fs/jbd/transaction.c

   1 /*
   2  * linux/fs/transaction.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
   5  *
   6  * Copyright 1998 Red Hat corp --- 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  * Generic filesystem transaction handling code; part of the ext2fs
  13  * journaling system.
  14  *
  15  * This file manages transactions (compound commits managed by the
  16  * journaling code) and handles (individual atomic operations by the
  17  * filesystem).
  18  */
  19
  20 #include <linux/time.h>
  21 #include <linux/fs.h>
  22 #include <linux/jbd.h>
  23 #include <linux/errno.h>
  24 #include <linux/slab.h>
  25 #include <linux/timer.h>
  26 #include <linux/smp_lock.h>
  27 #include <linux/mm.h>
  28 #include <linux/highmem.h>
  29
  30 /*
  31  * get_transaction: obtain a new transaction_t object.
  32  *
  33  * Simply allocate and initialise a new transaction.  Create it in
  34  * RUNNING state and add it to the current journal (which should not
  35  * have an existing running transaction: we only make a new transaction
  36  * once we have started to commit the old one).
  37  *
  38  * Preconditions:
  39  *      The journal MUST be locked.  We don't perform atomic mallocs on the
  40  *      new transaction and we can't block without protecting against other
  41  *      processes trying to touch the journal while it is in transition.
  42  *
  43  * Called under j_state_lock
  44  */
  45
  46 static transaction_t *
  47 get_transaction(journal_t *journal, transaction_t *transaction)
  48 {
  49         transaction->t_journal = journal;
  50         transaction->t_state = T_RUNNING;
  51         transaction->t_tid = journal->j_transaction_sequence++;
  52         transaction->t_expires = jiffies + journal->j_commit_interval;
  53         spin_lock_init(&transaction->t_handle_lock);
  54
  55         /* Set up the commit timer for the new transaction. */
  56         journal->j_commit_timer->expires = transaction->t_expires;
  57         add_timer(journal->j_commit_timer);
  58
  59         J_ASSERT(journal->j_running_transaction == NULL);
  60         journal->j_running_transaction = transaction;
  61
  62         return transaction;
  63 }
  64
  65 /*
  66  * Handle management.
  67  *
  68  * A handle_t is an object which represents a single atomic update to a
  69  * filesystem, and which tracks all of the modifications which form part
  70  * of that one update.
  71  */
  72
  73 /*
  74  * start_this_handle: Given a handle, deal with any locking or stalling
  75  * needed to make sure that there is enough journal space for the handle
  76  * to begin.  Attach the handle to a transaction and set up the
  77  * transaction's buffer credits.
  78  */
  79
  80 static int start_this_handle(journal_t *journal, handle_t *handle)
  81 {
  82         transaction_t *transaction;
  83         int needed;
  84         int nblocks = handle->h_buffer_credits;
  85         transaction_t *new_transaction = NULL;
  86         int ret = 0;
  87
  88         if (nblocks > journal->j_max_transaction_buffers) {
  89                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
  90                        current->comm, nblocks,
  91                        journal->j_max_transaction_buffers);
  92                 ret = -ENOSPC;
  93                 goto out;
  94         }
  95
  96 alloc_transaction:
  97         if (!journal->j_running_transaction) {
  98                 new_transaction = jbd_kmalloc(sizeof(*new_transaction),
  99                                                 GFP_NOFS);
 100                 if (!new_transaction) {
 101                         ret = -ENOMEM;
 102                         goto out;
 103                 }
 104                 memset(new_transaction, 0, sizeof(*new_transaction));
 105         }
 106
 107         jbd_debug(3, "New handle %p going live.\n", handle);
 108
 109 repeat:
 110
 111         /*
 112          * We need to hold j_state_lock until t_updates has been incremented,
 113          * for proper journal barrier handling
 114          */
 115         spin_lock(&journal->j_state_lock);
 116 repeat_locked:
 117         if (is_journal_aborted(journal) ||
 118             (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
 119                 spin_unlock(&journal->j_state_lock);
 120                 ret = -EROFS;
 121                 goto out;
 122         }
 123
 124         /* Wait on the journal's transaction barrier if necessary */
 125         if (journal->j_barrier_count) {
 126                 spin_unlock(&journal->j_state_lock);
 127                 wait_event(journal->j_wait_transaction_locked,
 128                                 journal->j_barrier_count == 0);
 129                 goto repeat;
 130         }
 131
 132         if (!journal->j_running_transaction) {
 133                 if (!new_transaction) {
 134                         spin_unlock(&journal->j_state_lock);
 135                         goto alloc_transaction;
 136                 }
 137                 get_transaction(journal, new_transaction);
 138                 new_transaction = NULL;
 139         }
 140
 141         transaction = journal->j_running_transaction;
 142
 143         /*
 144          * If the current transaction is locked down for commit, wait for the
 145          * lock to be released.
 146          */
 147         if (transaction->t_state == T_LOCKED) {
 148                 DEFINE_WAIT(wait);
 149
 150                 prepare_to_wait(&journal->j_wait_transaction_locked,
 151                                         &wait, TASK_UNINTERRUPTIBLE);
 152                 spin_unlock(&journal->j_state_lock);
 153                 schedule();
 154                 finish_wait(&journal->j_wait_transaction_locked, &wait);
 155                 goto repeat;
 156         }
 157
 158         /*
 159          * If there is not enough space left in the log to write all potential
 160          * buffers requested by this operation, we need to stall pending a log
 161          * checkpoint to free some more log space.
 162          */
 163         spin_lock(&transaction->t_handle_lock);
 164         needed = transaction->t_outstanding_credits + nblocks;
 165
 166         if (needed > journal->j_max_transaction_buffers) {
 167                 /*
 168                  * If the current transaction is already too large, then start
 169                  * to commit it: we can then go back and attach this handle to
 170                  * a new transaction.
 171                  */
 172                 DEFINE_WAIT(wait);
 173
 174                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
 175                 spin_unlock(&transaction->t_handle_lock);
 176                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
 177                                 TASK_UNINTERRUPTIBLE);
 178                 __log_start_commit(journal, transaction->t_tid);
 179                 spin_unlock(&journal->j_state_lock);
 180                 schedule();
 181                 finish_wait(&journal->j_wait_transaction_locked, &wait);
 182                 goto repeat;
 183         }
 184
 185         /*
 186          * The commit code assumes that it can get enough log space
 187          * without forcing a checkpoint.  This is *critical* for
 188          * correctness: a checkpoint of a buffer which is also
 189          * associated with a committing transaction creates a deadlock,
 190          * so commit simply cannot force through checkpoints.
 191          *
 192          * We must therefore ensure the necessary space in the journal
 193          * *before* starting to dirty potentially checkpointed buffers
 194          * in the new transaction.
 195          *
 196          * The worst part is, any transaction currently committing can
 197          * reduce the free space arbitrarily.  Be careful to account for
 198          * those buffers when checkpointing.
 199          */
 200
 201         /*
 202          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
 203          * a _lot_ of headroom: 1/4 of the journal plus the size of
 204          * the committing transaction.  Really, we only need to give it
 205          * committing_transaction->t_outstanding_credits plus "enough" for
 206          * the log control blocks.
 207          * Also, this test is inconsitent with the matching one in
 208          * journal_extend().
 209          */
 210         if (__log_space_left(journal) < jbd_space_needed(journal)) {
 211                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
 212                 spin_unlock(&transaction->t_handle_lock);
 213                 __log_wait_for_space(journal);
 214                 goto repeat_locked;
 215         }
 216
 217         /* OK, account for the buffers that this operation expects to
 218          * use and add the handle to the running transaction. */
 219
 220         handle->h_transaction = transaction;
 221         transaction->t_outstanding_credits += nblocks;
 222         transaction->t_updates++;
 223         transaction->t_handle_count++;
 224         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
 225                   handle, nblocks, transaction->t_outstanding_credits,
 226                   __log_space_left(journal));
 227         spin_unlock(&transaction->t_handle_lock);
 228         spin_unlock(&journal->j_state_lock);
 229 out:
 230         if (new_transaction)
 231                 kfree(new_transaction);
 232         return ret;
 233 }
 234
 235 /* Allocate a new handle.  This should probably be in a slab... */
 236 static handle_t *new_handle(int nblocks)
 237 {
 238         handle_t *handle = jbd_alloc_handle(GFP_NOFS);
 239         if (!handle)
 240                 return NULL;
 241         memset(handle, 0, sizeof(*handle));
 242         handle->h_buffer_credits = nblocks;
 243         handle->h_ref = 1;
 244
 245         return handle;
 246 }
 247
 248 /**
 249  * handle_t *journal_start() - Obtain a new handle.
 250  * @journal: Journal to start transaction on.
 251  * @nblocks: number of block buffer we might modify
 252  *
 253  * We make sure that the transaction can guarantee at least nblocks of
 254  * modified buffers in the log.  We block until the log can guarantee
 255  * that much space.
 256  *
 257  * This function is visible to journal users (like ext3fs), so is not
 258  * called with the journal already locked.
 259  *
 260  * Return a pointer to a newly allocated handle, or NULL on failure
 261  */
 262 handle_t *journal_start(journal_t *journal, int nblocks)
 263 {
 264         handle_t *handle = journal_current_handle();
 265         int err;
 266
 267         if (!journal)
 268                 return ERR_PTR(-EROFS);
 269
 270         if (handle) {
 271                 J_ASSERT(handle->h_transaction->t_journal == journal);
 272                 handle->h_ref++;
 273                 return handle;
 274         }
 275
 276         handle = new_handle(nblocks);
 277         if (!handle)
 278                 return ERR_PTR(-ENOMEM);
 279
 280         current->journal_info = handle;
 281
 282         err = start_this_handle(journal, handle);
 283         if (err < 0) {
 284                 jbd_free_handle(handle);
 285                 current->journal_info = NULL;
 286                 handle = ERR_PTR(err);
 287         }
 288         return handle;
 289 }
 290
 291 /**
 292  * int journal_extend() - extend buffer credits.
 293  * @handle:  handle to 'extend'
 294  * @nblocks: nr blocks to try to extend by.
 295  *
 296  * Some transactions, such as large extends and truncates, can be done
 297  * atomically all at once or in several stages.  The operation requests
 298  * a credit for a number of buffer modications in advance, but can
 299  * extend its credit if it needs more.
 300  *
 301  * journal_extend tries to give the running handle more buffer credits.
 302  * It does not guarantee that allocation - this is a best-effort only.
 303  * The calling process MUST be able to deal cleanly with a failure to
 304  * extend here.
 305  *
 306  * Return 0 on success, non-zero on failure.
 307  *
 308  * return code < 0 implies an error
 309  * return code > 0 implies normal transaction-full status.
 310  */
 311 int journal_extend(handle_t *handle, int nblocks)
 312 {
 313         transaction_t *transaction = handle->h_transaction;
 314         journal_t *journal = transaction->t_journal;
 315         int result;
 316         int wanted;
 317
 318         result = -EIO;
 319         if (is_handle_aborted(handle))
 320                 goto out;
 321
 322         result = 1;
 323
 324         spin_lock(&journal->j_state_lock);
 325
 326         /* Don't extend a locked-down transaction! */
 327         if (handle->h_transaction->t_state != T_RUNNING) {
 328                 jbd_debug(3, "denied handle %p %d blocks: "
 329                           "transaction not running\n", handle, nblocks);
 330                 goto error_out;
 331         }
 332
 333         spin_lock(&transaction->t_handle_lock);
 334         wanted = transaction->t_outstanding_credits + nblocks;
 335
 336         if (wanted > journal->j_max_transaction_buffers) {
 337                 jbd_debug(3, "denied handle %p %d blocks: "
 338                           "transaction too large\n", handle, nblocks);
 339                 goto unlock;
 340         }
 341
 342         if (wanted > __log_space_left(journal)) {
 343                 jbd_debug(3, "denied handle %p %d blocks: "
 344                           "insufficient log space\n", handle, nblocks);
 345                 goto unlock;
 346         }
 347
 348         handle->h_buffer_credits += nblocks;
 349         transaction->t_outstanding_credits += nblocks;
 350         result = 0;
 351
 352         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
 353 unlock:
 354         spin_unlock(&transaction->t_handle_lock);
 355 error_out:
 356         spin_unlock(&journal->j_state_lock);
 357 out:
 358         return result;
 359 }
 360
 361
 362 /**
 363  * int journal_restart() - restart a handle .
 364  * @handle:  handle to restart
 365  * @nblocks: nr credits requested
 366  *
 367  * Restart a handle for a multi-transaction filesystem
 368  * operation.
 369  *
 370  * If the journal_extend() call above fails to grant new buffer credits
 371  * to a running handle, a call to journal_restart will commit the
 372  * handle's transaction so far and reattach the handle to a new
 373  * transaction capabable of guaranteeing the requested number of
 374  * credits.
 375  */
 376
 377 int journal_restart(handle_t *handle, int nblocks)
 378 {
 379         transaction_t *transaction = handle->h_transaction;
 380         journal_t *journal = transaction->t_journal;
 381         int ret;
 382
 383         /* If we've had an abort of any type, don't even think about
 384          * actually doing the restart! */
 385         if (is_handle_aborted(handle))
 386                 return 0;
 387
 388         /*
 389          * First unlink the handle from its current transaction, and start the
 390          * commit on that.
 391          */
 392         J_ASSERT(transaction->t_updates > 0);
 393         J_ASSERT(journal_current_handle() == handle);
 394
 395         spin_lock(&journal->j_state_lock);
 396         spin_lock(&transaction->t_handle_lock);
 397         transaction->t_outstanding_credits -= handle->h_buffer_credits;
 398         transaction->t_updates--;
 399
 400         if (!transaction->t_updates)
 401                 wake_up(&journal->j_wait_updates);
 402         spin_unlock(&transaction->t_handle_lock);
 403
 404         jbd_debug(2, "restarting handle %p\n", handle);
 405         __log_start_commit(journal, transaction->t_tid);
 406         spin_unlock(&journal->j_state_lock);
 407
 408         handle->h_buffer_credits = nblocks;
 409         ret = start_this_handle(journal, handle);
 410         return ret;
 411 }
 412
 413
 414 /**
 415  * void journal_lock_updates () - establish a transaction barrier.
 416  * @journal:  Journal to establish a barrier on.
 417  *
 418  * This locks out any further updates from being started, and blocks
 419  * until all existing updates have completed, returning only once the
 420  * journal is in a quiescent state with no updates running.
 421  *
 422  * The journal lock should not be held on entry.
 423  */
 424 void journal_lock_updates(journal_t *journal)
 425 {
 426         DEFINE_WAIT(wait);
 427
 428         spin_lock(&journal->j_state_lock);
 429         ++journal->j_barrier_count;
 430
 431         /* Wait until there are no running updates */
 432         while (1) {
 433                 transaction_t *transaction = journal->j_running_transaction;
 434
 435                 if (!transaction)
 436                         break;
 437
 438                 spin_lock(&transaction->t_handle_lock);
 439                 if (!transaction->t_updates) {
 440                         spin_unlock(&transaction->t_handle_lock);
 441                         break;
 442                 }
 443                 prepare_to_wait(&journal->j_wait_updates, &wait,
 444                                 TASK_UNINTERRUPTIBLE);
 445                 spin_unlock(&transaction->t_handle_lock);
 446                 spin_unlock(&journal->j_state_lock);
 447                 schedule();
 448                 finish_wait(&journal->j_wait_updates, &wait);
 449                 spin_lock(&journal->j_state_lock);
 450         }
 451         spin_unlock(&journal->j_state_lock);
 452
 453         /*
 454          * We have now established a barrier against other normal updates, but
 455          * we also need to barrier against other journal_lock_updates() calls
 456          * to make sure that we serialise special journal-locked operations
 457          * too.
 458          */
 459         down(&journal->j_barrier);
 460 }
 461
 462 /**
 463  * void journal_unlock_updates (journal_t* journal) - release barrier
 464  * @journal:  Journal to release the barrier on.
 465  *
 466  * Release a transaction barrier obtained with journal_lock_updates().
 467  *
 468  * Should be called without the journal lock held.
 469  */
 470 void journal_unlock_updates (journal_t *journal)
 471 {
 472         J_ASSERT(journal->j_barrier_count != 0);
 473
 474         up(&journal->j_barrier);
 475         spin_lock(&journal->j_state_lock);
 476         --journal->j_barrier_count;
 477         spin_unlock(&journal->j_state_lock);
 478         wake_up(&journal->j_wait_transaction_locked);
 479 }
 480
 481 /*
 482  * Report any unexpected dirty buffers which turn up.  Normally those
 483  * indicate an error, but they can occur if the user is running (say)
 484  * tune2fs to modify the live filesystem, so we need the option of
 485  * continuing as gracefully as possible.  #
 486  *
 487  * The caller should already hold the journal lock and
 488  * j_list_lock spinlock: most callers will need those anyway
 489  * in order to probe the buffer's journaling state safely.
 490  */
 491 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
 492 {
 493         struct buffer_head *bh = jh2bh(jh);
 494         int jlist;
 495
 496         if (buffer_dirty(bh)) {
 497                 /* If this buffer is one which might reasonably be dirty
 498                  * --- ie. data, or not part of this journal --- then
 499                  * we're OK to leave it alone, but otherwise we need to
 500                  * move the dirty bit to the journal's own internal
 501                  * JBDDirty bit. */
 502                 jlist = jh->b_jlist;
 503
 504                 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
 505                     jlist == BJ_Shadow || jlist == BJ_Forget) {
 506                         if (test_clear_buffer_dirty(jh2bh(jh))) {
 507                                 set_bit(BH_JBDDirty, &jh2bh(jh)->b_state);
 508                         }
 509                 }
 510         }
 511 }
 512
 513 /*
 514  * If the buffer is already part of the current transaction, then there
 515  * is nothing we need to do.  If it is already part of a prior
 516  * transaction which we are still committing to disk, then we need to
 517  * make sure that we do not overwrite the old copy: we do copy-out to
 518  * preserve the copy going to disk.  We also account the buffer against
 519  * the handle's metadata buffer credits (unless the buffer is already
 520  * part of the transaction, that is).
 521  *
 522  */
 523 static int
 524 do_get_write_access(handle_t *handle, struct journal_head *jh,
 525                         int force_copy, int *credits)
 526 {
 527         struct buffer_head *bh;
 528         transaction_t *transaction;
 529         journal_t *journal;
 530         int error;
 531         char *frozen_buffer = NULL;
 532         int need_copy = 0;
 533
 534         if (is_handle_aborted(handle))
 535                 return -EROFS;
 536
 537         transaction = handle->h_transaction;
 538         journal = transaction->t_journal;
 539
 540         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
 541
 542         JBUFFER_TRACE(jh, "entry");
 543 repeat:
 544         bh = jh2bh(jh);
 545
 546         /* @@@ Need to check for errors here at some point. */
 547
 548         lock_buffer(bh);
 549         jbd_lock_bh_state(bh);
 550
 551         /* We now hold the buffer lock so it is safe to query the buffer
 552          * state.  Is the buffer dirty?
 553          *
 554          * If so, there are two possibilities.  The buffer may be
 555          * non-journaled, and undergoing a quite legitimate writeback.
 556          * Otherwise, it is journaled, and we don't expect dirty buffers
 557          * in that state (the buffers should be marked JBD_Dirty
 558          * instead.)  So either the IO is being done under our own
 559          * control and this is a bug, or it's a third party IO such as
 560          * dump(8) (which may leave the buffer scheduled for read ---
 561          * ie. locked but not dirty) or tune2fs (which may actually have
 562          * the buffer dirtied, ugh.)  */
 563
 564         if (buffer_dirty(bh)) {
 565                 /*
 566                  * First question: is this buffer already part of the current
 567                  * transaction or the existing committing transaction?
 568                  */
 569                 if (jh->b_transaction) {
 570                         J_ASSERT_JH(jh,
 571                                 jh->b_transaction == transaction ||
 572                                 jh->b_transaction ==
 573                                         journal->j_committing_transaction);
 574                         if (jh->b_next_transaction)
 575                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
 576                                                         transaction);
 577                         JBUFFER_TRACE(jh, "Unexpected dirty buffer");
 578                         jbd_unexpected_dirty_buffer(jh);
 579                 }
 580         }
 581
 582         unlock_buffer(bh);
 583
 584         error = -EROFS;
 585         if (is_handle_aborted(handle)) {
 586                 jbd_unlock_bh_state(bh);
 587                 goto out;
 588         }
 589         error = 0;
 590
 591         /*
 592          * The buffer is already part of this transaction if b_transaction or
 593          * b_next_transaction points to it
 594          */
 595         if (jh->b_transaction == transaction ||
 596             jh->b_next_transaction == transaction)
 597                 goto done;
 598
 599         /*
 600          * If there is already a copy-out version of this buffer, then we don't
 601          * need to make another one
 602          */
 603         if (jh->b_frozen_data) {
 604                 JBUFFER_TRACE(jh, "has frozen data");
 605                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 606                 jh->b_next_transaction = transaction;
 607
 608                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
 609                 handle->h_buffer_credits--;
 610                 if (credits)
 611                         (*credits)++;
 612                 goto done;
 613         }
 614
 615         /* Is there data here we need to preserve? */
 616
 617         if (jh->b_transaction && jh->b_transaction != transaction) {
 618                 JBUFFER_TRACE(jh, "owned by older transaction");
 619                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 620                 J_ASSERT_JH(jh, jh->b_transaction ==
 621                                         journal->j_committing_transaction);
 622
 623                 /* There is one case we have to be very careful about.
 624                  * If the committing transaction is currently writing
 625                  * this buffer out to disk and has NOT made a copy-out,
 626                  * then we cannot modify the buffer contents at all
 627                  * right now.  The essence of copy-out is that it is the
 628                  * extra copy, not the primary copy, which gets
 629                  * journaled.  If the primary copy is already going to
 630                  * disk then we cannot do copy-out here. */
 631
 632                 if (jh->b_jlist == BJ_Shadow) {
 633                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
 634                         wait_queue_head_t *wqh;
 635
 636                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
 637
 638                         JBUFFER_TRACE(jh, "on shadow: sleep");
 639                         jbd_unlock_bh_state(bh);
 640                         /* commit wakes up all shadow buffers after IO */
 641                         for ( ; ; ) {
 642                                 prepare_to_wait(wqh, &wait.wait,
 643                                                 TASK_UNINTERRUPTIBLE);
 644                                 if (jh->b_jlist != BJ_Shadow)
 645                                         break;
 646                                 schedule();
 647                         }
 648                         finish_wait(wqh, &wait.wait);
 649                         goto repeat;
 650                 }
 651
 652                 /* Only do the copy if the currently-owning transaction
 653                  * still needs it.  If it is on the Forget list, the
 654                  * committing transaction is past that stage.  The
 655                  * buffer had better remain locked during the kmalloc,
 656                  * but that should be true --- we hold the journal lock
 657                  * still and the buffer is already on the BUF_JOURNAL
 658                  * list so won't be flushed.
 659                  *
 660                  * Subtle point, though: if this is a get_undo_access,
 661                  * then we will be relying on the frozen_data to contain
 662                  * the new value of the committed_data record after the
 663                  * transaction, so we HAVE to force the frozen_data copy
 664                  * in that case. */
 665
 666                 if (jh->b_jlist != BJ_Forget || force_copy) {
 667                         JBUFFER_TRACE(jh, "generate frozen data");
 668                         if (!frozen_buffer) {
 669                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
 670                                 jbd_unlock_bh_state(bh);
 671                                 frozen_buffer = jbd_kmalloc(jh2bh(jh)->b_size,
 672                                                             GFP_NOFS);
 673                                 if (!frozen_buffer) {
 674                                         printk(KERN_EMERG
 675                                                "%s: OOM for frozen_buffer\n",
 676                                                __FUNCTION__);
 677                                         JBUFFER_TRACE(jh, "oom!");
 678                                         error = -ENOMEM;
 679                                         jbd_lock_bh_state(bh);
 680                                         goto done;
 681                                 }
 682                                 goto repeat;
 683                         }
 684                         jh->b_frozen_data = frozen_buffer;
 685                         frozen_buffer = NULL;
 686                         need_copy = 1;
 687                 }
 688                 jh->b_next_transaction = transaction;
 689         }
 690
 691         J_ASSERT(handle->h_buffer_credits > 0);
 692         handle->h_buffer_credits--;
 693         if (credits)
 694                 (*credits)++;
 695
 696         /*
 697          * Finally, if the buffer is not journaled right now, we need to make
 698          * sure it doesn't get written to disk before the caller actually
 699          * commits the new data
 700          */
 701         if (!jh->b_transaction) {
 702                 JBUFFER_TRACE(jh, "no transaction");
 703                 J_ASSERT_JH(jh, !jh->b_next_transaction);
 704                 jh->b_transaction = transaction;
 705                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
 706                 spin_lock(&journal->j_list_lock);
 707                 __journal_file_buffer(jh, transaction, BJ_Reserved);
 708                 spin_unlock(&journal->j_list_lock);
 709         }
 710
 711 done:
 712         if (need_copy) {
 713                 struct page *page;
 714                 int offset;
 715                 char *source;
 716
 717                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
 718                             "Possible IO failure.\n");
 719                 page = jh2bh(jh)->b_page;
 720                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
 721                 source = kmap_atomic(page, KM_USER0);
 722                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
 723                 kunmap_atomic(source, KM_USER0);
 724         }
 725         jbd_unlock_bh_state(bh);
 726
 727         /*
 728          * If we are about to journal a buffer, then any revoke pending on it is
 729          * no longer valid
 730          */
 731         journal_cancel_revoke(handle, jh);
 732
 733 out:
 734         if (frozen_buffer)
 735                 kfree(frozen_buffer);
 736
 737         JBUFFER_TRACE(jh, "exit");
 738         return error;
 739 }
 740
 741 /**
 742  * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
 743  * @handle: transaction to add buffer modifications to
 744  * @bh:     bh to be used for metadata writes
 745  *
 746  * Returns an error code or 0 on success.
 747  *
 748  * In full data journalling mode the buffer may be of type BJ_AsyncData,
 749  * because we're write()ing a buffer which is also part of a shared mapping.
 750  */
 751
 752 int journal_get_write_access(handle_t *handle,
 753                         struct buffer_head *bh, int *credits)
 754 {
 755         struct journal_head *jh = journal_add_journal_head(bh);
 756         int rc;
 757
 758         /* We do not want to get caught playing with fields which the
 759          * log thread also manipulates.  Make sure that the buffer
 760          * completes any outstanding IO before proceeding. */
 761         rc = do_get_write_access(handle, jh, 0, credits);
 762         journal_put_journal_head(jh);
 763         return rc;
 764 }
 765
 766
 767 /*
 768  * When the user wants to journal a newly created buffer_head
 769  * (ie. getblk() returned a new buffer and we are going to populate it
 770  * manually rather than reading off disk), then we need to keep the
 771  * buffer_head locked until it has been completely filled with new
 772  * data.  In this case, we should be able to make the assertion that
 773  * the bh is not already part of an existing transaction.
 774  *
 775  * The buffer should already be locked by the caller by this point.
 776  * There is no lock ranking violation: it was a newly created,
 777  * unlocked buffer beforehand. */
 778
 779 /**
 780  * int journal_get_create_access () - notify intent to use newly created bh
 781  * @handle: transaction to new buffer to
 782  * @bh: new buffer.
 783  *
 784  * Call this if you create a new bh.
 785  */
 786 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
 787 {
 788         transaction_t *transaction = handle->h_transaction;
 789         journal_t *journal = transaction->t_journal;
 790         struct journal_head *jh = journal_add_journal_head(bh);
 791         int err;
 792
 793         jbd_debug(5, "journal_head %p\n", jh);
 794         err = -EROFS;
 795         if (is_handle_aborted(handle))
 796                 goto out;
 797         err = 0;
 798
 799         JBUFFER_TRACE(jh, "entry");
 800         /*
 801          * The buffer may already belong to this transaction due to pre-zeroing
 802          * in the filesystem's new_block code.  It may also be on the previous,
 803          * committing transaction's lists, but it HAS to be in Forget state in
 804          * that case: the transaction must have deleted the buffer for it to be
 805          * reused here.
 806          */
 807         jbd_lock_bh_state(bh);
 808         spin_lock(&journal->j_list_lock);
 809         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
 810                 jh->b_transaction == NULL ||
 811                 (jh->b_transaction == journal->j_committing_transaction &&
 812                           jh->b_jlist == BJ_Forget)));
 813
 814         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 815         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
 816
 817         J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
 818         handle->h_buffer_credits--;
 819
 820         if (jh->b_transaction == NULL) {
 821                 jh->b_transaction = transaction;
 822                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
 823                 __journal_file_buffer(jh, transaction, BJ_Reserved);
 824         } else if (jh->b_transaction == journal->j_committing_transaction) {
 825                 JBUFFER_TRACE(jh, "set next transaction");
 826                 jh->b_next_transaction = transaction;
 827         }
 828         spin_unlock(&journal->j_list_lock);
 829         jbd_unlock_bh_state(bh);
 830
 831         /*
 832          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
 833          * blocks which contain freed but then revoked metadata.  We need
 834          * to cancel the revoke in case we end up freeing it yet again
 835          * and the reallocating as data - this would cause a second revoke,
 836          * which hits an assertion error.
 837          */
 838         JBUFFER_TRACE(jh, "cancelling revoke");
 839         journal_cancel_revoke(handle, jh);
 840         journal_put_journal_head(jh);
 841 out:
 842         return err;
 843 }
 844
 845 /**
 846  * int journal_get_undo_access() -  Notify intent to modify metadata with
 847  *     non-rewindable consequences
 848  * @handle: transaction
 849  * @bh: buffer to undo
 850  * @credits: store the number of taken credits here (if not NULL)
 851  *
 852  * Sometimes there is a need to distinguish between metadata which has
 853  * been committed to disk and that which has not.  The ext3fs code uses
 854  * this for freeing and allocating space, we have to make sure that we
 855  * do not reuse freed space until the deallocation has been committed,
 856  * since if we overwrote that space we would make the delete
 857  * un-rewindable in case of a crash.
 858  *
 859  * To deal with that, journal_get_undo_access requests write access to a
 860  * buffer for parts of non-rewindable operations such as delete
 861  * operations on the bitmaps.  The journaling code must keep a copy of
 862  * the buffer's contents prior to the undo_access call until such time
 863  * as we know that the buffer has definitely been committed to disk.
 864  *
 865  * We never need to know which transaction the committed data is part
 866  * of, buffers touched here are guaranteed to be dirtied later and so
 867  * will be committed to a new transaction in due course, at which point
 868  * we can discard the old committed data pointer.
 869  *
 870  * Returns error number or 0 on success.
 871  */
 872 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh,
 873                                 int *credits)
 874 {
 875         int err;
 876         struct journal_head *jh = journal_add_journal_head(bh);
 877         char *committed_data = NULL;
 878
 879         JBUFFER_TRACE(jh, "entry");
 880
 881         /*
 882          * Do this first --- it can drop the journal lock, so we want to
 883          * make sure that obtaining the committed_data is done
 884          * atomically wrt. completion of any outstanding commits.
 885          */
 886         err = do_get_write_access(handle, jh, 1, credits);
 887         if (err)
 888                 goto out;
 889
 890 repeat:
 891         if (!jh->b_committed_data) {
 892                 committed_data = jbd_kmalloc(jh2bh(jh)->b_size, GFP_NOFS);
 893                 if (!committed_data) {
 894                         printk(KERN_EMERG "%s: No memory for committed data\n",
 895                                 __FUNCTION__);
 896                         err = -ENOMEM;
 897                         goto out;
 898                 }
 899         }
 900
 901         jbd_lock_bh_state(bh);
 902         if (!jh->b_committed_data) {
 903                 /* Copy out the current buffer contents into the
 904                  * preserved, committed copy. */
 905                 JBUFFER_TRACE(jh, "generate b_committed data");
 906                 if (!committed_data) {
 907                         jbd_unlock_bh_state(bh);
 908                         goto repeat;
 909                 }
 910
 911                 jh->b_committed_data = committed_data;
 912                 committed_data = NULL;
 913                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
 914         }
 915         jbd_unlock_bh_state(bh);
 916 out:
 917         journal_put_journal_head(jh);
 918         if (committed_data)
 919                 kfree(committed_data);
 920         return err;
 921 }
 922
 923 /**
 924  * int journal_dirty_data() -  mark a buffer as containing dirty data which
 925  *                             needs to be flushed before we can commit the
 926  *                             current transaction.
 927  * @handle: transaction
 928  * @bh: bufferhead to mark
 929  *
 930  * The buffer is placed on the transaction's data list and is marked as
 931  * belonging to the transaction.
 932  *
 933  * Returns error number or 0 on success.
 934  *
 935  * journal_dirty_data() can be called via page_launder->ext3_writepage
 936  * by kswapd.
 937  */
 938 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
 939 {
 940         journal_t *journal = handle->h_transaction->t_journal;
 941         int need_brelse = 0;
 942         struct journal_head *jh;
 943
 944         if (is_handle_aborted(handle))
 945                 return 0;
 946
 947         jh = journal_add_journal_head(bh);
 948         JBUFFER_TRACE(jh, "entry");
 949
 950         /*
 951          * The buffer could *already* be dirty.  Writeout can start
 952          * at any time.
 953          */
 954         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
 955
 956         /*
 957          * What if the buffer is already part of a running transaction?
 958          *
 959          * There are two cases:
 960          * 1) It is part of the current running transaction.  Refile it,
 961          *    just in case we have allocated it as metadata, deallocated
 962          *    it, then reallocated it as data.
 963          * 2) It is part of the previous, still-committing transaction.
 964          *    If all we want to do is to guarantee that the buffer will be
 965          *    written to disk before this new transaction commits, then
 966          *    being sure that the *previous* transaction has this same
 967          *    property is sufficient for us!  Just leave it on its old
 968          *    transaction.
 969          *
 970          * In case (2), the buffer must not already exist as metadata
 971          * --- that would violate write ordering (a transaction is free
 972          * to write its data at any point, even before the previous
 973          * committing transaction has committed).  The caller must
 974          * never, ever allow this to happen: there's nothing we can do
 975          * about it in this layer.
 976          */
 977         jbd_lock_bh_state(bh);
 978         spin_lock(&journal->j_list_lock);
 979         if (jh->b_transaction) {
 980                 JBUFFER_TRACE(jh, "has transaction");
 981                 if (jh->b_transaction != handle->h_transaction) {
 982                         JBUFFER_TRACE(jh, "belongs to older transaction");
 983                         J_ASSERT_JH(jh, jh->b_transaction ==
 984                                         journal->j_committing_transaction);
 985
 986                         /* @@@ IS THIS TRUE  ? */
 987                         /*
 988                          * Not any more.  Scenario: someone does a write()
 989                          * in data=journal mode.  The buffer's transaction has
 990                          * moved into commit.  Then someone does another
 991                          * write() to the file.  We do the frozen data copyout
 992                          * and set b_next_transaction to point to j_running_t.
 993                          * And while we're in that state, someone does a
 994                          * writepage() in an attempt to pageout the same area
 995                          * of the file via a shared mapping.  At present that
 996                          * calls journal_dirty_data(), and we get right here.
 997                          * It may be too late to journal the data.  Simply
 998                          * falling through to the next test will suffice: the
 999                          * data will be dirty and wil be checkpointed.  The
1000                          * ordering comments in the next comment block still
1001                          * apply.
1002                          */
1003                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1004
1005                         /*
1006                          * If we're journalling data, and this buffer was
1007                          * subject to a write(), it could be metadata, forget
1008                          * or shadow against the committing transaction.  Now,
1009                          * someone has dirtied the same darn page via a mapping
1010                          * and it is being writepage()'d.
1011                          * We *could* just steal the page from commit, with some
1012                          * fancy locking there.  Instead, we just skip it -
1013                          * don't tie the page's buffers to the new transaction
1014                          * at all.
1015                          * Implication: if we crash before the writepage() data
1016                          * is written into the filesystem, recovery will replay
1017                          * the write() data.
1018                          */
1019                         if (jh->b_jlist != BJ_None &&
1020                                         jh->b_jlist != BJ_SyncData &&
1021                                         jh->b_jlist != BJ_Locked) {
1022                                 JBUFFER_TRACE(jh, "Not stealing");
1023                                 goto no_journal;
1024                         }
1025
1026                         /*
1027                          * This buffer may be undergoing writeout in commit.  We
1028                          * can't return from here and let the caller dirty it
1029                          * again because that can cause the write-out loop in
1030                          * commit to never terminate.
1031                          */
1032                         if (buffer_dirty(bh)) {
1033                                 get_bh(bh);
1034                                 spin_unlock(&journal->j_list_lock);
1035                                 jbd_unlock_bh_state(bh);
1036                                 need_brelse = 1;
1037                                 sync_dirty_buffer(bh);
1038                                 jbd_lock_bh_state(bh);
1039                                 spin_lock(&journal->j_list_lock);
1040                                 /* The buffer may become locked again at any
1041                                    time if it is redirtied */
1042                         }
1043
1044                         /* journal_clean_data_list() may have got there first */
1045                         if (jh->b_transaction != NULL) {
1046                                 JBUFFER_TRACE(jh, "unfile from commit");
1047                                 __journal_unfile_buffer(jh);
1048                         }
1049                         /* The buffer will be refiled below */
1050
1051                 }
1052                 /*
1053                  * Special case --- the buffer might actually have been
1054                  * allocated and then immediately deallocated in the previous,
1055                  * committing transaction, so might still be left on that
1056                  * transaction's metadata lists.
1057                  */
1058                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1059                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1060                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1061                         __journal_unfile_buffer(jh);
1062                         JBUFFER_TRACE(jh, "file as data");
1063                         __journal_file_buffer(jh, handle->h_transaction,
1064                                                 BJ_SyncData);
1065                 }
1066         } else {
1067                 JBUFFER_TRACE(jh, "not on a transaction");
1068                 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1069         }
1070 no_journal:
1071         spin_unlock(&journal->j_list_lock);
1072         jbd_unlock_bh_state(bh);
1073         if (need_brelse) {
1074                 BUFFER_TRACE(bh, "brelse");
1075                 __brelse(bh);
1076         }
1077         JBUFFER_TRACE(jh, "exit");
1078         journal_put_journal_head(jh);
1079         return 0;
1080 }
1081
1082 /**
1083  * int journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1084  * @handle: transaction to add buffer to.
1085  * @bh: buffer to mark
1086  *
1087  * mark dirty metadata which needs to be journaled as part of the current
1088  * transaction.
1089  *
1090  * The buffer is placed on the transaction's metadata list and is marked
1091  * as belonging to the transaction.
1092  *
1093  * Returns error number or 0 on success.
1094  *
1095  * Special care needs to be taken if the buffer already belongs to the
1096  * current committing transaction (in which case we should have frozen
1097  * data present for that commit).  In that case, we don't relink the
1098  * buffer: that only gets done when the old transaction finally
1099  * completes its commit.
1100  */
1101 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1102 {
1103         transaction_t *transaction = handle->h_transaction;
1104         journal_t *journal = transaction->t_journal;
1105         struct journal_head *jh = bh2jh(bh);
1106
1107         jbd_debug(5, "journal_head %p\n", jh);
1108         JBUFFER_TRACE(jh, "entry");
1109         if (is_handle_aborted(handle))
1110                 goto out;
1111
1112         jbd_lock_bh_state(bh);
1113
1114         /*
1115          * fastpath, to avoid expensive locking.  If this buffer is already
1116          * on the running transaction's metadata list there is nothing to do.
1117          * Nobody can take it off again because there is a handle open.
1118          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1119          * result in this test being false, so we go in and take the locks.
1120          */
1121         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1122                 JBUFFER_TRACE(jh, "fastpath");
1123                 J_ASSERT_JH(jh, jh->b_transaction ==
1124                                         journal->j_running_transaction);
1125                 goto out_unlock_bh;
1126         }
1127
1128         set_buffer_jbddirty(bh);
1129
1130         /*
1131          * Metadata already on the current transaction list doesn't
1132          * need to be filed.  Metadata on another transaction's list must
1133          * be committing, and will be refiled once the commit completes:
1134          * leave it alone for now.
1135          */
1136         if (jh->b_transaction != transaction) {
1137                 JBUFFER_TRACE(jh, "already on other transaction");
1138                 J_ASSERT_JH(jh, jh->b_transaction ==
1139                                         journal->j_committing_transaction);
1140                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1141                 /* And this case is illegal: we can't reuse another
1142                  * transaction's data buffer, ever. */
1143                 goto out_unlock_bh;
1144         }
1145
1146         /* That test should have eliminated the following case: */
1147         J_ASSERT_JH(jh, jh->b_frozen_data == 0);
1148
1149         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1150         spin_lock(&journal->j_list_lock);
1151         __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1152         spin_unlock(&journal->j_list_lock);
1153 out_unlock_bh:
1154         jbd_unlock_bh_state(bh);
1155 out:
1156         JBUFFER_TRACE(jh, "exit");
1157         return 0;
1158 }
1159
1160 /*
1161  * journal_release_buffer: undo a get_write_access without any buffer
1162  * updates, if the update decided in the end that it didn't need access.
1163  *
1164  * The caller passes in the number of credits which should be put back for
1165  * this buffer (zero or one).
1166  *
1167  * We leave the buffer attached to t_reserved_list because even though this
1168  * handle doesn't want it, some other concurrent handle may want to journal
1169  * this buffer.  If that handle is curently in between get_write_access() and
1170  * journal_dirty_metadata() then it expects the buffer to be reserved.  If
1171  * we were to rip it off t_reserved_list here, the other handle will explode
1172  * when journal_dirty_metadata is presented with a non-reserved buffer.
1173  *
1174  * If nobody really wants to journal this buffer then it will be thrown
1175  * away at the start of commit.
1176  */
1177 void
1178 journal_release_buffer(handle_t *handle, struct buffer_head *bh, int credits)
1179 {
1180         BUFFER_TRACE(bh, "entry");
1181         handle->h_buffer_credits += credits;
1182 }
1183
1184 /**
1185  * void journal_forget() - bforget() for potentially-journaled buffers.
1186  * @handle: transaction handle
1187  * @bh:     bh to 'forget'
1188  *
1189  * We can only do the bforget if there are no commits pending against the
1190  * buffer.  If the buffer is dirty in the current running transaction we
1191  * can safely unlink it.
1192  *
1193  * bh may not be a journalled buffer at all - it may be a non-JBD
1194  * buffer which came off the hashtable.  Check for this.
1195  *
1196  * Decrements bh->b_count by one.
1197  *
1198  * Allow this call even if the handle has aborted --- it may be part of
1199  * the caller's cleanup after an abort.
1200  */
1201 int journal_forget (handle_t *handle, struct buffer_head *bh)
1202 {
1203         transaction_t *transaction = handle->h_transaction;
1204         journal_t *journal = transaction->t_journal;
1205         struct journal_head *jh;
1206         int err = 0;
1207
1208         BUFFER_TRACE(bh, "entry");
1209
1210         jbd_lock_bh_state(bh);
1211         spin_lock(&journal->j_list_lock);
1212
1213         if (!buffer_jbd(bh))
1214                 goto not_jbd;
1215         jh = bh2jh(bh);
1216
1217         /* Critical error: attempting to delete a bitmap buffer, maybe?
1218          * Don't do any jbd operations, and return an error. */
1219         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1220                          "inconsistent data on disk")) {
1221                 err = -EIO;
1222                 goto not_jbd;
1223         }
1224
1225         if (jh->b_transaction == handle->h_transaction) {
1226                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1227
1228                 /* If we are forgetting a buffer which is already part
1229                  * of this transaction, then we can just drop it from
1230                  * the transaction immediately. */
1231                 clear_buffer_dirty(bh);
1232                 clear_buffer_jbddirty(bh);
1233
1234                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1235
1236                 __journal_unfile_buffer(jh);
1237
1238                 /*
1239                  * We are no longer going to journal this buffer.
1240                  * However, the commit of this transaction is still
1241                  * important to the buffer: the delete that we are now
1242                  * processing might obsolete an old log entry, so by
1243                  * committing, we can satisfy the buffer's checkpoint.
1244                  *
1245                  * So, if we have a checkpoint on the buffer, we should
1246                  * now refile the buffer on our BJ_Forget list so that
1247                  * we know to remove the checkpoint after we commit.
1248                  */
1249
1250                 if (jh->b_cp_transaction) {
1251                         __journal_file_buffer(jh, transaction, BJ_Forget);
1252                 } else {
1253                         journal_remove_journal_head(bh);
1254                         __brelse(bh);
1255                         if (!buffer_jbd(bh)) {
1256                                 spin_unlock(&journal->j_list_lock);
1257                                 jbd_unlock_bh_state(bh);
1258                                 __bforget(bh);
1259                                 return 0;
1260                         }
1261                 }
1262         } else if (jh->b_transaction) {
1263                 J_ASSERT_JH(jh, (jh->b_transaction ==
1264                                  journal->j_committing_transaction));
1265                 /* However, if the buffer is still owned by a prior
1266                  * (committing) transaction, we can't drop it yet... */
1267                 JBUFFER_TRACE(jh, "belongs to older transaction");
1268                 /* ... but we CAN drop it from the new transaction if we
1269                  * have also modified it since the original commit. */
1270
1271                 if (jh->b_next_transaction) {
1272                         J_ASSERT(jh->b_next_transaction == transaction);
1273                         jh->b_next_transaction = NULL;
1274                 }
1275         }
1276
1277 not_jbd:
1278         spin_unlock(&journal->j_list_lock);
1279         jbd_unlock_bh_state(bh);
1280         __brelse(bh);
1281         return err;
1282 }
1283
1284 /**
1285  * int journal_stop() - complete a transaction
1286  * @handle: tranaction to complete.
1287  *
1288  * All done for a particular handle.
1289  *
1290  * There is not much action needed here.  We just return any remaining
1291  * buffer credits to the transaction and remove the handle.  The only
1292  * complication is that we need to start a commit operation if the
1293  * filesystem is marked for synchronous update.
1294  *
1295  * journal_stop itself will not usually return an error, but it may
1296  * do so in unusual circumstances.  In particular, expect it to
1297  * return -EIO if a journal_abort has been executed since the
1298  * transaction began.
1299  */
1300 int journal_stop(handle_t *handle)
1301 {
1302         transaction_t *transaction = handle->h_transaction;
1303         journal_t *journal = transaction->t_journal;
1304         int old_handle_count, err;
1305
1306         J_ASSERT(transaction->t_updates > 0);
1307         J_ASSERT(journal_current_handle() == handle);
1308
1309         if (is_handle_aborted(handle))
1310                 err = -EIO;
1311         else
1312                 err = 0;
1313
1314         if (--handle->h_ref > 0) {
1315                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1316                           handle->h_ref);
1317                 return err;
1318         }
1319
1320         jbd_debug(4, "Handle %p going down\n", handle);
1321
1322         /*
1323          * Implement synchronous transaction batching.  If the handle
1324          * was synchronous, don't force a commit immediately.  Let's
1325          * yield and let another thread piggyback onto this transaction.
1326          * Keep doing that while new threads continue to arrive.
1327          * It doesn't cost much - we're about to run a commit and sleep
1328          * on IO anyway.  Speeds up many-threaded, many-dir operations
1329          * by 30x or more...
1330          */
1331         if (handle->h_sync) {
1332                 do {
1333                         old_handle_count = transaction->t_handle_count;
1334                         set_current_state(TASK_UNINTERRUPTIBLE);
1335                         schedule_timeout(1);
1336                 } while (old_handle_count != transaction->t_handle_count);
1337         }
1338
1339         current->journal_info = NULL;
1340         spin_lock(&journal->j_state_lock);
1341         spin_lock(&transaction->t_handle_lock);
1342         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1343         transaction->t_updates--;
1344         if (!transaction->t_updates) {
1345                 wake_up(&journal->j_wait_updates);
1346                 if (journal->j_barrier_count)
1347                         wake_up(&journal->j_wait_transaction_locked);
1348         }
1349
1350         /*
1351          * If the handle is marked SYNC, we need to set another commit
1352          * going!  We also want to force a commit if the current
1353          * transaction is occupying too much of the log, or if the
1354          * transaction is too old now.
1355          */
1356         if (handle->h_sync ||
1357                         transaction->t_outstanding_credits >
1358                                 journal->j_max_transaction_buffers ||
1359                         time_after_eq(jiffies, transaction->t_expires)) {
1360                 /* Do this even for aborted journals: an abort still
1361                  * completes the commit thread, it just doesn't write
1362                  * anything to disk. */
1363                 tid_t tid = transaction->t_tid;
1364
1365                 spin_unlock(&transaction->t_handle_lock);
1366                 jbd_debug(2, "transaction too old, requesting commit for "
1367                                         "handle %p\n", handle);
1368                 /* This is non-blocking */
1369                 __log_start_commit(journal, transaction->t_tid);
1370                 spin_unlock(&journal->j_state_lock);
1371
1372                 /*
1373                  * Special case: JFS_SYNC synchronous updates require us
1374                  * to wait for the commit to complete.
1375                  */
1376                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1377                         err = log_wait_commit(journal, tid);
1378         } else {
1379                 spin_unlock(&transaction->t_handle_lock);
1380                 spin_unlock(&journal->j_state_lock);
1381         }
1382
1383         jbd_free_handle(handle);
1384         return err;
1385 }
1386
1387 /**int journal_force_commit() - force any uncommitted transactions
1388  * @journal: journal to force
1389  *
1390  * For synchronous operations: force any uncommitted transactions
1391  * to disk.  May seem kludgy, but it reuses all the handle batching
1392  * code in a very simple manner.
1393  */
1394 int journal_force_commit(journal_t *journal)
1395 {
1396         handle_t *handle;
1397         int ret;
1398
1399         handle = journal_start(journal, 1);
1400         if (IS_ERR(handle)) {
1401                 ret = PTR_ERR(handle);
1402         } else {
1403                 handle->h_sync = 1;
1404                 ret = journal_stop(handle);
1405         }
1406         return ret;
1407 }
1408
1409 /*
1410  *
1411  * List management code snippets: various functions for manipulating the
1412  * transaction buffer lists.
1413  *
1414  */
1415
1416 /*
1417  * Append a buffer to a transaction list, given the transaction's list head
1418  * pointer.
1419  *
1420  * j_list_lock is held.
1421  *
1422  * jbd_lock_bh_state(jh2bh(jh)) is held.
1423  */
1424
1425 static inline void
1426 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1427 {
1428         if (!*list) {
1429                 jh->b_tnext = jh->b_tprev = jh;
1430                 *list = jh;
1431         } else {
1432                 /* Insert at the tail of the list to preserve order */
1433                 struct journal_head *first = *list, *last = first->b_tprev;
1434                 jh->b_tprev = last;
1435                 jh->b_tnext = first;
1436                 last->b_tnext = first->b_tprev = jh;
1437         }
1438 }
1439
1440 /*
1441  * Remove a buffer from a transaction list, given the transaction's list
1442  * head pointer.
1443  *
1444  * Called with j_list_lock held, and the journal may not be locked.
1445  *
1446  * jbd_lock_bh_state(jh2bh(jh)) is held.
1447  */
1448
1449 static inline void
1450 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1451 {
1452         if (*list == jh) {
1453                 *list = jh->b_tnext;
1454                 if (*list == jh)
1455                         *list = NULL;
1456         }
1457         jh->b_tprev->b_tnext = jh->b_tnext;
1458         jh->b_tnext->b_tprev = jh->b_tprev;
1459 }
1460
1461 /*
1462  * Remove a buffer from the appropriate transaction list.
1463  *
1464  * Note that this function can *change* the value of
1465  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1466  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1467  * is holding onto a copy of one of thee pointers, it could go bad.
1468  * Generally the caller needs to re-read the pointer from the transaction_t.
1469  *
1470  * Called under j_list_lock.  The journal may not be locked.
1471  */
1472 void __journal_unfile_buffer(struct journal_head *jh)
1473 {
1474         struct journal_head **list = NULL;
1475         transaction_t *transaction;
1476         struct buffer_head *bh = jh2bh(jh);
1477
1478         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1479         transaction = jh->b_transaction;
1480         if (transaction)
1481                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1482
1483         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1484         if (jh->b_jlist != BJ_None)
1485                 J_ASSERT_JH(jh, transaction != 0);
1486
1487         switch (jh->b_jlist) {
1488         case BJ_None:
1489                 goto out;
1490         case BJ_SyncData:
1491                 list = &transaction->t_sync_datalist;
1492                 break;
1493         case BJ_Metadata:
1494                 transaction->t_nr_buffers--;
1495                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1496                 list = &transaction->t_buffers;
1497                 break;
1498         case BJ_Forget:
1499                 list = &transaction->t_forget;
1500                 break;
1501         case BJ_IO:
1502                 list = &transaction->t_iobuf_list;
1503                 break;
1504         case BJ_Shadow:
1505                 list = &transaction->t_shadow_list;
1506                 break;
1507         case BJ_LogCtl:
1508                 list = &transaction->t_log_list;
1509                 break;
1510         case BJ_Reserved:
1511                 list = &transaction->t_reserved_list;
1512                 break;
1513         case BJ_Locked:
1514                 list = &transaction->t_locked_list;
1515                 break;
1516         }
1517
1518         __blist_del_buffer(list, jh);
1519         jh->b_jlist = BJ_None;
1520         if (test_clear_buffer_jbddirty(bh))
1521                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1522 out:
1523         jh->b_transaction = NULL;
1524 }
1525
1526 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1527 {
1528         jbd_lock_bh_state(jh2bh(jh));
1529         spin_lock(&journal->j_list_lock);
1530         __journal_unfile_buffer(jh);
1531         spin_unlock(&journal->j_list_lock);
1532         jbd_unlock_bh_state(jh2bh(jh));
1533 }
1534
1535 /*
1536  * Called from journal_try_to_free_buffers().
1537  *
1538  * Called under jbd_lock_bh_state(bh)
1539  */
1540 static void
1541 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1542 {
1543         struct journal_head *jh;
1544
1545         jh = bh2jh(bh);
1546
1547         if (buffer_locked(bh) || buffer_dirty(bh))
1548                 goto out;
1549
1550         if (jh->b_next_transaction != 0)
1551                 goto out;
1552
1553         spin_lock(&journal->j_list_lock);
1554         if (jh->b_transaction != 0 && jh->b_cp_transaction == 0) {
1555                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1556                         /* A written-back ordered data buffer */
1557                         JBUFFER_TRACE(jh, "release data");
1558                         __journal_unfile_buffer(jh);
1559                         journal_remove_journal_head(bh);
1560                         __brelse(bh);
1561                 }
1562         } else if (jh->b_cp_transaction != 0 && jh->b_transaction == 0) {
1563                 /* written-back checkpointed metadata buffer */
1564                 if (jh->b_jlist == BJ_None) {
1565                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1566                         __journal_remove_checkpoint(jh);
1567                         journal_remove_journal_head(bh);
1568                         __brelse(bh);
1569                 }
1570         }
1571         spin_unlock(&journal->j_list_lock);
1572 out:
1573         return;
1574 }
1575
1576
1577 /**
1578  * int journal_try_to_free_buffers() - try to free page buffers.
1579  * @journal: journal for operation
1580  * @page: to try and free
1581  * @gfp_mask: 'IO' mode for try_to_free_buffers()
1582  *
1583  *
1584  * For all the buffers on this page,
1585  * if they are fully written out ordered data, move them onto BUF_CLEAN
1586  * so try_to_free_buffers() can reap them.
1587  *
1588  * This function returns non-zero if we wish try_to_free_buffers()
1589  * to be called. We do this if the page is releasable by try_to_free_buffers().
1590  * We also do it if the page has locked or dirty buffers and the caller wants
1591  * us to perform sync or async writeout.
1592  *
1593  * This complicates JBD locking somewhat.  We aren't protected by the
1594  * BKL here.  We wish to remove the buffer from its committing or
1595  * running transaction's ->t_datalist via __journal_unfile_buffer.
1596  *
1597  * This may *change* the value of transaction_t->t_datalist, so anyone
1598  * who looks at t_datalist needs to lock against this function.
1599  *
1600  * Even worse, someone may be doing a journal_dirty_data on this
1601  * buffer.  So we need to lock against that.  journal_dirty_data()
1602  * will come out of the lock with the buffer dirty, which makes it
1603  * ineligible for release here.
1604  *
1605  * Who else is affected by this?  hmm...  Really the only contender
1606  * is do_get_write_access() - it could be looking at the buffer while
1607  * journal_try_to_free_buffer() is changing its state.  But that
1608  * cannot happen because we never reallocate freed data as metadata
1609  * while the data is part of a transaction.  Yes?
1610  */
1611 int journal_try_to_free_buffers(journal_t *journal,
1612                                 struct page *page, int unused_gfp_mask)
1613 {
1614         struct buffer_head *head;
1615         struct buffer_head *bh;
1616         int ret = 0;
1617
1618         J_ASSERT(PageLocked(page));
1619
1620         head = page_buffers(page);
1621         bh = head;
1622         do {
1623                 struct journal_head *jh;
1624
1625                 /*
1626                  * We take our own ref against the journal_head here to avoid
1627                  * having to add tons of locking around each instance of
1628                  * journal_remove_journal_head() and journal_put_journal_head().
1629                  */
1630                 jh = journal_grab_journal_head(bh);
1631                 if (!jh)
1632                         continue;
1633
1634                 jbd_lock_bh_state(bh);
1635                 __journal_try_to_free_buffer(journal, bh);
1636                 journal_put_journal_head(jh);
1637                 jbd_unlock_bh_state(bh);
1638                 if (buffer_jbd(bh))
1639                         goto busy;
1640         } while ((bh = bh->b_this_page) != head);
1641         ret = try_to_free_buffers(page);
1642 busy:
1643         return ret;
1644 }
1645
1646 /*
1647  * This buffer is no longer needed.  If it is on an older transaction's
1648  * checkpoint list we need to record it on this transaction's forget list
1649  * to pin this buffer (and hence its checkpointing transaction) down until
1650  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1651  * release it.
1652  * Returns non-zero if JBD no longer has an interest in the buffer.
1653  *
1654  * Called under j_list_lock.
1655  *
1656  * Called under jbd_lock_bh_state(bh).
1657  */
1658 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1659 {
1660         int may_free = 1;
1661         struct buffer_head *bh = jh2bh(jh);
1662
1663         __journal_unfile_buffer(jh);
1664
1665         if (jh->b_cp_transaction) {
1666                 JBUFFER_TRACE(jh, "on running+cp transaction");
1667                 __journal_file_buffer(jh, transaction, BJ_Forget);
1668                 clear_buffer_jbddirty(bh);
1669                 may_free = 0;
1670         } else {
1671                 JBUFFER_TRACE(jh, "on running transaction");
1672                 journal_remove_journal_head(bh);
1673                 __brelse(bh);
1674         }
1675         return may_free;
1676 }
1677
1678 /*
1679  * journal_invalidatepage
1680  *
1681  * This code is tricky.  It has a number of cases to deal with.
1682  *
1683  * There are two invariants which this code relies on:
1684  *
1685  * i_size must be updated on disk before we start calling invalidatepage on the
1686  * data.
1687  *
1688  *  This is done in ext3 by defining an ext3_setattr method which
1689  *  updates i_size before truncate gets going.  By maintaining this
1690  *  invariant, we can be sure that it is safe to throw away any buffers
1691  *  attached to the current transaction: once the transaction commits,
1692  *  we know that the data will not be needed.
1693  *
1694  *  Note however that we can *not* throw away data belonging to the
1695  *  previous, committing transaction!
1696  *
1697  * Any disk blocks which *are* part of the previous, committing
1698  * transaction (and which therefore cannot be discarded immediately) are
1699  * not going to be reused in the new running transaction
1700  *
1701  *  The bitmap committed_data images guarantee this: any block which is
1702  *  allocated in one transaction and removed in the next will be marked
1703  *  as in-use in the committed_data bitmap, so cannot be reused until
1704  *  the next transaction to delete the block commits.  This means that
1705  *  leaving committing buffers dirty is quite safe: the disk blocks
1706  *  cannot be reallocated to a different file and so buffer aliasing is
1707  *  not possible.
1708  *
1709  *
1710  * The above applies mainly to ordered data mode.  In writeback mode we
1711  * don't make guarantees about the order in which data hits disk --- in
1712  * particular we don't guarantee that new dirty data is flushed before
1713  * transaction commit --- so it is always safe just to discard data
1714  * immediately in that mode.  --sct
1715  */
1716
1717 /*
1718  * The journal_unmap_buffer helper function returns zero if the buffer
1719  * concerned remains pinned as an anonymous buffer belonging to an older
1720  * transaction.
1721  *
1722  * We're outside-transaction here.  Either or both of j_running_transaction
1723  * and j_committing_transaction may be NULL.
1724  */
1725 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1726 {
1727         transaction_t *transaction;
1728         struct journal_head *jh;
1729         int may_free = 1;
1730         int ret;
1731
1732         BUFFER_TRACE(bh, "entry");
1733
1734         /*
1735          * It is safe to proceed here without the j_list_lock because the
1736          * buffers cannot be stolen by try_to_free_buffers as long as we are
1737          * holding the page lock. --sct
1738          */
1739
1740         if (!buffer_jbd(bh))
1741                 goto zap_buffer_unlocked;
1742
1743         spin_lock(&journal->j_state_lock);
1744         jbd_lock_bh_state(bh);
1745         spin_lock(&journal->j_list_lock);
1746
1747         jh = journal_grab_journal_head(bh);
1748         if (!jh)
1749                 goto zap_buffer_no_jh;
1750
1751         transaction = jh->b_transaction;
1752         if (transaction == NULL) {
1753                 /* First case: not on any transaction.  If it
1754                  * has no checkpoint link, then we can zap it:
1755                  * it's a writeback-mode buffer so we don't care
1756                  * if it hits disk safely. */
1757                 if (!jh->b_cp_transaction) {
1758                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1759                         goto zap_buffer;
1760                 }
1761
1762                 if (!buffer_dirty(bh)) {
1763                         /* bdflush has written it.  We can drop it now */
1764                         goto zap_buffer;
1765                 }
1766
1767                 /* OK, it must be in the journal but still not
1768                  * written fully to disk: it's metadata or
1769                  * journaled data... */
1770
1771                 if (journal->j_running_transaction) {
1772                         /* ... and once the current transaction has
1773                          * committed, the buffer won't be needed any
1774                          * longer. */
1775                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1776                         ret = __dispose_buffer(jh,
1777                                         journal->j_running_transaction);
1778                         journal_put_journal_head(jh);
1779                         spin_unlock(&journal->j_list_lock);
1780                         jbd_unlock_bh_state(bh);
1781                         spin_unlock(&journal->j_state_lock);
1782                         return ret;
1783                 } else {
1784                         /* There is no currently-running transaction. So the
1785                          * orphan record which we wrote for this file must have
1786                          * passed into commit.  We must attach this buffer to
1787                          * the committing transaction, if it exists. */
1788                         if (journal->j_committing_transaction) {
1789                                 JBUFFER_TRACE(jh, "give to committing trans");
1790                                 ret = __dispose_buffer(jh,
1791                                         journal->j_committing_transaction);
1792                                 journal_put_journal_head(jh);
1793                                 spin_unlock(&journal->j_list_lock);
1794                                 jbd_unlock_bh_state(bh);
1795                                 spin_unlock(&journal->j_state_lock);
1796                                 return ret;
1797                         } else {
1798                                 /* The orphan record's transaction has
1799                                  * committed.  We can cleanse this buffer */
1800                                 clear_buffer_jbddirty(bh);
1801                                 goto zap_buffer;
1802                         }
1803                 }
1804         } else if (transaction == journal->j_committing_transaction) {
1805                 /* If it is committing, we simply cannot touch it.  We
1806                  * can remove it's next_transaction pointer from the
1807                  * running transaction if that is set, but nothing
1808                  * else. */
1809                 JBUFFER_TRACE(jh, "on committing transaction");
1810                 set_buffer_freed(bh);
1811                 if (jh->b_next_transaction) {
1812                         J_ASSERT(jh->b_next_transaction ==
1813                                         journal->j_running_transaction);
1814                         jh->b_next_transaction = NULL;
1815                 }
1816                 journal_put_journal_head(jh);
1817                 spin_unlock(&journal->j_list_lock);
1818                 jbd_unlock_bh_state(bh);
1819                 spin_unlock(&journal->j_state_lock);
1820                 return 0;
1821         } else {
1822                 /* Good, the buffer belongs to the running transaction.
1823                  * We are writing our own transaction's data, not any
1824                  * previous one's, so it is safe to throw it away
1825                  * (remember that we expect the filesystem to have set
1826                  * i_size already for this truncate so recovery will not
1827                  * expose the disk blocks we are discarding here.) */
1828                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1829                 may_free = __dispose_buffer(jh, transaction);
1830         }
1831
1832 zap_buffer:
1833         journal_put_journal_head(jh);
1834 zap_buffer_no_jh:
1835         spin_unlock(&journal->j_list_lock);
1836         jbd_unlock_bh_state(bh);
1837         spin_unlock(&journal->j_state_lock);
1838 zap_buffer_unlocked:
1839         clear_buffer_dirty(bh);
1840         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1841         clear_buffer_mapped(bh);
1842         clear_buffer_req(bh);
1843         clear_buffer_new(bh);
1844         bh->b_bdev = NULL;
1845         return may_free;
1846 }
1847
1848 /**
1849  * int journal_invalidatepage()
1850  * @journal: journal to use for flush...
1851  * @page:    page to flush
1852  * @offset:  length of page to invalidate.
1853  *
1854  * Reap page buffers containing data after offset in page.
1855  *
1856  * Return non-zero if the page's buffers were successfully reaped.
1857  */
1858 int journal_invalidatepage(journal_t *journal,
1859                       struct page *page,
1860                       unsigned long offset)
1861 {
1862         struct buffer_head *head, *bh, *next;
1863         unsigned int curr_off = 0;
1864         int may_free = 1;
1865
1866         if (!PageLocked(page))
1867                 BUG();
1868         if (!page_has_buffers(page))
1869                 return 1;
1870
1871         /* We will potentially be playing with lists other than just the
1872          * data lists (especially for journaled data mode), so be
1873          * cautious in our locking. */
1874
1875         head = bh = page_buffers(page);
1876         do {
1877                 unsigned int next_off = curr_off + bh->b_size;
1878                 next = bh->b_this_page;
1879
1880                 /* AKPM: doing lock_buffer here may be overly paranoid */
1881                 if (offset <= curr_off) {
1882                         /* This block is wholly outside the truncation point */
1883                         lock_buffer(bh);
1884                         may_free &= journal_unmap_buffer(journal, bh);
1885                         unlock_buffer(bh);
1886                 }
1887                 curr_off = next_off;
1888                 bh = next;
1889
1890         } while (bh != head);
1891
1892         if (!offset) {
1893                 if (!may_free || !try_to_free_buffers(page))
1894                         return 0;
1895                 J_ASSERT(!page_has_buffers(page));
1896         }
1897         return 1;
1898 }
1899
1900 /*
1901  * File a buffer on the given transaction list.
1902  */
1903 void __journal_file_buffer(struct journal_head *jh,
1904                         transaction_t *transaction, int jlist)
1905 {
1906         struct journal_head **list = NULL;
1907         int was_dirty = 0;
1908         struct buffer_head *bh = jh2bh(jh);
1909
1910         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1911         assert_spin_locked(&transaction->t_journal->j_list_lock);
1912
1913         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1914         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1915                                 jh->b_transaction == 0);
1916
1917         if (jh->b_transaction && jh->b_jlist == jlist)
1918                 return;
1919
1920         /* The following list of buffer states needs to be consistent
1921          * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1922          * state. */
1923
1924         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1925             jlist == BJ_Shadow || jlist == BJ_Forget) {
1926                 if (test_clear_buffer_dirty(bh) ||
1927                     test_clear_buffer_jbddirty(bh))
1928                         was_dirty = 1;
1929         }
1930
1931         if (jh->b_transaction)
1932                 __journal_unfile_buffer(jh);
1933         jh->b_transaction = transaction;
1934
1935         switch (jlist) {
1936         case BJ_None:
1937                 J_ASSERT_JH(jh, !jh->b_committed_data);
1938                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1939                 return;
1940         case BJ_SyncData:
1941                 list = &transaction->t_sync_datalist;
1942                 break;
1943         case BJ_Metadata:
1944                 transaction->t_nr_buffers++;
1945                 list = &transaction->t_buffers;
1946                 break;
1947         case BJ_Forget:
1948                 list = &transaction->t_forget;
1949                 break;
1950         case BJ_IO:
1951                 list = &transaction->t_iobuf_list;
1952                 break;
1953         case BJ_Shadow:
1954                 list = &transaction->t_shadow_list;
1955                 break;
1956         case BJ_LogCtl:
1957                 list = &transaction->t_log_list;
1958                 break;
1959         case BJ_Reserved:
1960                 list = &transaction->t_reserved_list;
1961                 break;
1962         case BJ_Locked:
1963                 list =  &transaction->t_locked_list;
1964                 break;
1965         }
1966
1967         __blist_add_buffer(list, jh);
1968         jh->b_jlist = jlist;
1969
1970         if (was_dirty)
1971                 set_buffer_jbddirty(bh);
1972 }
1973
1974 void journal_file_buffer(struct journal_head *jh,
1975                                 transaction_t *transaction, int jlist)
1976 {
1977         jbd_lock_bh_state(jh2bh(jh));
1978         spin_lock(&transaction->t_journal->j_list_lock);
1979         __journal_file_buffer(jh, transaction, jlist);
1980         spin_unlock(&transaction->t_journal->j_list_lock);
1981         jbd_unlock_bh_state(jh2bh(jh));
1982 }
1983
1984 /*
1985  * Remove a buffer from its current buffer list in preparation for
1986  * dropping it from its current transaction entirely.  If the buffer has
1987  * already started to be used by a subsequent transaction, refile the
1988  * buffer on that transaction's metadata list.
1989  *
1990  * Called under journal->j_list_lock
1991  *
1992  * Called under jbd_lock_bh_state(jh2bh(jh))
1993  */
1994 void __journal_refile_buffer(struct journal_head *jh)
1995 {
1996         int was_dirty;
1997         struct buffer_head *bh = jh2bh(jh);
1998
1999         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2000         if (jh->b_transaction)
2001                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2002
2003         /* If the buffer is now unused, just drop it. */
2004         if (jh->b_next_transaction == NULL) {
2005                 __journal_unfile_buffer(jh);
2006                 return;
2007         }
2008
2009         /*
2010          * It has been modified by a later transaction: add it to the new
2011          * transaction's metadata list.
2012          */
2013
2014         was_dirty = test_clear_buffer_jbddirty(bh);
2015         __journal_unfile_buffer(jh);
2016         jh->b_transaction = jh->b_next_transaction;
2017         jh->b_next_transaction = NULL;
2018         __journal_file_buffer(jh, jh->b_transaction, BJ_Metadata);
2019         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2020
2021         if (was_dirty)
2022                 set_buffer_jbddirty(bh);
2023 }
2024
2025 /*
2026  * For the unlocked version of this call, also make sure that any
2027  * hanging journal_head is cleaned up if necessary.
2028  *
2029  * __journal_refile_buffer is usually called as part of a single locked
2030  * operation on a buffer_head, in which the caller is probably going to
2031  * be hooking the journal_head onto other lists.  In that case it is up
2032  * to the caller to remove the journal_head if necessary.  For the
2033  * unlocked journal_refile_buffer call, the caller isn't going to be
2034  * doing anything else to the buffer so we need to do the cleanup
2035  * ourselves to avoid a jh leak.
2036  *
2037  * *** The journal_head may be freed by this call! ***
2038  */
2039 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2040 {
2041         struct buffer_head *bh = jh2bh(jh);
2042
2043         jbd_lock_bh_state(bh);
2044         spin_lock(&journal->j_list_lock);
2045
2046         __journal_refile_buffer(jh);
2047         jbd_unlock_bh_state(bh);
2048         journal_remove_journal_head(bh);
2049
2050         spin_unlock(&journal->j_list_lock);
2051         __brelse(bh);
2052 }