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