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