fs/gfs2/log.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   4  * Copyright (C) 2004-2007 Red Hat, Inc.  All rights reserved.
   5  */
   6
   7 #include <linux/sched.h>
   8 #include <linux/slab.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/completion.h>
  11 #include <linux/buffer_head.h>
  12 #include <linux/gfs2_ondisk.h>
  13 #include <linux/crc32.h>
  14 #include <linux/crc32c.h>
  15 #include <linux/delay.h>
  16 #include <linux/kthread.h>
  17 #include <linux/freezer.h>
  18 #include <linux/bio.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/writeback.h>
  21 #include <linux/list_sort.h>
  22
  23 #include "gfs2.h"
  24 #include "incore.h"
  25 #include "bmap.h"
  26 #include "glock.h"
  27 #include "log.h"
  28 #include "lops.h"
  29 #include "meta_io.h"
  30 #include "util.h"
  31 #include "dir.h"
  32 #include "trace_gfs2.h"
  33 #include "trans.h"
  34
  35 static void gfs2_log_shutdown(struct gfs2_sbd *sdp);
  36
  37 /**
  38  * gfs2_struct2blk - compute stuff
  39  * @sdp: the filesystem
  40  * @nstruct: the number of structures
  41  *
  42  * Compute the number of log descriptor blocks needed to hold a certain number
  43  * of structures of a certain size.
  44  *
  45  * Returns: the number of blocks needed (minimum is always 1)
  46  */
  47
  48 unsigned int gfs2_struct2blk(struct gfs2_sbd *sdp, unsigned int nstruct)
  49 {
  50         unsigned int blks;
  51         unsigned int first, second;
  52
  53         /* The initial struct gfs2_log_descriptor block */
  54         blks = 1;
  55         first = sdp->sd_ldptrs;
  56
  57         if (nstruct > first) {
  58                 /* Subsequent struct gfs2_meta_header blocks */
  59                 second = sdp->sd_inptrs;
  60                 blks += DIV_ROUND_UP(nstruct - first, second);
  61         }
  62
  63         return blks;
  64 }
  65
  66 /**
  67  * gfs2_remove_from_ail - Remove an entry from the ail lists, updating counters
  68  * @bd: The gfs2_bufdata to remove
  69  *
  70  * The ail lock _must_ be held when calling this function
  71  *
  72  */
  73
  74 void gfs2_remove_from_ail(struct gfs2_bufdata *bd)
  75 {
  76         bd->bd_tr = NULL;
  77         list_del_init(&bd->bd_ail_st_list);
  78         list_del_init(&bd->bd_ail_gl_list);
  79         atomic_dec(&bd->bd_gl->gl_ail_count);
  80         brelse(bd->bd_bh);
  81 }
  82
  83 static int __gfs2_writepage(struct folio *folio, struct writeback_control *wbc,
  84                        void *data)
  85 {
  86         struct address_space *mapping = data;
  87         int ret = mapping->a_ops->writepage(&folio->page, wbc);
  88         mapping_set_error(mapping, ret);
  89         return ret;
  90 }
  91
  92 /**
  93  * gfs2_ail1_start_one - Start I/O on a transaction
  94  * @sdp: The superblock
  95  * @wbc: The writeback control structure
  96  * @tr: The transaction to start I/O on
  97  * @plug: The block plug currently active
  98  */
  99
 100 static int gfs2_ail1_start_one(struct gfs2_sbd *sdp,
 101                                struct writeback_control *wbc,
 102                                struct gfs2_trans *tr, struct blk_plug *plug)
 103 __releases(&sdp->sd_ail_lock)
 104 __acquires(&sdp->sd_ail_lock)
 105 {
 106         struct gfs2_glock *gl = NULL;
 107         struct address_space *mapping;
 108         struct gfs2_bufdata *bd, *s;
 109         struct buffer_head *bh;
 110         int ret = 0;
 111
 112         list_for_each_entry_safe_reverse(bd, s, &tr->tr_ail1_list, bd_ail_st_list) {
 113                 bh = bd->bd_bh;
 114
 115                 gfs2_assert(sdp, bd->bd_tr == tr);
 116
 117                 if (!buffer_busy(bh)) {
 118                         if (buffer_uptodate(bh)) {
 119                                 list_move(&bd->bd_ail_st_list,
 120                                           &tr->tr_ail2_list);
 121                                 continue;
 122                         }
 123                         if (!cmpxchg(&sdp->sd_log_error, 0, -EIO)) {
 124                                 gfs2_io_error_bh(sdp, bh);
 125                                 gfs2_withdraw_delayed(sdp);
 126                         }
 127                 }
 128
 129                 if (gfs2_withdrawing_or_withdrawn(sdp)) {
 130                         gfs2_remove_from_ail(bd);
 131                         continue;
 132                 }
 133                 if (!buffer_dirty(bh))
 134                         continue;
 135                 if (gl == bd->bd_gl)
 136                         continue;
 137                 gl = bd->bd_gl;
 138                 list_move(&bd->bd_ail_st_list, &tr->tr_ail1_list);
 139                 mapping = bh->b_folio->mapping;
 140                 if (!mapping)
 141                         continue;
 142                 spin_unlock(&sdp->sd_ail_lock);
 143                 ret = write_cache_pages(mapping, wbc, __gfs2_writepage, mapping);
 144                 if (need_resched()) {
 145                         blk_finish_plug(plug);
 146                         cond_resched();
 147                         blk_start_plug(plug);
 148                 }
 149                 spin_lock(&sdp->sd_ail_lock);
 150                 if (ret == -ENODATA) /* if a jdata write into a new hole */
 151                         ret = 0; /* ignore it */
 152                 if (ret || wbc->nr_to_write <= 0)
 153                         break;
 154                 return -EBUSY;
 155         }
 156
 157         return ret;
 158 }
 159
 160 static void dump_ail_list(struct gfs2_sbd *sdp)
 161 {
 162         struct gfs2_trans *tr;
 163         struct gfs2_bufdata *bd;
 164         struct buffer_head *bh;
 165
 166         list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
 167                 list_for_each_entry_reverse(bd, &tr->tr_ail1_list,
 168                                             bd_ail_st_list) {
 169                         bh = bd->bd_bh;
 170                         fs_err(sdp, "bd %p: blk:0x%llx bh=%p ", bd,
 171                                (unsigned long long)bd->bd_blkno, bh);
 172                         if (!bh) {
 173                                 fs_err(sdp, "\n");
 174                                 continue;
 175                         }
 176                         fs_err(sdp, "0x%llx up2:%d dirt:%d lkd:%d req:%d "
 177                                "map:%d new:%d ar:%d aw:%d delay:%d "
 178                                "io err:%d unwritten:%d dfr:%d pin:%d esc:%d\n",
 179                                (unsigned long long)bh->b_blocknr,
 180                                buffer_uptodate(bh), buffer_dirty(bh),
 181                                buffer_locked(bh), buffer_req(bh),
 182                                buffer_mapped(bh), buffer_new(bh),
 183                                buffer_async_read(bh), buffer_async_write(bh),
 184                                buffer_delay(bh), buffer_write_io_error(bh),
 185                                buffer_unwritten(bh),
 186                                buffer_defer_completion(bh),
 187                                buffer_pinned(bh), buffer_escaped(bh));
 188                 }
 189         }
 190 }
 191
 192 /**
 193  * gfs2_ail1_flush - start writeback of some ail1 entries
 194  * @sdp: The super block
 195  * @wbc: The writeback control structure
 196  *
 197  * Writes back some ail1 entries, according to the limits in the
 198  * writeback control structure
 199  */
 200
 201 void gfs2_ail1_flush(struct gfs2_sbd *sdp, struct writeback_control *wbc)
 202 {
 203         struct list_head *head = &sdp->sd_ail1_list;
 204         struct gfs2_trans *tr;
 205         struct blk_plug plug;
 206         int ret;
 207         unsigned long flush_start = jiffies;
 208
 209         trace_gfs2_ail_flush(sdp, wbc, 1);
 210         blk_start_plug(&plug);
 211         spin_lock(&sdp->sd_ail_lock);
 212 restart:
 213         ret = 0;
 214         if (time_after(jiffies, flush_start + (HZ * 600))) {
 215                 fs_err(sdp, "Error: In %s for ten minutes! t=%d\n",
 216                        __func__, current->journal_info ? 1 : 0);
 217                 dump_ail_list(sdp);
 218                 goto out;
 219         }
 220         list_for_each_entry_reverse(tr, head, tr_list) {
 221                 if (wbc->nr_to_write <= 0)
 222                         break;
 223                 ret = gfs2_ail1_start_one(sdp, wbc, tr, &plug);
 224                 if (ret) {
 225                         if (ret == -EBUSY)
 226                                 goto restart;
 227                         break;
 228                 }
 229         }
 230 out:
 231         spin_unlock(&sdp->sd_ail_lock);
 232         blk_finish_plug(&plug);
 233         if (ret) {
 234                 gfs2_lm(sdp, "gfs2_ail1_start_one returned: %d\n", ret);
 235                 gfs2_withdraw(sdp);
 236         }
 237         trace_gfs2_ail_flush(sdp, wbc, 0);
 238 }
 239
 240 /**
 241  * gfs2_ail1_start - start writeback of all ail1 entries
 242  * @sdp: The superblock
 243  */
 244
 245 static void gfs2_ail1_start(struct gfs2_sbd *sdp)
 246 {
 247         struct writeback_control wbc = {
 248                 .sync_mode = WB_SYNC_NONE,
 249                 .nr_to_write = LONG_MAX,
 250                 .range_start = 0,
 251                 .range_end = LLONG_MAX,
 252         };
 253
 254         return gfs2_ail1_flush(sdp, &wbc);
 255 }
 256
 257 static void gfs2_log_update_flush_tail(struct gfs2_sbd *sdp)
 258 {
 259         unsigned int new_flush_tail = sdp->sd_log_head;
 260         struct gfs2_trans *tr;
 261
 262         if (!list_empty(&sdp->sd_ail1_list)) {
 263                 tr = list_last_entry(&sdp->sd_ail1_list,
 264                                      struct gfs2_trans, tr_list);
 265                 new_flush_tail = tr->tr_first;
 266         }
 267         sdp->sd_log_flush_tail = new_flush_tail;
 268 }
 269
 270 static void gfs2_log_update_head(struct gfs2_sbd *sdp)
 271 {
 272         unsigned int new_head = sdp->sd_log_flush_head;
 273
 274         if (sdp->sd_log_flush_tail == sdp->sd_log_head)
 275                 sdp->sd_log_flush_tail = new_head;
 276         sdp->sd_log_head = new_head;
 277 }
 278
 279 /*
 280  * gfs2_ail_empty_tr - empty one of the ail lists of a transaction
 281  */
 282
 283 static void gfs2_ail_empty_tr(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
 284                               struct list_head *head)
 285 {
 286         struct gfs2_bufdata *bd;
 287
 288         while (!list_empty(head)) {
 289                 bd = list_first_entry(head, struct gfs2_bufdata,
 290                                       bd_ail_st_list);
 291                 gfs2_assert(sdp, bd->bd_tr == tr);
 292                 gfs2_remove_from_ail(bd);
 293         }
 294 }
 295
 296 /**
 297  * gfs2_ail1_empty_one - Check whether or not a trans in the AIL has been synced
 298  * @sdp: the filesystem
 299  * @tr: the transaction
 300  * @max_revokes: If nonzero, issue revokes for the bd items for written buffers
 301  *
 302  * returns: the transaction's count of remaining active items
 303  */
 304
 305 static int gfs2_ail1_empty_one(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
 306                                 int *max_revokes)
 307 {
 308         struct gfs2_bufdata *bd, *s;
 309         struct buffer_head *bh;
 310         int active_count = 0;
 311
 312         list_for_each_entry_safe_reverse(bd, s, &tr->tr_ail1_list,
 313                                          bd_ail_st_list) {
 314                 bh = bd->bd_bh;
 315                 gfs2_assert(sdp, bd->bd_tr == tr);
 316                 /*
 317                  * If another process flagged an io error, e.g. writing to the
 318                  * journal, error all other bhs and move them off the ail1 to
 319                  * prevent a tight loop when unmount tries to flush ail1,
 320                  * regardless of whether they're still busy. If no outside
 321                  * errors were found and the buffer is busy, move to the next.
 322                  * If the ail buffer is not busy and caught an error, flag it
 323                  * for others.
 324                  */
 325                 if (!sdp->sd_log_error && buffer_busy(bh)) {
 326                         active_count++;
 327                         continue;
 328                 }
 329                 if (!buffer_uptodate(bh) &&
 330                     !cmpxchg(&sdp->sd_log_error, 0, -EIO)) {
 331                         gfs2_io_error_bh(sdp, bh);
 332                         gfs2_withdraw_delayed(sdp);
 333                 }
 334                 /*
 335                  * If we have space for revokes and the bd is no longer on any
 336                  * buf list, we can just add a revoke for it immediately and
 337                  * avoid having to put it on the ail2 list, where it would need
 338                  * to be revoked later.
 339                  */
 340                 if (*max_revokes && list_empty(&bd->bd_list)) {
 341                         gfs2_add_revoke(sdp, bd);
 342                         (*max_revokes)--;
 343                         continue;
 344                 }
 345                 list_move(&bd->bd_ail_st_list, &tr->tr_ail2_list);
 346         }
 347         return active_count;
 348 }
 349
 350 /**
 351  * gfs2_ail1_empty - Try to empty the ail1 lists
 352  * @sdp: The superblock
 353  * @max_revokes: If non-zero, add revokes where appropriate
 354  *
 355  * Tries to empty the ail1 lists, starting with the oldest first.
 356  * Returns %true if the ail1 list is now empty.
 357  */
 358
 359 static bool gfs2_ail1_empty(struct gfs2_sbd *sdp, int max_revokes)
 360 {
 361         struct gfs2_trans *tr, *s;
 362         int oldest_tr = 1;
 363         bool empty;
 364
 365         spin_lock(&sdp->sd_ail_lock);
 366         list_for_each_entry_safe_reverse(tr, s, &sdp->sd_ail1_list, tr_list) {
 367                 if (!gfs2_ail1_empty_one(sdp, tr, &max_revokes) && oldest_tr)
 368                         list_move(&tr->tr_list, &sdp->sd_ail2_list);
 369                 else
 370                         oldest_tr = 0;
 371         }
 372         gfs2_log_update_flush_tail(sdp);
 373         empty = list_empty(&sdp->sd_ail1_list);
 374         spin_unlock(&sdp->sd_ail_lock);
 375
 376         return empty;
 377 }
 378
 379 static void gfs2_ail1_wait(struct gfs2_sbd *sdp)
 380 {
 381         struct gfs2_trans *tr;
 382         struct gfs2_bufdata *bd;
 383         struct buffer_head *bh;
 384
 385         spin_lock(&sdp->sd_ail_lock);
 386         list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
 387                 list_for_each_entry(bd, &tr->tr_ail1_list, bd_ail_st_list) {
 388                         bh = bd->bd_bh;
 389                         if (!buffer_locked(bh))
 390                                 continue;
 391                         get_bh(bh);
 392                         spin_unlock(&sdp->sd_ail_lock);
 393                         wait_on_buffer(bh);
 394                         brelse(bh);
 395                         return;
 396                 }
 397         }
 398         spin_unlock(&sdp->sd_ail_lock);
 399 }
 400
 401 static void __ail2_empty(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 402 {
 403         gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
 404         list_del(&tr->tr_list);
 405         gfs2_assert_warn(sdp, list_empty(&tr->tr_ail1_list));
 406         gfs2_assert_warn(sdp, list_empty(&tr->tr_ail2_list));
 407         gfs2_trans_free(sdp, tr);
 408 }
 409
 410 static void ail2_empty(struct gfs2_sbd *sdp, unsigned int new_tail)
 411 {
 412         struct list_head *ail2_list = &sdp->sd_ail2_list;
 413         unsigned int old_tail = sdp->sd_log_tail;
 414         struct gfs2_trans *tr, *safe;
 415
 416         spin_lock(&sdp->sd_ail_lock);
 417         if (old_tail <= new_tail) {
 418                 list_for_each_entry_safe(tr, safe, ail2_list, tr_list) {
 419                         if (old_tail <= tr->tr_first && tr->tr_first < new_tail)
 420                                 __ail2_empty(sdp, tr);
 421                 }
 422         } else {
 423                 list_for_each_entry_safe(tr, safe, ail2_list, tr_list) {
 424                         if (old_tail <= tr->tr_first || tr->tr_first < new_tail)
 425                                 __ail2_empty(sdp, tr);
 426                 }
 427         }
 428         spin_unlock(&sdp->sd_ail_lock);
 429 }
 430
 431 /**
 432  * gfs2_log_is_empty - Check if the log is empty
 433  * @sdp: The GFS2 superblock
 434  */
 435
 436 bool gfs2_log_is_empty(struct gfs2_sbd *sdp) {
 437         return atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks;
 438 }
 439
 440 static bool __gfs2_log_try_reserve_revokes(struct gfs2_sbd *sdp, unsigned int revokes)
 441 {
 442         unsigned int available;
 443
 444         available = atomic_read(&sdp->sd_log_revokes_available);
 445         while (available >= revokes) {
 446                 if (atomic_try_cmpxchg(&sdp->sd_log_revokes_available,
 447                                        &available, available - revokes))
 448                         return true;
 449         }
 450         return false;
 451 }
 452
 453 /**
 454  * gfs2_log_release_revokes - Release a given number of revokes
 455  * @sdp: The GFS2 superblock
 456  * @revokes: The number of revokes to release
 457  *
 458  * sdp->sd_log_flush_lock must be held.
 459  */
 460 void gfs2_log_release_revokes(struct gfs2_sbd *sdp, unsigned int revokes)
 461 {
 462         if (revokes)
 463                 atomic_add(revokes, &sdp->sd_log_revokes_available);
 464 }
 465
 466 /**
 467  * gfs2_log_release - Release a given number of log blocks
 468  * @sdp: The GFS2 superblock
 469  * @blks: The number of blocks
 470  *
 471  */
 472
 473 void gfs2_log_release(struct gfs2_sbd *sdp, unsigned int blks)
 474 {
 475         atomic_add(blks, &sdp->sd_log_blks_free);
 476         trace_gfs2_log_blocks(sdp, blks);
 477         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 478                                   sdp->sd_jdesc->jd_blocks);
 479         if (atomic_read(&sdp->sd_log_blks_needed))
 480                 wake_up(&sdp->sd_log_waitq);
 481 }
 482
 483 /**
 484  * __gfs2_log_try_reserve - Try to make a log reservation
 485  * @sdp: The GFS2 superblock
 486  * @blks: The number of blocks to reserve
 487  * @taboo_blks: The number of blocks to leave free
 488  *
 489  * Try to do the same as __gfs2_log_reserve(), but fail if no more log
 490  * space is immediately available.
 491  */
 492 static bool __gfs2_log_try_reserve(struct gfs2_sbd *sdp, unsigned int blks,
 493                                    unsigned int taboo_blks)
 494 {
 495         unsigned wanted = blks + taboo_blks;
 496         unsigned int free_blocks;
 497
 498         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 499         while (free_blocks >= wanted) {
 500                 if (atomic_try_cmpxchg(&sdp->sd_log_blks_free, &free_blocks,
 501                                        free_blocks - blks)) {
 502                         trace_gfs2_log_blocks(sdp, -blks);
 503                         return true;
 504                 }
 505         }
 506         return false;
 507 }
 508
 509 /**
 510  * __gfs2_log_reserve - Make a log reservation
 511  * @sdp: The GFS2 superblock
 512  * @blks: The number of blocks to reserve
 513  * @taboo_blks: The number of blocks to leave free
 514  *
 515  * @taboo_blks is set to 0 for logd, and to GFS2_LOG_FLUSH_MIN_BLOCKS
 516  * for all other processes.  This ensures that when the log is almost full,
 517  * logd will still be able to call gfs2_log_flush one more time  without
 518  * blocking, which will advance the tail and make some more log space
 519  * available.
 520  *
 521  * We no longer flush the log here, instead we wake up logd to do that
 522  * for us. To avoid the thundering herd and to ensure that we deal fairly
 523  * with queued waiters, we use an exclusive wait. This means that when we
 524  * get woken with enough journal space to get our reservation, we need to
 525  * wake the next waiter on the list.
 526  */
 527
 528 static void __gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks,
 529                                unsigned int taboo_blks)
 530 {
 531         unsigned wanted = blks + taboo_blks;
 532         unsigned int free_blocks;
 533
 534         atomic_add(blks, &sdp->sd_log_blks_needed);
 535         for (;;) {
 536                 if (current != sdp->sd_logd_process)
 537                         wake_up(&sdp->sd_logd_waitq);
 538                 io_wait_event(sdp->sd_log_waitq,
 539                         (free_blocks = atomic_read(&sdp->sd_log_blks_free),
 540                          free_blocks >= wanted));
 541                 do {
 542                         if (atomic_try_cmpxchg(&sdp->sd_log_blks_free,
 543                                                &free_blocks,
 544                                                free_blocks - blks))
 545                                 goto reserved;
 546                 } while (free_blocks >= wanted);
 547         }
 548
 549 reserved:
 550         trace_gfs2_log_blocks(sdp, -blks);
 551         if (atomic_sub_return(blks, &sdp->sd_log_blks_needed))
 552                 wake_up(&sdp->sd_log_waitq);
 553 }
 554
 555 /**
 556  * gfs2_log_try_reserve - Try to make a log reservation
 557  * @sdp: The GFS2 superblock
 558  * @tr: The transaction
 559  * @extra_revokes: The number of additional revokes reserved (output)
 560  *
 561  * This is similar to gfs2_log_reserve, but sdp->sd_log_flush_lock must be
 562  * held for correct revoke accounting.
 563  */
 564
 565 bool gfs2_log_try_reserve(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
 566                           unsigned int *extra_revokes)
 567 {
 568         unsigned int blks = tr->tr_reserved;
 569         unsigned int revokes = tr->tr_revokes;
 570         unsigned int revoke_blks = 0;
 571
 572         *extra_revokes = 0;
 573         if (revokes && !__gfs2_log_try_reserve_revokes(sdp, revokes)) {
 574                 revoke_blks = DIV_ROUND_UP(revokes, sdp->sd_inptrs);
 575                 *extra_revokes = revoke_blks * sdp->sd_inptrs - revokes;
 576                 blks += revoke_blks;
 577         }
 578         if (!blks)
 579                 return true;
 580         if (__gfs2_log_try_reserve(sdp, blks, GFS2_LOG_FLUSH_MIN_BLOCKS))
 581                 return true;
 582         if (!revoke_blks)
 583                 gfs2_log_release_revokes(sdp, revokes);
 584         return false;
 585 }
 586
 587 /**
 588  * gfs2_log_reserve - Make a log reservation
 589  * @sdp: The GFS2 superblock
 590  * @tr: The transaction
 591  * @extra_revokes: The number of additional revokes reserved (output)
 592  *
 593  * sdp->sd_log_flush_lock must not be held.
 594  */
 595
 596 void gfs2_log_reserve(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
 597                       unsigned int *extra_revokes)
 598 {
 599         unsigned int blks = tr->tr_reserved;
 600         unsigned int revokes = tr->tr_revokes;
 601         unsigned int revoke_blks;
 602
 603         *extra_revokes = 0;
 604         if (revokes) {
 605                 revoke_blks = DIV_ROUND_UP(revokes, sdp->sd_inptrs);
 606                 *extra_revokes = revoke_blks * sdp->sd_inptrs - revokes;
 607                 blks += revoke_blks;
 608         }
 609         __gfs2_log_reserve(sdp, blks, GFS2_LOG_FLUSH_MIN_BLOCKS);
 610 }
 611
 612 /**
 613  * log_distance - Compute distance between two journal blocks
 614  * @sdp: The GFS2 superblock
 615  * @newer: The most recent journal block of the pair
 616  * @older: The older journal block of the pair
 617  *
 618  *   Compute the distance (in the journal direction) between two
 619  *   blocks in the journal
 620  *
 621  * Returns: the distance in blocks
 622  */
 623
 624 static inline unsigned int log_distance(struct gfs2_sbd *sdp, unsigned int newer,
 625                                         unsigned int older)
 626 {
 627         int dist;
 628
 629         dist = newer - older;
 630         if (dist < 0)
 631                 dist += sdp->sd_jdesc->jd_blocks;
 632
 633         return dist;
 634 }
 635
 636 /**
 637  * calc_reserved - Calculate the number of blocks to keep reserved
 638  * @sdp: The GFS2 superblock
 639  *
 640  * This is complex.  We need to reserve room for all our currently used
 641  * metadata blocks (e.g. normal file I/O rewriting file time stamps) and
 642  * all our journaled data blocks for journaled files (e.g. files in the
 643  * meta_fs like rindex, or files for which chattr +j was done.)
 644  * If we don't reserve enough space, corruption will follow.
 645  *
 646  * We can have metadata blocks and jdata blocks in the same journal.  Each
 647  * type gets its own log descriptor, for which we need to reserve a block.
 648  * In fact, each type has the potential for needing more than one log descriptor
 649  * in cases where we have more blocks than will fit in a log descriptor.
 650  * Metadata journal entries take up half the space of journaled buffer entries.
 651  *
 652  * Also, we need to reserve blocks for revoke journal entries and one for an
 653  * overall header for the lot.
 654  *
 655  * Returns: the number of blocks reserved
 656  */
 657 static unsigned int calc_reserved(struct gfs2_sbd *sdp)
 658 {
 659         unsigned int reserved = GFS2_LOG_FLUSH_MIN_BLOCKS;
 660         unsigned int blocks;
 661         struct gfs2_trans *tr = sdp->sd_log_tr;
 662
 663         if (tr) {
 664                 blocks = tr->tr_num_buf_new - tr->tr_num_buf_rm;
 665                 reserved += blocks + DIV_ROUND_UP(blocks, buf_limit(sdp));
 666                 blocks = tr->tr_num_databuf_new - tr->tr_num_databuf_rm;
 667                 reserved += blocks + DIV_ROUND_UP(blocks, databuf_limit(sdp));
 668         }
 669         return reserved;
 670 }
 671
 672 static void log_pull_tail(struct gfs2_sbd *sdp)
 673 {
 674         unsigned int new_tail = sdp->sd_log_flush_tail;
 675         unsigned int dist;
 676
 677         if (new_tail == sdp->sd_log_tail)
 678                 return;
 679         dist = log_distance(sdp, new_tail, sdp->sd_log_tail);
 680         ail2_empty(sdp, new_tail);
 681         gfs2_log_release(sdp, dist);
 682         sdp->sd_log_tail = new_tail;
 683 }
 684
 685
 686 void log_flush_wait(struct gfs2_sbd *sdp)
 687 {
 688         DEFINE_WAIT(wait);
 689
 690         if (atomic_read(&sdp->sd_log_in_flight)) {
 691                 do {
 692                         prepare_to_wait(&sdp->sd_log_flush_wait, &wait,
 693                                         TASK_UNINTERRUPTIBLE);
 694                         if (atomic_read(&sdp->sd_log_in_flight))
 695                                 io_schedule();
 696                 } while(atomic_read(&sdp->sd_log_in_flight));
 697                 finish_wait(&sdp->sd_log_flush_wait, &wait);
 698         }
 699 }
 700
 701 static int ip_cmp(void *priv, const struct list_head *a, const struct list_head *b)
 702 {
 703         struct gfs2_inode *ipa, *ipb;
 704
 705         ipa = list_entry(a, struct gfs2_inode, i_ordered);
 706         ipb = list_entry(b, struct gfs2_inode, i_ordered);
 707
 708         if (ipa->i_no_addr < ipb->i_no_addr)
 709                 return -1;
 710         if (ipa->i_no_addr > ipb->i_no_addr)
 711                 return 1;
 712         return 0;
 713 }
 714
 715 static void __ordered_del_inode(struct gfs2_inode *ip)
 716 {
 717         if (!list_empty(&ip->i_ordered))
 718                 list_del_init(&ip->i_ordered);
 719 }
 720
 721 static void gfs2_ordered_write(struct gfs2_sbd *sdp)
 722 {
 723         struct gfs2_inode *ip;
 724         LIST_HEAD(written);
 725
 726         spin_lock(&sdp->sd_ordered_lock);
 727         list_sort(NULL, &sdp->sd_log_ordered, &ip_cmp);
 728         while (!list_empty(&sdp->sd_log_ordered)) {
 729                 ip = list_first_entry(&sdp->sd_log_ordered, struct gfs2_inode, i_ordered);
 730                 if (ip->i_inode.i_mapping->nrpages == 0) {
 731                         __ordered_del_inode(ip);
 732                         continue;
 733                 }
 734                 list_move(&ip->i_ordered, &written);
 735                 spin_unlock(&sdp->sd_ordered_lock);
 736                 filemap_fdatawrite(ip->i_inode.i_mapping);
 737                 spin_lock(&sdp->sd_ordered_lock);
 738         }
 739         list_splice(&written, &sdp->sd_log_ordered);
 740         spin_unlock(&sdp->sd_ordered_lock);
 741 }
 742
 743 static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
 744 {
 745         struct gfs2_inode *ip;
 746
 747         spin_lock(&sdp->sd_ordered_lock);
 748         while (!list_empty(&sdp->sd_log_ordered)) {
 749                 ip = list_first_entry(&sdp->sd_log_ordered, struct gfs2_inode, i_ordered);
 750                 __ordered_del_inode(ip);
 751                 if (ip->i_inode.i_mapping->nrpages == 0)
 752                         continue;
 753                 spin_unlock(&sdp->sd_ordered_lock);
 754                 filemap_fdatawait(ip->i_inode.i_mapping);
 755                 spin_lock(&sdp->sd_ordered_lock);
 756         }
 757         spin_unlock(&sdp->sd_ordered_lock);
 758 }
 759
 760 void gfs2_ordered_del_inode(struct gfs2_inode *ip)
 761 {
 762         struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
 763
 764         spin_lock(&sdp->sd_ordered_lock);
 765         __ordered_del_inode(ip);
 766         spin_unlock(&sdp->sd_ordered_lock);
 767 }
 768
 769 void gfs2_add_revoke(struct gfs2_sbd *sdp, struct gfs2_bufdata *bd)
 770 {
 771         struct buffer_head *bh = bd->bd_bh;
 772         struct gfs2_glock *gl = bd->bd_gl;
 773
 774         sdp->sd_log_num_revoke++;
 775         if (atomic_inc_return(&gl->gl_revokes) == 1)
 776                 gfs2_glock_hold(gl);
 777         bh->b_private = NULL;
 778         bd->bd_blkno = bh->b_blocknr;
 779         gfs2_remove_from_ail(bd); /* drops ref on bh */
 780         bd->bd_bh = NULL;
 781         set_bit(GLF_LFLUSH, &gl->gl_flags);
 782         list_add(&bd->bd_list, &sdp->sd_log_revokes);
 783 }
 784
 785 void gfs2_glock_remove_revoke(struct gfs2_glock *gl)
 786 {
 787         if (atomic_dec_return(&gl->gl_revokes) == 0) {
 788                 clear_bit(GLF_LFLUSH, &gl->gl_flags);
 789                 gfs2_glock_put_async(gl);
 790         }
 791 }
 792
 793 /**
 794  * gfs2_flush_revokes - Add as many revokes to the system transaction as we can
 795  * @sdp: The GFS2 superblock
 796  *
 797  * Our usual strategy is to defer writing revokes as much as we can in the hope
 798  * that we'll eventually overwrite the journal, which will make those revokes
 799  * go away.  This changes when we flush the log: at that point, there will
 800  * likely be some left-over space in the last revoke block of that transaction.
 801  * We can fill that space with additional revokes for blocks that have already
 802  * been written back.  This will basically come at no cost now, and will save
 803  * us from having to keep track of those blocks on the AIL2 list later.
 804  */
 805 void gfs2_flush_revokes(struct gfs2_sbd *sdp)
 806 {
 807         /* number of revokes we still have room for */
 808         unsigned int max_revokes = atomic_read(&sdp->sd_log_revokes_available);
 809
 810         gfs2_log_lock(sdp);
 811         gfs2_ail1_empty(sdp, max_revokes);
 812         gfs2_log_unlock(sdp);
 813
 814         if (gfs2_withdrawing(sdp))
 815                 gfs2_withdraw(sdp);
 816 }
 817
 818 /**
 819  * gfs2_write_log_header - Write a journal log header buffer at lblock
 820  * @sdp: The GFS2 superblock
 821  * @jd: journal descriptor of the journal to which we are writing
 822  * @seq: sequence number
 823  * @tail: tail of the log
 824  * @lblock: value for lh_blkno (block number relative to start of journal)
 825  * @flags: log header flags GFS2_LOG_HEAD_*
 826  * @op_flags: flags to pass to the bio
 827  *
 828  * Returns: the initialized log buffer descriptor
 829  */
 830
 831 void gfs2_write_log_header(struct gfs2_sbd *sdp, struct gfs2_jdesc *jd,
 832                            u64 seq, u32 tail, u32 lblock, u32 flags,
 833                            blk_opf_t op_flags)
 834 {
 835         struct gfs2_log_header *lh;
 836         u32 hash, crc;
 837         struct page *page;
 838         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 839         struct timespec64 tv;
 840         struct super_block *sb = sdp->sd_vfs;
 841         u64 dblock;
 842
 843         if (gfs2_withdrawing_or_withdrawn(sdp))
 844                 return;
 845
 846         page = mempool_alloc(gfs2_page_pool, GFP_NOIO);
 847         lh = page_address(page);
 848         clear_page(lh);
 849
 850         lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
 851         lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
 852         lh->lh_header.__pad0 = cpu_to_be64(0);
 853         lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
 854         lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
 855         lh->lh_sequence = cpu_to_be64(seq);
 856         lh->lh_flags = cpu_to_be32(flags);
 857         lh->lh_tail = cpu_to_be32(tail);
 858         lh->lh_blkno = cpu_to_be32(lblock);
 859         hash = ~crc32(~0, lh, LH_V1_SIZE);
 860         lh->lh_hash = cpu_to_be32(hash);
 861
 862         ktime_get_coarse_real_ts64(&tv);
 863         lh->lh_nsec = cpu_to_be32(tv.tv_nsec);
 864         lh->lh_sec = cpu_to_be64(tv.tv_sec);
 865         if (!list_empty(&jd->extent_list))
 866                 dblock = gfs2_log_bmap(jd, lblock);
 867         else {
 868                 unsigned int extlen;
 869                 int ret;
 870
 871                 extlen = 1;
 872                 ret = gfs2_get_extent(jd->jd_inode, lblock, &dblock, &extlen);
 873                 if (gfs2_assert_withdraw(sdp, ret == 0))
 874                         return;
 875         }
 876         lh->lh_addr = cpu_to_be64(dblock);
 877         lh->lh_jinode = cpu_to_be64(GFS2_I(jd->jd_inode)->i_no_addr);
 878
 879         /* We may only write local statfs, quota, etc., when writing to our
 880            own journal. The values are left 0 when recovering a journal
 881            different from our own. */
 882         if (!(flags & GFS2_LOG_HEAD_RECOVERY)) {
 883                 lh->lh_statfs_addr =
 884                         cpu_to_be64(GFS2_I(sdp->sd_sc_inode)->i_no_addr);
 885                 lh->lh_quota_addr =
 886                         cpu_to_be64(GFS2_I(sdp->sd_qc_inode)->i_no_addr);
 887
 888                 spin_lock(&sdp->sd_statfs_spin);
 889                 lh->lh_local_total = cpu_to_be64(l_sc->sc_total);
 890                 lh->lh_local_free = cpu_to_be64(l_sc->sc_free);
 891                 lh->lh_local_dinodes = cpu_to_be64(l_sc->sc_dinodes);
 892                 spin_unlock(&sdp->sd_statfs_spin);
 893         }
 894
 895         BUILD_BUG_ON(offsetof(struct gfs2_log_header, lh_crc) != LH_V1_SIZE);
 896
 897         crc = crc32c(~0, (void *)lh + LH_V1_SIZE + 4,
 898                      sb->s_blocksize - LH_V1_SIZE - 4);
 899         lh->lh_crc = cpu_to_be32(crc);
 900
 901         gfs2_log_write(sdp, jd, page, sb->s_blocksize, 0, dblock);
 902         gfs2_log_submit_bio(&jd->jd_log_bio, REQ_OP_WRITE | op_flags);
 903 }
 904
 905 /**
 906  * log_write_header - Get and initialize a journal header buffer
 907  * @sdp: The GFS2 superblock
 908  * @flags: The log header flags, including log header origin
 909  *
 910  * Returns: the initialized log buffer descriptor
 911  */
 912
 913 static void log_write_header(struct gfs2_sbd *sdp, u32 flags)
 914 {
 915         blk_opf_t op_flags = REQ_PREFLUSH | REQ_FUA | REQ_META | REQ_SYNC;
 916         struct super_block *sb = sdp->sd_vfs;
 917
 918         gfs2_assert_withdraw(sdp, sb->s_writers.frozen != SB_FREEZE_COMPLETE);
 919
 920         if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags)) {
 921                 gfs2_ordered_wait(sdp);
 922                 log_flush_wait(sdp);
 923                 op_flags = REQ_SYNC | REQ_META | REQ_PRIO;
 924         }
 925         sdp->sd_log_idle = (sdp->sd_log_flush_tail == sdp->sd_log_flush_head);
 926         gfs2_write_log_header(sdp, sdp->sd_jdesc, sdp->sd_log_sequence++,
 927                               sdp->sd_log_flush_tail, sdp->sd_log_flush_head,
 928                               flags, op_flags);
 929         gfs2_log_incr_head(sdp);
 930         log_flush_wait(sdp);
 931         log_pull_tail(sdp);
 932         gfs2_log_update_head(sdp);
 933 }
 934
 935 /**
 936  * gfs2_ail_drain - drain the ail lists after a withdraw
 937  * @sdp: Pointer to GFS2 superblock
 938  */
 939 void gfs2_ail_drain(struct gfs2_sbd *sdp)
 940 {
 941         struct gfs2_trans *tr;
 942
 943         spin_lock(&sdp->sd_ail_lock);
 944         /*
 945          * For transactions on the sd_ail1_list we need to drain both the
 946          * ail1 and ail2 lists. That's because function gfs2_ail1_start_one
 947          * (temporarily) moves items from its tr_ail1 list to tr_ail2 list
 948          * before revokes are sent for that block. Items on the sd_ail2_list
 949          * should have already gotten beyond that point, so no need.
 950          */
 951         while (!list_empty(&sdp->sd_ail1_list)) {
 952                 tr = list_first_entry(&sdp->sd_ail1_list, struct gfs2_trans,
 953                                       tr_list);
 954                 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail1_list);
 955                 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
 956                 list_del(&tr->tr_list);
 957                 gfs2_trans_free(sdp, tr);
 958         }
 959         while (!list_empty(&sdp->sd_ail2_list)) {
 960                 tr = list_first_entry(&sdp->sd_ail2_list, struct gfs2_trans,
 961                                       tr_list);
 962                 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
 963                 list_del(&tr->tr_list);
 964                 gfs2_trans_free(sdp, tr);
 965         }
 966         gfs2_drain_revokes(sdp);
 967         spin_unlock(&sdp->sd_ail_lock);
 968 }
 969
 970 /**
 971  * empty_ail1_list - try to start IO and empty the ail1 list
 972  * @sdp: Pointer to GFS2 superblock
 973  */
 974 static void empty_ail1_list(struct gfs2_sbd *sdp)
 975 {
 976         unsigned long start = jiffies;
 977         bool empty = false;
 978
 979         while (!empty) {
 980                 if (time_after(jiffies, start + (HZ * 600))) {
 981                         fs_err(sdp, "Error: In %s for 10 minutes! t=%d\n",
 982                                __func__, current->journal_info ? 1 : 0);
 983                         dump_ail_list(sdp);
 984                         return;
 985                 }
 986                 gfs2_ail1_start(sdp);
 987                 gfs2_ail1_wait(sdp);
 988                 empty = gfs2_ail1_empty(sdp, 0);
 989
 990                 if (gfs2_withdrawing_or_withdrawn(sdp))
 991                         break;
 992         }
 993
 994         if (gfs2_withdrawing(sdp))
 995                 gfs2_withdraw(sdp);
 996 }
 997
 998 /**
 999  * trans_drain - drain the buf and databuf queue for a failed transaction
1000  * @tr: the transaction to drain
1001  *
1002  * When this is called, we're taking an error exit for a log write that failed
1003  * but since we bypassed the after_commit functions, we need to remove the
1004  * items from the buf and databuf queue.
1005  */
1006 static void trans_drain(struct gfs2_trans *tr)
1007 {
1008         struct gfs2_bufdata *bd;
1009         struct list_head *head;
1010
1011         if (!tr)
1012                 return;
1013
1014         head = &tr->tr_buf;
1015         while (!list_empty(head)) {
1016                 bd = list_first_entry(head, struct gfs2_bufdata, bd_list);
1017                 list_del_init(&bd->bd_list);
1018                 if (!list_empty(&bd->bd_ail_st_list))
1019                         gfs2_remove_from_ail(bd);
1020                 kmem_cache_free(gfs2_bufdata_cachep, bd);
1021         }
1022         head = &tr->tr_databuf;
1023         while (!list_empty(head)) {
1024                 bd = list_first_entry(head, struct gfs2_bufdata, bd_list);
1025                 list_del_init(&bd->bd_list);
1026                 if (!list_empty(&bd->bd_ail_st_list))
1027                         gfs2_remove_from_ail(bd);
1028                 kmem_cache_free(gfs2_bufdata_cachep, bd);
1029         }
1030 }
1031
1032 /**
1033  * gfs2_log_flush - flush incore transaction(s)
1034  * @sdp: The filesystem
1035  * @gl: The glock structure to flush.  If NULL, flush the whole incore log
1036  * @flags: The log header flags: GFS2_LOG_HEAD_FLUSH_* and debug flags
1037  *
1038  */
1039
1040 void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl, u32 flags)
1041 {
1042         struct gfs2_trans *tr = NULL;
1043         unsigned int reserved_blocks = 0, used_blocks = 0;
1044         bool frozen = test_bit(SDF_FROZEN, &sdp->sd_flags);
1045         unsigned int first_log_head;
1046         unsigned int reserved_revokes = 0;
1047
1048         down_write(&sdp->sd_log_flush_lock);
1049         trace_gfs2_log_flush(sdp, 1, flags);
1050
1051 repeat:
1052         /*
1053          * Do this check while holding the log_flush_lock to prevent new
1054          * buffers from being added to the ail via gfs2_pin()
1055          */
1056         if (gfs2_withdrawing_or_withdrawn(sdp) ||
1057             !test_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags))
1058                 goto out;
1059
1060         /* Log might have been flushed while we waited for the flush lock */
1061         if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags))
1062                 goto out;
1063
1064         first_log_head = sdp->sd_log_head;
1065         sdp->sd_log_flush_head = first_log_head;
1066
1067         tr = sdp->sd_log_tr;
1068         if (tr || sdp->sd_log_num_revoke) {
1069                 if (reserved_blocks)
1070                         gfs2_log_release(sdp, reserved_blocks);
1071                 reserved_blocks = sdp->sd_log_blks_reserved;
1072                 reserved_revokes = sdp->sd_log_num_revoke;
1073                 if (tr) {
1074                         sdp->sd_log_tr = NULL;
1075                         tr->tr_first = first_log_head;
1076                         if (unlikely(frozen)) {
1077                                 if (gfs2_assert_withdraw_delayed(sdp,
1078                                        !tr->tr_num_buf_new && !tr->tr_num_databuf_new))
1079                                         goto out_withdraw;
1080                         }
1081                 }
1082         } else if (!reserved_blocks) {
1083                 unsigned int taboo_blocks = GFS2_LOG_FLUSH_MIN_BLOCKS;
1084
1085                 reserved_blocks = GFS2_LOG_FLUSH_MIN_BLOCKS;
1086                 if (current == sdp->sd_logd_process)
1087                         taboo_blocks = 0;
1088
1089                 if (!__gfs2_log_try_reserve(sdp, reserved_blocks, taboo_blocks)) {
1090                         up_write(&sdp->sd_log_flush_lock);
1091                         __gfs2_log_reserve(sdp, reserved_blocks, taboo_blocks);
1092                         down_write(&sdp->sd_log_flush_lock);
1093                         goto repeat;
1094                 }
1095                 BUG_ON(sdp->sd_log_num_revoke);
1096         }
1097
1098         if (flags & GFS2_LOG_HEAD_FLUSH_SHUTDOWN)
1099                 clear_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags);
1100
1101         if (unlikely(frozen))
1102                 if (gfs2_assert_withdraw_delayed(sdp, !reserved_revokes))
1103                         goto out_withdraw;
1104
1105         gfs2_ordered_write(sdp);
1106         if (gfs2_withdrawing_or_withdrawn(sdp))
1107                 goto out_withdraw;
1108         lops_before_commit(sdp, tr);
1109         if (gfs2_withdrawing_or_withdrawn(sdp))
1110                 goto out_withdraw;
1111         if (sdp->sd_jdesc)
1112                 gfs2_log_submit_bio(&sdp->sd_jdesc->jd_log_bio, REQ_OP_WRITE);
1113         if (gfs2_withdrawing_or_withdrawn(sdp))
1114                 goto out_withdraw;
1115
1116         if (sdp->sd_log_head != sdp->sd_log_flush_head) {
1117                 log_write_header(sdp, flags);
1118         } else if (sdp->sd_log_tail != sdp->sd_log_flush_tail && !sdp->sd_log_idle) {
1119                 log_write_header(sdp, flags);
1120         }
1121         if (gfs2_withdrawing_or_withdrawn(sdp))
1122                 goto out_withdraw;
1123         lops_after_commit(sdp, tr);
1124
1125         gfs2_log_lock(sdp);
1126         sdp->sd_log_blks_reserved = 0;
1127
1128         spin_lock(&sdp->sd_ail_lock);
1129         if (tr && !list_empty(&tr->tr_ail1_list)) {
1130                 list_add(&tr->tr_list, &sdp->sd_ail1_list);
1131                 tr = NULL;
1132         }
1133         spin_unlock(&sdp->sd_ail_lock);
1134         gfs2_log_unlock(sdp);
1135
1136         if (!(flags & GFS2_LOG_HEAD_FLUSH_NORMAL)) {
1137                 if (!sdp->sd_log_idle) {
1138                         empty_ail1_list(sdp);
1139                         if (gfs2_withdrawing_or_withdrawn(sdp))
1140                                 goto out_withdraw;
1141                         log_write_header(sdp, flags);
1142                 }
1143                 if (flags & (GFS2_LOG_HEAD_FLUSH_SHUTDOWN |
1144                              GFS2_LOG_HEAD_FLUSH_FREEZE))
1145                         gfs2_log_shutdown(sdp);
1146         }
1147
1148 out_end:
1149         used_blocks = log_distance(sdp, sdp->sd_log_flush_head, first_log_head);
1150         reserved_revokes += atomic_read(&sdp->sd_log_revokes_available);
1151         atomic_set(&sdp->sd_log_revokes_available, sdp->sd_ldptrs);
1152         gfs2_assert_withdraw(sdp, reserved_revokes % sdp->sd_inptrs == sdp->sd_ldptrs);
1153         if (reserved_revokes > sdp->sd_ldptrs)
1154                 reserved_blocks += (reserved_revokes - sdp->sd_ldptrs) / sdp->sd_inptrs;
1155 out:
1156         if (used_blocks != reserved_blocks) {
1157                 gfs2_assert_withdraw_delayed(sdp, used_blocks < reserved_blocks);
1158                 gfs2_log_release(sdp, reserved_blocks - used_blocks);
1159         }
1160         up_write(&sdp->sd_log_flush_lock);
1161         gfs2_trans_free(sdp, tr);
1162         if (gfs2_withdrawing(sdp))
1163                 gfs2_withdraw(sdp);
1164         trace_gfs2_log_flush(sdp, 0, flags);
1165         return;
1166
1167 out_withdraw:
1168         trans_drain(tr);
1169         /**
1170          * If the tr_list is empty, we're withdrawing during a log
1171          * flush that targets a transaction, but the transaction was
1172          * never queued onto any of the ail lists. Here we add it to
1173          * ail1 just so that ail_drain() will find and free it.
1174          */
1175         spin_lock(&sdp->sd_ail_lock);
1176         if (tr && list_empty(&tr->tr_list))
1177                 list_add(&tr->tr_list, &sdp->sd_ail1_list);
1178         spin_unlock(&sdp->sd_ail_lock);
1179         tr = NULL;
1180         goto out_end;
1181 }
1182
1183 /**
1184  * gfs2_merge_trans - Merge a new transaction into a cached transaction
1185  * @sdp: the filesystem
1186  * @new: New transaction to be merged
1187  */
1188
1189 static void gfs2_merge_trans(struct gfs2_sbd *sdp, struct gfs2_trans *new)
1190 {
1191         struct gfs2_trans *old = sdp->sd_log_tr;
1192
1193         WARN_ON_ONCE(!test_bit(TR_ATTACHED, &old->tr_flags));
1194
1195         old->tr_num_buf_new     += new->tr_num_buf_new;
1196         old->tr_num_databuf_new += new->tr_num_databuf_new;
1197         old->tr_num_buf_rm      += new->tr_num_buf_rm;
1198         old->tr_num_databuf_rm  += new->tr_num_databuf_rm;
1199         old->tr_revokes         += new->tr_revokes;
1200         old->tr_num_revoke      += new->tr_num_revoke;
1201
1202         list_splice_tail_init(&new->tr_databuf, &old->tr_databuf);
1203         list_splice_tail_init(&new->tr_buf, &old->tr_buf);
1204
1205         spin_lock(&sdp->sd_ail_lock);
1206         list_splice_tail_init(&new->tr_ail1_list, &old->tr_ail1_list);
1207         list_splice_tail_init(&new->tr_ail2_list, &old->tr_ail2_list);
1208         spin_unlock(&sdp->sd_ail_lock);
1209 }
1210
1211 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
1212 {
1213         unsigned int reserved;
1214         unsigned int unused;
1215         unsigned int maxres;
1216
1217         gfs2_log_lock(sdp);
1218
1219         if (sdp->sd_log_tr) {
1220                 gfs2_merge_trans(sdp, tr);
1221         } else if (tr->tr_num_buf_new || tr->tr_num_databuf_new) {
1222                 gfs2_assert_withdraw(sdp, !test_bit(TR_ONSTACK, &tr->tr_flags));
1223                 sdp->sd_log_tr = tr;
1224                 set_bit(TR_ATTACHED, &tr->tr_flags);
1225         }
1226
1227         reserved = calc_reserved(sdp);
1228         maxres = sdp->sd_log_blks_reserved + tr->tr_reserved;
1229         gfs2_assert_withdraw(sdp, maxres >= reserved);
1230         unused = maxres - reserved;
1231         if (unused)
1232                 gfs2_log_release(sdp, unused);
1233         sdp->sd_log_blks_reserved = reserved;
1234
1235         gfs2_log_unlock(sdp);
1236 }
1237
1238 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
1239 {
1240         return atomic_read(&sdp->sd_log_pinned) +
1241                atomic_read(&sdp->sd_log_blks_needed) >=
1242                atomic_read(&sdp->sd_log_thresh1);
1243 }
1244
1245 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
1246 {
1247         return sdp->sd_jdesc->jd_blocks -
1248                atomic_read(&sdp->sd_log_blks_free) +
1249                atomic_read(&sdp->sd_log_blks_needed) >=
1250                atomic_read(&sdp->sd_log_thresh2);
1251 }
1252
1253 /**
1254  * gfs2_log_commit - Commit a transaction to the log
1255  * @sdp: the filesystem
1256  * @tr: the transaction
1257  *
1258  * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
1259  * or the total number of used blocks (pinned blocks plus AIL blocks)
1260  * is greater than thresh2.
1261  *
1262  * At mount time thresh1 is 2/5ths of journal size, thresh2 is 4/5ths of
1263  * journal size.
1264  *
1265  * Returns: errno
1266  */
1267
1268 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
1269 {
1270         log_refund(sdp, tr);
1271
1272         if (gfs2_ail_flush_reqd(sdp) || gfs2_jrnl_flush_reqd(sdp))
1273                 wake_up(&sdp->sd_logd_waitq);
1274 }
1275
1276 /**
1277  * gfs2_log_shutdown - write a shutdown header into a journal
1278  * @sdp: the filesystem
1279  *
1280  */
1281
1282 static void gfs2_log_shutdown(struct gfs2_sbd *sdp)
1283 {
1284         gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
1285         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
1286         gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
1287
1288         log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT | GFS2_LFC_SHUTDOWN);
1289         log_pull_tail(sdp);
1290
1291         gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
1292         gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
1293 }
1294
1295 /**
1296  * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
1297  * @data: Pointer to GFS2 superblock
1298  *
1299  * Also, periodically check to make sure that we're using the most recent
1300  * journal index.
1301  */
1302
1303 int gfs2_logd(void *data)
1304 {
1305         struct gfs2_sbd *sdp = data;
1306         unsigned long t = 1;
1307
1308         set_freezable();
1309         while (!kthread_should_stop()) {
1310                 if (gfs2_withdrawing_or_withdrawn(sdp))
1311                         break;
1312
1313                 /* Check for errors writing to the journal */
1314                 if (sdp->sd_log_error) {
1315                         gfs2_lm(sdp,
1316                                 "GFS2: fsid=%s: error %d: "
1317                                 "withdrawing the file system to "
1318                                 "prevent further damage.\n",
1319                                 sdp->sd_fsname, sdp->sd_log_error);
1320                         gfs2_withdraw(sdp);
1321                         break;
1322                 }
1323
1324                 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
1325                         gfs2_ail1_empty(sdp, 0);
1326                         gfs2_log_flush(sdp, NULL, GFS2_LOG_HEAD_FLUSH_NORMAL |
1327                                                   GFS2_LFC_LOGD_JFLUSH_REQD);
1328                 }
1329
1330                 if (test_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags) ||
1331                     gfs2_ail_flush_reqd(sdp)) {
1332                         clear_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
1333                         gfs2_ail1_start(sdp);
1334                         gfs2_ail1_wait(sdp);
1335                         gfs2_ail1_empty(sdp, 0);
1336                         gfs2_log_flush(sdp, NULL, GFS2_LOG_HEAD_FLUSH_NORMAL |
1337                                                   GFS2_LFC_LOGD_AIL_FLUSH_REQD);
1338                 }
1339
1340                 t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
1341
1342                 t = wait_event_freezable_timeout(sdp->sd_logd_waitq,
1343                                 test_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags) ||
1344                                 gfs2_ail_flush_reqd(sdp) ||
1345                                 gfs2_jrnl_flush_reqd(sdp) ||
1346                                 sdp->sd_log_error ||
1347                                 gfs2_withdrawing_or_withdrawn(sdp) ||
1348                                 kthread_should_stop(),
1349                                 t);
1350         }
1351
1352         if (gfs2_withdrawing(sdp))
1353                 gfs2_withdraw(sdp);
1354
1355         return 0;
1356 }
1357