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         INIT_LIST_HEAD(&transaction->t_jcb);
  54         spin_lock_init(&transaction->t_handle_lock);
  55         spin_lock_init(&transaction->t_jcb_lock);
  56
  57         /* Set up the commit timer for the new transaction. */
  58         journal->j_commit_timer->expires = transaction->t_expires;
  59         add_timer(journal->j_commit_timer);
  60
  61         J_ASSERT(journal->j_running_transaction == NULL);
  62         journal->j_running_transaction = transaction;
  63
  64         return transaction;
  65 }
  66
  67 /*
  68  * Handle management.
  69  *
  70  * A handle_t is an object which represents a single atomic update to a
  71  * filesystem, and which tracks all of the modifications which form part
  72  * of that one update.
  73  */
  74
  75 /*
  76  * start_this_handle: Given a handle, deal with any locking or stalling
  77  * needed to make sure that there is enough journal space for the handle
  78  * to begin.  Attach the handle to a transaction and set up the
  79  * transaction's buffer credits.
  80  */
  81
  82 static int start_this_handle(journal_t *journal, handle_t *handle)
  83 {
  84         transaction_t *transaction;
  85         int needed;
  86         int nblocks = handle->h_buffer_credits;
  87         transaction_t *new_transaction = NULL;
  88         int ret = 0;
  89
  90         if (nblocks > journal->j_max_transaction_buffers) {
  91                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
  92                        current->comm, nblocks,
  93                        journal->j_max_transaction_buffers);
  94                 ret = -ENOSPC;
  95                 goto out;
  96         }
  97
  98 alloc_transaction:
  99         if (!journal->j_running_transaction) {
 100                 new_transaction = jbd_kmalloc(sizeof(*new_transaction),
 101                                                 GFP_NOFS);
 102                 if (!new_transaction) {
 103                         ret = -ENOMEM;
 104                         goto out;
 105                 }
 106                 memset(new_transaction, 0, sizeof(*new_transaction));
 107         }
 108
 109         jbd_debug(3, "New handle %p going live.\n", handle);
 110
 111 repeat:
 112
 113         /*
 114          * We need to hold j_state_lock until t_updates has been incremented,
 115          * for proper journal barrier handling
 116          */
 117         spin_lock(&journal->j_state_lock);
 118 repeat_locked:
 119         if (is_journal_aborted(journal) ||
 120             (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
 121                 spin_unlock(&journal->j_state_lock);
 122                 ret = -EROFS;
 123                 goto out;
 124         }
 125
 126         /* Wait on the journal's transaction barrier if necessary */
 127         if (journal->j_barrier_count) {
 128                 spin_unlock(&journal->j_state_lock);
 129                 wait_event(journal->j_wait_transaction_locked,
 130                                 journal->j_barrier_count == 0);
 131                 goto repeat;
 132         }
 133
 134         if (!journal->j_running_transaction) {
 135                 if (!new_transaction) {
 136                         spin_unlock(&journal->j_state_lock);
 137                         goto alloc_transaction;
 138                 }
 139                 get_transaction(journal, new_transaction);
 140                 new_transaction = NULL;
 141         }
 142
 143         transaction = journal->j_running_transaction;
 144
 145         /*
 146          * If the current transaction is locked down for commit, wait for the
 147          * lock to be released.
 148          */
 149         if (transaction->t_state == T_LOCKED) {
 150                 DEFINE_WAIT(wait);
 151
 152                 prepare_to_wait(&journal->j_wait_transaction_locked,
 153                                         &wait, TASK_UNINTERRUPTIBLE);
 154                 spin_unlock(&journal->j_state_lock);
 155                 schedule();
 156                 finish_wait(&journal->j_wait_transaction_locked, &wait);
 157                 goto repeat;
 158         }
 159
 160         /*
 161          * If there is not enough space left in the log to write all potential
 162          * buffers requested by this operation, we need to stall pending a log
 163          * checkpoint to free some more log space.
 164          */
 165         spin_lock(&transaction->t_handle_lock);
 166         needed = transaction->t_outstanding_credits + nblocks;
 167
 168         if (needed > journal->j_max_transaction_buffers) {
 169                 /*
 170                  * If the current transaction is already too large, then start
 171                  * to commit it: we can then go back and attach this handle to
 172                  * a new transaction.
 173                  */
 174                 DEFINE_WAIT(wait);
 175
 176                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
 177                 spin_unlock(&transaction->t_handle_lock);
 178                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
 179                                 TASK_UNINTERRUPTIBLE);
 180                 __log_start_commit(journal, transaction->t_tid);
 181                 spin_unlock(&journal->j_state_lock);
 182                 schedule();
 183                 finish_wait(&journal->j_wait_transaction_locked, &wait);
 184                 goto repeat;
 185         }
 186
 187         /*
 188          * The commit code assumes that it can get enough log space
 189          * without forcing a checkpoint.  This is *critical* for
 190          * correctness: a checkpoint of a buffer which is also
 191          * associated with a committing transaction creates a deadlock,
 192          * so commit simply cannot force through checkpoints.
 193          *
 194          * We must therefore ensure the necessary space in the journal
 195          * *before* starting to dirty potentially checkpointed buffers
 196          * in the new transaction.
 197          *
 198          * The worst part is, any transaction currently committing can
 199          * reduce the free space arbitrarily.  Be careful to account for
 200          * those buffers when checkpointing.
 201          */
 202
 203         /*
 204          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
 205          * a _lot_ of headroom: 1/4 of the journal plus the size of
 206          * the committing transaction.  Really, we only need to give it
 207          * committing_transaction->t_outstanding_credits plus "enough" for
 208          * the log control blocks.
 209          * Also, this test is inconsitent with the matching one in
 210          * journal_extend().
 211          */
 212         if (__log_space_left(journal) < jbd_space_needed(journal)) {
 213                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
 214                 spin_unlock(&transaction->t_handle_lock);
 215                 __log_wait_for_space(journal);
 216                 goto repeat_locked;
 217         }
 218
 219         /* OK, account for the buffers that this operation expects to
 220          * use and add the handle to the running transaction. */
 221
 222         handle->h_transaction = transaction;
 223         transaction->t_outstanding_credits += nblocks;
 224         transaction->t_updates++;
 225         transaction->t_handle_count++;
 226         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
 227                   handle, nblocks, transaction->t_outstanding_credits,
 228                   __log_space_left(journal));
 229         spin_unlock(&transaction->t_handle_lock);
 230         spin_unlock(&journal->j_state_lock);
 231 out:
 232         if (new_transaction)
 233                 kfree(new_transaction);
 234         return ret;
 235 }
 236
 237 /* Allocate a new handle.  This should probably be in a slab... */
 238 static handle_t *new_handle(int nblocks)
 239 {
 240         handle_t *handle = jbd_alloc_handle(GFP_NOFS);
 241         if (!handle)
 242                 return NULL;
 243         memset(handle, 0, sizeof(*handle));
 244         handle->h_buffer_credits = nblocks;
 245         handle->h_ref = 1;
 246         INIT_LIST_HEAD(&handle->h_jcb);
 247
 248         return handle;
 249 }
 250
 251 /**
 252  * handle_t *journal_start() - Obtain a new handle.
 253  * @journal: Journal to start transaction on.
 254  * @nblocks: number of block buffer we might modify
 255  *
 256  * We make sure that the transaction can guarantee at least nblocks of
 257  * modified buffers in the log.  We block until the log can guarantee
 258  * that much space.
 259  *
 260  * This function is visible to journal users (like ext3fs), so is not
 261  * called with the journal already locked.
 262  *
 263  * Return a pointer to a newly allocated handle, or NULL on failure
 264  */
 265 handle_t *journal_start(journal_t *journal, int nblocks)
 266 {
 267         handle_t *handle = journal_current_handle();
 268         int err;
 269
 270         if (!journal)
 271                 return ERR_PTR(-EROFS);
 272
 273         if (handle) {
 274                 J_ASSERT(handle->h_transaction->t_journal == journal);
 275                 handle->h_ref++;
 276                 return handle;
 277         }
 278
 279         handle = new_handle(nblocks);
 280         if (!handle)
 281                 return ERR_PTR(-ENOMEM);
 282
 283         current->journal_info = handle;
 284
 285         err = start_this_handle(journal, handle);
 286         if (err < 0) {
 287                 jbd_free_handle(handle);
 288                 current->journal_info = NULL;
 289                 handle = ERR_PTR(err);
 290         }
 291         return handle;
 292 }
 293
 294 /**
 295  * int journal_extend() - extend buffer credits.
 296  * @handle:  handle to 'extend'
 297  * @nblocks: nr blocks to try to extend by.
 298  *
 299  * Some transactions, such as large extends and truncates, can be done
 300  * atomically all at once or in several stages.  The operation requests
 301  * a credit for a number of buffer modications in advance, but can
 302  * extend its credit if it needs more.
 303  *
 304  * journal_extend tries to give the running handle more buffer credits.
 305  * It does not guarantee that allocation - this is a best-effort only.
 306  * The calling process MUST be able to deal cleanly with a failure to
 307  * extend here.
 308  *
 309  * Return 0 on success, non-zero on failure.
 310  *
 311  * return code < 0 implies an error
 312  * return code > 0 implies normal transaction-full status.
 313  */
 314 int journal_extend(handle_t *handle, int nblocks)
 315 {
 316         transaction_t *transaction = handle->h_transaction;
 317         journal_t *journal = transaction->t_journal;
 318         int result;
 319         int wanted;
 320
 321         result = -EIO;
 322         if (is_handle_aborted(handle))
 323                 goto error_out;
 324
 325         result = 1;
 326
 327         spin_lock(&journal->j_state_lock);
 328
 329         /* Don't extend a locked-down transaction! */
 330         if (handle->h_transaction->t_state != T_RUNNING) {
 331                 jbd_debug(3, "denied handle %p %d blocks: "
 332                           "transaction not running\n", handle, nblocks);
 333                 goto error_out;
 334         }
 335
 336         spin_lock(&transaction->t_handle_lock);
 337         wanted = transaction->t_outstanding_credits + nblocks;
 338
 339         if (wanted > journal->j_max_transaction_buffers) {
 340                 jbd_debug(3, "denied handle %p %d blocks: "
 341                           "transaction too large\n", handle, nblocks);
 342                 goto unlock;
 343         }
 344
 345         if (wanted > __log_space_left(journal)) {
 346                 jbd_debug(3, "denied handle %p %d blocks: "
 347                           "insufficient log space\n", handle, nblocks);
 348                 goto unlock;
 349         }
 350
 351         handle->h_buffer_credits += nblocks;
 352         transaction->t_outstanding_credits += nblocks;
 353         result = 0;
 354
 355         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
 356 unlock:
 357         spin_unlock(&transaction->t_handle_lock);
 358 error_out:
 359         spin_unlock(&journal->j_state_lock);
 360         return result;
 361 }
 362
 363
 364 /**
 365  * int journal_restart() - restart a handle .
 366  * @handle:  handle to restart
 367  * @nblocks: nr credits requested
 368  *
 369  * Restart a handle for a multi-transaction filesystem
 370  * operation.
 371  *
 372  * If the journal_extend() call above fails to grant new buffer credits
 373  * to a running handle, a call to journal_restart will commit the
 374  * handle's transaction so far and reattach the handle to a new
 375  * transaction capabable of guaranteeing the requested number of
 376  * credits.
 377  */
 378
 379 int journal_restart(handle_t *handle, int nblocks)
 380 {
 381         transaction_t *transaction = handle->h_transaction;
 382         journal_t *journal = transaction->t_journal;
 383         int ret;
 384
 385         /* If we've had an abort of any type, don't even think about
 386          * actually doing the restart! */
 387         if (is_handle_aborted(handle))
 388                 return 0;
 389
 390         /*
 391          * First unlink the handle from its current transaction, and start the
 392          * commit on that.
 393          */
 394         J_ASSERT(transaction->t_updates > 0);
 395         J_ASSERT(journal_current_handle() == handle);
 396
 397         spin_lock(&journal->j_state_lock);
 398         spin_lock(&transaction->t_handle_lock);
 399         transaction->t_outstanding_credits -= handle->h_buffer_credits;
 400         transaction->t_updates--;
 401
 402         if (!transaction->t_updates)
 403                 wake_up(&journal->j_wait_updates);
 404         spin_unlock(&transaction->t_handle_lock);
 405
 406         jbd_debug(2, "restarting handle %p\n", handle);
 407         __log_start_commit(journal, transaction->t_tid);
 408         spin_unlock(&journal->j_state_lock);
 409
 410         handle->h_buffer_credits = nblocks;
 411         ret = start_this_handle(journal, handle);
 412         return ret;
 413 }
 414
 415
 416 /**
 417  * void journal_lock_updates () - establish a transaction barrier.
 418  * @journal:  Journal to establish a barrier on.
 419  *
 420  * This locks out any further updates from being started, and blocks
 421  * until all existing updates have completed, returning only once the
 422  * journal is in a quiescent state with no updates running.
 423  *
 424  * The journal lock should not be held on entry.
 425  */
 426 void journal_lock_updates(journal_t *journal)
 427 {
 428         DEFINE_WAIT(wait);
 429
 430         spin_lock(&journal->j_state_lock);
 431         ++journal->j_barrier_count;
 432
 433         /* Wait until there are no running updates */
 434         while (1) {
 435                 transaction_t *transaction = journal->j_running_transaction;
 436
 437                 if (!transaction)
 438                         break;
 439
 440                 spin_lock(&transaction->t_handle_lock);
 441                 if (!transaction->t_updates) {
 442                         spin_unlock(&transaction->t_handle_lock);
 443                         break;
 444                 }
 445                 prepare_to_wait(&journal->j_wait_updates, &wait,
 446                                 TASK_UNINTERRUPTIBLE);
 447                 spin_unlock(&transaction->t_handle_lock);
 448                 spin_unlock(&journal->j_state_lock);
 449                 schedule();
 450                 finish_wait(&journal->j_wait_updates, &wait);
 451                 spin_lock(&journal->j_state_lock);
 452         }
 453         spin_unlock(&journal->j_state_lock);
 454
 455         /*
 456          * We have now established a barrier against other normal updates, but
 457          * we also need to barrier against other journal_lock_updates() calls
 458          * to make sure that we serialise special journal-locked operations
 459          * too.
 460          */
 461         down(&journal->j_barrier);
 462 }
 463
 464 /**
 465  * void journal_unlock_updates (journal_t* journal) - release barrier
 466  * @journal:  Journal to release the barrier on.
 467  *
 468  * Release a transaction barrier obtained with journal_lock_updates().
 469  *
 470  * Should be called without the journal lock held.
 471  */
 472 void journal_unlock_updates (journal_t *journal)
 473 {
 474         J_ASSERT(journal->j_barrier_count != 0);
 475
 476         up(&journal->j_barrier);
 477         spin_lock(&journal->j_state_lock);
 478         --journal->j_barrier_count;
 479         spin_unlock(&journal->j_state_lock);
 480         wake_up(&journal->j_wait_transaction_locked);
 481 }
 482
 483 /*
 484  * Report any unexpected dirty buffers which turn up.  Normally those
 485  * indicate an error, but they can occur if the user is running (say)
 486  * tune2fs to modify the live filesystem, so we need the option of
 487  * continuing as gracefully as possible.  #
 488  *
 489  * The caller should already hold the journal lock and
 490  * j_list_lock spinlock: most callers will need those anyway
 491  * in order to probe the buffer's journaling state safely.
 492  */
 493 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
 494 {
 495         struct buffer_head *bh = jh2bh(jh);
 496         int jlist;
 497
 498         if (buffer_dirty(bh)) {
 499                 /* If this buffer is one which might reasonably be dirty
 500                  * --- ie. data, or not part of this journal --- then
 501                  * we're OK to leave it alone, but otherwise we need to
 502                  * move the dirty bit to the journal's own internal
 503                  * JBDDirty bit. */
 504                 jlist = jh->b_jlist;
 505
 506                 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
 507                     jlist == BJ_Shadow || jlist == BJ_Forget) {
 508                         if (test_clear_buffer_dirty(jh2bh(jh))) {
 509                                 set_bit(BH_JBDDirty, &jh2bh(jh)->b_state);
 510                         }
 511                 }
 512         }
 513 }
 514
 515 /*
 516  * If the buffer is already part of the current transaction, then there
 517  * is nothing we need to do.  If it is already part of a prior
 518  * transaction which we are still committing to disk, then we need to
 519  * make sure that we do not overwrite the old copy: we do copy-out to
 520  * preserve the copy going to disk.  We also account the buffer against
 521  * the handle's metadata buffer credits (unless the buffer is already
 522  * part of the transaction, that is).
 523  *
 524  */
 525 static int
 526 do_get_write_access(handle_t *handle, struct journal_head *jh,
 527                         int force_copy, int *credits)
 528 {
 529         struct buffer_head *bh;
 530         transaction_t *transaction;
 531         journal_t *journal;
 532         int error;
 533         char *frozen_buffer = NULL;
 534         int need_copy = 0;
 535
 536         if (is_handle_aborted(handle))
 537                 return -EROFS;
 538
 539         transaction = handle->h_transaction;
 540         journal = transaction->t_journal;
 541
 542         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
 543
 544         JBUFFER_TRACE(jh, "entry");
 545 repeat:
 546         bh = jh2bh(jh);
 547
 548         /* @@@ Need to check for errors here at some point. */
 549
 550         lock_buffer(bh);
 551         jbd_lock_bh_state(bh);
 552
 553         /* We now hold the buffer lock so it is safe to query the buffer
 554          * state.  Is the buffer dirty?
 555          *
 556          * If so, there are two possibilities.  The buffer may be
 557          * non-journaled, and undergoing a quite legitimate writeback.
 558          * Otherwise, it is journaled, and we don't expect dirty buffers
 559          * in that state (the buffers should be marked JBD_Dirty
 560          * instead.)  So either the IO is being done under our own
 561          * control and this is a bug, or it's a third party IO such as
 562          * dump(8) (which may leave the buffer scheduled for read ---
 563          * ie. locked but not dirty) or tune2fs (which may actually have
 564          * the buffer dirtied, ugh.)  */
 565
 566         if (buffer_dirty(bh)) {
 567                 /*
 568                  * First question: is this buffer already part of the current
 569                  * transaction or the existing committing transaction?
 570                  */
 571                 if (jh->b_transaction) {
 572                         J_ASSERT_JH(jh,
 573                                 jh->b_transaction == transaction ||
 574                                 jh->b_transaction ==
 575                                         journal->j_committing_transaction);
 576                         if (jh->b_next_transaction)
 577                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
 578                                                         transaction);
 579                         JBUFFER_TRACE(jh, "Unexpected dirty buffer");
 580                         jbd_unexpected_dirty_buffer(jh);
 581                 }
 582         }
 583
 584         unlock_buffer(bh);
 585
 586         error = -EROFS;
 587         if (is_handle_aborted(handle)) {
 588                 jbd_unlock_bh_state(bh);
 589                 goto out;
 590         }
 591         error = 0;
 592
 593         /*
 594          * The buffer is already part of this transaction if b_transaction or
 595          * b_next_transaction points to it
 596          */
 597         if (jh->b_transaction == transaction ||
 598             jh->b_next_transaction == transaction)
 599                 goto done;
 600
 601         /*
 602          * If there is already a copy-out version of this buffer, then we don't
 603          * need to make another one
 604          */
 605         if (jh->b_frozen_data) {
 606                 JBUFFER_TRACE(jh, "has frozen data");
 607                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 608                 jh->b_next_transaction = transaction;
 609
 610                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
 611                 handle->h_buffer_credits--;
 612                 if (credits)
 613                         (*credits)++;
 614                 goto done;
 615         }
 616
 617         /* Is there data here we need to preserve? */
 618
 619         if (jh->b_transaction && jh->b_transaction != transaction) {
 620                 JBUFFER_TRACE(jh, "owned by older transaction");
 621                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 622                 J_ASSERT_JH(jh, jh->b_transaction ==
 623                                         journal->j_committing_transaction);
 624
 625                 /* There is one case we have to be very careful about.
 626                  * If the committing transaction is currently writing
 627                  * this buffer out to disk and has NOT made a copy-out,
 628                  * then we cannot modify the buffer contents at all
 629                  * right now.  The essence of copy-out is that it is the
 630                  * extra copy, not the primary copy, which gets
 631                  * journaled.  If the primary copy is already going to
 632                  * disk then we cannot do copy-out here. */
 633
 634                 if (jh->b_jlist == BJ_Shadow) {
 635                         wait_queue_head_t *wqh;
 636
 637                         JBUFFER_TRACE(jh, "on shadow: sleep");
 638                         jbd_unlock_bh_state(bh);
 639                         /* commit wakes up all shadow buffers after IO */
 640                         wqh = bh_waitq_head(jh2bh(jh));
 641                         wait_event(*wqh, (jh->b_jlist != BJ_Shadow));
 642                         goto repeat;
 643                 }
 644
 645                 /* Only do the copy if the currently-owning transaction
 646                  * still needs it.  If it is on the Forget list, the
 647                  * committing transaction is past that stage.  The
 648                  * buffer had better remain locked during the kmalloc,
 649                  * but that should be true --- we hold the journal lock
 650                  * still and the buffer is already on the BUF_JOURNAL
 651                  * list so won't be flushed.
 652                  *
 653                  * Subtle point, though: if this is a get_undo_access,
 654                  * then we will be relying on the frozen_data to contain
 655                  * the new value of the committed_data record after the
 656                  * transaction, so we HAVE to force the frozen_data copy
 657                  * in that case. */
 658
 659                 if (jh->b_jlist != BJ_Forget || force_copy) {
 660                         JBUFFER_TRACE(jh, "generate frozen data");
 661                         if (!frozen_buffer) {
 662                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
 663                                 jbd_unlock_bh_state(bh);
 664                                 frozen_buffer = jbd_kmalloc(jh2bh(jh)->b_size,
 665                                                             GFP_NOFS);
 666                                 if (!frozen_buffer) {
 667                                         printk(KERN_EMERG
 668                                                "%s: OOM for frozen_buffer\n",
 669                                                __FUNCTION__);
 670                                         JBUFFER_TRACE(jh, "oom!");
 671                                         error = -ENOMEM;
 672                                         jbd_lock_bh_state(bh);
 673                                         goto done;
 674                                 }
 675                                 goto repeat;
 676                         }
 677                         jh->b_frozen_data = frozen_buffer;
 678                         frozen_buffer = NULL;
 679                         need_copy = 1;
 680                 }
 681                 jh->b_next_transaction = transaction;
 682         }
 683
 684         J_ASSERT(handle->h_buffer_credits > 0);
 685         handle->h_buffer_credits--;
 686         if (credits)
 687                 (*credits)++;
 688
 689         /*
 690          * Finally, if the buffer is not journaled right now, we need to make
 691          * sure it doesn't get written to disk before the caller actually
 692          * commits the new data
 693          */
 694         if (!jh->b_transaction) {
 695                 JBUFFER_TRACE(jh, "no transaction");
 696                 J_ASSERT_JH(jh, !jh->b_next_transaction);
 697                 jh->b_transaction = transaction;
 698                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
 699                 spin_lock(&journal->j_list_lock);
 700                 __journal_file_buffer(jh, transaction, BJ_Reserved);
 701                 spin_unlock(&journal->j_list_lock);
 702         }
 703
 704 done:
 705         if (need_copy) {
 706                 struct page *page;
 707                 int offset;
 708                 char *source;
 709
 710                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
 711                             "Possible IO failure.\n");
 712                 page = jh2bh(jh)->b_page;
 713                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
 714                 source = kmap_atomic(page, KM_USER0);
 715                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
 716                 kunmap_atomic(source, KM_USER0);
 717         }
 718         jbd_unlock_bh_state(bh);
 719
 720         /*
 721          * If we are about to journal a buffer, then any revoke pending on it is
 722          * no longer valid
 723          */
 724         journal_cancel_revoke(handle, jh);
 725
 726 out:
 727         if (frozen_buffer)
 728                 kfree(frozen_buffer);
 729
 730         JBUFFER_TRACE(jh, "exit");
 731         return error;
 732 }
 733
 734 /**
 735  * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
 736  * @handle: transaction to add buffer modifications to
 737  * @bh:     bh to be used for metadata writes
 738  *
 739  * Returns an error code or 0 on success.
 740  *
 741  * In full data journalling mode the buffer may be of type BJ_AsyncData,
 742  * because we're write()ing a buffer which is also part of a shared mapping.
 743  */
 744
 745 int journal_get_write_access(handle_t *handle,
 746                         struct buffer_head *bh, int *credits)
 747 {
 748         struct journal_head *jh = journal_add_journal_head(bh);
 749         int rc;
 750
 751         /* We do not want to get caught playing with fields which the
 752          * log thread also manipulates.  Make sure that the buffer
 753          * completes any outstanding IO before proceeding. */
 754         rc = do_get_write_access(handle, jh, 0, credits);
 755         journal_put_journal_head(jh);
 756         return rc;
 757 }
 758
 759
 760 /*
 761  * When the user wants to journal a newly created buffer_head
 762  * (ie. getblk() returned a new buffer and we are going to populate it
 763  * manually rather than reading off disk), then we need to keep the
 764  * buffer_head locked until it has been completely filled with new
 765  * data.  In this case, we should be able to make the assertion that
 766  * the bh is not already part of an existing transaction.
 767  *
 768  * The buffer should already be locked by the caller by this point.
 769  * There is no lock ranking violation: it was a newly created,
 770  * unlocked buffer beforehand. */
 771
 772 /**
 773  * int journal_get_create_access () - notify intent to use newly created bh
 774  * @handle: transaction to new buffer to
 775  * @bh: new buffer.
 776  *
 777  * Call this if you create a new bh.
 778  */
 779 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
 780 {
 781         transaction_t *transaction = handle->h_transaction;
 782         journal_t *journal = transaction->t_journal;
 783         struct journal_head *jh = journal_add_journal_head(bh);
 784         int err;
 785
 786         jbd_debug(5, "journal_head %p\n", jh);
 787         err = -EROFS;
 788         if (is_handle_aborted(handle))
 789                 goto out;
 790         err = 0;
 791
 792         JBUFFER_TRACE(jh, "entry");
 793         /*
 794          * The buffer may already belong to this transaction due to pre-zeroing
 795          * in the filesystem's new_block code.  It may also be on the previous,
 796          * committing transaction's lists, but it HAS to be in Forget state in
 797          * that case: the transaction must have deleted the buffer for it to be
 798          * reused here.
 799          */
 800         jbd_lock_bh_state(bh);
 801         spin_lock(&journal->j_list_lock);
 802         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
 803                 jh->b_transaction == NULL ||
 804                 (jh->b_transaction == journal->j_committing_transaction &&
 805                           jh->b_jlist == BJ_Forget)));
 806
 807         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 808         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
 809
 810         J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
 811         handle->h_buffer_credits--;
 812
 813         if (jh->b_transaction == NULL) {
 814                 jh->b_transaction = transaction;
 815                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
 816                 __journal_file_buffer(jh, transaction, BJ_Reserved);
 817         } else if (jh->b_transaction == journal->j_committing_transaction) {
 818                 JBUFFER_TRACE(jh, "set next transaction");
 819                 jh->b_next_transaction = transaction;
 820         }
 821         spin_unlock(&journal->j_list_lock);
 822         jbd_unlock_bh_state(bh);
 823
 824         /*
 825          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
 826          * blocks which contain freed but then revoked metadata.  We need
 827          * to cancel the revoke in case we end up freeing it yet again
 828          * and the reallocating as data - this would cause a second revoke,
 829          * which hits an assertion error.
 830          */
 831         JBUFFER_TRACE(jh, "cancelling revoke");
 832         journal_cancel_revoke(handle, jh);
 833         journal_put_journal_head(jh);
 834 out:
 835         return err;
 836 }
 837
 838 /**
 839  * int journal_get_undo_access() -  Notify intent to modify metadata with
 840  *     non-rewindable consequences
 841  * @handle: transaction
 842  * @bh: buffer to undo
 843  * @credits: store the number of taken credits here (if not NULL)
 844  *
 845  * Sometimes there is a need to distinguish between metadata which has
 846  * been committed to disk and that which has not.  The ext3fs code uses
 847  * this for freeing and allocating space, we have to make sure that we
 848  * do not reuse freed space until the deallocation has been committed,
 849  * since if we overwrote that space we would make the delete
 850  * un-rewindable in case of a crash.
 851  *
 852  * To deal with that, journal_get_undo_access requests write access to a
 853  * buffer for parts of non-rewindable operations such as delete
 854  * operations on the bitmaps.  The journaling code must keep a copy of
 855  * the buffer's contents prior to the undo_access call until such time
 856  * as we know that the buffer has definitely been committed to disk.
 857  *
 858  * We never need to know which transaction the committed data is part
 859  * of, buffers touched here are guaranteed to be dirtied later and so
 860  * will be committed to a new transaction in due course, at which point
 861  * we can discard the old committed data pointer.
 862  *
 863  * Returns error number or 0 on success.
 864  */
 865 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh,
 866                                 int *credits)
 867 {
 868         int err;
 869         struct journal_head *jh = journal_add_journal_head(bh);
 870         char *committed_data = NULL;
 871
 872         JBUFFER_TRACE(jh, "entry");
 873
 874         /*
 875          * Do this first --- it can drop the journal lock, so we want to
 876          * make sure that obtaining the committed_data is done
 877          * atomically wrt. completion of any outstanding commits.
 878          */
 879         err = do_get_write_access(handle, jh, 1, credits);
 880         if (err)
 881                 goto out;
 882
 883 repeat:
 884         if (!jh->b_committed_data) {
 885                 committed_data = jbd_kmalloc(jh2bh(jh)->b_size, GFP_NOFS);
 886                 if (!committed_data) {
 887                         printk(KERN_EMERG "%s: No memory for committed data\n",
 888                                 __FUNCTION__);
 889                         err = -ENOMEM;
 890                         goto out;
 891                 }
 892         }
 893
 894         jbd_lock_bh_state(bh);
 895         if (!jh->b_committed_data) {
 896                 /* Copy out the current buffer contents into the
 897                  * preserved, committed copy. */
 898                 JBUFFER_TRACE(jh, "generate b_committed data");
 899                 if (!committed_data) {
 900                         jbd_unlock_bh_state(bh);
 901                         goto repeat;
 902                 }
 903
 904                 jh->b_committed_data = committed_data;
 905                 committed_data = NULL;
 906                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
 907         }
 908         jbd_unlock_bh_state(bh);
 909 out:
 910         journal_put_journal_head(jh);
 911         if (committed_data)
 912                 kfree(committed_data);
 913         return err;
 914 }
 915
 916 /**
 917  * int journal_dirty_data() -  mark a buffer as containing dirty data which
 918  *                             needs to be flushed before we can commit the
 919  *                             current transaction.
 920  * @handle: transaction
 921  * @bh: bufferhead to mark
 922  *
 923  * The buffer is placed on the transaction's data list and is marked as
 924  * belonging to the transaction.
 925  *
 926  * Returns error number or 0 on success.
 927  *
 928  * journal_dirty_data() can be called via page_launder->ext3_writepage
 929  * by kswapd.
 930  */
 931 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
 932 {
 933         journal_t *journal = handle->h_transaction->t_journal;
 934         int need_brelse = 0;
 935         struct journal_head *jh;
 936
 937         if (is_handle_aborted(handle))
 938                 return 0;
 939
 940         jh = journal_add_journal_head(bh);
 941         JBUFFER_TRACE(jh, "entry");
 942
 943         /*
 944          * The buffer could *already* be dirty.  Writeout can start
 945          * at any time.
 946          */
 947         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
 948
 949         /*
 950          * What if the buffer is already part of a running transaction?
 951          *
 952          * There are two cases:
 953          * 1) It is part of the current running transaction.  Refile it,
 954          *    just in case we have allocated it as metadata, deallocated
 955          *    it, then reallocated it as data.
 956          * 2) It is part of the previous, still-committing transaction.
 957          *    If all we want to do is to guarantee that the buffer will be
 958          *    written to disk before this new transaction commits, then
 959          *    being sure that the *previous* transaction has this same
 960          *    property is sufficient for us!  Just leave it on its old
 961          *    transaction.
 962          *
 963          * In case (2), the buffer must not already exist as metadata
 964          * --- that would violate write ordering (a transaction is free
 965          * to write its data at any point, even before the previous
 966          * committing transaction has committed).  The caller must
 967          * never, ever allow this to happen: there's nothing we can do
 968          * about it in this layer.
 969          */
 970         jbd_lock_bh_state(bh);
 971         spin_lock(&journal->j_list_lock);
 972         if (jh->b_transaction) {
 973                 JBUFFER_TRACE(jh, "has transaction");
 974                 if (jh->b_transaction != handle->h_transaction) {
 975                         JBUFFER_TRACE(jh, "belongs to older transaction");
 976                         J_ASSERT_JH(jh, jh->b_transaction ==
 977                                         journal->j_committing_transaction);
 978
 979                         /* @@@ IS THIS TRUE  ? */
 980                         /*
 981                          * Not any more.  Scenario: someone does a write()
 982                          * in data=journal mode.  The buffer's transaction has
 983                          * moved into commit.  Then someone does another
 984                          * write() to the file.  We do the frozen data copyout
 985                          * and set b_next_transaction to point to j_running_t.
 986                          * And while we're in that state, someone does a
 987                          * writepage() in an attempt to pageout the same area
 988                          * of the file via a shared mapping.  At present that
 989                          * calls journal_dirty_data(), and we get right here.
 990                          * It may be too late to journal the data.  Simply
 991                          * falling through to the next test will suffice: the
 992                          * data will be dirty and wil be checkpointed.  The
 993                          * ordering comments in the next comment block still
 994                          * apply.
 995                          */
 996                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 997
 998                         /*
 999                          * If we're journalling data, and this buffer was
1000                          * subject to a write(), it could be metadata, forget
1001                          * or shadow against the committing transaction.  Now,
1002                          * someone has dirtied the same darn page via a mapping
1003                          * and it is being writepage()'d.
1004                          * We *could* just steal the page from commit, with some
1005                          * fancy locking there.  Instead, we just skip it -
1006                          * don't tie the page's buffers to the new transaction
1007                          * at all.
1008                          * Implication: if we crash before the writepage() data
1009                          * is written into the filesystem, recovery will replay
1010                          * the write() data.
1011                          */
1012                         if (jh->b_jlist != BJ_None &&
1013                                         jh->b_jlist != BJ_SyncData &&
1014                                         jh->b_jlist != BJ_Locked) {
1015                                 JBUFFER_TRACE(jh, "Not stealing");
1016                                 goto no_journal;
1017                         }
1018
1019                         /*
1020                          * This buffer may be undergoing writeout in commit.  We
1021                          * can't return from here and let the caller dirty it
1022                          * again because that can cause the write-out loop in
1023                          * commit to never terminate.
1024                          */
1025                         if (buffer_dirty(bh)) {
1026                                 get_bh(bh);
1027                                 spin_unlock(&journal->j_list_lock);
1028                                 jbd_unlock_bh_state(bh);
1029                                 need_brelse = 1;
1030                                 sync_dirty_buffer(bh);
1031                                 jbd_lock_bh_state(bh);
1032                                 spin_lock(&journal->j_list_lock);
1033                                 /* The buffer may become locked again at any
1034                                    time if it is redirtied */
1035                         }
1036
1037                         /* journal_clean_data_list() may have got there first */
1038                         if (jh->b_transaction != NULL) {
1039                                 JBUFFER_TRACE(jh, "unfile from commit");
1040                                 __journal_unfile_buffer(jh);
1041                         }
1042                         /* The buffer will be refiled below */
1043
1044                 }
1045                 /*
1046                  * Special case --- the buffer might actually have been
1047                  * allocated and then immediately deallocated in the previous,
1048                  * committing transaction, so might still be left on that
1049                  * transaction's metadata lists.
1050                  */
1051                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1052                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1053                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1054                         __journal_unfile_buffer(jh);
1055                         JBUFFER_TRACE(jh, "file as data");
1056                         __journal_file_buffer(jh, handle->h_transaction,
1057                                                 BJ_SyncData);
1058                 }
1059         } else {
1060                 JBUFFER_TRACE(jh, "not on a transaction");
1061                 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1062         }
1063 no_journal:
1064         spin_unlock(&journal->j_list_lock);
1065         jbd_unlock_bh_state(bh);
1066         if (need_brelse) {
1067                 BUFFER_TRACE(bh, "brelse");
1068                 __brelse(bh);
1069         }
1070         JBUFFER_TRACE(jh, "exit");
1071         journal_put_journal_head(jh);
1072         return 0;
1073 }
1074
1075 /**
1076  * int journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1077  * @handle: transaction to add buffer to.
1078  * @bh: buffer to mark
1079  *
1080  * mark dirty metadata which needs to be journaled as part of the current
1081  * transaction.
1082  *
1083  * The buffer is placed on the transaction's metadata list and is marked
1084  * as belonging to the transaction.
1085  *
1086  * Returns error number or 0 on success.
1087  *
1088  * Special care needs to be taken if the buffer already belongs to the
1089  * current committing transaction (in which case we should have frozen
1090  * data present for that commit).  In that case, we don't relink the
1091  * buffer: that only gets done when the old transaction finally
1092  * completes its commit.
1093  */
1094 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1095 {
1096         transaction_t *transaction = handle->h_transaction;
1097         journal_t *journal = transaction->t_journal;
1098         struct journal_head *jh = bh2jh(bh);
1099
1100         jbd_debug(5, "journal_head %p\n", jh);
1101         JBUFFER_TRACE(jh, "entry");
1102         if (is_handle_aborted(handle))
1103                 goto out;
1104
1105         jbd_lock_bh_state(bh);
1106
1107         /*
1108          * fastpath, to avoid expensive locking.  If this buffer is already
1109          * on the running transaction's metadata list there is nothing to do.
1110          * Nobody can take it off again because there is a handle open.
1111          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1112          * result in this test being false, so we go in and take the locks.
1113          */
1114         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1115                 JBUFFER_TRACE(jh, "fastpath");
1116                 J_ASSERT_JH(jh, jh->b_transaction ==
1117                                         journal->j_running_transaction);
1118                 goto out_unlock_bh;
1119         }
1120
1121         set_buffer_jbddirty(bh);
1122
1123         /*
1124          * Metadata already on the current transaction list doesn't
1125          * need to be filed.  Metadata on another transaction's list must
1126          * be committing, and will be refiled once the commit completes:
1127          * leave it alone for now.
1128          */
1129         if (jh->b_transaction != transaction) {
1130                 JBUFFER_TRACE(jh, "already on other transaction");
1131                 J_ASSERT_JH(jh, jh->b_transaction ==
1132                                         journal->j_committing_transaction);
1133                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1134                 /* And this case is illegal: we can't reuse another
1135                  * transaction's data buffer, ever. */
1136                 goto out_unlock_bh;
1137         }
1138
1139         /* That test should have eliminated the following case: */
1140         J_ASSERT_JH(jh, jh->b_frozen_data == 0);
1141
1142         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1143         spin_lock(&journal->j_list_lock);
1144         __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1145         spin_unlock(&journal->j_list_lock);
1146 out_unlock_bh:
1147         jbd_unlock_bh_state(bh);
1148 out:
1149         JBUFFER_TRACE(jh, "exit");
1150         return 0;
1151 }
1152
1153 /*
1154  * journal_release_buffer: undo a get_write_access without any buffer
1155  * updates, if the update decided in the end that it didn't need access.
1156  *
1157  * The caller passes in the number of credits which should be put back for
1158  * this buffer (zero or one).
1159  *
1160  * We leave the buffer attached to t_reserved_list because even though this
1161  * handle doesn't want it, some other concurrent handle may want to journal
1162  * this buffer.  If that handle is curently in between get_write_access() and
1163  * journal_dirty_metadata() then it expects the buffer to be reserved.  If
1164  * we were to rip it off t_reserved_list here, the other handle will explode
1165  * when journal_dirty_metadata is presented with a non-reserved buffer.
1166  *
1167  * If nobody really wants to journal this buffer then it will be thrown
1168  * away at the start of commit.
1169  */
1170 void
1171 journal_release_buffer(handle_t *handle, struct buffer_head *bh, int credits)
1172 {
1173         BUFFER_TRACE(bh, "entry");
1174         handle->h_buffer_credits += credits;
1175 }
1176
1177 /**
1178  * void journal_forget() - bforget() for potentially-journaled buffers.
1179  * @handle: transaction handle
1180  * @bh:     bh to 'forget'
1181  *
1182  * We can only do the bforget if there are no commits pending against the
1183  * buffer.  If the buffer is dirty in the current running transaction we
1184  * can safely unlink it.
1185  *
1186  * bh may not be a journalled buffer at all - it may be a non-JBD
1187  * buffer which came off the hashtable.  Check for this.
1188  *
1189  * Decrements bh->b_count by one.
1190  *
1191  * Allow this call even if the handle has aborted --- it may be part of
1192  * the caller's cleanup after an abort.
1193  */
1194 void journal_forget(handle_t *handle, struct buffer_head *bh)
1195 {
1196         transaction_t *transaction = handle->h_transaction;
1197         journal_t *journal = transaction->t_journal;
1198         struct journal_head *jh;
1199
1200         BUFFER_TRACE(bh, "entry");
1201
1202         jbd_lock_bh_state(bh);
1203         spin_lock(&journal->j_list_lock);
1204
1205         if (!buffer_jbd(bh))
1206                 goto not_jbd;
1207         jh = bh2jh(bh);
1208
1209         if (jh->b_transaction == handle->h_transaction) {
1210                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1211
1212                 /* If we are forgetting a buffer which is already part
1213                  * of this transaction, then we can just drop it from
1214                  * the transaction immediately. */
1215                 clear_buffer_dirty(bh);
1216                 clear_buffer_jbddirty(bh);
1217
1218                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1219                 J_ASSERT_JH(jh, !jh->b_committed_data);
1220
1221                 __journal_unfile_buffer(jh);
1222
1223                 /*
1224                  * We are no longer going to journal this buffer.
1225                  * However, the commit of this transaction is still
1226                  * important to the buffer: the delete that we are now
1227                  * processing might obsolete an old log entry, so by
1228                  * committing, we can satisfy the buffer's checkpoint.
1229                  *
1230                  * So, if we have a checkpoint on the buffer, we should
1231                  * now refile the buffer on our BJ_Forget list so that
1232                  * we know to remove the checkpoint after we commit.
1233                  */
1234
1235                 if (jh->b_cp_transaction) {
1236                         __journal_file_buffer(jh, transaction, BJ_Forget);
1237                 } else {
1238                         journal_remove_journal_head(bh);
1239                         __brelse(bh);
1240                         if (!buffer_jbd(bh)) {
1241                                 spin_unlock(&journal->j_list_lock);
1242                                 jbd_unlock_bh_state(bh);
1243                                 __bforget(bh);
1244                                 return;
1245                         }
1246                 }
1247         } else if (jh->b_transaction) {
1248                 J_ASSERT_JH(jh, (jh->b_transaction ==
1249                                  journal->j_committing_transaction));
1250                 /* However, if the buffer is still owned by a prior
1251                  * (committing) transaction, we can't drop it yet... */
1252                 JBUFFER_TRACE(jh, "belongs to older transaction");
1253                 /* ... but we CAN drop it from the new transaction if we
1254                  * have also modified it since the original commit. */
1255
1256                 if (jh->b_next_transaction) {
1257                         J_ASSERT(jh->b_next_transaction == transaction);
1258                         jh->b_next_transaction = NULL;
1259                 }
1260         }
1261
1262 not_jbd:
1263         spin_unlock(&journal->j_list_lock);
1264         jbd_unlock_bh_state(bh);
1265         __brelse(bh);
1266         return;
1267 }
1268
1269 /**
1270  * void journal_callback_set() -  Register a callback function for this handle.
1271  * @handle: handle to attach the callback to.
1272  * @func: function to callback.
1273  * @jcb:  structure with additional information required by func() , and
1274  *        some space for jbd internal information.
1275  *
1276  * The function will be
1277  * called when the transaction that this handle is part of has been
1278  * committed to disk with the original callback data struct and the
1279  * error status of the journal as parameters.  There is no guarantee of
1280  * ordering between handles within a single transaction, nor between
1281  * callbacks registered on the same handle.
1282  *
1283  * The caller is responsible for allocating the journal_callback struct.
1284  * This is to allow the caller to add as much extra data to the callback
1285  * as needed, but reduce the overhead of multiple allocations.  The caller
1286  * allocated struct must start with a struct journal_callback at offset 0,
1287  * and has the caller-specific data afterwards.
1288  */
1289 void journal_callback_set(handle_t *handle,
1290                         void (*func)(struct journal_callback *jcb, int error),
1291                         struct journal_callback *jcb)
1292 {
1293         spin_lock(&handle->h_transaction->t_jcb_lock);
1294         list_add_tail(&jcb->jcb_list, &handle->h_jcb);
1295         spin_unlock(&handle->h_transaction->t_jcb_lock);
1296         jcb->jcb_func = func;
1297 }
1298
1299 /**
1300  * int journal_stop() - complete a transaction
1301  * @handle: tranaction to complete.
1302  *
1303  * All done for a particular handle.
1304  *
1305  * There is not much action needed here.  We just return any remaining
1306  * buffer credits to the transaction and remove the handle.  The only
1307  * complication is that we need to start a commit operation if the
1308  * filesystem is marked for synchronous update.
1309  *
1310  * journal_stop itself will not usually return an error, but it may
1311  * do so in unusual circumstances.  In particular, expect it to
1312  * return -EIO if a journal_abort has been executed since the
1313  * transaction began.
1314  */
1315 int journal_stop(handle_t *handle)
1316 {
1317         transaction_t *transaction = handle->h_transaction;
1318         journal_t *journal = transaction->t_journal;
1319         int old_handle_count, err;
1320
1321         J_ASSERT(transaction->t_updates > 0);
1322         J_ASSERT(journal_current_handle() == handle);
1323
1324         if (is_handle_aborted(handle))
1325                 err = -EIO;
1326         else
1327                 err = 0;
1328
1329         if (--handle->h_ref > 0) {
1330                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1331                           handle->h_ref);
1332                 return err;
1333         }
1334
1335         jbd_debug(4, "Handle %p going down\n", handle);
1336
1337         /*
1338          * Implement synchronous transaction batching.  If the handle
1339          * was synchronous, don't force a commit immediately.  Let's
1340          * yield and let another thread piggyback onto this transaction.
1341          * Keep doing that while new threads continue to arrive.
1342          * It doesn't cost much - we're about to run a commit and sleep
1343          * on IO anyway.  Speeds up many-threaded, many-dir operations
1344          * by 30x or more...
1345          */
1346         if (handle->h_sync) {
1347                 do {
1348                         old_handle_count = transaction->t_handle_count;
1349                         set_current_state(TASK_UNINTERRUPTIBLE);
1350                         schedule_timeout(1);
1351                 } while (old_handle_count != transaction->t_handle_count);
1352         }
1353
1354         current->journal_info = NULL;
1355         spin_lock(&journal->j_state_lock);
1356         spin_lock(&transaction->t_handle_lock);
1357         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1358         transaction->t_updates--;
1359         if (!transaction->t_updates) {
1360                 wake_up(&journal->j_wait_updates);
1361                 if (journal->j_barrier_count)
1362                         wake_up(&journal->j_wait_transaction_locked);
1363         }
1364
1365         /* Move callbacks from the handle to the transaction. */
1366         spin_lock(&transaction->t_jcb_lock);
1367         list_splice(&handle->h_jcb, &transaction->t_jcb);
1368         spin_unlock(&transaction->t_jcb_lock);
1369
1370         /*
1371          * If the handle is marked SYNC, we need to set another commit
1372          * going!  We also want to force a commit if the current
1373          * transaction is occupying too much of the log, or if the
1374          * transaction is too old now.
1375          */
1376         if (handle->h_sync ||
1377                         transaction->t_outstanding_credits >
1378                                 journal->j_max_transaction_buffers ||
1379                         time_after_eq(jiffies, transaction->t_expires)) {
1380                 /* Do this even for aborted journals: an abort still
1381                  * completes the commit thread, it just doesn't write
1382                  * anything to disk. */
1383                 tid_t tid = transaction->t_tid;
1384
1385                 spin_unlock(&transaction->t_handle_lock);
1386                 jbd_debug(2, "transaction too old, requesting commit for "
1387                                         "handle %p\n", handle);
1388                 /* This is non-blocking */
1389                 __log_start_commit(journal, transaction->t_tid);
1390                 spin_unlock(&journal->j_state_lock);
1391
1392                 /*
1393                  * Special case: JFS_SYNC synchronous updates require us
1394                  * to wait for the commit to complete.
1395                  */
1396                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1397                         err = log_wait_commit(journal, tid);
1398         } else {
1399                 spin_unlock(&transaction->t_handle_lock);
1400                 spin_unlock(&journal->j_state_lock);
1401         }
1402
1403         jbd_free_handle(handle);
1404         return err;
1405 }
1406
1407 /**int journal_force_commit() - force any uncommitted transactions
1408  * @journal: journal to force
1409  *
1410  * For synchronous operations: force any uncommitted transactions
1411  * to disk.  May seem kludgy, but it reuses all the handle batching
1412  * code in a very simple manner.
1413  */
1414 int journal_force_commit(journal_t *journal)
1415 {
1416         handle_t *handle;
1417         int ret;
1418
1419         handle = journal_start(journal, 1);
1420         if (IS_ERR(handle)) {
1421                 ret = PTR_ERR(handle);
1422         } else {
1423                 handle->h_sync = 1;
1424                 ret = journal_stop(handle);
1425         }
1426         return ret;
1427 }
1428
1429 /*
1430  *
1431  * List management code snippets: various functions for manipulating the
1432  * transaction buffer lists.
1433  *
1434  */
1435
1436 /*
1437  * Append a buffer to a transaction list, given the transaction's list head
1438  * pointer.
1439  *
1440  * j_list_lock is held.
1441  *
1442  * jbd_lock_bh_state(jh2bh(jh)) is held.
1443  */
1444
1445 static inline void
1446 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1447 {
1448         if (!*list) {
1449                 jh->b_tnext = jh->b_tprev = jh;
1450                 *list = jh;
1451         } else {
1452                 /* Insert at the tail of the list to preserve order */
1453                 struct journal_head *first = *list, *last = first->b_tprev;
1454                 jh->b_tprev = last;
1455                 jh->b_tnext = first;
1456                 last->b_tnext = first->b_tprev = jh;
1457         }
1458 }
1459
1460 /*
1461  * Remove a buffer from a transaction list, given the transaction's list
1462  * head pointer.
1463  *
1464  * Called with j_list_lock held, and the journal may not be locked.
1465  *
1466  * jbd_lock_bh_state(jh2bh(jh)) is held.
1467  */
1468
1469 static inline void
1470 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1471 {
1472         if (*list == jh) {
1473                 *list = jh->b_tnext;
1474                 if (*list == jh)
1475                         *list = 0;
1476         }
1477         jh->b_tprev->b_tnext = jh->b_tnext;
1478         jh->b_tnext->b_tprev = jh->b_tprev;
1479 }
1480
1481 /*
1482  * Remove a buffer from the appropriate transaction list.
1483  *
1484  * Note that this function can *change* the value of
1485  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1486  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1487  * is holding onto a copy of one of thee pointers, it could go bad.
1488  * Generally the caller needs to re-read the pointer from the transaction_t.
1489  *
1490  * Called under j_list_lock.  The journal may not be locked.
1491  */
1492 void __journal_unfile_buffer(struct journal_head *jh)
1493 {
1494         struct journal_head **list = 0;
1495         transaction_t *transaction;
1496         struct buffer_head *bh = jh2bh(jh);
1497
1498         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1499         transaction = jh->b_transaction;
1500         if (transaction)
1501                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1502
1503         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1504         if (jh->b_jlist != BJ_None)
1505                 J_ASSERT_JH(jh, transaction != 0);
1506
1507         switch (jh->b_jlist) {
1508         case BJ_None:
1509                 goto out;
1510         case BJ_SyncData:
1511                 list = &transaction->t_sync_datalist;
1512                 break;
1513         case BJ_Metadata:
1514                 transaction->t_nr_buffers--;
1515                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1516                 list = &transaction->t_buffers;
1517                 break;
1518         case BJ_Forget:
1519                 list = &transaction->t_forget;
1520                 break;
1521         case BJ_IO:
1522                 list = &transaction->t_iobuf_list;
1523                 break;
1524         case BJ_Shadow:
1525                 list = &transaction->t_shadow_list;
1526                 break;
1527         case BJ_LogCtl:
1528                 list = &transaction->t_log_list;
1529                 break;
1530         case BJ_Reserved:
1531                 list = &transaction->t_reserved_list;
1532                 break;
1533         case BJ_Locked:
1534                 list = &transaction->t_locked_list;
1535                 break;
1536         }
1537
1538         __blist_del_buffer(list, jh);
1539         jh->b_jlist = BJ_None;
1540         if (test_clear_buffer_jbddirty(bh))
1541                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1542 out:
1543         jh->b_transaction = NULL;
1544 }
1545
1546 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1547 {
1548         jbd_lock_bh_state(jh2bh(jh));
1549         spin_lock(&journal->j_list_lock);
1550         __journal_unfile_buffer(jh);
1551         spin_unlock(&journal->j_list_lock);
1552         jbd_unlock_bh_state(jh2bh(jh));
1553 }
1554
1555 /*
1556  * Called from journal_try_to_free_buffers().
1557  *
1558  * Called under jbd_lock_bh_state(bh)
1559  */
1560 static void
1561 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1562 {
1563         struct journal_head *jh;
1564
1565         jh = bh2jh(bh);
1566
1567         if (buffer_locked(bh) || buffer_dirty(bh))
1568                 goto out;
1569
1570         if (jh->b_next_transaction != 0)
1571                 goto out;
1572
1573         spin_lock(&journal->j_list_lock);
1574         if (jh->b_transaction != 0 && jh->b_cp_transaction == 0) {
1575                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1576                         /* A written-back ordered data buffer */
1577                         JBUFFER_TRACE(jh, "release data");
1578                         __journal_unfile_buffer(jh);
1579                         journal_remove_journal_head(bh);
1580                         __brelse(bh);
1581                 }
1582         } else if (jh->b_cp_transaction != 0 && jh->b_transaction == 0) {
1583                 /* written-back checkpointed metadata buffer */
1584                 if (jh->b_jlist == BJ_None) {
1585                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1586                         __journal_remove_checkpoint(jh);
1587                         journal_remove_journal_head(bh);
1588                         __brelse(bh);
1589                 }
1590         }
1591         spin_unlock(&journal->j_list_lock);
1592 out:
1593         return;
1594 }
1595
1596
1597 /**
1598  * int journal_try_to_free_buffers() - try to free page buffers.
1599  * @journal: journal for operation
1600  * @page: to try and free
1601  * @gfp_mask: 'IO' mode for try_to_free_buffers()
1602  *
1603  *
1604  * For all the buffers on this page,
1605  * if they are fully written out ordered data, move them onto BUF_CLEAN
1606  * so try_to_free_buffers() can reap them.
1607  *
1608  * This function returns non-zero if we wish try_to_free_buffers()
1609  * to be called. We do this if the page is releasable by try_to_free_buffers().
1610  * We also do it if the page has locked or dirty buffers and the caller wants
1611  * us to perform sync or async writeout.
1612  *
1613  * This complicates JBD locking somewhat.  We aren't protected by the
1614  * BKL here.  We wish to remove the buffer from its committing or
1615  * running transaction's ->t_datalist via __journal_unfile_buffer.
1616  *
1617  * This may *change* the value of transaction_t->t_datalist, so anyone
1618  * who looks at t_datalist needs to lock against this function.
1619  *
1620  * Even worse, someone may be doing a journal_dirty_data on this
1621  * buffer.  So we need to lock against that.  journal_dirty_data()
1622  * will come out of the lock with the buffer dirty, which makes it
1623  * ineligible for release here.
1624  *
1625  * Who else is affected by this?  hmm...  Really the only contender
1626  * is do_get_write_access() - it could be looking at the buffer while
1627  * journal_try_to_free_buffer() is changing its state.  But that
1628  * cannot happen because we never reallocate freed data as metadata
1629  * while the data is part of a transaction.  Yes?
1630  */
1631 int journal_try_to_free_buffers(journal_t *journal,
1632                                 struct page *page, int unused_gfp_mask)
1633 {
1634         struct buffer_head *head;
1635         struct buffer_head *bh;
1636         int ret = 0;
1637
1638         J_ASSERT(PageLocked(page));
1639
1640         head = page_buffers(page);
1641         bh = head;
1642         do {
1643                 struct journal_head *jh;
1644
1645                 /*
1646                  * We take our own ref against the journal_head here to avoid
1647                  * having to add tons of locking around each instance of
1648                  * journal_remove_journal_head() and journal_put_journal_head().
1649                  */
1650                 jh = journal_grab_journal_head(bh);
1651                 if (!jh)
1652                         continue;
1653
1654                 jbd_lock_bh_state(bh);
1655                 __journal_try_to_free_buffer(journal, bh);
1656                 journal_put_journal_head(jh);
1657                 jbd_unlock_bh_state(bh);
1658                 if (buffer_jbd(bh))
1659                         goto busy;
1660         } while ((bh = bh->b_this_page) != head);
1661         ret = try_to_free_buffers(page);
1662 busy:
1663         return ret;
1664 }
1665
1666 /*
1667  * This buffer is no longer needed.  If it is on an older transaction's
1668  * checkpoint list we need to record it on this transaction's forget list
1669  * to pin this buffer (and hence its checkpointing transaction) down until
1670  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1671  * release it.
1672  * Returns non-zero if JBD no longer has an interest in the buffer.
1673  *
1674  * Called under j_list_lock.
1675  *
1676  * Called under jbd_lock_bh_state(bh).
1677  */
1678 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1679 {
1680         int may_free = 1;
1681         struct buffer_head *bh = jh2bh(jh);
1682
1683         __journal_unfile_buffer(jh);
1684
1685         if (jh->b_cp_transaction) {
1686                 JBUFFER_TRACE(jh, "on running+cp transaction");
1687                 __journal_file_buffer(jh, transaction, BJ_Forget);
1688                 clear_buffer_jbddirty(bh);
1689                 may_free = 0;
1690         } else {
1691                 JBUFFER_TRACE(jh, "on running transaction");
1692                 journal_remove_journal_head(bh);
1693                 __brelse(bh);
1694         }
1695         return may_free;
1696 }
1697
1698 /*
1699  * journal_invalidatepage
1700  *
1701  * This code is tricky.  It has a number of cases to deal with.
1702  *
1703  * There are two invariants which this code relies on:
1704  *
1705  * i_size must be updated on disk before we start calling invalidatepage on the
1706  * data.
1707  *
1708  *  This is done in ext3 by defining an ext3_setattr method which
1709  *  updates i_size before truncate gets going.  By maintaining this
1710  *  invariant, we can be sure that it is safe to throw away any buffers
1711  *  attached to the current transaction: once the transaction commits,
1712  *  we know that the data will not be needed.
1713  *
1714  *  Note however that we can *not* throw away data belonging to the
1715  *  previous, committing transaction!
1716  *
1717  * Any disk blocks which *are* part of the previous, committing
1718  * transaction (and which therefore cannot be discarded immediately) are
1719  * not going to be reused in the new running transaction
1720  *
1721  *  The bitmap committed_data images guarantee this: any block which is
1722  *  allocated in one transaction and removed in the next will be marked
1723  *  as in-use in the committed_data bitmap, so cannot be reused until
1724  *  the next transaction to delete the block commits.  This means that
1725  *  leaving committing buffers dirty is quite safe: the disk blocks
1726  *  cannot be reallocated to a different file and so buffer aliasing is
1727  *  not possible.
1728  *
1729  *
1730  * The above applies mainly to ordered data mode.  In writeback mode we
1731  * don't make guarantees about the order in which data hits disk --- in
1732  * particular we don't guarantee that new dirty data is flushed before
1733  * transaction commit --- so it is always safe just to discard data
1734  * immediately in that mode.  --sct
1735  */
1736
1737 /*
1738  * The journal_unmap_buffer helper function returns zero if the buffer
1739  * concerned remains pinned as an anonymous buffer belonging to an older
1740  * transaction.
1741  *
1742  * We're outside-transaction here.  Either or both of j_running_transaction
1743  * and j_committing_transaction may be NULL.
1744  */
1745 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1746 {
1747         transaction_t *transaction;
1748         struct journal_head *jh;
1749         int may_free = 1;
1750         int ret;
1751
1752         BUFFER_TRACE(bh, "entry");
1753
1754         /*
1755          * It is safe to proceed here without the j_list_lock because the
1756          * buffers cannot be stolen by try_to_free_buffers as long as we are
1757          * holding the page lock. --sct
1758          */
1759
1760         if (!buffer_jbd(bh))
1761                 goto zap_buffer_unlocked;
1762
1763         spin_lock(&journal->j_state_lock);
1764         jbd_lock_bh_state(bh);
1765         spin_lock(&journal->j_list_lock);
1766
1767         /*
1768          * Now we have the locks, check again to see whether kjournald has
1769          * taken the buffer off the transaction.
1770          */
1771         if (!buffer_jbd(bh))
1772                 goto zap_buffer;
1773
1774         jh = bh2jh(bh);
1775         transaction = jh->b_transaction;
1776         if (transaction == NULL) {
1777                 /* First case: not on any transaction.  If it
1778                  * has no checkpoint link, then we can zap it:
1779                  * it's a writeback-mode buffer so we don't care
1780                  * if it hits disk safely. */
1781                 if (!jh->b_cp_transaction) {
1782                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1783                         goto zap_buffer;
1784                 }
1785
1786                 if (!buffer_dirty(bh)) {
1787                         /* bdflush has written it.  We can drop it now */
1788                         goto zap_buffer;
1789                 }
1790
1791                 /* OK, it must be in the journal but still not
1792                  * written fully to disk: it's metadata or
1793                  * journaled data... */
1794
1795                 if (journal->j_running_transaction) {
1796                         /* ... and once the current transaction has
1797                          * committed, the buffer won't be needed any
1798                          * longer. */
1799                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1800                         ret = __dispose_buffer(jh,
1801                                         journal->j_running_transaction);
1802                         spin_unlock(&journal->j_list_lock);
1803                         jbd_unlock_bh_state(bh);
1804                         spin_unlock(&journal->j_state_lock);
1805                         return ret;
1806                 } else {
1807                         /* There is no currently-running transaction. So the
1808                          * orphan record which we wrote for this file must have
1809                          * passed into commit.  We must attach this buffer to
1810                          * the committing transaction, if it exists. */
1811                         if (journal->j_committing_transaction) {
1812                                 JBUFFER_TRACE(jh, "give to committing trans");
1813                                 ret = __dispose_buffer(jh,
1814                                         journal->j_committing_transaction);
1815                                 spin_unlock(&journal->j_list_lock);
1816                                 jbd_unlock_bh_state(bh);
1817                                 spin_unlock(&journal->j_state_lock);
1818                                 return ret;
1819                         } else {
1820                                 /* The orphan record's transaction has
1821                                  * committed.  We can cleanse this buffer */
1822                                 clear_buffer_jbddirty(bh);
1823                                 goto zap_buffer;
1824                         }
1825                 }
1826         } else if (transaction == journal->j_committing_transaction) {
1827                 /* If it is committing, we simply cannot touch it.  We
1828                  * can remove it's next_transaction pointer from the
1829                  * running transaction if that is set, but nothing
1830                  * else. */
1831                 JBUFFER_TRACE(jh, "on committing transaction");
1832                 set_buffer_freed(bh);
1833                 if (jh->b_next_transaction) {
1834                         J_ASSERT(jh->b_next_transaction ==
1835                                         journal->j_running_transaction);
1836                         jh->b_next_transaction = NULL;
1837                 }
1838                 spin_unlock(&journal->j_list_lock);
1839                 jbd_unlock_bh_state(bh);
1840                 spin_unlock(&journal->j_state_lock);
1841                 return 0;
1842         } else {
1843                 /* Good, the buffer belongs to the running transaction.
1844                  * We are writing our own transaction's data, not any
1845                  * previous one's, so it is safe to throw it away
1846                  * (remember that we expect the filesystem to have set
1847                  * i_size already for this truncate so recovery will not
1848                  * expose the disk blocks we are discarding here.) */
1849                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1850                 may_free = __dispose_buffer(jh, transaction);
1851         }
1852
1853 zap_buffer:
1854         spin_unlock(&journal->j_list_lock);
1855         jbd_unlock_bh_state(bh);
1856         spin_unlock(&journal->j_state_lock);
1857 zap_buffer_unlocked:
1858         clear_buffer_dirty(bh);
1859         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1860         clear_buffer_mapped(bh);
1861         clear_buffer_req(bh);
1862         clear_buffer_new(bh);
1863         bh->b_bdev = NULL;
1864         return may_free;
1865 }
1866
1867 /**
1868  * int journal_invalidatepage()
1869  * @journal: journal to use for flush...
1870  * @page:    page to flush
1871  * @offset:  length of page to invalidate.
1872  *
1873  * Reap page buffers containing data after offset in page.
1874  *
1875  * Return non-zero if the page's buffers were successfully reaped.
1876  */
1877 int journal_invalidatepage(journal_t *journal,
1878                       struct page *page,
1879                       unsigned long offset)
1880 {
1881         struct buffer_head *head, *bh, *next;
1882         unsigned int curr_off = 0;
1883         int may_free = 1;
1884
1885         if (!PageLocked(page))
1886                 BUG();
1887         if (!page_has_buffers(page))
1888                 return 1;
1889
1890         /* We will potentially be playing with lists other than just the
1891          * data lists (especially for journaled data mode), so be
1892          * cautious in our locking. */
1893
1894         head = bh = page_buffers(page);
1895         do {
1896                 unsigned int next_off = curr_off + bh->b_size;
1897                 next = bh->b_this_page;
1898
1899                 /* AKPM: doing lock_buffer here may be overly paranoid */
1900                 if (offset <= curr_off) {
1901                         /* This block is wholly outside the truncation point */
1902                         lock_buffer(bh);
1903                         may_free &= journal_unmap_buffer(journal, bh);
1904                         unlock_buffer(bh);
1905                 }
1906                 curr_off = next_off;
1907                 bh = next;
1908
1909         } while (bh != head);
1910
1911         if (!offset) {
1912                 if (!may_free || !try_to_free_buffers(page))
1913                         return 0;
1914                 J_ASSERT(!page_has_buffers(page));
1915         }
1916         return 1;
1917 }
1918
1919 /*
1920  * File a buffer on the given transaction list.
1921  */
1922 void __journal_file_buffer(struct journal_head *jh,
1923                         transaction_t *transaction, int jlist)
1924 {
1925         struct journal_head **list = 0;
1926         int was_dirty = 0;
1927         struct buffer_head *bh = jh2bh(jh);
1928
1929         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1930         assert_spin_locked(&transaction->t_journal->j_list_lock);
1931
1932         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1933         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1934                                 jh->b_transaction == 0);
1935
1936         if (jh->b_transaction && jh->b_jlist == jlist)
1937                 return;
1938
1939         /* The following list of buffer states needs to be consistent
1940          * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1941          * state. */
1942
1943         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1944             jlist == BJ_Shadow || jlist == BJ_Forget) {
1945                 if (test_clear_buffer_dirty(bh) ||
1946                     test_clear_buffer_jbddirty(bh))
1947                         was_dirty = 1;
1948         }
1949
1950         if (jh->b_transaction)
1951                 __journal_unfile_buffer(jh);
1952         jh->b_transaction = transaction;
1953
1954         switch (jlist) {
1955         case BJ_None:
1956                 J_ASSERT_JH(jh, !jh->b_committed_data);
1957                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1958                 return;
1959         case BJ_SyncData:
1960                 list = &transaction->t_sync_datalist;
1961                 break;
1962         case BJ_Metadata:
1963                 transaction->t_nr_buffers++;
1964                 list = &transaction->t_buffers;
1965                 break;
1966         case BJ_Forget:
1967                 list = &transaction->t_forget;
1968                 break;
1969         case BJ_IO:
1970                 list = &transaction->t_iobuf_list;
1971                 break;
1972         case BJ_Shadow:
1973                 list = &transaction->t_shadow_list;
1974                 break;
1975         case BJ_LogCtl:
1976                 list = &transaction->t_log_list;
1977                 break;
1978         case BJ_Reserved:
1979                 list = &transaction->t_reserved_list;
1980                 break;
1981         case BJ_Locked:
1982                 list =  &transaction->t_locked_list;
1983                 break;
1984         }
1985
1986         __blist_add_buffer(list, jh);
1987         jh->b_jlist = jlist;
1988
1989         if (was_dirty)
1990                 set_buffer_jbddirty(bh);
1991 }
1992
1993 void journal_file_buffer(struct journal_head *jh,
1994                                 transaction_t *transaction, int jlist)
1995 {
1996         jbd_lock_bh_state(jh2bh(jh));
1997         spin_lock(&transaction->t_journal->j_list_lock);
1998         __journal_file_buffer(jh, transaction, jlist);
1999         spin_unlock(&transaction->t_journal->j_list_lock);
2000         jbd_unlock_bh_state(jh2bh(jh));
2001 }
2002
2003 /*
2004  * Remove a buffer from its current buffer list in preparation for
2005  * dropping it from its current transaction entirely.  If the buffer has
2006  * already started to be used by a subsequent transaction, refile the
2007  * buffer on that transaction's metadata list.
2008  *
2009  * Called under journal->j_list_lock
2010  *
2011  * Called under jbd_lock_bh_state(jh2bh(jh))
2012  */
2013 void __journal_refile_buffer(struct journal_head *jh)
2014 {
2015         int was_dirty;
2016         struct buffer_head *bh = jh2bh(jh);
2017
2018         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2019         if (jh->b_transaction)
2020                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2021
2022         /* If the buffer is now unused, just drop it. */
2023         if (jh->b_next_transaction == NULL) {
2024                 __journal_unfile_buffer(jh);
2025                 return;
2026         }
2027
2028         /*
2029          * It has been modified by a later transaction: add it to the new
2030          * transaction's metadata list.
2031          */
2032
2033         was_dirty = test_clear_buffer_jbddirty(bh);
2034         __journal_unfile_buffer(jh);
2035         jh->b_transaction = jh->b_next_transaction;
2036         jh->b_next_transaction = NULL;
2037         __journal_file_buffer(jh, jh->b_transaction, BJ_Metadata);
2038         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2039
2040         if (was_dirty)
2041                 set_buffer_jbddirty(bh);
2042 }
2043
2044 /*
2045  * For the unlocked version of this call, also make sure that any
2046  * hanging journal_head is cleaned up if necessary.
2047  *
2048  * __journal_refile_buffer is usually called as part of a single locked
2049  * operation on a buffer_head, in which the caller is probably going to
2050  * be hooking the journal_head onto other lists.  In that case it is up
2051  * to the caller to remove the journal_head if necessary.  For the
2052  * unlocked journal_refile_buffer call, the caller isn't going to be
2053  * doing anything else to the buffer so we need to do the cleanup
2054  * ourselves to avoid a jh leak.
2055  *
2056  * *** The journal_head may be freed by this call! ***
2057  */
2058 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2059 {
2060         struct buffer_head *bh = jh2bh(jh);
2061
2062         jbd_lock_bh_state(bh);
2063         spin_lock(&journal->j_list_lock);
2064
2065         __journal_refile_buffer(jh);
2066         jbd_unlock_bh_state(bh);
2067         journal_remove_journal_head(bh);
2068
2069         spin_unlock(&journal->j_list_lock);
2070         __brelse(bh);
2071 }