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_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_page->mapping;
 131                 if (!mapping)
 132                         continue;
 133                 spin_unlock(&sdp->sd_ail_lock);
 134                 ret = generic_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                 if (ret || wbc->nr_to_write <= 0)
 144                         break;
 145                 return -EBUSY;
 146         }
 147
 148         return ret;
 149 }
 150
 151 static void dump_ail_list(struct gfs2_sbd *sdp)
 152 {
 153         struct gfs2_trans *tr;
 154         struct gfs2_bufdata *bd;
 155         struct buffer_head *bh;
 156
 157         list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
 158                 list_for_each_entry_reverse(bd, &tr->tr_ail1_list,
 159                                             bd_ail_st_list) {
 160                         bh = bd->bd_bh;
 161                         fs_err(sdp, "bd %p: blk:0x%llx bh=%p ", bd,
 162                                (unsigned long long)bd->bd_blkno, bh);
 163                         if (!bh) {
 164                                 fs_err(sdp, "\n");
 165                                 continue;
 166                         }
 167                         fs_err(sdp, "0x%llx up2:%d dirt:%d lkd:%d req:%d "
 168                                "map:%d new:%d ar:%d aw:%d delay:%d "
 169                                "io err:%d unwritten:%d dfr:%d pin:%d esc:%d\n",
 170                                (unsigned long long)bh->b_blocknr,
 171                                buffer_uptodate(bh), buffer_dirty(bh),
 172                                buffer_locked(bh), buffer_req(bh),
 173                                buffer_mapped(bh), buffer_new(bh),
 174                                buffer_async_read(bh), buffer_async_write(bh),
 175                                buffer_delay(bh), buffer_write_io_error(bh),
 176                                buffer_unwritten(bh),
 177                                buffer_defer_completion(bh),
 178                                buffer_pinned(bh), buffer_escaped(bh));
 179                 }
 180         }
 181 }
 182
 183 /**
 184  * gfs2_ail1_flush - start writeback of some ail1 entries
 185  * @sdp: The super block
 186  * @wbc: The writeback control structure
 187  *
 188  * Writes back some ail1 entries, according to the limits in the
 189  * writeback control structure
 190  */
 191
 192 void gfs2_ail1_flush(struct gfs2_sbd *sdp, struct writeback_control *wbc)
 193 {
 194         struct list_head *head = &sdp->sd_ail1_list;
 195         struct gfs2_trans *tr;
 196         struct blk_plug plug;
 197         int ret;
 198         unsigned long flush_start = jiffies;
 199
 200         trace_gfs2_ail_flush(sdp, wbc, 1);
 201         blk_start_plug(&plug);
 202         spin_lock(&sdp->sd_ail_lock);
 203 restart:
 204         ret = 0;
 205         if (time_after(jiffies, flush_start + (HZ * 600))) {
 206                 fs_err(sdp, "Error: In %s for ten minutes! t=%d\n",
 207                        __func__, current->journal_info ? 1 : 0);
 208                 dump_ail_list(sdp);
 209                 goto out;
 210         }
 211         list_for_each_entry_reverse(tr, head, tr_list) {
 212                 if (wbc->nr_to_write <= 0)
 213                         break;
 214                 ret = gfs2_ail1_start_one(sdp, wbc, tr, &plug);
 215                 if (ret) {
 216                         if (ret == -EBUSY)
 217                                 goto restart;
 218                         break;
 219                 }
 220         }
 221 out:
 222         spin_unlock(&sdp->sd_ail_lock);
 223         blk_finish_plug(&plug);
 224         if (ret) {
 225                 gfs2_lm(sdp, "gfs2_ail1_start_one (generic_writepages) "
 226                         "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  */
 349
 350 static int gfs2_ail1_empty(struct gfs2_sbd *sdp, int max_revokes)
 351 {
 352         struct gfs2_trans *tr, *s;
 353         int oldest_tr = 1;
 354         int ret;
 355
 356         spin_lock(&sdp->sd_ail_lock);
 357         list_for_each_entry_safe_reverse(tr, s, &sdp->sd_ail1_list, tr_list) {
 358                 if (!gfs2_ail1_empty_one(sdp, tr, &max_revokes) && oldest_tr)
 359                         list_move(&tr->tr_list, &sdp->sd_ail2_list);
 360                 else
 361                         oldest_tr = 0;
 362         }
 363         gfs2_log_update_flush_tail(sdp);
 364         ret = list_empty(&sdp->sd_ail1_list);
 365         spin_unlock(&sdp->sd_ail_lock);
 366
 367         if (test_bit(SDF_WITHDRAWING, &sdp->sd_flags)) {
 368                 gfs2_lm(sdp, "fatal: I/O error(s)\n");
 369                 gfs2_withdraw(sdp);
 370         }
 371
 372         return ret;
 373 }
 374
 375 static void gfs2_ail1_wait(struct gfs2_sbd *sdp)
 376 {
 377         struct gfs2_trans *tr;
 378         struct gfs2_bufdata *bd;
 379         struct buffer_head *bh;
 380
 381         spin_lock(&sdp->sd_ail_lock);
 382         list_for_each_entry_reverse(tr, &sdp->sd_ail1_list, tr_list) {
 383                 list_for_each_entry(bd, &tr->tr_ail1_list, bd_ail_st_list) {
 384                         bh = bd->bd_bh;
 385                         if (!buffer_locked(bh))
 386                                 continue;
 387                         get_bh(bh);
 388                         spin_unlock(&sdp->sd_ail_lock);
 389                         wait_on_buffer(bh);
 390                         brelse(bh);
 391                         return;
 392                 }
 393         }
 394         spin_unlock(&sdp->sd_ail_lock);
 395 }
 396
 397 static void __ail2_empty(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
 398 {
 399         gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
 400         list_del(&tr->tr_list);
 401         gfs2_assert_warn(sdp, list_empty(&tr->tr_ail1_list));
 402         gfs2_assert_warn(sdp, list_empty(&tr->tr_ail2_list));
 403         gfs2_trans_free(sdp, tr);
 404 }
 405
 406 static void ail2_empty(struct gfs2_sbd *sdp, unsigned int new_tail)
 407 {
 408         struct list_head *ail2_list = &sdp->sd_ail2_list;
 409         unsigned int old_tail = sdp->sd_log_tail;
 410         struct gfs2_trans *tr, *safe;
 411
 412         spin_lock(&sdp->sd_ail_lock);
 413         if (old_tail <= new_tail) {
 414                 list_for_each_entry_safe(tr, safe, ail2_list, tr_list) {
 415                         if (old_tail <= tr->tr_first && tr->tr_first < new_tail)
 416                                 __ail2_empty(sdp, tr);
 417                 }
 418         } else {
 419                 list_for_each_entry_safe(tr, safe, ail2_list, tr_list) {
 420                         if (old_tail <= tr->tr_first || tr->tr_first < new_tail)
 421                                 __ail2_empty(sdp, tr);
 422                 }
 423         }
 424         spin_unlock(&sdp->sd_ail_lock);
 425 }
 426
 427 /**
 428  * gfs2_log_is_empty - Check if the log is empty
 429  * @sdp: The GFS2 superblock
 430  */
 431
 432 bool gfs2_log_is_empty(struct gfs2_sbd *sdp) {
 433         return atomic_read(&sdp->sd_log_blks_free) == sdp->sd_jdesc->jd_blocks;
 434 }
 435
 436 static bool __gfs2_log_try_reserve_revokes(struct gfs2_sbd *sdp, unsigned int revokes)
 437 {
 438         unsigned int available;
 439
 440         available = atomic_read(&sdp->sd_log_revokes_available);
 441         while (available >= revokes) {
 442                 if (atomic_try_cmpxchg(&sdp->sd_log_revokes_available,
 443                                        &available, available - revokes))
 444                         return true;
 445         }
 446         return false;
 447 }
 448
 449 /**
 450  * gfs2_log_release_revokes - Release a given number of revokes
 451  * @sdp: The GFS2 superblock
 452  * @revokes: The number of revokes to release
 453  *
 454  * sdp->sd_log_flush_lock must be held.
 455  */
 456 void gfs2_log_release_revokes(struct gfs2_sbd *sdp, unsigned int revokes)
 457 {
 458         if (revokes)
 459                 atomic_add(revokes, &sdp->sd_log_revokes_available);
 460 }
 461
 462 /**
 463  * gfs2_log_release - Release a given number of log blocks
 464  * @sdp: The GFS2 superblock
 465  * @blks: The number of blocks
 466  *
 467  */
 468
 469 void gfs2_log_release(struct gfs2_sbd *sdp, unsigned int blks)
 470 {
 471         atomic_add(blks, &sdp->sd_log_blks_free);
 472         trace_gfs2_log_blocks(sdp, blks);
 473         gfs2_assert_withdraw(sdp, atomic_read(&sdp->sd_log_blks_free) <=
 474                                   sdp->sd_jdesc->jd_blocks);
 475         if (atomic_read(&sdp->sd_log_blks_needed))
 476                 wake_up(&sdp->sd_log_waitq);
 477 }
 478
 479 /**
 480  * __gfs2_log_try_reserve - Try to make a log reservation
 481  * @sdp: The GFS2 superblock
 482  * @blks: The number of blocks to reserve
 483  * @taboo_blks: The number of blocks to leave free
 484  *
 485  * Try to do the same as __gfs2_log_reserve(), but fail if no more log
 486  * space is immediately available.
 487  */
 488 static bool __gfs2_log_try_reserve(struct gfs2_sbd *sdp, unsigned int blks,
 489                                    unsigned int taboo_blks)
 490 {
 491         unsigned wanted = blks + taboo_blks;
 492         unsigned int free_blocks;
 493
 494         free_blocks = atomic_read(&sdp->sd_log_blks_free);
 495         while (free_blocks >= wanted) {
 496                 if (atomic_try_cmpxchg(&sdp->sd_log_blks_free, &free_blocks,
 497                                        free_blocks - blks)) {
 498                         trace_gfs2_log_blocks(sdp, -blks);
 499                         return true;
 500                 }
 501         }
 502         return false;
 503 }
 504
 505 /**
 506  * __gfs2_log_reserve - Make a log reservation
 507  * @sdp: The GFS2 superblock
 508  * @blks: The number of blocks to reserve
 509  * @taboo_blks: The number of blocks to leave free
 510  *
 511  * @taboo_blks is set to 0 for logd, and to GFS2_LOG_FLUSH_MIN_BLOCKS
 512  * for all other processes.  This ensures that when the log is almost full,
 513  * logd will still be able to call gfs2_log_flush one more time  without
 514  * blocking, which will advance the tail and make some more log space
 515  * available.
 516  *
 517  * We no longer flush the log here, instead we wake up logd to do that
 518  * for us. To avoid the thundering herd and to ensure that we deal fairly
 519  * with queued waiters, we use an exclusive wait. This means that when we
 520  * get woken with enough journal space to get our reservation, we need to
 521  * wake the next waiter on the list.
 522  */
 523
 524 static void __gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks,
 525                                unsigned int taboo_blks)
 526 {
 527         unsigned wanted = blks + taboo_blks;
 528         unsigned int free_blocks;
 529
 530         atomic_add(blks, &sdp->sd_log_blks_needed);
 531         for (;;) {
 532                 if (current != sdp->sd_logd_process)
 533                         wake_up(&sdp->sd_logd_waitq);
 534                 io_wait_event(sdp->sd_log_waitq,
 535                         (free_blocks = atomic_read(&sdp->sd_log_blks_free),
 536                          free_blocks >= wanted));
 537                 do {
 538                         if (atomic_try_cmpxchg(&sdp->sd_log_blks_free,
 539                                                &free_blocks,
 540                                                free_blocks - blks))
 541                                 goto reserved;
 542                 } while (free_blocks >= wanted);
 543         }
 544
 545 reserved:
 546         trace_gfs2_log_blocks(sdp, -blks);
 547         if (atomic_sub_return(blks, &sdp->sd_log_blks_needed))
 548                 wake_up(&sdp->sd_log_waitq);
 549 }
 550
 551 /**
 552  * gfs2_log_try_reserve - Try to make a log reservation
 553  * @sdp: The GFS2 superblock
 554  * @tr: The transaction
 555  * @extra_revokes: The number of additional revokes reserved (output)
 556  *
 557  * This is similar to gfs2_log_reserve, but sdp->sd_log_flush_lock must be
 558  * held for correct revoke accounting.
 559  */
 560
 561 bool gfs2_log_try_reserve(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
 562                           unsigned int *extra_revokes)
 563 {
 564         unsigned int blks = tr->tr_reserved;
 565         unsigned int revokes = tr->tr_revokes;
 566         unsigned int revoke_blks = 0;
 567
 568         *extra_revokes = 0;
 569         if (revokes && !__gfs2_log_try_reserve_revokes(sdp, revokes)) {
 570                 revoke_blks = DIV_ROUND_UP(revokes, sdp->sd_inptrs);
 571                 *extra_revokes = revoke_blks * sdp->sd_inptrs - revokes;
 572                 blks += revoke_blks;
 573         }
 574         if (!blks)
 575                 return true;
 576         if (__gfs2_log_try_reserve(sdp, blks, GFS2_LOG_FLUSH_MIN_BLOCKS))
 577                 return true;
 578         if (!revoke_blks)
 579                 gfs2_log_release_revokes(sdp, revokes);
 580         return false;
 581 }
 582
 583 /**
 584  * gfs2_log_reserve - Make a log reservation
 585  * @sdp: The GFS2 superblock
 586  * @tr: The transaction
 587  * @extra_revokes: The number of additional revokes reserved (output)
 588  *
 589  * sdp->sd_log_flush_lock must not be held.
 590  */
 591
 592 void gfs2_log_reserve(struct gfs2_sbd *sdp, struct gfs2_trans *tr,
 593                       unsigned int *extra_revokes)
 594 {
 595         unsigned int blks = tr->tr_reserved;
 596         unsigned int revokes = tr->tr_revokes;
 597         unsigned int revoke_blks;
 598
 599         *extra_revokes = 0;
 600         if (revokes) {
 601                 revoke_blks = DIV_ROUND_UP(revokes, sdp->sd_inptrs);
 602                 *extra_revokes = revoke_blks * sdp->sd_inptrs - revokes;
 603                 blks += revoke_blks;
 604         }
 605         __gfs2_log_reserve(sdp, blks, GFS2_LOG_FLUSH_MIN_BLOCKS);
 606 }
 607
 608 /**
 609  * log_distance - Compute distance between two journal blocks
 610  * @sdp: The GFS2 superblock
 611  * @newer: The most recent journal block of the pair
 612  * @older: The older journal block of the pair
 613  *
 614  *   Compute the distance (in the journal direction) between two
 615  *   blocks in the journal
 616  *
 617  * Returns: the distance in blocks
 618  */
 619
 620 static inline unsigned int log_distance(struct gfs2_sbd *sdp, unsigned int newer,
 621                                         unsigned int older)
 622 {
 623         int dist;
 624
 625         dist = newer - older;
 626         if (dist < 0)
 627                 dist += sdp->sd_jdesc->jd_blocks;
 628
 629         return dist;
 630 }
 631
 632 /**
 633  * calc_reserved - Calculate the number of blocks to keep reserved
 634  * @sdp: The GFS2 superblock
 635  *
 636  * This is complex.  We need to reserve room for all our currently used
 637  * metadata blocks (e.g. normal file I/O rewriting file time stamps) and
 638  * all our journaled data blocks for journaled files (e.g. files in the
 639  * meta_fs like rindex, or files for which chattr +j was done.)
 640  * If we don't reserve enough space, corruption will follow.
 641  *
 642  * We can have metadata blocks and jdata blocks in the same journal.  Each
 643  * type gets its own log descriptor, for which we need to reserve a block.
 644  * In fact, each type has the potential for needing more than one log descriptor
 645  * in cases where we have more blocks than will fit in a log descriptor.
 646  * Metadata journal entries take up half the space of journaled buffer entries.
 647  *
 648  * Also, we need to reserve blocks for revoke journal entries and one for an
 649  * overall header for the lot.
 650  *
 651  * Returns: the number of blocks reserved
 652  */
 653 static unsigned int calc_reserved(struct gfs2_sbd *sdp)
 654 {
 655         unsigned int reserved = GFS2_LOG_FLUSH_MIN_BLOCKS;
 656         unsigned int blocks;
 657         struct gfs2_trans *tr = sdp->sd_log_tr;
 658
 659         if (tr) {
 660                 blocks = tr->tr_num_buf_new - tr->tr_num_buf_rm;
 661                 reserved += blocks + DIV_ROUND_UP(blocks, buf_limit(sdp));
 662                 blocks = tr->tr_num_databuf_new - tr->tr_num_databuf_rm;
 663                 reserved += blocks + DIV_ROUND_UP(blocks, databuf_limit(sdp));
 664         }
 665         return reserved;
 666 }
 667
 668 static void log_pull_tail(struct gfs2_sbd *sdp)
 669 {
 670         unsigned int new_tail = sdp->sd_log_flush_tail;
 671         unsigned int dist;
 672
 673         if (new_tail == sdp->sd_log_tail)
 674                 return;
 675         dist = log_distance(sdp, new_tail, sdp->sd_log_tail);
 676         ail2_empty(sdp, new_tail);
 677         gfs2_log_release(sdp, dist);
 678         sdp->sd_log_tail = new_tail;
 679 }
 680
 681
 682 void log_flush_wait(struct gfs2_sbd *sdp)
 683 {
 684         DEFINE_WAIT(wait);
 685
 686         if (atomic_read(&sdp->sd_log_in_flight)) {
 687                 do {
 688                         prepare_to_wait(&sdp->sd_log_flush_wait, &wait,
 689                                         TASK_UNINTERRUPTIBLE);
 690                         if (atomic_read(&sdp->sd_log_in_flight))
 691                                 io_schedule();
 692                 } while(atomic_read(&sdp->sd_log_in_flight));
 693                 finish_wait(&sdp->sd_log_flush_wait, &wait);
 694         }
 695 }
 696
 697 static int ip_cmp(void *priv, const struct list_head *a, const struct list_head *b)
 698 {
 699         struct gfs2_inode *ipa, *ipb;
 700
 701         ipa = list_entry(a, struct gfs2_inode, i_ordered);
 702         ipb = list_entry(b, struct gfs2_inode, i_ordered);
 703
 704         if (ipa->i_no_addr < ipb->i_no_addr)
 705                 return -1;
 706         if (ipa->i_no_addr > ipb->i_no_addr)
 707                 return 1;
 708         return 0;
 709 }
 710
 711 static void __ordered_del_inode(struct gfs2_inode *ip)
 712 {
 713         if (!list_empty(&ip->i_ordered))
 714                 list_del_init(&ip->i_ordered);
 715 }
 716
 717 static void gfs2_ordered_write(struct gfs2_sbd *sdp)
 718 {
 719         struct gfs2_inode *ip;
 720         LIST_HEAD(written);
 721
 722         spin_lock(&sdp->sd_ordered_lock);
 723         list_sort(NULL, &sdp->sd_log_ordered, &ip_cmp);
 724         while (!list_empty(&sdp->sd_log_ordered)) {
 725                 ip = list_first_entry(&sdp->sd_log_ordered, struct gfs2_inode, i_ordered);
 726                 if (ip->i_inode.i_mapping->nrpages == 0) {
 727                         __ordered_del_inode(ip);
 728                         continue;
 729                 }
 730                 list_move(&ip->i_ordered, &written);
 731                 spin_unlock(&sdp->sd_ordered_lock);
 732                 filemap_fdatawrite(ip->i_inode.i_mapping);
 733                 spin_lock(&sdp->sd_ordered_lock);
 734         }
 735         list_splice(&written, &sdp->sd_log_ordered);
 736         spin_unlock(&sdp->sd_ordered_lock);
 737 }
 738
 739 static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
 740 {
 741         struct gfs2_inode *ip;
 742
 743         spin_lock(&sdp->sd_ordered_lock);
 744         while (!list_empty(&sdp->sd_log_ordered)) {
 745                 ip = list_first_entry(&sdp->sd_log_ordered, struct gfs2_inode, i_ordered);
 746                 __ordered_del_inode(ip);
 747                 if (ip->i_inode.i_mapping->nrpages == 0)
 748                         continue;
 749                 spin_unlock(&sdp->sd_ordered_lock);
 750                 filemap_fdatawait(ip->i_inode.i_mapping);
 751                 spin_lock(&sdp->sd_ordered_lock);
 752         }
 753         spin_unlock(&sdp->sd_ordered_lock);
 754 }
 755
 756 void gfs2_ordered_del_inode(struct gfs2_inode *ip)
 757 {
 758         struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
 759
 760         spin_lock(&sdp->sd_ordered_lock);
 761         __ordered_del_inode(ip);
 762         spin_unlock(&sdp->sd_ordered_lock);
 763 }
 764
 765 void gfs2_add_revoke(struct gfs2_sbd *sdp, struct gfs2_bufdata *bd)
 766 {
 767         struct buffer_head *bh = bd->bd_bh;
 768         struct gfs2_glock *gl = bd->bd_gl;
 769
 770         sdp->sd_log_num_revoke++;
 771         if (atomic_inc_return(&gl->gl_revokes) == 1)
 772                 gfs2_glock_hold(gl);
 773         bh->b_private = NULL;
 774         bd->bd_blkno = bh->b_blocknr;
 775         gfs2_remove_from_ail(bd); /* drops ref on bh */
 776         bd->bd_bh = NULL;
 777         set_bit(GLF_LFLUSH, &gl->gl_flags);
 778         list_add(&bd->bd_list, &sdp->sd_log_revokes);
 779 }
 780
 781 void gfs2_glock_remove_revoke(struct gfs2_glock *gl)
 782 {
 783         if (atomic_dec_return(&gl->gl_revokes) == 0) {
 784                 clear_bit(GLF_LFLUSH, &gl->gl_flags);
 785                 gfs2_glock_queue_put(gl);
 786         }
 787 }
 788
 789 /**
 790  * gfs2_flush_revokes - Add as many revokes to the system transaction as we can
 791  * @sdp: The GFS2 superblock
 792  *
 793  * Our usual strategy is to defer writing revokes as much as we can in the hope
 794  * that we'll eventually overwrite the journal, which will make those revokes
 795  * go away.  This changes when we flush the log: at that point, there will
 796  * likely be some left-over space in the last revoke block of that transaction.
 797  * We can fill that space with additional revokes for blocks that have already
 798  * been written back.  This will basically come at no cost now, and will save
 799  * us from having to keep track of those blocks on the AIL2 list later.
 800  */
 801 void gfs2_flush_revokes(struct gfs2_sbd *sdp)
 802 {
 803         /* number of revokes we still have room for */
 804         unsigned int max_revokes = atomic_read(&sdp->sd_log_revokes_available);
 805
 806         gfs2_log_lock(sdp);
 807         gfs2_ail1_empty(sdp, max_revokes);
 808         gfs2_log_unlock(sdp);
 809 }
 810
 811 /**
 812  * gfs2_write_log_header - Write a journal log header buffer at lblock
 813  * @sdp: The GFS2 superblock
 814  * @jd: journal descriptor of the journal to which we are writing
 815  * @seq: sequence number
 816  * @tail: tail of the log
 817  * @lblock: value for lh_blkno (block number relative to start of journal)
 818  * @flags: log header flags GFS2_LOG_HEAD_*
 819  * @op_flags: flags to pass to the bio
 820  *
 821  * Returns: the initialized log buffer descriptor
 822  */
 823
 824 void gfs2_write_log_header(struct gfs2_sbd *sdp, struct gfs2_jdesc *jd,
 825                            u64 seq, u32 tail, u32 lblock, u32 flags,
 826                            int op_flags)
 827 {
 828         struct gfs2_log_header *lh;
 829         u32 hash, crc;
 830         struct page *page;
 831         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 832         struct timespec64 tv;
 833         struct super_block *sb = sdp->sd_vfs;
 834         u64 dblock;
 835
 836         if (gfs2_withdrawn(sdp))
 837                 return;
 838
 839         page = mempool_alloc(gfs2_page_pool, GFP_NOIO);
 840         lh = page_address(page);
 841         clear_page(lh);
 842
 843         lh->lh_header.mh_magic = cpu_to_be32(GFS2_MAGIC);
 844         lh->lh_header.mh_type = cpu_to_be32(GFS2_METATYPE_LH);
 845         lh->lh_header.__pad0 = cpu_to_be64(0);
 846         lh->lh_header.mh_format = cpu_to_be32(GFS2_FORMAT_LH);
 847         lh->lh_header.mh_jid = cpu_to_be32(sdp->sd_jdesc->jd_jid);
 848         lh->lh_sequence = cpu_to_be64(seq);
 849         lh->lh_flags = cpu_to_be32(flags);
 850         lh->lh_tail = cpu_to_be32(tail);
 851         lh->lh_blkno = cpu_to_be32(lblock);
 852         hash = ~crc32(~0, lh, LH_V1_SIZE);
 853         lh->lh_hash = cpu_to_be32(hash);
 854
 855         ktime_get_coarse_real_ts64(&tv);
 856         lh->lh_nsec = cpu_to_be32(tv.tv_nsec);
 857         lh->lh_sec = cpu_to_be64(tv.tv_sec);
 858         if (!list_empty(&jd->extent_list))
 859                 dblock = gfs2_log_bmap(jd, lblock);
 860         else {
 861                 unsigned int extlen;
 862                 int ret;
 863
 864                 extlen = 1;
 865                 ret = gfs2_get_extent(jd->jd_inode, lblock, &dblock, &extlen);
 866                 if (gfs2_assert_withdraw(sdp, ret == 0))
 867                         return;
 868         }
 869         lh->lh_addr = cpu_to_be64(dblock);
 870         lh->lh_jinode = cpu_to_be64(GFS2_I(jd->jd_inode)->i_no_addr);
 871
 872         /* We may only write local statfs, quota, etc., when writing to our
 873            own journal. The values are left 0 when recovering a journal
 874            different from our own. */
 875         if (!(flags & GFS2_LOG_HEAD_RECOVERY)) {
 876                 lh->lh_statfs_addr =
 877                         cpu_to_be64(GFS2_I(sdp->sd_sc_inode)->i_no_addr);
 878                 lh->lh_quota_addr =
 879                         cpu_to_be64(GFS2_I(sdp->sd_qc_inode)->i_no_addr);
 880
 881                 spin_lock(&sdp->sd_statfs_spin);
 882                 lh->lh_local_total = cpu_to_be64(l_sc->sc_total);
 883                 lh->lh_local_free = cpu_to_be64(l_sc->sc_free);
 884                 lh->lh_local_dinodes = cpu_to_be64(l_sc->sc_dinodes);
 885                 spin_unlock(&sdp->sd_statfs_spin);
 886         }
 887
 888         BUILD_BUG_ON(offsetof(struct gfs2_log_header, lh_crc) != LH_V1_SIZE);
 889
 890         crc = crc32c(~0, (void *)lh + LH_V1_SIZE + 4,
 891                      sb->s_blocksize - LH_V1_SIZE - 4);
 892         lh->lh_crc = cpu_to_be32(crc);
 893
 894         gfs2_log_write(sdp, jd, page, sb->s_blocksize, 0, dblock);
 895         gfs2_log_submit_bio(&jd->jd_log_bio, REQ_OP_WRITE | op_flags);
 896 }
 897
 898 /**
 899  * log_write_header - Get and initialize a journal header buffer
 900  * @sdp: The GFS2 superblock
 901  * @flags: The log header flags, including log header origin
 902  *
 903  * Returns: the initialized log buffer descriptor
 904  */
 905
 906 static void log_write_header(struct gfs2_sbd *sdp, u32 flags)
 907 {
 908         int op_flags = REQ_PREFLUSH | REQ_FUA | REQ_META | REQ_SYNC;
 909         enum gfs2_freeze_state state = atomic_read(&sdp->sd_freeze_state);
 910
 911         gfs2_assert_withdraw(sdp, (state != SFS_FROZEN));
 912
 913         if (test_bit(SDF_NOBARRIERS, &sdp->sd_flags)) {
 914                 gfs2_ordered_wait(sdp);
 915                 log_flush_wait(sdp);
 916                 op_flags = REQ_SYNC | REQ_META | REQ_PRIO;
 917         }
 918         sdp->sd_log_idle = (sdp->sd_log_flush_tail == sdp->sd_log_flush_head);
 919         gfs2_write_log_header(sdp, sdp->sd_jdesc, sdp->sd_log_sequence++,
 920                               sdp->sd_log_flush_tail, sdp->sd_log_flush_head,
 921                               flags, op_flags);
 922         gfs2_log_incr_head(sdp);
 923         log_flush_wait(sdp);
 924         log_pull_tail(sdp);
 925         gfs2_log_update_head(sdp);
 926 }
 927
 928 /**
 929  * gfs2_ail_drain - drain the ail lists after a withdraw
 930  * @sdp: Pointer to GFS2 superblock
 931  */
 932 void gfs2_ail_drain(struct gfs2_sbd *sdp)
 933 {
 934         struct gfs2_trans *tr;
 935
 936         spin_lock(&sdp->sd_ail_lock);
 937         /*
 938          * For transactions on the sd_ail1_list we need to drain both the
 939          * ail1 and ail2 lists. That's because function gfs2_ail1_start_one
 940          * (temporarily) moves items from its tr_ail1 list to tr_ail2 list
 941          * before revokes are sent for that block. Items on the sd_ail2_list
 942          * should have already gotten beyond that point, so no need.
 943          */
 944         while (!list_empty(&sdp->sd_ail1_list)) {
 945                 tr = list_first_entry(&sdp->sd_ail1_list, struct gfs2_trans,
 946                                       tr_list);
 947                 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail1_list);
 948                 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
 949                 list_del(&tr->tr_list);
 950                 gfs2_trans_free(sdp, tr);
 951         }
 952         while (!list_empty(&sdp->sd_ail2_list)) {
 953                 tr = list_first_entry(&sdp->sd_ail2_list, struct gfs2_trans,
 954                                       tr_list);
 955                 gfs2_ail_empty_tr(sdp, tr, &tr->tr_ail2_list);
 956                 list_del(&tr->tr_list);
 957                 gfs2_trans_free(sdp, tr);
 958         }
 959         gfs2_drain_revokes(sdp);
 960         spin_unlock(&sdp->sd_ail_lock);
 961 }
 962
 963 /**
 964  * empty_ail1_list - try to start IO and empty the ail1 list
 965  * @sdp: Pointer to GFS2 superblock
 966  */
 967 static void empty_ail1_list(struct gfs2_sbd *sdp)
 968 {
 969         unsigned long start = jiffies;
 970
 971         for (;;) {
 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                 if (gfs2_ail1_empty(sdp, 0))
 981                         return;
 982         }
 983 }
 984
 985 /**
 986  * trans_drain - drain the buf and databuf queue for a failed transaction
 987  * @tr: the transaction to drain
 988  *
 989  * When this is called, we're taking an error exit for a log write that failed
 990  * but since we bypassed the after_commit functions, we need to remove the
 991  * items from the buf and databuf queue.
 992  */
 993 static void trans_drain(struct gfs2_trans *tr)
 994 {
 995         struct gfs2_bufdata *bd;
 996         struct list_head *head;
 997
 998         if (!tr)
 999                 return;
1000
1001         head = &tr->tr_buf;
1002         while (!list_empty(head)) {
1003                 bd = list_first_entry(head, struct gfs2_bufdata, bd_list);
1004                 list_del_init(&bd->bd_list);
1005                 if (!list_empty(&bd->bd_ail_st_list))
1006                         gfs2_remove_from_ail(bd);
1007                 kmem_cache_free(gfs2_bufdata_cachep, bd);
1008         }
1009         head = &tr->tr_databuf;
1010         while (!list_empty(head)) {
1011                 bd = list_first_entry(head, struct gfs2_bufdata, bd_list);
1012                 list_del_init(&bd->bd_list);
1013                 if (!list_empty(&bd->bd_ail_st_list))
1014                         gfs2_remove_from_ail(bd);
1015                 kmem_cache_free(gfs2_bufdata_cachep, bd);
1016         }
1017 }
1018
1019 /**
1020  * gfs2_log_flush - flush incore transaction(s)
1021  * @sdp: The filesystem
1022  * @gl: The glock structure to flush.  If NULL, flush the whole incore log
1023  * @flags: The log header flags: GFS2_LOG_HEAD_FLUSH_* and debug flags
1024  *
1025  */
1026
1027 void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl, u32 flags)
1028 {
1029         struct gfs2_trans *tr = NULL;
1030         unsigned int reserved_blocks = 0, used_blocks = 0;
1031         enum gfs2_freeze_state state = atomic_read(&sdp->sd_freeze_state);
1032         unsigned int first_log_head;
1033         unsigned int reserved_revokes = 0;
1034
1035         down_write(&sdp->sd_log_flush_lock);
1036         trace_gfs2_log_flush(sdp, 1, flags);
1037
1038 repeat:
1039         /*
1040          * Do this check while holding the log_flush_lock to prevent new
1041          * buffers from being added to the ail via gfs2_pin()
1042          */
1043         if (gfs2_withdrawn(sdp) || !test_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags))
1044                 goto out;
1045
1046         /* Log might have been flushed while we waited for the flush lock */
1047         if (gl && !test_bit(GLF_LFLUSH, &gl->gl_flags))
1048                 goto out;
1049
1050         first_log_head = sdp->sd_log_head;
1051         sdp->sd_log_flush_head = first_log_head;
1052
1053         tr = sdp->sd_log_tr;
1054         if (tr || sdp->sd_log_num_revoke) {
1055                 if (reserved_blocks)
1056                         gfs2_log_release(sdp, reserved_blocks);
1057                 reserved_blocks = sdp->sd_log_blks_reserved;
1058                 reserved_revokes = sdp->sd_log_num_revoke;
1059                 if (tr) {
1060                         sdp->sd_log_tr = NULL;
1061                         tr->tr_first = first_log_head;
1062                         if (unlikely (state == SFS_FROZEN)) {
1063                                 if (gfs2_assert_withdraw_delayed(sdp,
1064                                        !tr->tr_num_buf_new && !tr->tr_num_databuf_new))
1065                                         goto out_withdraw;
1066                         }
1067                 }
1068         } else if (!reserved_blocks) {
1069                 unsigned int taboo_blocks = GFS2_LOG_FLUSH_MIN_BLOCKS;
1070
1071                 reserved_blocks = GFS2_LOG_FLUSH_MIN_BLOCKS;
1072                 if (current == sdp->sd_logd_process)
1073                         taboo_blocks = 0;
1074
1075                 if (!__gfs2_log_try_reserve(sdp, reserved_blocks, taboo_blocks)) {
1076                         up_write(&sdp->sd_log_flush_lock);
1077                         __gfs2_log_reserve(sdp, reserved_blocks, taboo_blocks);
1078                         down_write(&sdp->sd_log_flush_lock);
1079                         goto repeat;
1080                 }
1081                 BUG_ON(sdp->sd_log_num_revoke);
1082         }
1083
1084         if (flags & GFS2_LOG_HEAD_FLUSH_SHUTDOWN)
1085                 clear_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags);
1086
1087         if (unlikely(state == SFS_FROZEN))
1088                 if (gfs2_assert_withdraw_delayed(sdp, !reserved_revokes))
1089                         goto out_withdraw;
1090
1091         gfs2_ordered_write(sdp);
1092         if (gfs2_withdrawn(sdp))
1093                 goto out_withdraw;
1094         lops_before_commit(sdp, tr);
1095         if (gfs2_withdrawn(sdp))
1096                 goto out_withdraw;
1097         gfs2_log_submit_bio(&sdp->sd_jdesc->jd_log_bio, REQ_OP_WRITE);
1098         if (gfs2_withdrawn(sdp))
1099                 goto out_withdraw;
1100
1101         if (sdp->sd_log_head != sdp->sd_log_flush_head) {
1102                 log_write_header(sdp, flags);
1103         } else if (sdp->sd_log_tail != sdp->sd_log_flush_tail && !sdp->sd_log_idle) {
1104                 log_write_header(sdp, flags);
1105         }
1106         if (gfs2_withdrawn(sdp))
1107                 goto out_withdraw;
1108         lops_after_commit(sdp, tr);
1109
1110         gfs2_log_lock(sdp);
1111         sdp->sd_log_blks_reserved = 0;
1112
1113         spin_lock(&sdp->sd_ail_lock);
1114         if (tr && !list_empty(&tr->tr_ail1_list)) {
1115                 list_add(&tr->tr_list, &sdp->sd_ail1_list);
1116                 tr = NULL;
1117         }
1118         spin_unlock(&sdp->sd_ail_lock);
1119         gfs2_log_unlock(sdp);
1120
1121         if (!(flags & GFS2_LOG_HEAD_FLUSH_NORMAL)) {
1122                 if (!sdp->sd_log_idle) {
1123                         empty_ail1_list(sdp);
1124                         if (gfs2_withdrawn(sdp))
1125                                 goto out_withdraw;
1126                         log_write_header(sdp, flags);
1127                 }
1128                 if (flags & (GFS2_LOG_HEAD_FLUSH_SHUTDOWN |
1129                              GFS2_LOG_HEAD_FLUSH_FREEZE))
1130                         gfs2_log_shutdown(sdp);
1131                 if (flags & GFS2_LOG_HEAD_FLUSH_FREEZE)
1132                         atomic_set(&sdp->sd_freeze_state, SFS_FROZEN);
1133         }
1134
1135 out_end:
1136         used_blocks = log_distance(sdp, sdp->sd_log_flush_head, first_log_head);
1137         reserved_revokes += atomic_read(&sdp->sd_log_revokes_available);
1138         atomic_set(&sdp->sd_log_revokes_available, sdp->sd_ldptrs);
1139         gfs2_assert_withdraw(sdp, reserved_revokes % sdp->sd_inptrs == sdp->sd_ldptrs);
1140         if (reserved_revokes > sdp->sd_ldptrs)
1141                 reserved_blocks += (reserved_revokes - sdp->sd_ldptrs) / sdp->sd_inptrs;
1142 out:
1143         if (used_blocks != reserved_blocks) {
1144                 gfs2_assert_withdraw_delayed(sdp, used_blocks < reserved_blocks);
1145                 gfs2_log_release(sdp, reserved_blocks - used_blocks);
1146         }
1147         up_write(&sdp->sd_log_flush_lock);
1148         gfs2_trans_free(sdp, tr);
1149         if (gfs2_withdrawing(sdp))
1150                 gfs2_withdraw(sdp);
1151         trace_gfs2_log_flush(sdp, 0, flags);
1152         return;
1153
1154 out_withdraw:
1155         trans_drain(tr);
1156         /**
1157          * If the tr_list is empty, we're withdrawing during a log
1158          * flush that targets a transaction, but the transaction was
1159          * never queued onto any of the ail lists. Here we add it to
1160          * ail1 just so that ail_drain() will find and free it.
1161          */
1162         spin_lock(&sdp->sd_ail_lock);
1163         if (tr && list_empty(&tr->tr_list))
1164                 list_add(&tr->tr_list, &sdp->sd_ail1_list);
1165         spin_unlock(&sdp->sd_ail_lock);
1166         tr = NULL;
1167         goto out_end;
1168 }
1169
1170 /**
1171  * gfs2_merge_trans - Merge a new transaction into a cached transaction
1172  * @sdp: the filesystem
1173  * @new: New transaction to be merged
1174  */
1175
1176 static void gfs2_merge_trans(struct gfs2_sbd *sdp, struct gfs2_trans *new)
1177 {
1178         struct gfs2_trans *old = sdp->sd_log_tr;
1179
1180         WARN_ON_ONCE(!test_bit(TR_ATTACHED, &old->tr_flags));
1181
1182         old->tr_num_buf_new     += new->tr_num_buf_new;
1183         old->tr_num_databuf_new += new->tr_num_databuf_new;
1184         old->tr_num_buf_rm      += new->tr_num_buf_rm;
1185         old->tr_num_databuf_rm  += new->tr_num_databuf_rm;
1186         old->tr_revokes         += new->tr_revokes;
1187         old->tr_num_revoke      += new->tr_num_revoke;
1188
1189         list_splice_tail_init(&new->tr_databuf, &old->tr_databuf);
1190         list_splice_tail_init(&new->tr_buf, &old->tr_buf);
1191
1192         spin_lock(&sdp->sd_ail_lock);
1193         list_splice_tail_init(&new->tr_ail1_list, &old->tr_ail1_list);
1194         list_splice_tail_init(&new->tr_ail2_list, &old->tr_ail2_list);
1195         spin_unlock(&sdp->sd_ail_lock);
1196 }
1197
1198 static void log_refund(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
1199 {
1200         unsigned int reserved;
1201         unsigned int unused;
1202         unsigned int maxres;
1203
1204         gfs2_log_lock(sdp);
1205
1206         if (sdp->sd_log_tr) {
1207                 gfs2_merge_trans(sdp, tr);
1208         } else if (tr->tr_num_buf_new || tr->tr_num_databuf_new) {
1209                 gfs2_assert_withdraw(sdp, !test_bit(TR_ONSTACK, &tr->tr_flags));
1210                 sdp->sd_log_tr = tr;
1211                 set_bit(TR_ATTACHED, &tr->tr_flags);
1212         }
1213
1214         reserved = calc_reserved(sdp);
1215         maxres = sdp->sd_log_blks_reserved + tr->tr_reserved;
1216         gfs2_assert_withdraw(sdp, maxres >= reserved);
1217         unused = maxres - reserved;
1218         if (unused)
1219                 gfs2_log_release(sdp, unused);
1220         sdp->sd_log_blks_reserved = reserved;
1221
1222         gfs2_log_unlock(sdp);
1223 }
1224
1225 /**
1226  * gfs2_log_commit - Commit a transaction to the log
1227  * @sdp: the filesystem
1228  * @tr: the transaction
1229  *
1230  * We wake up gfs2_logd if the number of pinned blocks exceed thresh1
1231  * or the total number of used blocks (pinned blocks plus AIL blocks)
1232  * is greater than thresh2.
1233  *
1234  * At mount time thresh1 is 2/5ths of journal size, thresh2 is 4/5ths of
1235  * journal size.
1236  *
1237  * Returns: errno
1238  */
1239
1240 void gfs2_log_commit(struct gfs2_sbd *sdp, struct gfs2_trans *tr)
1241 {
1242         log_refund(sdp, tr);
1243
1244         if (atomic_read(&sdp->sd_log_pinned) > atomic_read(&sdp->sd_log_thresh1) ||
1245             ((sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free)) >
1246             atomic_read(&sdp->sd_log_thresh2)))
1247                 wake_up(&sdp->sd_logd_waitq);
1248 }
1249
1250 /**
1251  * gfs2_log_shutdown - write a shutdown header into a journal
1252  * @sdp: the filesystem
1253  *
1254  */
1255
1256 static void gfs2_log_shutdown(struct gfs2_sbd *sdp)
1257 {
1258         gfs2_assert_withdraw(sdp, !sdp->sd_log_blks_reserved);
1259         gfs2_assert_withdraw(sdp, !sdp->sd_log_num_revoke);
1260         gfs2_assert_withdraw(sdp, list_empty(&sdp->sd_ail1_list));
1261
1262         log_write_header(sdp, GFS2_LOG_HEAD_UNMOUNT | GFS2_LFC_SHUTDOWN);
1263         log_pull_tail(sdp);
1264
1265         gfs2_assert_warn(sdp, sdp->sd_log_head == sdp->sd_log_tail);
1266         gfs2_assert_warn(sdp, list_empty(&sdp->sd_ail2_list));
1267 }
1268
1269 static inline int gfs2_jrnl_flush_reqd(struct gfs2_sbd *sdp)
1270 {
1271         return (atomic_read(&sdp->sd_log_pinned) +
1272                 atomic_read(&sdp->sd_log_blks_needed) >=
1273                 atomic_read(&sdp->sd_log_thresh1));
1274 }
1275
1276 static inline int gfs2_ail_flush_reqd(struct gfs2_sbd *sdp)
1277 {
1278         unsigned int used_blocks = sdp->sd_jdesc->jd_blocks - atomic_read(&sdp->sd_log_blks_free);
1279
1280         if (test_and_clear_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags))
1281                 return 1;
1282
1283         return used_blocks + atomic_read(&sdp->sd_log_blks_needed) >=
1284                 atomic_read(&sdp->sd_log_thresh2);
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         DEFINE_WAIT(wait);
1300
1301         while (!kthread_should_stop()) {
1302
1303                 if (gfs2_withdrawn(sdp)) {
1304                         msleep_interruptible(HZ);
1305                         continue;
1306                 }
1307                 /* Check for errors writing to the journal */
1308                 if (sdp->sd_log_error) {
1309                         gfs2_lm(sdp,
1310                                 "GFS2: fsid=%s: error %d: "
1311                                 "withdrawing the file system to "
1312                                 "prevent further damage.\n",
1313                                 sdp->sd_fsname, sdp->sd_log_error);
1314                         gfs2_withdraw(sdp);
1315                         continue;
1316                 }
1317
1318                 if (gfs2_jrnl_flush_reqd(sdp) || t == 0) {
1319                         gfs2_ail1_empty(sdp, 0);
1320                         gfs2_log_flush(sdp, NULL, GFS2_LOG_HEAD_FLUSH_NORMAL |
1321                                                   GFS2_LFC_LOGD_JFLUSH_REQD);
1322                 }
1323
1324                 if (gfs2_ail_flush_reqd(sdp)) {
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                 try_to_freeze();
1335
1336                 do {
1337                         prepare_to_wait(&sdp->sd_logd_waitq, &wait,
1338                                         TASK_INTERRUPTIBLE);
1339                         if (!gfs2_ail_flush_reqd(sdp) &&
1340                             !gfs2_jrnl_flush_reqd(sdp) &&
1341                             !kthread_should_stop())
1342                                 t = schedule_timeout(t);
1343                 } while(t && !gfs2_ail_flush_reqd(sdp) &&
1344                         !gfs2_jrnl_flush_reqd(sdp) &&
1345                         !kthread_should_stop());
1346                 finish_wait(&sdp->sd_logd_waitq, &wait);
1347         }
1348
1349         return 0;
1350 }
1351