fs/jbd/transaction.c

   1 /*
   2  * linux/fs/transaction.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
   5  *
   6  * Copyright 1998 Red Hat corp --- All Rights Reserved
   7  *
   8  * This file is part of the Linux kernel and is made available under
   9  * the terms of the GNU General Public License, version 2, or at your
  10  * option, any later version, incorporated herein by reference.
  11  *
  12  * Generic filesystem transaction handling code; part of the ext2fs
  13  * journaling system.
  14  *
  15  * This file manages transactions (compound commits managed by the
  16  * journaling code) and handles (individual atomic operations by the
  17  * filesystem).
  18  */
  19
  20 #include <linux/time.h>
  21 #include <linux/fs.h>
  22 #include <linux/jbd.h>
  23 #include <linux/errno.h>
  24 #include <linux/slab.h>
  25 #include <linux/timer.h>
  26 #include <linux/smp_lock.h>
  27 #include <linux/mm.h>
  28 #include <linux/highmem.h>
  29
  30 /*
  31  * get_transaction: obtain a new transaction_t object.
  32  *
  33  * Simply allocate and initialise a new transaction.  Create it in
  34  * RUNNING state and add it to the current journal (which should not
  35  * have an existing running transaction: we only make a new transaction
  36  * once we have started to commit the old one).
  37  *
  38  * Preconditions:
  39  *      The journal MUST be locked.  We don't perform atomic mallocs on the
  40  *      new transaction and we can't block without protecting against other
  41  *      processes trying to touch the journal while it is in transition.
  42  *
  43  * Called under j_state_lock
  44  */
  45
  46 static transaction_t *
  47 get_transaction(journal_t *journal, transaction_t *transaction)
  48 {
  49         transaction->t_journal = journal;
  50         transaction->t_state = T_RUNNING;
  51         transaction->t_tid = journal->j_transaction_sequence++;
  52         transaction->t_expires = jiffies + journal->j_commit_interval;
  53         spin_lock_init(&transaction->t_handle_lock);
  54
  55         /* Set up the commit timer for the new transaction. */
  56         journal->j_commit_timer->expires = transaction->t_expires;
  57         add_timer(journal->j_commit_timer);
  58
  59         J_ASSERT(journal->j_running_transaction == NULL);
  60         journal->j_running_transaction = transaction;
  61
  62         return transaction;
  63 }
  64
  65 /*
  66  * Handle management.
  67  *
  68  * A handle_t is an object which represents a single atomic update to a
  69  * filesystem, and which tracks all of the modifications which form part
  70  * of that one update.
  71  */
  72
  73 /*
  74  * start_this_handle: Given a handle, deal with any locking or stalling
  75  * needed to make sure that there is enough journal space for the handle
  76  * to begin.  Attach the handle to a transaction and set up the
  77  * transaction's buffer credits.
  78  */
  79
  80 static int start_this_handle(journal_t *journal, handle_t *handle)
  81 {
  82         transaction_t *transaction;
  83         int needed;
  84         int nblocks = handle->h_buffer_credits;
  85         transaction_t *new_transaction = NULL;
  86         int ret = 0;
  87
  88         if (nblocks > journal->j_max_transaction_buffers) {
  89                 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
  90                        current->comm, nblocks,
  91                        journal->j_max_transaction_buffers);
  92                 ret = -ENOSPC;
  93                 goto out;
  94         }
  95
  96 alloc_transaction:
  97         if (!journal->j_running_transaction) {
  98                 new_transaction = jbd_kmalloc(sizeof(*new_transaction),
  99                                                 GFP_NOFS);
 100                 if (!new_transaction) {
 101                         ret = -ENOMEM;
 102                         goto out;
 103                 }
 104                 memset(new_transaction, 0, sizeof(*new_transaction));
 105         }
 106
 107         jbd_debug(3, "New handle %p going live.\n", handle);
 108
 109 repeat:
 110
 111         /*
 112          * We need to hold j_state_lock until t_updates has been incremented,
 113          * for proper journal barrier handling
 114          */
 115         spin_lock(&journal->j_state_lock);
 116 repeat_locked:
 117         if (is_journal_aborted(journal) ||
 118             (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
 119                 spin_unlock(&journal->j_state_lock);
 120                 ret = -EROFS;
 121                 goto out;
 122         }
 123
 124         /* Wait on the journal's transaction barrier if necessary */
 125         if (journal->j_barrier_count) {
 126                 spin_unlock(&journal->j_state_lock);
 127                 wait_event(journal->j_wait_transaction_locked,
 128                                 journal->j_barrier_count == 0);
 129                 goto repeat;
 130         }
 131
 132         if (!journal->j_running_transaction) {
 133                 if (!new_transaction) {
 134                         spin_unlock(&journal->j_state_lock);
 135                         goto alloc_transaction;
 136                 }
 137                 get_transaction(journal, new_transaction);
 138                 new_transaction = NULL;
 139         }
 140
 141         transaction = journal->j_running_transaction;
 142
 143         /*
 144          * If the current transaction is locked down for commit, wait for the
 145          * lock to be released.
 146          */
 147         if (transaction->t_state == T_LOCKED) {
 148                 DEFINE_WAIT(wait);
 149
 150                 prepare_to_wait(&journal->j_wait_transaction_locked,
 151                                         &wait, TASK_UNINTERRUPTIBLE);
 152                 spin_unlock(&journal->j_state_lock);
 153                 schedule();
 154                 finish_wait(&journal->j_wait_transaction_locked, &wait);
 155                 goto repeat;
 156         }
 157
 158         /*
 159          * If there is not enough space left in the log to write all potential
 160          * buffers requested by this operation, we need to stall pending a log
 161          * checkpoint to free some more log space.
 162          */
 163         spin_lock(&transaction->t_handle_lock);
 164         needed = transaction->t_outstanding_credits + nblocks;
 165
 166         if (needed > journal->j_max_transaction_buffers) {
 167                 /*
 168                  * If the current transaction is already too large, then start
 169                  * to commit it: we can then go back and attach this handle to
 170                  * a new transaction.
 171                  */
 172                 DEFINE_WAIT(wait);
 173
 174                 jbd_debug(2, "Handle %p starting new commit...\n", handle);
 175                 spin_unlock(&transaction->t_handle_lock);
 176                 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
 177                                 TASK_UNINTERRUPTIBLE);
 178                 __log_start_commit(journal, transaction->t_tid);
 179                 spin_unlock(&journal->j_state_lock);
 180                 schedule();
 181                 finish_wait(&journal->j_wait_transaction_locked, &wait);
 182                 goto repeat;
 183         }
 184
 185         /*
 186          * The commit code assumes that it can get enough log space
 187          * without forcing a checkpoint.  This is *critical* for
 188          * correctness: a checkpoint of a buffer which is also
 189          * associated with a committing transaction creates a deadlock,
 190          * so commit simply cannot force through checkpoints.
 191          *
 192          * We must therefore ensure the necessary space in the journal
 193          * *before* starting to dirty potentially checkpointed buffers
 194          * in the new transaction.
 195          *
 196          * The worst part is, any transaction currently committing can
 197          * reduce the free space arbitrarily.  Be careful to account for
 198          * those buffers when checkpointing.
 199          */
 200
 201         /*
 202          * @@@ AKPM: This seems rather over-defensive.  We're giving commit
 203          * a _lot_ of headroom: 1/4 of the journal plus the size of
 204          * the committing transaction.  Really, we only need to give it
 205          * committing_transaction->t_outstanding_credits plus "enough" for
 206          * the log control blocks.
 207          * Also, this test is inconsitent with the matching one in
 208          * journal_extend().
 209          */
 210         if (__log_space_left(journal) < jbd_space_needed(journal)) {
 211                 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
 212                 spin_unlock(&transaction->t_handle_lock);
 213                 __log_wait_for_space(journal);
 214                 goto repeat_locked;
 215         }
 216
 217         /* OK, account for the buffers that this operation expects to
 218          * use and add the handle to the running transaction. */
 219
 220         handle->h_transaction = transaction;
 221         transaction->t_outstanding_credits += nblocks;
 222         transaction->t_updates++;
 223         transaction->t_handle_count++;
 224         jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
 225                   handle, nblocks, transaction->t_outstanding_credits,
 226                   __log_space_left(journal));
 227         spin_unlock(&transaction->t_handle_lock);
 228         spin_unlock(&journal->j_state_lock);
 229 out:
 230         if (new_transaction)
 231                 kfree(new_transaction);
 232         return ret;
 233 }
 234
 235 /* Allocate a new handle.  This should probably be in a slab... */
 236 static handle_t *new_handle(int nblocks)
 237 {
 238         handle_t *handle = jbd_alloc_handle(GFP_NOFS);
 239         if (!handle)
 240                 return NULL;
 241         memset(handle, 0, sizeof(*handle));
 242         handle->h_buffer_credits = nblocks;
 243         handle->h_ref = 1;
 244
 245         return handle;
 246 }
 247
 248 /**
 249  * handle_t *journal_start() - Obtain a new handle.
 250  * @journal: Journal to start transaction on.
 251  * @nblocks: number of block buffer we might modify
 252  *
 253  * We make sure that the transaction can guarantee at least nblocks of
 254  * modified buffers in the log.  We block until the log can guarantee
 255  * that much space.
 256  *
 257  * This function is visible to journal users (like ext3fs), so is not
 258  * called with the journal already locked.
 259  *
 260  * Return a pointer to a newly allocated handle, or NULL on failure
 261  */
 262 handle_t *journal_start(journal_t *journal, int nblocks)
 263 {
 264         handle_t *handle = journal_current_handle();
 265         int err;
 266
 267         if (!journal)
 268                 return ERR_PTR(-EROFS);
 269
 270         if (handle) {
 271                 J_ASSERT(handle->h_transaction->t_journal == journal);
 272                 handle->h_ref++;
 273                 return handle;
 274         }
 275
 276         handle = new_handle(nblocks);
 277         if (!handle)
 278                 return ERR_PTR(-ENOMEM);
 279
 280         current->journal_info = handle;
 281
 282         err = start_this_handle(journal, handle);
 283         if (err < 0) {
 284                 jbd_free_handle(handle);
 285                 current->journal_info = NULL;
 286                 handle = ERR_PTR(err);
 287         }
 288         return handle;
 289 }
 290
 291 /**
 292  * int journal_extend() - extend buffer credits.
 293  * @handle:  handle to 'extend'
 294  * @nblocks: nr blocks to try to extend by.
 295  *
 296  * Some transactions, such as large extends and truncates, can be done
 297  * atomically all at once or in several stages.  The operation requests
 298  * a credit for a number of buffer modications in advance, but can
 299  * extend its credit if it needs more.
 300  *
 301  * journal_extend tries to give the running handle more buffer credits.
 302  * It does not guarantee that allocation - this is a best-effort only.
 303  * The calling process MUST be able to deal cleanly with a failure to
 304  * extend here.
 305  *
 306  * Return 0 on success, non-zero on failure.
 307  *
 308  * return code < 0 implies an error
 309  * return code > 0 implies normal transaction-full status.
 310  */
 311 int journal_extend(handle_t *handle, int nblocks)
 312 {
 313         transaction_t *transaction = handle->h_transaction;
 314         journal_t *journal = transaction->t_journal;
 315         int result;
 316         int wanted;
 317
 318         result = -EIO;
 319         if (is_handle_aborted(handle))
 320                 goto out;
 321
 322         result = 1;
 323
 324         spin_lock(&journal->j_state_lock);
 325
 326         /* Don't extend a locked-down transaction! */
 327         if (handle->h_transaction->t_state != T_RUNNING) {
 328                 jbd_debug(3, "denied handle %p %d blocks: "
 329                           "transaction not running\n", handle, nblocks);
 330                 goto error_out;
 331         }
 332
 333         spin_lock(&transaction->t_handle_lock);
 334         wanted = transaction->t_outstanding_credits + nblocks;
 335
 336         if (wanted > journal->j_max_transaction_buffers) {
 337                 jbd_debug(3, "denied handle %p %d blocks: "
 338                           "transaction too large\n", handle, nblocks);
 339                 goto unlock;
 340         }
 341
 342         if (wanted > __log_space_left(journal)) {
 343                 jbd_debug(3, "denied handle %p %d blocks: "
 344                           "insufficient log space\n", handle, nblocks);
 345                 goto unlock;
 346         }
 347
 348         handle->h_buffer_credits += nblocks;
 349         transaction->t_outstanding_credits += nblocks;
 350         result = 0;
 351
 352         jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
 353 unlock:
 354         spin_unlock(&transaction->t_handle_lock);
 355 error_out:
 356         spin_unlock(&journal->j_state_lock);
 357 out:
 358         return result;
 359 }
 360
 361
 362 /**
 363  * int journal_restart() - restart a handle .
 364  * @handle:  handle to restart
 365  * @nblocks: nr credits requested
 366  *
 367  * Restart a handle for a multi-transaction filesystem
 368  * operation.
 369  *
 370  * If the journal_extend() call above fails to grant new buffer credits
 371  * to a running handle, a call to journal_restart will commit the
 372  * handle's transaction so far and reattach the handle to a new
 373  * transaction capabable of guaranteeing the requested number of
 374  * credits.
 375  */
 376
 377 int journal_restart(handle_t *handle, int nblocks)
 378 {
 379         transaction_t *transaction = handle->h_transaction;
 380         journal_t *journal = transaction->t_journal;
 381         int ret;
 382
 383         /* If we've had an abort of any type, don't even think about
 384          * actually doing the restart! */
 385         if (is_handle_aborted(handle))
 386                 return 0;
 387
 388         /*
 389          * First unlink the handle from its current transaction, and start the
 390          * commit on that.
 391          */
 392         J_ASSERT(transaction->t_updates > 0);
 393         J_ASSERT(journal_current_handle() == handle);
 394
 395         spin_lock(&journal->j_state_lock);
 396         spin_lock(&transaction->t_handle_lock);
 397         transaction->t_outstanding_credits -= handle->h_buffer_credits;
 398         transaction->t_updates--;
 399
 400         if (!transaction->t_updates)
 401                 wake_up(&journal->j_wait_updates);
 402         spin_unlock(&transaction->t_handle_lock);
 403
 404         jbd_debug(2, "restarting handle %p\n", handle);
 405         __log_start_commit(journal, transaction->t_tid);
 406         spin_unlock(&journal->j_state_lock);
 407
 408         handle->h_buffer_credits = nblocks;
 409         ret = start_this_handle(journal, handle);
 410         return ret;
 411 }
 412
 413
 414 /**
 415  * void journal_lock_updates () - establish a transaction barrier.
 416  * @journal:  Journal to establish a barrier on.
 417  *
 418  * This locks out any further updates from being started, and blocks
 419  * until all existing updates have completed, returning only once the
 420  * journal is in a quiescent state with no updates running.
 421  *
 422  * The journal lock should not be held on entry.
 423  */
 424 void journal_lock_updates(journal_t *journal)
 425 {
 426         DEFINE_WAIT(wait);
 427
 428         spin_lock(&journal->j_state_lock);
 429         ++journal->j_barrier_count;
 430
 431         /* Wait until there are no running updates */
 432         while (1) {
 433                 transaction_t *transaction = journal->j_running_transaction;
 434
 435                 if (!transaction)
 436                         break;
 437
 438                 spin_lock(&transaction->t_handle_lock);
 439                 if (!transaction->t_updates) {
 440                         spin_unlock(&transaction->t_handle_lock);
 441                         break;
 442                 }
 443                 prepare_to_wait(&journal->j_wait_updates, &wait,
 444                                 TASK_UNINTERRUPTIBLE);
 445                 spin_unlock(&transaction->t_handle_lock);
 446                 spin_unlock(&journal->j_state_lock);
 447                 schedule();
 448                 finish_wait(&journal->j_wait_updates, &wait);
 449                 spin_lock(&journal->j_state_lock);
 450         }
 451         spin_unlock(&journal->j_state_lock);
 452
 453         /*
 454          * We have now established a barrier against other normal updates, but
 455          * we also need to barrier against other journal_lock_updates() calls
 456          * to make sure that we serialise special journal-locked operations
 457          * too.
 458          */
 459         down(&journal->j_barrier);
 460 }
 461
 462 /**
 463  * void journal_unlock_updates (journal_t* journal) - release barrier
 464  * @journal:  Journal to release the barrier on.
 465  *
 466  * Release a transaction barrier obtained with journal_lock_updates().
 467  *
 468  * Should be called without the journal lock held.
 469  */
 470 void journal_unlock_updates (journal_t *journal)
 471 {
 472         J_ASSERT(journal->j_barrier_count != 0);
 473
 474         up(&journal->j_barrier);
 475         spin_lock(&journal->j_state_lock);
 476         --journal->j_barrier_count;
 477         spin_unlock(&journal->j_state_lock);
 478         wake_up(&journal->j_wait_transaction_locked);
 479 }
 480
 481 /*
 482  * Report any unexpected dirty buffers which turn up.  Normally those
 483  * indicate an error, but they can occur if the user is running (say)
 484  * tune2fs to modify the live filesystem, so we need the option of
 485  * continuing as gracefully as possible.  #
 486  *
 487  * The caller should already hold the journal lock and
 488  * j_list_lock spinlock: most callers will need those anyway
 489  * in order to probe the buffer's journaling state safely.
 490  */
 491 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
 492 {
 493         struct buffer_head *bh = jh2bh(jh);
 494         int jlist;
 495
 496         if (buffer_dirty(bh)) {
 497                 /* If this buffer is one which might reasonably be dirty
 498                  * --- ie. data, or not part of this journal --- then
 499                  * we're OK to leave it alone, but otherwise we need to
 500                  * move the dirty bit to the journal's own internal
 501                  * JBDDirty bit. */
 502                 jlist = jh->b_jlist;
 503
 504                 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
 505                     jlist == BJ_Shadow || jlist == BJ_Forget) {
 506                         if (test_clear_buffer_dirty(jh2bh(jh))) {
 507                                 set_bit(BH_JBDDirty, &jh2bh(jh)->b_state);
 508                         }
 509                 }
 510         }
 511 }
 512
 513 /*
 514  * If the buffer is already part of the current transaction, then there
 515  * is nothing we need to do.  If it is already part of a prior
 516  * transaction which we are still committing to disk, then we need to
 517  * make sure that we do not overwrite the old copy: we do copy-out to
 518  * preserve the copy going to disk.  We also account the buffer against
 519  * the handle's metadata buffer credits (unless the buffer is already
 520  * part of the transaction, that is).
 521  *
 522  */
 523 static int
 524 do_get_write_access(handle_t *handle, struct journal_head *jh,
 525                         int force_copy)
 526 {
 527         struct buffer_head *bh;
 528         transaction_t *transaction;
 529         journal_t *journal;
 530         int error;
 531         char *frozen_buffer = NULL;
 532         int need_copy = 0;
 533
 534         if (is_handle_aborted(handle))
 535                 return -EROFS;
 536
 537         transaction = handle->h_transaction;
 538         journal = transaction->t_journal;
 539
 540         jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
 541
 542         JBUFFER_TRACE(jh, "entry");
 543 repeat:
 544         bh = jh2bh(jh);
 545
 546         /* @@@ Need to check for errors here at some point. */
 547
 548         lock_buffer(bh);
 549         jbd_lock_bh_state(bh);
 550
 551         /* We now hold the buffer lock so it is safe to query the buffer
 552          * state.  Is the buffer dirty?
 553          *
 554          * If so, there are two possibilities.  The buffer may be
 555          * non-journaled, and undergoing a quite legitimate writeback.
 556          * Otherwise, it is journaled, and we don't expect dirty buffers
 557          * in that state (the buffers should be marked JBD_Dirty
 558          * instead.)  So either the IO is being done under our own
 559          * control and this is a bug, or it's a third party IO such as
 560          * dump(8) (which may leave the buffer scheduled for read ---
 561          * ie. locked but not dirty) or tune2fs (which may actually have
 562          * the buffer dirtied, ugh.)  */
 563
 564         if (buffer_dirty(bh)) {
 565                 /*
 566                  * First question: is this buffer already part of the current
 567                  * transaction or the existing committing transaction?
 568                  */
 569                 if (jh->b_transaction) {
 570                         J_ASSERT_JH(jh,
 571                                 jh->b_transaction == transaction ||
 572                                 jh->b_transaction ==
 573                                         journal->j_committing_transaction);
 574                         if (jh->b_next_transaction)
 575                                 J_ASSERT_JH(jh, jh->b_next_transaction ==
 576                                                         transaction);
 577                         JBUFFER_TRACE(jh, "Unexpected dirty buffer");
 578                         jbd_unexpected_dirty_buffer(jh);
 579                 }
 580         }
 581
 582         unlock_buffer(bh);
 583
 584         error = -EROFS;
 585         if (is_handle_aborted(handle)) {
 586                 jbd_unlock_bh_state(bh);
 587                 goto out;
 588         }
 589         error = 0;
 590
 591         /*
 592          * The buffer is already part of this transaction if b_transaction or
 593          * b_next_transaction points to it
 594          */
 595         if (jh->b_transaction == transaction ||
 596             jh->b_next_transaction == transaction)
 597                 goto done;
 598
 599         /*
 600          * If there is already a copy-out version of this buffer, then we don't
 601          * need to make another one
 602          */
 603         if (jh->b_frozen_data) {
 604                 JBUFFER_TRACE(jh, "has frozen data");
 605                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 606                 jh->b_next_transaction = transaction;
 607                 goto done;
 608         }
 609
 610         /* Is there data here we need to preserve? */
 611
 612         if (jh->b_transaction && jh->b_transaction != transaction) {
 613                 JBUFFER_TRACE(jh, "owned by older transaction");
 614                 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 615                 J_ASSERT_JH(jh, jh->b_transaction ==
 616                                         journal->j_committing_transaction);
 617
 618                 /* There is one case we have to be very careful about.
 619                  * If the committing transaction is currently writing
 620                  * this buffer out to disk and has NOT made a copy-out,
 621                  * then we cannot modify the buffer contents at all
 622                  * right now.  The essence of copy-out is that it is the
 623                  * extra copy, not the primary copy, which gets
 624                  * journaled.  If the primary copy is already going to
 625                  * disk then we cannot do copy-out here. */
 626
 627                 if (jh->b_jlist == BJ_Shadow) {
 628                         DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
 629                         wait_queue_head_t *wqh;
 630
 631                         wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
 632
 633                         JBUFFER_TRACE(jh, "on shadow: sleep");
 634                         jbd_unlock_bh_state(bh);
 635                         /* commit wakes up all shadow buffers after IO */
 636                         for ( ; ; ) {
 637                                 prepare_to_wait(wqh, &wait.wait,
 638                                                 TASK_UNINTERRUPTIBLE);
 639                                 if (jh->b_jlist != BJ_Shadow)
 640                                         break;
 641                                 schedule();
 642                         }
 643                         finish_wait(wqh, &wait.wait);
 644                         goto repeat;
 645                 }
 646
 647                 /* Only do the copy if the currently-owning transaction
 648                  * still needs it.  If it is on the Forget list, the
 649                  * committing transaction is past that stage.  The
 650                  * buffer had better remain locked during the kmalloc,
 651                  * but that should be true --- we hold the journal lock
 652                  * still and the buffer is already on the BUF_JOURNAL
 653                  * list so won't be flushed.
 654                  *
 655                  * Subtle point, though: if this is a get_undo_access,
 656                  * then we will be relying on the frozen_data to contain
 657                  * the new value of the committed_data record after the
 658                  * transaction, so we HAVE to force the frozen_data copy
 659                  * in that case. */
 660
 661                 if (jh->b_jlist != BJ_Forget || force_copy) {
 662                         JBUFFER_TRACE(jh, "generate frozen data");
 663                         if (!frozen_buffer) {
 664                                 JBUFFER_TRACE(jh, "allocate memory for buffer");
 665                                 jbd_unlock_bh_state(bh);
 666                                 frozen_buffer = jbd_kmalloc(jh2bh(jh)->b_size,
 667                                                             GFP_NOFS);
 668                                 if (!frozen_buffer) {
 669                                         printk(KERN_EMERG
 670                                                "%s: OOM for frozen_buffer\n",
 671                                                __FUNCTION__);
 672                                         JBUFFER_TRACE(jh, "oom!");
 673                                         error = -ENOMEM;
 674                                         jbd_lock_bh_state(bh);
 675                                         goto done;
 676                                 }
 677                                 goto repeat;
 678                         }
 679                         jh->b_frozen_data = frozen_buffer;
 680                         frozen_buffer = NULL;
 681                         need_copy = 1;
 682                 }
 683                 jh->b_next_transaction = transaction;
 684         }
 685
 686
 687         /*
 688          * Finally, if the buffer is not journaled right now, we need to make
 689          * sure it doesn't get written to disk before the caller actually
 690          * commits the new data
 691          */
 692         if (!jh->b_transaction) {
 693                 JBUFFER_TRACE(jh, "no transaction");
 694                 J_ASSERT_JH(jh, !jh->b_next_transaction);
 695                 jh->b_transaction = transaction;
 696                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
 697                 spin_lock(&journal->j_list_lock);
 698                 __journal_file_buffer(jh, transaction, BJ_Reserved);
 699                 spin_unlock(&journal->j_list_lock);
 700         }
 701
 702 done:
 703         if (need_copy) {
 704                 struct page *page;
 705                 int offset;
 706                 char *source;
 707
 708                 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
 709                             "Possible IO failure.\n");
 710                 page = jh2bh(jh)->b_page;
 711                 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
 712                 source = kmap_atomic(page, KM_USER0);
 713                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
 714                 kunmap_atomic(source, KM_USER0);
 715         }
 716         jbd_unlock_bh_state(bh);
 717
 718         /*
 719          * If we are about to journal a buffer, then any revoke pending on it is
 720          * no longer valid
 721          */
 722         journal_cancel_revoke(handle, jh);
 723
 724 out:
 725         if (frozen_buffer)
 726                 kfree(frozen_buffer);
 727
 728         JBUFFER_TRACE(jh, "exit");
 729         return error;
 730 }
 731
 732 /**
 733  * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
 734  * @handle: transaction to add buffer modifications to
 735  * @bh:     bh to be used for metadata writes
 736  * @credits: variable that will receive credits for the buffer
 737  *
 738  * Returns an error code or 0 on success.
 739  *
 740  * In full data journalling mode the buffer may be of type BJ_AsyncData,
 741  * because we're write()ing a buffer which is also part of a shared mapping.
 742  */
 743
 744 int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
 745 {
 746         struct journal_head *jh = journal_add_journal_head(bh);
 747         int rc;
 748
 749         /* We do not want to get caught playing with fields which the
 750          * log thread also manipulates.  Make sure that the buffer
 751          * completes any outstanding IO before proceeding. */
 752         rc = do_get_write_access(handle, jh, 0);
 753         journal_put_journal_head(jh);
 754         return rc;
 755 }
 756
 757
 758 /*
 759  * When the user wants to journal a newly created buffer_head
 760  * (ie. getblk() returned a new buffer and we are going to populate it
 761  * manually rather than reading off disk), then we need to keep the
 762  * buffer_head locked until it has been completely filled with new
 763  * data.  In this case, we should be able to make the assertion that
 764  * the bh is not already part of an existing transaction.
 765  *
 766  * The buffer should already be locked by the caller by this point.
 767  * There is no lock ranking violation: it was a newly created,
 768  * unlocked buffer beforehand. */
 769
 770 /**
 771  * int journal_get_create_access () - notify intent to use newly created bh
 772  * @handle: transaction to new buffer to
 773  * @bh: new buffer.
 774  *
 775  * Call this if you create a new bh.
 776  */
 777 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
 778 {
 779         transaction_t *transaction = handle->h_transaction;
 780         journal_t *journal = transaction->t_journal;
 781         struct journal_head *jh = journal_add_journal_head(bh);
 782         int err;
 783
 784         jbd_debug(5, "journal_head %p\n", jh);
 785         err = -EROFS;
 786         if (is_handle_aborted(handle))
 787                 goto out;
 788         err = 0;
 789
 790         JBUFFER_TRACE(jh, "entry");
 791         /*
 792          * The buffer may already belong to this transaction due to pre-zeroing
 793          * in the filesystem's new_block code.  It may also be on the previous,
 794          * committing transaction's lists, but it HAS to be in Forget state in
 795          * that case: the transaction must have deleted the buffer for it to be
 796          * reused here.
 797          */
 798         jbd_lock_bh_state(bh);
 799         spin_lock(&journal->j_list_lock);
 800         J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
 801                 jh->b_transaction == NULL ||
 802                 (jh->b_transaction == journal->j_committing_transaction &&
 803                           jh->b_jlist == BJ_Forget)));
 804
 805         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 806         J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
 807
 808         if (jh->b_transaction == NULL) {
 809                 jh->b_transaction = transaction;
 810                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
 811                 __journal_file_buffer(jh, transaction, BJ_Reserved);
 812         } else if (jh->b_transaction == journal->j_committing_transaction) {
 813                 JBUFFER_TRACE(jh, "set next transaction");
 814                 jh->b_next_transaction = transaction;
 815         }
 816         spin_unlock(&journal->j_list_lock);
 817         jbd_unlock_bh_state(bh);
 818
 819         /*
 820          * akpm: I added this.  ext3_alloc_branch can pick up new indirect
 821          * blocks which contain freed but then revoked metadata.  We need
 822          * to cancel the revoke in case we end up freeing it yet again
 823          * and the reallocating as data - this would cause a second revoke,
 824          * which hits an assertion error.
 825          */
 826         JBUFFER_TRACE(jh, "cancelling revoke");
 827         journal_cancel_revoke(handle, jh);
 828         journal_put_journal_head(jh);
 829 out:
 830         return err;
 831 }
 832
 833 /**
 834  * int journal_get_undo_access() -  Notify intent to modify metadata with
 835  *     non-rewindable consequences
 836  * @handle: transaction
 837  * @bh: buffer to undo
 838  * @credits: store the number of taken credits here (if not NULL)
 839  *
 840  * Sometimes there is a need to distinguish between metadata which has
 841  * been committed to disk and that which has not.  The ext3fs code uses
 842  * this for freeing and allocating space, we have to make sure that we
 843  * do not reuse freed space until the deallocation has been committed,
 844  * since if we overwrote that space we would make the delete
 845  * un-rewindable in case of a crash.
 846  *
 847  * To deal with that, journal_get_undo_access requests write access to a
 848  * buffer for parts of non-rewindable operations such as delete
 849  * operations on the bitmaps.  The journaling code must keep a copy of
 850  * the buffer's contents prior to the undo_access call until such time
 851  * as we know that the buffer has definitely been committed to disk.
 852  *
 853  * We never need to know which transaction the committed data is part
 854  * of, buffers touched here are guaranteed to be dirtied later and so
 855  * will be committed to a new transaction in due course, at which point
 856  * we can discard the old committed data pointer.
 857  *
 858  * Returns error number or 0 on success.
 859  */
 860 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
 861 {
 862         int err;
 863         struct journal_head *jh = journal_add_journal_head(bh);
 864         char *committed_data = NULL;
 865
 866         JBUFFER_TRACE(jh, "entry");
 867
 868         /*
 869          * Do this first --- it can drop the journal lock, so we want to
 870          * make sure that obtaining the committed_data is done
 871          * atomically wrt. completion of any outstanding commits.
 872          */
 873         err = do_get_write_access(handle, jh, 1);
 874         if (err)
 875                 goto out;
 876
 877 repeat:
 878         if (!jh->b_committed_data) {
 879                 committed_data = jbd_kmalloc(jh2bh(jh)->b_size, GFP_NOFS);
 880                 if (!committed_data) {
 881                         printk(KERN_EMERG "%s: No memory for committed data\n",
 882                                 __FUNCTION__);
 883                         err = -ENOMEM;
 884                         goto out;
 885                 }
 886         }
 887
 888         jbd_lock_bh_state(bh);
 889         if (!jh->b_committed_data) {
 890                 /* Copy out the current buffer contents into the
 891                  * preserved, committed copy. */
 892                 JBUFFER_TRACE(jh, "generate b_committed data");
 893                 if (!committed_data) {
 894                         jbd_unlock_bh_state(bh);
 895                         goto repeat;
 896                 }
 897
 898                 jh->b_committed_data = committed_data;
 899                 committed_data = NULL;
 900                 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
 901         }
 902         jbd_unlock_bh_state(bh);
 903 out:
 904         journal_put_journal_head(jh);
 905         if (committed_data)
 906                 kfree(committed_data);
 907         return err;
 908 }
 909
 910 /**
 911  * int journal_dirty_data() -  mark a buffer as containing dirty data which
 912  *                             needs to be flushed before we can commit the
 913  *                             current transaction.
 914  * @handle: transaction
 915  * @bh: bufferhead to mark
 916  *
 917  * The buffer is placed on the transaction's data list and is marked as
 918  * belonging to the transaction.
 919  *
 920  * Returns error number or 0 on success.
 921  *
 922  * journal_dirty_data() can be called via page_launder->ext3_writepage
 923  * by kswapd.
 924  */
 925 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
 926 {
 927         journal_t *journal = handle->h_transaction->t_journal;
 928         int need_brelse = 0;
 929         struct journal_head *jh;
 930
 931         if (is_handle_aborted(handle))
 932                 return 0;
 933
 934         jh = journal_add_journal_head(bh);
 935         JBUFFER_TRACE(jh, "entry");
 936
 937         /*
 938          * The buffer could *already* be dirty.  Writeout can start
 939          * at any time.
 940          */
 941         jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
 942
 943         /*
 944          * What if the buffer is already part of a running transaction?
 945          *
 946          * There are two cases:
 947          * 1) It is part of the current running transaction.  Refile it,
 948          *    just in case we have allocated it as metadata, deallocated
 949          *    it, then reallocated it as data.
 950          * 2) It is part of the previous, still-committing transaction.
 951          *    If all we want to do is to guarantee that the buffer will be
 952          *    written to disk before this new transaction commits, then
 953          *    being sure that the *previous* transaction has this same
 954          *    property is sufficient for us!  Just leave it on its old
 955          *    transaction.
 956          *
 957          * In case (2), the buffer must not already exist as metadata
 958          * --- that would violate write ordering (a transaction is free
 959          * to write its data at any point, even before the previous
 960          * committing transaction has committed).  The caller must
 961          * never, ever allow this to happen: there's nothing we can do
 962          * about it in this layer.
 963          */
 964         jbd_lock_bh_state(bh);
 965         spin_lock(&journal->j_list_lock);
 966         if (jh->b_transaction) {
 967                 JBUFFER_TRACE(jh, "has transaction");
 968                 if (jh->b_transaction != handle->h_transaction) {
 969                         JBUFFER_TRACE(jh, "belongs to older transaction");
 970                         J_ASSERT_JH(jh, jh->b_transaction ==
 971                                         journal->j_committing_transaction);
 972
 973                         /* @@@ IS THIS TRUE  ? */
 974                         /*
 975                          * Not any more.  Scenario: someone does a write()
 976                          * in data=journal mode.  The buffer's transaction has
 977                          * moved into commit.  Then someone does another
 978                          * write() to the file.  We do the frozen data copyout
 979                          * and set b_next_transaction to point to j_running_t.
 980                          * And while we're in that state, someone does a
 981                          * writepage() in an attempt to pageout the same area
 982                          * of the file via a shared mapping.  At present that
 983                          * calls journal_dirty_data(), and we get right here.
 984                          * It may be too late to journal the data.  Simply
 985                          * falling through to the next test will suffice: the
 986                          * data will be dirty and wil be checkpointed.  The
 987                          * ordering comments in the next comment block still
 988                          * apply.
 989                          */
 990                         //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 991
 992                         /*
 993                          * If we're journalling data, and this buffer was
 994                          * subject to a write(), it could be metadata, forget
 995                          * or shadow against the committing transaction.  Now,
 996                          * someone has dirtied the same darn page via a mapping
 997                          * and it is being writepage()'d.
 998                          * We *could* just steal the page from commit, with some
 999                          * fancy locking there.  Instead, we just skip it -
1000                          * don't tie the page's buffers to the new transaction
1001                          * at all.
1002                          * Implication: if we crash before the writepage() data
1003                          * is written into the filesystem, recovery will replay
1004                          * the write() data.
1005                          */
1006                         if (jh->b_jlist != BJ_None &&
1007                                         jh->b_jlist != BJ_SyncData &&
1008                                         jh->b_jlist != BJ_Locked) {
1009                                 JBUFFER_TRACE(jh, "Not stealing");
1010                                 goto no_journal;
1011                         }
1012
1013                         /*
1014                          * This buffer may be undergoing writeout in commit.  We
1015                          * can't return from here and let the caller dirty it
1016                          * again because that can cause the write-out loop in
1017                          * commit to never terminate.
1018                          */
1019                         if (buffer_dirty(bh)) {
1020                                 get_bh(bh);
1021                                 spin_unlock(&journal->j_list_lock);
1022                                 jbd_unlock_bh_state(bh);
1023                                 need_brelse = 1;
1024                                 sync_dirty_buffer(bh);
1025                                 jbd_lock_bh_state(bh);
1026                                 spin_lock(&journal->j_list_lock);
1027                                 /* The buffer may become locked again at any
1028                                    time if it is redirtied */
1029                         }
1030
1031                         /* journal_clean_data_list() may have got there first */
1032                         if (jh->b_transaction != NULL) {
1033                                 JBUFFER_TRACE(jh, "unfile from commit");
1034                                 __journal_temp_unlink_buffer(jh);
1035                                 /* It still points to the committing
1036                                  * transaction; move it to this one so
1037                                  * that the refile assert checks are
1038                                  * happy. */
1039                                 jh->b_transaction = handle->h_transaction;
1040                         }
1041                         /* The buffer will be refiled below */
1042
1043                 }
1044                 /*
1045                  * Special case --- the buffer might actually have been
1046                  * allocated and then immediately deallocated in the previous,
1047                  * committing transaction, so might still be left on that
1048                  * transaction's metadata lists.
1049                  */
1050                 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1051                         JBUFFER_TRACE(jh, "not on correct data list: unfile");
1052                         J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1053                         __journal_temp_unlink_buffer(jh);
1054                         jh->b_transaction = handle->h_transaction;
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         if (jh->b_modified == 0) {
1108                 /*
1109                  * This buffer's got modified and becoming part
1110                  * of the transaction. This needs to be done
1111                  * once a transaction -bzzz
1112                  */
1113                 jh->b_modified = 1;
1114                 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1115                 handle->h_buffer_credits--;
1116         }
1117
1118         /*
1119          * fastpath, to avoid expensive locking.  If this buffer is already
1120          * on the running transaction's metadata list there is nothing to do.
1121          * Nobody can take it off again because there is a handle open.
1122          * I _think_ we're OK here with SMP barriers - a mistaken decision will
1123          * result in this test being false, so we go in and take the locks.
1124          */
1125         if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1126                 JBUFFER_TRACE(jh, "fastpath");
1127                 J_ASSERT_JH(jh, jh->b_transaction ==
1128                                         journal->j_running_transaction);
1129                 goto out_unlock_bh;
1130         }
1131
1132         set_buffer_jbddirty(bh);
1133
1134         /*
1135          * Metadata already on the current transaction list doesn't
1136          * need to be filed.  Metadata on another transaction's list must
1137          * be committing, and will be refiled once the commit completes:
1138          * leave it alone for now.
1139          */
1140         if (jh->b_transaction != transaction) {
1141                 JBUFFER_TRACE(jh, "already on other transaction");
1142                 J_ASSERT_JH(jh, jh->b_transaction ==
1143                                         journal->j_committing_transaction);
1144                 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1145                 /* And this case is illegal: we can't reuse another
1146                  * transaction's data buffer, ever. */
1147                 goto out_unlock_bh;
1148         }
1149
1150         /* That test should have eliminated the following case: */
1151         J_ASSERT_JH(jh, jh->b_frozen_data == 0);
1152
1153         JBUFFER_TRACE(jh, "file as BJ_Metadata");
1154         spin_lock(&journal->j_list_lock);
1155         __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1156         spin_unlock(&journal->j_list_lock);
1157 out_unlock_bh:
1158         jbd_unlock_bh_state(bh);
1159 out:
1160         JBUFFER_TRACE(jh, "exit");
1161         return 0;
1162 }
1163
1164 /*
1165  * journal_release_buffer: undo a get_write_access without any buffer
1166  * updates, if the update decided in the end that it didn't need access.
1167  *
1168  */
1169 void
1170 journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1171 {
1172         BUFFER_TRACE(bh, "entry");
1173 }
1174
1175 /**
1176  * void journal_forget() - bforget() for potentially-journaled buffers.
1177  * @handle: transaction handle
1178  * @bh:     bh to 'forget'
1179  *
1180  * We can only do the bforget if there are no commits pending against the
1181  * buffer.  If the buffer is dirty in the current running transaction we
1182  * can safely unlink it.
1183  *
1184  * bh may not be a journalled buffer at all - it may be a non-JBD
1185  * buffer which came off the hashtable.  Check for this.
1186  *
1187  * Decrements bh->b_count by one.
1188  *
1189  * Allow this call even if the handle has aborted --- it may be part of
1190  * the caller's cleanup after an abort.
1191  */
1192 int journal_forget (handle_t *handle, struct buffer_head *bh)
1193 {
1194         transaction_t *transaction = handle->h_transaction;
1195         journal_t *journal = transaction->t_journal;
1196         struct journal_head *jh;
1197         int drop_reserve = 0;
1198         int err = 0;
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         /* Critical error: attempting to delete a bitmap buffer, maybe?
1210          * Don't do any jbd operations, and return an error. */
1211         if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1212                          "inconsistent data on disk")) {
1213                 err = -EIO;
1214                 goto not_jbd;
1215         }
1216
1217         /*
1218          * The buffer's going from the transaction, we must drop
1219          * all references -bzzz
1220          */
1221         jh->b_modified = 0;
1222
1223         if (jh->b_transaction == handle->h_transaction) {
1224                 J_ASSERT_JH(jh, !jh->b_frozen_data);
1225
1226                 /* If we are forgetting a buffer which is already part
1227                  * of this transaction, then we can just drop it from
1228                  * the transaction immediately. */
1229                 clear_buffer_dirty(bh);
1230                 clear_buffer_jbddirty(bh);
1231
1232                 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1233
1234                 drop_reserve = 1;
1235
1236                 /*
1237                  * We are no longer going to journal this buffer.
1238                  * However, the commit of this transaction is still
1239                  * important to the buffer: the delete that we are now
1240                  * processing might obsolete an old log entry, so by
1241                  * committing, we can satisfy the buffer's checkpoint.
1242                  *
1243                  * So, if we have a checkpoint on the buffer, we should
1244                  * now refile the buffer on our BJ_Forget list so that
1245                  * we know to remove the checkpoint after we commit.
1246                  */
1247
1248                 if (jh->b_cp_transaction) {
1249                         __journal_temp_unlink_buffer(jh);
1250                         __journal_file_buffer(jh, transaction, BJ_Forget);
1251                 } else {
1252                         __journal_unfile_buffer(jh);
1253                         journal_remove_journal_head(bh);
1254                         __brelse(bh);
1255                         if (!buffer_jbd(bh)) {
1256                                 spin_unlock(&journal->j_list_lock);
1257                                 jbd_unlock_bh_state(bh);
1258                                 __bforget(bh);
1259                                 goto drop;
1260                         }
1261                 }
1262         } else if (jh->b_transaction) {
1263                 J_ASSERT_JH(jh, (jh->b_transaction ==
1264                                  journal->j_committing_transaction));
1265                 /* However, if the buffer is still owned by a prior
1266                  * (committing) transaction, we can't drop it yet... */
1267                 JBUFFER_TRACE(jh, "belongs to older transaction");
1268                 /* ... but we CAN drop it from the new transaction if we
1269                  * have also modified it since the original commit. */
1270
1271                 if (jh->b_next_transaction) {
1272                         J_ASSERT(jh->b_next_transaction == transaction);
1273                         jh->b_next_transaction = NULL;
1274                         drop_reserve = 1;
1275                 }
1276         }
1277
1278 not_jbd:
1279         spin_unlock(&journal->j_list_lock);
1280         jbd_unlock_bh_state(bh);
1281         __brelse(bh);
1282 drop:
1283         if (drop_reserve) {
1284                 /* no need to reserve log space for this block -bzzz */
1285                 handle->h_buffer_credits++;
1286         }
1287         return err;
1288 }
1289
1290 /**
1291  * int journal_stop() - complete a transaction
1292  * @handle: tranaction to complete.
1293  *
1294  * All done for a particular handle.
1295  *
1296  * There is not much action needed here.  We just return any remaining
1297  * buffer credits to the transaction and remove the handle.  The only
1298  * complication is that we need to start a commit operation if the
1299  * filesystem is marked for synchronous update.
1300  *
1301  * journal_stop itself will not usually return an error, but it may
1302  * do so in unusual circumstances.  In particular, expect it to
1303  * return -EIO if a journal_abort has been executed since the
1304  * transaction began.
1305  */
1306 int journal_stop(handle_t *handle)
1307 {
1308         transaction_t *transaction = handle->h_transaction;
1309         journal_t *journal = transaction->t_journal;
1310         int old_handle_count, err;
1311
1312         J_ASSERT(transaction->t_updates > 0);
1313         J_ASSERT(journal_current_handle() == handle);
1314
1315         if (is_handle_aborted(handle))
1316                 err = -EIO;
1317         else
1318                 err = 0;
1319
1320         if (--handle->h_ref > 0) {
1321                 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1322                           handle->h_ref);
1323                 return err;
1324         }
1325
1326         jbd_debug(4, "Handle %p going down\n", handle);
1327
1328         /*
1329          * Implement synchronous transaction batching.  If the handle
1330          * was synchronous, don't force a commit immediately.  Let's
1331          * yield and let another thread piggyback onto this transaction.
1332          * Keep doing that while new threads continue to arrive.
1333          * It doesn't cost much - we're about to run a commit and sleep
1334          * on IO anyway.  Speeds up many-threaded, many-dir operations
1335          * by 30x or more...
1336          */
1337         if (handle->h_sync) {
1338                 do {
1339                         old_handle_count = transaction->t_handle_count;
1340                         set_current_state(TASK_UNINTERRUPTIBLE);
1341                         schedule_timeout(1);
1342                 } while (old_handle_count != transaction->t_handle_count);
1343         }
1344
1345         current->journal_info = NULL;
1346         spin_lock(&journal->j_state_lock);
1347         spin_lock(&transaction->t_handle_lock);
1348         transaction->t_outstanding_credits -= handle->h_buffer_credits;
1349         transaction->t_updates--;
1350         if (!transaction->t_updates) {
1351                 wake_up(&journal->j_wait_updates);
1352                 if (journal->j_barrier_count)
1353                         wake_up(&journal->j_wait_transaction_locked);
1354         }
1355
1356         /*
1357          * If the handle is marked SYNC, we need to set another commit
1358          * going!  We also want to force a commit if the current
1359          * transaction is occupying too much of the log, or if the
1360          * transaction is too old now.
1361          */
1362         if (handle->h_sync ||
1363                         transaction->t_outstanding_credits >
1364                                 journal->j_max_transaction_buffers ||
1365                         time_after_eq(jiffies, transaction->t_expires)) {
1366                 /* Do this even for aborted journals: an abort still
1367                  * completes the commit thread, it just doesn't write
1368                  * anything to disk. */
1369                 tid_t tid = transaction->t_tid;
1370
1371                 spin_unlock(&transaction->t_handle_lock);
1372                 jbd_debug(2, "transaction too old, requesting commit for "
1373                                         "handle %p\n", handle);
1374                 /* This is non-blocking */
1375                 __log_start_commit(journal, transaction->t_tid);
1376                 spin_unlock(&journal->j_state_lock);
1377
1378                 /*
1379                  * Special case: JFS_SYNC synchronous updates require us
1380                  * to wait for the commit to complete.
1381                  */
1382                 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1383                         err = log_wait_commit(journal, tid);
1384         } else {
1385                 spin_unlock(&transaction->t_handle_lock);
1386                 spin_unlock(&journal->j_state_lock);
1387         }
1388
1389         jbd_free_handle(handle);
1390         return err;
1391 }
1392
1393 /**int journal_force_commit() - force any uncommitted transactions
1394  * @journal: journal to force
1395  *
1396  * For synchronous operations: force any uncommitted transactions
1397  * to disk.  May seem kludgy, but it reuses all the handle batching
1398  * code in a very simple manner.
1399  */
1400 int journal_force_commit(journal_t *journal)
1401 {
1402         handle_t *handle;
1403         int ret;
1404
1405         handle = journal_start(journal, 1);
1406         if (IS_ERR(handle)) {
1407                 ret = PTR_ERR(handle);
1408         } else {
1409                 handle->h_sync = 1;
1410                 ret = journal_stop(handle);
1411         }
1412         return ret;
1413 }
1414
1415 /*
1416  *
1417  * List management code snippets: various functions for manipulating the
1418  * transaction buffer lists.
1419  *
1420  */
1421
1422 /*
1423  * Append a buffer to a transaction list, given the transaction's list head
1424  * pointer.
1425  *
1426  * j_list_lock is held.
1427  *
1428  * jbd_lock_bh_state(jh2bh(jh)) is held.
1429  */
1430
1431 static inline void
1432 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1433 {
1434         if (!*list) {
1435                 jh->b_tnext = jh->b_tprev = jh;
1436                 *list = jh;
1437         } else {
1438                 /* Insert at the tail of the list to preserve order */
1439                 struct journal_head *first = *list, *last = first->b_tprev;
1440                 jh->b_tprev = last;
1441                 jh->b_tnext = first;
1442                 last->b_tnext = first->b_tprev = jh;
1443         }
1444 }
1445
1446 /*
1447  * Remove a buffer from a transaction list, given the transaction's list
1448  * head pointer.
1449  *
1450  * Called with j_list_lock held, and the journal may not be locked.
1451  *
1452  * jbd_lock_bh_state(jh2bh(jh)) is held.
1453  */
1454
1455 static inline void
1456 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1457 {
1458         if (*list == jh) {
1459                 *list = jh->b_tnext;
1460                 if (*list == jh)
1461                         *list = NULL;
1462         }
1463         jh->b_tprev->b_tnext = jh->b_tnext;
1464         jh->b_tnext->b_tprev = jh->b_tprev;
1465 }
1466
1467 /*
1468  * Remove a buffer from the appropriate transaction list.
1469  *
1470  * Note that this function can *change* the value of
1471  * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1472  * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list.  If the caller
1473  * is holding onto a copy of one of thee pointers, it could go bad.
1474  * Generally the caller needs to re-read the pointer from the transaction_t.
1475  *
1476  * Called under j_list_lock.  The journal may not be locked.
1477  */
1478 void __journal_temp_unlink_buffer(struct journal_head *jh)
1479 {
1480         struct journal_head **list = NULL;
1481         transaction_t *transaction;
1482         struct buffer_head *bh = jh2bh(jh);
1483
1484         J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1485         transaction = jh->b_transaction;
1486         if (transaction)
1487                 assert_spin_locked(&transaction->t_journal->j_list_lock);
1488
1489         J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1490         if (jh->b_jlist != BJ_None)
1491                 J_ASSERT_JH(jh, transaction != 0);
1492
1493         switch (jh->b_jlist) {
1494         case BJ_None:
1495                 return;
1496         case BJ_SyncData:
1497                 list = &transaction->t_sync_datalist;
1498                 break;
1499         case BJ_Metadata:
1500                 transaction->t_nr_buffers--;
1501                 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1502                 list = &transaction->t_buffers;
1503                 break;
1504         case BJ_Forget:
1505                 list = &transaction->t_forget;
1506                 break;
1507         case BJ_IO:
1508                 list = &transaction->t_iobuf_list;
1509                 break;
1510         case BJ_Shadow:
1511                 list = &transaction->t_shadow_list;
1512                 break;
1513         case BJ_LogCtl:
1514                 list = &transaction->t_log_list;
1515                 break;
1516         case BJ_Reserved:
1517                 list = &transaction->t_reserved_list;
1518                 break;
1519         case BJ_Locked:
1520                 list = &transaction->t_locked_list;
1521                 break;
1522         }
1523
1524         __blist_del_buffer(list, jh);
1525         jh->b_jlist = BJ_None;
1526         if (test_clear_buffer_jbddirty(bh))
1527                 mark_buffer_dirty(bh);  /* Expose it to the VM */
1528 }
1529
1530 void __journal_unfile_buffer(struct journal_head *jh)
1531 {
1532         __journal_temp_unlink_buffer(jh);
1533         jh->b_transaction = NULL;
1534 }
1535
1536 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1537 {
1538         jbd_lock_bh_state(jh2bh(jh));
1539         spin_lock(&journal->j_list_lock);
1540         __journal_unfile_buffer(jh);
1541         spin_unlock(&journal->j_list_lock);
1542         jbd_unlock_bh_state(jh2bh(jh));
1543 }
1544
1545 /*
1546  * Called from journal_try_to_free_buffers().
1547  *
1548  * Called under jbd_lock_bh_state(bh)
1549  */
1550 static void
1551 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1552 {
1553         struct journal_head *jh;
1554
1555         jh = bh2jh(bh);
1556
1557         if (buffer_locked(bh) || buffer_dirty(bh))
1558                 goto out;
1559
1560         if (jh->b_next_transaction != 0)
1561                 goto out;
1562
1563         spin_lock(&journal->j_list_lock);
1564         if (jh->b_transaction != 0 && jh->b_cp_transaction == 0) {
1565                 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1566                         /* A written-back ordered data buffer */
1567                         JBUFFER_TRACE(jh, "release data");
1568                         __journal_unfile_buffer(jh);
1569                         journal_remove_journal_head(bh);
1570                         __brelse(bh);
1571                 }
1572         } else if (jh->b_cp_transaction != 0 && jh->b_transaction == 0) {
1573                 /* written-back checkpointed metadata buffer */
1574                 if (jh->b_jlist == BJ_None) {
1575                         JBUFFER_TRACE(jh, "remove from checkpoint list");
1576                         __journal_remove_checkpoint(jh);
1577                         journal_remove_journal_head(bh);
1578                         __brelse(bh);
1579                 }
1580         }
1581         spin_unlock(&journal->j_list_lock);
1582 out:
1583         return;
1584 }
1585
1586
1587 /**
1588  * int journal_try_to_free_buffers() - try to free page buffers.
1589  * @journal: journal for operation
1590  * @page: to try and free
1591  * @unused_gfp_mask: unused
1592  *
1593  *
1594  * For all the buffers on this page,
1595  * if they are fully written out ordered data, move them onto BUF_CLEAN
1596  * so try_to_free_buffers() can reap them.
1597  *
1598  * This function returns non-zero if we wish try_to_free_buffers()
1599  * to be called. We do this if the page is releasable by try_to_free_buffers().
1600  * We also do it if the page has locked or dirty buffers and the caller wants
1601  * us to perform sync or async writeout.
1602  *
1603  * This complicates JBD locking somewhat.  We aren't protected by the
1604  * BKL here.  We wish to remove the buffer from its committing or
1605  * running transaction's ->t_datalist via __journal_unfile_buffer.
1606  *
1607  * This may *change* the value of transaction_t->t_datalist, so anyone
1608  * who looks at t_datalist needs to lock against this function.
1609  *
1610  * Even worse, someone may be doing a journal_dirty_data on this
1611  * buffer.  So we need to lock against that.  journal_dirty_data()
1612  * will come out of the lock with the buffer dirty, which makes it
1613  * ineligible for release here.
1614  *
1615  * Who else is affected by this?  hmm...  Really the only contender
1616  * is do_get_write_access() - it could be looking at the buffer while
1617  * journal_try_to_free_buffer() is changing its state.  But that
1618  * cannot happen because we never reallocate freed data as metadata
1619  * while the data is part of a transaction.  Yes?
1620  */
1621 int journal_try_to_free_buffers(journal_t *journal,
1622                                 struct page *page, int unused_gfp_mask)
1623 {
1624         struct buffer_head *head;
1625         struct buffer_head *bh;
1626         int ret = 0;
1627
1628         J_ASSERT(PageLocked(page));
1629
1630         head = page_buffers(page);
1631         bh = head;
1632         do {
1633                 struct journal_head *jh;
1634
1635                 /*
1636                  * We take our own ref against the journal_head here to avoid
1637                  * having to add tons of locking around each instance of
1638                  * journal_remove_journal_head() and journal_put_journal_head().
1639                  */
1640                 jh = journal_grab_journal_head(bh);
1641                 if (!jh)
1642                         continue;
1643
1644                 jbd_lock_bh_state(bh);
1645                 __journal_try_to_free_buffer(journal, bh);
1646                 journal_put_journal_head(jh);
1647                 jbd_unlock_bh_state(bh);
1648                 if (buffer_jbd(bh))
1649                         goto busy;
1650         } while ((bh = bh->b_this_page) != head);
1651         ret = try_to_free_buffers(page);
1652 busy:
1653         return ret;
1654 }
1655
1656 /*
1657  * This buffer is no longer needed.  If it is on an older transaction's
1658  * checkpoint list we need to record it on this transaction's forget list
1659  * to pin this buffer (and hence its checkpointing transaction) down until
1660  * this transaction commits.  If the buffer isn't on a checkpoint list, we
1661  * release it.
1662  * Returns non-zero if JBD no longer has an interest in the buffer.
1663  *
1664  * Called under j_list_lock.
1665  *
1666  * Called under jbd_lock_bh_state(bh).
1667  */
1668 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1669 {
1670         int may_free = 1;
1671         struct buffer_head *bh = jh2bh(jh);
1672
1673         __journal_unfile_buffer(jh);
1674
1675         if (jh->b_cp_transaction) {
1676                 JBUFFER_TRACE(jh, "on running+cp transaction");
1677                 __journal_file_buffer(jh, transaction, BJ_Forget);
1678                 clear_buffer_jbddirty(bh);
1679                 may_free = 0;
1680         } else {
1681                 JBUFFER_TRACE(jh, "on running transaction");
1682                 journal_remove_journal_head(bh);
1683                 __brelse(bh);
1684         }
1685         return may_free;
1686 }
1687
1688 /*
1689  * journal_invalidatepage
1690  *
1691  * This code is tricky.  It has a number of cases to deal with.
1692  *
1693  * There are two invariants which this code relies on:
1694  *
1695  * i_size must be updated on disk before we start calling invalidatepage on the
1696  * data.
1697  *
1698  *  This is done in ext3 by defining an ext3_setattr method which
1699  *  updates i_size before truncate gets going.  By maintaining this
1700  *  invariant, we can be sure that it is safe to throw away any buffers
1701  *  attached to the current transaction: once the transaction commits,
1702  *  we know that the data will not be needed.
1703  *
1704  *  Note however that we can *not* throw away data belonging to the
1705  *  previous, committing transaction!
1706  *
1707  * Any disk blocks which *are* part of the previous, committing
1708  * transaction (and which therefore cannot be discarded immediately) are
1709  * not going to be reused in the new running transaction
1710  *
1711  *  The bitmap committed_data images guarantee this: any block which is
1712  *  allocated in one transaction and removed in the next will be marked
1713  *  as in-use in the committed_data bitmap, so cannot be reused until
1714  *  the next transaction to delete the block commits.  This means that
1715  *  leaving committing buffers dirty is quite safe: the disk blocks
1716  *  cannot be reallocated to a different file and so buffer aliasing is
1717  *  not possible.
1718  *
1719  *
1720  * The above applies mainly to ordered data mode.  In writeback mode we
1721  * don't make guarantees about the order in which data hits disk --- in
1722  * particular we don't guarantee that new dirty data is flushed before
1723  * transaction commit --- so it is always safe just to discard data
1724  * immediately in that mode.  --sct
1725  */
1726
1727 /*
1728  * The journal_unmap_buffer helper function returns zero if the buffer
1729  * concerned remains pinned as an anonymous buffer belonging to an older
1730  * transaction.
1731  *
1732  * We're outside-transaction here.  Either or both of j_running_transaction
1733  * and j_committing_transaction may be NULL.
1734  */
1735 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1736 {
1737         transaction_t *transaction;
1738         struct journal_head *jh;
1739         int may_free = 1;
1740         int ret;
1741
1742         BUFFER_TRACE(bh, "entry");
1743
1744         /*
1745          * It is safe to proceed here without the j_list_lock because the
1746          * buffers cannot be stolen by try_to_free_buffers as long as we are
1747          * holding the page lock. --sct
1748          */
1749
1750         if (!buffer_jbd(bh))
1751                 goto zap_buffer_unlocked;
1752
1753         spin_lock(&journal->j_state_lock);
1754         jbd_lock_bh_state(bh);
1755         spin_lock(&journal->j_list_lock);
1756
1757         jh = journal_grab_journal_head(bh);
1758         if (!jh)
1759                 goto zap_buffer_no_jh;
1760
1761         transaction = jh->b_transaction;
1762         if (transaction == NULL) {
1763                 /* First case: not on any transaction.  If it
1764                  * has no checkpoint link, then we can zap it:
1765                  * it's a writeback-mode buffer so we don't care
1766                  * if it hits disk safely. */
1767                 if (!jh->b_cp_transaction) {
1768                         JBUFFER_TRACE(jh, "not on any transaction: zap");
1769                         goto zap_buffer;
1770                 }
1771
1772                 if (!buffer_dirty(bh)) {
1773                         /* bdflush has written it.  We can drop it now */
1774                         goto zap_buffer;
1775                 }
1776
1777                 /* OK, it must be in the journal but still not
1778                  * written fully to disk: it's metadata or
1779                  * journaled data... */
1780
1781                 if (journal->j_running_transaction) {
1782                         /* ... and once the current transaction has
1783                          * committed, the buffer won't be needed any
1784                          * longer. */
1785                         JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1786                         ret = __dispose_buffer(jh,
1787                                         journal->j_running_transaction);
1788                         journal_put_journal_head(jh);
1789                         spin_unlock(&journal->j_list_lock);
1790                         jbd_unlock_bh_state(bh);
1791                         spin_unlock(&journal->j_state_lock);
1792                         return ret;
1793                 } else {
1794                         /* There is no currently-running transaction. So the
1795                          * orphan record which we wrote for this file must have
1796                          * passed into commit.  We must attach this buffer to
1797                          * the committing transaction, if it exists. */
1798                         if (journal->j_committing_transaction) {
1799                                 JBUFFER_TRACE(jh, "give to committing trans");
1800                                 ret = __dispose_buffer(jh,
1801                                         journal->j_committing_transaction);
1802                                 journal_put_journal_head(jh);
1803                                 spin_unlock(&journal->j_list_lock);
1804                                 jbd_unlock_bh_state(bh);
1805                                 spin_unlock(&journal->j_state_lock);
1806                                 return ret;
1807                         } else {
1808                                 /* The orphan record's transaction has
1809                                  * committed.  We can cleanse this buffer */
1810                                 clear_buffer_jbddirty(bh);
1811                                 goto zap_buffer;
1812                         }
1813                 }
1814         } else if (transaction == journal->j_committing_transaction) {
1815                 if (jh->b_jlist == BJ_Locked) {
1816                         /*
1817                          * The buffer is on the committing transaction's locked
1818                          * list.  We have the buffer locked, so I/O has
1819                          * completed.  So we can nail the buffer now.
1820                          */
1821                         may_free = __dispose_buffer(jh, transaction);
1822                         goto zap_buffer;
1823                 }
1824                 /*
1825                  * If it is committing, we simply cannot touch it.  We
1826                  * can remove it's next_transaction pointer from the
1827                  * running transaction if that is set, but nothing
1828                  * else. */
1829                 JBUFFER_TRACE(jh, "on committing transaction");
1830                 set_buffer_freed(bh);
1831                 if (jh->b_next_transaction) {
1832                         J_ASSERT(jh->b_next_transaction ==
1833                                         journal->j_running_transaction);
1834                         jh->b_next_transaction = NULL;
1835                 }
1836                 journal_put_journal_head(jh);
1837                 spin_unlock(&journal->j_list_lock);
1838                 jbd_unlock_bh_state(bh);
1839                 spin_unlock(&journal->j_state_lock);
1840                 return 0;
1841         } else {
1842                 /* Good, the buffer belongs to the running transaction.
1843                  * We are writing our own transaction's data, not any
1844                  * previous one's, so it is safe to throw it away
1845                  * (remember that we expect the filesystem to have set
1846                  * i_size already for this truncate so recovery will not
1847                  * expose the disk blocks we are discarding here.) */
1848                 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1849                 may_free = __dispose_buffer(jh, transaction);
1850         }
1851
1852 zap_buffer:
1853         journal_put_journal_head(jh);
1854 zap_buffer_no_jh:
1855         spin_unlock(&journal->j_list_lock);
1856         jbd_unlock_bh_state(bh);
1857         spin_unlock(&journal->j_state_lock);
1858 zap_buffer_unlocked:
1859         clear_buffer_dirty(bh);
1860         J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1861         clear_buffer_mapped(bh);
1862         clear_buffer_req(bh);
1863         clear_buffer_new(bh);
1864         bh->b_bdev = NULL;
1865         return may_free;
1866 }
1867
1868 /**
1869  * int journal_invalidatepage()
1870  * @journal: journal to use for flush...
1871  * @page:    page to flush
1872  * @offset:  length of page to invalidate.
1873  *
1874  * Reap page buffers containing data after offset in page.
1875  *
1876  * Return non-zero if the page's buffers were successfully reaped.
1877  */
1878 int journal_invalidatepage(journal_t *journal,
1879                       struct page *page,
1880                       unsigned long offset)
1881 {
1882         struct buffer_head *head, *bh, *next;
1883         unsigned int curr_off = 0;
1884         int may_free = 1;
1885
1886         if (!PageLocked(page))
1887                 BUG();
1888         if (!page_has_buffers(page))
1889                 return 1;
1890
1891         /* We will potentially be playing with lists other than just the
1892          * data lists (especially for journaled data mode), so be
1893          * cautious in our locking. */
1894
1895         head = bh = page_buffers(page);
1896         do {
1897                 unsigned int next_off = curr_off + bh->b_size;
1898                 next = bh->b_this_page;
1899
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 = NULL;
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_temp_unlink_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_temp_unlink_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 }