linux/fs/jbd2/transaction.c
<<
>>
Prefs
   1/*
   2 * linux/fs/jbd2/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/jbd2.h>
  23#include <linux/errno.h>
  24#include <linux/slab.h>
  25#include <linux/timer.h>
  26#include <linux/mm.h>
  27#include <linux/highmem.h>
  28#include <linux/hrtimer.h>
  29#include <linux/backing-dev.h>
  30#include <linux/bug.h>
  31#include <linux/module.h>
  32
  33static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
  34static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
  35
  36static struct kmem_cache *transaction_cache;
  37int __init jbd2_journal_init_transaction_cache(void)
  38{
  39        J_ASSERT(!transaction_cache);
  40        transaction_cache = kmem_cache_create("jbd2_transaction_s",
  41                                        sizeof(transaction_t),
  42                                        0,
  43                                        SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
  44                                        NULL);
  45        if (transaction_cache)
  46                return 0;
  47        return -ENOMEM;
  48}
  49
  50void jbd2_journal_destroy_transaction_cache(void)
  51{
  52        if (transaction_cache) {
  53                kmem_cache_destroy(transaction_cache);
  54                transaction_cache = NULL;
  55        }
  56}
  57
  58void jbd2_journal_free_transaction(transaction_t *transaction)
  59{
  60        if (unlikely(ZERO_OR_NULL_PTR(transaction)))
  61                return;
  62        kmem_cache_free(transaction_cache, transaction);
  63}
  64
  65/*
  66 * jbd2_get_transaction: obtain a new transaction_t object.
  67 *
  68 * Simply allocate and initialise a new transaction.  Create it in
  69 * RUNNING state and add it to the current journal (which should not
  70 * have an existing running transaction: we only make a new transaction
  71 * once we have started to commit the old one).
  72 *
  73 * Preconditions:
  74 *      The journal MUST be locked.  We don't perform atomic mallocs on the
  75 *      new transaction and we can't block without protecting against other
  76 *      processes trying to touch the journal while it is in transition.
  77 *
  78 */
  79
  80static transaction_t *
  81jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
  82{
  83        transaction->t_journal = journal;
  84        transaction->t_state = T_RUNNING;
  85        transaction->t_start_time = ktime_get();
  86        transaction->t_tid = journal->j_transaction_sequence++;
  87        transaction->t_expires = jiffies + journal->j_commit_interval;
  88        spin_lock_init(&transaction->t_handle_lock);
  89        atomic_set(&transaction->t_updates, 0);
  90        atomic_set(&transaction->t_outstanding_credits, 0);
  91        atomic_set(&transaction->t_handle_count, 0);
  92        INIT_LIST_HEAD(&transaction->t_inode_list);
  93        INIT_LIST_HEAD(&transaction->t_private_list);
  94
  95        /* Set up the commit timer for the new transaction. */
  96        journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
  97        add_timer(&journal->j_commit_timer);
  98
  99        J_ASSERT(journal->j_running_transaction == NULL);
 100        journal->j_running_transaction = transaction;
 101        transaction->t_max_wait = 0;
 102        transaction->t_start = jiffies;
 103
 104        return transaction;
 105}
 106
 107/*
 108 * Handle management.
 109 *
 110 * A handle_t is an object which represents a single atomic update to a
 111 * filesystem, and which tracks all of the modifications which form part
 112 * of that one update.
 113 */
 114
 115/*
 116 * Update transaction's maximum wait time, if debugging is enabled.
 117 *
 118 * In order for t_max_wait to be reliable, it must be protected by a
 119 * lock.  But doing so will mean that start_this_handle() can not be
 120 * run in parallel on SMP systems, which limits our scalability.  So
 121 * unless debugging is enabled, we no longer update t_max_wait, which
 122 * means that maximum wait time reported by the jbd2_run_stats
 123 * tracepoint will always be zero.
 124 */
 125static inline void update_t_max_wait(transaction_t *transaction,
 126                                     unsigned long ts)
 127{
 128#ifdef CONFIG_JBD2_DEBUG
 129        if (jbd2_journal_enable_debug &&
 130            time_after(transaction->t_start, ts)) {
 131                ts = jbd2_time_diff(ts, transaction->t_start);
 132                spin_lock(&transaction->t_handle_lock);
 133                if (ts > transaction->t_max_wait)
 134                        transaction->t_max_wait = ts;
 135                spin_unlock(&transaction->t_handle_lock);
 136        }
 137#endif
 138}
 139
 140/*
 141 * start_this_handle: Given a handle, deal with any locking or stalling
 142 * needed to make sure that there is enough journal space for the handle
 143 * to begin.  Attach the handle to a transaction and set up the
 144 * transaction's buffer credits.
 145 */
 146
 147static int start_this_handle(journal_t *journal, handle_t *handle,
 148                             gfp_t gfp_mask)
 149{
 150        transaction_t   *transaction, *new_transaction = NULL;
 151        tid_t           tid;
 152        int             needed, need_to_start;
 153        int             nblocks = handle->h_buffer_credits;
 154        unsigned long ts = jiffies;
 155
 156        if (nblocks > journal->j_max_transaction_buffers) {
 157                printk(KERN_ERR "JBD2: %s wants too many credits (%d > %d)\n",
 158                       current->comm, nblocks,
 159                       journal->j_max_transaction_buffers);
 160                return -ENOSPC;
 161        }
 162
 163alloc_transaction:
 164        if (!journal->j_running_transaction) {
 165                new_transaction = kmem_cache_zalloc(transaction_cache,
 166                                                    gfp_mask);
 167                if (!new_transaction) {
 168                        /*
 169                         * If __GFP_FS is not present, then we may be
 170                         * being called from inside the fs writeback
 171                         * layer, so we MUST NOT fail.  Since
 172                         * __GFP_NOFAIL is going away, we will arrange
 173                         * to retry the allocation ourselves.
 174                         */
 175                        if ((gfp_mask & __GFP_FS) == 0) {
 176                                congestion_wait(BLK_RW_ASYNC, HZ/50);
 177                                goto alloc_transaction;
 178                        }
 179                        return -ENOMEM;
 180                }
 181        }
 182
 183        jbd_debug(3, "New handle %p going live.\n", handle);
 184
 185        /*
 186         * We need to hold j_state_lock until t_updates has been incremented,
 187         * for proper journal barrier handling
 188         */
 189repeat:
 190        read_lock(&journal->j_state_lock);
 191        BUG_ON(journal->j_flags & JBD2_UNMOUNT);
 192        if (is_journal_aborted(journal) ||
 193            (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
 194                read_unlock(&journal->j_state_lock);
 195                jbd2_journal_free_transaction(new_transaction);
 196                return -EROFS;
 197        }
 198
 199        /* Wait on the journal's transaction barrier if necessary */
 200        if (journal->j_barrier_count) {
 201                read_unlock(&journal->j_state_lock);
 202                wait_event(journal->j_wait_transaction_locked,
 203                                journal->j_barrier_count == 0);
 204                goto repeat;
 205        }
 206
 207        if (!journal->j_running_transaction) {
 208                read_unlock(&journal->j_state_lock);
 209                if (!new_transaction)
 210                        goto alloc_transaction;
 211                write_lock(&journal->j_state_lock);
 212                if (!journal->j_running_transaction) {
 213                        jbd2_get_transaction(journal, new_transaction);
 214                        new_transaction = NULL;
 215                }
 216                write_unlock(&journal->j_state_lock);
 217                goto repeat;
 218        }
 219
 220        transaction = journal->j_running_transaction;
 221
 222        /*
 223         * If the current transaction is locked down for commit, wait for the
 224         * lock to be released.
 225         */
 226        if (transaction->t_state == T_LOCKED) {
 227                DEFINE_WAIT(wait);
 228
 229                prepare_to_wait(&journal->j_wait_transaction_locked,
 230                                        &wait, TASK_UNINTERRUPTIBLE);
 231                read_unlock(&journal->j_state_lock);
 232                schedule();
 233                finish_wait(&journal->j_wait_transaction_locked, &wait);
 234                goto repeat;
 235        }
 236
 237        /*
 238         * If there is not enough space left in the log to write all potential
 239         * buffers requested by this operation, we need to stall pending a log
 240         * checkpoint to free some more log space.
 241         */
 242        needed = atomic_add_return(nblocks,
 243                                   &transaction->t_outstanding_credits);
 244
 245        if (needed > journal->j_max_transaction_buffers) {
 246                /*
 247                 * If the current transaction is already too large, then start
 248                 * to commit it: we can then go back and attach this handle to
 249                 * a new transaction.
 250                 */
 251                DEFINE_WAIT(wait);
 252
 253                jbd_debug(2, "Handle %p starting new commit...\n", handle);
 254                atomic_sub(nblocks, &transaction->t_outstanding_credits);
 255                prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
 256                                TASK_UNINTERRUPTIBLE);
 257                tid = transaction->t_tid;
 258                need_to_start = !tid_geq(journal->j_commit_request, tid);
 259                read_unlock(&journal->j_state_lock);
 260                if (need_to_start)
 261                        jbd2_log_start_commit(journal, tid);
 262                schedule();
 263                finish_wait(&journal->j_wait_transaction_locked, &wait);
 264                goto repeat;
 265        }
 266
 267        /*
 268         * The commit code assumes that it can get enough log space
 269         * without forcing a checkpoint.  This is *critical* for
 270         * correctness: a checkpoint of a buffer which is also
 271         * associated with a committing transaction creates a deadlock,
 272         * so commit simply cannot force through checkpoints.
 273         *
 274         * We must therefore ensure the necessary space in the journal
 275         * *before* starting to dirty potentially checkpointed buffers
 276         * in the new transaction.
 277         *
 278         * The worst part is, any transaction currently committing can
 279         * reduce the free space arbitrarily.  Be careful to account for
 280         * those buffers when checkpointing.
 281         */
 282
 283        /*
 284         * @@@ AKPM: This seems rather over-defensive.  We're giving commit
 285         * a _lot_ of headroom: 1/4 of the journal plus the size of
 286         * the committing transaction.  Really, we only need to give it
 287         * committing_transaction->t_outstanding_credits plus "enough" for
 288         * the log control blocks.
 289         * Also, this test is inconsistent with the matching one in
 290         * jbd2_journal_extend().
 291         */
 292        if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
 293                jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
 294                atomic_sub(nblocks, &transaction->t_outstanding_credits);
 295                read_unlock(&journal->j_state_lock);
 296                write_lock(&journal->j_state_lock);
 297                if (__jbd2_log_space_left(journal) < jbd_space_needed(journal))
 298                        __jbd2_log_wait_for_space(journal);
 299                write_unlock(&journal->j_state_lock);
 300                goto repeat;
 301        }
 302
 303        /* OK, account for the buffers that this operation expects to
 304         * use and add the handle to the running transaction. 
 305         */
 306        update_t_max_wait(transaction, ts);
 307        handle->h_transaction = transaction;
 308        atomic_inc(&transaction->t_updates);
 309        atomic_inc(&transaction->t_handle_count);
 310        jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
 311                  handle, nblocks,
 312                  atomic_read(&transaction->t_outstanding_credits),
 313                  __jbd2_log_space_left(journal));
 314        read_unlock(&journal->j_state_lock);
 315
 316        lock_map_acquire(&handle->h_lockdep_map);
 317        jbd2_journal_free_transaction(new_transaction);
 318        return 0;
 319}
 320
 321static struct lock_class_key jbd2_handle_key;
 322
 323/* Allocate a new handle.  This should probably be in a slab... */
 324static handle_t *new_handle(int nblocks)
 325{
 326        handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
 327        if (!handle)
 328                return NULL;
 329        memset(handle, 0, sizeof(*handle));
 330        handle->h_buffer_credits = nblocks;
 331        handle->h_ref = 1;
 332
 333        lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
 334                                                &jbd2_handle_key, 0);
 335
 336        return handle;
 337}
 338
 339/**
 340 * handle_t *jbd2_journal_start() - Obtain a new handle.
 341 * @journal: Journal to start transaction on.
 342 * @nblocks: number of block buffer we might modify
 343 *
 344 * We make sure that the transaction can guarantee at least nblocks of
 345 * modified buffers in the log.  We block until the log can guarantee
 346 * that much space.
 347 *
 348 * This function is visible to journal users (like ext3fs), so is not
 349 * called with the journal already locked.
 350 *
 351 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
 352 * on failure.
 353 */
 354handle_t *jbd2__journal_start(journal_t *journal, int nblocks, gfp_t gfp_mask)
 355{
 356        handle_t *handle = journal_current_handle();
 357        int err;
 358
 359        if (!journal)
 360                return ERR_PTR(-EROFS);
 361
 362        if (handle) {
 363                J_ASSERT(handle->h_transaction->t_journal == journal);
 364                handle->h_ref++;
 365                return handle;
 366        }
 367
 368        handle = new_handle(nblocks);
 369        if (!handle)
 370                return ERR_PTR(-ENOMEM);
 371
 372        current->journal_info = handle;
 373
 374        err = start_this_handle(journal, handle, gfp_mask);
 375        if (err < 0) {
 376                jbd2_free_handle(handle);
 377                current->journal_info = NULL;
 378                handle = ERR_PTR(err);
 379        }
 380        return handle;
 381}
 382EXPORT_SYMBOL(jbd2__journal_start);
 383
 384
 385handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
 386{
 387        return jbd2__journal_start(journal, nblocks, GFP_NOFS);
 388}
 389EXPORT_SYMBOL(jbd2_journal_start);
 390
 391
 392/**
 393 * int jbd2_journal_extend() - extend buffer credits.
 394 * @handle:  handle to 'extend'
 395 * @nblocks: nr blocks to try to extend by.
 396 *
 397 * Some transactions, such as large extends and truncates, can be done
 398 * atomically all at once or in several stages.  The operation requests
 399 * a credit for a number of buffer modications in advance, but can
 400 * extend its credit if it needs more.
 401 *
 402 * jbd2_journal_extend tries to give the running handle more buffer credits.
 403 * It does not guarantee that allocation - this is a best-effort only.
 404 * The calling process MUST be able to deal cleanly with a failure to
 405 * extend here.
 406 *
 407 * Return 0 on success, non-zero on failure.
 408 *
 409 * return code < 0 implies an error
 410 * return code > 0 implies normal transaction-full status.
 411 */
 412int jbd2_journal_extend(handle_t *handle, int nblocks)
 413{
 414        transaction_t *transaction = handle->h_transaction;
 415        journal_t *journal = transaction->t_journal;
 416        int result;
 417        int wanted;
 418
 419        result = -EIO;
 420        if (is_handle_aborted(handle))
 421                goto out;
 422
 423        result = 1;
 424
 425        read_lock(&journal->j_state_lock);
 426
 427        /* Don't extend a locked-down transaction! */
 428        if (handle->h_transaction->t_state != T_RUNNING) {
 429                jbd_debug(3, "denied handle %p %d blocks: "
 430                          "transaction not running\n", handle, nblocks);
 431                goto error_out;
 432        }
 433
 434        spin_lock(&transaction->t_handle_lock);
 435        wanted = atomic_read(&transaction->t_outstanding_credits) + nblocks;
 436
 437        if (wanted > journal->j_max_transaction_buffers) {
 438                jbd_debug(3, "denied handle %p %d blocks: "
 439                          "transaction too large\n", handle, nblocks);
 440                goto unlock;
 441        }
 442
 443        if (wanted > __jbd2_log_space_left(journal)) {
 444                jbd_debug(3, "denied handle %p %d blocks: "
 445                          "insufficient log space\n", handle, nblocks);
 446                goto unlock;
 447        }
 448
 449        handle->h_buffer_credits += nblocks;
 450        atomic_add(nblocks, &transaction->t_outstanding_credits);
 451        result = 0;
 452
 453        jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
 454unlock:
 455        spin_unlock(&transaction->t_handle_lock);
 456error_out:
 457        read_unlock(&journal->j_state_lock);
 458out:
 459        return result;
 460}
 461
 462
 463/**
 464 * int jbd2_journal_restart() - restart a handle .
 465 * @handle:  handle to restart
 466 * @nblocks: nr credits requested
 467 *
 468 * Restart a handle for a multi-transaction filesystem
 469 * operation.
 470 *
 471 * If the jbd2_journal_extend() call above fails to grant new buffer credits
 472 * to a running handle, a call to jbd2_journal_restart will commit the
 473 * handle's transaction so far and reattach the handle to a new
 474 * transaction capabable of guaranteeing the requested number of
 475 * credits.
 476 */
 477int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
 478{
 479        transaction_t *transaction = handle->h_transaction;
 480        journal_t *journal = transaction->t_journal;
 481        tid_t           tid;
 482        int             need_to_start, ret;
 483
 484        /* If we've had an abort of any type, don't even think about
 485         * actually doing the restart! */
 486        if (is_handle_aborted(handle))
 487                return 0;
 488
 489        /*
 490         * First unlink the handle from its current transaction, and start the
 491         * commit on that.
 492         */
 493        J_ASSERT(atomic_read(&transaction->t_updates) > 0);
 494        J_ASSERT(journal_current_handle() == handle);
 495
 496        read_lock(&journal->j_state_lock);
 497        spin_lock(&transaction->t_handle_lock);
 498        atomic_sub(handle->h_buffer_credits,
 499                   &transaction->t_outstanding_credits);
 500        if (atomic_dec_and_test(&transaction->t_updates))
 501                wake_up(&journal->j_wait_updates);
 502        spin_unlock(&transaction->t_handle_lock);
 503
 504        jbd_debug(2, "restarting handle %p\n", handle);
 505        tid = transaction->t_tid;
 506        need_to_start = !tid_geq(journal->j_commit_request, tid);
 507        read_unlock(&journal->j_state_lock);
 508        if (need_to_start)
 509                jbd2_log_start_commit(journal, tid);
 510
 511        lock_map_release(&handle->h_lockdep_map);
 512        handle->h_buffer_credits = nblocks;
 513        ret = start_this_handle(journal, handle, gfp_mask);
 514        return ret;
 515}
 516EXPORT_SYMBOL(jbd2__journal_restart);
 517
 518
 519int jbd2_journal_restart(handle_t *handle, int nblocks)
 520{
 521        return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
 522}
 523EXPORT_SYMBOL(jbd2_journal_restart);
 524
 525/**
 526 * void jbd2_journal_lock_updates () - establish a transaction barrier.
 527 * @journal:  Journal to establish a barrier on.
 528 *
 529 * This locks out any further updates from being started, and blocks
 530 * until all existing updates have completed, returning only once the
 531 * journal is in a quiescent state with no updates running.
 532 *
 533 * The journal lock should not be held on entry.
 534 */
 535void jbd2_journal_lock_updates(journal_t *journal)
 536{
 537        DEFINE_WAIT(wait);
 538
 539        write_lock(&journal->j_state_lock);
 540        ++journal->j_barrier_count;
 541
 542        /* Wait until there are no running updates */
 543        while (1) {
 544                transaction_t *transaction = journal->j_running_transaction;
 545
 546                if (!transaction)
 547                        break;
 548
 549                spin_lock(&transaction->t_handle_lock);
 550                prepare_to_wait(&journal->j_wait_updates, &wait,
 551                                TASK_UNINTERRUPTIBLE);
 552                if (!atomic_read(&transaction->t_updates)) {
 553                        spin_unlock(&transaction->t_handle_lock);
 554                        finish_wait(&journal->j_wait_updates, &wait);
 555                        break;
 556                }
 557                spin_unlock(&transaction->t_handle_lock);
 558                write_unlock(&journal->j_state_lock);
 559                schedule();
 560                finish_wait(&journal->j_wait_updates, &wait);
 561                write_lock(&journal->j_state_lock);
 562        }
 563        write_unlock(&journal->j_state_lock);
 564
 565        /*
 566         * We have now established a barrier against other normal updates, but
 567         * we also need to barrier against other jbd2_journal_lock_updates() calls
 568         * to make sure that we serialise special journal-locked operations
 569         * too.
 570         */
 571        mutex_lock(&journal->j_barrier);
 572}
 573
 574/**
 575 * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
 576 * @journal:  Journal to release the barrier on.
 577 *
 578 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
 579 *
 580 * Should be called without the journal lock held.
 581 */
 582void jbd2_journal_unlock_updates (journal_t *journal)
 583{
 584        J_ASSERT(journal->j_barrier_count != 0);
 585
 586        mutex_unlock(&journal->j_barrier);
 587        write_lock(&journal->j_state_lock);
 588        --journal->j_barrier_count;
 589        write_unlock(&journal->j_state_lock);
 590        wake_up(&journal->j_wait_transaction_locked);
 591}
 592
 593static void warn_dirty_buffer(struct buffer_head *bh)
 594{
 595        char b[BDEVNAME_SIZE];
 596
 597        printk(KERN_WARNING
 598               "JBD2: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
 599               "There's a risk of filesystem corruption in case of system "
 600               "crash.\n",
 601               bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
 602}
 603
 604/*
 605 * If the buffer is already part of the current transaction, then there
 606 * is nothing we need to do.  If it is already part of a prior
 607 * transaction which we are still committing to disk, then we need to
 608 * make sure that we do not overwrite the old copy: we do copy-out to
 609 * preserve the copy going to disk.  We also account the buffer against
 610 * the handle's metadata buffer credits (unless the buffer is already
 611 * part of the transaction, that is).
 612 *
 613 */
 614static int
 615do_get_write_access(handle_t *handle, struct journal_head *jh,
 616                        int force_copy)
 617{
 618        struct buffer_head *bh;
 619        transaction_t *transaction;
 620        journal_t *journal;
 621        int error;
 622        char *frozen_buffer = NULL;
 623        int need_copy = 0;
 624
 625        if (is_handle_aborted(handle))
 626                return -EROFS;
 627
 628        transaction = handle->h_transaction;
 629        journal = transaction->t_journal;
 630
 631        jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
 632
 633        JBUFFER_TRACE(jh, "entry");
 634repeat:
 635        bh = jh2bh(jh);
 636
 637        /* @@@ Need to check for errors here at some point. */
 638
 639        lock_buffer(bh);
 640        jbd_lock_bh_state(bh);
 641
 642        /* We now hold the buffer lock so it is safe to query the buffer
 643         * state.  Is the buffer dirty?
 644         *
 645         * If so, there are two possibilities.  The buffer may be
 646         * non-journaled, and undergoing a quite legitimate writeback.
 647         * Otherwise, it is journaled, and we don't expect dirty buffers
 648         * in that state (the buffers should be marked JBD_Dirty
 649         * instead.)  So either the IO is being done under our own
 650         * control and this is a bug, or it's a third party IO such as
 651         * dump(8) (which may leave the buffer scheduled for read ---
 652         * ie. locked but not dirty) or tune2fs (which may actually have
 653         * the buffer dirtied, ugh.)  */
 654
 655        if (buffer_dirty(bh)) {
 656                /*
 657                 * First question: is this buffer already part of the current
 658                 * transaction or the existing committing transaction?
 659                 */
 660                if (jh->b_transaction) {
 661                        J_ASSERT_JH(jh,
 662                                jh->b_transaction == transaction ||
 663                                jh->b_transaction ==
 664                                        journal->j_committing_transaction);
 665                        if (jh->b_next_transaction)
 666                                J_ASSERT_JH(jh, jh->b_next_transaction ==
 667                                                        transaction);
 668                        warn_dirty_buffer(bh);
 669                }
 670                /*
 671                 * In any case we need to clean the dirty flag and we must
 672                 * do it under the buffer lock to be sure we don't race
 673                 * with running write-out.
 674                 */
 675                JBUFFER_TRACE(jh, "Journalling dirty buffer");
 676                clear_buffer_dirty(bh);
 677                set_buffer_jbddirty(bh);
 678        }
 679
 680        unlock_buffer(bh);
 681
 682        error = -EROFS;
 683        if (is_handle_aborted(handle)) {
 684                jbd_unlock_bh_state(bh);
 685                goto out;
 686        }
 687        error = 0;
 688
 689        /*
 690         * The buffer is already part of this transaction if b_transaction or
 691         * b_next_transaction points to it
 692         */
 693        if (jh->b_transaction == transaction ||
 694            jh->b_next_transaction == transaction)
 695                goto done;
 696
 697        /*
 698         * this is the first time this transaction is touching this buffer,
 699         * reset the modified flag
 700         */
 701       jh->b_modified = 0;
 702
 703        /*
 704         * If there is already a copy-out version of this buffer, then we don't
 705         * need to make another one
 706         */
 707        if (jh->b_frozen_data) {
 708                JBUFFER_TRACE(jh, "has frozen data");
 709                J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 710                jh->b_next_transaction = transaction;
 711                goto done;
 712        }
 713
 714        /* Is there data here we need to preserve? */
 715
 716        if (jh->b_transaction && jh->b_transaction != transaction) {
 717                JBUFFER_TRACE(jh, "owned by older transaction");
 718                J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 719                J_ASSERT_JH(jh, jh->b_transaction ==
 720                                        journal->j_committing_transaction);
 721
 722                /* There is one case we have to be very careful about.
 723                 * If the committing transaction is currently writing
 724                 * this buffer out to disk and has NOT made a copy-out,
 725                 * then we cannot modify the buffer contents at all
 726                 * right now.  The essence of copy-out is that it is the
 727                 * extra copy, not the primary copy, which gets
 728                 * journaled.  If the primary copy is already going to
 729                 * disk then we cannot do copy-out here. */
 730
 731                if (jh->b_jlist == BJ_Shadow) {
 732                        DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
 733                        wait_queue_head_t *wqh;
 734
 735                        wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
 736
 737                        JBUFFER_TRACE(jh, "on shadow: sleep");
 738                        jbd_unlock_bh_state(bh);
 739                        /* commit wakes up all shadow buffers after IO */
 740                        for ( ; ; ) {
 741                                prepare_to_wait(wqh, &wait.wait,
 742                                                TASK_UNINTERRUPTIBLE);
 743                                if (jh->b_jlist != BJ_Shadow)
 744                                        break;
 745                                schedule();
 746                        }
 747                        finish_wait(wqh, &wait.wait);
 748                        goto repeat;
 749                }
 750
 751                /* Only do the copy if the currently-owning transaction
 752                 * still needs it.  If it is on the Forget list, the
 753                 * committing transaction is past that stage.  The
 754                 * buffer had better remain locked during the kmalloc,
 755                 * but that should be true --- we hold the journal lock
 756                 * still and the buffer is already on the BUF_JOURNAL
 757                 * list so won't be flushed.
 758                 *
 759                 * Subtle point, though: if this is a get_undo_access,
 760                 * then we will be relying on the frozen_data to contain
 761                 * the new value of the committed_data record after the
 762                 * transaction, so we HAVE to force the frozen_data copy
 763                 * in that case. */
 764
 765                if (jh->b_jlist != BJ_Forget || force_copy) {
 766                        JBUFFER_TRACE(jh, "generate frozen data");
 767                        if (!frozen_buffer) {
 768                                JBUFFER_TRACE(jh, "allocate memory for buffer");
 769                                jbd_unlock_bh_state(bh);
 770                                frozen_buffer =
 771                                        jbd2_alloc(jh2bh(jh)->b_size,
 772                                                         GFP_NOFS);
 773                                if (!frozen_buffer) {
 774                                        printk(KERN_EMERG
 775                                               "%s: OOM for frozen_buffer\n",
 776                                               __func__);
 777                                        JBUFFER_TRACE(jh, "oom!");
 778                                        error = -ENOMEM;
 779                                        jbd_lock_bh_state(bh);
 780                                        goto done;
 781                                }
 782                                goto repeat;
 783                        }
 784                        jh->b_frozen_data = frozen_buffer;
 785                        frozen_buffer = NULL;
 786                        need_copy = 1;
 787                }
 788                jh->b_next_transaction = transaction;
 789        }
 790
 791
 792        /*
 793         * Finally, if the buffer is not journaled right now, we need to make
 794         * sure it doesn't get written to disk before the caller actually
 795         * commits the new data
 796         */
 797        if (!jh->b_transaction) {
 798                JBUFFER_TRACE(jh, "no transaction");
 799                J_ASSERT_JH(jh, !jh->b_next_transaction);
 800                JBUFFER_TRACE(jh, "file as BJ_Reserved");
 801                spin_lock(&journal->j_list_lock);
 802                __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
 803                spin_unlock(&journal->j_list_lock);
 804        }
 805
 806done:
 807        if (need_copy) {
 808                struct page *page;
 809                int offset;
 810                char *source;
 811
 812                J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
 813                            "Possible IO failure.\n");
 814                page = jh2bh(jh)->b_page;
 815                offset = offset_in_page(jh2bh(jh)->b_data);
 816                source = kmap_atomic(page);
 817                /* Fire data frozen trigger just before we copy the data */
 818                jbd2_buffer_frozen_trigger(jh, source + offset,
 819                                           jh->b_triggers);
 820                memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
 821                kunmap_atomic(source);
 822
 823                /*
 824                 * Now that the frozen data is saved off, we need to store
 825                 * any matching triggers.
 826                 */
 827                jh->b_frozen_triggers = jh->b_triggers;
 828        }
 829        jbd_unlock_bh_state(bh);
 830
 831        /*
 832         * If we are about to journal a buffer, then any revoke pending on it is
 833         * no longer valid
 834         */
 835        jbd2_journal_cancel_revoke(handle, jh);
 836
 837out:
 838        if (unlikely(frozen_buffer))    /* It's usually NULL */
 839                jbd2_free(frozen_buffer, bh->b_size);
 840
 841        JBUFFER_TRACE(jh, "exit");
 842        return error;
 843}
 844
 845/**
 846 * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
 847 * @handle: transaction to add buffer modifications to
 848 * @bh:     bh to be used for metadata writes
 849 *
 850 * Returns an error code or 0 on success.
 851 *
 852 * In full data journalling mode the buffer may be of type BJ_AsyncData,
 853 * because we're write()ing a buffer which is also part of a shared mapping.
 854 */
 855
 856int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
 857{
 858        struct journal_head *jh = jbd2_journal_add_journal_head(bh);
 859        int rc;
 860
 861        /* We do not want to get caught playing with fields which the
 862         * log thread also manipulates.  Make sure that the buffer
 863         * completes any outstanding IO before proceeding. */
 864        rc = do_get_write_access(handle, jh, 0);
 865        jbd2_journal_put_journal_head(jh);
 866        return rc;
 867}
 868
 869
 870/*
 871 * When the user wants to journal a newly created buffer_head
 872 * (ie. getblk() returned a new buffer and we are going to populate it
 873 * manually rather than reading off disk), then we need to keep the
 874 * buffer_head locked until it has been completely filled with new
 875 * data.  In this case, we should be able to make the assertion that
 876 * the bh is not already part of an existing transaction.
 877 *
 878 * The buffer should already be locked by the caller by this point.
 879 * There is no lock ranking violation: it was a newly created,
 880 * unlocked buffer beforehand. */
 881
 882/**
 883 * int jbd2_journal_get_create_access () - notify intent to use newly created bh
 884 * @handle: transaction to new buffer to
 885 * @bh: new buffer.
 886 *
 887 * Call this if you create a new bh.
 888 */
 889int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
 890{
 891        transaction_t *transaction = handle->h_transaction;
 892        journal_t *journal = transaction->t_journal;
 893        struct journal_head *jh = jbd2_journal_add_journal_head(bh);
 894        int err;
 895
 896        jbd_debug(5, "journal_head %p\n", jh);
 897        err = -EROFS;
 898        if (is_handle_aborted(handle))
 899                goto out;
 900        err = 0;
 901
 902        JBUFFER_TRACE(jh, "entry");
 903        /*
 904         * The buffer may already belong to this transaction due to pre-zeroing
 905         * in the filesystem's new_block code.  It may also be on the previous,
 906         * committing transaction's lists, but it HAS to be in Forget state in
 907         * that case: the transaction must have deleted the buffer for it to be
 908         * reused here.
 909         */
 910        jbd_lock_bh_state(bh);
 911        spin_lock(&journal->j_list_lock);
 912        J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
 913                jh->b_transaction == NULL ||
 914                (jh->b_transaction == journal->j_committing_transaction &&
 915                          jh->b_jlist == BJ_Forget)));
 916
 917        J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 918        J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
 919
 920        if (jh->b_transaction == NULL) {
 921                /*
 922                 * Previous jbd2_journal_forget() could have left the buffer
 923                 * with jbddirty bit set because it was being committed. When
 924                 * the commit finished, we've filed the buffer for
 925                 * checkpointing and marked it dirty. Now we are reallocating
 926                 * the buffer so the transaction freeing it must have
 927                 * committed and so it's safe to clear the dirty bit.
 928                 */
 929                clear_buffer_dirty(jh2bh(jh));
 930                /* first access by this transaction */
 931                jh->b_modified = 0;
 932
 933                JBUFFER_TRACE(jh, "file as BJ_Reserved");
 934                __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
 935        } else if (jh->b_transaction == journal->j_committing_transaction) {
 936                /* first access by this transaction */
 937                jh->b_modified = 0;
 938
 939                JBUFFER_TRACE(jh, "set next transaction");
 940                jh->b_next_transaction = transaction;
 941        }
 942        spin_unlock(&journal->j_list_lock);
 943        jbd_unlock_bh_state(bh);
 944
 945        /*
 946         * akpm: I added this.  ext3_alloc_branch can pick up new indirect
 947         * blocks which contain freed but then revoked metadata.  We need
 948         * to cancel the revoke in case we end up freeing it yet again
 949         * and the reallocating as data - this would cause a second revoke,
 950         * which hits an assertion error.
 951         */
 952        JBUFFER_TRACE(jh, "cancelling revoke");
 953        jbd2_journal_cancel_revoke(handle, jh);
 954out:
 955        jbd2_journal_put_journal_head(jh);
 956        return err;
 957}
 958
 959/**
 960 * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
 961 *     non-rewindable consequences
 962 * @handle: transaction
 963 * @bh: buffer to undo
 964 *
 965 * Sometimes there is a need to distinguish between metadata which has
 966 * been committed to disk and that which has not.  The ext3fs code uses
 967 * this for freeing and allocating space, we have to make sure that we
 968 * do not reuse freed space until the deallocation has been committed,
 969 * since if we overwrote that space we would make the delete
 970 * un-rewindable in case of a crash.
 971 *
 972 * To deal with that, jbd2_journal_get_undo_access requests write access to a
 973 * buffer for parts of non-rewindable operations such as delete
 974 * operations on the bitmaps.  The journaling code must keep a copy of
 975 * the buffer's contents prior to the undo_access call until such time
 976 * as we know that the buffer has definitely been committed to disk.
 977 *
 978 * We never need to know which transaction the committed data is part
 979 * of, buffers touched here are guaranteed to be dirtied later and so
 980 * will be committed to a new transaction in due course, at which point
 981 * we can discard the old committed data pointer.
 982 *
 983 * Returns error number or 0 on success.
 984 */
 985int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
 986{
 987        int err;
 988        struct journal_head *jh = jbd2_journal_add_journal_head(bh);
 989        char *committed_data = NULL;
 990
 991        JBUFFER_TRACE(jh, "entry");
 992
 993        /*
 994         * Do this first --- it can drop the journal lock, so we want to
 995         * make sure that obtaining the committed_data is done
 996         * atomically wrt. completion of any outstanding commits.
 997         */
 998        err = do_get_write_access(handle, jh, 1);
 999        if (err)
1000                goto out;
1001
1002repeat:
1003        if (!jh->b_committed_data) {
1004                committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
1005                if (!committed_data) {
1006                        printk(KERN_EMERG "%s: No memory for committed data\n",
1007                                __func__);
1008                        err = -ENOMEM;
1009                        goto out;
1010                }
1011        }
1012
1013        jbd_lock_bh_state(bh);
1014        if (!jh->b_committed_data) {
1015                /* Copy out the current buffer contents into the
1016                 * preserved, committed copy. */
1017                JBUFFER_TRACE(jh, "generate b_committed data");
1018                if (!committed_data) {
1019                        jbd_unlock_bh_state(bh);
1020                        goto repeat;
1021                }
1022
1023                jh->b_committed_data = committed_data;
1024                committed_data = NULL;
1025                memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1026        }
1027        jbd_unlock_bh_state(bh);
1028out:
1029        jbd2_journal_put_journal_head(jh);
1030        if (unlikely(committed_data))
1031                jbd2_free(committed_data, bh->b_size);
1032        return err;
1033}
1034
1035/**
1036 * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1037 * @bh: buffer to trigger on
1038 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1039 *
1040 * Set any triggers on this journal_head.  This is always safe, because
1041 * triggers for a committing buffer will be saved off, and triggers for
1042 * a running transaction will match the buffer in that transaction.
1043 *
1044 * Call with NULL to clear the triggers.
1045 */
1046void jbd2_journal_set_triggers(struct buffer_head *bh,
1047                               struct jbd2_buffer_trigger_type *type)
1048{
1049        struct journal_head *jh = bh2jh(bh);
1050
1051        jh->b_triggers = type;
1052}
1053
1054void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1055                                struct jbd2_buffer_trigger_type *triggers)
1056{
1057        struct buffer_head *bh = jh2bh(jh);
1058
1059        if (!triggers || !triggers->t_frozen)
1060                return;
1061
1062        triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1063}
1064
1065void jbd2_buffer_abort_trigger(struct journal_head *jh,
1066                               struct jbd2_buffer_trigger_type *triggers)
1067{
1068        if (!triggers || !triggers->t_abort)
1069                return;
1070
1071        triggers->t_abort(triggers, jh2bh(jh));
1072}
1073
1074
1075
1076/**
1077 * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1078 * @handle: transaction to add buffer to.
1079 * @bh: buffer to mark
1080 *
1081 * mark dirty metadata which needs to be journaled as part of the current
1082 * transaction.
1083 *
1084 * The buffer must have previously had jbd2_journal_get_write_access()
1085 * called so that it has a valid journal_head attached to the buffer
1086 * head.
1087 *
1088 * The buffer is placed on the transaction's metadata list and is marked
1089 * as belonging to the transaction.
1090 *
1091 * Returns error number or 0 on success.
1092 *
1093 * Special care needs to be taken if the buffer already belongs to the
1094 * current committing transaction (in which case we should have frozen
1095 * data present for that commit).  In that case, we don't relink the
1096 * buffer: that only gets done when the old transaction finally
1097 * completes its commit.
1098 */
1099int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1100{
1101        transaction_t *transaction = handle->h_transaction;
1102        journal_t *journal = transaction->t_journal;
1103        struct journal_head *jh = bh2jh(bh);
1104        int ret = 0;
1105
1106        jbd_debug(5, "journal_head %p\n", jh);
1107        JBUFFER_TRACE(jh, "entry");
1108        if (is_handle_aborted(handle))
1109                goto out;
1110        if (!buffer_jbd(bh)) {
1111                ret = -EUCLEAN;
1112                goto out;
1113        }
1114
1115        jbd_lock_bh_state(bh);
1116
1117        if (jh->b_modified == 0) {
1118                /*
1119                 * This buffer's got modified and becoming part
1120                 * of the transaction. This needs to be done
1121                 * once a transaction -bzzz
1122                 */
1123                jh->b_modified = 1;
1124                J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1125                handle->h_buffer_credits--;
1126        }
1127
1128        /*
1129         * fastpath, to avoid expensive locking.  If this buffer is already
1130         * on the running transaction's metadata list there is nothing to do.
1131         * Nobody can take it off again because there is a handle open.
1132         * I _think_ we're OK here with SMP barriers - a mistaken decision will
1133         * result in this test being false, so we go in and take the locks.
1134         */
1135        if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1136                JBUFFER_TRACE(jh, "fastpath");
1137                if (unlikely(jh->b_transaction !=
1138                             journal->j_running_transaction)) {
1139                        printk(KERN_EMERG "JBD: %s: "
1140                               "jh->b_transaction (%llu, %p, %u) != "
1141                               "journal->j_running_transaction (%p, %u)",
1142                               journal->j_devname,
1143                               (unsigned long long) bh->b_blocknr,
1144                               jh->b_transaction,
1145                               jh->b_transaction ? jh->b_transaction->t_tid : 0,
1146                               journal->j_running_transaction,
1147                               journal->j_running_transaction ?
1148                               journal->j_running_transaction->t_tid : 0);
1149                        ret = -EINVAL;
1150                }
1151                goto out_unlock_bh;
1152        }
1153
1154        set_buffer_jbddirty(bh);
1155
1156        /*
1157         * Metadata already on the current transaction list doesn't
1158         * need to be filed.  Metadata on another transaction's list must
1159         * be committing, and will be refiled once the commit completes:
1160         * leave it alone for now.
1161         */
1162        if (jh->b_transaction != transaction) {
1163                JBUFFER_TRACE(jh, "already on other transaction");
1164                if (unlikely(jh->b_transaction !=
1165                             journal->j_committing_transaction)) {
1166                        printk(KERN_EMERG "JBD: %s: "
1167                               "jh->b_transaction (%llu, %p, %u) != "
1168                               "journal->j_committing_transaction (%p, %u)",
1169                               journal->j_devname,
1170                               (unsigned long long) bh->b_blocknr,
1171                               jh->b_transaction,
1172                               jh->b_transaction ? jh->b_transaction->t_tid : 0,
1173                               journal->j_committing_transaction,
1174                               journal->j_committing_transaction ?
1175                               journal->j_committing_transaction->t_tid : 0);
1176                        ret = -EINVAL;
1177                }
1178                if (unlikely(jh->b_next_transaction != transaction)) {
1179                        printk(KERN_EMERG "JBD: %s: "
1180                               "jh->b_next_transaction (%llu, %p, %u) != "
1181                               "transaction (%p, %u)",
1182                               journal->j_devname,
1183                               (unsigned long long) bh->b_blocknr,
1184                               jh->b_next_transaction,
1185                               jh->b_next_transaction ?
1186                               jh->b_next_transaction->t_tid : 0,
1187                               transaction, transaction->t_tid);
1188                        ret = -EINVAL;
1189                }
1190                /* And this case is illegal: we can't reuse another
1191                 * transaction's data buffer, ever. */
1192                goto out_unlock_bh;
1193        }
1194
1195        /* That test should have eliminated the following case: */
1196        J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1197
1198        JBUFFER_TRACE(jh, "file as BJ_Metadata");
1199        spin_lock(&journal->j_list_lock);
1200        __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1201        spin_unlock(&journal->j_list_lock);
1202out_unlock_bh:
1203        jbd_unlock_bh_state(bh);
1204out:
1205        JBUFFER_TRACE(jh, "exit");
1206        WARN_ON(ret);   /* All errors are bugs, so dump the stack */
1207        return ret;
1208}
1209
1210/*
1211 * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1212 * updates, if the update decided in the end that it didn't need access.
1213 *
1214 */
1215void
1216jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1217{
1218        BUFFER_TRACE(bh, "entry");
1219}
1220
1221/**
1222 * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1223 * @handle: transaction handle
1224 * @bh:     bh to 'forget'
1225 *
1226 * We can only do the bforget if there are no commits pending against the
1227 * buffer.  If the buffer is dirty in the current running transaction we
1228 * can safely unlink it.
1229 *
1230 * bh may not be a journalled buffer at all - it may be a non-JBD
1231 * buffer which came off the hashtable.  Check for this.
1232 *
1233 * Decrements bh->b_count by one.
1234 *
1235 * Allow this call even if the handle has aborted --- it may be part of
1236 * the caller's cleanup after an abort.
1237 */
1238int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1239{
1240        transaction_t *transaction = handle->h_transaction;
1241        journal_t *journal = transaction->t_journal;
1242        struct journal_head *jh;
1243        int drop_reserve = 0;
1244        int err = 0;
1245        int was_modified = 0;
1246
1247        BUFFER_TRACE(bh, "entry");
1248
1249        jbd_lock_bh_state(bh);
1250        spin_lock(&journal->j_list_lock);
1251
1252        if (!buffer_jbd(bh))
1253                goto not_jbd;
1254        jh = bh2jh(bh);
1255
1256        /* Critical error: attempting to delete a bitmap buffer, maybe?
1257         * Don't do any jbd operations, and return an error. */
1258        if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1259                         "inconsistent data on disk")) {
1260                err = -EIO;
1261                goto not_jbd;
1262        }
1263
1264        /* keep track of wether or not this transaction modified us */
1265        was_modified = jh->b_modified;
1266
1267        /*
1268         * The buffer's going from the transaction, we must drop
1269         * all references -bzzz
1270         */
1271        jh->b_modified = 0;
1272
1273        if (jh->b_transaction == handle->h_transaction) {
1274                J_ASSERT_JH(jh, !jh->b_frozen_data);
1275
1276                /* If we are forgetting a buffer which is already part
1277                 * of this transaction, then we can just drop it from
1278                 * the transaction immediately. */
1279                clear_buffer_dirty(bh);
1280                clear_buffer_jbddirty(bh);
1281
1282                JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1283
1284                /*
1285                 * we only want to drop a reference if this transaction
1286                 * modified the buffer
1287                 */
1288                if (was_modified)
1289                        drop_reserve = 1;
1290
1291                /*
1292                 * We are no longer going to journal this buffer.
1293                 * However, the commit of this transaction is still
1294                 * important to the buffer: the delete that we are now
1295                 * processing might obsolete an old log entry, so by
1296                 * committing, we can satisfy the buffer's checkpoint.
1297                 *
1298                 * So, if we have a checkpoint on the buffer, we should
1299                 * now refile the buffer on our BJ_Forget list so that
1300                 * we know to remove the checkpoint after we commit.
1301                 */
1302
1303                if (jh->b_cp_transaction) {
1304                        __jbd2_journal_temp_unlink_buffer(jh);
1305                        __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1306                } else {
1307                        __jbd2_journal_unfile_buffer(jh);
1308                        if (!buffer_jbd(bh)) {
1309                                spin_unlock(&journal->j_list_lock);
1310                                jbd_unlock_bh_state(bh);
1311                                __bforget(bh);
1312                                goto drop;
1313                        }
1314                }
1315        } else if (jh->b_transaction) {
1316                J_ASSERT_JH(jh, (jh->b_transaction ==
1317                                 journal->j_committing_transaction));
1318                /* However, if the buffer is still owned by a prior
1319                 * (committing) transaction, we can't drop it yet... */
1320                JBUFFER_TRACE(jh, "belongs to older transaction");
1321                /* ... but we CAN drop it from the new transaction if we
1322                 * have also modified it since the original commit. */
1323
1324                if (jh->b_next_transaction) {
1325                        J_ASSERT(jh->b_next_transaction == transaction);
1326                        jh->b_next_transaction = NULL;
1327
1328                        /*
1329                         * only drop a reference if this transaction modified
1330                         * the buffer
1331                         */
1332                        if (was_modified)
1333                                drop_reserve = 1;
1334                }
1335        }
1336
1337not_jbd:
1338        spin_unlock(&journal->j_list_lock);
1339        jbd_unlock_bh_state(bh);
1340        __brelse(bh);
1341drop:
1342        if (drop_reserve) {
1343                /* no need to reserve log space for this block -bzzz */
1344                handle->h_buffer_credits++;
1345        }
1346        return err;
1347}
1348
1349/**
1350 * int jbd2_journal_stop() - complete a transaction
1351 * @handle: tranaction to complete.
1352 *
1353 * All done for a particular handle.
1354 *
1355 * There is not much action needed here.  We just return any remaining
1356 * buffer credits to the transaction and remove the handle.  The only
1357 * complication is that we need to start a commit operation if the
1358 * filesystem is marked for synchronous update.
1359 *
1360 * jbd2_journal_stop itself will not usually return an error, but it may
1361 * do so in unusual circumstances.  In particular, expect it to
1362 * return -EIO if a jbd2_journal_abort has been executed since the
1363 * transaction began.
1364 */
1365int jbd2_journal_stop(handle_t *handle)
1366{
1367        transaction_t *transaction = handle->h_transaction;
1368        journal_t *journal = transaction->t_journal;
1369        int err, wait_for_commit = 0;
1370        tid_t tid;
1371        pid_t pid;
1372
1373        J_ASSERT(journal_current_handle() == handle);
1374
1375        if (is_handle_aborted(handle))
1376                err = -EIO;
1377        else {
1378                J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1379                err = 0;
1380        }
1381
1382        if (--handle->h_ref > 0) {
1383                jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1384                          handle->h_ref);
1385                return err;
1386        }
1387
1388        jbd_debug(4, "Handle %p going down\n", handle);
1389
1390        /*
1391         * Implement synchronous transaction batching.  If the handle
1392         * was synchronous, don't force a commit immediately.  Let's
1393         * yield and let another thread piggyback onto this
1394         * transaction.  Keep doing that while new threads continue to
1395         * arrive.  It doesn't cost much - we're about to run a commit
1396         * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1397         * operations by 30x or more...
1398         *
1399         * We try and optimize the sleep time against what the
1400         * underlying disk can do, instead of having a static sleep
1401         * time.  This is useful for the case where our storage is so
1402         * fast that it is more optimal to go ahead and force a flush
1403         * and wait for the transaction to be committed than it is to
1404         * wait for an arbitrary amount of time for new writers to
1405         * join the transaction.  We achieve this by measuring how
1406         * long it takes to commit a transaction, and compare it with
1407         * how long this transaction has been running, and if run time
1408         * < commit time then we sleep for the delta and commit.  This
1409         * greatly helps super fast disks that would see slowdowns as
1410         * more threads started doing fsyncs.
1411         *
1412         * But don't do this if this process was the most recent one
1413         * to perform a synchronous write.  We do this to detect the
1414         * case where a single process is doing a stream of sync
1415         * writes.  No point in waiting for joiners in that case.
1416         */
1417        pid = current->pid;
1418        if (handle->h_sync && journal->j_last_sync_writer != pid) {
1419                u64 commit_time, trans_time;
1420
1421                journal->j_last_sync_writer = pid;
1422
1423                read_lock(&journal->j_state_lock);
1424                commit_time = journal->j_average_commit_time;
1425                read_unlock(&journal->j_state_lock);
1426
1427                trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1428                                                   transaction->t_start_time));
1429
1430                commit_time = max_t(u64, commit_time,
1431                                    1000*journal->j_min_batch_time);
1432                commit_time = min_t(u64, commit_time,
1433                                    1000*journal->j_max_batch_time);
1434
1435                if (trans_time < commit_time) {
1436                        ktime_t expires = ktime_add_ns(ktime_get(),
1437                                                       commit_time);
1438                        set_current_state(TASK_UNINTERRUPTIBLE);
1439                        schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1440                }
1441        }
1442
1443        if (handle->h_sync)
1444                transaction->t_synchronous_commit = 1;
1445        current->journal_info = NULL;
1446        atomic_sub(handle->h_buffer_credits,
1447                   &transaction->t_outstanding_credits);
1448
1449        /*
1450         * If the handle is marked SYNC, we need to set another commit
1451         * going!  We also want to force a commit if the current
1452         * transaction is occupying too much of the log, or if the
1453         * transaction is too old now.
1454         */
1455        if (handle->h_sync ||
1456            (atomic_read(&transaction->t_outstanding_credits) >
1457             journal->j_max_transaction_buffers) ||
1458            time_after_eq(jiffies, transaction->t_expires)) {
1459                /* Do this even for aborted journals: an abort still
1460                 * completes the commit thread, it just doesn't write
1461                 * anything to disk. */
1462
1463                jbd_debug(2, "transaction too old, requesting commit for "
1464                                        "handle %p\n", handle);
1465                /* This is non-blocking */
1466                jbd2_log_start_commit(journal, transaction->t_tid);
1467
1468                /*
1469                 * Special case: JBD2_SYNC synchronous updates require us
1470                 * to wait for the commit to complete.
1471                 */
1472                if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1473                        wait_for_commit = 1;
1474        }
1475
1476        /*
1477         * Once we drop t_updates, if it goes to zero the transaction
1478         * could start committing on us and eventually disappear.  So
1479         * once we do this, we must not dereference transaction
1480         * pointer again.
1481         */
1482        tid = transaction->t_tid;
1483        if (atomic_dec_and_test(&transaction->t_updates)) {
1484                wake_up(&journal->j_wait_updates);
1485                if (journal->j_barrier_count)
1486                        wake_up(&journal->j_wait_transaction_locked);
1487        }
1488
1489        if (wait_for_commit)
1490                err = jbd2_log_wait_commit(journal, tid);
1491
1492        lock_map_release(&handle->h_lockdep_map);
1493
1494        jbd2_free_handle(handle);
1495        return err;
1496}
1497
1498/**
1499 * int jbd2_journal_force_commit() - force any uncommitted transactions
1500 * @journal: journal to force
1501 *
1502 * For synchronous operations: force any uncommitted transactions
1503 * to disk.  May seem kludgy, but it reuses all the handle batching
1504 * code in a very simple manner.
1505 */
1506int jbd2_journal_force_commit(journal_t *journal)
1507{
1508        handle_t *handle;
1509        int ret;
1510
1511        handle = jbd2_journal_start(journal, 1);
1512        if (IS_ERR(handle)) {
1513                ret = PTR_ERR(handle);
1514        } else {
1515                handle->h_sync = 1;
1516                ret = jbd2_journal_stop(handle);
1517        }
1518        return ret;
1519}
1520
1521/*
1522 *
1523 * List management code snippets: various functions for manipulating the
1524 * transaction buffer lists.
1525 *
1526 */
1527
1528/*
1529 * Append a buffer to a transaction list, given the transaction's list head
1530 * pointer.
1531 *
1532 * j_list_lock is held.
1533 *
1534 * jbd_lock_bh_state(jh2bh(jh)) is held.
1535 */
1536
1537static inline void
1538__blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1539{
1540        if (!*list) {
1541                jh->b_tnext = jh->b_tprev = jh;
1542                *list = jh;
1543        } else {
1544                /* Insert at the tail of the list to preserve order */
1545                struct journal_head *first = *list, *last = first->b_tprev;
1546                jh->b_tprev = last;
1547                jh->b_tnext = first;
1548                last->b_tnext = first->b_tprev = jh;
1549        }
1550}
1551
1552/*
1553 * Remove a buffer from a transaction list, given the transaction's list
1554 * head pointer.
1555 *
1556 * Called with j_list_lock held, and the journal may not be locked.
1557 *
1558 * jbd_lock_bh_state(jh2bh(jh)) is held.
1559 */
1560
1561static inline void
1562__blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1563{
1564        if (*list == jh) {
1565                *list = jh->b_tnext;
1566                if (*list == jh)
1567                        *list = NULL;
1568        }
1569        jh->b_tprev->b_tnext = jh->b_tnext;
1570        jh->b_tnext->b_tprev = jh->b_tprev;
1571}
1572
1573/*
1574 * Remove a buffer from the appropriate transaction list.
1575 *
1576 * Note that this function can *change* the value of
1577 * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1578 * t_log_list or t_reserved_list.  If the caller is holding onto a copy of one
1579 * of these pointers, it could go bad.  Generally the caller needs to re-read
1580 * the pointer from the transaction_t.
1581 *
1582 * Called under j_list_lock.
1583 */
1584static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1585{
1586        struct journal_head **list = NULL;
1587        transaction_t *transaction;
1588        struct buffer_head *bh = jh2bh(jh);
1589
1590        J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1591        transaction = jh->b_transaction;
1592        if (transaction)
1593                assert_spin_locked(&transaction->t_journal->j_list_lock);
1594
1595        J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1596        if (jh->b_jlist != BJ_None)
1597                J_ASSERT_JH(jh, transaction != NULL);
1598
1599        switch (jh->b_jlist) {
1600        case BJ_None:
1601                return;
1602        case BJ_Metadata:
1603                transaction->t_nr_buffers--;
1604                J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1605                list = &transaction->t_buffers;
1606                break;
1607        case BJ_Forget:
1608                list = &transaction->t_forget;
1609                break;
1610        case BJ_IO:
1611                list = &transaction->t_iobuf_list;
1612                break;
1613        case BJ_Shadow:
1614                list = &transaction->t_shadow_list;
1615                break;
1616        case BJ_LogCtl:
1617                list = &transaction->t_log_list;
1618                break;
1619        case BJ_Reserved:
1620                list = &transaction->t_reserved_list;
1621                break;
1622        }
1623
1624        __blist_del_buffer(list, jh);
1625        jh->b_jlist = BJ_None;
1626        if (test_clear_buffer_jbddirty(bh))
1627                mark_buffer_dirty(bh);  /* Expose it to the VM */
1628}
1629
1630/*
1631 * Remove buffer from all transactions.
1632 *
1633 * Called with bh_state lock and j_list_lock
1634 *
1635 * jh and bh may be already freed when this function returns.
1636 */
1637static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1638{
1639        __jbd2_journal_temp_unlink_buffer(jh);
1640        jh->b_transaction = NULL;
1641        jbd2_journal_put_journal_head(jh);
1642}
1643
1644void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1645{
1646        struct buffer_head *bh = jh2bh(jh);
1647
1648        /* Get reference so that buffer cannot be freed before we unlock it */
1649        get_bh(bh);
1650        jbd_lock_bh_state(bh);
1651        spin_lock(&journal->j_list_lock);
1652        __jbd2_journal_unfile_buffer(jh);
1653        spin_unlock(&journal->j_list_lock);
1654        jbd_unlock_bh_state(bh);
1655        __brelse(bh);
1656}
1657
1658/*
1659 * Called from jbd2_journal_try_to_free_buffers().
1660 *
1661 * Called under jbd_lock_bh_state(bh)
1662 */
1663static void
1664__journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1665{
1666        struct journal_head *jh;
1667
1668        jh = bh2jh(bh);
1669
1670        if (buffer_locked(bh) || buffer_dirty(bh))
1671                goto out;
1672
1673        if (jh->b_next_transaction != NULL)
1674                goto out;
1675
1676        spin_lock(&journal->j_list_lock);
1677        if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1678                /* written-back checkpointed metadata buffer */
1679                JBUFFER_TRACE(jh, "remove from checkpoint list");
1680                __jbd2_journal_remove_checkpoint(jh);
1681        }
1682        spin_unlock(&journal->j_list_lock);
1683out:
1684        return;
1685}
1686
1687/**
1688 * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1689 * @journal: journal for operation
1690 * @page: to try and free
1691 * @gfp_mask: we use the mask to detect how hard should we try to release
1692 * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1693 * release the buffers.
1694 *
1695 *
1696 * For all the buffers on this page,
1697 * if they are fully written out ordered data, move them onto BUF_CLEAN
1698 * so try_to_free_buffers() can reap them.
1699 *
1700 * This function returns non-zero if we wish try_to_free_buffers()
1701 * to be called. We do this if the page is releasable by try_to_free_buffers().
1702 * We also do it if the page has locked or dirty buffers and the caller wants
1703 * us to perform sync or async writeout.
1704 *
1705 * This complicates JBD locking somewhat.  We aren't protected by the
1706 * BKL here.  We wish to remove the buffer from its committing or
1707 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1708 *
1709 * This may *change* the value of transaction_t->t_datalist, so anyone
1710 * who looks at t_datalist needs to lock against this function.
1711 *
1712 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1713 * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1714 * will come out of the lock with the buffer dirty, which makes it
1715 * ineligible for release here.
1716 *
1717 * Who else is affected by this?  hmm...  Really the only contender
1718 * is do_get_write_access() - it could be looking at the buffer while
1719 * journal_try_to_free_buffer() is changing its state.  But that
1720 * cannot happen because we never reallocate freed data as metadata
1721 * while the data is part of a transaction.  Yes?
1722 *
1723 * Return 0 on failure, 1 on success
1724 */
1725int jbd2_journal_try_to_free_buffers(journal_t *journal,
1726                                struct page *page, gfp_t gfp_mask)
1727{
1728        struct buffer_head *head;
1729        struct buffer_head *bh;
1730        int ret = 0;
1731
1732        J_ASSERT(PageLocked(page));
1733
1734        head = page_buffers(page);
1735        bh = head;
1736        do {
1737                struct journal_head *jh;
1738
1739                /*
1740                 * We take our own ref against the journal_head here to avoid
1741                 * having to add tons of locking around each instance of
1742                 * jbd2_journal_put_journal_head().
1743                 */
1744                jh = jbd2_journal_grab_journal_head(bh);
1745                if (!jh)
1746                        continue;
1747
1748                jbd_lock_bh_state(bh);
1749                __journal_try_to_free_buffer(journal, bh);
1750                jbd2_journal_put_journal_head(jh);
1751                jbd_unlock_bh_state(bh);
1752                if (buffer_jbd(bh))
1753                        goto busy;
1754        } while ((bh = bh->b_this_page) != head);
1755
1756        ret = try_to_free_buffers(page);
1757
1758busy:
1759        return ret;
1760}
1761
1762/*
1763 * This buffer is no longer needed.  If it is on an older transaction's
1764 * checkpoint list we need to record it on this transaction's forget list
1765 * to pin this buffer (and hence its checkpointing transaction) down until
1766 * this transaction commits.  If the buffer isn't on a checkpoint list, we
1767 * release it.
1768 * Returns non-zero if JBD no longer has an interest in the buffer.
1769 *
1770 * Called under j_list_lock.
1771 *
1772 * Called under jbd_lock_bh_state(bh).
1773 */
1774static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1775{
1776        int may_free = 1;
1777        struct buffer_head *bh = jh2bh(jh);
1778
1779        if (jh->b_cp_transaction) {
1780                JBUFFER_TRACE(jh, "on running+cp transaction");
1781                __jbd2_journal_temp_unlink_buffer(jh);
1782                /*
1783                 * We don't want to write the buffer anymore, clear the
1784                 * bit so that we don't confuse checks in
1785                 * __journal_file_buffer
1786                 */
1787                clear_buffer_dirty(bh);
1788                __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1789                may_free = 0;
1790        } else {
1791                JBUFFER_TRACE(jh, "on running transaction");
1792                __jbd2_journal_unfile_buffer(jh);
1793        }
1794        return may_free;
1795}
1796
1797/*
1798 * jbd2_journal_invalidatepage
1799 *
1800 * This code is tricky.  It has a number of cases to deal with.
1801 *
1802 * There are two invariants which this code relies on:
1803 *
1804 * i_size must be updated on disk before we start calling invalidatepage on the
1805 * data.
1806 *
1807 *  This is done in ext3 by defining an ext3_setattr method which
1808 *  updates i_size before truncate gets going.  By maintaining this
1809 *  invariant, we can be sure that it is safe to throw away any buffers
1810 *  attached to the current transaction: once the transaction commits,
1811 *  we know that the data will not be needed.
1812 *
1813 *  Note however that we can *not* throw away data belonging to the
1814 *  previous, committing transaction!
1815 *
1816 * Any disk blocks which *are* part of the previous, committing
1817 * transaction (and which therefore cannot be discarded immediately) are
1818 * not going to be reused in the new running transaction
1819 *
1820 *  The bitmap committed_data images guarantee this: any block which is
1821 *  allocated in one transaction and removed in the next will be marked
1822 *  as in-use in the committed_data bitmap, so cannot be reused until
1823 *  the next transaction to delete the block commits.  This means that
1824 *  leaving committing buffers dirty is quite safe: the disk blocks
1825 *  cannot be reallocated to a different file and so buffer aliasing is
1826 *  not possible.
1827 *
1828 *
1829 * The above applies mainly to ordered data mode.  In writeback mode we
1830 * don't make guarantees about the order in which data hits disk --- in
1831 * particular we don't guarantee that new dirty data is flushed before
1832 * transaction commit --- so it is always safe just to discard data
1833 * immediately in that mode.  --sct
1834 */
1835
1836/*
1837 * The journal_unmap_buffer helper function returns zero if the buffer
1838 * concerned remains pinned as an anonymous buffer belonging to an older
1839 * transaction.
1840 *
1841 * We're outside-transaction here.  Either or both of j_running_transaction
1842 * and j_committing_transaction may be NULL.
1843 */
1844static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1845{
1846        transaction_t *transaction;
1847        struct journal_head *jh;
1848        int may_free = 1;
1849        int ret;
1850
1851        BUFFER_TRACE(bh, "entry");
1852
1853        /*
1854         * It is safe to proceed here without the j_list_lock because the
1855         * buffers cannot be stolen by try_to_free_buffers as long as we are
1856         * holding the page lock. --sct
1857         */
1858
1859        if (!buffer_jbd(bh))
1860                goto zap_buffer_unlocked;
1861
1862        /* OK, we have data buffer in journaled mode */
1863        write_lock(&journal->j_state_lock);
1864        jbd_lock_bh_state(bh);
1865        spin_lock(&journal->j_list_lock);
1866
1867        jh = jbd2_journal_grab_journal_head(bh);
1868        if (!jh)
1869                goto zap_buffer_no_jh;
1870
1871        /*
1872         * We cannot remove the buffer from checkpoint lists until the
1873         * transaction adding inode to orphan list (let's call it T)
1874         * is committed.  Otherwise if the transaction changing the
1875         * buffer would be cleaned from the journal before T is
1876         * committed, a crash will cause that the correct contents of
1877         * the buffer will be lost.  On the other hand we have to
1878         * clear the buffer dirty bit at latest at the moment when the
1879         * transaction marking the buffer as freed in the filesystem
1880         * structures is committed because from that moment on the
1881         * buffer can be reallocated and used by a different page.
1882         * Since the block hasn't been freed yet but the inode has
1883         * already been added to orphan list, it is safe for us to add
1884         * the buffer to BJ_Forget list of the newest transaction.
1885         */
1886        transaction = jh->b_transaction;
1887        if (transaction == NULL) {
1888                /* First case: not on any transaction.  If it
1889                 * has no checkpoint link, then we can zap it:
1890                 * it's a writeback-mode buffer so we don't care
1891                 * if it hits disk safely. */
1892                if (!jh->b_cp_transaction) {
1893                        JBUFFER_TRACE(jh, "not on any transaction: zap");
1894                        goto zap_buffer;
1895                }
1896
1897                if (!buffer_dirty(bh)) {
1898                        /* bdflush has written it.  We can drop it now */
1899                        goto zap_buffer;
1900                }
1901
1902                /* OK, it must be in the journal but still not
1903                 * written fully to disk: it's metadata or
1904                 * journaled data... */
1905
1906                if (journal->j_running_transaction) {
1907                        /* ... and once the current transaction has
1908                         * committed, the buffer won't be needed any
1909                         * longer. */
1910                        JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1911                        ret = __dispose_buffer(jh,
1912                                        journal->j_running_transaction);
1913                        jbd2_journal_put_journal_head(jh);
1914                        spin_unlock(&journal->j_list_lock);
1915                        jbd_unlock_bh_state(bh);
1916                        write_unlock(&journal->j_state_lock);
1917                        return ret;
1918                } else {
1919                        /* There is no currently-running transaction. So the
1920                         * orphan record which we wrote for this file must have
1921                         * passed into commit.  We must attach this buffer to
1922                         * the committing transaction, if it exists. */
1923                        if (journal->j_committing_transaction) {
1924                                JBUFFER_TRACE(jh, "give to committing trans");
1925                                ret = __dispose_buffer(jh,
1926                                        journal->j_committing_transaction);
1927                                jbd2_journal_put_journal_head(jh);
1928                                spin_unlock(&journal->j_list_lock);
1929                                jbd_unlock_bh_state(bh);
1930                                write_unlock(&journal->j_state_lock);
1931                                return ret;
1932                        } else {
1933                                /* The orphan record's transaction has
1934                                 * committed.  We can cleanse this buffer */
1935                                clear_buffer_jbddirty(bh);
1936                                goto zap_buffer;
1937                        }
1938                }
1939        } else if (transaction == journal->j_committing_transaction) {
1940                JBUFFER_TRACE(jh, "on committing transaction");
1941                /*
1942                 * The buffer is committing, we simply cannot touch
1943                 * it. So we just set j_next_transaction to the
1944                 * running transaction (if there is one) and mark
1945                 * buffer as freed so that commit code knows it should
1946                 * clear dirty bits when it is done with the buffer.
1947                 */
1948                set_buffer_freed(bh);
1949                if (journal->j_running_transaction && buffer_jbddirty(bh))
1950                        jh->b_next_transaction = journal->j_running_transaction;
1951                jbd2_journal_put_journal_head(jh);
1952                spin_unlock(&journal->j_list_lock);
1953                jbd_unlock_bh_state(bh);
1954                write_unlock(&journal->j_state_lock);
1955                return 0;
1956        } else {
1957                /* Good, the buffer belongs to the running transaction.
1958                 * We are writing our own transaction's data, not any
1959                 * previous one's, so it is safe to throw it away
1960                 * (remember that we expect the filesystem to have set
1961                 * i_size already for this truncate so recovery will not
1962                 * expose the disk blocks we are discarding here.) */
1963                J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1964                JBUFFER_TRACE(jh, "on running transaction");
1965                may_free = __dispose_buffer(jh, transaction);
1966        }
1967
1968zap_buffer:
1969        jbd2_journal_put_journal_head(jh);
1970zap_buffer_no_jh:
1971        spin_unlock(&journal->j_list_lock);
1972        jbd_unlock_bh_state(bh);
1973        write_unlock(&journal->j_state_lock);
1974zap_buffer_unlocked:
1975        clear_buffer_dirty(bh);
1976        J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1977        clear_buffer_mapped(bh);
1978        clear_buffer_req(bh);
1979        clear_buffer_new(bh);
1980        clear_buffer_delay(bh);
1981        clear_buffer_unwritten(bh);
1982        bh->b_bdev = NULL;
1983        return may_free;
1984}
1985
1986/**
1987 * void jbd2_journal_invalidatepage()
1988 * @journal: journal to use for flush...
1989 * @page:    page to flush
1990 * @offset:  length of page to invalidate.
1991 *
1992 * Reap page buffers containing data after offset in page.
1993 *
1994 */
1995void jbd2_journal_invalidatepage(journal_t *journal,
1996                      struct page *page,
1997                      unsigned long offset)
1998{
1999        struct buffer_head *head, *bh, *next;
2000        unsigned int curr_off = 0;
2001        int may_free = 1;
2002
2003        if (!PageLocked(page))
2004                BUG();
2005        if (!page_has_buffers(page))
2006                return;
2007
2008        /* We will potentially be playing with lists other than just the
2009         * data lists (especially for journaled data mode), so be
2010         * cautious in our locking. */
2011
2012        head = bh = page_buffers(page);
2013        do {
2014                unsigned int next_off = curr_off + bh->b_size;
2015                next = bh->b_this_page;
2016
2017                if (offset <= curr_off) {
2018                        /* This block is wholly outside the truncation point */
2019                        lock_buffer(bh);
2020                        may_free &= journal_unmap_buffer(journal, bh);
2021                        unlock_buffer(bh);
2022                }
2023                curr_off = next_off;
2024                bh = next;
2025
2026        } while (bh != head);
2027
2028        if (!offset) {
2029                if (may_free && try_to_free_buffers(page))
2030                        J_ASSERT(!page_has_buffers(page));
2031        }
2032}
2033
2034/*
2035 * File a buffer on the given transaction list.
2036 */
2037void __jbd2_journal_file_buffer(struct journal_head *jh,
2038                        transaction_t *transaction, int jlist)
2039{
2040        struct journal_head **list = NULL;
2041        int was_dirty = 0;
2042        struct buffer_head *bh = jh2bh(jh);
2043
2044        J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2045        assert_spin_locked(&transaction->t_journal->j_list_lock);
2046
2047        J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2048        J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2049                                jh->b_transaction == NULL);
2050
2051        if (jh->b_transaction && jh->b_jlist == jlist)
2052                return;
2053
2054        if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2055            jlist == BJ_Shadow || jlist == BJ_Forget) {
2056                /*
2057                 * For metadata buffers, we track dirty bit in buffer_jbddirty
2058                 * instead of buffer_dirty. We should not see a dirty bit set
2059                 * here because we clear it in do_get_write_access but e.g.
2060                 * tune2fs can modify the sb and set the dirty bit at any time
2061                 * so we try to gracefully handle that.
2062                 */
2063                if (buffer_dirty(bh))
2064                        warn_dirty_buffer(bh);
2065                if (test_clear_buffer_dirty(bh) ||
2066                    test_clear_buffer_jbddirty(bh))
2067                        was_dirty = 1;
2068        }
2069
2070        if (jh->b_transaction)
2071                __jbd2_journal_temp_unlink_buffer(jh);
2072        else
2073                jbd2_journal_grab_journal_head(bh);
2074        jh->b_transaction = transaction;
2075
2076        switch (jlist) {
2077        case BJ_None:
2078                J_ASSERT_JH(jh, !jh->b_committed_data);
2079                J_ASSERT_JH(jh, !jh->b_frozen_data);
2080                return;
2081        case BJ_Metadata:
2082                transaction->t_nr_buffers++;
2083                list = &transaction->t_buffers;
2084                break;
2085        case BJ_Forget:
2086                list = &transaction->t_forget;
2087                break;
2088        case BJ_IO:
2089                list = &transaction->t_iobuf_list;
2090                break;
2091        case BJ_Shadow:
2092                list = &transaction->t_shadow_list;
2093                break;
2094        case BJ_LogCtl:
2095                list = &transaction->t_log_list;
2096                break;
2097        case BJ_Reserved:
2098                list = &transaction->t_reserved_list;
2099                break;
2100        }
2101
2102        __blist_add_buffer(list, jh);
2103        jh->b_jlist = jlist;
2104
2105        if (was_dirty)
2106                set_buffer_jbddirty(bh);
2107}
2108
2109void jbd2_journal_file_buffer(struct journal_head *jh,
2110                                transaction_t *transaction, int jlist)
2111{
2112        jbd_lock_bh_state(jh2bh(jh));
2113        spin_lock(&transaction->t_journal->j_list_lock);
2114        __jbd2_journal_file_buffer(jh, transaction, jlist);
2115        spin_unlock(&transaction->t_journal->j_list_lock);
2116        jbd_unlock_bh_state(jh2bh(jh));
2117}
2118
2119/*
2120 * Remove a buffer from its current buffer list in preparation for
2121 * dropping it from its current transaction entirely.  If the buffer has
2122 * already started to be used by a subsequent transaction, refile the
2123 * buffer on that transaction's metadata list.
2124 *
2125 * Called under j_list_lock
2126 * Called under jbd_lock_bh_state(jh2bh(jh))
2127 *
2128 * jh and bh may be already free when this function returns
2129 */
2130void __jbd2_journal_refile_buffer(struct journal_head *jh)
2131{
2132        int was_dirty, jlist;
2133        struct buffer_head *bh = jh2bh(jh);
2134
2135        J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2136        if (jh->b_transaction)
2137                assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2138
2139        /* If the buffer is now unused, just drop it. */
2140        if (jh->b_next_transaction == NULL) {
2141                __jbd2_journal_unfile_buffer(jh);
2142                return;
2143        }
2144
2145        /*
2146         * It has been modified by a later transaction: add it to the new
2147         * transaction's metadata list.
2148         */
2149
2150        was_dirty = test_clear_buffer_jbddirty(bh);
2151        __jbd2_journal_temp_unlink_buffer(jh);
2152        /*
2153         * We set b_transaction here because b_next_transaction will inherit
2154         * our jh reference and thus __jbd2_journal_file_buffer() must not
2155         * take a new one.
2156         */
2157        jh->b_transaction = jh->b_next_transaction;
2158        jh->b_next_transaction = NULL;
2159        if (buffer_freed(bh))
2160                jlist = BJ_Forget;
2161        else if (jh->b_modified)
2162                jlist = BJ_Metadata;
2163        else
2164                jlist = BJ_Reserved;
2165        __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2166        J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2167
2168        if (was_dirty)
2169                set_buffer_jbddirty(bh);
2170}
2171
2172/*
2173 * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2174 * bh reference so that we can safely unlock bh.
2175 *
2176 * The jh and bh may be freed by this call.
2177 */
2178void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2179{
2180        struct buffer_head *bh = jh2bh(jh);
2181
2182        /* Get reference so that buffer cannot be freed before we unlock it */
2183        get_bh(bh);
2184        jbd_lock_bh_state(bh);
2185        spin_lock(&journal->j_list_lock);
2186        __jbd2_journal_refile_buffer(jh);
2187        jbd_unlock_bh_state(bh);
2188        spin_unlock(&journal->j_list_lock);
2189        __brelse(bh);
2190}
2191
2192/*
2193 * File inode in the inode list of the handle's transaction
2194 */
2195int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2196{
2197        transaction_t *transaction = handle->h_transaction;
2198        journal_t *journal = transaction->t_journal;
2199
2200        if (is_handle_aborted(handle))
2201                return -EIO;
2202
2203        jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2204                        transaction->t_tid);
2205
2206        /*
2207         * First check whether inode isn't already on the transaction's
2208         * lists without taking the lock. Note that this check is safe
2209         * without the lock as we cannot race with somebody removing inode
2210         * from the transaction. The reason is that we remove inode from the
2211         * transaction only in journal_release_jbd_inode() and when we commit
2212         * the transaction. We are guarded from the first case by holding
2213         * a reference to the inode. We are safe against the second case
2214         * because if jinode->i_transaction == transaction, commit code
2215         * cannot touch the transaction because we hold reference to it,
2216         * and if jinode->i_next_transaction == transaction, commit code
2217         * will only file the inode where we want it.
2218         */
2219        if (jinode->i_transaction == transaction ||
2220            jinode->i_next_transaction == transaction)
2221                return 0;
2222
2223        spin_lock(&journal->j_list_lock);
2224
2225        if (jinode->i_transaction == transaction ||
2226            jinode->i_next_transaction == transaction)
2227                goto done;
2228
2229        /*
2230         * We only ever set this variable to 1 so the test is safe. Since
2231         * t_need_data_flush is likely to be set, we do the test to save some
2232         * cacheline bouncing
2233         */
2234        if (!transaction->t_need_data_flush)
2235                transaction->t_need_data_flush = 1;
2236        /* On some different transaction's list - should be
2237         * the committing one */
2238        if (jinode->i_transaction) {
2239                J_ASSERT(jinode->i_next_transaction == NULL);
2240                J_ASSERT(jinode->i_transaction ==
2241                                        journal->j_committing_transaction);
2242                jinode->i_next_transaction = transaction;
2243                goto done;
2244        }
2245        /* Not on any transaction list... */
2246        J_ASSERT(!jinode->i_next_transaction);
2247        jinode->i_transaction = transaction;
2248        list_add(&jinode->i_list, &transaction->t_inode_list);
2249done:
2250        spin_unlock(&journal->j_list_lock);
2251
2252        return 0;
2253}
2254
2255/*
2256 * File truncate and transaction commit interact with each other in a
2257 * non-trivial way.  If a transaction writing data block A is
2258 * committing, we cannot discard the data by truncate until we have
2259 * written them.  Otherwise if we crashed after the transaction with
2260 * write has committed but before the transaction with truncate has
2261 * committed, we could see stale data in block A.  This function is a
2262 * helper to solve this problem.  It starts writeout of the truncated
2263 * part in case it is in the committing transaction.
2264 *
2265 * Filesystem code must call this function when inode is journaled in
2266 * ordered mode before truncation happens and after the inode has been
2267 * placed on orphan list with the new inode size. The second condition
2268 * avoids the race that someone writes new data and we start
2269 * committing the transaction after this function has been called but
2270 * before a transaction for truncate is started (and furthermore it
2271 * allows us to optimize the case where the addition to orphan list
2272 * happens in the same transaction as write --- we don't have to write
2273 * any data in such case).
2274 */
2275int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2276                                        struct jbd2_inode *jinode,
2277                                        loff_t new_size)
2278{
2279        transaction_t *inode_trans, *commit_trans;
2280        int ret = 0;
2281
2282        /* This is a quick check to avoid locking if not necessary */
2283        if (!jinode->i_transaction)
2284                goto out;
2285        /* Locks are here just to force reading of recent values, it is
2286         * enough that the transaction was not committing before we started
2287         * a transaction adding the inode to orphan list */
2288        read_lock(&journal->j_state_lock);
2289        commit_trans = journal->j_committing_transaction;
2290        read_unlock(&journal->j_state_lock);
2291        spin_lock(&journal->j_list_lock);
2292        inode_trans = jinode->i_transaction;
2293        spin_unlock(&journal->j_list_lock);
2294        if (inode_trans == commit_trans) {
2295                ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2296                        new_size, LLONG_MAX);
2297                if (ret)
2298                        jbd2_journal_abort(journal, ret);
2299        }
2300out:
2301        return ret;
2302}
2303
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.