fs/jbd/checkpoint.c

   1 /*
   2  * linux/fs/checkpoint.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
   5  *
   6  * Copyright 1999 Red Hat Software --- All Rights Reserved
   7  *
   8  * This file is part of the Linux kernel and is made available under
   9  * the terms of the GNU General Public License, version 2, or at your
  10  * option, any later version, incorporated herein by reference.
  11  *
  12  * Checkpoint routines for the generic filesystem journaling code.
  13  * Part of the ext2fs journaling system.
  14  *
  15  * Checkpointing is the process of ensuring that a section of the log is
  16  * committed fully to disk, so that that portion of the log can be
  17  * reused.
  18  */
  19
  20 #include <linux/sched.h>
  21 #include <linux/fs.h>
  22 #include <linux/jbd.h>
  23 #include <linux/errno.h>
  24 #include <linux/slab.h>
  25 #include <linux/locks.h>
  26
  27 extern spinlock_t journal_datalist_lock;
  28
  29 /*
  30  * Unlink a buffer from a transaction.
  31  *
  32  * Called with journal_datalist_lock held.
  33  */
  34
  35 static inline void __buffer_unlink(struct journal_head *jh)
  36 {
  37         transaction_t *transaction;
  38
  39         transaction = jh->b_cp_transaction;
  40         jh->b_cp_transaction = NULL;
  41
  42         jh->b_cpnext->b_cpprev = jh->b_cpprev;
  43         jh->b_cpprev->b_cpnext = jh->b_cpnext;
  44         if (transaction->t_checkpoint_list == jh)
  45                 transaction->t_checkpoint_list = jh->b_cpnext;
  46         if (transaction->t_checkpoint_list == jh)
  47                 transaction->t_checkpoint_list = NULL;
  48 }
  49
  50 /*
  51  * Try to release a checkpointed buffer from its transaction.
  52  * Returns 1 if we released it.
  53  * Requires journal_datalist_lock
  54  */
  55 static int __try_to_free_cp_buf(struct journal_head *jh)
  56 {
  57         int ret = 0;
  58         struct buffer_head *bh = jh2bh(jh);
  59
  60         if (jh->b_jlist == BJ_None && !buffer_locked(bh) && !buffer_dirty(bh)) {
  61                 JBUFFER_TRACE(jh, "remove from checkpoint list");
  62                 __journal_remove_checkpoint(jh);
  63                 __journal_remove_journal_head(bh);
  64                 BUFFER_TRACE(bh, "release");
  65                 /* BUF_LOCKED -> BUF_CLEAN (fwiw) */
  66                 refile_buffer(bh);
  67                 __brelse(bh);
  68                 ret = 1;
  69         }
  70         return ret;
  71 }
  72
  73 /*
  74  * log_wait_for_space: wait until there is space in the journal.
  75  *
  76  * Called with the journal already locked, but it will be unlocked if we have
  77  * to wait for a checkpoint to free up some space in the log.
  78  */
  79
  80 void log_wait_for_space(journal_t *journal, int nblocks)
  81 {
  82         while (log_space_left(journal) < nblocks) {
  83                 if (journal->j_flags & JFS_ABORT)
  84                         return;
  85                 unlock_journal(journal);
  86                 down(&journal->j_checkpoint_sem);
  87                 lock_journal(journal);
  88
  89                 /* Test again, another process may have checkpointed
  90                  * while we were waiting for the checkpoint lock */
  91                 if (log_space_left(journal) < nblocks) {
  92                         log_do_checkpoint(journal, nblocks);
  93                 }
  94                 up(&journal->j_checkpoint_sem);
  95         }
  96 }
  97
  98 /*
  99  * Clean up a transaction's checkpoint list.
 100  *
 101  * We wait for any pending IO to complete and make sure any clean
 102  * buffers are removed from the transaction.
 103  *
 104  * Return 1 if we performed any actions which might have destroyed the
 105  * checkpoint.  (journal_remove_checkpoint() deletes the transaction when
 106  * the last checkpoint buffer is cleansed)
 107  *
 108  * Called with the journal locked.
 109  * Called with journal_datalist_lock held.
 110  */
 111 static int __cleanup_transaction(journal_t *journal, transaction_t *transaction)
 112 {
 113         struct journal_head *jh, *next_jh, *last_jh;
 114         struct buffer_head *bh;
 115         int ret = 0;
 116
 117         assert_spin_locked(&journal_datalist_lock);
 118         jh = transaction->t_checkpoint_list;
 119         if (!jh)
 120                 return 0;
 121
 122         last_jh = jh->b_cpprev;
 123         next_jh = jh;
 124         do {
 125                 jh = next_jh;
 126                 bh = jh2bh(jh);
 127                 if (buffer_locked(bh)) {
 128                         atomic_inc(&bh->b_count);
 129                         spin_unlock(&journal_datalist_lock);
 130                         unlock_journal(journal);
 131                         wait_on_buffer(bh);
 132                         /* the journal_head may have gone by now */
 133                         BUFFER_TRACE(bh, "brelse");
 134                         __brelse(bh);
 135                         goto out_return_1;
 136                 }
 137
 138                 if (jh->b_transaction != NULL) {
 139                         transaction_t *transaction = jh->b_transaction;
 140                         tid_t tid = transaction->t_tid;
 141
 142                         spin_unlock(&journal_datalist_lock);
 143                         log_start_commit(journal, transaction);
 144                         unlock_journal(journal);
 145                         log_wait_commit(journal, tid);
 146                         goto out_return_1;
 147                 }
 148
 149                 /*
 150                  * We used to test for (jh->b_list != BUF_CLEAN) here.
 151                  * But unmap_underlying_metadata() can place buffer onto
 152                  * BUF_CLEAN. Since refile_buffer() no longer takes buffers
 153                  * off checkpoint lists, we cope with it here
 154                  */
 155                 /*
 156                  * AKPM: I think the buffer_jdirty test is redundant - it
 157                  * shouldn't have NULL b_transaction?
 158                  */
 159                 next_jh = jh->b_cpnext;
 160                 if (!buffer_dirty(bh) && !buffer_jdirty(bh)) {
 161                         BUFFER_TRACE(bh, "remove from checkpoint");
 162                         __journal_remove_checkpoint(jh);
 163                         __journal_remove_journal_head(bh);
 164                         refile_buffer(bh);
 165                         __brelse(bh);
 166                         ret = 1;
 167                 }
 168
 169                 jh = next_jh;
 170         } while (jh != last_jh);
 171
 172         return ret;
 173 out_return_1:
 174         lock_journal(journal);
 175         spin_lock(&journal_datalist_lock);
 176         return 1;
 177 }
 178
 179 #define NR_BATCH        64
 180
 181 static void __flush_batch(struct buffer_head **bhs, int *batch_count)
 182 {
 183         int i;
 184
 185         spin_unlock(&journal_datalist_lock);
 186         ll_rw_block(WRITE, *batch_count, bhs);
 187         run_task_queue(&tq_disk);
 188         spin_lock(&journal_datalist_lock);
 189         for (i = 0; i < *batch_count; i++) {
 190                 struct buffer_head *bh = bhs[i];
 191                 clear_bit(BH_JWrite, &bh->b_state);
 192                 BUFFER_TRACE(bh, "brelse");
 193                 __brelse(bh);
 194         }
 195         *batch_count = 0;
 196 }
 197
 198 /*
 199  * Try to flush one buffer from the checkpoint list to disk.
 200  *
 201  * Return 1 if something happened which requires us to abort the current
 202  * scan of the checkpoint list.
 203  *
 204  * Called with journal_datalist_lock held.
 205  */
 206 static int __flush_buffer(journal_t *journal, struct journal_head *jh,
 207                         struct buffer_head **bhs, int *batch_count,
 208                         int *drop_count)
 209 {
 210         struct buffer_head *bh = jh2bh(jh);
 211         int ret = 0;
 212
 213         if (buffer_dirty(bh) && !buffer_locked(bh) && jh->b_jlist == BJ_None) {
 214                 J_ASSERT_JH(jh, jh->b_transaction == NULL);
 215
 216                 /*
 217                  * Important: we are about to write the buffer, and
 218                  * possibly block, while still holding the journal lock.
 219                  * We cannot afford to let the transaction logic start
 220                  * messing around with this buffer before we write it to
 221                  * disk, as that would break recoverability.
 222                  */
 223                 BUFFER_TRACE(bh, "queue");
 224                 atomic_inc(&bh->b_count);
 225                 J_ASSERT_BH(bh, !test_bit(BH_JWrite, &bh->b_state));
 226                 set_bit(BH_JWrite, &bh->b_state);
 227                 bhs[*batch_count] = bh;
 228                 (*batch_count)++;
 229                 if (*batch_count == NR_BATCH) {
 230                         __flush_batch(bhs, batch_count);
 231                         ret = 1;
 232                 }
 233         } else {
 234                 int last_buffer = 0;
 235                 if (jh->b_cpnext == jh) {
 236                         /* We may be about to drop the transaction.  Tell the
 237                          * caller that the lists have changed.
 238                          */
 239                         last_buffer = 1;
 240                 }
 241                 if (__try_to_free_cp_buf(jh)) {
 242                         (*drop_count)++;
 243                         ret = last_buffer;
 244                 }
 245         }
 246         return ret;
 247 }
 248
 249
 250 /*
 251  * Perform an actual checkpoint.  We don't write out only enough to
 252  * satisfy the current blocked requests: rather we submit a reasonably
 253  * sized chunk of the outstanding data to disk at once for
 254  * efficiency.  log_wait_for_space() will retry if we didn't free enough.
 255  *
 256  * However, we _do_ take into account the amount requested so that once
 257  * the IO has been queued, we can return as soon as enough of it has
 258  * completed to disk.
 259  *
 260  * The journal should be locked before calling this function.
 261  */
 262
 263 /* @@@ `nblocks' is unused.  Should it be used? */
 264 int log_do_checkpoint (journal_t *journal, int nblocks)
 265 {
 266         transaction_t *transaction, *last_transaction, *next_transaction;
 267         int result;
 268         int target;
 269         int batch_count = 0;
 270         struct buffer_head *bhs[NR_BATCH];
 271
 272         jbd_debug(1, "Start checkpoint\n");
 273
 274         /*
 275          * First thing: if there are any transactions in the log which
 276          * don't need checkpointing, just eliminate them from the
 277          * journal straight away.
 278          */
 279         result = cleanup_journal_tail(journal);
 280         jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
 281         if (result <= 0)
 282                 return result;
 283
 284         /*
 285          * OK, we need to start writing disk blocks.  Try to free up a
 286          * quarter of the log in a single checkpoint if we can.
 287          */
 288         /*
 289          * AKPM: check this code.  I had a feeling a while back that it
 290          * degenerates into a busy loop at unmount time.
 291          */
 292         target = (journal->j_last - journal->j_first) / 4;
 293
 294         spin_lock(&journal_datalist_lock);
 295 repeat:
 296         transaction = journal->j_checkpoint_transactions;
 297         if (transaction == NULL)
 298                 goto done;
 299         last_transaction = transaction->t_cpprev;
 300         next_transaction = transaction;
 301
 302         do {
 303                 struct journal_head *jh, *last_jh, *next_jh;
 304                 int drop_count = 0;
 305                 int cleanup_ret, retry = 0;
 306
 307                 transaction = next_transaction;
 308                 next_transaction = transaction->t_cpnext;
 309                 jh = transaction->t_checkpoint_list;
 310                 last_jh = jh->b_cpprev;
 311                 next_jh = jh;
 312                 do {
 313                         jh = next_jh;
 314                         next_jh = jh->b_cpnext;
 315                         retry = __flush_buffer(journal, jh, bhs, &batch_count,
 316                                                 &drop_count);
 317                 } while (jh != last_jh && !retry);
 318                 if (batch_count) {
 319                         __flush_batch(bhs, &batch_count);
 320                         goto repeat;
 321                 }
 322                 if (retry)
 323                         goto repeat;
 324                 /*
 325                  * We have walked the whole transaction list without
 326                  * finding anything to write to disk.  We had better be
 327                  * able to make some progress or we are in trouble.
 328                  */
 329                 cleanup_ret = __cleanup_transaction(journal, transaction);
 330                 J_ASSERT(drop_count != 0 || cleanup_ret != 0);
 331                 goto repeat;    /* __cleanup may have dropped lock */
 332         } while (transaction != last_transaction);
 333
 334 done:
 335         spin_unlock(&journal_datalist_lock);
 336         result = cleanup_journal_tail(journal);
 337         if (result < 0)
 338                 return result;
 339
 340         return 0;
 341 }
 342
 343 /*
 344  * Check the list of checkpoint transactions for the journal to see if
 345  * we have already got rid of any since the last update of the log tail
 346  * in the journal superblock.  If so, we can instantly roll the
 347  * superblock forward to remove those transactions from the log.
 348  *
 349  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
 350  *
 351  * Called with the journal lock held.
 352  *
 353  * This is the only part of the journaling code which really needs to be
 354  * aware of transaction aborts.  Checkpointing involves writing to the
 355  * main filesystem area rather than to the journal, so it can proceed
 356  * even in abort state, but we must not update the journal superblock if
 357  * we have an abort error outstanding.
 358  */
 359
 360 int cleanup_journal_tail(journal_t *journal)
 361 {
 362         transaction_t * transaction;
 363         tid_t           first_tid;
 364         unsigned long   blocknr, freed;
 365
 366         /* OK, work out the oldest transaction remaining in the log, and
 367          * the log block it starts at.
 368          *
 369          * If the log is now empty, we need to work out which is the
 370          * next transaction ID we will write, and where it will
 371          * start. */
 372
 373         /* j_checkpoint_transactions needs locking */
 374         spin_lock(&journal_datalist_lock);
 375         transaction = journal->j_checkpoint_transactions;
 376         if (transaction) {
 377                 first_tid = transaction->t_tid;
 378                 blocknr = transaction->t_log_start;
 379         } else if ((transaction = journal->j_committing_transaction) != NULL) {
 380                 first_tid = transaction->t_tid;
 381                 blocknr = transaction->t_log_start;
 382         } else if ((transaction = journal->j_running_transaction) != NULL) {
 383                 first_tid = transaction->t_tid;
 384                 blocknr = journal->j_head;
 385         } else {
 386                 first_tid = journal->j_transaction_sequence;
 387                 blocknr = journal->j_head;
 388         }
 389         spin_unlock(&journal_datalist_lock);
 390         J_ASSERT (blocknr != 0);
 391
 392         /* If the oldest pinned transaction is at the tail of the log
 393            already then there's not much we can do right now. */
 394         if (journal->j_tail_sequence == first_tid)
 395                 return 1;
 396
 397         /* OK, update the superblock to recover the freed space.
 398          * Physical blocks come first: have we wrapped beyond the end of
 399          * the log?  */
 400         freed = blocknr - journal->j_tail;
 401         if (blocknr < journal->j_tail)
 402                 freed = freed + journal->j_last - journal->j_first;
 403
 404         jbd_debug(1,
 405                   "Cleaning journal tail from %d to %d (offset %lu), "
 406                   "freeing %lu\n",
 407                   journal->j_tail_sequence, first_tid, blocknr, freed);
 408
 409         journal->j_free += freed;
 410         journal->j_tail_sequence = first_tid;
 411         journal->j_tail = blocknr;
 412         if (!(journal->j_flags & JFS_ABORT))
 413                 journal_update_superblock(journal, 1);
 414         return 0;
 415 }
 416
 417
 418 /* Checkpoint list management */
 419
 420 /*
 421  * journal_clean_checkpoint_list
 422  *
 423  * Find all the written-back checkpoint buffers in the journal and release them.
 424  *
 425  * Called with the journal locked.
 426  * Called with journal_datalist_lock held.
 427  * Returns number of bufers reaped (for debug)
 428  */
 429
 430 int __journal_clean_checkpoint_list(journal_t *journal)
 431 {
 432         transaction_t *transaction, *last_transaction, *next_transaction;
 433         int ret = 0;
 434
 435         transaction = journal->j_checkpoint_transactions;
 436         if (transaction == 0)
 437                 goto out;
 438
 439         last_transaction = transaction->t_cpprev;
 440         next_transaction = transaction;
 441         do {
 442                 struct journal_head *jh;
 443
 444                 transaction = next_transaction;
 445                 next_transaction = transaction->t_cpnext;
 446                 jh = transaction->t_checkpoint_list;
 447                 if (jh) {
 448                         struct journal_head *last_jh = jh->b_cpprev;
 449                         struct journal_head *next_jh = jh;
 450                         do {
 451                                 jh = next_jh;
 452                                 next_jh = jh->b_cpnext;
 453                                 ret += __try_to_free_cp_buf(jh);
 454                         } while (jh != last_jh);
 455                 }
 456         } while (transaction != last_transaction);
 457 out:
 458         return ret;
 459 }
 460
 461 /*
 462  * journal_remove_checkpoint: called after a buffer has been committed
 463  * to disk (either by being write-back flushed to disk, or being
 464  * committed to the log).
 465  *
 466  * We cannot safely clean a transaction out of the log until all of the
 467  * buffer updates committed in that transaction have safely been stored
 468  * elsewhere on disk.  To achieve this, all of the buffers in a
 469  * transaction need to be maintained on the transaction's checkpoint
 470  * list until they have been rewritten, at which point this function is
 471  * called to remove the buffer from the existing transaction's
 472  * checkpoint list.
 473  *
 474  * This function is called with the journal locked.
 475  * This function is called with journal_datalist_lock held.
 476  */
 477
 478 void __journal_remove_checkpoint(struct journal_head *jh)
 479 {
 480         transaction_t *transaction;
 481         journal_t *journal;
 482
 483         JBUFFER_TRACE(jh, "entry");
 484
 485         if ((transaction = jh->b_cp_transaction) == NULL) {
 486                 JBUFFER_TRACE(jh, "not on transaction");
 487                 goto out;
 488         }
 489
 490         journal = transaction->t_journal;
 491
 492         __buffer_unlink(jh);
 493
 494         if (transaction->t_checkpoint_list != NULL)
 495                 goto out;
 496         JBUFFER_TRACE(jh, "transaction has no more buffers");
 497
 498         /* There is one special case to worry about: if we have just
 499            pulled the buffer off a committing transaction's forget list,
 500            then even if the checkpoint list is empty, the transaction
 501            obviously cannot be dropped! */
 502
 503         if (transaction == journal->j_committing_transaction) {
 504                 JBUFFER_TRACE(jh, "belongs to committing transaction");
 505                 goto out;
 506         }
 507
 508         /* OK, that was the last buffer for the transaction: we can now
 509            safely remove this transaction from the log */
 510
 511         __journal_drop_transaction(journal, transaction);
 512
 513         /* Just in case anybody was waiting for more transactions to be
 514            checkpointed... */
 515         wake_up(&journal->j_wait_logspace);
 516 out:
 517         JBUFFER_TRACE(jh, "exit");
 518 }
 519
 520 void journal_remove_checkpoint(struct journal_head *jh)
 521 {
 522         spin_lock(&journal_datalist_lock);
 523         __journal_remove_checkpoint(jh);
 524         spin_unlock(&journal_datalist_lock);
 525 }
 526
 527 /*
 528  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
 529  * list so that we know when it is safe to clean the transaction out of
 530  * the log.
 531  *
 532  * Called with the journal locked.
 533  * Called with journal_datalist_lock held.
 534  */
 535 void __journal_insert_checkpoint(struct journal_head *jh,
 536                                transaction_t *transaction)
 537 {
 538         JBUFFER_TRACE(jh, "entry");
 539         J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jdirty(jh2bh(jh)));
 540         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
 541
 542         assert_spin_locked(&journal_datalist_lock);
 543         jh->b_cp_transaction = transaction;
 544
 545         if (!transaction->t_checkpoint_list) {
 546                 jh->b_cpnext = jh->b_cpprev = jh;
 547         } else {
 548                 jh->b_cpnext = transaction->t_checkpoint_list;
 549                 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
 550                 jh->b_cpprev->b_cpnext = jh;
 551                 jh->b_cpnext->b_cpprev = jh;
 552         }
 553         transaction->t_checkpoint_list = jh;
 554 }
 555
 556 void journal_insert_checkpoint(struct journal_head *jh,
 557                                transaction_t *transaction)
 558 {
 559         spin_lock(&journal_datalist_lock);
 560         __journal_insert_checkpoint(jh, transaction);
 561         spin_unlock(&journal_datalist_lock);
 562 }
 563
 564 /*
 565  * We've finished with this transaction structure: adios...
 566  *
 567  * The transaction must have no links except for the checkpoint by this
 568  * point.
 569  *
 570  * Called with the journal locked.
 571  * Called with journal_datalist_lock held.
 572  */
 573
 574 void __journal_drop_transaction(journal_t *journal, transaction_t *transaction)
 575 {
 576         assert_spin_locked(&journal_datalist_lock);
 577         if (transaction->t_cpnext) {
 578                 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
 579                 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
 580                 if (journal->j_checkpoint_transactions == transaction)
 581                         journal->j_checkpoint_transactions =
 582                                 transaction->t_cpnext;
 583                 if (journal->j_checkpoint_transactions == transaction)
 584                         journal->j_checkpoint_transactions = NULL;
 585         }
 586
 587         J_ASSERT (transaction->t_ilist == NULL);
 588         J_ASSERT (transaction->t_buffers == NULL);
 589         J_ASSERT (transaction->t_sync_datalist == NULL);
 590         J_ASSERT (transaction->t_async_datalist == NULL);
 591         J_ASSERT (transaction->t_forget == NULL);
 592         J_ASSERT (transaction->t_iobuf_list == NULL);
 593         J_ASSERT (transaction->t_shadow_list == NULL);
 594         J_ASSERT (transaction->t_log_list == NULL);
 595         J_ASSERT (transaction->t_checkpoint_list == NULL);
 596         J_ASSERT (transaction->t_updates == 0);
 597         J_ASSERT (list_empty(&transaction->t_jcb));
 598
 599         J_ASSERT (transaction->t_journal->j_committing_transaction !=
 600                                         transaction);
 601
 602         jbd_debug (1, "Dropping transaction %d, all done\n",
 603                    transaction->t_tid);
 604         kfree (transaction);
 605 }
 606