executed as part of source-bag while the
 664  * deallocation must not be done until source-end.
 665  */
 666
 667 /*
 668  * Unmount record used to have a string "Unmount filesystem--" in the
 669  * data section where the "Un" was really a magic number (XLOG_UNMOUNT_TYPE).
 670  * We just write the magic number now since that particular field isn't
 671  * currently architecture converted and "nUmount" is a bit foo.
 672  * As far as I know, there weren't any dependencies on the old behaviour.
 673  */
 674
 675 int
 676 xfs_log_unmount_write(xfs_mount_t *mp)
 677 {
 678         xlog_t           *log = mp->m_log;
 679         xlog_in_core_t   *iclog;
 680 #ifdef DEBUG
 681         xlog_in_core_t   *first_iclog;
 682 #endif
 683         xlog_ticket_t   *tic = NULL;
 684         xfs_lsn_t        lsn;
 685         int              error;
 686
 687         /*
 688          * Don't write out unmount record on read-only mounts.
 689          * Or, if we are doing a forced umount (typically because of IO errors).
 690          */
 691         if (mp->m_flags & XFS_MOUNT_RDONLY)
 692                 return 0;
 693
 694         error = _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
 695         ASSERT(error || !(XLOG_FORCED_SHUTDOWN(log)));
 696
 697 #ifdef DEBUG
 698         first_iclog = iclog = log->l_iclog;
 699         do {
 700                 if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
 701                         ASSERT(iclog->ic_state & XLOG_STATE_ACTIVE);
 702                         ASSERT(iclog->ic_offset == 0);
 703                 }
 704                 iclog = iclog->ic_next;
 705         } while (iclog != first_iclog);
 706 #endif
 707         if (! (XLOG_FORCED_SHUTDOWN(log))) {
 708                 error = xfs_log_reserve(mp, 600, 1, &tic,
 709                                         XFS_LOG, 0, XLOG_UNMOUNT_REC_TYPE);
 710                 if (!error) {
 711                         /* the data section must be 32 bit size aligned */
 712                         struct {
 713                             __uint16_t magic;
 714                             __uint16_t pad1;
 715                             __uint32_t pad2; /* may as well make it 64 bits */
 716                         } magic = {
 717                                 .magic = XLOG_UNMOUNT_TYPE,
 718                         };
 719                         struct xfs_log_iovec reg = {
 720                                 .i_addr = &magic,
 721                                 .i_len = sizeof(magic),
 722                                 .i_type = XLOG_REG_TYPE_UNMOUNT,
 723                         };
 724                         struct xfs_log_vec vec = {
 725                                 .lv_niovecs = 1,
 726                                 .lv_iovecp = &reg,
 727                         };
 728
 729                         /* remove inited flag, and account for space used */
 730                         tic->t_flags = 0;
 731                         tic->t_curr_res -= sizeof(magic);
 732                         error = xlog_write(log, &vec, tic, &lsn,
 733                                            NULL, XLOG_UNMOUNT_TRANS);
 734                         /*
 735                          * At this point, we're umounting anyway,
 736                          * so there's no point in transitioning log state
 737                          * to IOERROR. Just continue...
 738                          */
 739                 }
 740
 741                 if (error)
 742                         xfs_alert(mp, "%s: unmount record failed", __func__);
 743
 744
 745                 spin_lock(&log->l_icloglock);
 746                 iclog = log->l_iclog;
 747                 atomic_inc(&iclog->ic_refcnt);
 748                 xlog_state_want_sync(log, iclog);
 749                 spin_unlock(&log->l_icloglock);
 750                 error = xlog_state_release_iclog(log, iclog);
 751
 752                 spin_lock(&log->l_icloglock);
 753                 if (!(iclog->ic_state == XLOG_STATE_ACTIVE ||
 754                       iclog->ic_state == XLOG_STATE_DIRTY)) {
 755                         if (!XLOG_FORCED_SHUTDOWN(log)) {
 756                                 xlog_wait(&iclog->ic_force_wait,
 757                                                         &log->l_icloglock);
 758                         } else {
 759                                 spin_unlock(&log->l_icloglock);
 760                         }
 761                 } else {
 762                         spin_unlock(&log->l_icloglock);
 763                 }
 764                 if (tic) {
 765                         trace_xfs_log_umount_write(log, tic);
 766                         xlog_ungrant_log_space(log, tic);
 767                         xfs_log_ticket_put(tic);
 768                 }
 769         } else {
 770                 /*
 771                  * We're already in forced_shutdown mode, couldn't
 772                  * even attempt to write out the unmount transaction.
 773                  *
 774                  * Go through the motions of sync'ing and releasing
 775                  * the iclog, even though no I/O will actually happen,
 776                  * we need to wait for other log I/Os that may already
 777                  * be in progress.  Do this as a separate section of
 778                  * code so we'll know if we ever get stuck here that
 779                  * we're in this odd situation of trying to unmount
 780                  * a file system that went into forced_shutdown as
 781                  * the result of an unmount..
 782                  */
 783                 spin_lock(&log->l_icloglock);
 784                 iclog = log->l_iclog;
 785                 atomic_inc(&iclog->ic_refcnt);
 786
 787                 xlog_state_want_sync(log, iclog);
 788                 spin_unlock(&log->l_icloglock);
 789                 error =  xlog_state_release_iclog(log, iclog);
 790
 791                 spin_lock(&log->l_icloglock);
 792
 793                 if ( ! (   iclog->ic_state == XLOG_STATE_ACTIVE
 794                         || iclog->ic_state == XLOG_STATE_DIRTY
 795                         || iclog->ic_state == XLOG_STATE_IOERROR) ) {
 796
 797                                 xlog_wait(&iclog->ic_force_wait,
 798                                                         &log->l_icloglock);
 799                 } else {
 800                         spin_unlock(&log->l_icloglock);
 801                 }
 802         }
 803
 804         return error;
 805 }       /* xfs_log_unmount_write */
 806
 807 /*
 808  * Deallocate log structures for unmount/relocation.
 809  *
 810  * We need to stop the aild from running before we destroy
 811  * and deallocate the log as the aild references the log.
 812  */
 813 void
 814 xfs_log_unmount(xfs_mount_t *mp)
 815 {
 816         xfs_trans_ail_destroy(mp);
 817         xlog_dealloc_log(mp->m_log);
 818 }
 819
 820 void
 821 xfs_log_item_init(
 822         struct xfs_mount        *mp,
 823         struct xfs_log_item     *item,
 824         int                     type,
 825         const struct xfs_item_ops *ops)
 826 {
 827         item->li_mountp = mp;
 828         item->li_ailp = mp->m_ail;
 829         item->li_type = type;
 830         item->li_ops = ops;
 831         item->li_lv = NULL;
 832
 833         INIT_LIST_HEAD(&item->li_ail);
 834         INIT_LIST_HEAD(&item->li_cil);
 835 }
 836
 837 /*
 838  * Wake up processes waiting for log space after we have moved the log tail.
 839  */
 840 void
 841 xfs_log_space_wake(
 842         struct xfs_mount        *mp)
 843 {
 844         struct log              *log = mp->m_log;
 845         int                     free_bytes;
 846
 847         if (XLOG_FORCED_SHUTDOWN(log))
 848                 return;
 849
 850         if (!list_empty_careful(&log->l_write_head.waiters)) {
 851                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 852
 853                 spin_lock(&log->l_write_head.lock);
 854                 free_bytes = xlog_space_left(log, &log->l_write_head.grant);
 855                 xlog_grant_head_wake(log, &log->l_write_head, &free_bytes);
 856                 spin_unlock(&log->l_write_head.lock);
 857         }
 858
 859         if (!list_empty_careful(&log->l_reserve_head.waiters)) {
 860                 ASSERT(!(log->l_flags & XLOG_ACTIVE_RECOVERY));
 861
 862                 spin_lock(&log->l_reserve_head.lock);
 863                 free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
 864                 xlog_grant_head_wake(log, &log->l_reserve_head, &free_bytes);
 865                 spin_unlock(&log->l_reserve_head.lock);
 866         }
 867 }
 868
 869 /*
 870  * Determine if we have a transaction that has gone to disk
 871  * that needs to be covered. To begin the transition to the idle state
 872  * firstly the log needs to be idle (no AIL and nothing in the iclogs).
 873  * If we are then in a state where covering is needed, the caller is informed
 874  * that dummy transactions are required to move the log into the idle state.
 875  *
 876  * Because this is called as part of the sync process, we should also indicate
 877  * that dummy transactions should be issued in anything but the covered or
 878  * idle states. This ensures that the log tail is accurately reflected in
 879  * the log at the end of the sync, hence if a crash occurrs avoids replay
 880  * of transactions where the metadata is already on disk.
 881  */
 882 int
 883 xfs_log_need_covered(xfs_mount_t *mp)
 884 {
 885         int             needed = 0;
 886         xlog_t          *log = mp->m_log;
 887
 888         if (!xfs_fs_writable(mp))
 889                 return 0;
 890
 891         spin_lock(&log->l_icloglock);
 892         switch (log->l_covered_state) {
 893         case XLOG_STATE_COVER_DONE:
 894         case XLOG_STATE_COVER_DONE2:
 895         case XLOG_STATE_COVER_IDLE:
 896                 break;
 897         case XLOG_STATE_COVER_NEED:
 898         case XLOG_STATE_COVER_NEED2:
 899                 if (!xfs_ail_min_lsn(log->l_ailp) &&
 900                     xlog_iclogs_empty(log)) {
 901                         if (log->l_covered_state == XLOG_STATE_COVER_NEED)
 902                                 log->l_covered_state = XLOG_STATE_COVER_DONE;
 903                         else
 904                                 log->l_covered_state = XLOG_STATE_COVER_DONE2;
 905                 }
 906                 /* FALLTHRU */
 907         default:
 908                 needed = 1;
 909                 break;
 910         }
 911         spin_unlock(&log->l_icloglock);
 912         return needed;
 913 }
 914
 915 /*
 916  * We may be holding the log iclog lock upon entering this routine.
 917  */
 918 xfs_lsn_t
 919 xlog_assign_tail_lsn_locked(
 920         struct xfs_mount        *mp)
 921 {
 922         struct log              *log = mp->m_log;
 923         struct xfs_log_item     *lip;
 924         xfs_lsn_t               tail_lsn;
 925
 926         assert_spin_locked(&mp->m_ail->xa_lock);
 927
 928         /*
 929          * To make sure we always have a valid LSN for the log tail we keep
 930          * track of the last LSN which was committed in log->l_last_sync_lsn,
 931          * and use that when the AIL was empty.
 932          */
 933         lip = xfs_ail_min(mp->m_ail);
 934         if (lip)
 935                 tail_lsn = lip->li_lsn;
 936         else
 937                 tail_lsn = atomic64_read(&log->l_last_sync_lsn);
 938         atomic64_set(&log->l_tail_lsn, tail_lsn);
 939         return tail_lsn;
 940 }
 941
 942 xfs_lsn_t
 943 xlog_assign_tail_lsn(
 944         struct xfs_mount        *mp)
 945 {
 946         xfs_lsn_t               tail_lsn;
 947
 948         spin_lock(&mp->m_ail->xa_lock);
 949         tail_lsn = xlog_assign_tail_lsn_locked(mp);
 950         spin_unlock(&mp->m_ail->xa_lock);
 951
 952         return tail_lsn;
 953 }
 954
 955 /*
 956  * Return the space in the log between the tail and the head.  The head
 957  * is passed in the cycle/bytes formal parms.  In the special case where
 958  * the reserve head has wrapped passed the tail, this calculation is no
 959  * longer valid.  In this case, just return 0 which means there is no space
 960  * in the log.  This works for all places where this function is called
 961  * with the reserve head.  Of course, if the write head were to ever
 962  * wrap the tail, we should blow up.  Rather than catch this case here,
 963  * we depend on other ASSERTions in other parts of the code.   XXXmiken
 964  *
 965  * This code also handles the case where the reservation head is behind
 966  * the tail.  The details of this case are described below, but the end
 967  * result is that we return the size of the log as the amount of space left.
 968  */
 969 STATIC int
 970 xlog_space_left(
 971         struct log      *log,
 972         atomic64_t      *head)
 973 {
 974         int             free_bytes;
 975         int             tail_bytes;
 976         int             tail_cycle;
 977         int             head_cycle;
 978         int             head_bytes;
 979
 980         xlog_crack_grant_head(head, &head_cycle, &head_bytes);
 981         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_bytes);
 982         tail_bytes = BBTOB(tail_bytes);
 983         if (tail_cycle == head_cycle && head_bytes >= tail_bytes)
 984                 free_bytes = log->l_logsize - (head_bytes - tail_bytes);
 985         else if (tail_cycle + 1 < head_cycle)
 986                 return 0;
 987         else if (tail_cycle < head_cycle) {
 988                 ASSERT(tail_cycle == (head_cycle - 1));
 989                 free_bytes = tail_bytes - head_bytes;
 990         } else {
 991                 /*
 992                  * The reservation head is behind the tail.
 993                  * In this case we just want to return the size of the
 994                  * log as the amount of space left.
 995                  */
 996                 xfs_alert(log->l_mp,
 997                         "xlog_space_left: head behind tail\n"
 998                         "  tail_cycle = %d, tail_bytes = %d\n"
 999                         "  GH   cycle = %d, GH   bytes = %d",
1000                         tail_cycle, tail_bytes, head_cycle, head_bytes);
1001                 ASSERT(0);
1002                 free_bytes = log->l_logsize;
1003         }
1004         return free_bytes;
1005 }
1006
1007
1008 /*
1009  * Log function which is called when an io completes.
1010  *
1011  * The log manager needs its own routine, in order to control what
1012  * happens with the buffer after the write completes.
1013  */
1014 void
1015 xlog_iodone(xfs_buf_t *bp)
1016 {
1017         xlog_in_core_t  *iclog = bp->b_fspriv;
1018         xlog_t          *l = iclog->ic_log;
1019         int             aborted = 0;
1020
1021         /*
1022          * Race to shutdown the filesystem if we see an error.
1023          */
1024         if (XFS_TEST_ERROR((xfs_buf_geterror(bp)), l->l_mp,
1025                         XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
1026                 xfs_buf_ioerror_alert(bp, __func__);
1027                 xfs_buf_stale(bp);
1028                 xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);
1029                 /*
1030                  * This flag will be propagated to the trans-committed
1031                  * callback routines to let them know that the log-commit
1032                  * didn't succeed.
1033                  */
1034                 aborted = XFS_LI_ABORTED;
1035         } else if (iclog->ic_state & XLOG_STATE_IOERROR) {
1036                 aborted = XFS_LI_ABORTED;
1037         }
1038
1039         /* log I/O is always issued ASYNC */
1040         ASSERT(XFS_BUF_ISASYNC(bp));
1041         xlog_state_done_syncing(iclog, aborted);
1042         /*
1043          * do not reference the buffer (bp) here as we could race
1044          * with it being freed after writing the unmount record to the
1045          * log.
1046          */
1047
1048 }       /* xlog_iodone */
1049
1050 /*
1051  * Return size of each in-core log record buffer.
1052  *
1053  * All machines get 8 x 32kB buffers by default, unless tuned otherwise.
1054  *
1055  * If the filesystem blocksize is too large, we may need to choose a
1056  * larger size since the directory code currently logs entire blocks.
1057  */
1058
1059 STATIC void
1060 xlog_get_iclog_buffer_size(xfs_mount_t  *mp,
1061                            xlog_t       *log)
1062 {
1063         int size;
1064         int xhdrs;
1065
1066         if (mp->m_logbufs <= 0)
1067                 log->l_iclog_bufs = XLOG_MAX_ICLOGS;
1068         else
1069                 log->l_iclog_bufs = mp->m_logbufs;
1070
1071         /*
1072          * Buffer size passed in from mount system call.
1073          */
1074         if (mp->m_logbsize > 0) {
1075                 size = log->l_iclog_size = mp->m_logbsize;
1076                 log->l_iclog_size_log = 0;
1077                 while (size != 1) {
1078                         log->l_iclog_size_log++;
1079                         size >>= 1;
1080                 }
1081
1082                 if (xfs_sb_version_haslogv2(&mp->m_sb)) {
1083                         /* # headers = size / 32k
1084                          * one header holds cycles from 32k of data
1085                          */
1086
1087                         xhdrs = mp->m_logbsize / XLOG_HEADER_CYCLE_SIZE;
1088                         if (mp->m_logbsize % XLOG_HEADER_CYCLE_SIZE)
1089                                 xhdrs++;
1090                         log->l_iclog_hsize = xhdrs << BBSHIFT;
1091                         log->l_iclog_heads = xhdrs;
1092                 } else {
1093                         ASSERT(mp->m_logbsize <= XLOG_BIG_RECORD_BSIZE);
1094                         log->l_iclog_hsize = BBSIZE;
1095                         log->l_iclog_heads = 1;
1096                 }
1097                 goto done;
1098         }
1099
1100         /* All machines use 32kB buffers by default. */
1101         log->l_iclog_size = XLOG_BIG_RECORD_BSIZE;
1102         log->l_iclog_size_log = XLOG_BIG_RECORD_BSHIFT;
1103
1104         /* the default log size is 16k or 32k which is one header sector */
1105         log->l_iclog_hsize = BBSIZE;
1106         log->l_iclog_heads = 1;
1107
1108 done:
1109         /* are we being asked to make the sizes selected above visible? */
1110         if (mp->m_logbufs == 0)
1111                 mp->m_logbufs = log->l_iclog_bufs;
1112         if (mp->m_logbsize == 0)
1113                 mp->m_logbsize = log->l_iclog_size;
1114 }       /* xlog_get_iclog_buffer_size */
1115
1116
1117 /*
1118  * This routine initializes some of the log structure for a given mount point.
1119  * Its primary purpose is to fill in enough, so recovery can occur.  However,
1120  * some other stuff may be filled in too.
1121  */
1122 STATIC xlog_t *
1123 xlog_alloc_log(xfs_mount_t      *mp,
1124                xfs_buftarg_t    *log_target,
1125                xfs_daddr_t      blk_offset,
1126                int              num_bblks)
1127 {
1128         xlog_t                  *log;
1129         xlog_rec_header_t       *head;
1130         xlog_in_core_t          **iclogp;
1131         xlog_in_core_t          *iclog, *prev_iclog=NULL;
1132         xfs_buf_t               *bp;
1133         int                     i;
1134         int                     error = ENOMEM;
1135         uint                    log2_size = 0;
1136
1137         log = kmem_zalloc(sizeof(xlog_t), KM_MAYFAIL);
1138         if (!log) {
1139                 xfs_warn(mp, "Log allocation failed: No memory!");
1140                 goto out;
1141         }
1142
1143         log->l_mp          = mp;
1144         log->l_targ        = log_target;
1145         log->l_logsize     = BBTOB(num_bblks);
1146         log->l_logBBstart  = blk_offset;
1147         log->l_logBBsize   = num_bblks;
1148         log->l_covered_state = XLOG_STATE_COVER_IDLE;
1149         log->l_flags       |= XLOG_ACTIVE_RECOVERY;
1150
1151         log->l_prev_block  = -1;
1152         /* log->l_tail_lsn = 0x100000000LL; cycle = 1; current block = 0 */
1153         xlog_assign_atomic_lsn(&log->l_tail_lsn, 1, 0);
1154         xlog_assign_atomic_lsn(&log->l_last_sync_lsn, 1, 0);
1155         log->l_curr_cycle  = 1;     /* 0 is bad since this is initial value */
1156
1157         xlog_grant_head_init(&log->l_reserve_head);
1158         xlog_grant_head_init(&log->l_write_head);
1159
1160         error = EFSCORRUPTED;
1161         if (xfs_sb_version_hassector(&mp->m_sb)) {
1162                 log2_size = mp->m_sb.sb_logsectlog;
1163                 if (log2_size < BBSHIFT) {
1164                         xfs_warn(mp, "Log sector size too small (0x%x < 0x%x)",
1165                                 log2_size, BBSHIFT);
1166                         goto out_free_log;
1167                 }
1168
1169                 log2_size -= BBSHIFT;
1170                 if (log2_size > mp->m_sectbb_log) {
1171                         xfs_warn(mp, "Log sector size too large (0x%x > 0x%x)",
1172                                 log2_size, mp->m_sectbb_log);
1173                         goto out_free_log;
1174                 }
1175
1176                 /* for larger sector sizes, must have v2 or external log */
1177                 if (log2_size && log->l_logBBstart > 0 &&
1178                             !xfs_sb_version_haslogv2(&mp->m_sb)) {
1179                         xfs_warn(mp,
1180                 "log sector size (0x%x) invalid for configuration.",
1181                                 log2_size);
1182                         goto out_free_log;
1183                 }
1184         }
1185         log->l_sectBBsize = 1 << log2_size;
1186
1187         xlog_get_iclog_buffer_size(mp, log);
1188
1189         error = ENOMEM;
1190         bp = xfs_buf_alloc(mp->m_logdev_targp, 0, BTOBB(log->l_iclog_size), 0);
1191         if (!bp)
1192                 goto out_free_log;
1193         bp->b_iodone = xlog_iodone;
1194         ASSERT(xfs_buf_islocked(bp));
1195         log->l_xbuf = bp;
1196
1197         spin_lock_init(&log->l_icloglock);
1198         init_waitqueue_head(&log->l_flush_wait);
1199
1200         /* log record size must be multiple of BBSIZE; see xlog_rec_header_t */
1201         ASSERT((XFS_BUF_SIZE(bp) & BBMASK) == 0);
1202
1203         iclogp = &log->l_iclog;
1204         /*
1205          * The amount of memory to allocate for the iclog structure is
1206          * rather funky due to the way the structure is defined.  It is
1207          * done this way so that we can use different sizes for machines
1208          * with different amounts of memory.  See the definition of
1209          * xlog_in_core_t in xfs_log_priv.h for details.
1210          */
1211         ASSERT(log->l_iclog_size >= 4096);
1212         for (i=0; i < log->l_iclog_bufs; i++) {
1213                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1214                 if (!*iclogp)
1215                         goto out_free_iclog;
1216
1217                 iclog = *iclogp;
1218                 iclog->ic_prev = prev_iclog;
1219                 prev_iclog = iclog;
1220
1221                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1222                                                 BTOBB(log->l_iclog_size), 0);
1223                 if (!bp)
1224                         goto out_free_iclog;
1225
1226                 bp->b_iodone = xlog_iodone;
1227                 iclog->ic_bp = bp;
1228                 iclog->ic_data = bp->b_addr;
1229 #ifdef DEBUG
1230                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1231 #endif
1232                 head = &iclog->ic_header;
1233                 memset(head, 0, sizeof(xlog_rec_header_t));
1234                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1235                 head->h_version = cpu_to_be32(
1236                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1237                 head->h_size = cpu_to_be32(log->l_iclog_size);
1238                 /* new fields */
1239                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1240                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1241
1242                 iclog->ic_size = XFS_BUF_SIZE(bp) - log->l_iclog_hsize;
1243                 iclog->ic_state = XLOG_STATE_ACTIVE;
1244                 iclog->ic_log = log;
1245                 atomic_set(&iclog->ic_refcnt, 0);
1246                 spin_lock_init(&iclog->ic_callback_lock);
1247                 iclog->ic_callback_tail = &(iclog->ic_callback);
1248                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1249
1250                 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1251                 init_waitqueue_head(&iclog->ic_force_wait);
1252                 init_waitqueue_head(&iclog->ic_write_wait);
1253
1254                 iclogp = &iclog->ic_next;
1255         }
1256         *iclogp = log->l_iclog;                 /* complete ring */
1257         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1258
1259         error = xlog_cil_init(log);
1260         if (error)
1261                 goto out_free_iclog;
1262         return log;
1263
1264 out_free_iclog:
1265         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1266                 prev_iclog = iclog->ic_next;
1267                 if (iclog->ic_bp)
1268                         xfs_buf_free(iclog->ic_bp);
1269                 kmem_free(iclog);
1270         }
1271         spinlock_destroy(&log->l_icloglock);
1272         xfs_buf_free(log->l_xbuf);
1273 out_free_log:
1274         kmem_free(log);
1275 out:
1276         return ERR_PTR(-error);
1277 }       /* xlog_alloc_log */
1278
1279
1280 /*
1281  * Write out the commit record of a transaction associated with the given
1282  * ticket.  Return the lsn of the commit record.
1283  */
1284 STATIC int
1285 xlog_commit_record(
1286         struct log              *log,
1287         struct xlog_ticket      *ticket,
1288         struct xlog_in_core     **iclog,
1289         xfs_lsn_t               *commitlsnp)
1290 {
1291         struct xfs_mount *mp = log->l_mp;
1292         int     error;
1293         struct xfs_log_iovec reg = {
1294                 .i_addr = NULL,
1295                 .i_len = 0,
1296                 .i_type = XLOG_REG_TYPE_COMMIT,
1297         };
1298         struct xfs_log_vec vec = {
1299                 .lv_niovecs = 1,
1300                 .lv_iovecp = &reg,
1301         };
1302
1303         ASSERT_ALWAYS(iclog);
1304         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1305                                         XLOG_COMMIT_TRANS);
1306         if (error)
1307                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1308         return error;
1309 }
1310
1311 /*
1312  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1313  * log space.  This code pushes on the lsn which would supposedly free up
1314  * the 25% which we want to leave free.  We may need to adopt a policy which
1315  * pushes on an lsn which is further along in the log once we reach the high
1316  * water mark.  In this manner, we would be creating a low water mark.
1317  */
1318 STATIC void
1319 xlog_grant_push_ail(
1320         struct log      *log,
1321         int             need_bytes)
1322 {
1323         xfs_lsn_t       threshold_lsn = 0;
1324         xfs_lsn_t       last_sync_lsn;
1325         int             free_blocks;
1326         int             free_bytes;
1327         int             threshold_block;
1328         int             threshold_cycle;
1329         int             free_threshold;
1330
1331         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1332
1333         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1334         free_blocks = BTOBBT(free_bytes);
1335
1336         /*
1337          * Set the threshold for the minimum number of free blocks in the
1338          * log to the maximum of what the caller needs, one quarter of the
1339          * log, and 256 blocks.
1340          */
1341         free_threshold = BTOBB(need_bytes);
1342         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1343         free_threshold = MAX(free_threshold, 256);
1344         if (free_blocks >= free_threshold)
1345                 return;
1346
1347         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1348                                                 &threshold_block);
1349         threshold_block += free_threshold;
1350         if (threshold_block >= log->l_logBBsize) {
1351                 threshold_block -= log->l_logBBsize;
1352                 threshold_cycle += 1;
1353         }
1354         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1355                                         threshold_block);
1356         /*
1357          * Don't pass in an lsn greater than the lsn of the last
1358          * log record known to be on disk. Use a snapshot of the last sync lsn
1359          * so that it doesn't change between the compare and the set.
1360          */
1361         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1362         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1363                 threshold_lsn = last_sync_lsn;
1364
1365         /*
1366          * Get the transaction layer to kick the dirty buffers out to
1367          * disk asynchronously. No point in trying to do this if
1368          * the filesystem is shutting down.
1369          */
1370         if (!XLOG_FORCED_SHUTDOWN(log))
1371                 xfs_ail_push(log->l_ailp, threshold_lsn);
1372 }
1373
1374 /*
1375  * The bdstrat callback function for log bufs. This gives us a central
1376  * place to trap bufs in case we get hit by a log I/O error and need to
1377  * shutdown. Actually, in practice, even when we didn't get a log error,
1378  * we transition the iclogs to IOERROR state *after* flushing all existing
1379  * iclogs to disk. This is because we don't want anymore new transactions to be
1380  * started or completed afterwards.
1381  */
1382 STATIC int
1383 xlog_bdstrat(
1384         struct xfs_buf          *bp)
1385 {
1386         struct xlog_in_core     *iclog = bp->b_fspriv;
1387
1388         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1389                 xfs_buf_ioerror(bp, EIO);
1390                 xfs_buf_stale(bp);
1391                 xfs_buf_ioend(bp, 0);
1392                 /*
1393                  * It would seem logical to return EIO here, but we rely on
1394                  * the log state machine to propagate I/O errors instead of
1395                  * doing it here.
1396                  */
1397                 return 0;
1398         }
1399
1400         xfs_buf_iorequest(bp);
1401         return 0;
1402 }
1403
1404 /*
1405  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1406  * fashion.  Previously, we should have moved the current iclog
1407  * ptr in the log to point to the next available iclog.  This allows further
1408  * write to continue while this code syncs out an iclog ready to go.
1409  * Before an in-core log can be written out, the data section must be scanned
1410  * to save away the 1st word of each BBSIZE block into the header.  We replace
1411  * it with the current cycle count.  Each BBSIZE block is tagged with the
1412  * cycle count because there in an implicit assumption that drives will
1413  * guarantee that entire 512 byte blocks get written at once.  In other words,
1414  * we can't have part of a 512 byte block written and part not written.  By
1415  * tagging each block, we will know which blocks are valid when recovering
1416  * after an unclean shutdown.
1417  *
1418  * This routine is single threaded on the iclog.  No other thread can be in
1419  * this routine with the same iclog.  Changing contents of iclog can there-
1420  * fore be done without grabbing the state machine lock.  Updating the global
1421  * log will require grabbing the lock though.
1422  *
1423  * The entire log manager uses a logical block numbering scheme.  Only
1424  * log_sync (and then only bwrite()) know about the fact that the log may
1425  * not start with block zero on a given device.  The log block start offset
1426  * is added immediately before calling bwrite().
1427  */
1428
1429 STATIC int
1430 xlog_sync(xlog_t                *log,
1431           xlog_in_core_t        *iclog)
1432 {
1433         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1434         xfs_buf_t       *bp;
1435         int             i;
1436         uint            count;          /* byte count of bwrite */
1437         uint            count_init;     /* initial count before roundup */
1438         int             roundoff;       /* roundoff to BB or stripe */
1439         int             split = 0;      /* split write into two regions */
1440         int             error;
1441         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1442
1443         XFS_STATS_INC(xs_log_writes);
1444         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1445
1446         /* Add for LR header */
1447         count_init = log->l_iclog_hsize + iclog->ic_offset;
1448
1449         /* Round out the log write size */
1450         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1451                 /* we have a v2 stripe unit to use */
1452                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1453         } else {
1454                 count = BBTOB(BTOBB(count_init));
1455         }
1456         roundoff = count - count_init;
1457         ASSERT(roundoff >= 0);
1458         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1459                 roundoff < log->l_mp->m_sb.sb_logsunit)
1460                 ||
1461                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1462                  roundoff < BBTOB(1)));
1463
1464         /* move grant heads by roundoff in sync */
1465         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1466         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1467
1468         /* put cycle number in every block */
1469         xlog_pack_data(log, iclog, roundoff);
1470
1471         /* real byte length */
1472         if (v2) {
1473                 iclog->ic_header.h_len =
1474                         cpu_to_be32(iclog->ic_offset + roundoff);
1475         } else {
1476                 iclog->ic_header.h_len =
1477                         cpu_to_be32(iclog->ic_offset);
1478         }
1479
1480         bp = iclog->ic_bp;
1481         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1482
1483         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1484
1485         /* Do we need to split this write into 2 parts? */
1486         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1487                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1488                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1489                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1490         } else {
1491                 iclog->ic_bwritecnt = 1;
1492         }
1493         XFS_BUF_SET_COUNT(bp, count);
1494         bp->b_fspriv = iclog;
1495         XFS_BUF_ZEROFLAGS(bp);
1496         XFS_BUF_ASYNC(bp);
1497         bp->b_flags |= XBF_SYNCIO;
1498
1499         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1500                 bp->b_flags |= XBF_FUA;
1501
1502                 /*
1503                  * Flush the data device before flushing the log to make
1504                  * sure all meta data written back from the AIL actually made
1505                  * it to disk before stamping the new log tail LSN into the
1506                  * log buffer.  For an external log we need to issue the
1507                  * flush explicitly, and unfortunately synchronously here;
1508                  * for an internal log we can simply use the block layer
1509                  * state machine for preflushes.
1510                  */
1511                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1512                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1513                 else
1514                         bp->b_flags |= XBF_FLUSH;
1515         }
1516
1517         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1518         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1519
1520         xlog_verify_iclog(log, iclog, count, B_TRUE);
1521
1522         /* account for log which doesn't start at block #0 */
1523         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1524         /*
1525          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1526          * is shutting down.
1527          */
1528         XFS_BUF_WRITE(bp);
1529
1530         error = xlog_bdstrat(bp);
1531         if (error) {
1532                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1533                 return error;
1534         }
1535         if (split) {
1536                 bp = iclog->ic_log->l_xbuf;
1537                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1538                 xfs_buf_associate_memory(bp,
1539                                 (char *)&iclog->ic_header + count, split);
1540                 bp->b_fspriv = iclog;
1541                 XFS_BUF_ZEROFLAGS(bp);
1542                 XFS_BUF_ASYNC(bp);
1543                 bp->b_flags |= XBF_SYNCIO;
1544                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1545                         bp->b_flags |= XBF_FUA;
1546                 dptr = bp->b_addr;
1547                 /*
1548                  * Bump the cycle numbers at the start of each block
1549                  * since this part of the buffer is at the start of
1550                  * a new cycle.  Watch out for the header magic number
1551                  * case, though.
1552                  */
1553                 for (i = 0; i < split; i += BBSIZE) {
1554                         be32_add_cpu((__be32 *)dptr, 1);
1555                         if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1556                                 be32_add_cpu((__be32 *)dptr, 1);
1557                         dptr += BBSIZE;
1558                 }
1559
1560                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1561                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1562
1563                 /* account for internal log which doesn't start at block #0 */
1564                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1565                 XFS_BUF_WRITE(bp);
1566                 error = xlog_bdstrat(bp);
1567                 if (error) {
1568                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1569                         return error;
1570                 }
1571         }
1572         return 0;
1573 }       /* xlog_sync */
1574
1575
1576 /*
1577  * Deallocate a log structure
1578  */
1579 STATIC void
1580 xlog_dealloc_log(xlog_t *log)
1581 {
1582         xlog_in_core_t  *iclog, *next_iclog;
1583         int             i;
1584
1585         xlog_cil_destroy(log);
1586
1587         /*
1588          * always need to ensure that the extra buffer does not point to memory
1589          * owned by another log buffer before we free it.
1590          */
1591         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1592         xfs_buf_free(log->l_xbuf);
1593
1594         iclog = log->l_iclog;
1595         for (i=0; i<log->l_iclog_bufs; i++) {
1596                 xfs_buf_free(iclog->ic_bp);
1597                 next_iclog = iclog->ic_next;
1598                 kmem_free(iclog);
1599                 iclog = next_iclog;
1600         }
1601         spinlock_destroy(&log->l_icloglock);
1602
1603         log->l_mp->m_log = NULL;
1604         kmem_free(log);
1605 }       /* xlog_dealloc_log */
1606
1607 /*
1608  * Update counters atomically now that memcpy is done.
1609  */
1610 /* ARGSUSED */
1611 static inline void
1612 xlog_state_finish_copy(xlog_t           *log,
1613                        xlog_in_core_t   *iclog,
1614                        int              record_cnt,
1615                        int              copy_bytes)
1616 {
1617         spin_lock(&log->l_icloglock);
1618
1619         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1620         iclog->ic_offset += copy_bytes;
1621
1622         spin_unlock(&log->l_icloglock);
1623 }       /* xlog_state_finish_copy */
1624
1625
1626
1627
1628 /*
1629  * print out info relating to regions written which consume
1630  * the reservation
1631  */
1632 void
1633 xlog_print_tic_res(
1634         struct xfs_mount        *mp,
1635         struct xlog_ticket      *ticket)
1636 {
1637         uint i;
1638         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1639
1640         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1641         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1642             "bformat",
1643             "bchunk",
1644             "efi_format",
1645             "efd_format",
1646             "iformat",
1647             "icore",
1648             "iext",
1649             "ibroot",
1650             "ilocal",
1651             "iattr_ext",
1652             "iattr_broot",
1653             "iattr_local",
1654             "qformat",
1655             "dquot",
1656             "quotaoff",
1657             "LR header",
1658             "unmount",
1659             "commit",
1660             "trans header"
1661         };
1662         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1663             "SETATTR_NOT_SIZE",
1664             "SETATTR_SIZE",
1665             "INACTIVE",
1666             "CREATE",
1667             "CREATE_TRUNC",
1668             "TRUNCATE_FILE",
1669             "REMOVE",
1670             "LINK",
1671             "RENAME",
1672             "MKDIR",
1673             "RMDIR",
1674             "SYMLINK",
1675             "SET_DMATTRS",
1676             "GROWFS",
1677             "STRAT_WRITE",
1678             "DIOSTRAT",
1679             "WRITE_SYNC",
1680             "WRITEID",
1681             "ADDAFORK",
1682             "ATTRINVAL",
1683             "ATRUNCATE",
1684             "ATTR_SET",
1685             "ATTR_RM",
1686             "ATTR_FLAG",
1687             "CLEAR_AGI_BUCKET",
1688             "QM_SBCHANGE",
1689             "DUMMY1",
1690             "DUMMY2",
1691             "QM_QUOTAOFF",
1692             "QM_DQALLOC",
1693             "QM_SETQLIM",
1694             "QM_DQCLUSTER",
1695             "QM_QINOCREATE",
1696             "QM_QUOTAOFF_END",
1697             "SB_UNIT",
1698             "FSYNC_TS",
1699             "GROWFSRT_ALLOC",
1700             "GROWFSRT_ZERO",
1701             "GROWFSRT_FREE",
1702             "SWAPEXT"
1703         };
1704
1705         xfs_warn(mp,
1706                 "xlog_write: reservation summary:\n"
1707                 "  trans type  = %s (%u)\n"
1708                 "  unit res    = %d bytes\n"
1709                 "  current res = %d bytes\n"
1710                 "  total reg   = %u bytes (o/flow = %u bytes)\n"
1711                 "  ophdrs      = %u (ophdr space = %u bytes)\n"
1712                 "  ophdr + reg = %u bytes\n"
1713                 "  num regions = %u\n",
1714                 ((ticket->t_trans_type <= 0 ||
1715                   ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1716                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1717                 ticket->t_trans_type,
1718                 ticket->t_unit_res,
1719                 ticket->t_curr_res,
1720                 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1721                 ticket->t_res_num_ophdrs, ophdr_spc,
1722                 ticket->t_res_arr_sum +
1723                 ticket->t_res_o_flow + ophdr_spc,
1724                 ticket->t_res_num);
1725
1726         for (i = 0; i < ticket->t_res_num; i++) {
1727                 uint r_type = ticket->t_res_arr[i].r_type;
1728                 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1729                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1730                             "bad-rtype" : res_type_str[r_type-1]),
1731                             ticket->t_res_arr[i].r_len);
1732         }
1733
1734         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1735                 "xlog_write: reservation ran out. Need to up reservation");
1736         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1737 }
1738
1739 /*
1740  * Calculate the potential space needed by the log vector.  Each region gets
1741  * its own xlog_op_header_t and may need to be double word aligned.
1742  */
1743 static int
1744 xlog_write_calc_vec_length(
1745         struct xlog_ticket      *ticket,
1746         struct xfs_log_vec      *log_vector)
1747 {
1748         struct xfs_log_vec      *lv;
1749         int                     headers = 0;
1750         int                     len = 0;
1751         int                     i;
1752
1753         /* acct for start rec of xact */
1754         if (ticket->t_flags & XLOG_TIC_INITED)
1755                 headers++;
1756
1757         for (lv = log_vector; lv; lv = lv->lv_next) {
1758                 headers += lv->lv_niovecs;
1759
1760                 for (i = 0; i < lv->lv_niovecs; i++) {
1761                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
1762
1763                         len += vecp->i_len;
1764                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1765                 }
1766         }
1767
1768         ticket->t_res_num_ophdrs += headers;
1769         len += headers * sizeof(struct xlog_op_header);
1770
1771         return len;
1772 }
1773
1774 /*
1775  * If first write for transaction, insert start record  We can't be trying to
1776  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
1777  */
1778 static int
1779 xlog_write_start_rec(
1780         struct xlog_op_header   *ophdr,
1781         struct xlog_ticket      *ticket)
1782 {
1783         if (!(ticket->t_flags & XLOG_TIC_INITED))
1784                 return 0;
1785
1786         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
1787         ophdr->oh_clientid = ticket->t_clientid;
1788         ophdr->oh_len = 0;
1789         ophdr->oh_flags = XLOG_START_TRANS;
1790         ophdr->oh_res2 = 0;
1791
1792         ticket->t_flags &= ~XLOG_TIC_INITED;
1793
1794         return sizeof(struct xlog_op_header);
1795 }
1796
1797 static xlog_op_header_t *
1798 xlog_write_setup_ophdr(
1799         struct log              *log,
1800         struct xlog_op_header   *ophdr,
1801         struct xlog_ticket      *ticket,
1802         uint                    flags)
1803 {
1804         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1805         ophdr->oh_clientid = ticket->t_clientid;
1806         ophdr->oh_res2 = 0;
1807
1808         /* are we copying a commit or unmount record? */
1809         ophdr->oh_flags = flags;
1810
1811         /*
1812          * We've seen logs corrupted with bad transaction client ids.  This
1813          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
1814          * and shut down the filesystem.
1815          */
1816         switch (ophdr->oh_clientid)  {
1817         case XFS_TRANSACTION:
1818         case XFS_VOLUME:
1819         case XFS_LOG:
1820                 break;
1821         default:
1822                 xfs_warn(log->l_mp,
1823                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1824                         ophdr->oh_clientid, ticket);
1825                 return NULL;
1826         }
1827
1828         return ophdr;
1829 }
1830
1831 /*
1832  * Set up the parameters of the region copy into the log. This has
1833  * to handle region write split across multiple log buffers - this
1834  * state is kept external to this function so that this code can
1835  * can be written in an obvious, self documenting manner.
1836  */
1837 static int
1838 xlog_write_setup_copy(
1839         struct xlog_ticket      *ticket,
1840         struct xlog_op_header   *ophdr,
1841         int                     space_available,
1842         int                     space_required,
1843         int                     *copy_off,
1844         int                     *copy_len,
1845         int                     *last_was_partial_copy,
1846         int                     *bytes_consumed)
1847 {
1848         int                     still_to_copy;
1849
1850         still_to_copy = space_required - *bytes_consumed;
1851         *copy_off = *bytes_consumed;
1852
1853         if (still_to_copy <= space_available) {
1854                 /* write of region completes here */
1855                 *copy_len = still_to_copy;
1856                 ophdr->oh_len = cpu_to_be32(*copy_len);
1857                 if (*last_was_partial_copy)
1858                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1859                 *last_was_partial_copy = 0;
1860                 *bytes_consumed = 0;
1861                 return 0;
1862         }
1863
1864         /* partial write of region, needs extra log op header reservation */
1865         *copy_len = space_available;
1866         ophdr->oh_len = cpu_to_be32(*copy_len);
1867         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1868         if (*last_was_partial_copy)
1869                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1870         *bytes_consumed += *copy_len;
1871         (*last_was_partial_copy)++;
1872
1873         /* account for new log op header */
1874         ticket->t_curr_res -= sizeof(struct xlog_op_header);
1875         ticket->t_res_num_ophdrs++;
1876
1877         return sizeof(struct xlog_op_header);
1878 }
1879
1880 static int
1881 xlog_write_copy_finish(
1882         struct log              *log,
1883         struct xlog_in_core     *iclog,
1884         uint                    flags,
1885         int                     *record_cnt,
1886         int                     *data_cnt,
1887         int                     *partial_copy,
1888         int                     *partial_copy_len,
1889         int                     log_offset,
1890         struct xlog_in_core     **commit_iclog)
1891 {
1892         if (*partial_copy) {
1893                 /*
1894                  * This iclog has already been marked WANT_SYNC by
1895                  * xlog_state_get_iclog_space.
1896                  */
1897                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1898                 *record_cnt = 0;
1899                 *data_cnt = 0;
1900                 return xlog_state_release_iclog(log, iclog);
1901         }
1902
1903         *partial_copy = 0;
1904         *partial_copy_len = 0;
1905
1906         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1907                 /* no more space in this iclog - push it. */
1908                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1909                 *record_cnt = 0;
1910                 *data_cnt = 0;
1911
1912                 spin_lock(&log->l_icloglock);
1913                 xlog_state_want_sync(log, iclog);
1914                 spin_unlock(&log->l_icloglock);
1915
1916                 if (!commit_iclog)
1917                         return xlog_state_release_iclog(log, iclog);
1918                 ASSERT(flags & XLOG_COMMIT_TRANS);
1919                 *commit_iclog = iclog;
1920         }
1921
1922         return 0;
1923 }
1924
1925 /*
1926  * Write some region out to in-core log
1927  *
1928  * This will be called when writing externally provided regions or when
1929  * writing out a commit record for a given transaction.
1930  *
1931  * General algorithm:
1932  *      1. Find total length of this write.  This may include adding to the
1933  *              lengths passed in.
1934  *      2. Check whether we violate the tickets reservation.
1935  *      3. While writing to this iclog
1936  *          A. Reserve as much space in this iclog as can get
1937  *          B. If this is first write, save away start lsn
1938  *          C. While writing this region:
1939  *              1. If first write of transaction, write start record
1940  *              2. Write log operation header (header per region)
1941  *              3. Find out if we can fit entire region into this iclog
1942  *              4. Potentially, verify destination memcpy ptr
1943  *              5. Memcpy (partial) region
1944  *              6. If partial copy, release iclog; otherwise, continue
1945  *                      copying more regions into current iclog
1946  *      4. Mark want sync bit (in simulation mode)
1947  *      5. Release iclog for potential flush to on-disk log.
1948  *
1949  * ERRORS:
1950  * 1.   Panic if reservation is overrun.  This should never happen since
1951  *      reservation amounts are generated internal to the filesystem.
1952  * NOTES:
1953  * 1. Tickets are single threaded data structures.
1954  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1955  *      syncing routine.  When a single log_write region needs to span
1956  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1957  *      on all log operation writes which don't contain the end of the
1958  *      region.  The XLOG_END_TRANS bit is used for the in-core log
1959  *      operation which contains the end of the continued log_write region.
1960  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1961  *      we don't really know exactly how much space will be used.  As a result,
1962  *      we don't update ic_offset until the end when we know exactly how many
1963  *      bytes have been written out.
1964  */
1965 int
1966 xlog_write(
1967         struct log              *log,
1968         struct xfs_log_vec      *log_vector,
1969         struct xlog_ticket      *ticket,
1970         xfs_lsn_t               *start_lsn,
1971         struct xlog_in_core     **commit_iclog,
1972         uint                    flags)
1973 {
1974         struct xlog_in_core     *iclog = NULL;
1975         struct xfs_log_iovec    *vecp;
1976         struct xfs_log_vec      *lv;
1977         int                     len;
1978         int                     index;
1979         int                     partial_copy = 0;
1980         int                     partial_copy_len = 0;
1981         int                     contwr = 0;
1982         int                     record_cnt = 0;
1983         int                     data_cnt = 0;
1984         int                     error;
1985
1986         *start_lsn = 0;
1987
1988         len = xlog_write_calc_vec_length(ticket, log_vector);
1989
1990         /*
1991          * Region headers and bytes are already accounted for.
1992          * We only need to take into account start records and
1993          * split regions in this function.
1994          */
1995         if (ticket->t_flags & XLOG_TIC_INITED)
1996                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1997
1998         /*
1999          * Commit record headers need to be accounted for. These
2000          * come in as separate writes so are easy to detect.
2001          */
2002         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2003                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2004
2005         if (ticket->t_curr_res < 0)
2006                 xlog_print_tic_res(log->l_mp, ticket);
2007
2008         index = 0;
2009         lv = log_vector;
2010         vecp = lv->lv_iovecp;
2011         while (lv && index < lv->lv_niovecs) {
2012                 void            *ptr;
2013                 int             log_offset;
2014
2015                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2016                                                    &contwr, &log_offset);
2017                 if (error)
2018                         return error;
2019
2020                 ASSERT(log_offset <= iclog->ic_size - 1);
2021                 ptr = iclog->ic_datap + log_offset;
2022
2023                 /* start_lsn is the first lsn written to. That's all we need. */
2024                 if (!*start_lsn)
2025                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2026
2027                 /*
2028                  * This loop writes out as many regions as can fit in the amount
2029                  * of space which was allocated by xlog_state_get_iclog_space().
2030                  */
2031                 while (lv && index < lv->lv_niovecs) {
2032                         struct xfs_log_iovec    *reg = &vecp[index];
2033                         struct xlog_op_header   *ophdr;
2034                         int                     start_rec_copy;
2035                         int                     copy_len;
2036                         int                     copy_off;
2037
2038                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2039                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2040
2041                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2042                         if (start_rec_copy) {
2043                                 record_cnt++;
2044                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2045                                                    start_rec_copy);
2046                         }
2047
2048                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2049                         if (!ophdr)
2050                                 return XFS_ERROR(EIO);
2051
2052                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2053                                            sizeof(struct xlog_op_header));
2054
2055                         len += xlog_write_setup_copy(ticket, ophdr,
2056                                                      iclog->ic_size-log_offset,
2057                                                      reg->i_len,
2058                                                      &copy_off, &copy_len,
2059                                                      &partial_copy,
2060                                                      &partial_copy_len);
2061                         xlog_verify_dest_ptr(log, ptr);
2062
2063                         /* copy region */
2064                         ASSERT(copy_len >= 0);
2065                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
2066                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2067
2068                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2069                         record_cnt++;
2070                         data_cnt += contwr ? copy_len : 0;
2071
2072                         error = xlog_write_copy_finish(log, iclog, flags,
2073                                                        &record_cnt, &data_cnt,
2074                                                        &partial_copy,
2075                                                        &partial_copy_len,
2076                                                        log_offset,
2077                                                        commit_iclog);
2078                         if (error)
2079                                 return error;
2080
2081                         /*
2082                          * if we had a partial copy, we need to get more iclog
2083                          * space but we don't want to increment the region
2084                          * index because there is still more is this region to
2085                          * write.
2086                          *
2087                          * If we completed writing this region, and we flushed
2088                          * the iclog (indicated by resetting of the record
2089                          * count), then we also need to get more log space. If
2090                          * this was the last record, though, we are done and
2091                          * can just return.
2092                          */
2093                         if (partial_copy)
2094                                 break;
2095
2096                         if (++index == lv->lv_niovecs) {
2097                                 lv = lv->lv_next;
2098                                 index = 0;
2099                                 if (lv)
2100                                         vecp = lv->lv_iovecp;
2101                         }
2102                         if (record_cnt == 0) {
2103                                 if (!lv)
2104                                         return 0;
2105                                 break;
2106                         }
2107                 }
2108         }
2109
2110         ASSERT(len == 0);
2111
2112         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2113         if (!commit_iclog)
2114                 return xlog_state_release_iclog(log, iclog);
2115
2116         ASSERT(flags & XLOG_COMMIT_TRANS);
2117         *commit_iclog = iclog;
2118         return 0;
2119 }
2120
2121
2122 /*****************************************************************************
2123  *
2124  *              State Machine functions
2125  *
2126  *****************************************************************************
2127  */
2128
2129 /* Clean iclogs starting from the head.  This ordering must be
2130  * maintained, so an iclog doesn't become ACTIVE beyond one that
2131  * is SYNCING.  This is also required to maintain the notion that we use
2132  * a ordered wait queue to hold off would be writers to the log when every
2133  * iclog is trying to sync to disk.
2134  *
2135  * State Change: DIRTY -> ACTIVE
2136  */
2137 STATIC void
2138 xlog_state_clean_log(xlog_t *log)
2139 {
2140         xlog_in_core_t  *iclog;
2141         int changed = 0;
2142
2143         iclog = log->l_iclog;
2144         do {
2145                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2146                         iclog->ic_state = XLOG_STATE_ACTIVE;
2147                         iclog->ic_offset       = 0;
2148                         ASSERT(iclog->ic_callback == NULL);
2149                         /*
2150                          * If the number of ops in this iclog indicate it just
2151                          * contains the dummy transaction, we can
2152                          * change state into IDLE (the second time around).
2153                          * Otherwise we should change the state into
2154                          * NEED a dummy.
2155                          * We don't need to cover the dummy.
2156                          */
2157                         if (!changed &&
2158                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2159                                         XLOG_COVER_OPS)) {
2160                                 changed = 1;
2161                         } else {
2162                                 /*
2163                                  * We have two dirty iclogs so start over
2164                                  * This could also be num of ops indicates
2165                                  * this is not the dummy going out.
2166                                  */
2167                                 changed = 2;
2168                         }
2169                         iclog->ic_header.h_num_logops = 0;
2170                         memset(iclog->ic_header.h_cycle_data, 0,
2171                               sizeof(iclog->ic_header.h_cycle_data));
2172                         iclog->ic_header.h_lsn = 0;
2173                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2174                         /* do nothing */;
2175                 else
2176                         break;  /* stop cleaning */
2177                 iclog = iclog->ic_next;
2178         } while (iclog != log->l_iclog);
2179
2180         /* log is locked when we are called */
2181         /*
2182          * Change state for the dummy log recording.
2183          * We usually go to NEED. But we go to NEED2 if the changed indicates
2184          * we are done writing the dummy record.
2185          * If we are done with the second dummy recored (DONE2), then
2186          * we go to IDLE.
2187          */
2188         if (changed) {
2189                 switch (log->l_covered_state) {
2190                 case XLOG_STATE_COVER_IDLE:
2191                 case XLOG_STATE_COVER_NEED:
2192                 case XLOG_STATE_COVER_NEED2:
2193                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2194                         break;
2195
2196                 case XLOG_STATE_COVER_DONE:
2197                         if (changed == 1)
2198                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2199                         else
2200                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2201                         break;
2202
2203                 case XLOG_STATE_COVER_DONE2:
2204                         if (changed == 1)
2205                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2206                         else
2207                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2208                         break;
2209
2210                 default:
2211                         ASSERT(0);
2212                 }
2213         }
2214 }       /* xlog_state_clean_log */
2215
2216 STATIC xfs_lsn_t
2217 xlog_get_lowest_lsn(
2218         xlog_t          *log)
2219 {
2220         xlog_in_core_t  *lsn_log;
2221         xfs_lsn_t       lowest_lsn, lsn;
2222
2223         lsn_log = log->l_iclog;
2224         lowest_lsn = 0;
2225         do {
2226             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2227                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2228                 if ((lsn && !lowest_lsn) ||
2229                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2230                         lowest_lsn = lsn;
2231                 }
2232             }
2233             lsn_log = lsn_log->ic_next;
2234         } while (lsn_log != log->l_iclog);
2235         return lowest_lsn;
2236 }
2237
2238
2239 STATIC void
2240 xlog_state_do_callback(
2241         xlog_t          *log,
2242         int             aborted,
2243         xlog_in_core_t  *ciclog)
2244 {
2245         xlog_in_core_t     *iclog;
2246         xlog_in_core_t     *first_iclog;        /* used to know when we've
2247                                                  * processed all iclogs once */
2248         xfs_log_callback_t *cb, *cb_next;
2249         int                flushcnt = 0;
2250         xfs_lsn_t          lowest_lsn;
2251         int                ioerrors;    /* counter: iclogs with errors */
2252         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2253         int                funcdidcallbacks; /* flag: function did callbacks */
2254         int                repeats;     /* for issuing console warnings if
2255                                          * looping too many times */
2256         int                wake = 0;
2257
2258         spin_lock(&log->l_icloglock);
2259         first_iclog = iclog = log->l_iclog;
2260         ioerrors = 0;
2261         funcdidcallbacks = 0;
2262         repeats = 0;
2263
2264         do {
2265                 /*
2266                  * Scan all iclogs starting with the one pointed to by the
2267                  * log.  Reset this starting point each time the log is
2268                  * unlocked (during callbacks).
2269                  *
2270                  * Keep looping through iclogs until one full pass is made
2271                  * without running any callbacks.
2272                  */
2273                 first_iclog = log->l_iclog;
2274                 iclog = log->l_iclog;
2275                 loopdidcallbacks = 0;
2276                 repeats++;
2277
2278                 do {
2279
2280                         /* skip all iclogs in the ACTIVE & DIRTY states */
2281                         if (iclog->ic_state &
2282                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2283                                 iclog = iclog->ic_next;
2284                                 continue;
2285                         }
2286
2287                         /*
2288                          * Between marking a filesystem SHUTDOWN and stopping
2289                          * the log, we do flush all iclogs to disk (if there
2290                          * wasn't a log I/O error). So, we do want things to
2291                          * go smoothly in case of just a SHUTDOWN  w/o a
2292                          * LOG_IO_ERROR.
2293                          */
2294                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2295                                 /*
2296                                  * Can only perform callbacks in order.  Since
2297                                  * this iclog is not in the DONE_SYNC/
2298                                  * DO_CALLBACK state, we skip the rest and
2299                                  * just try to clean up.  If we set our iclog
2300                                  * to DO_CALLBACK, we will not process it when
2301                                  * we retry since a previous iclog is in the
2302                                  * CALLBACK and the state cannot change since
2303                                  * we are holding the l_icloglock.
2304                                  */
2305                                 if (!(iclog->ic_state &
2306                                         (XLOG_STATE_DONE_SYNC |
2307                                                  XLOG_STATE_DO_CALLBACK))) {
2308                                         if (ciclog && (ciclog->ic_state ==
2309                                                         XLOG_STATE_DONE_SYNC)) {
2310                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2311                                         }
2312                                         break;
2313                                 }
2314                                 /*
2315                                  * We now have an iclog that is in either the
2316                                  * DO_CALLBACK or DONE_SYNC states. The other
2317                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2318                                  * caught by the above if and are going to
2319                                  * clean (i.e. we aren't doing their callbacks)
2320                                  * see the above if.
2321                                  */
2322
2323                                 /*
2324                                  * We will do one more check here to see if we
2325                                  * have chased our tail around.
2326                                  */
2327
2328                                 lowest_lsn = xlog_get_lowest_lsn(log);
2329                                 if (lowest_lsn &&
2330                                     XFS_LSN_CMP(lowest_lsn,
2331                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2332                                         iclog = iclog->ic_next;
2333                                         continue; /* Leave this iclog for
2334                                                    * another thread */
2335                                 }
2336
2337                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2338
2339
2340                                 /*
2341                                  * update the last_sync_lsn before we drop the
2342                                  * icloglock to ensure we are the only one that
2343                                  * can update it.
2344                                  */
2345                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2346                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2347                                 atomic64_set(&log->l_last_sync_lsn,
2348                                         be64_to_cpu(iclog->ic_header.h_lsn));
2349
2350                         } else
2351                                 ioerrors++;
2352
2353                         spin_unlock(&log->l_icloglock);
2354
2355                         /*
2356                          * Keep processing entries in the callback list until
2357                          * we come around and it is empty.  We need to
2358                          * atomically see that the list is empty and change the
2359                          * state to DIRTY so that we don't miss any more
2360                          * callbacks being added.
2361                          */
2362                         spin_lock(&iclog->ic_callback_lock);
2363                         cb = iclog->ic_callback;
2364                         while (cb) {
2365                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2366                                 iclog->ic_callback = NULL;
2367                                 spin_unlock(&iclog->ic_callback_lock);
2368
2369                                 /* perform callbacks in the order given */
2370                                 for (; cb; cb = cb_next) {
2371                                         cb_next = cb->cb_next;
2372                                         cb->cb_func(cb->cb_arg, aborted);
2373                                 }
2374                                 spin_lock(&iclog->ic_callback_lock);
2375                                 cb = iclog->ic_callback;
2376                         }
2377
2378                         loopdidcallbacks++;
2379                         funcdidcallbacks++;
2380
2381                         spin_lock(&log->l_icloglock);
2382                         ASSERT(iclog->ic_callback == NULL);
2383                         spin_unlock(&iclog->ic_callback_lock);
2384                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2385                                 iclog->ic_state = XLOG_STATE_DIRTY;
2386
2387                         /*
2388                          * Transition from DIRTY to ACTIVE if applicable.
2389                          * NOP if STATE_IOERROR.
2390                          */
2391                         xlog_state_clean_log(log);
2392
2393                         /* wake up threads waiting in xfs_log_force() */
2394                         wake_up_all(&iclog->ic_force_wait);
2395
2396                         iclog = iclog->ic_next;
2397                 } while (first_iclog != iclog);
2398
2399                 if (repeats > 5000) {
2400                         flushcnt += repeats;
2401                         repeats = 0;
2402                         xfs_warn(log->l_mp,
2403                                 "%s: possible infinite loop (%d iterations)",
2404                                 __func__, flushcnt);
2405                 }
2406         } while (!ioerrors && loopdidcallbacks);
2407
2408         /*
2409          * make one last gasp attempt to see if iclogs are being left in
2410          * limbo..
2411          */
2412 #ifdef DEBUG
2413         if (funcdidcallbacks) {
2414                 first_iclog = iclog = log->l_iclog;
2415                 do {
2416                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2417                         /*
2418                          * Terminate the loop if iclogs are found in states
2419                          * which will cause other threads to clean up iclogs.
2420                          *
2421                          * SYNCING - i/o completion will go through logs
2422                          * DONE_SYNC - interrupt thread should be waiting for
2423                          *              l_icloglock
2424                          * IOERROR - give up hope all ye who enter here
2425                          */
2426                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2427                             iclog->ic_state == XLOG_STATE_SYNCING ||
2428                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2429                             iclog->ic_state == XLOG_STATE_IOERROR )
2430                                 break;
2431                         iclog = iclog->ic_next;
2432                 } while (first_iclog != iclog);
2433         }
2434 #endif
2435
2436         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2437                 wake = 1;
2438         spin_unlock(&log->l_icloglock);
2439
2440         if (wake)
2441                 wake_up_all(&log->l_flush_wait);
2442 }
2443
2444
2445 /*
2446  * Finish transitioning this iclog to the dirty state.
2447  *
2448  * Make sure that we completely execute this routine only when this is
2449  * the last call to the iclog.  There is a good chance that iclog flushes,
2450  * when we reach the end of the physical log, get turned into 2 separate
2451  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2452  * routine.  By using the reference count bwritecnt, we guarantee that only
2453  * the second completion goes through.
2454  *
2455  * Callbacks could take time, so they are done outside the scope of the
2456  * global state machine log lock.
2457  */
2458 STATIC void
2459 xlog_state_done_syncing(
2460         xlog_in_core_t  *iclog,
2461         int             aborted)
2462 {
2463         xlog_t             *log = iclog->ic_log;
2464
2465         spin_lock(&log->l_icloglock);
2466
2467         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2468                iclog->ic_state == XLOG_STATE_IOERROR);
2469         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2470         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2471
2472
2473         /*
2474          * If we got an error, either on the first buffer, or in the case of
2475          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2476          * and none should ever be attempted to be written to disk
2477          * again.
2478          */
2479         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2480                 if (--iclog->ic_bwritecnt == 1) {
2481                         spin_unlock(&log->l_icloglock);
2482                         return;
2483                 }
2484                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2485         }
2486
2487         /*
2488          * Someone could be sleeping prior to writing out the next
2489          * iclog buffer, we wake them all, one will get to do the
2490          * I/O, the others get to wait for the result.
2491          */
2492         wake_up_all(&iclog->ic_write_wait);
2493         spin_unlock(&log->l_icloglock);
2494         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2495 }       /* xlog_state_done_syncing */
2496
2497
2498 /*
2499  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2500  * sleep.  We wait on the flush queue on the head iclog as that should be
2501  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2502  * we will wait here and all new writes will sleep until a sync completes.
2503  *
2504  * The in-core logs are used in a circular fashion. They are not used
2505  * out-of-order even when an iclog past the head is free.
2506  *
2507  * return:
2508  *      * log_offset where xlog_write() can start writing into the in-core
2509  *              log's data space.
2510  *      * in-core log pointer to which xlog_write() should write.
2511  *      * boolean indicating this is a continued write to an in-core log.
2512  *              If this is the last write, then the in-core log's offset field
2513  *              needs to be incremented, depending on the amount of data which
2514  *              is copied.
2515  */
2516 STATIC int
2517 xlog_state_get_iclog_space(xlog_t         *log,
2518                            int            len,
2519                            xlog_in_core_t **iclogp,
2520                            xlog_ticket_t  *ticket,
2521                            int            *continued_write,
2522                            int            *logoffsetp)
2523 {
2524         int               log_offset;
2525         xlog_rec_header_t *head;
2526         xlog_in_core_t    *iclog;
2527         int               error;
2528
2529 restart:
2530         spin_lock(&log->l_icloglock);
2531         if (XLOG_FORCED_SHUTDOWN(log)) {
2532                 spin_unlock(&log->l_icloglock);
2533                 return XFS_ERROR(EIO);
2534         }
2535
2536         iclog = log->l_iclog;
2537         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2538                 XFS_STATS_INC(xs_log_noiclogs);
2539
2540                 /* Wait for log writes to have flushed */
2541                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2542                 goto restart;
2543         }
2544
2545         head = &iclog->ic_header;
2546
2547         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2548         log_offset = iclog->ic_offset;
2549
2550         /* On the 1st write to an iclog, figure out lsn.  This works
2551          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2552          * committing to.  If the offset is set, that's how many blocks
2553          * must be written.
2554          */
2555         if (log_offset == 0) {
2556                 ticket->t_curr_res -= log->l_iclog_hsize;
2557                 xlog_tic_add_region(ticket,
2558                                     log->l_iclog_hsize,
2559                                     XLOG_REG_TYPE_LRHEADER);
2560                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2561                 head->h_lsn = cpu_to_be64(
2562                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2563                 ASSERT(log->l_curr_block >= 0);
2564         }
2565
2566         /* If there is enough room to write everything, then do it.  Otherwise,
2567          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2568          * bit is on, so this will get flushed out.  Don't update ic_offset
2569          * until you know exactly how many bytes get copied.  Therefore, wait
2570          * until later to update ic_offset.
2571          *
2572          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2573          * can fit into remaining data section.
2574          */
2575         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2576                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2577
2578                 /*
2579                  * If I'm the only one writing to this iclog, sync it to disk.
2580                  * We need to do an atomic compare and decrement here to avoid
2581                  * racing with concurrent atomic_dec_and_lock() calls in
2582                  * xlog_state_release_iclog() when there is more than one
2583                  * reference to the iclog.
2584                  */
2585                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2586                         /* we are the only one */
2587                         spin_unlock(&log->l_icloglock);
2588                         error = xlog_state_release_iclog(log, iclog);
2589                         if (error)
2590                                 return error;
2591                 } else {
2592                         spin_unlock(&log->l_icloglock);
2593                 }
2594                 goto restart;
2595         }
2596
2597         /* Do we have enough room to write the full amount in the remainder
2598          * of this iclog?  Or must we continue a write on the next iclog and
2599          * mark this iclog as completely taken?  In the case where we switch
2600          * iclogs (to mark it taken), this particular iclog will release/sync
2601          * to disk in xlog_write().
2602          */
2603         if (len <= iclog->ic_size - iclog->ic_offset) {
2604                 *continued_write = 0;
2605                 iclog->ic_offset += len;
2606         } else {
2607                 *continued_write = 1;
2608                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2609         }
2610         *iclogp = iclog;
2611
2612         ASSERT(iclog->ic_offset <= iclog->ic_size);
2613         spin_unlock(&log->l_icloglock);
2614
2615         *logoffsetp = log_offset;
2616         return 0;
2617 }       /* xlog_state_get_iclog_space */
2618
2619 /* The first cnt-1 times through here we don't need to
2620  * move the grant write head because the permanent
2621  * reservation has reserved cnt times the unit amount.
2622  * Release part of current permanent unit reservation and
2623  * reset current reservation to be one units worth.  Also
2624  * move grant reservation head forward.
2625  */
2626 STATIC void
2627 xlog_regrant_reserve_log_space(xlog_t        *log,
2628                                xlog_ticket_t *ticket)
2629 {
2630         trace_xfs_log_regrant_reserve_enter(log, ticket);
2631
2632         if (ticket->t_cnt > 0)
2633                 ticket->t_cnt--;
2634
2635         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2636                                         ticket->t_curr_res);
2637         xlog_grant_sub_space(log, &log->l_write_head.grant,
2638                                         ticket->t_curr_res);
2639         ticket->t_curr_res = ticket->t_unit_res;
2640         xlog_tic_reset_res(ticket);
2641
2642         trace_xfs_log_regrant_reserve_sub(log, ticket);
2643
2644         /* just return if we still have some of the pre-reserved space */
2645         if (ticket->t_cnt > 0)
2646                 return;
2647
2648         xlog_grant_add_space(log, &log->l_reserve_head.grant,
2649                                         ticket->t_unit_res);
2650
2651         trace_xfs_log_regrant_reserve_exit(log, ticket);
2652
2653         ticket->t_curr_res = ticket->t_unit_res;
2654         xlog_tic_reset_res(ticket);
2655 }       /* xlog_regrant_reserve_log_space */
2656
2657
2658 /*
2659  * Give back the space left from a reservation.
2660  *
2661  * All the information we need to make a correct determination of space left
2662  * is present.  For non-permanent reservations, things are quite easy.  The
2663  * count should have been decremented to zero.  We only need to deal with the
2664  * space remaining in the current reservation part of the ticket.  If the
2665  * ticket contains a permanent reservation, there may be left over space which
2666  * needs to be released.  A count of N means that N-1 refills of the current
2667  * reservation can be done before we need to ask for more space.  The first
2668  * one goes to fill up the first current reservation.  Once we run out of
2669  * space, the count will stay at zero and the only space remaining will be
2670  * in the current reservation field.
2671  */
2672 STATIC void
2673 xlog_ungrant_log_space(xlog_t        *log,
2674                        xlog_ticket_t *ticket)
2675 {
2676         int     bytes;
2677
2678         if (ticket->t_cnt > 0)
2679                 ticket->t_cnt--;
2680
2681         trace_xfs_log_ungrant_enter(log, ticket);
2682         trace_xfs_log_ungrant_sub(log, ticket);
2683
2684         /*
2685          * If this is a permanent reservation ticket, we may be able to free
2686          * up more space based on the remaining count.
2687          */
2688         bytes = ticket->t_curr_res;
2689         if (ticket->t_cnt > 0) {
2690                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2691                 bytes += ticket->t_unit_res*ticket->t_cnt;
2692         }
2693
2694         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2695         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2696
2697         trace_xfs_log_ungrant_exit(log, ticket);
2698
2699         xfs_log_space_wake(log->l_mp);
2700 }
2701
2702 /*
2703  * Flush iclog to disk if this is the last reference to the given iclog and
2704  * the WANT_SYNC bit is set.
2705  *
2706  * When this function is entered, the iclog is not necessarily in the
2707  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2708  *
2709  *
2710  */
2711 STATIC int
2712 xlog_state_release_iclog(
2713         xlog_t          *log,
2714         xlog_in_core_t  *iclog)
2715 {
2716         int             sync = 0;       /* do we sync? */
2717
2718         if (iclog->ic_state & XLOG_STATE_IOERROR)
2719                 return XFS_ERROR(EIO);
2720
2721         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2722         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2723                 return 0;
2724
2725         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2726                 spin_unlock(&log->l_icloglock);
2727                 return XFS_ERROR(EIO);
2728         }
2729         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2730                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2731
2732         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2733                 /* update tail before writing to iclog */
2734                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2735                 sync++;
2736                 iclog->ic_state = XLOG_STATE_SYNCING;
2737                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2738                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2739                 /* cycle incremented when incrementing curr_block */
2740         }
2741         spin_unlock(&log->l_icloglock);
2742
2743         /*
2744          * We let the log lock go, so it's possible that we hit a log I/O
2745          * error or some other SHUTDOWN condition that marks the iclog
2746          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2747          * this iclog has consistent data, so we ignore IOERROR
2748          * flags after this point.
2749          */
2750         if (sync)
2751                 return xlog_sync(log, iclog);
2752         return 0;
2753 }       /* xlog_state_release_iclog */
2754
2755
2756 /*
2757  * This routine will mark the current iclog in the ring as WANT_SYNC
2758  * and move the current iclog pointer to the next iclog in the ring.
2759  * When this routine is called from xlog_state_get_iclog_space(), the
2760  * exact size of the iclog has not yet been determined.  All we know is
2761  * that every data block.  We have run out of space in this log record.
2762  */
2763 STATIC void
2764 xlog_state_switch_iclogs(xlog_t         *log,
2765                          xlog_in_core_t *iclog,
2766                          int            eventual_size)
2767 {
2768         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2769         if (!eventual_size)
2770                 eventual_size = iclog->ic_offset;
2771         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2772         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2773         log->l_prev_block = log->l_curr_block;
2774         log->l_prev_cycle = log->l_curr_cycle;
2775
2776         /* roll log?: ic_offset changed later */
2777         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2778
2779         /* Round up to next log-sunit */
2780         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2781             log->l_mp->m_sb.sb_logsunit > 1) {
2782                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2783                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2784         }
2785
2786         if (log->l_curr_block >= log->l_logBBsize) {
2787                 log->l_curr_cycle++;
2788                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2789                         log->l_curr_cycle++;
2790                 log->l_curr_block -= log->l_logBBsize;
2791                 ASSERT(log->l_curr_block >= 0);
2792         }
2793         ASSERT(iclog == log->l_iclog);
2794         log->l_iclog = iclog->ic_next;
2795 }       /* xlog_state_switch_iclogs */
2796
2797 /*
2798  * Write out all data in the in-core log as of this exact moment in time.
2799  *
2800  * Data may be written to the in-core log during this call.  However,
2801  * we don't guarantee this data will be written out.  A change from past
2802  * implementation means this routine will *not* write out zero length LRs.
2803  *
2804  * Basically, we try and perform an intelligent scan of the in-core logs.
2805  * If we determine there is no flushable data, we just return.  There is no
2806  * flushable data if:
2807  *
2808  *      1. the current iclog is active and has no data; the previous iclog
2809  *              is in the active or dirty state.
2810  *      2. the current iclog is drity, and the previous iclog is in the
2811  *              active or dirty state.
2812  *
2813  * We may sleep if:
2814  *
2815  *      1. the current iclog is not in the active nor dirty state.
2816  *      2. the current iclog dirty, and the previous iclog is not in the
2817  *              active nor dirty state.
2818  *      3. the current iclog is active, and there is another thread writing
2819  *              to this particular iclog.
2820  *      4. a) the current iclog is active and has no other writers
2821  *         b) when we return from flushing out this iclog, it is still
2822  *              not in the active nor dirty state.
2823  */
2824 int
2825 _xfs_log_force(
2826         struct xfs_mount        *mp,
2827         uint                    flags,
2828         int                     *log_flushed)
2829 {
2830         struct log              *log = mp->m_log;
2831         struct xlog_in_core     *iclog;
2832         xfs_lsn_t               lsn;
2833
2834         XFS_STATS_INC(xs_log_force);
2835
2836         xlog_cil_force(log);
2837
2838         spin_lock(&log->l_icloglock);
2839
2840         iclog = log->l_iclog;
2841         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2842                 spin_unlock(&log->l_icloglock);
2843                 return XFS_ERROR(EIO);
2844         }
2845
2846         /* If the head iclog is not active nor dirty, we just attach
2847          * ourselves to the head and go to sleep.
2848          */
2849         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2850             iclog->ic_state == XLOG_STATE_DIRTY) {
2851                 /*
2852                  * If the head is dirty or (active and empty), then
2853                  * we need to look at the previous iclog.  If the previous
2854                  * iclog is active or dirty we are done.  There is nothing
2855                  * to sync out.  Otherwise, we attach ourselves to the
2856                  * previous iclog and go to sleep.
2857                  */
2858                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2859                     (atomic_read(&iclog->ic_refcnt) == 0
2860                      && iclog->ic_offset == 0)) {
2861                         iclog = iclog->ic_prev;
2862                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2863                             iclog->ic_state == XLOG_STATE_DIRTY)
2864                                 goto no_sleep;
2865                         else
2866                                 goto maybe_sleep;
2867                 } else {
2868                         if (atomic_read(&iclog->ic_refcnt) == 0) {
2869                                 /* We are the only one with access to this
2870                                  * iclog.  Flush it out now.  There should
2871                                  * be a roundoff of zero to show that someone
2872                                  * has already taken care of the roundoff from
2873                                  * the previous sync.
2874                                  */
2875                                 atomic_inc(&iclog->ic_refcnt);
2876                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2877                                 xlog_state_switch_iclogs(log, iclog, 0);
2878                                 spin_unlock(&log->l_icloglock);
2879
2880                                 if (xlog_state_release_iclog(log, iclog))
2881                                         return XFS_ERROR(EIO);
2882
2883                                 if (log_flushed)
2884                                         *log_flushed = 1;
2885                                 spin_lock(&log->l_icloglock);
2886                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2887                                     iclog->ic_state != XLOG_STATE_DIRTY)
2888                                         goto maybe_sleep;
2889                                 else
2890                                         goto no_sleep;
2891                         } else {
2892                                 /* Someone else is writing to this iclog.
2893                                  * Use its call to flush out the data.  However,
2894                                  * the other thread may not force out this LR,
2895                                  * so we mark it WANT_SYNC.
2896                                  */
2897                                 xlog_state_switch_iclogs(log, iclog, 0);
2898                                 goto maybe_sleep;
2899                         }
2900                 }
2901         }
2902
2903         /* By the time we come around again, the iclog could've been filled
2904          * which would give it another lsn.  If we have a new lsn, just
2905          * return because the relevant data has been flushed.
2906          */
2907 maybe_sleep:
2908         if (flags & XFS_LOG_SYNC) {
2909                 /*
2910                  * We must check if we're shutting down here, before
2911                  * we wait, while we're holding the l_icloglock.
2912                  * Then we check again after waking up, in case our
2913                  * sleep was disturbed by a bad news.
2914                  */
2915                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2916                         spin_unlock(&log->l_icloglock);
2917                         return XFS_ERROR(EIO);
2918                 }
2919                 XFS_STATS_INC(xs_log_force_sleep);
2920                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
2921                 /*
2922                  * No need to grab the log lock here since we're
2923                  * only deciding whether or not to return EIO
2924                  * and the memory read should be atomic.
2925                  */
2926                 if (iclog->ic_state & XLOG_STATE_IOERROR)
2927                         return XFS_ERROR(EIO);
2928                 if (log_flushed)
2929                         *log_flushed = 1;
2930         } else {
2931
2932 no_sleep:
2933                 spin_unlock(&log->l_icloglock);
2934         }
2935         return 0;
2936 }
2937
2938 /*
2939  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
2940  * about errors or whether the log was flushed or not. This is the normal
2941  * interface to use when trying to unpin items or move the log forward.
2942  */
2943 void
2944 xfs_log_force(
2945         xfs_mount_t     *mp,
2946         uint            flags)
2947 {
2948         int     error;
2949
2950         error = _xfs_log_force(mp, flags, NULL);
2951         if (error)
2952                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
2953 }
2954
2955 /*
2956  * Force the in-core log to disk for a specific LSN.
2957  *
2958  * Find in-core log with lsn.
2959  *      If it is in the DIRTY state, just return.
2960  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2961  *              state and go to sleep or return.
2962  *      If it is in any other state, go to sleep or return.
2963  *
2964  * Synchronous forces are implemented with a signal variable. All callers
2965  * to force a given lsn to disk will wait on a the sv attached to the
2966  * specific in-core log.  When given in-core log finally completes its
2967  * write to disk, that thread will wake up all threads waiting on the
2968  * sv.
2969  */
2970 int
2971 _xfs_log_force_lsn(
2972         struct xfs_mount        *mp,
2973         xfs_lsn_t               lsn,
2974         uint                    flags,
2975         int                     *log_flushed)
2976 {
2977         struct log              *log = mp->m_log;
2978         struct xlog_in_core     *iclog;
2979         int                     already_slept = 0;
2980
2981         ASSERT(lsn != 0);
2982
2983         XFS_STATS_INC(xs_log_force);
2984
2985         lsn = xlog_cil_force_lsn(log, lsn);
2986         if (lsn == NULLCOMMITLSN)
2987                 return 0;
2988
2989 try_again:
2990         spin_lock(&log->l_icloglock);
2991         iclog = log->l_iclog;
2992         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2993                 spin_unlock(&log->l_icloglock);
2994                 return XFS_ERROR(EIO);
2995         }
2996
2997         do {
2998                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
2999                         iclog = iclog->ic_next;
3000                         continue;
3001                 }
3002
3003                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3004                         spin_unlock(&log->l_icloglock);
3005                         return 0;
3006                 }
3007
3008                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3009                         /*
3010                          * We sleep here if we haven't already slept (e.g.
3011                          * this is the first time we've looked at the correct
3012                          * iclog buf) and the buffer before us is going to
3013                          * be sync'ed. The reason for this is that if we
3014                          * are doing sync transactions here, by waiting for
3015                          * the previous I/O to complete, we can allow a few
3016                          * more transactions into this iclog before we close
3017                          * it down.
3018                          *
3019                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3020                          * up the refcnt so we can release the log (which
3021                          * drops the ref count).  The state switch keeps new
3022                          * transaction commits from using this buffer.  When
3023                          * the current commits finish writing into the buffer,
3024                          * the refcount will drop to zero and the buffer will
3025                          * go out then.
3026                          */
3027                         if (!already_slept &&
3028                             (iclog->ic_prev->ic_state &
3029                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3030                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3031
3032                                 XFS_STATS_INC(xs_log_force_sleep);
3033
3034                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3035                                                         &log->l_icloglock);
3036                                 if (log_flushed)
3037                                         *log_flushed = 1;
3038                                 already_slept = 1;
3039                                 goto try_again;
3040                         }
3041                         atomic_inc(&iclog->ic_refcnt);
3042                         xlog_state_switch_iclogs(log, iclog, 0);
3043                         spin_unlock(&log->l_icloglock);
3044                         if (xlog_state_release_iclog(log, iclog))
3045                                 return XFS_ERROR(EIO);
3046                         if (log_flushed)
3047                                 *log_flushed = 1;
3048                         spin_lock(&log->l_icloglock);
3049                 }
3050
3051                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3052                     !(iclog->ic_state &
3053                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3054                         /*
3055                          * Don't wait on completion if we know that we've
3056                          * gotten a log write error.
3057                          */
3058                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3059                                 spin_unlock(&log->l_icloglock);
3060                                 return XFS_ERROR(EIO);
3061                         }
3062                         XFS_STATS_INC(xs_log_force_sleep);
3063                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3064                         /*
3065                          * No need to grab the log lock here since we're
3066                          * only deciding whether or not to return EIO
3067                          * and the memory read should be atomic.
3068                          */
3069                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3070                                 return XFS_ERROR(EIO);
3071
3072                         if (log_flushed)
3073                                 *log_flushed = 1;
3074                 } else {                /* just return */
3075                         spin_unlock(&log->l_icloglock);
3076                 }
3077
3078                 return 0;
3079         } while (iclog != log->l_iclog);
3080
3081         spin_unlock(&log->l_icloglock);
3082         return 0;
3083 }
3084
3085 /*
3086  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3087  * about errors or whether the log was flushed or not. This is the normal
3088  * interface to use when trying to unpin items or move the log forward.
3089  */
3090 void
3091 xfs_log_force_lsn(
3092         xfs_mount_t     *mp,
3093         xfs_lsn_t       lsn,
3094         uint            flags)
3095 {
3096         int     error;
3097
3098         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3099         if (error)
3100                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3101 }
3102
3103 /*
3104  * Called when we want to mark the current iclog as being ready to sync to
3105  * disk.
3106  */
3107 STATIC void
3108 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3109 {
3110         assert_spin_locked(&log->l_icloglock);
3111
3112         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3113                 xlog_state_switch_iclogs(log, iclog, 0);
3114         } else {
3115                 ASSERT(iclog->ic_state &
3116                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3117         }
3118 }
3119
3120
3121 /*****************************************************************************
3122  *
3123  *              TICKET functions
3124  *
3125  *****************************************************************************
3126  */
3127
3128 /*
3129  * Free a used ticket when its refcount falls to zero.
3130  */
3131 void
3132 xfs_log_ticket_put(
3133         xlog_ticket_t   *ticket)
3134 {
3135         ASSERT(atomic_read(&ticket->t_ref) > 0);
3136         if (atomic_dec_and_test(&ticket->t_ref))
3137                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3138 }
3139
3140 xlog_ticket_t *
3141 xfs_log_ticket_get(
3142         xlog_ticket_t   *ticket)
3143 {
3144         ASSERT(atomic_read(&ticket->t_ref) > 0);
3145         atomic_inc(&ticket->t_ref);
3146         return ticket;
3147 }
3148
3149 /*
3150  * Allocate and initialise a new log ticket.
3151  */
3152 xlog_ticket_t *
3153 xlog_ticket_alloc(
3154         struct log      *log,
3155         int             unit_bytes,
3156         int             cnt,
3157         char            client,
3158         bool            permanent,
3159         int             alloc_flags)
3160 {
3161         struct xlog_ticket *tic;
3162         uint            num_headers;
3163         int             iclog_space;
3164
3165         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3166         if (!tic)
3167                 return NULL;
3168
3169         /*
3170          * Permanent reservations have up to 'cnt'-1 active log operations
3171          * in the log.  A unit in this case is the amount of space for one
3172          * of these log operations.  Normal reservations have a cnt of 1
3173          * and their unit amount is the total amount of space required.
3174          *
3175          * The following lines of code account for non-transaction data
3176          * which occupy space in the on-disk log.
3177          *
3178          * Normal form of a transaction is:
3179          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3180          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3181          *
3182          * We need to account for all the leadup data and trailer data
3183          * around the transaction data.
3184          * And then we need to account for the worst case in terms of using
3185          * more space.
3186          * The worst case will happen if:
3187          * - the placement of the transaction happens to be such that the
3188          *   roundoff is at its maximum
3189          * - the transaction data is synced before the commit record is synced
3190          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3191          *   Therefore the commit record is in its own Log Record.
3192          *   This can happen as the commit record is called with its
3193          *   own region to xlog_write().
3194          *   This then means that in the worst case, roundoff can happen for
3195          *   the commit-rec as well.
3196          *   The commit-rec is smaller than padding in this scenario and so it is
3197          *   not added separately.
3198          */
3199
3200         /* for trans header */
3201         unit_bytes += sizeof(xlog_op_header_t);
3202         unit_bytes += sizeof(xfs_trans_header_t);
3203
3204         /* for start-rec */
3205         unit_bytes += sizeof(xlog_op_header_t);
3206
3207         /*
3208          * for LR headers - the space for data in an iclog is the size minus
3209          * the space used for the headers. If we use the iclog size, then we
3210          * undercalculate the number of headers required.
3211          *
3212          * Furthermore - the addition of op headers for split-recs might
3213          * increase the space required enough to require more log and op
3214          * headers, so take that into account too.
3215          *
3216          * IMPORTANT: This reservation makes the assumption that if this
3217          * transaction is the first in an iclog and hence has the LR headers
3218          * accounted to it, then the remaining space in the iclog is
3219          * exclusively for this transaction.  i.e. if the transaction is larger
3220          * than the iclog, it will be the only thing in that iclog.
3221          * Fundamentally, this means we must pass the entire log vector to
3222          * xlog_write to guarantee this.
3223          */
3224         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3225         num_headers = howmany(unit_bytes, iclog_space);
3226
3227         /* for split-recs - ophdrs added when data split over LRs */
3228         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3229
3230         /* add extra header reservations if we overrun */
3231         while (!num_headers ||
3232                howmany(unit_bytes, iclog_space) > num_headers) {
3233                 unit_bytes += sizeof(xlog_op_header_t);
3234                 num_headers++;
3235         }
3236         unit_bytes += log->l_iclog_hsize * num_headers;
3237
3238         /* for commit-rec LR header - note: padding will subsume the ophdr */
3239         unit_bytes += log->l_iclog_hsize;
3240
3241         /* for roundoff padding for transaction data and one for commit record */
3242         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3243             log->l_mp->m_sb.sb_logsunit > 1) {
3244                 /* log su roundoff */
3245                 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3246         } else {
3247                 /* BB roundoff */
3248                 unit_bytes += 2*BBSIZE;
3249         }
3250
3251         atomic_set(&tic->t_ref, 1);
3252         tic->t_task             = current;
3253         INIT_LIST_HEAD(&tic->t_queue);
3254         tic->t_unit_res         = unit_bytes;
3255         tic->t_curr_res         = unit_bytes;
3256         tic->t_cnt              = cnt;
3257         tic->t_ocnt             = cnt;
3258         tic->t_tid              = random32();
3259         tic->t_clientid         = client;
3260         tic->t_flags            = XLOG_TIC_INITED;
3261         tic->t_trans_type       = 0;
3262         if (permanent)
3263                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3264
3265         xlog_tic_reset_res(tic);
3266
3267         return tic;
3268 }
3269
3270
3271 /******************************************************************************
3272  *
3273  *              Log debug routines
3274  *
3275  ******************************************************************************
3276  */
3277 #if defined(DEBUG)
3278 /*
3279  * Make sure that the destination ptr is within the valid data region of
3280  * one of the iclogs.  This uses backup pointers stored in a different
3281  * part of the log in case we trash the log structure.
3282  */
3283 void
3284 xlog_verify_dest_ptr(
3285         struct log      *log,
3286         char            *ptr)
3287 {
3288         int i;
3289         int good_ptr = 0;
3290
3291         for (i = 0; i < log->l_iclog_bufs; i++) {
3292                 if (ptr >= log->l_iclog_bak[i] &&
3293                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3294                         good_ptr++;
3295         }
3296
3297         if (!good_ptr)
3298                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3299 }
3300
3301 /*
3302  * Check to make sure the grant write head didn't just over lap the tail.  If
3303  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3304  * the cycles differ by exactly one and check the byte count.
3305  *
3306  * This check is run unlocked, so can give false positives. Rather than assert
3307  * on failures, use a warn-once flag and a panic tag to allow the admin to
3308  * determine if they want to panic the machine when such an error occurs. For
3309  * debug kernels this will have the same effect as using an assert but, unlinke
3310  * an assert, it can be turned off at runtime.
3311  */
3312 STATIC void
3313 xlog_verify_grant_tail(
3314         struct log      *log)
3315 {
3316         int             tail_cycle, tail_blocks;
3317         int             cycle, space;
3318
3319         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3320         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3321         if (tail_cycle != cycle) {
3322                 if (cycle - 1 != tail_cycle &&
3323                     !(log->l_flags & XLOG_TAIL_WARN)) {
3324                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3325                                 "%s: cycle - 1 != tail_cycle", __func__);
3326                         log->l_flags |= XLOG_TAIL_WARN;
3327                 }
3328
3329                 if (space > BBTOB(tail_blocks) &&
3330                     !(log->l_flags & XLOG_TAIL_WARN)) {
3331                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3332                                 "%s: space > BBTOB(tail_blocks)", __func__);
3333                         log->l_flags |= XLOG_TAIL_WARN;
3334                 }
3335         }
3336 }
3337
3338 /* check if it will fit */
3339 STATIC void
3340 xlog_verify_tail_lsn(xlog_t         *log,
3341                      xlog_in_core_t *iclog,
3342                      xfs_lsn_t      tail_lsn)
3343 {
3344     int blocks;
3345
3346     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3347         blocks =
3348             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3349         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3350                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3351     } else {
3352         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3353
3354         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3355                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3356
3357         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3358         if (blocks < BTOBB(iclog->ic_offset) + 1)
3359                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3360     }
3361 }       /* xlog_verify_tail_lsn */
3362
3363 /*
3364  * Perform a number of checks on the iclog before writing to disk.
3365  *
3366  * 1. Make sure the iclogs are still circular
3367  * 2. Make sure we have a good magic number
3368  * 3. Make sure we don't have magic numbers in the data
3369  * 4. Check fields of each log operation header for:
3370  *      A. Valid client identifier
3371  *      B. tid ptr value falls in valid ptr space (user space code)
3372  *      C. Length in log record header is correct according to the
3373  *              individual operation headers within record.
3374  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3375  *      log, check the preceding blocks of the physical log to make sure all
3376  *      the cycle numbers agree with the current cycle number.
3377  */
3378 STATIC void
3379 xlog_verify_iclog(xlog_t         *log,
3380                   xlog_in_core_t *iclog,
3381                   int            count,
3382                   boolean_t      syncing)
3383 {
3384         xlog_op_header_t        *ophead;
3385         xlog_in_core_t          *icptr;
3386         xlog_in_core_2_t        *xhdr;
3387         xfs_caddr_t             ptr;
3388         xfs_caddr_t             base_ptr;
3389         __psint_t               field_offset;
3390         __uint8_t               clientid;
3391         int                     len, i, j, k, op_len;
3392         int                     idx;
3393
3394         /* check validity of iclog pointers */
3395         spin_lock(&log->l_icloglock);
3396         icptr = log->l_iclog;
3397         for (i=0; i < log->l_iclog_bufs; i++) {
3398                 if (icptr == NULL)
3399                         xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3400                 icptr = icptr->ic_next;
3401         }
3402         if (icptr != log->l_iclog)
3403                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3404         spin_unlock(&log->l_icloglock);
3405
3406         /* check log magic numbers */
3407         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3408                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3409
3410         ptr = (xfs_caddr_t) &iclog->ic_header;
3411         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3412              ptr += BBSIZE) {
3413                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3414                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3415                                 __func__);
3416         }
3417
3418         /* check fields */
3419         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3420         ptr = iclog->ic_datap;
3421         base_ptr = ptr;
3422         ophead = (xlog_op_header_t *)ptr;
3423         xhdr = iclog->ic_data;
3424         for (i = 0; i < len; i++) {
3425                 ophead = (xlog_op_header_t *)ptr;
3426
3427                 /* clientid is only 1 byte */
3428                 field_offset = (__psint_t)
3429                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3430                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3431                         clientid = ophead->oh_clientid;
3432                 } else {
3433                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3434                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3435                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3436                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3437                                 clientid = xlog_get_client_id(
3438                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3439                         } else {
3440                                 clientid = xlog_get_client_id(
3441                                         iclog->ic_header.h_cycle_data[idx]);
3442                         }
3443                 }
3444                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3445                         xfs_warn(log->l_mp,
3446                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3447                                 __func__, clientid, ophead,
3448                                 (unsigned long)field_offset);
3449
3450                 /* check length */
3451                 field_offset = (__psint_t)
3452                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3453                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3454                         op_len = be32_to_cpu(ophead->oh_len);
3455                 } else {
3456                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3457                                     (__psint_t)iclog->ic_datap);
3458                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3459                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3460                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3461                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3462                         } else {
3463                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3464                         }
3465                 }
3466                 ptr += sizeof(xlog_op_header_t) + op_len;
3467         }
3468 }       /* xlog_verify_iclog */
3469 #endif
3470
3471 /*
3472  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3473  */
3474 STATIC int
3475 xlog_state_ioerror(
3476         xlog_t  *log)
3477 {
3478         xlog_in_core_t  *iclog, *ic;
3479
3480         iclog = log->l_iclog;
3481         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3482                 /*
3483                  * Mark all the incore logs IOERROR.
3484                  * From now on, no log flushes will result.
3485                  */
3486                 ic = iclog;
3487                 do {
3488                         ic->ic_state = XLOG_STATE_IOERROR;
3489                         ic = ic->ic_next;
3490                 } while (ic != iclog);
3491                 return 0;
3492         }
3493         /*
3494          * Return non-zero, if state transition has already happened.
3495          */
3496         return 1;
3497 }
3498
3499 /*
3500  * This is called from xfs_force_shutdown, when we're forcibly
3501  * shutting down the filesystem, typically because of an IO error.
3502  * Our main objectives here are to make sure that:
3503  *      a. the filesystem gets marked 'SHUTDOWN' for all interested
3504  *         parties to find out, 'atomically'.
3505  *      b. those who're sleeping on log reservations, pinned objects and
3506  *          other resources get woken up, and be told the bad news.
3507  *      c. nothing new gets queued up after (a) and (b) are done.
3508  *      d. if !logerror, flush the iclogs to disk, then seal them off
3509  *         for business.
3510  *
3511  * Note: for delayed logging the !logerror case needs to flush the regions
3512  * held in memory out to the iclogs before flushing them to disk. This needs
3513  * to be done before the log is marked as shutdown, otherwise the flush to the
3514  * iclogs will fail.
3515  */
3516 int
3517 xfs_log_force_umount(
3518         struct xfs_mount        *mp,
3519         int                     logerror)
3520 {
3521         xlog_t          *log;
3522         int             retval;
3523
3524         log = mp->m_log;
3525
3526         /*
3527          * If this happens during log recovery, don't worry about
3528          * locking; the log isn't open for business yet.
3529          */
3530         if (!log ||
3531             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3532                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3533                 if (mp->m_sb_bp)
3534                         XFS_BUF_DONE(mp->m_sb_bp);
3535                 return 0;
3536         }
3537
3538         /*
3539          * Somebody could've already done the hard work for us.
3540          * No need to get locks for this.
3541          */
3542         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3543                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3544                 return 1;
3545         }
3546         retval = 0;
3547
3548         /*
3549          * Flush the in memory commit item list before marking the log as
3550          * being shut down. We need to do it in this order to ensure all the
3551          * completed transactions are flushed to disk with the xfs_log_force()
3552          * call below.
3553          */
3554         if (!logerror)
3555                 xlog_cil_force(log);
3556
3557         /*
3558          * mark the filesystem and the as in a shutdown state and wake
3559          * everybody up to tell them the bad news.
3560          */
3561         spin_lock(&log->l_icloglock);
3562         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3563         if (mp->m_sb_bp)
3564                 XFS_BUF_DONE(mp->m_sb_bp);
3565
3566         /*
3567          * This flag is sort of redundant because of the mount flag, but
3568          * it's good to maintain the separation between the log and the rest
3569          * of XFS.
3570          */
3571         log->l_flags |= XLOG_IO_ERROR;
3572
3573         /*
3574          * If we hit a log error, we want to mark all the iclogs IOERROR
3575          * while we're still holding the loglock.
3576          */
3577         if (logerror)
3578                 retval = xlog_state_ioerror(log);
3579         spin_unlock(&log->l_icloglock);
3580
3581         /*
3582          * We don't want anybody waiting for log reservations after this. That
3583          * means we have to wake up everybody queued up on reserveq as well as
3584          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3585          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3586          * action is protected by the grant locks.
3587          */
3588         xlog_grant_head_wake_all(&log->l_reserve_head);
3589         xlog_grant_head_wake_all(&log->l_write_head);
3590
3591         if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3592                 ASSERT(!logerror);
3593                 /*
3594                  * Force the incore logs to disk before shutting the
3595                  * log down completely.
3596                  */
3597                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3598
3599                 spin_lock(&log->l_icloglock);
3600                 retval = xlog_state_ioerror(log);
3601                 spin_unlock(&log->l_icloglock);
3602         }
3603         /*
3604          * Wake up everybody waiting on xfs_log_force.
3605          * Callback all log item committed functions as if the
3606          * log writes were completed.
3607          */
3608         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3609
3610 #ifdef XFSERRORDEBUG
3611         {
3612                 xlog_in_core_t  *iclog;
3613
3614                 spin_lock(&log->l_icloglock);
3615                 iclog = log->l_iclog;
3616                 do {
3617                         ASSERT(iclog->ic_callback == 0);
3618                         iclog = iclog->ic_next;
3619                 } while (iclog != log->l_iclog);
3620                 spin_unlock(&log->l_icloglock);
3621         }
3622 #endif
3623         /* return non-zero if log IOERROR transition had already happened */
3624         return retval;
3625 }
3626
3627 STATIC int
3628 xlog_iclogs_empty(xlog_t *log)
3629 {
3630         xlog_in_core_t  *iclog;
3631
3632         iclog = log->l_iclog;
3633         do {
3634                 /* endianness does not matter here, zero is zero in
3635                  * any language.
3636                  */
3637                 if (iclog->ic_header.h_num_logops)
3638                         return 0;
3639                 iclog = iclog->ic_next;
3640         } while (iclog != log->l_iclog);
3641         return 1;
3642 }