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  * @mapping: The associated mapping (maybe NULL)
  69  * @bd: The gfs2_bufdata to remove
  70  *
  71  * The ail lock _must_ be held when calling this function
  72  *
  73  */
  74
  75 void gfs2_remove_from_ail(struct gfs2_bufdata *bd)
  76 {
  77         bd->bd_tr = NULL;
  78         list_del_init(&bd->bd_ail_st_list);
  79         list_del_init(&bd->bd_ail_gl_list);
  80         atomic_dec(&bd->bd_gl->gl_ail_count);
  81         brelse(bd->bd_bh);
  82 }
  83
  84 /**
  85  * gfs2_ail1_start_one - Start I/O on a part of the AIL
  86  * @sdp: the filesystem
  87  * @wbc: The writeback control structure
  88  * @ai: The ail structure
  89  *
  90  */
  91
  92 static int gfs2_ail1_start_one(struct gfs2_sbd *sdp,
  93                                struct writeback_control *wbc,
  94                                struct gfs2_trans *tr, struct blk_plug *plug)
  95 __releases(&sdp->sd_ail_lock)
  96 __acquires(&sdp->sd_ail_lock)
  97 {
  98         struct gfs2_glock *gl = NULL;
  99         struct address_space *mapping;
 100         struct gfs2_bufdata *bd, *s;
 101         struct buffer_head *bh;
 102         int ret = 0;
 103
 104         list_for_each_entry_safe_reverse(bd, s, &tr->tr_ail1_list, bd_ail_st_list) {
 105                 bh = bd->bd_bh;
 106
 107                 gfs2_assert(sdp, bd->bd_tr == tr);
 108
 109                 if (!buffer_busy(bh)) {
 110                         if (buffer_uptodate(bh)) {
 111                                 list_move(&bd->bd_ail_st_list,
 112                                           &tr->tr_ail2_list);
 113                                 continue;
 114                         }
 115                         if (!cmpxchg(&sdp->sd_log_error, 0, -EIO)) {
 116                                 gfs2_io_error_bh(sdp, bh);
 117                                 gfs2_withdraw_delayed(sdp);
 118                         }
 119                 }
 120
 121                 if (gfs2_withdrawn(sdp)) {
 122                         gfs2_remove_from_ail(bd);
 123                         continue;
 124                 }
 125                 if (!buffer_dirty(bh))
 126                         continue;
 127                 if (gl == bd->bd_gl)
 128                         continue;
 129                 gl = bd->bd_gl;
 130                 list_move(&bd->bd_ail_st_list, &tr->tr_ail1_list);
 131                 mapping = bh->b_page->mapping;
 132                 if (!mapping)
 133                         continue;
 134                 spin_unlock(&sdp->sd_ail_lock);
 135                 ret = generic_writepages(mapping, wbc);
 136                 if (need_resched()) {
 137                         blk_finish_plug(plug);
 138                         cond_resched();
 139                         blk_start_plug(plug);
 140                 }
 141                 spin_lock(&sdp->sd_ail_lock);
 142                 if (ret == -ENODATA) /* if a jdata write into a new hole */
 143                         ret = 0; /* ignore it */
 144                 if (ret || wbc->nr_to_write <= 0)
 145                         break;
 146                 return -EBUSY;
 147         }
 148
 149         return ret;
 150 }
 151
 152 static void dump_ail_list(struct gfs2_sbd *sdp)
 153 {
 154         struct gfs2_trans *tr;
 155         struct gfs2_bufdata *bd;
 156         struct buffer_head *bh;
 157
 158         list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
 159                 list_for_each_entry_reverse(bd, &tr->tr_ail1_list,
 160                                             bd_ail_st_list) {
 161                         bh = bd->bd_bh;
 162                         fs_err(sdp, "bd %p: blk:0x%llx bh=%p ", bd,
 163                                (unsigned long long)bd->bd_blkno, bh);
 164                         if (!bh) {
 165                                 fs_err(sdp, "\n");
 166                                 continue;
 167                         }
 168                         fs_err(sdp, "0x%llx up2:%d dirt:%d lkd:%d req:%d "
 169                                "map:%d new:%d ar:%d aw:%d delay:%d "
 170                                "io err:%d unwritten:%d dfr:%d pin:%d esc:%d\n",
 171                                (unsigned long long)bh->b_blocknr,
 172                                buffer_uptodate(bh), buffer_dirty(bh),
 173                                buffer_locked(bh), buffer_req(bh),
 174                                buffer_mapped(bh), buffer_new(bh),
 175                                buffer_async_read(bh), buffer_async_write(bh),
 176                                buffer_delay(bh), buffer_write_io_error(bh),
 177                                buffer_unwritten(bh),
 178                                buffer_defer_completion(bh),
 179                                buffer_pinned(bh), buffer_escaped(bh));
 180                 }
 181         }
 182 }
 183
 184 /**
 185  * gfs2_ail1_flush - start writeback of some ail1 entries
 186  * @sdp: The super block
 187  * @wbc: The writeback control structure
 188  *
 189  * Writes back some ail1 entries, according to the limits in the
 190  * writeback control structure
 191  */
 192
 193 void gfs2_ail1_flush(struct gfs2_sbd *sdp, struct writeback_control *wbc)
 194 {
 195         struct list_head *head = &sdp->sd_ail1_list;
 196         struct gfs2_trans *tr;
 197         struct blk_plug plug;
 198         int ret;
 199         unsigned long flush_start = jiffies;
 200
 201         trace_gfs2_ail_flush(sdp, wbc, 1);
 202         blk_start_plug(&plug);
 203         spin_lock(&sdp->sd_ail_lock);
 204 restart:
 205         ret = 0;
 206         if (time_after(jiffies, flush_start + (HZ * 600))) {
 207                 fs_err(sdp, "Error: In %s for ten minutes! t=%d\n",
 208                        __func__, current->journal_info ? 1 : 0);
 209                 dump_ail_list(sdp);
 210                 goto out;
 211         }
 212         list_for_each_entry_reverse(tr, head, tr_list) {
 213                 if (wbc->nr_to_write <= 0)
 214                         break;
 215                 ret = gfs2_ail1_start_one(sdp, wbc, tr, &plug);
 216                 if (ret) {
 217                         if (ret == -EBUSY)
 218                                 goto restart;
 219                         break;
 220                 }
 221         }
 222 out:
 223         spin_unlock(&sdp->sd_ail_lock);
 224         blk_finish_plug(&plug);
 225         if (ret) {
 226                 gfs2_lm(sdp, "gfs2_ail1_start_one (generic_writepages) "
 227                         "returned: %d\n", ret);
 228                 gfs2_withdraw(sdp);
 229         }
 230         trace_gfs2_ail_flush(sdp, wbc, 0);
 231 }
 232
 233 /**
 234  * gfs2_ail1_start - start writeback of all ail1 entries
 235  * @sdp: The superblock
 236  */
 237
 238 static void gfs2_ail1_start(struct gfs2_sbd *sdp)
 239 {
 240         struct writeback_control wbc = {
 241                 .sync_mode = WB_SYNC_NONE,
 242                 .nr_to_write = LONG_MAX,
 243                 .range_start = 0,
 244                 .range_end = LLONG_MAX,
 245         };
 246
 247         return gfs2_ail1_flush(sdp, &wbc);
 248 }
 249
 250 static void gfs2_log_update_flush_tail(struct gfs2_sbd *sdp)
 251 {
 252         unsigned int new_flush_tail = sdp->sd_log_head;
 253         struct gfs2_trans *tr;
 254
 255         if (!list_empty(&sdp->sd_ail1_list)) {
 256                 tr = list_last_entry(&sdp->sd_ail1_list,
 257                                      struct gfs2_trans, tr_list);
 258                 new_flush_tail = tr->tr_first;
 259         }
 260         sdp->sd_log_flush_tail = new_flush_tail;
 261 }
 262
 263 static void gfs2_log_update_head(struct gfs2_sbd *sdp)
 264 {
 265         unsigned int new_head = sdp->sd_log_flush_head;
 266
 267         if (sdp->sd_log_flush_tail == sdp->sd_log_head)
 268                 sdp->sd_log_flush_tail = new_head;
 269         sdp->sd_log_head = new_head;
 270 }
 271
 272 /**
 273  * gfs2_ail_empty_tr - empty one of the ail lists of a transaction
 274  */
 275
 276 static void gfs2_ail_empty_tr(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
 277                               struct list_head *head)
 278 {
 279         struct gfs2_bufdata *bd;
 280
 281         while (!list_empty(head)) {
 282                 bd = list_first_entry(head, struct gfs2_bufdata,
 283                                       bd_ail_st_list);
 284                 gfs2_assert(sdp, bd->bd_tr == tr);
 285                 gfs2_remove_from_ail(bd);
 286         }
 287 }
 288
 289 /**
 290  * gfs2_ail1_empty_one - Check whether or not a trans in the AIL has been synced
 291  * @sdp: the filesystem
 292  * @tr: the transaction
 293  * @max_revokes: If nonzero, issue revokes for the bd items for written buffers
 294  *
 295  * returns: the transaction's count of remaining active items
 296  */
 297
 298 static int gfs2_ail1_empty_one(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
 299                                 int *max_revokes)
 300 {
 301         struct gfs2_bufdata *bd, *s;
 302         struct buffer_head *bh;
 303         int active_count = 0;
 304
 305         list_for_each_entry_safe_reverse(bd, s, &tr->tr_ail1_list,
 306                                          bd_ail_st_list) {
 307                 bh = bd->bd_bh;
 308                 gfs2_assert(sdp, bd->bd_tr == tr);
 309                 /*
 310                  * If another process flagged an io error, e.g. writing to the
 311                  * journal, error all other bhs and move them off the ail1 to
 312                  * prevent a tight loop when unmount tries to flush ail1,
 313                  * regardless of whether they're still busy. If no outside
 314                  * errors were found and the buffer is busy, move to the next.
 315                  * If the ail buffer is not busy and caught an error, flag it
 316                  * for others.
 317                  */
 318                 if (!sdp->sd_log_error && buffer_busy(bh)) {
 319                         active_count++;
 320                         continue;
 321                 }
 322                 if (!buffer_uptodate(bh) &&
 323                     !cmpxchg(&sdp->sd_log_error, 0, -EIO)) {
 324                         gfs2_io_error_bh(sdp, bh);
 325                         gfs2_withdraw_delayed(sdp);
 326                 }
 327                 /*
 328                  * If we have space for revokes and the bd is no longer on any
 329                  * buf list, we can just add a revoke for it immediately and
 330                  * avoid having to put it on the ail2 list, where it would need
 331                  * to be revoked later.
 332                  */
 333                 if (*max_revokes && list_empty(&bd->bd_list)) {
 334                         gfs2_add_revoke(sdp, bd);
 335                         (*max_revokes)--;
 336                         continue;
 337                 }
 338                 list_move(&bd->bd_ail_st_list, &tr->tr_ail2_list);
 339         }
 340         return active_count;
 341 }
 342
 343 /**
 344  * gfs2_ail1_empty - Try to empty the ail1 lists
 345  * @sdp: The superblock
 346  * @max_revokes: If non-zero, add revokes where appropriate
 347  *
 348  * Tries to empty the ail1 lists, starting with the oldest first
 349  */
 350
 351 static int gfs2_ail1_empty(struct gfs2_sbd *sdp, int max_revokes)
 352 {
 353         struct gfs2_trans *tr, *s;
 354         int oldest_tr = 1;
 355         int ret;
 356
 357         spin_lock(&sdp->sd_ail_lock);
 358         list_for_each_entry_safe_reverse(tr, s, &sdp->sd_ail1_list, tr_list) {
 359                 if (!gfs2_ail1_empty_one(sdp, tr, &max_revokes) && oldest_tr)
 360                         list_move(&tr->tr_list, &sdp->sd_ail2_list);
 361                 else
 362                         oldest_tr = 0;
 363         }
 364         gfs2_log_update_flush_tail(sdp);
 365         ret = list_empty(&sdp->sd_ail1_list);
 366         spin_unlock(&sdp->sd_ail_lock);
 367
 368         if (test_bit(SDF_WITHDRAWING, &sdp->sd_flags)) {
 369                 gfs2_lm(sdp, "fatal: I/O error(s)\n");
 370                 gfs2_withdraw(sdp);
 371         }
 372
 373         return ret;
 374 }
 375
 376 static void gfs2_ail1_wait(struct gfs2_sbd *sdp)
 377 {
 378         struct gfs2_trans *tr;
 379         struct gfs2_bufdata *bd;
 380         struct buffer_head *bh;
 381
 382         spin_lock(&sdp->sd_ail_lock);
 383         list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
 384                 list_for_each_entry(bd, &tr->tr_ail1_list, bd_ail_st_list) {
 385                         bh = bd->bd_bh;
 386                         if (!buffer_locked(bh))
 387                                 continue;
 388                         get_bh(bh);
 389                         spin_unlock(&sdp->sd_ail_lock);
 390                         wait_on_buffer(bh);
 391                         brelse(bh);
 392                         return;
 393                 }
 394         }
 395         spin_unlock(&sdp->sd_ail_lock);
 396 }
 397
 398 static void __ail2_empty(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 399 {
 400         gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
 401         list_del(&tr->tr_list);
 402         gfs2_assert_warn(sdp, list_empty(&tr->tr_ail1_list));
 403         gfs2_assert_warn(sdp, list_empty(&tr->tr_ail2_list));
 404         gfs2_trans_free(sdp, tr);
 405 }
 406
 407 static void ail2_empty(struct gfs2_sbd *sdp, unsigned int new_tail)
 408 {
 409         struct list_head *ail2_list = &sdp->sd_ail2_list;
 410         unsigned int old_tail = sdp->sd_log_tail;
 411         struct gfs2_trans *tr, *safe;
 412
 413         spin_lock(&sdp->sd_ail_lock);
 414         if (old_tail <= new_tail) {
 415                 list_for_each_entry_safe(tr, safe, ail2_list, tr_list) {
 416                         if (old_tail <= tr->tr_first && tr->tr_first < new_tail)
 417                                 __ail2_empty(sdp, tr);
 418                 }
 419         } else {
 420                 list_for_each_entry_safe(tr, safe, ail2_list, tr_list) {
 421                         if (old_tail <= tr->tr_first || tr->tr_first < new_tail)
 422                                 __ail2_empty(sdp, tr);
 423                 }
 424         }
 425         spin_unlock(&sdp->sd_ail_lock);
 426 }
 427
 428 /**
 429  * gfs2_log_is_empty - Check if the log is empty
 430  * @sdp: The GFS2 superblock
 431  */
 432
 433 bool gfs2_log_is_empty(struct gfs2_sbd *sdp) {
 434         return atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks;
 435 }
 436
 437 static bool __gfs2_log_try_reserve_revokes(struct gfs2_sbd *sdp, unsigned int revokes)
 438 {
 439         unsigned int available;
 440
 441         available = atomic_read(&sdp->sd_log_revokes_available);
 442         while (available >= revokes) {
 443                 if (atomic_try_cmpxchg(&sdp->sd_log_revokes_available,
 444                                        &available, available - revokes))
 445                         return true;
 446         }
 447         return false;
 448 }
 449
 450 /**
 451  * gfs2_log_release_revokes - Release a given number of revokes
 452  * @sdp: The GFS2 superblock
 453  * @revokes: The number of revokes to release
 454  *
 455  * sdp->sd_log_flush_lock must be held.
 456  */
 457 void gfs2_log_release_revokes(struct gfs2_sbd *sdp, unsigned int revokes)
 458 {
 459         if (revokes)
 460                 atomic_add(revokes, &sdp->sd_log_revokes_available);
 461 }
 462
 463 /**
 464  * gfs2_log_release - Release a given number of log blocks
 465  * @sdp: The GFS2 superblock
 466  * @blks: The number of blocks
 467  *
 468  */
 469
 470 void gfs2_log_release(struct gfs2_sbd *sdp, unsigned int blks)
 471 {
 472         atomic_add(blks, &sdp->sd_log_blks_free);
 473         trace_gfs2_log_blocks(sdp, blks);
 474         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 475                                   sdp->sd_jdesc->jd_blocks);
 476         if (atomic_read(&sdp->sd_log_blks_needed))
 477                 wake_up(&sdp->sd_log_waitq);
 478 }
 479
 480 /**
 481  * __gfs2_log_try_reserve - Try to make a log reservation
 482  * @sdp: The GFS2 superblock
 483  * @blks: The number of blocks to reserve
 484  * @taboo_blks: The number of blocks to leave free
 485  *
 486  * Try to do the same as __gfs2_log_reserve(), but fail if no more log
 487  * space is immediately available.
 488  */
 489 static bool __gfs2_log_try_reserve(struct gfs2_sbd *sdp, unsigned int blks,
 490                                    unsigned int taboo_blks)
 491 {
 492         unsigned wanted = blks + taboo_blks;
 493         unsigned int free_blocks;
 494
 495         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 496         while (free_blocks >= wanted) {
 497                 if (atomic_try_cmpxchg(&sdp->sd_log_blks_free, &free_blocks,
 498                                        free_blocks - blks)) {
 499                         trace_gfs2_log_blocks(sdp, -blks);
 500                         return true;
 501                 }
 502         }
 503         return false;
 504 }
 505
 506 /**
 507  * __gfs2_log_reserve - Make a log reservation
 508  * @sdp: The GFS2 superblock
 509  * @blks: The number of blocks to reserve
 510  * @taboo_blks: The number of blocks to leave free
 511  *
 512  * @taboo_blks is set to 0 for logd, and to GFS2_LOG_FLUSH_MIN_BLOCKS
 513  * for all other processes.  This ensures that when the log is almost full,
 514  * logd will still be able to call gfs2_log_flush one more time  without
 515  * blocking, which will advance the tail and make some more log space
 516  * available.
 517  *
 518  * We no longer flush the log here, instead we wake up logd to do that
 519  * for us. To avoid the thundering herd and to ensure that we deal fairly
 520  * with queued waiters, we use an exclusive wait. This means that when we
 521  * get woken with enough journal space to get our reservation, we need to
 522  * wake the next waiter on the list.
 523  */
 524
 525 static void __gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks,
 526                                unsigned int taboo_blks)
 527 {
 528         unsigned wanted = blks + taboo_blks;
 529         unsigned int free_blocks;
 530
 531         atomic_add(blks, &sdp->sd_log_blks_needed);
 532         for (;;) {
 533                 if (current != sdp->sd_logd_process)
 534                         wake_up(&sdp->sd_logd_waitq);
 535                 io_wait_event(sdp->sd_log_waitq,
 536                         (free_blocks = atomic_read(&sdp->sd_log_blks_free),
 537                          free_blocks >= wanted));
 538                 do {
 539                         if (atomic_try_cmpxchg(&sdp->sd_log_blks_free,
 540                                                &free_blocks,
 541                                                free_blocks - blks))
 542                                 goto reserved;
 543                 } while (free_blocks >= wanted);
 544         }
 545
 546 reserved:
 547         trace_gfs2_log_blocks(sdp, -blks);
 548         if (atomic_sub_return(blks, &sdp->sd_log_blks_needed))
 549                 wake_up(&sdp->sd_log_waitq);
 550 }
 551
 552 /**
 553  * gfs2_log_try_reserve - Try to make a log reservation
 554  * @sdp: The GFS2 superblock
 555  * @tr: The transaction
 556  * @extra_revokes: The number of additional revokes reserved (output)
 557  *
 558  * This is similar to gfs2_log_reserve, but sdp->sd_log_flush_lock must be
 559  * held for correct revoke accounting.
 560  */
 561
 562 bool gfs2_log_try_reserve(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
 563                           unsigned int *extra_revokes)
 564 {
 565         unsigned int blks = tr->tr_reserved;
 566         unsigned int revokes = tr->tr_revokes;
 567         unsigned int revoke_blks = 0;
 568
 569         *extra_revokes = 0;
 570         if (revokes && !__gfs2_log_try_reserve_revokes(sdp, revokes)) {
 571                 revoke_blks = DIV_ROUND_UP(revokes, sdp->sd_inptrs);
 572                 *extra_revokes = revoke_blks * sdp->sd_inptrs - revokes;
 573                 blks += revoke_blks;
 574         }
 575         if (!blks)
 576                 return true;
 577         if (__gfs2_log_try_reserve(sdp, blks, GFS2_LOG_FLUSH_MIN_BLOCKS))
 578                 return true;
 579         if (!revoke_blks)
 580                 gfs2_log_release_revokes(sdp, revokes);
 581         return false;
 582 }
 583
 584 /**
 585  * gfs2_log_reserve - Make a log reservation
 586  * @sdp: The GFS2 superblock
 587  * @tr: The transaction
 588  * @extra_revokes: The number of additional revokes reserved (output)
 589  *
 590  * sdp->sd_log_flush_lock must not be held.
 591  */
 592
 593 void gfs2_log_reserve(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
 594                       unsigned int *extra_revokes)
 595 {
 596         unsigned int blks = tr->tr_reserved;
 597         unsigned int revokes = tr->tr_revokes;
 598         unsigned int revoke_blks = 0;
 599
 600         *extra_revokes = 0;
 601         if (revokes) {
 602                 revoke_blks = DIV_ROUND_UP(revokes, sdp->sd_inptrs);
 603                 *extra_revokes = revoke_blks * sdp->sd_inptrs - revokes;
 604                 blks += revoke_blks;
 605         }
 606         __gfs2_log_reserve(sdp, blks, GFS2_LOG_FLUSH_MIN_BLOCKS);
 607 }
 608
 609 /**
 610  * log_distance - Compute distance between two journal blocks
 611  * @sdp: The GFS2 superblock
 612  * @newer: The most recent journal block of the pair
 613  * @older: The older journal block of the pair
 614  *
 615  *   Compute the distance (in the journal direction) between two
 616  *   blocks in the journal
 617  *
 618  * Returns: the distance in blocks
 619  */
 620
 621 static inline unsigned int log_distance(struct gfs2_sbd *sdp, unsigned int newer,
 622                                         unsigned int older)
 623 {
 624         int dist;
 625
 626         dist = newer - older;
 627         if (dist < 0)
 628                 dist += sdp->sd_jdesc->jd_blocks;
 629
 630         return dist;
 631 }
 632
 633 /**
 634  * calc_reserved - Calculate the number of blocks to keep reserved
 635  * @sdp: The GFS2 superblock
 636  *
 637  * This is complex.  We need to reserve room for all our currently used
 638  * metadata blocks (e.g. normal file I/O rewriting file time stamps) and
 639  * all our journaled data blocks for journaled files (e.g. files in the
 640  * meta_fs like rindex, or files for which chattr +j was done.)
 641  * If we don't reserve enough space, corruption will follow.
 642  *
 643  * We can have metadata blocks and jdata blocks in the same journal.  Each
 644  * type gets its own log descriptor, for which we need to reserve a block.
 645  * In fact, each type has the potential for needing more than one log descriptor
 646  * in cases where we have more blocks than will fit in a log descriptor.
 647  * Metadata journal entries take up half the space of journaled buffer entries.
 648  *
 649  * Also, we need to reserve blocks for revoke journal entries and one for an
 650  * overall header for the lot.
 651  *
 652  * Returns: the number of blocks reserved
 653  */
 654 static unsigned int calc_reserved(struct gfs2_sbd *sdp)
 655 {
 656         unsigned int reserved = GFS2_LOG_FLUSH_MIN_BLOCKS;
 657         unsigned int blocks;
 658         struct gfs2_trans *tr = sdp->sd_log_tr;
 659
 660         if (tr) {
 661                 blocks = tr->tr_num_buf_new - tr->tr_num_buf_rm;
 662                 reserved += blocks + DIV_ROUND_UP(blocks, buf_limit(sdp));
 663                 blocks = tr->tr_num_databuf_new - tr->tr_num_databuf_rm;
 664                 reserved += blocks + DIV_ROUND_UP(blocks, databuf_limit(sdp));
 665         }
 666         return reserved;
 667 }
 668
 669 static void log_pull_tail(struct gfs2_sbd *sdp)
 670 {
 671         unsigned int new_tail = sdp->sd_log_flush_tail;
 672         unsigned int dist;
 673
 674         if (new_tail == sdp->sd_log_tail)
 675                 return;
 676         dist = log_distance(sdp, new_tail, sdp->sd_log_tail);
 677         ail2_empty(sdp, new_tail);
 678         gfs2_log_release(sdp, dist);
 679         sdp->sd_log_tail = new_tail;
 680 }
 681
 682
 683 void log_flush_wait(struct gfs2_sbd *sdp)
 684 {
 685         DEFINE_WAIT(wait);
 686
 687         if (atomic_read(&sdp->sd_log_in_flight)) {
 688                 do {
 689                         prepare_to_wait(&sdp->sd_log_flush_wait, &wait,
 690                                         TASK_UNINTERRUPTIBLE);
 691                         if (atomic_read(&sdp->sd_log_in_flight))
 692                                 io_schedule();
 693                 } while(atomic_read(&sdp->sd_log_in_flight));
 694                 finish_wait(&sdp->sd_log_flush_wait, &wait);
 695         }
 696 }
 697
 698 static int ip_cmp(void *priv, struct list_head *a, struct list_head *b)
 699 {
 700         struct gfs2_inode *ipa, *ipb;
 701
 702         ipa = list_entry(a, struct gfs2_inode, i_ordered);
 703         ipb = list_entry(b, struct gfs2_inode, i_ordered);
 704
 705         if (ipa->i_no_addr < ipb->i_no_addr)
 706                 return -1;
 707         if (ipa->i_no_addr > ipb->i_no_addr)
 708                 return 1;
 709         return 0;
 710 }
 711
 712 static void __ordered_del_inode(struct gfs2_inode *ip)
 713 {
 714         if (!list_empty(&ip->i_ordered))
 715                 list_del_init(&ip->i_ordered);
 716 }
 717
 718 static void gfs2_ordered_write(struct gfs2_sbd *sdp)
 719 {
 720         struct gfs2_inode *ip;
 721         LIST_HEAD(written);
 722
 723         spin_lock(&sdp->sd_ordered_lock);
 724         list_sort(NULL, &sdp->sd_log_ordered, &ip_cmp);
 725         while (!list_empty(&sdp->sd_log_ordered)) {
 726                 ip = list_first_entry(&sdp->sd_log_ordered, struct gfs2_inode, i_ordered);
 727                 if (ip->i_inode.i_mapping->nrpages == 0) {
 728                         __ordered_del_inode(ip);
 729                         continue;
 730                 }
 731                 list_move(&ip->i_ordered, &written);
 732                 spin_unlock(&sdp->sd_ordered_lock);
 733                 filemap_fdatawrite(ip->i_inode.i_mapping);
 734                 spin_lock(&sdp->sd_ordered_lock);
 735         }
 736         list_splice(&written, &sdp->sd_log_ordered);
 737         spin_unlock(&sdp->sd_ordered_lock);
 738 }
 739
 740 static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
 741 {
 742         struct gfs2_inode *ip;
 743
 744         spin_lock(&sdp->sd_ordered_lock);
 745         while (!list_empty(&sdp->sd_log_ordered)) {
 746                 ip = list_first_entry(&sdp->sd_log_ordered, struct gfs2_inode, i_ordered);
 747                 __ordered_del_inode(ip);
 748                 if (ip->i_inode.i_mapping->nrpages == 0)
 749                         continue;
 750                 spin_unlock(&sdp->sd_ordered_lock);
 751                 filemap_fdatawait(ip->i_inode.i_mapping);
 752                 spin_lock(&sdp->sd_ordered_lock);
 753         }
 754         spin_unlock(&sdp->sd_ordered_lock);
 755 }
 756
 757 void gfs2_ordered_del_inode(struct gfs2_inode *ip)
 758 {
 759         struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
 760
 761         spin_lock(&sdp->sd_ordered_lock);
 762         __ordered_del_inode(ip);
 763         spin_unlock(&sdp->sd_ordered_lock);
 764 }
 765
 766 void gfs2_add_revoke(struct gfs2_sbd *sdp, struct gfs2_bufdata *bd)
 767 {
 768         struct buffer_head *bh = bd->bd_bh;
 769         struct gfs2_glock *gl = bd->bd_gl;
 770
 771         sdp->sd_log_num_revoke++;
 772         if (atomic_inc_return(&gl->gl_revokes) == 1)
 773                 gfs2_glock_hold(gl);
 774         bh->b_private = NULL;
 775         bd->bd_blkno = bh->b_blocknr;
 776         gfs2_remove_from_ail(bd); /* drops ref on bh */
 777         bd->bd_bh = NULL;
 778         set_bit(GLF_LFLUSH, &gl->gl_flags);
 779         list_add(&bd->bd_list, &sdp->sd_log_revokes);
 780 }
 781
 782 void gfs2_glock_remove_revoke(struct gfs2_glock *gl)
 783 {
 784         if (atomic_dec_return(&gl->gl_revokes) == 0) {
 785                 clear_bit(GLF_LFLUSH, &gl->gl_flags);
 786                 gfs2_glock_queue_put(gl);
 787         }
 788 }
 789
 790 /**
 791  * gfs2_flush_revokes - Add as many revokes to the system transaction as we can
 792  * @sdp: The GFS2 superblock
 793  *
 794  * Our usual strategy is to defer writing revokes as much as we can in the hope
 795  * that we'll eventually overwrite the journal, which will make those revokes
 796  * go away.  This changes when we flush the log: at that point, there will
 797  * likely be some left-over space in the last revoke block of that transaction.
 798  * We can fill that space with additional revokes for blocks that have already
 799  * been written back.  This will basically come at no cost now, and will save
 800  * us from having to keep track of those blocks on the AIL2 list later.
 801  */
 802 void gfs2_flush_revokes(struct gfs2_sbd *sdp)
 803 {
 804         /* number of revokes we still have room for */
 805         unsigned int max_revokes = atomic_read(&sdp->sd_log_revokes_available);
 806
 807         gfs2_log_lock(sdp);
 808         gfs2_ail1_empty(sdp, max_revokes);
 809         gfs2_log_unlock(sdp);
 810 }
 811
 812 /**
 813  * gfs2_write_log_header - Write a journal log header buffer at lblock
 814  * @sdp: The GFS2 superblock
 815  * @jd: journal descriptor of the journal to which we are writing
 816  * @seq: sequence number
 817  * @tail: tail of the log
 818  * @lblock: value for lh_blkno (block number relative to start of journal)
 819  * @flags: log header flags GFS2_LOG_HEAD_*
 820  * @op_flags: flags to pass to the bio
 821  *
 822  * Returns: the initialized log buffer descriptor
 823  */
 824
 825 void gfs2_write_log_header(struct gfs2_sbd *sdp, struct gfs2_jdesc *jd,
 826                            u64 seq, u32 tail, u32 lblock, u32 flags,
 827                            int op_flags)
 828 {
 829         struct gfs2_log_header *lh;
 830         u32 hash, crc;
 831         struct page *page;
 832         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 833         struct timespec64 tv;
 834         struct super_block *sb = sdp->sd_vfs;
 835         u64 dblock;
 836
 837         if (gfs2_withdrawn(sdp))
 838                 return;
 839
 840         page = mempool_alloc(gfs2_page_pool, GFP_NOIO);
 841         lh = page_address(page);
 842         clear_page(lh);
 843
 844         lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
 845         lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
 846         lh->lh_header.__pad0 = cpu_to_be64(0);
 847         lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
 848         lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
 849         lh->lh_sequence = cpu_to_be64(seq);
 850         lh->lh_flags = cpu_to_be32(flags);
 851         lh->lh_tail = cpu_to_be32(tail);
 852         lh->lh_blkno = cpu_to_be32(lblock);
 853         hash = ~crc32(~0, lh, LH_V1_SIZE);
 854         lh->lh_hash = cpu_to_be32(hash);
 855
 856         ktime_get_coarse_real_ts64(&tv);
 857         lh->lh_nsec = cpu_to_be32(tv.tv_nsec);
 858         lh->lh_sec = cpu_to_be64(tv.tv_sec);
 859         if (!list_empty(&jd->extent_list))
 860                 dblock = gfs2_log_bmap(jd, lblock);
 861         else {
 862                 int ret = gfs2_lblk_to_dblk(jd->jd_inode, lblock, &dblock);
 863                 if (gfs2_assert_withdraw(sdp, ret == 0))
 864                         return;
 865         }
 866         lh->lh_addr = cpu_to_be64(dblock);
 867         lh->lh_jinode = cpu_to_be64(GFS2_I(jd->jd_inode)->i_no_addr);
 868
 869         /* We may only write local statfs, quota, etc., when writing to our
 870            own journal. The values are left 0 when recovering a journal
 871            different from our own. */
 872         if (!(flags & GFS2_LOG_HEAD_RECOVERY)) {
 873                 lh->lh_statfs_addr =
 874                         cpu_to_be64(GFS2_I(sdp->sd_sc_inode)->i_no_addr);
 875                 lh->lh_quota_addr =
 876                         cpu_to_be64(GFS2_I(sdp->sd_qc_inode)->i_no_addr);
 877
 878                 spin_lock(&sdp->sd_statfs_spin);
 879                 lh->lh_local_total = cpu_to_be64(l_sc->sc_total);
 880                 lh->lh_local_free = cpu_to_be64(l_sc->sc_free);
 881                 lh->lh_local_dinodes = cpu_to_be64(l_sc->sc_dinodes);
 882                 spin_unlock(&sdp->sd_statfs_spin);
 883         }
 884
 885         BUILD_BUG_ON(offsetof(struct gfs2_log_header, lh_crc) != LH_V1_SIZE);
 886
 887         crc = crc32c(~0, (void *)lh + LH_V1_SIZE + 4,
 888                      sb->s_blocksize - LH_V1_SIZE - 4);
 889         lh->lh_crc = cpu_to_be32(crc);
 890
 891         gfs2_log_write(sdp, jd, page, sb->s_blocksize, 0, dblock);
 892         gfs2_log_submit_bio(&jd->jd_log_bio, REQ_OP_WRITE | op_flags);
 893 }
 894
 895 /**
 896  * log_write_header - Get and initialize a journal header buffer
 897  * @sdp: The GFS2 superblock
 898  * @flags: The log header flags, including log header origin
 899  *
 900  * Returns: the initialized log buffer descriptor
 901  */
 902
 903 static void log_write_header(struct gfs2_sbd *sdp, u32 flags)
 904 {
 905         int op_flags = REQ_PREFLUSH | REQ_FUA | REQ_META | REQ_SYNC;
 906         enum gfs2_freeze_state state = atomic_read(&sdp->sd_freeze_state);
 907
 908         gfs2_assert_withdraw(sdp, (state != SFS_FROZEN));
 909
 910         if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags)) {
 911                 gfs2_ordered_wait(sdp);
 912                 log_flush_wait(sdp);
 913                 op_flags = REQ_SYNC | REQ_META | REQ_PRIO;
 914         }
 915         sdp->sd_log_idle = (sdp->sd_log_flush_tail == sdp->sd_log_flush_head);
 916         gfs2_write_log_header(sdp, sdp->sd_jdesc, sdp->sd_log_sequence++,
 917                               sdp->sd_log_flush_tail, sdp->sd_log_flush_head,
 918                               flags, op_flags);
 919         gfs2_log_incr_head(sdp);
 920         log_flush_wait(sdp);
 921         log_pull_tail(sdp);
 922         gfs2_log_update_head(sdp);
 923 }
 924
 925 /**
 926  * ail_drain - drain the ail lists after a withdraw
 927  * @sdp: Pointer to GFS2 superblock
 928  */
 929 static void ail_drain(struct gfs2_sbd *sdp)
 930 {
 931         struct gfs2_trans *tr;
 932
 933         spin_lock(&sdp->sd_ail_lock);
 934         /*
 935          * For transactions on the sd_ail1_list we need to drain both the
 936          * ail1 and ail2 lists. That's because function gfs2_ail1_start_one
 937          * (temporarily) moves items from its tr_ail1 list to tr_ail2 list
 938          * before revokes are sent for that block. Items on the sd_ail2_list
 939          * should have already gotten beyond that point, so no need.
 940          */
 941         while (!list_empty(&sdp->sd_ail1_list)) {
 942                 tr = list_first_entry(&sdp->sd_ail1_list, struct gfs2_trans,
 943                                       tr_list);
 944                 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail1_list);
 945                 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
 946                 list_del(&tr->tr_list);
 947                 gfs2_trans_free(sdp, tr);
 948         }
 949         while (!list_empty(&sdp->sd_ail2_list)) {
 950                 tr = list_first_entry(&sdp->sd_ail2_list, struct gfs2_trans,
 951                                       tr_list);
 952                 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
 953                 list_del(&tr->tr_list);
 954                 gfs2_trans_free(sdp, tr);
 955         }
 956         spin_unlock(&sdp->sd_ail_lock);
 957 }
 958
 959 /**
 960  * empty_ail1_list - try to start IO and empty the ail1 list
 961  * @sdp: Pointer to GFS2 superblock
 962  */
 963 static void empty_ail1_list(struct gfs2_sbd *sdp)
 964 {
 965         unsigned long start = jiffies;
 966
 967         for (;;) {
 968                 if (time_after(jiffies, start + (HZ * 600))) {
 969                         fs_err(sdp, "Error: In %s for 10 minutes! t=%d\n",
 970                                __func__, current->journal_info ? 1 : 0);
 971                         dump_ail_list(sdp);
 972                         return;
 973                 }
 974                 gfs2_ail1_start(sdp);
 975                 gfs2_ail1_wait(sdp);
 976                 if (gfs2_ail1_empty(sdp, 0))
 977                         return;
 978         }
 979 }
 980
 981 /**
 982  * trans_drain - drain the buf and databuf queue for a failed transaction
 983  * @tr: the transaction to drain
 984  *
 985  * When this is called, we're taking an error exit for a log write that failed
 986  * but since we bypassed the after_commit functions, we need to remove the
 987  * items from the buf and databuf queue.
 988  */
 989 static void trans_drain(struct gfs2_trans *tr)
 990 {
 991         struct gfs2_bufdata *bd;
 992         struct list_head *head;
 993
 994         if (!tr)
 995                 return;
 996
 997         head = &tr->tr_buf;
 998         while (!list_empty(head)) {
 999                 bd = list_first_entry(head, struct gfs2_bufdata, bd_list);
1000                 list_del_init(&bd->bd_list);
1001                 if (!list_empty(&bd->bd_ail_st_list))
1002                         gfs2_remove_from_ail(bd);
1003                 kmem_cache_free(gfs2_bufdata_cachep, bd);
1004         }
1005         head = &tr->tr_databuf;
1006         while (!list_empty(head)) {
1007                 bd = list_first_entry(head, struct gfs2_bufdata, bd_list);
1008                 list_del_init(&bd->bd_list);
1009                 if (!list_empty(&bd->bd_ail_st_list))
1010                         gfs2_remove_from_ail(bd);
1011                 kmem_cache_free(gfs2_bufdata_cachep, bd);
1012         }
1013 }
1014
1015 /**
1016  * gfs2_log_flush - flush incore transaction(s)
1017  * @sdp: the filesystem
1018  * @gl: The glock structure to flush.  If NULL, flush the whole incore log
1019  * @flags: The log header flags: GFS2_LOG_HEAD_FLUSH_* and debug flags
1020  *
1021  */
1022
1023 void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl, u32 flags)
1024 {
1025         struct gfs2_trans *tr = NULL;
1026         unsigned int reserved_blocks = 0, used_blocks = 0;
1027         enum gfs2_freeze_state state = atomic_read(&sdp->sd_freeze_state);
1028         unsigned int first_log_head;
1029         unsigned int reserved_revokes = 0;
1030
1031         down_write(&sdp->sd_log_flush_lock);
1032         trace_gfs2_log_flush(sdp, 1, flags);
1033
1034 repeat:
1035         /*
1036          * Do this check while holding the log_flush_lock to prevent new
1037          * buffers from being added to the ail via gfs2_pin()
1038          */
1039         if (gfs2_withdrawn(sdp))
1040                 goto out;
1041
1042         /* Log might have been flushed while we waited for the flush lock */
1043         if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags))
1044                 goto out;
1045
1046         first_log_head = sdp->sd_log_head;
1047         sdp->sd_log_flush_head = first_log_head;
1048
1049         tr = sdp->sd_log_tr;
1050         if (tr || sdp->sd_log_num_revoke) {
1051                 if (reserved_blocks)
1052                         gfs2_log_release(sdp, reserved_blocks);
1053                 reserved_blocks = sdp->sd_log_blks_reserved;
1054                 reserved_revokes = sdp->sd_log_num_revoke;
1055                 if (tr) {
1056                         sdp->sd_log_tr = NULL;
1057                         tr->tr_first = first_log_head;
1058                         if (unlikely (state == SFS_FROZEN)) {
1059                                 if (gfs2_assert_withdraw_delayed(sdp,
1060                                        !tr->tr_num_buf_new && !tr->tr_num_databuf_new))
1061                                         goto out_withdraw;
1062                         }
1063                 }
1064         } else if (!reserved_blocks) {
1065                 unsigned int taboo_blocks = GFS2_LOG_FLUSH_MIN_BLOCKS;
1066
1067                 reserved_blocks = GFS2_LOG_FLUSH_MIN_BLOCKS;
1068                 if (current == sdp->sd_logd_process)
1069                         taboo_blocks = 0;
1070
1071                 if (!__gfs2_log_try_reserve(sdp, reserved_blocks, taboo_blocks)) {
1072                         up_write(&sdp->sd_log_flush_lock);
1073                         __gfs2_log_reserve(sdp, reserved_blocks, taboo_blocks);
1074                         down_write(&sdp->sd_log_flush_lock);
1075                         goto repeat;
1076                 }
1077                 BUG_ON(sdp->sd_log_num_revoke);
1078         }
1079
1080         if (flags & GFS2_LOG_HEAD_FLUSH_SHUTDOWN)
1081                 clear_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags);
1082
1083         if (unlikely(state == SFS_FROZEN))
1084                 if (gfs2_assert_withdraw_delayed(sdp, !reserved_revokes))
1085                         goto out_withdraw;
1086
1087         gfs2_ordered_write(sdp);
1088         if (gfs2_withdrawn(sdp))
1089                 goto out_withdraw;
1090         lops_before_commit(sdp, tr);
1091         if (gfs2_withdrawn(sdp))
1092                 goto out_withdraw;
1093         gfs2_log_submit_bio(&sdp->sd_jdesc->jd_log_bio, REQ_OP_WRITE);
1094         if (gfs2_withdrawn(sdp))
1095                 goto out_withdraw;
1096
1097         if (sdp->sd_log_head != sdp->sd_log_flush_head) {
1098                 log_write_header(sdp, flags);
1099         } else if (sdp->sd_log_tail != sdp->sd_log_flush_tail && !sdp->sd_log_idle) {
1100                 log_write_header(sdp, flags);
1101         }
1102         if (gfs2_withdrawn(sdp))
1103                 goto out_withdraw;
1104         lops_after_commit(sdp, tr);
1105
1106         gfs2_log_lock(sdp);
1107         sdp->sd_log_blks_reserved = 0;
1108
1109         spin_lock(&sdp->sd_ail_lock);
1110         if (tr && !list_empty(&tr->tr_ail1_list)) {
1111                 list_add(&tr->tr_list, &sdp->sd_ail1_list);
1112                 tr = NULL;
1113         }
1114         spin_unlock(&sdp->sd_ail_lock);
1115         gfs2_log_unlock(sdp);
1116
1117         if (!(flags & GFS2_LOG_HEAD_FLUSH_NORMAL)) {
1118                 if (!sdp->sd_log_idle) {
1119                         empty_ail1_list(sdp);
1120                         if (gfs2_withdrawn(sdp))
1121                                 goto out_withdraw;
1122                         log_write_header(sdp, flags);
1123                 }
1124                 if (flags & (GFS2_LOG_HEAD_FLUSH_SHUTDOWN |
1125                              GFS2_LOG_HEAD_FLUSH_FREEZE))
1126                         gfs2_log_shutdown(sdp);
1127                 if (flags & GFS2_LOG_HEAD_FLUSH_FREEZE)
1128                         atomic_set(&sdp->sd_freeze_state, SFS_FROZEN);
1129         }
1130
1131 out_end:
1132         used_blocks = log_distance(sdp, sdp->sd_log_flush_head, first_log_head);
1133         reserved_revokes += atomic_read(&sdp->sd_log_revokes_available);
1134         atomic_set(&sdp->sd_log_revokes_available, sdp->sd_ldptrs);
1135         gfs2_assert_withdraw(sdp, reserved_revokes % sdp->sd_inptrs == sdp->sd_ldptrs);
1136         if (reserved_revokes > sdp->sd_ldptrs)
1137                 reserved_blocks += (reserved_revokes - sdp->sd_ldptrs) / sdp->sd_inptrs;
1138 out:
1139         if (used_blocks != reserved_blocks) {
1140                 gfs2_assert_withdraw_delayed(sdp, used_blocks < reserved_blocks);
1141                 gfs2_log_release(sdp, reserved_blocks - used_blocks);
1142         }
1143         up_write(&sdp->sd_log_flush_lock);
1144         gfs2_trans_free(sdp, tr);
1145         if (gfs2_withdrawing(sdp))
1146                 gfs2_withdraw(sdp);
1147         trace_gfs2_log_flush(sdp, 0, flags);
1148         return;
1149
1150 out_withdraw:
1151         trans_drain(tr);
1152         /**
1153          * If the tr_list is empty, we're withdrawing during a log
1154          * flush that targets a transaction, but the transaction was
1155          * never queued onto any of the ail lists. Here we add it to
1156          * ail1 just so that ail_drain() will find and free it.
1157          */
1158         spin_lock(&sdp->sd_ail_lock);
1159         if (tr && list_empty(&tr->tr_list))
1160                 list_add(&tr->tr_list, &sdp->sd_ail1_list);
1161         spin_unlock(&sdp->sd_ail_lock);
1162         ail_drain(sdp); /* frees all transactions */
1163         tr = NULL;
1164         goto out_end;
1165 }
1166
1167 /**
1168  * gfs2_merge_trans - Merge a new transaction into a cached transaction
1169  * @old: Original transaction to be expanded
1170  * @new: New transaction to be merged
1171  */
1172
1173 static void gfs2_merge_trans(struct gfs2_sbd *sdp, struct gfs2_trans *new)
1174 {
1175         struct gfs2_trans *old = sdp->sd_log_tr;
1176
1177         WARN_ON_ONCE(!test_bit(TR_ATTACHED, &old->tr_flags));
1178
1179         old->tr_num_buf_new     += new->tr_num_buf_new;
1180         old->tr_num_databuf_new += new->tr_num_databuf_new;
1181         old->tr_num_buf_rm      += new->tr_num_buf_rm;
1182         old->tr_num_databuf_rm  += new->tr_num_databuf_rm;
1183         old->tr_revokes         += new->tr_revokes;
1184         old->tr_num_revoke      += new->tr_num_revoke;
1185
1186         list_splice_tail_init(&new->tr_databuf, &old->tr_databuf);
1187         list_splice_tail_init(&new->tr_buf, &old->tr_buf);
1188
1189         spin_lock(&sdp->sd_ail_lock);
1190         list_splice_tail_init(&new->tr_ail1_list, &old->tr_ail1_list);
1191         list_splice_tail_init(&new->tr_ail2_list, &old->tr_ail2_list);
1192         spin_unlock(&sdp->sd_ail_lock);
1193 }
1194
1195 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
1196 {
1197         unsigned int reserved;
1198         unsigned int unused;
1199         unsigned int maxres;
1200
1201         gfs2_log_lock(sdp);
1202
1203         if (sdp->sd_log_tr) {
1204                 gfs2_merge_trans(sdp, tr);
1205         } else if (tr->tr_num_buf_new || tr->tr_num_databuf_new) {
1206                 gfs2_assert_withdraw(sdp, !test_bit(TR_ONSTACK, &tr->tr_flags));
1207                 sdp->sd_log_tr = tr;
1208                 set_bit(TR_ATTACHED, &tr->tr_flags);
1209         }
1210
1211         reserved = calc_reserved(sdp);
1212         maxres = sdp->sd_log_blks_reserved + tr->tr_reserved;
1213         gfs2_assert_withdraw(sdp, maxres >= reserved);
1214         unused = maxres - reserved;
1215         if (unused)
1216                 gfs2_log_release(sdp, unused);
1217         sdp->sd_log_blks_reserved = reserved;
1218
1219         gfs2_log_unlock(sdp);
1220 }
1221
1222 /**
1223  * gfs2_log_commit - Commit a transaction to the log
1224  * @sdp: the filesystem
1225  * @tr: the transaction
1226  *
1227  * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
1228  * or the total number of used blocks (pinned blocks plus AIL blocks)
1229  * is greater than thresh2.
1230  *
1231  * At mount time thresh1 is 2/5ths of journal size, thresh2 is 4/5ths of
1232  * journal size.
1233  *
1234  * Returns: errno
1235  */
1236
1237 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
1238 {
1239         log_refund(sdp, tr);
1240
1241         if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
1242             ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
1243             atomic_read(&sdp->sd_log_thresh2)))
1244                 wake_up(&sdp->sd_logd_waitq);
1245 }
1246
1247 /**
1248  * gfs2_log_shutdown - write a shutdown header into a journal
1249  * @sdp: the filesystem
1250  *
1251  */
1252
1253 static void gfs2_log_shutdown(struct gfs2_sbd *sdp)
1254 {
1255         gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
1256         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
1257         gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
1258
1259         log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT | GFS2_LFC_SHUTDOWN);
1260         log_pull_tail(sdp);
1261
1262         gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
1263         gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
1264 }
1265
1266 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
1267 {
1268         return (atomic_read(&sdp->sd_log_pinned) +
1269                 atomic_read(&sdp->sd_log_blks_needed) >=
1270                 atomic_read(&sdp->sd_log_thresh1));
1271 }
1272
1273 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
1274 {
1275         unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
1276
1277         if (test_and_clear_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags))
1278                 return 1;
1279
1280         return used_blocks + atomic_read(&sdp->sd_log_blks_needed) >=
1281                 atomic_read(&sdp->sd_log_thresh2);
1282 }
1283
1284 /**
1285  * gfs2_logd - Update log tail as Active Items get flushed to in-place blocks
1286  * @sdp: Pointer to GFS2 superblock
1287  *
1288  * Also, periodically check to make sure that we're using the most recent
1289  * journal index.
1290  */
1291
1292 int gfs2_logd(void *data)
1293 {
1294         struct gfs2_sbd *sdp = data;
1295         unsigned long t = 1;
1296         DEFINE_WAIT(wait);
1297
1298         while (!kthread_should_stop()) {
1299
1300                 if (gfs2_withdrawn(sdp)) {
1301                         msleep_interruptible(HZ);
1302                         continue;
1303                 }
1304                 /* Check for errors writing to the journal */
1305                 if (sdp->sd_log_error) {
1306                         gfs2_lm(sdp,
1307                                 "GFS2: fsid=%s: error %d: "
1308                                 "withdrawing the file system to "
1309                                 "prevent further damage.\n",
1310                                 sdp->sd_fsname, sdp->sd_log_error);
1311                         gfs2_withdraw(sdp);
1312                         continue;
1313                 }
1314
1315                 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
1316                         gfs2_ail1_empty(sdp, 0);
1317                         gfs2_log_flush(sdp, NULL, GFS2_LOG_HEAD_FLUSH_NORMAL |
1318                                                   GFS2_LFC_LOGD_JFLUSH_REQD);
1319                 }
1320
1321                 if (gfs2_ail_flush_reqd(sdp)) {
1322                         gfs2_ail1_start(sdp);
1323                         gfs2_ail1_wait(sdp);
1324                         gfs2_ail1_empty(sdp, 0);
1325                         gfs2_log_flush(sdp, NULL, GFS2_LOG_HEAD_FLUSH_NORMAL |
1326                                                   GFS2_LFC_LOGD_AIL_FLUSH_REQD);
1327                 }
1328
1329                 t = gfs2_tune_get(sdp, gt_logd_secs) * HZ;
1330
1331                 try_to_freeze();
1332
1333                 do {
1334                         prepare_to_wait(&sdp->sd_logd_waitq, &wait,
1335                                         TASK_INTERRUPTIBLE);
1336                         if (!gfs2_ail_flush_reqd(sdp) &&
1337                             !gfs2_jrnl_flush_reqd(sdp) &&
1338                             !kthread_should_stop())
1339                                 t = schedule_timeout(t);
1340                 } while(t && !gfs2_ail_flush_reqd(sdp) &&
1341                         !gfs2_jrnl_flush_reqd(sdp) &&
1342                         !kthread_should_stop());
1343                 finish_wait(&sdp->sd_logd_waitq, &wait);
1344         }
1345
1346         return 0;
1347 }
1348