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         iclogp = &log->l_iclog;
1201         /*
1202          * The amount of memory to allocate for the iclog structure is
1203          * rather funky due to the way the structure is defined.  It is
1204          * done this way so that we can use different sizes for machines
1205          * with different amounts of memory.  See the definition of
1206          * xlog_in_core_t in xfs_log_priv.h for details.
1207          */
1208         ASSERT(log->l_iclog_size >= 4096);
1209         for (i=0; i < log->l_iclog_bufs; i++) {
1210                 *iclogp = kmem_zalloc(sizeof(xlog_in_core_t), KM_MAYFAIL);
1211                 if (!*iclogp)
1212                         goto out_free_iclog;
1213
1214                 iclog = *iclogp;
1215                 iclog->ic_prev = prev_iclog;
1216                 prev_iclog = iclog;
1217
1218                 bp = xfs_buf_get_uncached(mp->m_logdev_targp,
1219                                                 BTOBB(log->l_iclog_size), 0);
1220                 if (!bp)
1221                         goto out_free_iclog;
1222
1223                 bp->b_iodone = xlog_iodone;
1224                 iclog->ic_bp = bp;
1225                 iclog->ic_data = bp->b_addr;
1226 #ifdef DEBUG
1227                 log->l_iclog_bak[i] = (xfs_caddr_t)&(iclog->ic_header);
1228 #endif
1229                 head = &iclog->ic_header;
1230                 memset(head, 0, sizeof(xlog_rec_header_t));
1231                 head->h_magicno = cpu_to_be32(XLOG_HEADER_MAGIC_NUM);
1232                 head->h_version = cpu_to_be32(
1233                         xfs_sb_version_haslogv2(&log->l_mp->m_sb) ? 2 : 1);
1234                 head->h_size = cpu_to_be32(log->l_iclog_size);
1235                 /* new fields */
1236                 head->h_fmt = cpu_to_be32(XLOG_FMT);
1237                 memcpy(&head->h_fs_uuid, &mp->m_sb.sb_uuid, sizeof(uuid_t));
1238
1239                 iclog->ic_size = BBTOB(bp->b_length) - log->l_iclog_hsize;
1240                 iclog->ic_state = XLOG_STATE_ACTIVE;
1241                 iclog->ic_log = log;
1242                 atomic_set(&iclog->ic_refcnt, 0);
1243                 spin_lock_init(&iclog->ic_callback_lock);
1244                 iclog->ic_callback_tail = &(iclog->ic_callback);
1245                 iclog->ic_datap = (char *)iclog->ic_data + log->l_iclog_hsize;
1246
1247                 ASSERT(xfs_buf_islocked(iclog->ic_bp));
1248                 init_waitqueue_head(&iclog->ic_force_wait);
1249                 init_waitqueue_head(&iclog->ic_write_wait);
1250
1251                 iclogp = &iclog->ic_next;
1252         }
1253         *iclogp = log->l_iclog;                 /* complete ring */
1254         log->l_iclog->ic_prev = prev_iclog;     /* re-write 1st prev ptr */
1255
1256         error = xlog_cil_init(log);
1257         if (error)
1258                 goto out_free_iclog;
1259         return log;
1260
1261 out_free_iclog:
1262         for (iclog = log->l_iclog; iclog; iclog = prev_iclog) {
1263                 prev_iclog = iclog->ic_next;
1264                 if (iclog->ic_bp)
1265                         xfs_buf_free(iclog->ic_bp);
1266                 kmem_free(iclog);
1267         }
1268         spinlock_destroy(&log->l_icloglock);
1269         xfs_buf_free(log->l_xbuf);
1270 out_free_log:
1271         kmem_free(log);
1272 out:
1273         return ERR_PTR(-error);
1274 }       /* xlog_alloc_log */
1275
1276
1277 /*
1278  * Write out the commit record of a transaction associated with the given
1279  * ticket.  Return the lsn of the commit record.
1280  */
1281 STATIC int
1282 xlog_commit_record(
1283         struct log              *log,
1284         struct xlog_ticket      *ticket,
1285         struct xlog_in_core     **iclog,
1286         xfs_lsn_t               *commitlsnp)
1287 {
1288         struct xfs_mount *mp = log->l_mp;
1289         int     error;
1290         struct xfs_log_iovec reg = {
1291                 .i_addr = NULL,
1292                 .i_len = 0,
1293                 .i_type = XLOG_REG_TYPE_COMMIT,
1294         };
1295         struct xfs_log_vec vec = {
1296                 .lv_niovecs = 1,
1297                 .lv_iovecp = &reg,
1298         };
1299
1300         ASSERT_ALWAYS(iclog);
1301         error = xlog_write(log, &vec, ticket, commitlsnp, iclog,
1302                                         XLOG_COMMIT_TRANS);
1303         if (error)
1304                 xfs_force_shutdown(mp, SHUTDOWN_LOG_IO_ERROR);
1305         return error;
1306 }
1307
1308 /*
1309  * Push on the buffer cache code if we ever use more than 75% of the on-disk
1310  * log space.  This code pushes on the lsn which would supposedly free up
1311  * the 25% which we want to leave free.  We may need to adopt a policy which
1312  * pushes on an lsn which is further along in the log once we reach the high
1313  * water mark.  In this manner, we would be creating a low water mark.
1314  */
1315 STATIC void
1316 xlog_grant_push_ail(
1317         struct log      *log,
1318         int             need_bytes)
1319 {
1320         xfs_lsn_t       threshold_lsn = 0;
1321         xfs_lsn_t       last_sync_lsn;
1322         int             free_blocks;
1323         int             free_bytes;
1324         int             threshold_block;
1325         int             threshold_cycle;
1326         int             free_threshold;
1327
1328         ASSERT(BTOBB(need_bytes) < log->l_logBBsize);
1329
1330         free_bytes = xlog_space_left(log, &log->l_reserve_head.grant);
1331         free_blocks = BTOBBT(free_bytes);
1332
1333         /*
1334          * Set the threshold for the minimum number of free blocks in the
1335          * log to the maximum of what the caller needs, one quarter of the
1336          * log, and 256 blocks.
1337          */
1338         free_threshold = BTOBB(need_bytes);
1339         free_threshold = MAX(free_threshold, (log->l_logBBsize >> 2));
1340         free_threshold = MAX(free_threshold, 256);
1341         if (free_blocks >= free_threshold)
1342                 return;
1343
1344         xlog_crack_atomic_lsn(&log->l_tail_lsn, &threshold_cycle,
1345                                                 &threshold_block);
1346         threshold_block += free_threshold;
1347         if (threshold_block >= log->l_logBBsize) {
1348                 threshold_block -= log->l_logBBsize;
1349                 threshold_cycle += 1;
1350         }
1351         threshold_lsn = xlog_assign_lsn(threshold_cycle,
1352                                         threshold_block);
1353         /*
1354          * Don't pass in an lsn greater than the lsn of the last
1355          * log record known to be on disk. Use a snapshot of the last sync lsn
1356          * so that it doesn't change between the compare and the set.
1357          */
1358         last_sync_lsn = atomic64_read(&log->l_last_sync_lsn);
1359         if (XFS_LSN_CMP(threshold_lsn, last_sync_lsn) > 0)
1360                 threshold_lsn = last_sync_lsn;
1361
1362         /*
1363          * Get the transaction layer to kick the dirty buffers out to
1364          * disk asynchronously. No point in trying to do this if
1365          * the filesystem is shutting down.
1366          */
1367         if (!XLOG_FORCED_SHUTDOWN(log))
1368                 xfs_ail_push(log->l_ailp, threshold_lsn);
1369 }
1370
1371 /*
1372  * The bdstrat callback function for log bufs. This gives us a central
1373  * place to trap bufs in case we get hit by a log I/O error and need to
1374  * shutdown. Actually, in practice, even when we didn't get a log error,
1375  * we transition the iclogs to IOERROR state *after* flushing all existing
1376  * iclogs to disk. This is because we don't want anymore new transactions to be
1377  * started or completed afterwards.
1378  */
1379 STATIC int
1380 xlog_bdstrat(
1381         struct xfs_buf          *bp)
1382 {
1383         struct xlog_in_core     *iclog = bp->b_fspriv;
1384
1385         if (iclog->ic_state & XLOG_STATE_IOERROR) {
1386                 xfs_buf_ioerror(bp, EIO);
1387                 xfs_buf_stale(bp);
1388                 xfs_buf_ioend(bp, 0);
1389                 /*
1390                  * It would seem logical to return EIO here, but we rely on
1391                  * the log state machine to propagate I/O errors instead of
1392                  * doing it here.
1393                  */
1394                 return 0;
1395         }
1396
1397         xfs_buf_iorequest(bp);
1398         return 0;
1399 }
1400
1401 /*
1402  * Flush out the in-core log (iclog) to the on-disk log in an asynchronous
1403  * fashion.  Previously, we should have moved the current iclog
1404  * ptr in the log to point to the next available iclog.  This allows further
1405  * write to continue while this code syncs out an iclog ready to go.
1406  * Before an in-core log can be written out, the data section must be scanned
1407  * to save away the 1st word of each BBSIZE block into the header.  We replace
1408  * it with the current cycle count.  Each BBSIZE block is tagged with the
1409  * cycle count because there in an implicit assumption that drives will
1410  * guarantee that entire 512 byte blocks get written at once.  In other words,
1411  * we can't have part of a 512 byte block written and part not written.  By
1412  * tagging each block, we will know which blocks are valid when recovering
1413  * after an unclean shutdown.
1414  *
1415  * This routine is single threaded on the iclog.  No other thread can be in
1416  * this routine with the same iclog.  Changing contents of iclog can there-
1417  * fore be done without grabbing the state machine lock.  Updating the global
1418  * log will require grabbing the lock though.
1419  *
1420  * The entire log manager uses a logical block numbering scheme.  Only
1421  * log_sync (and then only bwrite()) know about the fact that the log may
1422  * not start with block zero on a given device.  The log block start offset
1423  * is added immediately before calling bwrite().
1424  */
1425
1426 STATIC int
1427 xlog_sync(xlog_t                *log,
1428           xlog_in_core_t        *iclog)
1429 {
1430         xfs_caddr_t     dptr;           /* pointer to byte sized element */
1431         xfs_buf_t       *bp;
1432         int             i;
1433         uint            count;          /* byte count of bwrite */
1434         uint            count_init;     /* initial count before roundup */
1435         int             roundoff;       /* roundoff to BB or stripe */
1436         int             split = 0;      /* split write into two regions */
1437         int             error;
1438         int             v2 = xfs_sb_version_haslogv2(&log->l_mp->m_sb);
1439
1440         XFS_STATS_INC(xs_log_writes);
1441         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
1442
1443         /* Add for LR header */
1444         count_init = log->l_iclog_hsize + iclog->ic_offset;
1445
1446         /* Round out the log write size */
1447         if (v2 && log->l_mp->m_sb.sb_logsunit > 1) {
1448                 /* we have a v2 stripe unit to use */
1449                 count = XLOG_LSUNITTOB(log, XLOG_BTOLSUNIT(log, count_init));
1450         } else {
1451                 count = BBTOB(BTOBB(count_init));
1452         }
1453         roundoff = count - count_init;
1454         ASSERT(roundoff >= 0);
1455         ASSERT((v2 && log->l_mp->m_sb.sb_logsunit > 1 &&
1456                 roundoff < log->l_mp->m_sb.sb_logsunit)
1457                 ||
1458                 (log->l_mp->m_sb.sb_logsunit <= 1 &&
1459                  roundoff < BBTOB(1)));
1460
1461         /* move grant heads by roundoff in sync */
1462         xlog_grant_add_space(log, &log->l_reserve_head.grant, roundoff);
1463         xlog_grant_add_space(log, &log->l_write_head.grant, roundoff);
1464
1465         /* put cycle number in every block */
1466         xlog_pack_data(log, iclog, roundoff);
1467
1468         /* real byte length */
1469         if (v2) {
1470                 iclog->ic_header.h_len =
1471                         cpu_to_be32(iclog->ic_offset + roundoff);
1472         } else {
1473                 iclog->ic_header.h_len =
1474                         cpu_to_be32(iclog->ic_offset);
1475         }
1476
1477         bp = iclog->ic_bp;
1478         XFS_BUF_SET_ADDR(bp, BLOCK_LSN(be64_to_cpu(iclog->ic_header.h_lsn)));
1479
1480         XFS_STATS_ADD(xs_log_blocks, BTOBB(count));
1481
1482         /* Do we need to split this write into 2 parts? */
1483         if (XFS_BUF_ADDR(bp) + BTOBB(count) > log->l_logBBsize) {
1484                 split = count - (BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp)));
1485                 count = BBTOB(log->l_logBBsize - XFS_BUF_ADDR(bp));
1486                 iclog->ic_bwritecnt = 2;        /* split into 2 writes */
1487         } else {
1488                 iclog->ic_bwritecnt = 1;
1489         }
1490         bp->b_io_length = BTOBB(count);
1491         bp->b_fspriv = iclog;
1492         XFS_BUF_ZEROFLAGS(bp);
1493         XFS_BUF_ASYNC(bp);
1494         bp->b_flags |= XBF_SYNCIO;
1495
1496         if (log->l_mp->m_flags & XFS_MOUNT_BARRIER) {
1497                 bp->b_flags |= XBF_FUA;
1498
1499                 /*
1500                  * Flush the data device before flushing the log to make
1501                  * sure all meta data written back from the AIL actually made
1502                  * it to disk before stamping the new log tail LSN into the
1503                  * log buffer.  For an external log we need to issue the
1504                  * flush explicitly, and unfortunately synchronously here;
1505                  * for an internal log we can simply use the block layer
1506                  * state machine for preflushes.
1507                  */
1508                 if (log->l_mp->m_logdev_targp != log->l_mp->m_ddev_targp)
1509                         xfs_blkdev_issue_flush(log->l_mp->m_ddev_targp);
1510                 else
1511                         bp->b_flags |= XBF_FLUSH;
1512         }
1513
1514         ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1515         ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1516
1517         xlog_verify_iclog(log, iclog, count, B_TRUE);
1518
1519         /* account for log which doesn't start at block #0 */
1520         XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1521         /*
1522          * Don't call xfs_bwrite here. We do log-syncs even when the filesystem
1523          * is shutting down.
1524          */
1525         XFS_BUF_WRITE(bp);
1526
1527         error = xlog_bdstrat(bp);
1528         if (error) {
1529                 xfs_buf_ioerror_alert(bp, "xlog_sync");
1530                 return error;
1531         }
1532         if (split) {
1533                 bp = iclog->ic_log->l_xbuf;
1534                 XFS_BUF_SET_ADDR(bp, 0);             /* logical 0 */
1535                 xfs_buf_associate_memory(bp,
1536                                 (char *)&iclog->ic_header + count, split);
1537                 bp->b_fspriv = iclog;
1538                 XFS_BUF_ZEROFLAGS(bp);
1539                 XFS_BUF_ASYNC(bp);
1540                 bp->b_flags |= XBF_SYNCIO;
1541                 if (log->l_mp->m_flags & XFS_MOUNT_BARRIER)
1542                         bp->b_flags |= XBF_FUA;
1543                 dptr = bp->b_addr;
1544                 /*
1545                  * Bump the cycle numbers at the start of each block
1546                  * since this part of the buffer is at the start of
1547                  * a new cycle.  Watch out for the header magic number
1548                  * case, though.
1549                  */
1550                 for (i = 0; i < split; i += BBSIZE) {
1551                         be32_add_cpu((__be32 *)dptr, 1);
1552                         if (be32_to_cpu(*(__be32 *)dptr) == XLOG_HEADER_MAGIC_NUM)
1553                                 be32_add_cpu((__be32 *)dptr, 1);
1554                         dptr += BBSIZE;
1555                 }
1556
1557                 ASSERT(XFS_BUF_ADDR(bp) <= log->l_logBBsize-1);
1558                 ASSERT(XFS_BUF_ADDR(bp) + BTOBB(count) <= log->l_logBBsize);
1559
1560                 /* account for internal log which doesn't start at block #0 */
1561                 XFS_BUF_SET_ADDR(bp, XFS_BUF_ADDR(bp) + log->l_logBBstart);
1562                 XFS_BUF_WRITE(bp);
1563                 error = xlog_bdstrat(bp);
1564                 if (error) {
1565                         xfs_buf_ioerror_alert(bp, "xlog_sync (split)");
1566                         return error;
1567                 }
1568         }
1569         return 0;
1570 }       /* xlog_sync */
1571
1572
1573 /*
1574  * Deallocate a log structure
1575  */
1576 STATIC void
1577 xlog_dealloc_log(xlog_t *log)
1578 {
1579         xlog_in_core_t  *iclog, *next_iclog;
1580         int             i;
1581
1582         xlog_cil_destroy(log);
1583
1584         /*
1585          * always need to ensure that the extra buffer does not point to memory
1586          * owned by another log buffer before we free it.
1587          */
1588         xfs_buf_set_empty(log->l_xbuf, BTOBB(log->l_iclog_size));
1589         xfs_buf_free(log->l_xbuf);
1590
1591         iclog = log->l_iclog;
1592         for (i=0; i<log->l_iclog_bufs; i++) {
1593                 xfs_buf_free(iclog->ic_bp);
1594                 next_iclog = iclog->ic_next;
1595                 kmem_free(iclog);
1596                 iclog = next_iclog;
1597         }
1598         spinlock_destroy(&log->l_icloglock);
1599
1600         log->l_mp->m_log = NULL;
1601         kmem_free(log);
1602 }       /* xlog_dealloc_log */
1603
1604 /*
1605  * Update counters atomically now that memcpy is done.
1606  */
1607 /* ARGSUSED */
1608 static inline void
1609 xlog_state_finish_copy(xlog_t           *log,
1610                        xlog_in_core_t   *iclog,
1611                        int              record_cnt,
1612                        int              copy_bytes)
1613 {
1614         spin_lock(&log->l_icloglock);
1615
1616         be32_add_cpu(&iclog->ic_header.h_num_logops, record_cnt);
1617         iclog->ic_offset += copy_bytes;
1618
1619         spin_unlock(&log->l_icloglock);
1620 }       /* xlog_state_finish_copy */
1621
1622
1623
1624
1625 /*
1626  * print out info relating to regions written which consume
1627  * the reservation
1628  */
1629 void
1630 xlog_print_tic_res(
1631         struct xfs_mount        *mp,
1632         struct xlog_ticket      *ticket)
1633 {
1634         uint i;
1635         uint ophdr_spc = ticket->t_res_num_ophdrs * (uint)sizeof(xlog_op_header_t);
1636
1637         /* match with XLOG_REG_TYPE_* in xfs_log.h */
1638         static char *res_type_str[XLOG_REG_TYPE_MAX] = {
1639             "bformat",
1640             "bchunk",
1641             "efi_format",
1642             "efd_format",
1643             "iformat",
1644             "icore",
1645             "iext",
1646             "ibroot",
1647             "ilocal",
1648             "iattr_ext",
1649             "iattr_broot",
1650             "iattr_local",
1651             "qformat",
1652             "dquot",
1653             "quotaoff",
1654             "LR header",
1655             "unmount",
1656             "commit",
1657             "trans header"
1658         };
1659         static char *trans_type_str[XFS_TRANS_TYPE_MAX] = {
1660             "SETATTR_NOT_SIZE",
1661             "SETATTR_SIZE",
1662             "INACTIVE",
1663             "CREATE",
1664             "CREATE_TRUNC",
1665             "TRUNCATE_FILE",
1666             "REMOVE",
1667             "LINK",
1668             "RENAME",
1669             "MKDIR",
1670             "RMDIR",
1671             "SYMLINK",
1672             "SET_DMATTRS",
1673             "GROWFS",
1674             "STRAT_WRITE",
1675             "DIOSTRAT",
1676             "WRITE_SYNC",
1677             "WRITEID",
1678             "ADDAFORK",
1679             "ATTRINVAL",
1680             "ATRUNCATE",
1681             "ATTR_SET",
1682             "ATTR_RM",
1683             "ATTR_FLAG",
1684             "CLEAR_AGI_BUCKET",
1685             "QM_SBCHANGE",
1686             "DUMMY1",
1687             "DUMMY2",
1688             "QM_QUOTAOFF",
1689             "QM_DQALLOC",
1690             "QM_SETQLIM",
1691             "QM_DQCLUSTER",
1692             "QM_QINOCREATE",
1693             "QM_QUOTAOFF_END",
1694             "SB_UNIT",
1695             "FSYNC_TS",
1696             "GROWFSRT_ALLOC",
1697             "GROWFSRT_ZERO",
1698             "GROWFSRT_FREE",
1699             "SWAPEXT"
1700         };
1701
1702         xfs_warn(mp,
1703                 "xlog_write: reservation summary:\n"
1704                 "  trans type  = %s (%u)\n"
1705                 "  unit res    = %d bytes\n"
1706                 "  current res = %d bytes\n"
1707                 "  total reg   = %u bytes (o/flow = %u bytes)\n"
1708                 "  ophdrs      = %u (ophdr space = %u bytes)\n"
1709                 "  ophdr + reg = %u bytes\n"
1710                 "  num regions = %u\n",
1711                 ((ticket->t_trans_type <= 0 ||
1712                   ticket->t_trans_type > XFS_TRANS_TYPE_MAX) ?
1713                   "bad-trans-type" : trans_type_str[ticket->t_trans_type-1]),
1714                 ticket->t_trans_type,
1715                 ticket->t_unit_res,
1716                 ticket->t_curr_res,
1717                 ticket->t_res_arr_sum, ticket->t_res_o_flow,
1718                 ticket->t_res_num_ophdrs, ophdr_spc,
1719                 ticket->t_res_arr_sum +
1720                 ticket->t_res_o_flow + ophdr_spc,
1721                 ticket->t_res_num);
1722
1723         for (i = 0; i < ticket->t_res_num; i++) {
1724                 uint r_type = ticket->t_res_arr[i].r_type;
1725                 xfs_warn(mp, "region[%u]: %s - %u bytes\n", i,
1726                             ((r_type <= 0 || r_type > XLOG_REG_TYPE_MAX) ?
1727                             "bad-rtype" : res_type_str[r_type-1]),
1728                             ticket->t_res_arr[i].r_len);
1729         }
1730
1731         xfs_alert_tag(mp, XFS_PTAG_LOGRES,
1732                 "xlog_write: reservation ran out. Need to up reservation");
1733         xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
1734 }
1735
1736 /*
1737  * Calculate the potential space needed by the log vector.  Each region gets
1738  * its own xlog_op_header_t and may need to be double word aligned.
1739  */
1740 static int
1741 xlog_write_calc_vec_length(
1742         struct xlog_ticket      *ticket,
1743         struct xfs_log_vec      *log_vector)
1744 {
1745         struct xfs_log_vec      *lv;
1746         int                     headers = 0;
1747         int                     len = 0;
1748         int                     i;
1749
1750         /* acct for start rec of xact */
1751         if (ticket->t_flags & XLOG_TIC_INITED)
1752                 headers++;
1753
1754         for (lv = log_vector; lv; lv = lv->lv_next) {
1755                 headers += lv->lv_niovecs;
1756
1757                 for (i = 0; i < lv->lv_niovecs; i++) {
1758                         struct xfs_log_iovec    *vecp = &lv->lv_iovecp[i];
1759
1760                         len += vecp->i_len;
1761                         xlog_tic_add_region(ticket, vecp->i_len, vecp->i_type);
1762                 }
1763         }
1764
1765         ticket->t_res_num_ophdrs += headers;
1766         len += headers * sizeof(struct xlog_op_header);
1767
1768         return len;
1769 }
1770
1771 /*
1772  * If first write for transaction, insert start record  We can't be trying to
1773  * commit if we are inited.  We can't have any "partial_copy" if we are inited.
1774  */
1775 static int
1776 xlog_write_start_rec(
1777         struct xlog_op_header   *ophdr,
1778         struct xlog_ticket      *ticket)
1779 {
1780         if (!(ticket->t_flags & XLOG_TIC_INITED))
1781                 return 0;
1782
1783         ophdr->oh_tid   = cpu_to_be32(ticket->t_tid);
1784         ophdr->oh_clientid = ticket->t_clientid;
1785         ophdr->oh_len = 0;
1786         ophdr->oh_flags = XLOG_START_TRANS;
1787         ophdr->oh_res2 = 0;
1788
1789         ticket->t_flags &= ~XLOG_TIC_INITED;
1790
1791         return sizeof(struct xlog_op_header);
1792 }
1793
1794 static xlog_op_header_t *
1795 xlog_write_setup_ophdr(
1796         struct log              *log,
1797         struct xlog_op_header   *ophdr,
1798         struct xlog_ticket      *ticket,
1799         uint                    flags)
1800 {
1801         ophdr->oh_tid = cpu_to_be32(ticket->t_tid);
1802         ophdr->oh_clientid = ticket->t_clientid;
1803         ophdr->oh_res2 = 0;
1804
1805         /* are we copying a commit or unmount record? */
1806         ophdr->oh_flags = flags;
1807
1808         /*
1809          * We've seen logs corrupted with bad transaction client ids.  This
1810          * makes sure that XFS doesn't generate them on.  Turn this into an EIO
1811          * and shut down the filesystem.
1812          */
1813         switch (ophdr->oh_clientid)  {
1814         case XFS_TRANSACTION:
1815         case XFS_VOLUME:
1816         case XFS_LOG:
1817                 break;
1818         default:
1819                 xfs_warn(log->l_mp,
1820                         "Bad XFS transaction clientid 0x%x in ticket 0x%p",
1821                         ophdr->oh_clientid, ticket);
1822                 return NULL;
1823         }
1824
1825         return ophdr;
1826 }
1827
1828 /*
1829  * Set up the parameters of the region copy into the log. This has
1830  * to handle region write split across multiple log buffers - this
1831  * state is kept external to this function so that this code can
1832  * can be written in an obvious, self documenting manner.
1833  */
1834 static int
1835 xlog_write_setup_copy(
1836         struct xlog_ticket      *ticket,
1837         struct xlog_op_header   *ophdr,
1838         int                     space_available,
1839         int                     space_required,
1840         int                     *copy_off,
1841         int                     *copy_len,
1842         int                     *last_was_partial_copy,
1843         int                     *bytes_consumed)
1844 {
1845         int                     still_to_copy;
1846
1847         still_to_copy = space_required - *bytes_consumed;
1848         *copy_off = *bytes_consumed;
1849
1850         if (still_to_copy <= space_available) {
1851                 /* write of region completes here */
1852                 *copy_len = still_to_copy;
1853                 ophdr->oh_len = cpu_to_be32(*copy_len);
1854                 if (*last_was_partial_copy)
1855                         ophdr->oh_flags |= (XLOG_END_TRANS|XLOG_WAS_CONT_TRANS);
1856                 *last_was_partial_copy = 0;
1857                 *bytes_consumed = 0;
1858                 return 0;
1859         }
1860
1861         /* partial write of region, needs extra log op header reservation */
1862         *copy_len = space_available;
1863         ophdr->oh_len = cpu_to_be32(*copy_len);
1864         ophdr->oh_flags |= XLOG_CONTINUE_TRANS;
1865         if (*last_was_partial_copy)
1866                 ophdr->oh_flags |= XLOG_WAS_CONT_TRANS;
1867         *bytes_consumed += *copy_len;
1868         (*last_was_partial_copy)++;
1869
1870         /* account for new log op header */
1871         ticket->t_curr_res -= sizeof(struct xlog_op_header);
1872         ticket->t_res_num_ophdrs++;
1873
1874         return sizeof(struct xlog_op_header);
1875 }
1876
1877 static int
1878 xlog_write_copy_finish(
1879         struct log              *log,
1880         struct xlog_in_core     *iclog,
1881         uint                    flags,
1882         int                     *record_cnt,
1883         int                     *data_cnt,
1884         int                     *partial_copy,
1885         int                     *partial_copy_len,
1886         int                     log_offset,
1887         struct xlog_in_core     **commit_iclog)
1888 {
1889         if (*partial_copy) {
1890                 /*
1891                  * This iclog has already been marked WANT_SYNC by
1892                  * xlog_state_get_iclog_space.
1893                  */
1894                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1895                 *record_cnt = 0;
1896                 *data_cnt = 0;
1897                 return xlog_state_release_iclog(log, iclog);
1898         }
1899
1900         *partial_copy = 0;
1901         *partial_copy_len = 0;
1902
1903         if (iclog->ic_size - log_offset <= sizeof(xlog_op_header_t)) {
1904                 /* no more space in this iclog - push it. */
1905                 xlog_state_finish_copy(log, iclog, *record_cnt, *data_cnt);
1906                 *record_cnt = 0;
1907                 *data_cnt = 0;
1908
1909                 spin_lock(&log->l_icloglock);
1910                 xlog_state_want_sync(log, iclog);
1911                 spin_unlock(&log->l_icloglock);
1912
1913                 if (!commit_iclog)
1914                         return xlog_state_release_iclog(log, iclog);
1915                 ASSERT(flags & XLOG_COMMIT_TRANS);
1916                 *commit_iclog = iclog;
1917         }
1918
1919         return 0;
1920 }
1921
1922 /*
1923  * Write some region out to in-core log
1924  *
1925  * This will be called when writing externally provided regions or when
1926  * writing out a commit record for a given transaction.
1927  *
1928  * General algorithm:
1929  *      1. Find total length of this write.  This may include adding to the
1930  *              lengths passed in.
1931  *      2. Check whether we violate the tickets reservation.
1932  *      3. While writing to this iclog
1933  *          A. Reserve as much space in this iclog as can get
1934  *          B. If this is first write, save away start lsn
1935  *          C. While writing this region:
1936  *              1. If first write of transaction, write start record
1937  *              2. Write log operation header (header per region)
1938  *              3. Find out if we can fit entire region into this iclog
1939  *              4. Potentially, verify destination memcpy ptr
1940  *              5. Memcpy (partial) region
1941  *              6. If partial copy, release iclog; otherwise, continue
1942  *                      copying more regions into current iclog
1943  *      4. Mark want sync bit (in simulation mode)
1944  *      5. Release iclog for potential flush to on-disk log.
1945  *
1946  * ERRORS:
1947  * 1.   Panic if reservation is overrun.  This should never happen since
1948  *      reservation amounts are generated internal to the filesystem.
1949  * NOTES:
1950  * 1. Tickets are single threaded data structures.
1951  * 2. The XLOG_END_TRANS & XLOG_CONTINUE_TRANS flags are passed down to the
1952  *      syncing routine.  When a single log_write region needs to span
1953  *      multiple in-core logs, the XLOG_CONTINUE_TRANS bit should be set
1954  *      on all log operation writes which don't contain the end of the
1955  *      region.  The XLOG_END_TRANS bit is used for the in-core log
1956  *      operation which contains the end of the continued log_write region.
1957  * 3. When xlog_state_get_iclog_space() grabs the rest of the current iclog,
1958  *      we don't really know exactly how much space will be used.  As a result,
1959  *      we don't update ic_offset until the end when we know exactly how many
1960  *      bytes have been written out.
1961  */
1962 int
1963 xlog_write(
1964         struct log              *log,
1965         struct xfs_log_vec      *log_vector,
1966         struct xlog_ticket      *ticket,
1967         xfs_lsn_t               *start_lsn,
1968         struct xlog_in_core     **commit_iclog,
1969         uint                    flags)
1970 {
1971         struct xlog_in_core     *iclog = NULL;
1972         struct xfs_log_iovec    *vecp;
1973         struct xfs_log_vec      *lv;
1974         int                     len;
1975         int                     index;
1976         int                     partial_copy = 0;
1977         int                     partial_copy_len = 0;
1978         int                     contwr = 0;
1979         int                     record_cnt = 0;
1980         int                     data_cnt = 0;
1981         int                     error;
1982
1983         *start_lsn = 0;
1984
1985         len = xlog_write_calc_vec_length(ticket, log_vector);
1986
1987         /*
1988          * Region headers and bytes are already accounted for.
1989          * We only need to take into account start records and
1990          * split regions in this function.
1991          */
1992         if (ticket->t_flags & XLOG_TIC_INITED)
1993                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
1994
1995         /*
1996          * Commit record headers need to be accounted for. These
1997          * come in as separate writes so are easy to detect.
1998          */
1999         if (flags & (XLOG_COMMIT_TRANS | XLOG_UNMOUNT_TRANS))
2000                 ticket->t_curr_res -= sizeof(xlog_op_header_t);
2001
2002         if (ticket->t_curr_res < 0)
2003                 xlog_print_tic_res(log->l_mp, ticket);
2004
2005         index = 0;
2006         lv = log_vector;
2007         vecp = lv->lv_iovecp;
2008         while (lv && index < lv->lv_niovecs) {
2009                 void            *ptr;
2010                 int             log_offset;
2011
2012                 error = xlog_state_get_iclog_space(log, len, &iclog, ticket,
2013                                                    &contwr, &log_offset);
2014                 if (error)
2015                         return error;
2016
2017                 ASSERT(log_offset <= iclog->ic_size - 1);
2018                 ptr = iclog->ic_datap + log_offset;
2019
2020                 /* start_lsn is the first lsn written to. That's all we need. */
2021                 if (!*start_lsn)
2022                         *start_lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2023
2024                 /*
2025                  * This loop writes out as many regions as can fit in the amount
2026                  * of space which was allocated by xlog_state_get_iclog_space().
2027                  */
2028                 while (lv && index < lv->lv_niovecs) {
2029                         struct xfs_log_iovec    *reg = &vecp[index];
2030                         struct xlog_op_header   *ophdr;
2031                         int                     start_rec_copy;
2032                         int                     copy_len;
2033                         int                     copy_off;
2034
2035                         ASSERT(reg->i_len % sizeof(__int32_t) == 0);
2036                         ASSERT((unsigned long)ptr % sizeof(__int32_t) == 0);
2037
2038                         start_rec_copy = xlog_write_start_rec(ptr, ticket);
2039                         if (start_rec_copy) {
2040                                 record_cnt++;
2041                                 xlog_write_adv_cnt(&ptr, &len, &log_offset,
2042                                                    start_rec_copy);
2043                         }
2044
2045                         ophdr = xlog_write_setup_ophdr(log, ptr, ticket, flags);
2046                         if (!ophdr)
2047                                 return XFS_ERROR(EIO);
2048
2049                         xlog_write_adv_cnt(&ptr, &len, &log_offset,
2050                                            sizeof(struct xlog_op_header));
2051
2052                         len += xlog_write_setup_copy(ticket, ophdr,
2053                                                      iclog->ic_size-log_offset,
2054                                                      reg->i_len,
2055                                                      &copy_off, &copy_len,
2056                                                      &partial_copy,
2057                                                      &partial_copy_len);
2058                         xlog_verify_dest_ptr(log, ptr);
2059
2060                         /* copy region */
2061                         ASSERT(copy_len >= 0);
2062                         memcpy(ptr, reg->i_addr + copy_off, copy_len);
2063                         xlog_write_adv_cnt(&ptr, &len, &log_offset, copy_len);
2064
2065                         copy_len += start_rec_copy + sizeof(xlog_op_header_t);
2066                         record_cnt++;
2067                         data_cnt += contwr ? copy_len : 0;
2068
2069                         error = xlog_write_copy_finish(log, iclog, flags,
2070                                                        &record_cnt, &data_cnt,
2071                                                        &partial_copy,
2072                                                        &partial_copy_len,
2073                                                        log_offset,
2074                                                        commit_iclog);
2075                         if (error)
2076                                 return error;
2077
2078                         /*
2079                          * if we had a partial copy, we need to get more iclog
2080                          * space but we don't want to increment the region
2081                          * index because there is still more is this region to
2082                          * write.
2083                          *
2084                          * If we completed writing this region, and we flushed
2085                          * the iclog (indicated by resetting of the record
2086                          * count), then we also need to get more log space. If
2087                          * this was the last record, though, we are done and
2088                          * can just return.
2089                          */
2090                         if (partial_copy)
2091                                 break;
2092
2093                         if (++index == lv->lv_niovecs) {
2094                                 lv = lv->lv_next;
2095                                 index = 0;
2096                                 if (lv)
2097                                         vecp = lv->lv_iovecp;
2098                         }
2099                         if (record_cnt == 0) {
2100                                 if (!lv)
2101                                         return 0;
2102                                 break;
2103                         }
2104                 }
2105         }
2106
2107         ASSERT(len == 0);
2108
2109         xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
2110         if (!commit_iclog)
2111                 return xlog_state_release_iclog(log, iclog);
2112
2113         ASSERT(flags & XLOG_COMMIT_TRANS);
2114         *commit_iclog = iclog;
2115         return 0;
2116 }
2117
2118
2119 /*****************************************************************************
2120  *
2121  *              State Machine functions
2122  *
2123  *****************************************************************************
2124  */
2125
2126 /* Clean iclogs starting from the head.  This ordering must be
2127  * maintained, so an iclog doesn't become ACTIVE beyond one that
2128  * is SYNCING.  This is also required to maintain the notion that we use
2129  * a ordered wait queue to hold off would be writers to the log when every
2130  * iclog is trying to sync to disk.
2131  *
2132  * State Change: DIRTY -> ACTIVE
2133  */
2134 STATIC void
2135 xlog_state_clean_log(xlog_t *log)
2136 {
2137         xlog_in_core_t  *iclog;
2138         int changed = 0;
2139
2140         iclog = log->l_iclog;
2141         do {
2142                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
2143                         iclog->ic_state = XLOG_STATE_ACTIVE;
2144                         iclog->ic_offset       = 0;
2145                         ASSERT(iclog->ic_callback == NULL);
2146                         /*
2147                          * If the number of ops in this iclog indicate it just
2148                          * contains the dummy transaction, we can
2149                          * change state into IDLE (the second time around).
2150                          * Otherwise we should change the state into
2151                          * NEED a dummy.
2152                          * We don't need to cover the dummy.
2153                          */
2154                         if (!changed &&
2155                            (be32_to_cpu(iclog->ic_header.h_num_logops) ==
2156                                         XLOG_COVER_OPS)) {
2157                                 changed = 1;
2158                         } else {
2159                                 /*
2160                                  * We have two dirty iclogs so start over
2161                                  * This could also be num of ops indicates
2162                                  * this is not the dummy going out.
2163                                  */
2164                                 changed = 2;
2165                         }
2166                         iclog->ic_header.h_num_logops = 0;
2167                         memset(iclog->ic_header.h_cycle_data, 0,
2168                               sizeof(iclog->ic_header.h_cycle_data));
2169                         iclog->ic_header.h_lsn = 0;
2170                 } else if (iclog->ic_state == XLOG_STATE_ACTIVE)
2171                         /* do nothing */;
2172                 else
2173                         break;  /* stop cleaning */
2174                 iclog = iclog->ic_next;
2175         } while (iclog != log->l_iclog);
2176
2177         /* log is locked when we are called */
2178         /*
2179          * Change state for the dummy log recording.
2180          * We usually go to NEED. But we go to NEED2 if the changed indicates
2181          * we are done writing the dummy record.
2182          * If we are done with the second dummy recored (DONE2), then
2183          * we go to IDLE.
2184          */
2185         if (changed) {
2186                 switch (log->l_covered_state) {
2187                 case XLOG_STATE_COVER_IDLE:
2188                 case XLOG_STATE_COVER_NEED:
2189                 case XLOG_STATE_COVER_NEED2:
2190                         log->l_covered_state = XLOG_STATE_COVER_NEED;
2191                         break;
2192
2193                 case XLOG_STATE_COVER_DONE:
2194                         if (changed == 1)
2195                                 log->l_covered_state = XLOG_STATE_COVER_NEED2;
2196                         else
2197                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2198                         break;
2199
2200                 case XLOG_STATE_COVER_DONE2:
2201                         if (changed == 1)
2202                                 log->l_covered_state = XLOG_STATE_COVER_IDLE;
2203                         else
2204                                 log->l_covered_state = XLOG_STATE_COVER_NEED;
2205                         break;
2206
2207                 default:
2208                         ASSERT(0);
2209                 }
2210         }
2211 }       /* xlog_state_clean_log */
2212
2213 STATIC xfs_lsn_t
2214 xlog_get_lowest_lsn(
2215         xlog_t          *log)
2216 {
2217         xlog_in_core_t  *lsn_log;
2218         xfs_lsn_t       lowest_lsn, lsn;
2219
2220         lsn_log = log->l_iclog;
2221         lowest_lsn = 0;
2222         do {
2223             if (!(lsn_log->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY))) {
2224                 lsn = be64_to_cpu(lsn_log->ic_header.h_lsn);
2225                 if ((lsn && !lowest_lsn) ||
2226                     (XFS_LSN_CMP(lsn, lowest_lsn) < 0)) {
2227                         lowest_lsn = lsn;
2228                 }
2229             }
2230             lsn_log = lsn_log->ic_next;
2231         } while (lsn_log != log->l_iclog);
2232         return lowest_lsn;
2233 }
2234
2235
2236 STATIC void
2237 xlog_state_do_callback(
2238         xlog_t          *log,
2239         int             aborted,
2240         xlog_in_core_t  *ciclog)
2241 {
2242         xlog_in_core_t     *iclog;
2243         xlog_in_core_t     *first_iclog;        /* used to know when we've
2244                                                  * processed all iclogs once */
2245         xfs_log_callback_t *cb, *cb_next;
2246         int                flushcnt = 0;
2247         xfs_lsn_t          lowest_lsn;
2248         int                ioerrors;    /* counter: iclogs with errors */
2249         int                loopdidcallbacks; /* flag: inner loop did callbacks*/
2250         int                funcdidcallbacks; /* flag: function did callbacks */
2251         int                repeats;     /* for issuing console warnings if
2252                                          * looping too many times */
2253         int                wake = 0;
2254
2255         spin_lock(&log->l_icloglock);
2256         first_iclog = iclog = log->l_iclog;
2257         ioerrors = 0;
2258         funcdidcallbacks = 0;
2259         repeats = 0;
2260
2261         do {
2262                 /*
2263                  * Scan all iclogs starting with the one pointed to by the
2264                  * log.  Reset this starting point each time the log is
2265                  * unlocked (during callbacks).
2266                  *
2267                  * Keep looping through iclogs until one full pass is made
2268                  * without running any callbacks.
2269                  */
2270                 first_iclog = log->l_iclog;
2271                 iclog = log->l_iclog;
2272                 loopdidcallbacks = 0;
2273                 repeats++;
2274
2275                 do {
2276
2277                         /* skip all iclogs in the ACTIVE & DIRTY states */
2278                         if (iclog->ic_state &
2279                             (XLOG_STATE_ACTIVE|XLOG_STATE_DIRTY)) {
2280                                 iclog = iclog->ic_next;
2281                                 continue;
2282                         }
2283
2284                         /*
2285                          * Between marking a filesystem SHUTDOWN and stopping
2286                          * the log, we do flush all iclogs to disk (if there
2287                          * wasn't a log I/O error). So, we do want things to
2288                          * go smoothly in case of just a SHUTDOWN  w/o a
2289                          * LOG_IO_ERROR.
2290                          */
2291                         if (!(iclog->ic_state & XLOG_STATE_IOERROR)) {
2292                                 /*
2293                                  * Can only perform callbacks in order.  Since
2294                                  * this iclog is not in the DONE_SYNC/
2295                                  * DO_CALLBACK state, we skip the rest and
2296                                  * just try to clean up.  If we set our iclog
2297                                  * to DO_CALLBACK, we will not process it when
2298                                  * we retry since a previous iclog is in the
2299                                  * CALLBACK and the state cannot change since
2300                                  * we are holding the l_icloglock.
2301                                  */
2302                                 if (!(iclog->ic_state &
2303                                         (XLOG_STATE_DONE_SYNC |
2304                                                  XLOG_STATE_DO_CALLBACK))) {
2305                                         if (ciclog && (ciclog->ic_state ==
2306                                                         XLOG_STATE_DONE_SYNC)) {
2307                                                 ciclog->ic_state = XLOG_STATE_DO_CALLBACK;
2308                                         }
2309                                         break;
2310                                 }
2311                                 /*
2312                                  * We now have an iclog that is in either the
2313                                  * DO_CALLBACK or DONE_SYNC states. The other
2314                                  * states (WANT_SYNC, SYNCING, or CALLBACK were
2315                                  * caught by the above if and are going to
2316                                  * clean (i.e. we aren't doing their callbacks)
2317                                  * see the above if.
2318                                  */
2319
2320                                 /*
2321                                  * We will do one more check here to see if we
2322                                  * have chased our tail around.
2323                                  */
2324
2325                                 lowest_lsn = xlog_get_lowest_lsn(log);
2326                                 if (lowest_lsn &&
2327                                     XFS_LSN_CMP(lowest_lsn,
2328                                                 be64_to_cpu(iclog->ic_header.h_lsn)) < 0) {
2329                                         iclog = iclog->ic_next;
2330                                         continue; /* Leave this iclog for
2331                                                    * another thread */
2332                                 }
2333
2334                                 iclog->ic_state = XLOG_STATE_CALLBACK;
2335
2336
2337                                 /*
2338                                  * update the last_sync_lsn before we drop the
2339                                  * icloglock to ensure we are the only one that
2340                                  * can update it.
2341                                  */
2342                                 ASSERT(XFS_LSN_CMP(atomic64_read(&log->l_last_sync_lsn),
2343                                         be64_to_cpu(iclog->ic_header.h_lsn)) <= 0);
2344                                 atomic64_set(&log->l_last_sync_lsn,
2345                                         be64_to_cpu(iclog->ic_header.h_lsn));
2346
2347                         } else
2348                                 ioerrors++;
2349
2350                         spin_unlock(&log->l_icloglock);
2351
2352                         /*
2353                          * Keep processing entries in the callback list until
2354                          * we come around and it is empty.  We need to
2355                          * atomically see that the list is empty and change the
2356                          * state to DIRTY so that we don't miss any more
2357                          * callbacks being added.
2358                          */
2359                         spin_lock(&iclog->ic_callback_lock);
2360                         cb = iclog->ic_callback;
2361                         while (cb) {
2362                                 iclog->ic_callback_tail = &(iclog->ic_callback);
2363                                 iclog->ic_callback = NULL;
2364                                 spin_unlock(&iclog->ic_callback_lock);
2365
2366                                 /* perform callbacks in the order given */
2367                                 for (; cb; cb = cb_next) {
2368                                         cb_next = cb->cb_next;
2369                                         cb->cb_func(cb->cb_arg, aborted);
2370                                 }
2371                                 spin_lock(&iclog->ic_callback_lock);
2372                                 cb = iclog->ic_callback;
2373                         }
2374
2375                         loopdidcallbacks++;
2376                         funcdidcallbacks++;
2377
2378                         spin_lock(&log->l_icloglock);
2379                         ASSERT(iclog->ic_callback == NULL);
2380                         spin_unlock(&iclog->ic_callback_lock);
2381                         if (!(iclog->ic_state & XLOG_STATE_IOERROR))
2382                                 iclog->ic_state = XLOG_STATE_DIRTY;
2383
2384                         /*
2385                          * Transition from DIRTY to ACTIVE if applicable.
2386                          * NOP if STATE_IOERROR.
2387                          */
2388                         xlog_state_clean_log(log);
2389
2390                         /* wake up threads waiting in xfs_log_force() */
2391                         wake_up_all(&iclog->ic_force_wait);
2392
2393                         iclog = iclog->ic_next;
2394                 } while (first_iclog != iclog);
2395
2396                 if (repeats > 5000) {
2397                         flushcnt += repeats;
2398                         repeats = 0;
2399                         xfs_warn(log->l_mp,
2400                                 "%s: possible infinite loop (%d iterations)",
2401                                 __func__, flushcnt);
2402                 }
2403         } while (!ioerrors && loopdidcallbacks);
2404
2405         /*
2406          * make one last gasp attempt to see if iclogs are being left in
2407          * limbo..
2408          */
2409 #ifdef DEBUG
2410         if (funcdidcallbacks) {
2411                 first_iclog = iclog = log->l_iclog;
2412                 do {
2413                         ASSERT(iclog->ic_state != XLOG_STATE_DO_CALLBACK);
2414                         /*
2415                          * Terminate the loop if iclogs are found in states
2416                          * which will cause other threads to clean up iclogs.
2417                          *
2418                          * SYNCING - i/o completion will go through logs
2419                          * DONE_SYNC - interrupt thread should be waiting for
2420                          *              l_icloglock
2421                          * IOERROR - give up hope all ye who enter here
2422                          */
2423                         if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
2424                             iclog->ic_state == XLOG_STATE_SYNCING ||
2425                             iclog->ic_state == XLOG_STATE_DONE_SYNC ||
2426                             iclog->ic_state == XLOG_STATE_IOERROR )
2427                                 break;
2428                         iclog = iclog->ic_next;
2429                 } while (first_iclog != iclog);
2430         }
2431 #endif
2432
2433         if (log->l_iclog->ic_state & (XLOG_STATE_ACTIVE|XLOG_STATE_IOERROR))
2434                 wake = 1;
2435         spin_unlock(&log->l_icloglock);
2436
2437         if (wake)
2438                 wake_up_all(&log->l_flush_wait);
2439 }
2440
2441
2442 /*
2443  * Finish transitioning this iclog to the dirty state.
2444  *
2445  * Make sure that we completely execute this routine only when this is
2446  * the last call to the iclog.  There is a good chance that iclog flushes,
2447  * when we reach the end of the physical log, get turned into 2 separate
2448  * calls to bwrite.  Hence, one iclog flush could generate two calls to this
2449  * routine.  By using the reference count bwritecnt, we guarantee that only
2450  * the second completion goes through.
2451  *
2452  * Callbacks could take time, so they are done outside the scope of the
2453  * global state machine log lock.
2454  */
2455 STATIC void
2456 xlog_state_done_syncing(
2457         xlog_in_core_t  *iclog,
2458         int             aborted)
2459 {
2460         xlog_t             *log = iclog->ic_log;
2461
2462         spin_lock(&log->l_icloglock);
2463
2464         ASSERT(iclog->ic_state == XLOG_STATE_SYNCING ||
2465                iclog->ic_state == XLOG_STATE_IOERROR);
2466         ASSERT(atomic_read(&iclog->ic_refcnt) == 0);
2467         ASSERT(iclog->ic_bwritecnt == 1 || iclog->ic_bwritecnt == 2);
2468
2469
2470         /*
2471          * If we got an error, either on the first buffer, or in the case of
2472          * split log writes, on the second, we mark ALL iclogs STATE_IOERROR,
2473          * and none should ever be attempted to be written to disk
2474          * again.
2475          */
2476         if (iclog->ic_state != XLOG_STATE_IOERROR) {
2477                 if (--iclog->ic_bwritecnt == 1) {
2478                         spin_unlock(&log->l_icloglock);
2479                         return;
2480                 }
2481                 iclog->ic_state = XLOG_STATE_DONE_SYNC;
2482         }
2483
2484         /*
2485          * Someone could be sleeping prior to writing out the next
2486          * iclog buffer, we wake them all, one will get to do the
2487          * I/O, the others get to wait for the result.
2488          */
2489         wake_up_all(&iclog->ic_write_wait);
2490         spin_unlock(&log->l_icloglock);
2491         xlog_state_do_callback(log, aborted, iclog);    /* also cleans log */
2492 }       /* xlog_state_done_syncing */
2493
2494
2495 /*
2496  * If the head of the in-core log ring is not (ACTIVE or DIRTY), then we must
2497  * sleep.  We wait on the flush queue on the head iclog as that should be
2498  * the first iclog to complete flushing. Hence if all iclogs are syncing,
2499  * we will wait here and all new writes will sleep until a sync completes.
2500  *
2501  * The in-core logs are used in a circular fashion. They are not used
2502  * out-of-order even when an iclog past the head is free.
2503  *
2504  * return:
2505  *      * log_offset where xlog_write() can start writing into the in-core
2506  *              log's data space.
2507  *      * in-core log pointer to which xlog_write() should write.
2508  *      * boolean indicating this is a continued write to an in-core log.
2509  *              If this is the last write, then the in-core log's offset field
2510  *              needs to be incremented, depending on the amount of data which
2511  *              is copied.
2512  */
2513 STATIC int
2514 xlog_state_get_iclog_space(xlog_t         *log,
2515                            int            len,
2516                            xlog_in_core_t **iclogp,
2517                            xlog_ticket_t  *ticket,
2518                            int            *continued_write,
2519                            int            *logoffsetp)
2520 {
2521         int               log_offset;
2522         xlog_rec_header_t *head;
2523         xlog_in_core_t    *iclog;
2524         int               error;
2525
2526 restart:
2527         spin_lock(&log->l_icloglock);
2528         if (XLOG_FORCED_SHUTDOWN(log)) {
2529                 spin_unlock(&log->l_icloglock);
2530                 return XFS_ERROR(EIO);
2531         }
2532
2533         iclog = log->l_iclog;
2534         if (iclog->ic_state != XLOG_STATE_ACTIVE) {
2535                 XFS_STATS_INC(xs_log_noiclogs);
2536
2537                 /* Wait for log writes to have flushed */
2538                 xlog_wait(&log->l_flush_wait, &log->l_icloglock);
2539                 goto restart;
2540         }
2541
2542         head = &iclog->ic_header;
2543
2544         atomic_inc(&iclog->ic_refcnt);  /* prevents sync */
2545         log_offset = iclog->ic_offset;
2546
2547         /* On the 1st write to an iclog, figure out lsn.  This works
2548          * if iclogs marked XLOG_STATE_WANT_SYNC always write out what they are
2549          * committing to.  If the offset is set, that's how many blocks
2550          * must be written.
2551          */
2552         if (log_offset == 0) {
2553                 ticket->t_curr_res -= log->l_iclog_hsize;
2554                 xlog_tic_add_region(ticket,
2555                                     log->l_iclog_hsize,
2556                                     XLOG_REG_TYPE_LRHEADER);
2557                 head->h_cycle = cpu_to_be32(log->l_curr_cycle);
2558                 head->h_lsn = cpu_to_be64(
2559                         xlog_assign_lsn(log->l_curr_cycle, log->l_curr_block));
2560                 ASSERT(log->l_curr_block >= 0);
2561         }
2562
2563         /* If there is enough room to write everything, then do it.  Otherwise,
2564          * claim the rest of the region and make sure the XLOG_STATE_WANT_SYNC
2565          * bit is on, so this will get flushed out.  Don't update ic_offset
2566          * until you know exactly how many bytes get copied.  Therefore, wait
2567          * until later to update ic_offset.
2568          *
2569          * xlog_write() algorithm assumes that at least 2 xlog_op_header_t's
2570          * can fit into remaining data section.
2571          */
2572         if (iclog->ic_size - iclog->ic_offset < 2*sizeof(xlog_op_header_t)) {
2573                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2574
2575                 /*
2576                  * If I'm the only one writing to this iclog, sync it to disk.
2577                  * We need to do an atomic compare and decrement here to avoid
2578                  * racing with concurrent atomic_dec_and_lock() calls in
2579                  * xlog_state_release_iclog() when there is more than one
2580                  * reference to the iclog.
2581                  */
2582                 if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1)) {
2583                         /* we are the only one */
2584                         spin_unlock(&log->l_icloglock);
2585                         error = xlog_state_release_iclog(log, iclog);
2586                         if (error)
2587                                 return error;
2588                 } else {
2589                         spin_unlock(&log->l_icloglock);
2590                 }
2591                 goto restart;
2592         }
2593
2594         /* Do we have enough room to write the full amount in the remainder
2595          * of this iclog?  Or must we continue a write on the next iclog and
2596          * mark this iclog as completely taken?  In the case where we switch
2597          * iclogs (to mark it taken), this particular iclog will release/sync
2598          * to disk in xlog_write().
2599          */
2600         if (len <= iclog->ic_size - iclog->ic_offset) {
2601                 *continued_write = 0;
2602                 iclog->ic_offset += len;
2603         } else {
2604                 *continued_write = 1;
2605                 xlog_state_switch_iclogs(log, iclog, iclog->ic_size);
2606         }
2607         *iclogp = iclog;
2608
2609         ASSERT(iclog->ic_offset <= iclog->ic_size);
2610         spin_unlock(&log->l_icloglock);
2611
2612         *logoffsetp = log_offset;
2613         return 0;
2614 }       /* xlog_state_get_iclog_space */
2615
2616 /* The first cnt-1 times through here we don't need to
2617  * move the grant write head because the permanent
2618  * reservation has reserved cnt times the unit amount.
2619  * Release part of current permanent unit reservation and
2620  * reset current reservation to be one units worth.  Also
2621  * move grant reservation head forward.
2622  */
2623 STATIC void
2624 xlog_regrant_reserve_log_space(xlog_t        *log,
2625                                xlog_ticket_t *ticket)
2626 {
2627         trace_xfs_log_regrant_reserve_enter(log, ticket);
2628
2629         if (ticket->t_cnt > 0)
2630                 ticket->t_cnt--;
2631
2632         xlog_grant_sub_space(log, &log->l_reserve_head.grant,
2633                                         ticket->t_curr_res);
2634         xlog_grant_sub_space(log, &log->l_write_head.grant,
2635                                         ticket->t_curr_res);
2636         ticket->t_curr_res = ticket->t_unit_res;
2637         xlog_tic_reset_res(ticket);
2638
2639         trace_xfs_log_regrant_reserve_sub(log, ticket);
2640
2641         /* just return if we still have some of the pre-reserved space */
2642         if (ticket->t_cnt > 0)
2643                 return;
2644
2645         xlog_grant_add_space(log, &log->l_reserve_head.grant,
2646                                         ticket->t_unit_res);
2647
2648         trace_xfs_log_regrant_reserve_exit(log, ticket);
2649
2650         ticket->t_curr_res = ticket->t_unit_res;
2651         xlog_tic_reset_res(ticket);
2652 }       /* xlog_regrant_reserve_log_space */
2653
2654
2655 /*
2656  * Give back the space left from a reservation.
2657  *
2658  * All the information we need to make a correct determination of space left
2659  * is present.  For non-permanent reservations, things are quite easy.  The
2660  * count should have been decremented to zero.  We only need to deal with the
2661  * space remaining in the current reservation part of the ticket.  If the
2662  * ticket contains a permanent reservation, there may be left over space which
2663  * needs to be released.  A count of N means that N-1 refills of the current
2664  * reservation can be done before we need to ask for more space.  The first
2665  * one goes to fill up the first current reservation.  Once we run out of
2666  * space, the count will stay at zero and the only space remaining will be
2667  * in the current reservation field.
2668  */
2669 STATIC void
2670 xlog_ungrant_log_space(xlog_t        *log,
2671                        xlog_ticket_t *ticket)
2672 {
2673         int     bytes;
2674
2675         if (ticket->t_cnt > 0)
2676                 ticket->t_cnt--;
2677
2678         trace_xfs_log_ungrant_enter(log, ticket);
2679         trace_xfs_log_ungrant_sub(log, ticket);
2680
2681         /*
2682          * If this is a permanent reservation ticket, we may be able to free
2683          * up more space based on the remaining count.
2684          */
2685         bytes = ticket->t_curr_res;
2686         if (ticket->t_cnt > 0) {
2687                 ASSERT(ticket->t_flags & XLOG_TIC_PERM_RESERV);
2688                 bytes += ticket->t_unit_res*ticket->t_cnt;
2689         }
2690
2691         xlog_grant_sub_space(log, &log->l_reserve_head.grant, bytes);
2692         xlog_grant_sub_space(log, &log->l_write_head.grant, bytes);
2693
2694         trace_xfs_log_ungrant_exit(log, ticket);
2695
2696         xfs_log_space_wake(log->l_mp);
2697 }
2698
2699 /*
2700  * Flush iclog to disk if this is the last reference to the given iclog and
2701  * the WANT_SYNC bit is set.
2702  *
2703  * When this function is entered, the iclog is not necessarily in the
2704  * WANT_SYNC state.  It may be sitting around waiting to get filled.
2705  *
2706  *
2707  */
2708 STATIC int
2709 xlog_state_release_iclog(
2710         xlog_t          *log,
2711         xlog_in_core_t  *iclog)
2712 {
2713         int             sync = 0;       /* do we sync? */
2714
2715         if (iclog->ic_state & XLOG_STATE_IOERROR)
2716                 return XFS_ERROR(EIO);
2717
2718         ASSERT(atomic_read(&iclog->ic_refcnt) > 0);
2719         if (!atomic_dec_and_lock(&iclog->ic_refcnt, &log->l_icloglock))
2720                 return 0;
2721
2722         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2723                 spin_unlock(&log->l_icloglock);
2724                 return XFS_ERROR(EIO);
2725         }
2726         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE ||
2727                iclog->ic_state == XLOG_STATE_WANT_SYNC);
2728
2729         if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
2730                 /* update tail before writing to iclog */
2731                 xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
2732                 sync++;
2733                 iclog->ic_state = XLOG_STATE_SYNCING;
2734                 iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
2735                 xlog_verify_tail_lsn(log, iclog, tail_lsn);
2736                 /* cycle incremented when incrementing curr_block */
2737         }
2738         spin_unlock(&log->l_icloglock);
2739
2740         /*
2741          * We let the log lock go, so it's possible that we hit a log I/O
2742          * error or some other SHUTDOWN condition that marks the iclog
2743          * as XLOG_STATE_IOERROR before the bwrite. However, we know that
2744          * this iclog has consistent data, so we ignore IOERROR
2745          * flags after this point.
2746          */
2747         if (sync)
2748                 return xlog_sync(log, iclog);
2749         return 0;
2750 }       /* xlog_state_release_iclog */
2751
2752
2753 /*
2754  * This routine will mark the current iclog in the ring as WANT_SYNC
2755  * and move the current iclog pointer to the next iclog in the ring.
2756  * When this routine is called from xlog_state_get_iclog_space(), the
2757  * exact size of the iclog has not yet been determined.  All we know is
2758  * that every data block.  We have run out of space in this log record.
2759  */
2760 STATIC void
2761 xlog_state_switch_iclogs(xlog_t         *log,
2762                          xlog_in_core_t *iclog,
2763                          int            eventual_size)
2764 {
2765         ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
2766         if (!eventual_size)
2767                 eventual_size = iclog->ic_offset;
2768         iclog->ic_state = XLOG_STATE_WANT_SYNC;
2769         iclog->ic_header.h_prev_block = cpu_to_be32(log->l_prev_block);
2770         log->l_prev_block = log->l_curr_block;
2771         log->l_prev_cycle = log->l_curr_cycle;
2772
2773         /* roll log?: ic_offset changed later */
2774         log->l_curr_block += BTOBB(eventual_size)+BTOBB(log->l_iclog_hsize);
2775
2776         /* Round up to next log-sunit */
2777         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
2778             log->l_mp->m_sb.sb_logsunit > 1) {
2779                 __uint32_t sunit_bb = BTOBB(log->l_mp->m_sb.sb_logsunit);
2780                 log->l_curr_block = roundup(log->l_curr_block, sunit_bb);
2781         }
2782
2783         if (log->l_curr_block >= log->l_logBBsize) {
2784                 log->l_curr_cycle++;
2785                 if (log->l_curr_cycle == XLOG_HEADER_MAGIC_NUM)
2786                         log->l_curr_cycle++;
2787                 log->l_curr_block -= log->l_logBBsize;
2788                 ASSERT(log->l_curr_block >= 0);
2789         }
2790         ASSERT(iclog == log->l_iclog);
2791         log->l_iclog = iclog->ic_next;
2792 }       /* xlog_state_switch_iclogs */
2793
2794 /*
2795  * Write out all data in the in-core log as of this exact moment in time.
2796  *
2797  * Data may be written to the in-core log during this call.  However,
2798  * we don't guarantee this data will be written out.  A change from past
2799  * implementation means this routine will *not* write out zero length LRs.
2800  *
2801  * Basically, we try and perform an intelligent scan of the in-core logs.
2802  * If we determine there is no flushable data, we just return.  There is no
2803  * flushable data if:
2804  *
2805  *      1. the current iclog is active and has no data; the previous iclog
2806  *              is in the active or dirty state.
2807  *      2. the current iclog is drity, and the previous iclog is in the
2808  *              active or dirty state.
2809  *
2810  * We may sleep if:
2811  *
2812  *      1. the current iclog is not in the active nor dirty state.
2813  *      2. the current iclog dirty, and the previous iclog is not in the
2814  *              active nor dirty state.
2815  *      3. the current iclog is active, and there is another thread writing
2816  *              to this particular iclog.
2817  *      4. a) the current iclog is active and has no other writers
2818  *         b) when we return from flushing out this iclog, it is still
2819  *              not in the active nor dirty state.
2820  */
2821 int
2822 _xfs_log_force(
2823         struct xfs_mount        *mp,
2824         uint                    flags,
2825         int                     *log_flushed)
2826 {
2827         struct log              *log = mp->m_log;
2828         struct xlog_in_core     *iclog;
2829         xfs_lsn_t               lsn;
2830
2831         XFS_STATS_INC(xs_log_force);
2832
2833         xlog_cil_force(log);
2834
2835         spin_lock(&log->l_icloglock);
2836
2837         iclog = log->l_iclog;
2838         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2839                 spin_unlock(&log->l_icloglock);
2840                 return XFS_ERROR(EIO);
2841         }
2842
2843         /* If the head iclog is not active nor dirty, we just attach
2844          * ourselves to the head and go to sleep.
2845          */
2846         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2847             iclog->ic_state == XLOG_STATE_DIRTY) {
2848                 /*
2849                  * If the head is dirty or (active and empty), then
2850                  * we need to look at the previous iclog.  If the previous
2851                  * iclog is active or dirty we are done.  There is nothing
2852                  * to sync out.  Otherwise, we attach ourselves to the
2853                  * previous iclog and go to sleep.
2854                  */
2855                 if (iclog->ic_state == XLOG_STATE_DIRTY ||
2856                     (atomic_read(&iclog->ic_refcnt) == 0
2857                      && iclog->ic_offset == 0)) {
2858                         iclog = iclog->ic_prev;
2859                         if (iclog->ic_state == XLOG_STATE_ACTIVE ||
2860                             iclog->ic_state == XLOG_STATE_DIRTY)
2861                                 goto no_sleep;
2862                         else
2863                                 goto maybe_sleep;
2864                 } else {
2865                         if (atomic_read(&iclog->ic_refcnt) == 0) {
2866                                 /* We are the only one with access to this
2867                                  * iclog.  Flush it out now.  There should
2868                                  * be a roundoff of zero to show that someone
2869                                  * has already taken care of the roundoff from
2870                                  * the previous sync.
2871                                  */
2872                                 atomic_inc(&iclog->ic_refcnt);
2873                                 lsn = be64_to_cpu(iclog->ic_header.h_lsn);
2874                                 xlog_state_switch_iclogs(log, iclog, 0);
2875                                 spin_unlock(&log->l_icloglock);
2876
2877                                 if (xlog_state_release_iclog(log, iclog))
2878                                         return XFS_ERROR(EIO);
2879
2880                                 if (log_flushed)
2881                                         *log_flushed = 1;
2882                                 spin_lock(&log->l_icloglock);
2883                                 if (be64_to_cpu(iclog->ic_header.h_lsn) == lsn &&
2884                                     iclog->ic_state != XLOG_STATE_DIRTY)
2885                                         goto maybe_sleep;
2886                                 else
2887                                         goto no_sleep;
2888                         } else {
2889                                 /* Someone else is writing to this iclog.
2890                                  * Use its call to flush out the data.  However,
2891                                  * the other thread may not force out this LR,
2892                                  * so we mark it WANT_SYNC.
2893                                  */
2894                                 xlog_state_switch_iclogs(log, iclog, 0);
2895                                 goto maybe_sleep;
2896                         }
2897                 }
2898         }
2899
2900         /* By the time we come around again, the iclog could've been filled
2901          * which would give it another lsn.  If we have a new lsn, just
2902          * return because the relevant data has been flushed.
2903          */
2904 maybe_sleep:
2905         if (flags & XFS_LOG_SYNC) {
2906                 /*
2907                  * We must check if we're shutting down here, before
2908                  * we wait, while we're holding the l_icloglock.
2909                  * Then we check again after waking up, in case our
2910                  * sleep was disturbed by a bad news.
2911                  */
2912                 if (iclog->ic_state & XLOG_STATE_IOERROR) {
2913                         spin_unlock(&log->l_icloglock);
2914                         return XFS_ERROR(EIO);
2915                 }
2916                 XFS_STATS_INC(xs_log_force_sleep);
2917                 xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
2918                 /*
2919                  * No need to grab the log lock here since we're
2920                  * only deciding whether or not to return EIO
2921                  * and the memory read should be atomic.
2922                  */
2923                 if (iclog->ic_state & XLOG_STATE_IOERROR)
2924                         return XFS_ERROR(EIO);
2925                 if (log_flushed)
2926                         *log_flushed = 1;
2927         } else {
2928
2929 no_sleep:
2930                 spin_unlock(&log->l_icloglock);
2931         }
2932         return 0;
2933 }
2934
2935 /*
2936  * Wrapper for _xfs_log_force(), to be used when caller doesn't care
2937  * about errors or whether the log was flushed or not. This is the normal
2938  * interface to use when trying to unpin items or move the log forward.
2939  */
2940 void
2941 xfs_log_force(
2942         xfs_mount_t     *mp,
2943         uint            flags)
2944 {
2945         int     error;
2946
2947         error = _xfs_log_force(mp, flags, NULL);
2948         if (error)
2949                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
2950 }
2951
2952 /*
2953  * Force the in-core log to disk for a specific LSN.
2954  *
2955  * Find in-core log with lsn.
2956  *      If it is in the DIRTY state, just return.
2957  *      If it is in the ACTIVE state, move the in-core log into the WANT_SYNC
2958  *              state and go to sleep or return.
2959  *      If it is in any other state, go to sleep or return.
2960  *
2961  * Synchronous forces are implemented with a signal variable. All callers
2962  * to force a given lsn to disk will wait on a the sv attached to the
2963  * specific in-core log.  When given in-core log finally completes its
2964  * write to disk, that thread will wake up all threads waiting on the
2965  * sv.
2966  */
2967 int
2968 _xfs_log_force_lsn(
2969         struct xfs_mount        *mp,
2970         xfs_lsn_t               lsn,
2971         uint                    flags,
2972         int                     *log_flushed)
2973 {
2974         struct log              *log = mp->m_log;
2975         struct xlog_in_core     *iclog;
2976         int                     already_slept = 0;
2977
2978         ASSERT(lsn != 0);
2979
2980         XFS_STATS_INC(xs_log_force);
2981
2982         lsn = xlog_cil_force_lsn(log, lsn);
2983         if (lsn == NULLCOMMITLSN)
2984                 return 0;
2985
2986 try_again:
2987         spin_lock(&log->l_icloglock);
2988         iclog = log->l_iclog;
2989         if (iclog->ic_state & XLOG_STATE_IOERROR) {
2990                 spin_unlock(&log->l_icloglock);
2991                 return XFS_ERROR(EIO);
2992         }
2993
2994         do {
2995                 if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn) {
2996                         iclog = iclog->ic_next;
2997                         continue;
2998                 }
2999
3000                 if (iclog->ic_state == XLOG_STATE_DIRTY) {
3001                         spin_unlock(&log->l_icloglock);
3002                         return 0;
3003                 }
3004
3005                 if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3006                         /*
3007                          * We sleep here if we haven't already slept (e.g.
3008                          * this is the first time we've looked at the correct
3009                          * iclog buf) and the buffer before us is going to
3010                          * be sync'ed. The reason for this is that if we
3011                          * are doing sync transactions here, by waiting for
3012                          * the previous I/O to complete, we can allow a few
3013                          * more transactions into this iclog before we close
3014                          * it down.
3015                          *
3016                          * Otherwise, we mark the buffer WANT_SYNC, and bump
3017                          * up the refcnt so we can release the log (which
3018                          * drops the ref count).  The state switch keeps new
3019                          * transaction commits from using this buffer.  When
3020                          * the current commits finish writing into the buffer,
3021                          * the refcount will drop to zero and the buffer will
3022                          * go out then.
3023                          */
3024                         if (!already_slept &&
3025                             (iclog->ic_prev->ic_state &
3026                              (XLOG_STATE_WANT_SYNC | XLOG_STATE_SYNCING))) {
3027                                 ASSERT(!(iclog->ic_state & XLOG_STATE_IOERROR));
3028
3029                                 XFS_STATS_INC(xs_log_force_sleep);
3030
3031                                 xlog_wait(&iclog->ic_prev->ic_write_wait,
3032                                                         &log->l_icloglock);
3033                                 if (log_flushed)
3034                                         *log_flushed = 1;
3035                                 already_slept = 1;
3036                                 goto try_again;
3037                         }
3038                         atomic_inc(&iclog->ic_refcnt);
3039                         xlog_state_switch_iclogs(log, iclog, 0);
3040                         spin_unlock(&log->l_icloglock);
3041                         if (xlog_state_release_iclog(log, iclog))
3042                                 return XFS_ERROR(EIO);
3043                         if (log_flushed)
3044                                 *log_flushed = 1;
3045                         spin_lock(&log->l_icloglock);
3046                 }
3047
3048                 if ((flags & XFS_LOG_SYNC) && /* sleep */
3049                     !(iclog->ic_state &
3050                       (XLOG_STATE_ACTIVE | XLOG_STATE_DIRTY))) {
3051                         /*
3052                          * Don't wait on completion if we know that we've
3053                          * gotten a log write error.
3054                          */
3055                         if (iclog->ic_state & XLOG_STATE_IOERROR) {
3056                                 spin_unlock(&log->l_icloglock);
3057                                 return XFS_ERROR(EIO);
3058                         }
3059                         XFS_STATS_INC(xs_log_force_sleep);
3060                         xlog_wait(&iclog->ic_force_wait, &log->l_icloglock);
3061                         /*
3062                          * No need to grab the log lock here since we're
3063                          * only deciding whether or not to return EIO
3064                          * and the memory read should be atomic.
3065                          */
3066                         if (iclog->ic_state & XLOG_STATE_IOERROR)
3067                                 return XFS_ERROR(EIO);
3068
3069                         if (log_flushed)
3070                                 *log_flushed = 1;
3071                 } else {                /* just return */
3072                         spin_unlock(&log->l_icloglock);
3073                 }
3074
3075                 return 0;
3076         } while (iclog != log->l_iclog);
3077
3078         spin_unlock(&log->l_icloglock);
3079         return 0;
3080 }
3081
3082 /*
3083  * Wrapper for _xfs_log_force_lsn(), to be used when caller doesn't care
3084  * about errors or whether the log was flushed or not. This is the normal
3085  * interface to use when trying to unpin items or move the log forward.
3086  */
3087 void
3088 xfs_log_force_lsn(
3089         xfs_mount_t     *mp,
3090         xfs_lsn_t       lsn,
3091         uint            flags)
3092 {
3093         int     error;
3094
3095         error = _xfs_log_force_lsn(mp, lsn, flags, NULL);
3096         if (error)
3097                 xfs_warn(mp, "%s: error %d returned.", __func__, error);
3098 }
3099
3100 /*
3101  * Called when we want to mark the current iclog as being ready to sync to
3102  * disk.
3103  */
3104 STATIC void
3105 xlog_state_want_sync(xlog_t *log, xlog_in_core_t *iclog)
3106 {
3107         assert_spin_locked(&log->l_icloglock);
3108
3109         if (iclog->ic_state == XLOG_STATE_ACTIVE) {
3110                 xlog_state_switch_iclogs(log, iclog, 0);
3111         } else {
3112                 ASSERT(iclog->ic_state &
3113                         (XLOG_STATE_WANT_SYNC|XLOG_STATE_IOERROR));
3114         }
3115 }
3116
3117
3118 /*****************************************************************************
3119  *
3120  *              TICKET functions
3121  *
3122  *****************************************************************************
3123  */
3124
3125 /*
3126  * Free a used ticket when its refcount falls to zero.
3127  */
3128 void
3129 xfs_log_ticket_put(
3130         xlog_ticket_t   *ticket)
3131 {
3132         ASSERT(atomic_read(&ticket->t_ref) > 0);
3133         if (atomic_dec_and_test(&ticket->t_ref))
3134                 kmem_zone_free(xfs_log_ticket_zone, ticket);
3135 }
3136
3137 xlog_ticket_t *
3138 xfs_log_ticket_get(
3139         xlog_ticket_t   *ticket)
3140 {
3141         ASSERT(atomic_read(&ticket->t_ref) > 0);
3142         atomic_inc(&ticket->t_ref);
3143         return ticket;
3144 }
3145
3146 /*
3147  * Allocate and initialise a new log ticket.
3148  */
3149 xlog_ticket_t *
3150 xlog_ticket_alloc(
3151         struct log      *log,
3152         int             unit_bytes,
3153         int             cnt,
3154         char            client,
3155         bool            permanent,
3156         int             alloc_flags)
3157 {
3158         struct xlog_ticket *tic;
3159         uint            num_headers;
3160         int             iclog_space;
3161
3162         tic = kmem_zone_zalloc(xfs_log_ticket_zone, alloc_flags);
3163         if (!tic)
3164                 return NULL;
3165
3166         /*
3167          * Permanent reservations have up to 'cnt'-1 active log operations
3168          * in the log.  A unit in this case is the amount of space for one
3169          * of these log operations.  Normal reservations have a cnt of 1
3170          * and their unit amount is the total amount of space required.
3171          *
3172          * The following lines of code account for non-transaction data
3173          * which occupy space in the on-disk log.
3174          *
3175          * Normal form of a transaction is:
3176          * <oph><trans-hdr><start-oph><reg1-oph><reg1><reg2-oph>...<commit-oph>
3177          * and then there are LR hdrs, split-recs and roundoff at end of syncs.
3178          *
3179          * We need to account for all the leadup data and trailer data
3180          * around the transaction data.
3181          * And then we need to account for the worst case in terms of using
3182          * more space.
3183          * The worst case will happen if:
3184          * - the placement of the transaction happens to be such that the
3185          *   roundoff is at its maximum
3186          * - the transaction data is synced before the commit record is synced
3187          *   i.e. <transaction-data><roundoff> | <commit-rec><roundoff>
3188          *   Therefore the commit record is in its own Log Record.
3189          *   This can happen as the commit record is called with its
3190          *   own region to xlog_write().
3191          *   This then means that in the worst case, roundoff can happen for
3192          *   the commit-rec as well.
3193          *   The commit-rec is smaller than padding in this scenario and so it is
3194          *   not added separately.
3195          */
3196
3197         /* for trans header */
3198         unit_bytes += sizeof(xlog_op_header_t);
3199         unit_bytes += sizeof(xfs_trans_header_t);
3200
3201         /* for start-rec */
3202         unit_bytes += sizeof(xlog_op_header_t);
3203
3204         /*
3205          * for LR headers - the space for data in an iclog is the size minus
3206          * the space used for the headers. If we use the iclog size, then we
3207          * undercalculate the number of headers required.
3208          *
3209          * Furthermore - the addition of op headers for split-recs might
3210          * increase the space required enough to require more log and op
3211          * headers, so take that into account too.
3212          *
3213          * IMPORTANT: This reservation makes the assumption that if this
3214          * transaction is the first in an iclog and hence has the LR headers
3215          * accounted to it, then the remaining space in the iclog is
3216          * exclusively for this transaction.  i.e. if the transaction is larger
3217          * than the iclog, it will be the only thing in that iclog.
3218          * Fundamentally, this means we must pass the entire log vector to
3219          * xlog_write to guarantee this.
3220          */
3221         iclog_space = log->l_iclog_size - log->l_iclog_hsize;
3222         num_headers = howmany(unit_bytes, iclog_space);
3223
3224         /* for split-recs - ophdrs added when data split over LRs */
3225         unit_bytes += sizeof(xlog_op_header_t) * num_headers;
3226
3227         /* add extra header reservations if we overrun */
3228         while (!num_headers ||
3229                howmany(unit_bytes, iclog_space) > num_headers) {
3230                 unit_bytes += sizeof(xlog_op_header_t);
3231                 num_headers++;
3232         }
3233         unit_bytes += log->l_iclog_hsize * num_headers;
3234
3235         /* for commit-rec LR header - note: padding will subsume the ophdr */
3236         unit_bytes += log->l_iclog_hsize;
3237
3238         /* for roundoff padding for transaction data and one for commit record */
3239         if (xfs_sb_version_haslogv2(&log->l_mp->m_sb) &&
3240             log->l_mp->m_sb.sb_logsunit > 1) {
3241                 /* log su roundoff */
3242                 unit_bytes += 2*log->l_mp->m_sb.sb_logsunit;
3243         } else {
3244                 /* BB roundoff */
3245                 unit_bytes += 2*BBSIZE;
3246         }
3247
3248         atomic_set(&tic->t_ref, 1);
3249         tic->t_task             = current;
3250         INIT_LIST_HEAD(&tic->t_queue);
3251         tic->t_unit_res         = unit_bytes;
3252         tic->t_curr_res         = unit_bytes;
3253         tic->t_cnt              = cnt;
3254         tic->t_ocnt             = cnt;
3255         tic->t_tid              = random32();
3256         tic->t_clientid         = client;
3257         tic->t_flags            = XLOG_TIC_INITED;
3258         tic->t_trans_type       = 0;
3259         if (permanent)
3260                 tic->t_flags |= XLOG_TIC_PERM_RESERV;
3261
3262         xlog_tic_reset_res(tic);
3263
3264         return tic;
3265 }
3266
3267
3268 /******************************************************************************
3269  *
3270  *              Log debug routines
3271  *
3272  ******************************************************************************
3273  */
3274 #if defined(DEBUG)
3275 /*
3276  * Make sure that the destination ptr is within the valid data region of
3277  * one of the iclogs.  This uses backup pointers stored in a different
3278  * part of the log in case we trash the log structure.
3279  */
3280 void
3281 xlog_verify_dest_ptr(
3282         struct log      *log,
3283         char            *ptr)
3284 {
3285         int i;
3286         int good_ptr = 0;
3287
3288         for (i = 0; i < log->l_iclog_bufs; i++) {
3289                 if (ptr >= log->l_iclog_bak[i] &&
3290                     ptr <= log->l_iclog_bak[i] + log->l_iclog_size)
3291                         good_ptr++;
3292         }
3293
3294         if (!good_ptr)
3295                 xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3296 }
3297
3298 /*
3299  * Check to make sure the grant write head didn't just over lap the tail.  If
3300  * the cycles are the same, we can't be overlapping.  Otherwise, make sure that
3301  * the cycles differ by exactly one and check the byte count.
3302  *
3303  * This check is run unlocked, so can give false positives. Rather than assert
3304  * on failures, use a warn-once flag and a panic tag to allow the admin to
3305  * determine if they want to panic the machine when such an error occurs. For
3306  * debug kernels this will have the same effect as using an assert but, unlinke
3307  * an assert, it can be turned off at runtime.
3308  */
3309 STATIC void
3310 xlog_verify_grant_tail(
3311         struct log      *log)
3312 {
3313         int             tail_cycle, tail_blocks;
3314         int             cycle, space;
3315
3316         xlog_crack_grant_head(&log->l_write_head.grant, &cycle, &space);
3317         xlog_crack_atomic_lsn(&log->l_tail_lsn, &tail_cycle, &tail_blocks);
3318         if (tail_cycle != cycle) {
3319                 if (cycle - 1 != tail_cycle &&
3320                     !(log->l_flags & XLOG_TAIL_WARN)) {
3321                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3322                                 "%s: cycle - 1 != tail_cycle", __func__);
3323                         log->l_flags |= XLOG_TAIL_WARN;
3324                 }
3325
3326                 if (space > BBTOB(tail_blocks) &&
3327                     !(log->l_flags & XLOG_TAIL_WARN)) {
3328                         xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
3329                                 "%s: space > BBTOB(tail_blocks)", __func__);
3330                         log->l_flags |= XLOG_TAIL_WARN;
3331                 }
3332         }
3333 }
3334
3335 /* check if it will fit */
3336 STATIC void
3337 xlog_verify_tail_lsn(xlog_t         *log,
3338                      xlog_in_core_t *iclog,
3339                      xfs_lsn_t      tail_lsn)
3340 {
3341     int blocks;
3342
3343     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
3344         blocks =
3345             log->l_logBBsize - (log->l_prev_block - BLOCK_LSN(tail_lsn));
3346         if (blocks < BTOBB(iclog->ic_offset)+BTOBB(log->l_iclog_hsize))
3347                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3348     } else {
3349         ASSERT(CYCLE_LSN(tail_lsn)+1 == log->l_prev_cycle);
3350
3351         if (BLOCK_LSN(tail_lsn) == log->l_prev_block)
3352                 xfs_emerg(log->l_mp, "%s: tail wrapped", __func__);
3353
3354         blocks = BLOCK_LSN(tail_lsn) - log->l_prev_block;
3355         if (blocks < BTOBB(iclog->ic_offset) + 1)
3356                 xfs_emerg(log->l_mp, "%s: ran out of log space", __func__);
3357     }
3358 }       /* xlog_verify_tail_lsn */
3359
3360 /*
3361  * Perform a number of checks on the iclog before writing to disk.
3362  *
3363  * 1. Make sure the iclogs are still circular
3364  * 2. Make sure we have a good magic number
3365  * 3. Make sure we don't have magic numbers in the data
3366  * 4. Check fields of each log operation header for:
3367  *      A. Valid client identifier
3368  *      B. tid ptr value falls in valid ptr space (user space code)
3369  *      C. Length in log record header is correct according to the
3370  *              individual operation headers within record.
3371  * 5. When a bwrite will occur within 5 blocks of the front of the physical
3372  *      log, check the preceding blocks of the physical log to make sure all
3373  *      the cycle numbers agree with the current cycle number.
3374  */
3375 STATIC void
3376 xlog_verify_iclog(xlog_t         *log,
3377                   xlog_in_core_t *iclog,
3378                   int            count,
3379                   boolean_t      syncing)
3380 {
3381         xlog_op_header_t        *ophead;
3382         xlog_in_core_t          *icptr;
3383         xlog_in_core_2_t        *xhdr;
3384         xfs_caddr_t             ptr;
3385         xfs_caddr_t             base_ptr;
3386         __psint_t               field_offset;
3387         __uint8_t               clientid;
3388         int                     len, i, j, k, op_len;
3389         int                     idx;
3390
3391         /* check validity of iclog pointers */
3392         spin_lock(&log->l_icloglock);
3393         icptr = log->l_iclog;
3394         for (i=0; i < log->l_iclog_bufs; i++) {
3395                 if (icptr == NULL)
3396                         xfs_emerg(log->l_mp, "%s: invalid ptr", __func__);
3397                 icptr = icptr->ic_next;
3398         }
3399         if (icptr != log->l_iclog)
3400                 xfs_emerg(log->l_mp, "%s: corrupt iclog ring", __func__);
3401         spin_unlock(&log->l_icloglock);
3402
3403         /* check log magic numbers */
3404         if (iclog->ic_header.h_magicno != cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3405                 xfs_emerg(log->l_mp, "%s: invalid magic num", __func__);
3406
3407         ptr = (xfs_caddr_t) &iclog->ic_header;
3408         for (ptr += BBSIZE; ptr < ((xfs_caddr_t)&iclog->ic_header) + count;
3409              ptr += BBSIZE) {
3410                 if (*(__be32 *)ptr == cpu_to_be32(XLOG_HEADER_MAGIC_NUM))
3411                         xfs_emerg(log->l_mp, "%s: unexpected magic num",
3412                                 __func__);
3413         }
3414
3415         /* check fields */
3416         len = be32_to_cpu(iclog->ic_header.h_num_logops);
3417         ptr = iclog->ic_datap;
3418         base_ptr = ptr;
3419         ophead = (xlog_op_header_t *)ptr;
3420         xhdr = iclog->ic_data;
3421         for (i = 0; i < len; i++) {
3422                 ophead = (xlog_op_header_t *)ptr;
3423
3424                 /* clientid is only 1 byte */
3425                 field_offset = (__psint_t)
3426                                ((xfs_caddr_t)&(ophead->oh_clientid) - base_ptr);
3427                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3428                         clientid = ophead->oh_clientid;
3429                 } else {
3430                         idx = BTOBBT((xfs_caddr_t)&(ophead->oh_clientid) - iclog->ic_datap);
3431                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3432                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3433                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3434                                 clientid = xlog_get_client_id(
3435                                         xhdr[j].hic_xheader.xh_cycle_data[k]);
3436                         } else {
3437                                 clientid = xlog_get_client_id(
3438                                         iclog->ic_header.h_cycle_data[idx]);
3439                         }
3440                 }
3441                 if (clientid != XFS_TRANSACTION && clientid != XFS_LOG)
3442                         xfs_warn(log->l_mp,
3443                                 "%s: invalid clientid %d op 0x%p offset 0x%lx",
3444                                 __func__, clientid, ophead,
3445                                 (unsigned long)field_offset);
3446
3447                 /* check length */
3448                 field_offset = (__psint_t)
3449                                ((xfs_caddr_t)&(ophead->oh_len) - base_ptr);
3450                 if (syncing == B_FALSE || (field_offset & 0x1ff)) {
3451                         op_len = be32_to_cpu(ophead->oh_len);
3452                 } else {
3453                         idx = BTOBBT((__psint_t)&ophead->oh_len -
3454                                     (__psint_t)iclog->ic_datap);
3455                         if (idx >= (XLOG_HEADER_CYCLE_SIZE / BBSIZE)) {
3456                                 j = idx / (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3457                                 k = idx % (XLOG_HEADER_CYCLE_SIZE / BBSIZE);
3458                                 op_len = be32_to_cpu(xhdr[j].hic_xheader.xh_cycle_data[k]);
3459                         } else {
3460                                 op_len = be32_to_cpu(iclog->ic_header.h_cycle_data[idx]);
3461                         }
3462                 }
3463                 ptr += sizeof(xlog_op_header_t) + op_len;
3464         }
3465 }       /* xlog_verify_iclog */
3466 #endif
3467
3468 /*
3469  * Mark all iclogs IOERROR. l_icloglock is held by the caller.
3470  */
3471 STATIC int
3472 xlog_state_ioerror(
3473         xlog_t  *log)
3474 {
3475         xlog_in_core_t  *iclog, *ic;
3476
3477         iclog = log->l_iclog;
3478         if (! (iclog->ic_state & XLOG_STATE_IOERROR)) {
3479                 /*
3480                  * Mark all the incore logs IOERROR.
3481                  * From now on, no log flushes will result.
3482                  */
3483                 ic = iclog;
3484                 do {
3485                         ic->ic_state = XLOG_STATE_IOERROR;
3486                         ic = ic->ic_next;
3487                 } while (ic != iclog);
3488                 return 0;
3489         }
3490         /*
3491          * Return non-zero, if state transition has already happened.
3492          */
3493         return 1;
3494 }
3495
3496 /*
3497  * This is called from xfs_force_shutdown, when we're forcibly
3498  * shutting down the filesystem, typically because of an IO error.
3499  * Our main objectives here are to make sure that:
3500  *      a. the filesystem gets marked 'SHUTDOWN' for all interested
3501  *         parties to find out, 'atomically'.
3502  *      b. those who're sleeping on log reservations, pinned objects and
3503  *          other resources get woken up, and be told the bad news.
3504  *      c. nothing new gets queued up after (a) and (b) are done.
3505  *      d. if !logerror, flush the iclogs to disk, then seal them off
3506  *         for business.
3507  *
3508  * Note: for delayed logging the !logerror case needs to flush the regions
3509  * held in memory out to the iclogs before flushing them to disk. This needs
3510  * to be done before the log is marked as shutdown, otherwise the flush to the
3511  * iclogs will fail.
3512  */
3513 int
3514 xfs_log_force_umount(
3515         struct xfs_mount        *mp,
3516         int                     logerror)
3517 {
3518         xlog_t          *log;
3519         int             retval;
3520
3521         log = mp->m_log;
3522
3523         /*
3524          * If this happens during log recovery, don't worry about
3525          * locking; the log isn't open for business yet.
3526          */
3527         if (!log ||
3528             log->l_flags & XLOG_ACTIVE_RECOVERY) {
3529                 mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3530                 if (mp->m_sb_bp)
3531                         XFS_BUF_DONE(mp->m_sb_bp);
3532                 return 0;
3533         }
3534
3535         /*
3536          * Somebody could've already done the hard work for us.
3537          * No need to get locks for this.
3538          */
3539         if (logerror && log->l_iclog->ic_state & XLOG_STATE_IOERROR) {
3540                 ASSERT(XLOG_FORCED_SHUTDOWN(log));
3541                 return 1;
3542         }
3543         retval = 0;
3544
3545         /*
3546          * Flush the in memory commit item list before marking the log as
3547          * being shut down. We need to do it in this order to ensure all the
3548          * completed transactions are flushed to disk with the xfs_log_force()
3549          * call below.
3550          */
3551         if (!logerror)
3552                 xlog_cil_force(log);
3553
3554         /*
3555          * mark the filesystem and the as in a shutdown state and wake
3556          * everybody up to tell them the bad news.
3557          */
3558         spin_lock(&log->l_icloglock);
3559         mp->m_flags |= XFS_MOUNT_FS_SHUTDOWN;
3560         if (mp->m_sb_bp)
3561                 XFS_BUF_DONE(mp->m_sb_bp);
3562
3563         /*
3564          * This flag is sort of redundant because of the mount flag, but
3565          * it's good to maintain the separation between the log and the rest
3566          * of XFS.
3567          */
3568         log->l_flags |= XLOG_IO_ERROR;
3569
3570         /*
3571          * If we hit a log error, we want to mark all the iclogs IOERROR
3572          * while we're still holding the loglock.
3573          */
3574         if (logerror)
3575                 retval = xlog_state_ioerror(log);
3576         spin_unlock(&log->l_icloglock);
3577
3578         /*
3579          * We don't want anybody waiting for log reservations after this. That
3580          * means we have to wake up everybody queued up on reserveq as well as
3581          * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
3582          * we don't enqueue anything once the SHUTDOWN flag is set, and this
3583          * action is protected by the grant locks.
3584          */
3585         xlog_grant_head_wake_all(&log->l_reserve_head);
3586         xlog_grant_head_wake_all(&log->l_write_head);
3587
3588         if (!(log->l_iclog->ic_state & XLOG_STATE_IOERROR)) {
3589                 ASSERT(!logerror);
3590                 /*
3591                  * Force the incore logs to disk before shutting the
3592                  * log down completely.
3593                  */
3594                 _xfs_log_force(mp, XFS_LOG_SYNC, NULL);
3595
3596                 spin_lock(&log->l_icloglock);
3597                 retval = xlog_state_ioerror(log);
3598                 spin_unlock(&log->l_icloglock);
3599         }
3600         /*
3601          * Wake up everybody waiting on xfs_log_force.
3602          * Callback all log item committed functions as if the
3603          * log writes were completed.
3604          */
3605         xlog_state_do_callback(log, XFS_LI_ABORTED, NULL);
3606
3607 #ifdef XFSERRORDEBUG
3608         {
3609                 xlog_in_core_t  *iclog;
3610
3611                 spin_lock(&log->l_icloglock);
3612                 iclog = log->l_iclog;
3613                 do {
3614                         ASSERT(iclog->ic_callback == 0);
3615                         iclog = iclog->ic_next;
3616                 } while (iclog != log->l_iclog);
3617                 spin_unlock(&log->l_icloglock);
3618         }
3619 #endif
3620         /* return non-zero if log IOERROR transition had already happened */
3621         return retval;
3622 }
3623
3624 STATIC int
3625 xlog_iclogs_empty(xlog_t *log)
3626 {
3627         xlog_in_core_t  *iclog;
3628
3629         iclog = log->l_iclog;
3630         do {
3631                 /* endianness does not matter here, zero is zero in
3632                  * any language.
3633                  */
3634                 if (iclog->ic_header.h_num_logops)
3635                         return 0;
3636                 iclog = iclog->ic_next;
3637         } while (iclog != log->l_iclog);
3638         return 1;
3639 }