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 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 out:
 361         return result;
 362 }
 363
 364
 365 /**
 366  * int journal_restart() - restart a handle .
 367  * @handle:  handle to restart
 368  * @nblocks: nr credits requested
 369  *
 370  * Restart a handle for a multi-transaction filesystem
 371  * operation.
 372  *
 373  * If the journal_extend() call above fails to grant new buffer credits
 374  * to a running handle, a call to journal_restart will commit the
 375  * handle's transaction so far and reattach the handle to a new
 376  * transaction capabable of guaranteeing the requested number of
 377  * credits.
 378  */
 379
 380 int journal_restart(handle_t *handle, int nblocks)
 381 {
 382         transaction_t *transaction = handle->h_transaction;
 383         journal_t *journal = transaction->t_journal;
 384         int ret;
 385
 386         /* If we've had an abort of any type, don't even think about
 387          * actually doing the restart! */
 388         if (is_handle_aborted(handle))
 389                 return 0;
 390
 391         /*
 392          * First unlink the handle from its current transaction, and start the
 393          * commit on that.
 394          */
 395         J_ASSERT(transaction->t_updates > 0);
 396         J_ASSERT(journal_current_handle() == handle);
 397
 398         spin_lock(&journal->j_state_lock);
 399         spin_lock(&transaction->t_handle_lock);
 400         transaction->t_outstanding_credits -= handle->h_buffer_credits;
 401         transaction->t_updates--;
 402
 403         if (!transaction->t_updates)
 404                 wake_up(&journal->j_wait_updates);
 405         spin_unlock(&transaction->t_handle_lock);
 406
 407         jbd_debug(2, "restarting handle %p\n", handle);
 408         __log_start_commit(journal, transaction->t_tid);
 409         spin_unlock(&journal->j_state_lock);
 410
 411         handle->h_buffer_credits = nblocks;
 412         ret = start_this_handle(journal, handle);
 413         return ret;
 414 }
 415
 416
 417 /**
 418  * void journal_lock_updates () - establish a transaction barrier.
 419  * @journal:  Journal to establish a barrier on.
 420  *
 421  * This locks out any further updates from being started, and blocks
 422  * until all existing updates have completed, returning only once the
 423  * journal is in a quiescent state with no updates running.
 424  *
 425  * The journal lock should not be held on entry.
 426  */
 427 void journal_lock_updates(journal_t *journal)
 428 {
 429         DEFINE_WAIT(wait);
 430
 431         spin_lock(&journal->j_state_lock);
 432         ++journal->j_barrier_count;
 433
 434         /* Wait until there are no running updates */
 435         while (1) {
 436                 transaction_t *transaction = journal->j_running_transaction;
 437
 438                 if (!transaction)
 439                         break;
 440
 441                 spin_lock(&transaction->t_handle_lock);
 442                 if (!transaction->t_updates) {
 443                         spin_unlock(&transaction->t_handle_lock);
 444                         break;
 445                 }
 446                 prepare_to_wait(&journal->j_wait_updates, &wait,
 447                                 TASK_UNINTERRUPTIBLE);
 448                 spin_unlock(&transaction->t_handle_lock);
 449                 spin_unlock(&journal->j_state_lock);
 450                 schedule();
 451                 finish_wait(&journal->j_wait_updates, &wait);
 452                 spin_lock(&journal->j_state_lock);
 453         }
 454         spin_unlock(&journal->j_state_lock);
 455
 456         /*
 457          * We have now established a barrier against other normal updates, but
 458          * we also need to barrier against other journal_lock_updates() calls
 459          * to make sure that we serialise special journal-locked operations
 460          * too.
 461          */
 462         down(&journal->j_barrier);
 463 }
 464
 465 /**
 466  * void journal_unlock_updates (journal_t* journal) - release barrier
 467  * @journal:  Journal to release the barrier on.
 468  *
 469  * Release a transaction barrier obtained with journal_lock_updates().
 470  *
 471  * Should be called without the journal lock held.
 472  */
 473 void journal_unlock_updates (journal_t *journal)
 474 {
 475         J_ASSERT(journal->j_barrier_count != 0);
 476
 477         up(&journal->j_barrier);
 478         spin_lock(&journal->j_state_lock);
 479         --journal->j_barrier_count;
 480         spin_unlock(&journal->j_state_lock);
 481         wake_up(&journal->j_wait_transaction_locked);
 482 }
 483
 484 /*
 485  * Report any unexpected dirty buffers which turn up.  Normally those
 486  * indicate an error, but they can occur if the user is running (say)
 487  * tune2fs to modify the live filesystem, so we need the option of
 488  * continuing as gracefully as possible.  #
 489  *
 490  * The caller should already hold the journal lock and
 491  * j_list_lock spinlock: most callers will need those anyway
 492  * in order to probe the buffer's journaling state safely.
 493  */
 494 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
 495 {
 496         struct buffer_head *bh = jh2bh(jh);
 497         int jlist;
 498
 499         if (buffer_dirty(bh)) {
 500                 /* If this buffer is one which might reasonably be dirty
 501                  * --- ie. data, or not part of this journal --- then
 502                  * we're OK to leave it alone, but otherwise we need to
 503                  * move the dirty bit to the journal's own internal
 504                  * JBDDirty bit. */
 505                 jlist = jh->b_jlist;
 506
 507                 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
 508                     jlist == BJ_Shadow || jlist == BJ_Forget) {
 509                         if (test_clear_buffer_dirty(jh2bh(jh))) {
 510                                 set_bit(BH_JBDDirty, &jh2bh(jh)->b_state);
 511                         }
 512                 }
 513         }
 514 }
 515
 516 /*
 517  * If the buffer is already part of the current transaction, then there
 518  * is nothing we need to do.  If it is already part of a prior
 519  * transaction which we are still committing to disk, then we need to
 520  * make sure that we do not overwrite the old copy: we do copy-out to
 521  * preserve the copy going to disk.  We also account the buffer against
 522  * the handle's metadata buffer credits (unless the buffer is already
 523  * part of the transaction, that is).
 524  *
 525  */
 526 static int
 527 do_get_write_access(handle_t *handle, struct journal_head *jh,
 528                         int force_copy, int *credits)
 529 {
 530         struct buffer_head *bh;
 531         transaction_t *transaction;
 532         journal_t *journal;
 533         int error;
 534         char *frozen_buffer = NULL;
 535         int need_copy = 0;
 536
 537         if (is_handle_aborted(handle))
 538                 return -EROFS;
 539
 540         transaction = handle->h_transaction;
 541         journal = transaction->t_journal;
 542
 543         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
 544
 545         JBUFFER_TRACE(jh, "entry");
 546 repeat:
 547         bh = jh2bh(jh);
 548
 549         /* @@@ Need to check for errors here at some point. */
 550
 551         lock_buffer(bh);
 552         jbd_lock_bh_state(bh);
 553
 554         /* We now hold the buffer lock so it is safe to query the buffer
 555          * state.  Is the buffer dirty?
 556          *
 557          * If so, there are two possibilities.  The buffer may be
 558          * non-journaled, and undergoing a quite legitimate writeback.
 559          * Otherwise, it is journaled, and we don't expect dirty buffers
 560          * in that state (the buffers should be marked JBD_Dirty
 561          * instead.)  So either the IO is being done under our own
 562          * control and this is a bug, or it's a third party IO such as
 563          * dump(8) (which may leave the buffer scheduled for read ---
 564          * ie. locked but not dirty) or tune2fs (which may actually have
 565          * the buffer dirtied, ugh.)  */
 566
 567         if (buffer_dirty(bh)) {
 568                 /*
 569                  * First question: is this buffer already part of the current
 570                  * transaction or the existing committing transaction?
 571                  */
 572                 if (jh->b_transaction) {
 573                         J_ASSERT_JH(jh,
 574                                 jh->b_transaction == transaction ||
 575                                 jh->b_transaction ==
 576                                         journal->j_committing_transaction);
 577                         if (jh->b_next_transaction)
 578                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
 579                                                         transaction);
 580                         JBUFFER_TRACE(jh, "Unexpected dirty buffer");
 581                         jbd_unexpected_dirty_buffer(jh);
 582                 }
 583         }
 584
 585         unlock_buffer(bh);
 586
 587         error = -EROFS;
 588         if (is_handle_aborted(handle)) {
 589                 jbd_unlock_bh_state(bh);
 590                 goto out;
 591         }
 592         error = 0;
 593
 594         /*
 595          * The buffer is already part of this transaction if b_transaction or
 596          * b_next_transaction points to it
 597          */
 598         if (jh->b_transaction == transaction ||
 599             jh->b_next_transaction == transaction)
 600                 goto done;
 601
 602         /*
 603          * If there is already a copy-out version of this buffer, then we don't
 604          * need to make another one
 605          */
 606         if (jh->b_frozen_data) {
 607                 JBUFFER_TRACE(jh, "has frozen data");
 608                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 609                 jh->b_next_transaction = transaction;
 610
 611                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
 612                 handle->h_buffer_credits--;
 613                 if (credits)
 614                         (*credits)++;
 615                 goto done;
 616         }
 617
 618         /* Is there data here we need to preserve? */
 619
 620         if (jh->b_transaction && jh->b_transaction != transaction) {
 621                 JBUFFER_TRACE(jh, "owned by older transaction");
 622                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 623                 J_ASSERT_JH(jh, jh->b_transaction ==
 624                                         journal->j_committing_transaction);
 625
 626                 /* There is one case we have to be very careful about.
 627                  * If the committing transaction is currently writing
 628                  * this buffer out to disk and has NOT made a copy-out,
 629                  * then we cannot modify the buffer contents at all
 630                  * right now.  The essence of copy-out is that it is the
 631                  * extra copy, not the primary copy, which gets
 632                  * journaled.  If the primary copy is already going to
 633                  * disk then we cannot do copy-out here. */
 634
 635                 if (jh->b_jlist == BJ_Shadow) {
 636                         wait_queue_head_t *wqh;
 637                         DEFINE_WAIT(wait);
 638
 639                         JBUFFER_TRACE(jh, "on shadow: sleep");
 640                         jbd_unlock_bh_state(bh);
 641                         /* commit wakes up all shadow buffers after IO */
 642                         wqh = bh_waitq_head(bh);
 643                         for ( ; ; ) {
 644                                 prepare_to_wait(wqh, &wait,
 645                                                 TASK_UNINTERRUPTIBLE);
 646                                 if (jh->b_jlist != BJ_Shadow)
 647                                         break;
 648                                 schedule();
 649                         }
 650                         finish_wait(wqh, &wait);
 651                         goto repeat;
 652                 }
 653
 654                 /* Only do the copy if the currently-owning transaction
 655                  * still needs it.  If it is on the Forget list, the
 656                  * committing transaction is past that stage.  The
 657                  * buffer had better remain locked during the kmalloc,
 658                  * but that should be true --- we hold the journal lock
 659                  * still and the buffer is already on the BUF_JOURNAL
 660                  * list so won't be flushed.
 661                  *
 662                  * Subtle point, though: if this is a get_undo_access,
 663                  * then we will be relying on the frozen_data to contain
 664                  * the new value of the committed_data record after the
 665                  * transaction, so we HAVE to force the frozen_data copy
 666                  * in that case. */
 667
 668                 if (jh->b_jlist != BJ_Forget || force_copy) {
 669                         JBUFFER_TRACE(jh, "generate frozen data");
 670                         if (!frozen_buffer) {
 671                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
 672                                 jbd_unlock_bh_state(bh);
 673                                 frozen_buffer = jbd_kmalloc(jh2bh(jh)->b_size,
 674                                                             GFP_NOFS);
 675                                 if (!frozen_buffer) {
 676                                         printk(KERN_EMERG
 677                                                "%s: OOM for frozen_buffer\n",
 678                                                __FUNCTION__);
 679                                         JBUFFER_TRACE(jh, "oom!");
 680                                         error = -ENOMEM;
 681                                         jbd_lock_bh_state(bh);
 682                                         goto done;
 683                                 }
 684                                 goto repeat;
 685                         }
 686                         jh->b_frozen_data = frozen_buffer;
 687                         frozen_buffer = NULL;
 688                         need_copy = 1;
 689                 }
 690                 jh->b_next_transaction = transaction;
 691         }
 692
 693         J_ASSERT(handle->h_buffer_credits > 0);
 694         handle->h_buffer_credits--;
 695         if (credits)
 696                 (*credits)++;
 697
 698         /*
 699          * Finally, if the buffer is not journaled right now, we need to make
 700          * sure it doesn't get written to disk before the caller actually
 701          * commits the new data
 702          */
 703         if (!jh->b_transaction) {
 704                 JBUFFER_TRACE(jh, "no transaction");
 705                 J_ASSERT_JH(jh, !jh->b_next_transaction);
 706                 jh->b_transaction = transaction;
 707                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
 708                 spin_lock(&journal->j_list_lock);
 709                 __journal_file_buffer(jh, transaction, BJ_Reserved);
 710                 spin_unlock(&journal->j_list_lock);
 711         }
 712
 713 done:
 714         if (need_copy) {
 715                 struct page *page;
 716                 int offset;
 717                 char *source;
 718
 719                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
 720                             "Possible IO failure.\n");
 721                 page = jh2bh(jh)->b_page;
 722                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
 723                 source = kmap_atomic(page, KM_USER0);
 724                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
 725                 kunmap_atomic(source, KM_USER0);
 726         }
 727         jbd_unlock_bh_state(bh);
 728
 729         /*
 730          * If we are about to journal a buffer, then any revoke pending on it is
 731          * no longer valid
 732          */
 733         journal_cancel_revoke(handle, jh);
 734
 735 out:
 736         if (frozen_buffer)
 737                 kfree(frozen_buffer);
 738
 739         JBUFFER_TRACE(jh, "exit");
 740         return error;
 741 }
 742
 743 /**
 744  * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
 745  * @handle: transaction to add buffer modifications to
 746  * @bh:     bh to be used for metadata writes
 747  *
 748  * Returns an error code or 0 on success.
 749  *
 750  * In full data journalling mode the buffer may be of type BJ_AsyncData,
 751  * because we're write()ing a buffer which is also part of a shared mapping.
 752  */
 753
 754 int journal_get_write_access(handle_t *handle,
 755                         struct buffer_head *bh, int *credits)
 756 {
 757         struct journal_head *jh = journal_add_journal_head(bh);
 758         int rc;
 759
 760         /* We do not want to get caught playing with fields which the
 761          * log thread also manipulates.  Make sure that the buffer
 762          * completes any outstanding IO before proceeding. */
 763         rc = do_get_write_access(handle, jh, 0, credits);
 764         journal_put_journal_head(jh);
 765         return rc;
 766 }
 767
 768
 769 /*
 770  * When the user wants to journal a newly created buffer_head
 771  * (ie. getblk() returned a new buffer and we are going to populate it
 772  * manually rather than reading off disk), then we need to keep the
 773  * buffer_head locked until it has been completely filled with new
 774  * data.  In this case, we should be able to make the assertion that
 775  * the bh is not already part of an existing transaction.
 776  *
 777  * The buffer should already be locked by the caller by this point.
 778  * There is no lock ranking violation: it was a newly created,
 779  * unlocked buffer beforehand. */
 780
 781 /**
 782  * int journal_get_create_access () - notify intent to use newly created bh
 783  * @handle: transaction to new buffer to
 784  * @bh: new buffer.
 785  *
 786  * Call this if you create a new bh.
 787  */
 788 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
 789 {
 790         transaction_t *transaction = handle->h_transaction;
 791         journal_t *journal = transaction->t_journal;
 792         struct journal_head *jh = journal_add_journal_head(bh);
 793         int err;
 794
 795         jbd_debug(5, "journal_head %p\n", jh);
 796         err = -EROFS;
 797         if (is_handle_aborted(handle))
 798                 goto out;
 799         err = 0;
 800
 801         JBUFFER_TRACE(jh, "entry");
 802         /*
 803          * The buffer may already belong to this transaction due to pre-zeroing
 804          * in the filesystem's new_block code.  It may also be on the previous,
 805          * committing transaction's lists, but it HAS to be in Forget state in
 806          * that case: the transaction must have deleted the buffer for it to be
 807          * reused here.
 808          */
 809         jbd_lock_bh_state(bh);
 810         spin_lock(&journal->j_list_lock);
 811         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
 812                 jh->b_transaction == NULL ||
 813                 (jh->b_transaction == journal->j_committing_transaction &&
 814                           jh->b_jlist == BJ_Forget)));
 815
 816         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 817         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
 818
 819         J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
 820         handle->h_buffer_credits--;
 821
 822         if (jh->b_transaction == NULL) {
 823                 jh->b_transaction = transaction;
 824                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
 825                 __journal_file_buffer(jh, transaction, BJ_Reserved);
 826         } else if (jh->b_transaction == journal->j_committing_transaction) {
 827                 JBUFFER_TRACE(jh, "set next transaction");
 828                 jh->b_next_transaction = transaction;
 829         }
 830         spin_unlock(&journal->j_list_lock);
 831         jbd_unlock_bh_state(bh);
 832
 833         /*
 834          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
 835          * blocks which contain freed but then revoked metadata.  We need
 836          * to cancel the revoke in case we end up freeing it yet again
 837          * and the reallocating as data - this would cause a second revoke,
 838          * which hits an assertion error.
 839          */
 840         JBUFFER_TRACE(jh, "cancelling revoke");
 841         journal_cancel_revoke(handle, jh);
 842         journal_put_journal_head(jh);
 843 out:
 844         return err;
 845 }
 846
 847 /**
 848  * int journal_get_undo_access() -  Notify intent to modify metadata with
 849  *     non-rewindable consequences
 850  * @handle: transaction
 851  * @bh: buffer to undo
 852  * @credits: store the number of taken credits here (if not NULL)
 853  *
 854  * Sometimes there is a need to distinguish between metadata which has
 855  * been committed to disk and that which has not.  The ext3fs code uses
 856  * this for freeing and allocating space, we have to make sure that we
 857  * do not reuse freed space until the deallocation has been committed,
 858  * since if we overwrote that space we would make the delete
 859  * un-rewindable in case of a crash.
 860  *
 861  * To deal with that, journal_get_undo_access requests write access to a
 862  * buffer for parts of non-rewindable operations such as delete
 863  * operations on the bitmaps.  The journaling code must keep a copy of
 864  * the buffer's contents prior to the undo_access call until such time
 865  * as we know that the buffer has definitely been committed to disk.
 866  *
 867  * We never need to know which transaction the committed data is part
 868  * of, buffers touched here are guaranteed to be dirtied later and so
 869  * will be committed to a new transaction in due course, at which point
 870  * we can discard the old committed data pointer.
 871  *
 872  * Returns error number or 0 on success.
 873  */
 874 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh,
 875                                 int *credits)
 876 {
 877         int err;
 878         struct journal_head *jh = journal_add_journal_head(bh);
 879         char *committed_data = NULL;
 880
 881         JBUFFER_TRACE(jh, "entry");
 882
 883         /*
 884          * Do this first --- it can drop the journal lock, so we want to
 885          * make sure that obtaining the committed_data is done
 886          * atomically wrt. completion of any outstanding commits.
 887          */
 888         err = do_get_write_access(handle, jh, 1, credits);
 889         if (err)
 890                 goto out;
 891
 892 repeat:
 893         if (!jh->b_committed_data) {
 894                 committed_data = jbd_kmalloc(jh2bh(jh)->b_size, GFP_NOFS);
 895                 if (!committed_data) {
 896                         printk(KERN_EMERG "%s: No memory for committed data\n",
 897                                 __FUNCTION__);
 898                         err = -ENOMEM;
 899                         goto out;
 900                 }
 901         }
 902
 903         jbd_lock_bh_state(bh);
 904         if (!jh->b_committed_data) {
 905                 /* Copy out the current buffer contents into the
 906                  * preserved, committed copy. */
 907                 JBUFFER_TRACE(jh, "generate b_committed data");
 908                 if (!committed_data) {
 909                         jbd_unlock_bh_state(bh);
 910                         goto repeat;
 911                 }
 912
 913                 jh->b_committed_data = committed_data;
 914                 committed_data = NULL;
 915                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
 916         }
 917         jbd_unlock_bh_state(bh);
 918 out:
 919         journal_put_journal_head(jh);
 920         if (committed_data)
 921                 kfree(committed_data);
 922         return err;
 923 }
 924
 925 /**
 926  * int journal_dirty_data() -  mark a buffer as containing dirty data which
 927  *                             needs to be flushed before we can commit the
 928  *                             current transaction.
 929  * @handle: transaction
 930  * @bh: bufferhead to mark
 931  *
 932  * The buffer is placed on the transaction's data list and is marked as
 933  * belonging to the transaction.
 934  *
 935  * Returns error number or 0 on success.
 936  *
 937  * journal_dirty_data() can be called via page_launder->ext3_writepage
 938  * by kswapd.
 939  */
 940 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
 941 {
 942         journal_t *journal = handle->h_transaction->t_journal;
 943         int need_brelse = 0;
 944         struct journal_head *jh;
 945
 946         if (is_handle_aborted(handle))
 947                 return 0;
 948
 949         jh = journal_add_journal_head(bh);
 950         JBUFFER_TRACE(jh, "entry");
 951
 952         /*
 953          * The buffer could *already* be dirty.  Writeout can start
 954          * at any time.
 955          */
 956         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
 957
 958         /*
 959          * What if the buffer is already part of a running transaction?
 960          *
 961          * There are two cases:
 962          * 1) It is part of the current running transaction.  Refile it,
 963          *    just in case we have allocated it as metadata, deallocated
 964          *    it, then reallocated it as data.
 965          * 2) It is part of the previous, still-committing transaction.
 966          *    If all we want to do is to guarantee that the buffer will be
 967          *    written to disk before this new transaction commits, then
 968          *    being sure that the *previous* transaction has this same
 969          *    property is sufficient for us!  Just leave it on its old
 970          *    transaction.
 971          *
 972          * In case (2), the buffer must not already exist as metadata
 973          * --- that would violate write ordering (a transaction is free
 974          * to write its data at any point, even before the previous
 975          * committing transaction has committed).  The caller must
 976          * never, ever allow this to happen: there's nothing we can do
 977          * about it in this layer.
 978          */
 979         jbd_lock_bh_state(bh);
 980         spin_lock(&journal->j_list_lock);
 981         if (jh->b_transaction) {
 982                 JBUFFER_TRACE(jh, "has transaction");
 983                 if (jh->b_transaction != handle->h_transaction) {
 984                         JBUFFER_TRACE(jh, "belongs to older transaction");
 985                         J_ASSERT_JH(jh, jh->b_transaction ==
 986                                         journal->j_committing_transaction);
 987
 988                         /* @@@ IS THIS TRUE  ? */
 989                         /*
 990                          * Not any more.  Scenario: someone does a write()
 991                          * in data=journal mode.  The buffer's transaction has
 992                          * moved into commit.  Then someone does another
 993                          * write() to the file.  We do the frozen data copyout
 994                          * and set b_next_transaction to point to j_running_t.
 995                          * And while we're in that state, someone does a
 996                          * writepage() in an attempt to pageout the same area
 997                          * of the file via a shared mapping.  At present that
 998                          * calls journal_dirty_data(), and we get right here.
 999                          * It may be too late to journal the data.  Simply
1000                          * falling through to the next test will suffice: the
1001                          * data will be dirty and wil be checkpointed.  The
1002                          * ordering comments in the next comment block still
1003                          * apply.
1004                          */
1005                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1006
1007                         /*
1008                          * If we're journalling data, and this buffer was
1009                          * subject to a write(), it could be metadata, forget
1010                          * or shadow against the committing transaction.  Now,
1011                          * someone has dirtied the same darn page via a mapping
1012                          * and it is being writepage()'d.
1013                          * We *could* just steal the page from commit, with some
1014                          * fancy locking there.  Instead, we just skip it -
1015                          * don't tie the page's buffers to the new transaction
1016                          * at all.
1017                          * Implication: if we crash before the writepage() data
1018                          * is written into the filesystem, recovery will replay
1019                          * the write() data.
1020                          */
1021                         if (jh->b_jlist != BJ_None &&
1022                                         jh->b_jlist != BJ_SyncData &&
1023                                         jh->b_jlist != BJ_Locked) {
1024                                 JBUFFER_TRACE(jh, "Not stealing");
1025                                 goto no_journal;
1026                         }
1027
1028                         /*
1029                          * This buffer may be undergoing writeout in commit.  We
1030                          * can't return from here and let the caller dirty it
1031                          * again because that can cause the write-out loop in
1032                          * commit to never terminate.
1033                          */
1034                         if (buffer_dirty(bh)) {
1035                                 get_bh(bh);
1036                                 spin_unlock(&journal->j_list_lock);
1037                                 jbd_unlock_bh_state(bh);
1038                                 need_brelse = 1;
1039                                 sync_dirty_buffer(bh);
1040                                 jbd_lock_bh_state(bh);
1041                                 spin_lock(&journal->j_list_lock);
1042                                 /* The buffer may become locked again at any
1043                                    time if it is redirtied */
1044                         }
1045
1046                         /* journal_clean_data_list() may have got there first */
1047                         if (jh->b_transaction != NULL) {
1048                                 JBUFFER_TRACE(jh, "unfile from commit");
1049                                 __journal_unfile_buffer(jh);
1050                         }
1051                         /* The buffer will be refiled below */
1052
1053                 }
1054                 /*
1055                  * Special case --- the buffer might actually have been
1056                  * allocated and then immediately deallocated in the previous,
1057                  * committing transaction, so might still be left on that
1058                  * transaction's metadata lists.
1059                  */
1060                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1061                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1062                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1063                         __journal_unfile_buffer(jh);
1064                         JBUFFER_TRACE(jh, "file as data");
1065                         __journal_file_buffer(jh, handle->h_transaction,
1066                                                 BJ_SyncData);
1067                 }
1068         } else {
1069                 JBUFFER_TRACE(jh, "not on a transaction");
1070                 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1071         }
1072 no_journal:
1073         spin_unlock(&journal->j_list_lock);
1074         jbd_unlock_bh_state(bh);
1075         if (need_brelse) {
1076                 BUFFER_TRACE(bh, "brelse");
1077                 __brelse(bh);
1078         }
1079         JBUFFER_TRACE(jh, "exit");
1080         journal_put_journal_head(jh);
1081         return 0;
1082 }
1083
1084 /**
1085  * int journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1086  * @handle: transaction to add buffer to.
1087  * @bh: buffer to mark
1088  *
1089  * mark dirty metadata which needs to be journaled as part of the current
1090  * transaction.
1091  *
1092  * The buffer is placed on the transaction's metadata list and is marked
1093  * as belonging to the transaction.
1094  *
1095  * Returns error number or 0 on success.
1096  *
1097  * Special care needs to be taken if the buffer already belongs to the
1098  * current committing transaction (in which case we should have frozen
1099  * data present for that commit).  In that case, we don't relink the
1100  * buffer: that only gets done when the old transaction finally
1101  * completes its commit.
1102  */
1103 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1104 {
1105         transaction_t *transaction = handle->h_transaction;
1106         journal_t *journal = transaction->t_journal;
1107         struct journal_head *jh = bh2jh(bh);
1108
1109         jbd_debug(5, "journal_head %p\n", jh);
1110         JBUFFER_TRACE(jh, "entry");
1111         if (is_handle_aborted(handle))
1112                 goto out;
1113
1114         jbd_lock_bh_state(bh);
1115
1116         /*
1117          * fastpath, to avoid expensive locking.  If this buffer is already
1118          * on the running transaction's metadata list there is nothing to do.
1119          * Nobody can take it off again because there is a handle open.
1120          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1121          * result in this test being false, so we go in and take the locks.
1122          */
1123         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1124                 JBUFFER_TRACE(jh, "fastpath");
1125                 J_ASSERT_JH(jh, jh->b_transaction ==
1126                                         journal->j_running_transaction);
1127                 goto out_unlock_bh;
1128         }
1129
1130         set_buffer_jbddirty(bh);
1131
1132         /*
1133          * Metadata already on the current transaction list doesn't
1134          * need to be filed.  Metadata on another transaction's list must
1135          * be committing, and will be refiled once the commit completes:
1136          * leave it alone for now.
1137          */
1138         if (jh->b_transaction != transaction) {
1139                 JBUFFER_TRACE(jh, "already on other transaction");
1140                 J_ASSERT_JH(jh, jh->b_transaction ==
1141                                         journal->j_committing_transaction);
1142                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1143                 /* And this case is illegal: we can't reuse another
1144                  * transaction's data buffer, ever. */
1145                 goto out_unlock_bh;
1146         }
1147
1148         /* That test should have eliminated the following case: */
1149         J_ASSERT_JH(jh, jh->b_frozen_data == 0);
1150
1151         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1152         spin_lock(&journal->j_list_lock);
1153         __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1154         spin_unlock(&journal->j_list_lock);
1155 out_unlock_bh:
1156         jbd_unlock_bh_state(bh);
1157 out:
1158         JBUFFER_TRACE(jh, "exit");
1159         return 0;
1160 }
1161
1162 /*
1163  * journal_release_buffer: undo a get_write_access without any buffer
1164  * updates, if the update decided in the end that it didn't need access.
1165  *
1166  * The caller passes in the number of credits which should be put back for
1167  * this buffer (zero or one).
1168  *
1169  * We leave the buffer attached to t_reserved_list because even though this
1170  * handle doesn't want it, some other concurrent handle may want to journal
1171  * this buffer.  If that handle is curently in between get_write_access() and
1172  * journal_dirty_metadata() then it expects the buffer to be reserved.  If
1173  * we were to rip it off t_reserved_list here, the other handle will explode
1174  * when journal_dirty_metadata is presented with a non-reserved buffer.
1175  *
1176  * If nobody really wants to journal this buffer then it will be thrown
1177  * away at the start of commit.
1178  */
1179 void
1180 journal_release_buffer(handle_t *handle, struct buffer_head *bh, int credits)
1181 {
1182         BUFFER_TRACE(bh, "entry");
1183         handle->h_buffer_credits += credits;
1184 }
1185
1186 /**
1187  * void journal_forget() - bforget() for potentially-journaled buffers.
1188  * @handle: transaction handle
1189  * @bh:     bh to 'forget'
1190  *
1191  * We can only do the bforget if there are no commits pending against the
1192  * buffer.  If the buffer is dirty in the current running transaction we
1193  * can safely unlink it.
1194  *
1195  * bh may not be a journalled buffer at all - it may be a non-JBD
1196  * buffer which came off the hashtable.  Check for this.
1197  *
1198  * Decrements bh->b_count by one.
1199  *
1200  * Allow this call even if the handle has aborted --- it may be part of
1201  * the caller's cleanup after an abort.
1202  */
1203 void journal_forget(handle_t *handle, struct buffer_head *bh)
1204 {
1205         transaction_t *transaction = handle->h_transaction;
1206         journal_t *journal = transaction->t_journal;
1207         struct journal_head *jh;
1208
1209         BUFFER_TRACE(bh, "entry");
1210
1211         jbd_lock_bh_state(bh);
1212         spin_lock(&journal->j_list_lock);
1213
1214         if (!buffer_jbd(bh))
1215                 goto not_jbd;
1216         jh = bh2jh(bh);
1217
1218         if (jh->b_transaction == handle->h_transaction) {
1219                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1220
1221                 /* If we are forgetting a buffer which is already part
1222                  * of this transaction, then we can just drop it from
1223                  * the transaction immediately. */
1224                 clear_buffer_dirty(bh);
1225                 clear_buffer_jbddirty(bh);
1226
1227                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1228                 J_ASSERT_JH(jh, !jh->b_committed_data);
1229
1230                 __journal_unfile_buffer(jh);
1231
1232                 /*
1233                  * We are no longer going to journal this buffer.
1234                  * However, the commit of this transaction is still
1235                  * important to the buffer: the delete that we are now
1236                  * processing might obsolete an old log entry, so by
1237                  * committing, we can satisfy the buffer's checkpoint.
1238                  *
1239                  * So, if we have a checkpoint on the buffer, we should
1240                  * now refile the buffer on our BJ_Forget list so that
1241                  * we know to remove the checkpoint after we commit.
1242                  */
1243
1244                 if (jh->b_cp_transaction) {
1245                         __journal_file_buffer(jh, transaction, BJ_Forget);
1246                 } else {
1247                         journal_remove_journal_head(bh);
1248                         __brelse(bh);
1249                         if (!buffer_jbd(bh)) {
1250                                 spin_unlock(&journal->j_list_lock);
1251                                 jbd_unlock_bh_state(bh);
1252                                 __bforget(bh);
1253                                 return;
1254                         }
1255                 }
1256         } else if (jh->b_transaction) {
1257                 J_ASSERT_JH(jh, (jh->b_transaction ==
1258                                  journal->j_committing_transaction));
1259                 /* However, if the buffer is still owned by a prior
1260                  * (committing) transaction, we can't drop it yet... */
1261                 JBUFFER_TRACE(jh, "belongs to older transaction");
1262                 /* ... but we CAN drop it from the new transaction if we
1263                  * have also modified it since the original commit. */
1264
1265                 if (jh->b_next_transaction) {
1266                         J_ASSERT(jh->b_next_transaction == transaction);
1267                         jh->b_next_transaction = NULL;
1268                 }
1269         }
1270
1271 not_jbd:
1272         spin_unlock(&journal->j_list_lock);
1273         jbd_unlock_bh_state(bh);
1274         __brelse(bh);
1275         return;
1276 }
1277
1278 /**
1279  * void journal_callback_set() -  Register a callback function for this handle.
1280  * @handle: handle to attach the callback to.
1281  * @func: function to callback.
1282  * @jcb:  structure with additional information required by func() , and
1283  *        some space for jbd internal information.
1284  *
1285  * The function will be
1286  * called when the transaction that this handle is part of has been
1287  * committed to disk with the original callback data struct and the
1288  * error status of the journal as parameters.  There is no guarantee of
1289  * ordering between handles within a single transaction, nor between
1290  * callbacks registered on the same handle.
1291  *
1292  * The caller is responsible for allocating the journal_callback struct.
1293  * This is to allow the caller to add as much extra data to the callback
1294  * as needed, but reduce the overhead of multiple allocations.  The caller
1295  * allocated struct must start with a struct journal_callback at offset 0,
1296  * and has the caller-specific data afterwards.
1297  */
1298 void journal_callback_set(handle_t *handle,
1299                         void (*func)(struct journal_callback *jcb, int error),
1300                         struct journal_callback *jcb)
1301 {
1302         spin_lock(&handle->h_transaction->t_jcb_lock);
1303         list_add_tail(&jcb->jcb_list, &handle->h_jcb);
1304         spin_unlock(&handle->h_transaction->t_jcb_lock);
1305         jcb->jcb_func = func;
1306 }
1307
1308 /**
1309  * int journal_stop() - complete a transaction
1310  * @handle: tranaction to complete.
1311  *
1312  * All done for a particular handle.
1313  *
1314  * There is not much action needed here.  We just return any remaining
1315  * buffer credits to the transaction and remove the handle.  The only
1316  * complication is that we need to start a commit operation if the
1317  * filesystem is marked for synchronous update.
1318  *
1319  * journal_stop itself will not usually return an error, but it may
1320  * do so in unusual circumstances.  In particular, expect it to
1321  * return -EIO if a journal_abort has been executed since the
1322  * transaction began.
1323  */
1324 int journal_stop(handle_t *handle)
1325 {
1326         transaction_t *transaction = handle->h_transaction;
1327         journal_t *journal = transaction->t_journal;
1328         int old_handle_count, err;
1329
1330         J_ASSERT(transaction->t_updates > 0);
1331         J_ASSERT(journal_current_handle() == handle);
1332
1333         if (is_handle_aborted(handle))
1334                 err = -EIO;
1335         else
1336                 err = 0;
1337
1338         if (--handle->h_ref > 0) {
1339                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1340                           handle->h_ref);
1341                 return err;
1342         }
1343
1344         jbd_debug(4, "Handle %p going down\n", handle);
1345
1346         /*
1347          * Implement synchronous transaction batching.  If the handle
1348          * was synchronous, don't force a commit immediately.  Let's
1349          * yield and let another thread piggyback onto this transaction.
1350          * Keep doing that while new threads continue to arrive.
1351          * It doesn't cost much - we're about to run a commit and sleep
1352          * on IO anyway.  Speeds up many-threaded, many-dir operations
1353          * by 30x or more...
1354          */
1355         if (handle->h_sync) {
1356                 do {
1357                         old_handle_count = transaction->t_handle_count;
1358                         set_current_state(TASK_UNINTERRUPTIBLE);
1359                         schedule_timeout(1);
1360                 } while (old_handle_count != transaction->t_handle_count);
1361         }
1362
1363         current->journal_info = NULL;
1364         spin_lock(&journal->j_state_lock);
1365         spin_lock(&transaction->t_handle_lock);
1366         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1367         transaction->t_updates--;
1368         if (!transaction->t_updates) {
1369                 wake_up(&journal->j_wait_updates);
1370                 if (journal->j_barrier_count)
1371                         wake_up(&journal->j_wait_transaction_locked);
1372         }
1373
1374         /* Move callbacks from the handle to the transaction. */
1375         spin_lock(&transaction->t_jcb_lock);
1376         list_splice(&handle->h_jcb, &transaction->t_jcb);
1377         spin_unlock(&transaction->t_jcb_lock);
1378
1379         /*
1380          * If the handle is marked SYNC, we need to set another commit
1381          * going!  We also want to force a commit if the current
1382          * transaction is occupying too much of the log, or if the
1383          * transaction is too old now.
1384          */
1385         if (handle->h_sync ||
1386                         transaction->t_outstanding_credits >
1387                                 journal->j_max_transaction_buffers ||
1388                         time_after_eq(jiffies, transaction->t_expires)) {
1389                 /* Do this even for aborted journals: an abort still
1390                  * completes the commit thread, it just doesn't write
1391                  * anything to disk. */
1392                 tid_t tid = transaction->t_tid;
1393
1394                 spin_unlock(&transaction->t_handle_lock);
1395                 jbd_debug(2, "transaction too old, requesting commit for "
1396                                         "handle %p\n", handle);
1397                 /* This is non-blocking */
1398                 __log_start_commit(journal, transaction->t_tid);
1399                 spin_unlock(&journal->j_state_lock);
1400
1401                 /*
1402                  * Special case: JFS_SYNC synchronous updates require us
1403                  * to wait for the commit to complete.
1404                  */
1405                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1406                         err = log_wait_commit(journal, tid);
1407         } else {
1408                 spin_unlock(&transaction->t_handle_lock);
1409                 spin_unlock(&journal->j_state_lock);
1410         }
1411
1412         jbd_free_handle(handle);
1413         return err;
1414 }
1415
1416 /**int journal_force_commit() - force any uncommitted transactions
1417  * @journal: journal to force
1418  *
1419  * For synchronous operations: force any uncommitted transactions
1420  * to disk.  May seem kludgy, but it reuses all the handle batching
1421  * code in a very simple manner.
1422  */
1423 int journal_force_commit(journal_t *journal)
1424 {
1425         handle_t *handle;
1426         int ret;
1427
1428         handle = journal_start(journal, 1);
1429         if (IS_ERR(handle)) {
1430                 ret = PTR_ERR(handle);
1431         } else {
1432                 handle->h_sync = 1;
1433                 ret = journal_stop(handle);
1434         }
1435         return ret;
1436 }
1437
1438 /*
1439  *
1440  * List management code snippets: various functions for manipulating the
1441  * transaction buffer lists.
1442  *
1443  */
1444
1445 /*
1446  * Append a buffer to a transaction list, given the transaction's list head
1447  * pointer.
1448  *
1449  * j_list_lock is held.
1450  *
1451  * jbd_lock_bh_state(jh2bh(jh)) is held.
1452  */
1453
1454 static inline void
1455 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1456 {
1457         if (!*list) {
1458                 jh->b_tnext = jh->b_tprev = jh;
1459                 *list = jh;
1460         } else {
1461                 /* Insert at the tail of the list to preserve order */
1462                 struct journal_head *first = *list, *last = first->b_tprev;
1463                 jh->b_tprev = last;
1464                 jh->b_tnext = first;
1465                 last->b_tnext = first->b_tprev = jh;
1466         }
1467 }
1468
1469 /*
1470  * Remove a buffer from a transaction list, given the transaction's list
1471  * head pointer.
1472  *
1473  * Called with j_list_lock held, and the journal may not be locked.
1474  *
1475  * jbd_lock_bh_state(jh2bh(jh)) is held.
1476  */
1477
1478 static inline void
1479 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1480 {
1481         if (*list == jh) {
1482                 *list = jh->b_tnext;
1483                 if (*list == jh)
1484                         *list = NULL;
1485         }
1486         jh->b_tprev->b_tnext = jh->b_tnext;
1487         jh->b_tnext->b_tprev = jh->b_tprev;
1488 }
1489
1490 /*
1491  * Remove a buffer from the appropriate transaction list.
1492  *
1493  * Note that this function can *change* the value of
1494  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1495  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1496  * is holding onto a copy of one of thee pointers, it could go bad.
1497  * Generally the caller needs to re-read the pointer from the transaction_t.
1498  *
1499  * Called under j_list_lock.  The journal may not be locked.
1500  */
1501 void __journal_unfile_buffer(struct journal_head *jh)
1502 {
1503         struct journal_head **list = NULL;
1504         transaction_t *transaction;
1505         struct buffer_head *bh = jh2bh(jh);
1506
1507         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1508         transaction = jh->b_transaction;
1509         if (transaction)
1510                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1511
1512         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1513         if (jh->b_jlist != BJ_None)
1514                 J_ASSERT_JH(jh, transaction != 0);
1515
1516         switch (jh->b_jlist) {
1517         case BJ_None:
1518                 goto out;
1519         case BJ_SyncData:
1520                 list = &transaction->t_sync_datalist;
1521                 break;
1522         case BJ_Metadata:
1523                 transaction->t_nr_buffers--;
1524                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1525                 list = &transaction->t_buffers;
1526                 break;
1527         case BJ_Forget:
1528                 list = &transaction->t_forget;
1529                 break;
1530         case BJ_IO:
1531                 list = &transaction->t_iobuf_list;
1532                 break;
1533         case BJ_Shadow:
1534                 list = &transaction->t_shadow_list;
1535                 break;
1536         case BJ_LogCtl:
1537                 list = &transaction->t_log_list;
1538                 break;
1539         case BJ_Reserved:
1540                 list = &transaction->t_reserved_list;
1541                 break;
1542         case BJ_Locked:
1543                 list = &transaction->t_locked_list;
1544                 break;
1545         }
1546
1547         __blist_del_buffer(list, jh);
1548         jh->b_jlist = BJ_None;
1549         if (test_clear_buffer_jbddirty(bh))
1550                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1551 out:
1552         jh->b_transaction = NULL;
1553 }
1554
1555 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1556 {
1557         jbd_lock_bh_state(jh2bh(jh));
1558         spin_lock(&journal->j_list_lock);
1559         __journal_unfile_buffer(jh);
1560         spin_unlock(&journal->j_list_lock);
1561         jbd_unlock_bh_state(jh2bh(jh));
1562 }
1563
1564 /*
1565  * Called from journal_try_to_free_buffers().
1566  *
1567  * Called under jbd_lock_bh_state(bh)
1568  */
1569 static void
1570 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1571 {
1572         struct journal_head *jh;
1573
1574         jh = bh2jh(bh);
1575
1576         if (buffer_locked(bh) || buffer_dirty(bh))
1577                 goto out;
1578
1579         if (jh->b_next_transaction != 0)
1580                 goto out;
1581
1582         spin_lock(&journal->j_list_lock);
1583         if (jh->b_transaction != 0 && jh->b_cp_transaction == 0) {
1584                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1585                         /* A written-back ordered data buffer */
1586                         JBUFFER_TRACE(jh, "release data");
1587                         __journal_unfile_buffer(jh);
1588                         journal_remove_journal_head(bh);
1589                         __brelse(bh);
1590                 }
1591         } else if (jh->b_cp_transaction != 0 && jh->b_transaction == 0) {
1592                 /* written-back checkpointed metadata buffer */
1593                 if (jh->b_jlist == BJ_None) {
1594                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1595                         __journal_remove_checkpoint(jh);
1596                         journal_remove_journal_head(bh);
1597                         __brelse(bh);
1598                 }
1599         }
1600         spin_unlock(&journal->j_list_lock);
1601 out:
1602         return;
1603 }
1604
1605
1606 /**
1607  * int journal_try_to_free_buffers() - try to free page buffers.
1608  * @journal: journal for operation
1609  * @page: to try and free
1610  * @gfp_mask: 'IO' mode for try_to_free_buffers()
1611  *
1612  *
1613  * For all the buffers on this page,
1614  * if they are fully written out ordered data, move them onto BUF_CLEAN
1615  * so try_to_free_buffers() can reap them.
1616  *
1617  * This function returns non-zero if we wish try_to_free_buffers()
1618  * to be called. We do this if the page is releasable by try_to_free_buffers().
1619  * We also do it if the page has locked or dirty buffers and the caller wants
1620  * us to perform sync or async writeout.
1621  *
1622  * This complicates JBD locking somewhat.  We aren't protected by the
1623  * BKL here.  We wish to remove the buffer from its committing or
1624  * running transaction's ->t_datalist via __journal_unfile_buffer.
1625  *
1626  * This may *change* the value of transaction_t->t_datalist, so anyone
1627  * who looks at t_datalist needs to lock against this function.
1628  *
1629  * Even worse, someone may be doing a journal_dirty_data on this
1630  * buffer.  So we need to lock against that.  journal_dirty_data()
1631  * will come out of the lock with the buffer dirty, which makes it
1632  * ineligible for release here.
1633  *
1634  * Who else is affected by this?  hmm...  Really the only contender
1635  * is do_get_write_access() - it could be looking at the buffer while
1636  * journal_try_to_free_buffer() is changing its state.  But that
1637  * cannot happen because we never reallocate freed data as metadata
1638  * while the data is part of a transaction.  Yes?
1639  */
1640 int journal_try_to_free_buffers(journal_t *journal,
1641                                 struct page *page, int unused_gfp_mask)
1642 {
1643         struct buffer_head *head;
1644         struct buffer_head *bh;
1645         int ret = 0;
1646
1647         J_ASSERT(PageLocked(page));
1648
1649         head = page_buffers(page);
1650         bh = head;
1651         do {
1652                 struct journal_head *jh;
1653
1654                 /*
1655                  * We take our own ref against the journal_head here to avoid
1656                  * having to add tons of locking around each instance of
1657                  * journal_remove_journal_head() and journal_put_journal_head().
1658                  */
1659                 jh = journal_grab_journal_head(bh);
1660                 if (!jh)
1661                         continue;
1662
1663                 jbd_lock_bh_state(bh);
1664                 __journal_try_to_free_buffer(journal, bh);
1665                 journal_put_journal_head(jh);
1666                 jbd_unlock_bh_state(bh);
1667                 if (buffer_jbd(bh))
1668                         goto busy;
1669         } while ((bh = bh->b_this_page) != head);
1670         ret = try_to_free_buffers(page);
1671 busy:
1672         return ret;
1673 }
1674
1675 /*
1676  * This buffer is no longer needed.  If it is on an older transaction's
1677  * checkpoint list we need to record it on this transaction's forget list
1678  * to pin this buffer (and hence its checkpointing transaction) down until
1679  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1680  * release it.
1681  * Returns non-zero if JBD no longer has an interest in the buffer.
1682  *
1683  * Called under j_list_lock.
1684  *
1685  * Called under jbd_lock_bh_state(bh).
1686  */
1687 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1688 {
1689         int may_free = 1;
1690         struct buffer_head *bh = jh2bh(jh);
1691
1692         __journal_unfile_buffer(jh);
1693
1694         if (jh->b_cp_transaction) {
1695                 JBUFFER_TRACE(jh, "on running+cp transaction");
1696                 __journal_file_buffer(jh, transaction, BJ_Forget);
1697                 clear_buffer_jbddirty(bh);
1698                 may_free = 0;
1699         } else {
1700                 JBUFFER_TRACE(jh, "on running transaction");
1701                 journal_remove_journal_head(bh);
1702                 __brelse(bh);
1703         }
1704         return may_free;
1705 }
1706
1707 /*
1708  * journal_invalidatepage
1709  *
1710  * This code is tricky.  It has a number of cases to deal with.
1711  *
1712  * There are two invariants which this code relies on:
1713  *
1714  * i_size must be updated on disk before we start calling invalidatepage on the
1715  * data.
1716  *
1717  *  This is done in ext3 by defining an ext3_setattr method which
1718  *  updates i_size before truncate gets going.  By maintaining this
1719  *  invariant, we can be sure that it is safe to throw away any buffers
1720  *  attached to the current transaction: once the transaction commits,
1721  *  we know that the data will not be needed.
1722  *
1723  *  Note however that we can *not* throw away data belonging to the
1724  *  previous, committing transaction!
1725  *
1726  * Any disk blocks which *are* part of the previous, committing
1727  * transaction (and which therefore cannot be discarded immediately) are
1728  * not going to be reused in the new running transaction
1729  *
1730  *  The bitmap committed_data images guarantee this: any block which is
1731  *  allocated in one transaction and removed in the next will be marked
1732  *  as in-use in the committed_data bitmap, so cannot be reused until
1733  *  the next transaction to delete the block commits.  This means that
1734  *  leaving committing buffers dirty is quite safe: the disk blocks
1735  *  cannot be reallocated to a different file and so buffer aliasing is
1736  *  not possible.
1737  *
1738  *
1739  * The above applies mainly to ordered data mode.  In writeback mode we
1740  * don't make guarantees about the order in which data hits disk --- in
1741  * particular we don't guarantee that new dirty data is flushed before
1742  * transaction commit --- so it is always safe just to discard data
1743  * immediately in that mode.  --sct
1744  */
1745
1746 /*
1747  * The journal_unmap_buffer helper function returns zero if the buffer
1748  * concerned remains pinned as an anonymous buffer belonging to an older
1749  * transaction.
1750  *
1751  * We're outside-transaction here.  Either or both of j_running_transaction
1752  * and j_committing_transaction may be NULL.
1753  */
1754 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1755 {
1756         transaction_t *transaction;
1757         struct journal_head *jh;
1758         int may_free = 1;
1759         int ret;
1760
1761         BUFFER_TRACE(bh, "entry");
1762
1763         /*
1764          * It is safe to proceed here without the j_list_lock because the
1765          * buffers cannot be stolen by try_to_free_buffers as long as we are
1766          * holding the page lock. --sct
1767          */
1768
1769         if (!buffer_jbd(bh))
1770                 goto zap_buffer_unlocked;
1771
1772         spin_lock(&journal->j_state_lock);
1773         jbd_lock_bh_state(bh);
1774         spin_lock(&journal->j_list_lock);
1775
1776         jh = journal_grab_journal_head(bh);
1777         if (!jh)
1778                 goto zap_buffer_no_jh;
1779
1780         transaction = jh->b_transaction;
1781         if (transaction == NULL) {
1782                 /* First case: not on any transaction.  If it
1783                  * has no checkpoint link, then we can zap it:
1784                  * it's a writeback-mode buffer so we don't care
1785                  * if it hits disk safely. */
1786                 if (!jh->b_cp_transaction) {
1787                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1788                         goto zap_buffer;
1789                 }
1790
1791                 if (!buffer_dirty(bh)) {
1792                         /* bdflush has written it.  We can drop it now */
1793                         goto zap_buffer;
1794                 }
1795
1796                 /* OK, it must be in the journal but still not
1797                  * written fully to disk: it's metadata or
1798                  * journaled data... */
1799
1800                 if (journal->j_running_transaction) {
1801                         /* ... and once the current transaction has
1802                          * committed, the buffer won't be needed any
1803                          * longer. */
1804                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1805                         ret = __dispose_buffer(jh,
1806                                         journal->j_running_transaction);
1807                         spin_unlock(&journal->j_list_lock);
1808                         jbd_unlock_bh_state(bh);
1809                         spin_unlock(&journal->j_state_lock);
1810                         journal_put_journal_head(jh);
1811                         return ret;
1812                 } else {
1813                         /* There is no currently-running transaction. So the
1814                          * orphan record which we wrote for this file must have
1815                          * passed into commit.  We must attach this buffer to
1816                          * the committing transaction, if it exists. */
1817                         if (journal->j_committing_transaction) {
1818                                 JBUFFER_TRACE(jh, "give to committing trans");
1819                                 ret = __dispose_buffer(jh,
1820                                         journal->j_committing_transaction);
1821                                 spin_unlock(&journal->j_list_lock);
1822                                 jbd_unlock_bh_state(bh);
1823                                 spin_unlock(&journal->j_state_lock);
1824                                 journal_put_journal_head(jh);
1825                                 return ret;
1826                         } else {
1827                                 /* The orphan record's transaction has
1828                                  * committed.  We can cleanse this buffer */
1829                                 clear_buffer_jbddirty(bh);
1830                                 goto zap_buffer;
1831                         }
1832                 }
1833         } else if (transaction == journal->j_committing_transaction) {
1834                 /* If it is committing, we simply cannot touch it.  We
1835                  * can remove it's next_transaction pointer from the
1836                  * running transaction if that is set, but nothing
1837                  * else. */
1838                 JBUFFER_TRACE(jh, "on committing transaction");
1839                 set_buffer_freed(bh);
1840                 if (jh->b_next_transaction) {
1841                         J_ASSERT(jh->b_next_transaction ==
1842                                         journal->j_running_transaction);
1843                         jh->b_next_transaction = NULL;
1844                 }
1845                 spin_unlock(&journal->j_list_lock);
1846                 jbd_unlock_bh_state(bh);
1847                 spin_unlock(&journal->j_state_lock);
1848                 journal_put_journal_head(jh);
1849                 return 0;
1850         } else {
1851                 /* Good, the buffer belongs to the running transaction.
1852                  * We are writing our own transaction's data, not any
1853                  * previous one's, so it is safe to throw it away
1854                  * (remember that we expect the filesystem to have set
1855                  * i_size already for this truncate so recovery will not
1856                  * expose the disk blocks we are discarding here.) */
1857                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1858                 may_free = __dispose_buffer(jh, transaction);
1859         }
1860
1861 zap_buffer:
1862         journal_put_journal_head(jh);
1863 zap_buffer_no_jh:
1864         spin_unlock(&journal->j_list_lock);
1865         jbd_unlock_bh_state(bh);
1866         spin_unlock(&journal->j_state_lock);
1867 zap_buffer_unlocked:
1868         clear_buffer_dirty(bh);
1869         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1870         clear_buffer_mapped(bh);
1871         clear_buffer_req(bh);
1872         clear_buffer_new(bh);
1873         bh->b_bdev = NULL;
1874         return may_free;
1875 }
1876
1877 /**
1878  * int journal_invalidatepage()
1879  * @journal: journal to use for flush...
1880  * @page:    page to flush
1881  * @offset:  length of page to invalidate.
1882  *
1883  * Reap page buffers containing data after offset in page.
1884  *
1885  * Return non-zero if the page's buffers were successfully reaped.
1886  */
1887 int journal_invalidatepage(journal_t *journal,
1888                       struct page *page,
1889                       unsigned long offset)
1890 {
1891         struct buffer_head *head, *bh, *next;
1892         unsigned int curr_off = 0;
1893         int may_free = 1;
1894
1895         if (!PageLocked(page))
1896                 BUG();
1897         if (!page_has_buffers(page))
1898                 return 1;
1899
1900         /* We will potentially be playing with lists other than just the
1901          * data lists (especially for journaled data mode), so be
1902          * cautious in our locking. */
1903
1904         head = bh = page_buffers(page);
1905         do {
1906                 unsigned int next_off = curr_off + bh->b_size;
1907                 next = bh->b_this_page;
1908
1909                 /* AKPM: doing lock_buffer here may be overly paranoid */
1910                 if (offset <= curr_off) {
1911                         /* This block is wholly outside the truncation point */
1912                         lock_buffer(bh);
1913                         may_free &= journal_unmap_buffer(journal, bh);
1914                         unlock_buffer(bh);
1915                 }
1916                 curr_off = next_off;
1917                 bh = next;
1918
1919         } while (bh != head);
1920
1921         if (!offset) {
1922                 if (!may_free || !try_to_free_buffers(page))
1923                         return 0;
1924                 J_ASSERT(!page_has_buffers(page));
1925         }
1926         return 1;
1927 }
1928
1929 /*
1930  * File a buffer on the given transaction list.
1931  */
1932 void __journal_file_buffer(struct journal_head *jh,
1933                         transaction_t *transaction, int jlist)
1934 {
1935         struct journal_head **list = NULL;
1936         int was_dirty = 0;
1937         struct buffer_head *bh = jh2bh(jh);
1938
1939         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1940         assert_spin_locked(&transaction->t_journal->j_list_lock);
1941
1942         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1943         J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1944                                 jh->b_transaction == 0);
1945
1946         if (jh->b_transaction && jh->b_jlist == jlist)
1947                 return;
1948
1949         /* The following list of buffer states needs to be consistent
1950          * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1951          * state. */
1952
1953         if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1954             jlist == BJ_Shadow || jlist == BJ_Forget) {
1955                 if (test_clear_buffer_dirty(bh) ||
1956                     test_clear_buffer_jbddirty(bh))
1957                         was_dirty = 1;
1958         }
1959
1960         if (jh->b_transaction)
1961                 __journal_unfile_buffer(jh);
1962         jh->b_transaction = transaction;
1963
1964         switch (jlist) {
1965         case BJ_None:
1966                 J_ASSERT_JH(jh, !jh->b_committed_data);
1967                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1968                 return;
1969         case BJ_SyncData:
1970                 list = &transaction->t_sync_datalist;
1971                 break;
1972         case BJ_Metadata:
1973                 transaction->t_nr_buffers++;
1974                 list = &transaction->t_buffers;
1975                 break;
1976         case BJ_Forget:
1977                 list = &transaction->t_forget;
1978                 break;
1979         case BJ_IO:
1980                 list = &transaction->t_iobuf_list;
1981                 break;
1982         case BJ_Shadow:
1983                 list = &transaction->t_shadow_list;
1984                 break;
1985         case BJ_LogCtl:
1986                 list = &transaction->t_log_list;
1987                 break;
1988         case BJ_Reserved:
1989                 list = &transaction->t_reserved_list;
1990                 break;
1991         case BJ_Locked:
1992                 list =  &transaction->t_locked_list;
1993                 break;
1994         }
1995
1996         __blist_add_buffer(list, jh);
1997         jh->b_jlist = jlist;
1998
1999         if (was_dirty)
2000                 set_buffer_jbddirty(bh);
2001 }
2002
2003 void journal_file_buffer(struct journal_head *jh,
2004                                 transaction_t *transaction, int jlist)
2005 {
2006         jbd_lock_bh_state(jh2bh(jh));
2007         spin_lock(&transaction->t_journal->j_list_lock);
2008         __journal_file_buffer(jh, transaction, jlist);
2009         spin_unlock(&transaction->t_journal->j_list_lock);
2010         jbd_unlock_bh_state(jh2bh(jh));
2011 }
2012
2013 /*
2014  * Remove a buffer from its current buffer list in preparation for
2015  * dropping it from its current transaction entirely.  If the buffer has
2016  * already started to be used by a subsequent transaction, refile the
2017  * buffer on that transaction's metadata list.
2018  *
2019  * Called under journal->j_list_lock
2020  *
2021  * Called under jbd_lock_bh_state(jh2bh(jh))
2022  */
2023 void __journal_refile_buffer(struct journal_head *jh)
2024 {
2025         int was_dirty;
2026         struct buffer_head *bh = jh2bh(jh);
2027
2028         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2029         if (jh->b_transaction)
2030                 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2031
2032         /* If the buffer is now unused, just drop it. */
2033         if (jh->b_next_transaction == NULL) {
2034                 __journal_unfile_buffer(jh);
2035                 return;
2036         }
2037
2038         /*
2039          * It has been modified by a later transaction: add it to the new
2040          * transaction's metadata list.
2041          */
2042
2043         was_dirty = test_clear_buffer_jbddirty(bh);
2044         __journal_unfile_buffer(jh);
2045         jh->b_transaction = jh->b_next_transaction;
2046         jh->b_next_transaction = NULL;
2047         __journal_file_buffer(jh, jh->b_transaction, BJ_Metadata);
2048         J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2049
2050         if (was_dirty)
2051                 set_buffer_jbddirty(bh);
2052 }
2053
2054 /*
2055  * For the unlocked version of this call, also make sure that any
2056  * hanging journal_head is cleaned up if necessary.
2057  *
2058  * __journal_refile_buffer is usually called as part of a single locked
2059  * operation on a buffer_head, in which the caller is probably going to
2060  * be hooking the journal_head onto other lists.  In that case it is up
2061  * to the caller to remove the journal_head if necessary.  For the
2062  * unlocked journal_refile_buffer call, the caller isn't going to be
2063  * doing anything else to the buffer so we need to do the cleanup
2064  * ourselves to avoid a jh leak.
2065  *
2066  * *** The journal_head may be freed by this call! ***
2067  */
2068 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2069 {
2070         struct buffer_head *bh = jh2bh(jh);
2071
2072         jbd_lock_bh_state(bh);
2073         spin_lock(&journal->j_list_lock);
2074
2075         __journal_refile_buffer(jh);
2076         jbd_unlock_bh_state(bh);
2077         journal_remove_journal_head(bh);
2078
2079         spin_unlock(&journal->j_list_lock);
2080         __brelse(bh);
2081 }