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
  36/*
  37 * jbd2_get_transaction: obtain a new transaction_t object.
  38 *
  39 * Simply allocate and initialise a new transaction.  Create it in
  40 * RUNNING state and add it to the current journal (which should not
  41 * have an existing running transaction: we only make a new transaction
  42 * once we have started to commit the old one).
  43 *
  44 * Preconditions:
  45 *      The journal MUST be locked.  We don't perform atomic mallocs on the
  46 *      new transaction and we can't block without protecting against other
  47 *      processes trying to touch the journal while it is in transition.
  48 *
  49 */
  50
  51static transaction_t *
  52jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
  53{
  54        transaction->t_journal = journal;
  55        transaction->t_state = T_RUNNING;
  56        transaction->t_start_time = ktime_get();
  57        transaction->t_tid = journal->j_transaction_sequence++;
  58        transaction->t_expires = jiffies + journal->j_commit_interval;
  59        spin_lock_init(&transaction->t_handle_lock);
  60        atomic_set(&transaction->t_updates, 0);
  61        atomic_set(&transaction->t_outstanding_credits, 0);
  62        atomic_set(&transaction->t_handle_count, 0);
  63        INIT_LIST_HEAD(&transaction->t_inode_list);
  64        INIT_LIST_HEAD(&transaction->t_private_list);
  65
  66        /* Set up the commit timer for the new transaction. */
  67        journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
  68        add_timer(&journal->j_commit_timer);
  69
  70        J_ASSERT(journal->j_running_transaction == NULL);
  71        journal->j_running_transaction = transaction;
  72        transaction->t_max_wait = 0;
  73        transaction->t_start = jiffies;
  74
  75        return transaction;
  76}
  77
  78/*
  79 * Handle management.
  80 *
  81 * A handle_t is an object which represents a single atomic update to a
  82 * filesystem, and which tracks all of the modifications which form part
  83 * of that one update.
  84 */
  85
  86/*
  87 * Update transaction's maximum wait time, if debugging is enabled.
  88 *
  89 * In order for t_max_wait to be reliable, it must be protected by a
  90 * lock.  But doing so will mean that start_this_handle() can not be
  91 * run in parallel on SMP systems, which limits our scalability.  So
  92 * unless debugging is enabled, we no longer update t_max_wait, which
  93 * means that maximum wait time reported by the jbd2_run_stats
  94 * tracepoint will always be zero.
  95 */
  96static inline void update_t_max_wait(transaction_t *transaction,
  97                                     unsigned long ts)
  98{
  99#ifdef CONFIG_JBD2_DEBUG
 100        if (jbd2_journal_enable_debug &&
 101            time_after(transaction->t_start, ts)) {
 102                ts = jbd2_time_diff(ts, transaction->t_start);
 103                spin_lock(&transaction->t_handle_lock);
 104                if (ts > transaction->t_max_wait)
 105                        transaction->t_max_wait = ts;
 106                spin_unlock(&transaction->t_handle_lock);
 107        }
 108#endif
 109}
 110
 111/*
 112 * start_this_handle: Given a handle, deal with any locking or stalling
 113 * needed to make sure that there is enough journal space for the handle
 114 * to begin.  Attach the handle to a transaction and set up the
 115 * transaction's buffer credits.
 116 */
 117
 118static int start_this_handle(journal_t *journal, handle_t *handle,
 119                             gfp_t gfp_mask)
 120{
 121        transaction_t   *transaction, *new_transaction = NULL;
 122        tid_t           tid;
 123        int             needed, need_to_start;
 124        int             nblocks = handle->h_buffer_credits;
 125        unsigned long ts = jiffies;
 126
 127        if (nblocks > journal->j_max_transaction_buffers) {
 128                printk(KERN_ERR "JBD2: %s wants too many credits (%d > %d)\n",
 129                       current->comm, nblocks,
 130                       journal->j_max_transaction_buffers);
 131                return -ENOSPC;
 132        }
 133
 134alloc_transaction:
 135        if (!journal->j_running_transaction) {
 136                new_transaction = kzalloc(sizeof(*new_transaction), gfp_mask);
 137                if (!new_transaction) {
 138                        /*
 139                         * If __GFP_FS is not present, then we may be
 140                         * being called from inside the fs writeback
 141                         * layer, so we MUST NOT fail.  Since
 142                         * __GFP_NOFAIL is going away, we will arrange
 143                         * to retry the allocation ourselves.
 144                         */
 145                        if ((gfp_mask & __GFP_FS) == 0) {
 146                                congestion_wait(BLK_RW_ASYNC, HZ/50);
 147                                goto alloc_transaction;
 148                        }
 149                        return -ENOMEM;
 150                }
 151        }
 152
 153        jbd_debug(3, "New handle %p going live.\n", handle);
 154
 155        /*
 156         * We need to hold j_state_lock until t_updates has been incremented,
 157         * for proper journal barrier handling
 158         */
 159repeat:
 160        read_lock(&journal->j_state_lock);
 161        BUG_ON(journal->j_flags & JBD2_UNMOUNT);
 162        if (is_journal_aborted(journal) ||
 163            (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
 164                read_unlock(&journal->j_state_lock);
 165                kfree(new_transaction);
 166                return -EROFS;
 167        }
 168
 169        /* Wait on the journal's transaction barrier if necessary */
 170        if (journal->j_barrier_count) {
 171                read_unlock(&journal->j_state_lock);
 172                wait_event(journal->j_wait_transaction_locked,
 173                                journal->j_barrier_count == 0);
 174                goto repeat;
 175        }
 176
 177        if (!journal->j_running_transaction) {
 178                read_unlock(&journal->j_state_lock);
 179                if (!new_transaction)
 180                        goto alloc_transaction;
 181                write_lock(&journal->j_state_lock);
 182                if (!journal->j_running_transaction) {
 183                        jbd2_get_transaction(journal, new_transaction);
 184                        new_transaction = NULL;
 185                }
 186                write_unlock(&journal->j_state_lock);
 187                goto repeat;
 188        }
 189
 190        transaction = journal->j_running_transaction;
 191
 192        /*
 193         * If the current transaction is locked down for commit, wait for the
 194         * lock to be released.
 195         */
 196        if (transaction->t_state == T_LOCKED) {
 197                DEFINE_WAIT(wait);
 198
 199                prepare_to_wait(&journal->j_wait_transaction_locked,
 200                                        &wait, TASK_UNINTERRUPTIBLE);
 201                read_unlock(&journal->j_state_lock);
 202                schedule();
 203                finish_wait(&journal->j_wait_transaction_locked, &wait);
 204                goto repeat;
 205        }
 206
 207        /*
 208         * If there is not enough space left in the log to write all potential
 209         * buffers requested by this operation, we need to stall pending a log
 210         * checkpoint to free some more log space.
 211         */
 212        needed = atomic_add_return(nblocks,
 213                                   &transaction->t_outstanding_credits);
 214
 215        if (needed > journal->j_max_transaction_buffers) {
 216                /*
 217                 * If the current transaction is already too large, then start
 218                 * to commit it: we can then go back and attach this handle to
 219                 * a new transaction.
 220                 */
 221                DEFINE_WAIT(wait);
 222
 223                jbd_debug(2, "Handle %p starting new commit...\n", handle);
 224                atomic_sub(nblocks, &transaction->t_outstanding_credits);
 225                prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
 226                                TASK_UNINTERRUPTIBLE);
 227                tid = transaction->t_tid;
 228                need_to_start = !tid_geq(journal->j_commit_request, tid);
 229                read_unlock(&journal->j_state_lock);
 230                if (need_to_start)
 231                        jbd2_log_start_commit(journal, tid);
 232                schedule();
 233                finish_wait(&journal->j_wait_transaction_locked, &wait);
 234                goto repeat;
 235        }
 236
 237        /*
 238         * The commit code assumes that it can get enough log space
 239         * without forcing a checkpoint.  This is *critical* for
 240         * correctness: a checkpoint of a buffer which is also
 241         * associated with a committing transaction creates a deadlock,
 242         * so commit simply cannot force through checkpoints.
 243         *
 244         * We must therefore ensure the necessary space in the journal
 245         * *before* starting to dirty potentially checkpointed buffers
 246         * in the new transaction.
 247         *
 248         * The worst part is, any transaction currently committing can
 249         * reduce the free space arbitrarily.  Be careful to account for
 250         * those buffers when checkpointing.
 251         */
 252
 253        /*
 254         * @@@ AKPM: This seems rather over-defensive.  We're giving commit
 255         * a _lot_ of headroom: 1/4 of the journal plus the size of
 256         * the committing transaction.  Really, we only need to give it
 257         * committing_transaction->t_outstanding_credits plus "enough" for
 258         * the log control blocks.
 259         * Also, this test is inconsistent with the matching one in
 260         * jbd2_journal_extend().
 261         */
 262        if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
 263                jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
 264                atomic_sub(nblocks, &transaction->t_outstanding_credits);
 265                read_unlock(&journal->j_state_lock);
 266                write_lock(&journal->j_state_lock);
 267                if (__jbd2_log_space_left(journal) < jbd_space_needed(journal))
 268                        __jbd2_log_wait_for_space(journal);
 269                write_unlock(&journal->j_state_lock);
 270                goto repeat;
 271        }
 272
 273        /* OK, account for the buffers that this operation expects to
 274         * use and add the handle to the running transaction. 
 275         */
 276        update_t_max_wait(transaction, ts);
 277        handle->h_transaction = transaction;
 278        atomic_inc(&transaction->t_updates);
 279        atomic_inc(&transaction->t_handle_count);
 280        jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
 281                  handle, nblocks,
 282                  atomic_read(&transaction->t_outstanding_credits),
 283                  __jbd2_log_space_left(journal));
 284        read_unlock(&journal->j_state_lock);
 285
 286        lock_map_acquire(&handle->h_lockdep_map);
 287        kfree(new_transaction);
 288        return 0;
 289}
 290
 291static struct lock_class_key jbd2_handle_key;
 292
 293/* Allocate a new handle.  This should probably be in a slab... */
 294static handle_t *new_handle(int nblocks)
 295{
 296        handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
 297        if (!handle)
 298                return NULL;
 299        memset(handle, 0, sizeof(*handle));
 300        handle->h_buffer_credits = nblocks;
 301        handle->h_ref = 1;
 302
 303        lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
 304                                                &jbd2_handle_key, 0);
 305
 306        return handle;
 307}
 308
 309/**
 310 * handle_t *jbd2_journal_start() - Obtain a new handle.
 311 * @journal: Journal to start transaction on.
 312 * @nblocks: number of block buffer we might modify
 313 *
 314 * We make sure that the transaction can guarantee at least nblocks of
 315 * modified buffers in the log.  We block until the log can guarantee
 316 * that much space.
 317 *
 318 * This function is visible to journal users (like ext3fs), so is not
 319 * called with the journal already locked.
 320 *
 321 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
 322 * on failure.
 323 */
 324handle_t *jbd2__journal_start(journal_t *journal, int nblocks, gfp_t gfp_mask)
 325{
 326        handle_t *handle = journal_current_handle();
 327        int err;
 328
 329        if (!journal)
 330                return ERR_PTR(-EROFS);
 331
 332        if (handle) {
 333                J_ASSERT(handle->h_transaction->t_journal == journal);
 334                handle->h_ref++;
 335                return handle;
 336        }
 337
 338        handle = new_handle(nblocks);
 339        if (!handle)
 340                return ERR_PTR(-ENOMEM);
 341
 342        current->journal_info = handle;
 343
 344        err = start_this_handle(journal, handle, gfp_mask);
 345        if (err < 0) {
 346                jbd2_free_handle(handle);
 347                current->journal_info = NULL;
 348                handle = ERR_PTR(err);
 349        }
 350        return handle;
 351}
 352EXPORT_SYMBOL(jbd2__journal_start);
 353
 354
 355handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
 356{
 357        return jbd2__journal_start(journal, nblocks, GFP_NOFS);
 358}
 359EXPORT_SYMBOL(jbd2_journal_start);
 360
 361
 362/**
 363 * int jbd2_journal_extend() - extend buffer credits.
 364 * @handle:  handle to 'extend'
 365 * @nblocks: nr blocks to try to extend by.
 366 *
 367 * Some transactions, such as large extends and truncates, can be done
 368 * atomically all at once or in several stages.  The operation requests
 369 * a credit for a number of buffer modications in advance, but can
 370 * extend its credit if it needs more.
 371 *
 372 * jbd2_journal_extend tries to give the running handle more buffer credits.
 373 * It does not guarantee that allocation - this is a best-effort only.
 374 * The calling process MUST be able to deal cleanly with a failure to
 375 * extend here.
 376 *
 377 * Return 0 on success, non-zero on failure.
 378 *
 379 * return code < 0 implies an error
 380 * return code > 0 implies normal transaction-full status.
 381 */
 382int jbd2_journal_extend(handle_t *handle, int nblocks)
 383{
 384        transaction_t *transaction = handle->h_transaction;
 385        journal_t *journal = transaction->t_journal;
 386        int result;
 387        int wanted;
 388
 389        result = -EIO;
 390        if (is_handle_aborted(handle))
 391                goto out;
 392
 393        result = 1;
 394
 395        read_lock(&journal->j_state_lock);
 396
 397        /* Don't extend a locked-down transaction! */
 398        if (handle->h_transaction->t_state != T_RUNNING) {
 399                jbd_debug(3, "denied handle %p %d blocks: "
 400                          "transaction not running\n", handle, nblocks);
 401                goto error_out;
 402        }
 403
 404        spin_lock(&transaction->t_handle_lock);
 405        wanted = atomic_read(&transaction->t_outstanding_credits) + nblocks;
 406
 407        if (wanted > journal->j_max_transaction_buffers) {
 408                jbd_debug(3, "denied handle %p %d blocks: "
 409                          "transaction too large\n", handle, nblocks);
 410                goto unlock;
 411        }
 412
 413        if (wanted > __jbd2_log_space_left(journal)) {
 414                jbd_debug(3, "denied handle %p %d blocks: "
 415                          "insufficient log space\n", handle, nblocks);
 416                goto unlock;
 417        }
 418
 419        handle->h_buffer_credits += nblocks;
 420        atomic_add(nblocks, &transaction->t_outstanding_credits);
 421        result = 0;
 422
 423        jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
 424unlock:
 425        spin_unlock(&transaction->t_handle_lock);
 426error_out:
 427        read_unlock(&journal->j_state_lock);
 428out:
 429        return result;
 430}
 431
 432
 433/**
 434 * int jbd2_journal_restart() - restart a handle .
 435 * @handle:  handle to restart
 436 * @nblocks: nr credits requested
 437 *
 438 * Restart a handle for a multi-transaction filesystem
 439 * operation.
 440 *
 441 * If the jbd2_journal_extend() call above fails to grant new buffer credits
 442 * to a running handle, a call to jbd2_journal_restart will commit the
 443 * handle's transaction so far and reattach the handle to a new
 444 * transaction capabable of guaranteeing the requested number of
 445 * credits.
 446 */
 447int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
 448{
 449        transaction_t *transaction = handle->h_transaction;
 450        journal_t *journal = transaction->t_journal;
 451        tid_t           tid;
 452        int             need_to_start, ret;
 453
 454        /* If we've had an abort of any type, don't even think about
 455         * actually doing the restart! */
 456        if (is_handle_aborted(handle))
 457                return 0;
 458
 459        /*
 460         * First unlink the handle from its current transaction, and start the
 461         * commit on that.
 462         */
 463        J_ASSERT(atomic_read(&transaction->t_updates) > 0);
 464        J_ASSERT(journal_current_handle() == handle);
 465
 466        read_lock(&journal->j_state_lock);
 467        spin_lock(&transaction->t_handle_lock);
 468        atomic_sub(handle->h_buffer_credits,
 469                   &transaction->t_outstanding_credits);
 470        if (atomic_dec_and_test(&transaction->t_updates))
 471                wake_up(&journal->j_wait_updates);
 472        spin_unlock(&transaction->t_handle_lock);
 473
 474        jbd_debug(2, "restarting handle %p\n", handle);
 475        tid = transaction->t_tid;
 476        need_to_start = !tid_geq(journal->j_commit_request, tid);
 477        read_unlock(&journal->j_state_lock);
 478        if (need_to_start)
 479                jbd2_log_start_commit(journal, tid);
 480
 481        lock_map_release(&handle->h_lockdep_map);
 482        handle->h_buffer_credits = nblocks;
 483        ret = start_this_handle(journal, handle, gfp_mask);
 484        return ret;
 485}
 486EXPORT_SYMBOL(jbd2__journal_restart);
 487
 488
 489int jbd2_journal_restart(handle_t *handle, int nblocks)
 490{
 491        return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
 492}
 493EXPORT_SYMBOL(jbd2_journal_restart);
 494
 495/**
 496 * void jbd2_journal_lock_updates () - establish a transaction barrier.
 497 * @journal:  Journal to establish a barrier on.
 498 *
 499 * This locks out any further updates from being started, and blocks
 500 * until all existing updates have completed, returning only once the
 501 * journal is in a quiescent state with no updates running.
 502 *
 503 * The journal lock should not be held on entry.
 504 */
 505void jbd2_journal_lock_updates(journal_t *journal)
 506{
 507        DEFINE_WAIT(wait);
 508
 509        write_lock(&journal->j_state_lock);
 510        ++journal->j_barrier_count;
 511
 512        /* Wait until there are no running updates */
 513        while (1) {
 514                transaction_t *transaction = journal->j_running_transaction;
 515
 516                if (!transaction)
 517                        break;
 518
 519                spin_lock(&transaction->t_handle_lock);
 520                prepare_to_wait(&journal->j_wait_updates, &wait,
 521                                TASK_UNINTERRUPTIBLE);
 522                if (!atomic_read(&transaction->t_updates)) {
 523                        spin_unlock(&transaction->t_handle_lock);
 524                        finish_wait(&journal->j_wait_updates, &wait);
 525                        break;
 526                }
 527                spin_unlock(&transaction->t_handle_lock);
 528                write_unlock(&journal->j_state_lock);
 529                schedule();
 530                finish_wait(&journal->j_wait_updates, &wait);
 531                write_lock(&journal->j_state_lock);
 532        }
 533        write_unlock(&journal->j_state_lock);
 534
 535        /*
 536         * We have now established a barrier against other normal updates, but
 537         * we also need to barrier against other jbd2_journal_lock_updates() calls
 538         * to make sure that we serialise special journal-locked operations
 539         * too.
 540         */
 541        mutex_lock(&journal->j_barrier);
 542}
 543
 544/**
 545 * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
 546 * @journal:  Journal to release the barrier on.
 547 *
 548 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
 549 *
 550 * Should be called without the journal lock held.
 551 */
 552void jbd2_journal_unlock_updates (journal_t *journal)
 553{
 554        J_ASSERT(journal->j_barrier_count != 0);
 555
 556        mutex_unlock(&journal->j_barrier);
 557        write_lock(&journal->j_state_lock);
 558        --journal->j_barrier_count;
 559        write_unlock(&journal->j_state_lock);
 560        wake_up(&journal->j_wait_transaction_locked);
 561}
 562
 563static void warn_dirty_buffer(struct buffer_head *bh)
 564{
 565        char b[BDEVNAME_SIZE];
 566
 567        printk(KERN_WARNING
 568               "JBD2: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
 569               "There's a risk of filesystem corruption in case of system "
 570               "crash.\n",
 571               bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
 572}
 573
 574/*
 575 * If the buffer is already part of the current transaction, then there
 576 * is nothing we need to do.  If it is already part of a prior
 577 * transaction which we are still committing to disk, then we need to
 578 * make sure that we do not overwrite the old copy: we do copy-out to
 579 * preserve the copy going to disk.  We also account the buffer against
 580 * the handle's metadata buffer credits (unless the buffer is already
 581 * part of the transaction, that is).
 582 *
 583 */
 584static int
 585do_get_write_access(handle_t *handle, struct journal_head *jh,
 586                        int force_copy)
 587{
 588        struct buffer_head *bh;
 589        transaction_t *transaction;
 590        journal_t *journal;
 591        int error;
 592        char *frozen_buffer = NULL;
 593        int need_copy = 0;
 594
 595        if (is_handle_aborted(handle))
 596                return -EROFS;
 597
 598        transaction = handle->h_transaction;
 599        journal = transaction->t_journal;
 600
 601        jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
 602
 603        JBUFFER_TRACE(jh, "entry");
 604repeat:
 605        bh = jh2bh(jh);
 606
 607        /* @@@ Need to check for errors here at some point. */
 608
 609        lock_buffer(bh);
 610        jbd_lock_bh_state(bh);
 611
 612        /* We now hold the buffer lock so it is safe to query the buffer
 613         * state.  Is the buffer dirty?
 614         *
 615         * If so, there are two possibilities.  The buffer may be
 616         * non-journaled, and undergoing a quite legitimate writeback.
 617         * Otherwise, it is journaled, and we don't expect dirty buffers
 618         * in that state (the buffers should be marked JBD_Dirty
 619         * instead.)  So either the IO is being done under our own
 620         * control and this is a bug, or it's a third party IO such as
 621         * dump(8) (which may leave the buffer scheduled for read ---
 622         * ie. locked but not dirty) or tune2fs (which may actually have
 623         * the buffer dirtied, ugh.)  */
 624
 625        if (buffer_dirty(bh)) {
 626                /*
 627                 * First question: is this buffer already part of the current
 628                 * transaction or the existing committing transaction?
 629                 */
 630                if (jh->b_transaction) {
 631                        J_ASSERT_JH(jh,
 632                                jh->b_transaction == transaction ||
 633                                jh->b_transaction ==
 634                                        journal->j_committing_transaction);
 635                        if (jh->b_next_transaction)
 636                                J_ASSERT_JH(jh, jh->b_next_transaction ==
 637                                                        transaction);
 638                        warn_dirty_buffer(bh);
 639                }
 640                /*
 641                 * In any case we need to clean the dirty flag and we must
 642                 * do it under the buffer lock to be sure we don't race
 643                 * with running write-out.
 644                 */
 645                JBUFFER_TRACE(jh, "Journalling dirty buffer");
 646                clear_buffer_dirty(bh);
 647                set_buffer_jbddirty(bh);
 648        }
 649
 650        unlock_buffer(bh);
 651
 652        error = -EROFS;
 653        if (is_handle_aborted(handle)) {
 654                jbd_unlock_bh_state(bh);
 655                goto out;
 656        }
 657        error = 0;
 658
 659        /*
 660         * The buffer is already part of this transaction if b_transaction or
 661         * b_next_transaction points to it
 662         */
 663        if (jh->b_transaction == transaction ||
 664            jh->b_next_transaction == transaction)
 665                goto done;
 666
 667        /*
 668         * this is the first time this transaction is touching this buffer,
 669         * reset the modified flag
 670         */
 671       jh->b_modified = 0;
 672
 673        /*
 674         * If there is already a copy-out version of this buffer, then we don't
 675         * need to make another one
 676         */
 677        if (jh->b_frozen_data) {
 678                JBUFFER_TRACE(jh, "has frozen data");
 679                J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 680                jh->b_next_transaction = transaction;
 681                goto done;
 682        }
 683
 684        /* Is there data here we need to preserve? */
 685
 686        if (jh->b_transaction && jh->b_transaction != transaction) {
 687                JBUFFER_TRACE(jh, "owned by older transaction");
 688                J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 689                J_ASSERT_JH(jh, jh->b_transaction ==
 690                                        journal->j_committing_transaction);
 691
 692                /* There is one case we have to be very careful about.
 693                 * If the committing transaction is currently writing
 694                 * this buffer out to disk and has NOT made a copy-out,
 695                 * then we cannot modify the buffer contents at all
 696                 * right now.  The essence of copy-out is that it is the
 697                 * extra copy, not the primary copy, which gets
 698                 * journaled.  If the primary copy is already going to
 699                 * disk then we cannot do copy-out here. */
 700
 701                if (jh->b_jlist == BJ_Shadow) {
 702                        DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
 703                        wait_queue_head_t *wqh;
 704
 705                        wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
 706
 707                        JBUFFER_TRACE(jh, "on shadow: sleep");
 708                        jbd_unlock_bh_state(bh);
 709                        /* commit wakes up all shadow buffers after IO */
 710                        for ( ; ; ) {
 711                                prepare_to_wait(wqh, &wait.wait,
 712                                                TASK_UNINTERRUPTIBLE);
 713                                if (jh->b_jlist != BJ_Shadow)
 714                                        break;
 715                                schedule();
 716                        }
 717                        finish_wait(wqh, &wait.wait);
 718                        goto repeat;
 719                }
 720
 721                /* Only do the copy if the currently-owning transaction
 722                 * still needs it.  If it is on the Forget list, the
 723                 * committing transaction is past that stage.  The
 724                 * buffer had better remain locked during the kmalloc,
 725                 * but that should be true --- we hold the journal lock
 726                 * still and the buffer is already on the BUF_JOURNAL
 727                 * list so won't be flushed.
 728                 *
 729                 * Subtle point, though: if this is a get_undo_access,
 730                 * then we will be relying on the frozen_data to contain
 731                 * the new value of the committed_data record after the
 732                 * transaction, so we HAVE to force the frozen_data copy
 733                 * in that case. */
 734
 735                if (jh->b_jlist != BJ_Forget || force_copy) {
 736                        JBUFFER_TRACE(jh, "generate frozen data");
 737                        if (!frozen_buffer) {
 738                                JBUFFER_TRACE(jh, "allocate memory for buffer");
 739                                jbd_unlock_bh_state(bh);
 740                                frozen_buffer =
 741                                        jbd2_alloc(jh2bh(jh)->b_size,
 742                                                         GFP_NOFS);
 743                                if (!frozen_buffer) {
 744                                        printk(KERN_EMERG
 745                                               "%s: OOM for frozen_buffer\n",
 746                                               __func__);
 747                                        JBUFFER_TRACE(jh, "oom!");
 748                                        error = -ENOMEM;
 749                                        jbd_lock_bh_state(bh);
 750                                        goto done;
 751                                }
 752                                goto repeat;
 753                        }
 754                        jh->b_frozen_data = frozen_buffer;
 755                        frozen_buffer = NULL;
 756                        need_copy = 1;
 757                }
 758                jh->b_next_transaction = transaction;
 759        }
 760
 761
 762        /*
 763         * Finally, if the buffer is not journaled right now, we need to make
 764         * sure it doesn't get written to disk before the caller actually
 765         * commits the new data
 766         */
 767        if (!jh->b_transaction) {
 768                JBUFFER_TRACE(jh, "no transaction");
 769                J_ASSERT_JH(jh, !jh->b_next_transaction);
 770                JBUFFER_TRACE(jh, "file as BJ_Reserved");
 771                spin_lock(&journal->j_list_lock);
 772                __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
 773                spin_unlock(&journal->j_list_lock);
 774        }
 775
 776done:
 777        if (need_copy) {
 778                struct page *page;
 779                int offset;
 780                char *source;
 781
 782                J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
 783                            "Possible IO failure.\n");
 784                page = jh2bh(jh)->b_page;
 785                offset = offset_in_page(jh2bh(jh)->b_data);
 786                source = kmap_atomic(page, KM_USER0);
 787                /* Fire data frozen trigger just before we copy the data */
 788                jbd2_buffer_frozen_trigger(jh, source + offset,
 789                                           jh->b_triggers);
 790                memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
 791                kunmap_atomic(source, KM_USER0);
 792
 793                /*
 794                 * Now that the frozen data is saved off, we need to store
 795                 * any matching triggers.
 796                 */
 797                jh->b_frozen_triggers = jh->b_triggers;
 798        }
 799        jbd_unlock_bh_state(bh);
 800
 801        /*
 802         * If we are about to journal a buffer, then any revoke pending on it is
 803         * no longer valid
 804         */
 805        jbd2_journal_cancel_revoke(handle, jh);
 806
 807out:
 808        if (unlikely(frozen_buffer))    /* It's usually NULL */
 809                jbd2_free(frozen_buffer, bh->b_size);
 810
 811        JBUFFER_TRACE(jh, "exit");
 812        return error;
 813}
 814
 815/**
 816 * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
 817 * @handle: transaction to add buffer modifications to
 818 * @bh:     bh to be used for metadata writes
 819 *
 820 * Returns an error code or 0 on success.
 821 *
 822 * In full data journalling mode the buffer may be of type BJ_AsyncData,
 823 * because we're write()ing a buffer which is also part of a shared mapping.
 824 */
 825
 826int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
 827{
 828        struct journal_head *jh = jbd2_journal_add_journal_head(bh);
 829        int rc;
 830
 831        /* We do not want to get caught playing with fields which the
 832         * log thread also manipulates.  Make sure that the buffer
 833         * completes any outstanding IO before proceeding. */
 834        rc = do_get_write_access(handle, jh, 0);
 835        jbd2_journal_put_journal_head(jh);
 836        return rc;
 837}
 838
 839
 840/*
 841 * When the user wants to journal a newly created buffer_head
 842 * (ie. getblk() returned a new buffer and we are going to populate it
 843 * manually rather than reading off disk), then we need to keep the
 844 * buffer_head locked until it has been completely filled with new
 845 * data.  In this case, we should be able to make the assertion that
 846 * the bh is not already part of an existing transaction.
 847 *
 848 * The buffer should already be locked by the caller by this point.
 849 * There is no lock ranking violation: it was a newly created,
 850 * unlocked buffer beforehand. */
 851
 852/**
 853 * int jbd2_journal_get_create_access () - notify intent to use newly created bh
 854 * @handle: transaction to new buffer to
 855 * @bh: new buffer.
 856 *
 857 * Call this if you create a new bh.
 858 */
 859int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
 860{
 861        transaction_t *transaction = handle->h_transaction;
 862        journal_t *journal = transaction->t_journal;
 863        struct journal_head *jh = jbd2_journal_add_journal_head(bh);
 864        int err;
 865
 866        jbd_debug(5, "journal_head %p\n", jh);
 867        err = -EROFS;
 868        if (is_handle_aborted(handle))
 869                goto out;
 870        err = 0;
 871
 872        JBUFFER_TRACE(jh, "entry");
 873        /*
 874         * The buffer may already belong to this transaction due to pre-zeroing
 875         * in the filesystem's new_block code.  It may also be on the previous,
 876         * committing transaction's lists, but it HAS to be in Forget state in
 877         * that case: the transaction must have deleted the buffer for it to be
 878         * reused here.
 879         */
 880        jbd_lock_bh_state(bh);
 881        spin_lock(&journal->j_list_lock);
 882        J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
 883                jh->b_transaction == NULL ||
 884                (jh->b_transaction == journal->j_committing_transaction &&
 885                          jh->b_jlist == BJ_Forget)));
 886
 887        J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
 888        J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
 889
 890        if (jh->b_transaction == NULL) {
 891                /*
 892                 * Previous jbd2_journal_forget() could have left the buffer
 893                 * with jbddirty bit set because it was being committed. When
 894                 * the commit finished, we've filed the buffer for
 895                 * checkpointing and marked it dirty. Now we are reallocating
 896                 * the buffer so the transaction freeing it must have
 897                 * committed and so it's safe to clear the dirty bit.
 898                 */
 899                clear_buffer_dirty(jh2bh(jh));
 900                /* first access by this transaction */
 901                jh->b_modified = 0;
 902
 903                JBUFFER_TRACE(jh, "file as BJ_Reserved");
 904                __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
 905        } else if (jh->b_transaction == journal->j_committing_transaction) {
 906                /* first access by this transaction */
 907                jh->b_modified = 0;
 908
 909                JBUFFER_TRACE(jh, "set next transaction");
 910                jh->b_next_transaction = transaction;
 911        }
 912        spin_unlock(&journal->j_list_lock);
 913        jbd_unlock_bh_state(bh);
 914
 915        /*
 916         * akpm: I added this.  ext3_alloc_branch can pick up new indirect
 917         * blocks which contain freed but then revoked metadata.  We need
 918         * to cancel the revoke in case we end up freeing it yet again
 919         * and the reallocating as data - this would cause a second revoke,
 920         * which hits an assertion error.
 921         */
 922        JBUFFER_TRACE(jh, "cancelling revoke");
 923        jbd2_journal_cancel_revoke(handle, jh);
 924out:
 925        jbd2_journal_put_journal_head(jh);
 926        return err;
 927}
 928
 929/**
 930 * int jbd2_journal_get_undo_access() -  Notify intent to modify metadata with
 931 *     non-rewindable consequences
 932 * @handle: transaction
 933 * @bh: buffer to undo
 934 *
 935 * Sometimes there is a need to distinguish between metadata which has
 936 * been committed to disk and that which has not.  The ext3fs code uses
 937 * this for freeing and allocating space, we have to make sure that we
 938 * do not reuse freed space until the deallocation has been committed,
 939 * since if we overwrote that space we would make the delete
 940 * un-rewindable in case of a crash.
 941 *
 942 * To deal with that, jbd2_journal_get_undo_access requests write access to a
 943 * buffer for parts of non-rewindable operations such as delete
 944 * operations on the bitmaps.  The journaling code must keep a copy of
 945 * the buffer's contents prior to the undo_access call until such time
 946 * as we know that the buffer has definitely been committed to disk.
 947 *
 948 * We never need to know which transaction the committed data is part
 949 * of, buffers touched here are guaranteed to be dirtied later and so
 950 * will be committed to a new transaction in due course, at which point
 951 * we can discard the old committed data pointer.
 952 *
 953 * Returns error number or 0 on success.
 954 */
 955int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
 956{
 957        int err;
 958        struct journal_head *jh = jbd2_journal_add_journal_head(bh);
 959        char *committed_data = NULL;
 960
 961        JBUFFER_TRACE(jh, "entry");
 962
 963        /*
 964         * Do this first --- it can drop the journal lock, so we want to
 965         * make sure that obtaining the committed_data is done
 966         * atomically wrt. completion of any outstanding commits.
 967         */
 968        err = do_get_write_access(handle, jh, 1);
 969        if (err)
 970                goto out;
 971
 972repeat:
 973        if (!jh->b_committed_data) {
 974                committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
 975                if (!committed_data) {
 976                        printk(KERN_EMERG "%s: No memory for committed data\n",
 977                                __func__);
 978                        err = -ENOMEM;
 979                        goto out;
 980                }
 981        }
 982
 983        jbd_lock_bh_state(bh);
 984        if (!jh->b_committed_data) {
 985                /* Copy out the current buffer contents into the
 986                 * preserved, committed copy. */
 987                JBUFFER_TRACE(jh, "generate b_committed data");
 988                if (!committed_data) {
 989                        jbd_unlock_bh_state(bh);
 990                        goto repeat;
 991                }
 992
 993                jh->b_committed_data = committed_data;
 994                committed_data = NULL;
 995                memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
 996        }
 997        jbd_unlock_bh_state(bh);
 998out:
 999        jbd2_journal_put_journal_head(jh);
1000        if (unlikely(committed_data))
1001                jbd2_free(committed_data, bh->b_size);
1002        return err;
1003}
1004
1005/**
1006 * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1007 * @bh: buffer to trigger on
1008 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1009 *
1010 * Set any triggers on this journal_head.  This is always safe, because
1011 * triggers for a committing buffer will be saved off, and triggers for
1012 * a running transaction will match the buffer in that transaction.
1013 *
1014 * Call with NULL to clear the triggers.
1015 */
1016void jbd2_journal_set_triggers(struct buffer_head *bh,
1017                               struct jbd2_buffer_trigger_type *type)
1018{
1019        struct journal_head *jh = bh2jh(bh);
1020
1021        jh->b_triggers = type;
1022}
1023
1024void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1025                                struct jbd2_buffer_trigger_type *triggers)
1026{
1027        struct buffer_head *bh = jh2bh(jh);
1028
1029        if (!triggers || !triggers->t_frozen)
1030                return;
1031
1032        triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1033}
1034
1035void jbd2_buffer_abort_trigger(struct journal_head *jh,
1036                               struct jbd2_buffer_trigger_type *triggers)
1037{
1038        if (!triggers || !triggers->t_abort)
1039                return;
1040
1041        triggers->t_abort(triggers, jh2bh(jh));
1042}
1043
1044
1045
1046/**
1047 * int jbd2_journal_dirty_metadata() -  mark a buffer as containing dirty metadata
1048 * @handle: transaction to add buffer to.
1049 * @bh: buffer to mark
1050 *
1051 * mark dirty metadata which needs to be journaled as part of the current
1052 * transaction.
1053 *
1054 * The buffer must have previously had jbd2_journal_get_write_access()
1055 * called so that it has a valid journal_head attached to the buffer
1056 * head.
1057 *
1058 * The buffer is placed on the transaction's metadata list and is marked
1059 * as belonging to the transaction.
1060 *
1061 * Returns error number or 0 on success.
1062 *
1063 * Special care needs to be taken if the buffer already belongs to the
1064 * current committing transaction (in which case we should have frozen
1065 * data present for that commit).  In that case, we don't relink the
1066 * buffer: that only gets done when the old transaction finally
1067 * completes its commit.
1068 */
1069int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1070{
1071        transaction_t *transaction = handle->h_transaction;
1072        journal_t *journal = transaction->t_journal;
1073        struct journal_head *jh = bh2jh(bh);
1074        int ret = 0;
1075
1076        jbd_debug(5, "journal_head %p\n", jh);
1077        JBUFFER_TRACE(jh, "entry");
1078        if (is_handle_aborted(handle))
1079                goto out;
1080        if (!buffer_jbd(bh)) {
1081                ret = -EUCLEAN;
1082                goto out;
1083        }
1084
1085        jbd_lock_bh_state(bh);
1086
1087        if (jh->b_modified == 0) {
1088                /*
1089                 * This buffer's got modified and becoming part
1090                 * of the transaction. This needs to be done
1091                 * once a transaction -bzzz
1092                 */
1093                jh->b_modified = 1;
1094                J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1095                handle->h_buffer_credits--;
1096        }
1097
1098        /*
1099         * fastpath, to avoid expensive locking.  If this buffer is already
1100         * on the running transaction's metadata list there is nothing to do.
1101         * Nobody can take it off again because there is a handle open.
1102         * I _think_ we're OK here with SMP barriers - a mistaken decision will
1103         * result in this test being false, so we go in and take the locks.
1104         */
1105        if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1106                JBUFFER_TRACE(jh, "fastpath");
1107                if (unlikely(jh->b_transaction !=
1108                             journal->j_running_transaction)) {
1109                        printk(KERN_EMERG "JBD: %s: "
1110                               "jh->b_transaction (%llu, %p, %u) != "
1111                               "journal->j_running_transaction (%p, %u)",
1112                               journal->j_devname,
1113                               (unsigned long long) bh->b_blocknr,
1114                               jh->b_transaction,
1115                               jh->b_transaction ? jh->b_transaction->t_tid : 0,
1116                               journal->j_running_transaction,
1117                               journal->j_running_transaction ?
1118                               journal->j_running_transaction->t_tid : 0);
1119                        ret = -EINVAL;
1120                }
1121                goto out_unlock_bh;
1122        }
1123
1124        set_buffer_jbddirty(bh);
1125
1126        /*
1127         * Metadata already on the current transaction list doesn't
1128         * need to be filed.  Metadata on another transaction's list must
1129         * be committing, and will be refiled once the commit completes:
1130         * leave it alone for now.
1131         */
1132        if (jh->b_transaction != transaction) {
1133                JBUFFER_TRACE(jh, "already on other transaction");
1134                if (unlikely(jh->b_transaction !=
1135                             journal->j_committing_transaction)) {
1136                        printk(KERN_EMERG "JBD: %s: "
1137                               "jh->b_transaction (%llu, %p, %u) != "
1138                               "journal->j_committing_transaction (%p, %u)",
1139                               journal->j_devname,
1140                               (unsigned long long) bh->b_blocknr,
1141                               jh->b_transaction,
1142                               jh->b_transaction ? jh->b_transaction->t_tid : 0,
1143                               journal->j_committing_transaction,
1144                               journal->j_committing_transaction ?
1145                               journal->j_committing_transaction->t_tid : 0);
1146                        ret = -EINVAL;
1147                }
1148                if (unlikely(jh->b_next_transaction != transaction)) {
1149                        printk(KERN_EMERG "JBD: %s: "
1150                               "jh->b_next_transaction (%llu, %p, %u) != "
1151                               "transaction (%p, %u)",
1152                               journal->j_devname,
1153                               (unsigned long long) bh->b_blocknr,
1154                               jh->b_next_transaction,
1155                               jh->b_next_transaction ?
1156                               jh->b_next_transaction->t_tid : 0,
1157                               transaction, transaction->t_tid);
1158                        ret = -EINVAL;
1159                }
1160                /* And this case is illegal: we can't reuse another
1161                 * transaction's data buffer, ever. */
1162                goto out_unlock_bh;
1163        }
1164
1165        /* That test should have eliminated the following case: */
1166        J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1167
1168        JBUFFER_TRACE(jh, "file as BJ_Metadata");
1169        spin_lock(&journal->j_list_lock);
1170        __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1171        spin_unlock(&journal->j_list_lock);
1172out_unlock_bh:
1173        jbd_unlock_bh_state(bh);
1174out:
1175        JBUFFER_TRACE(jh, "exit");
1176        WARN_ON(ret);   /* All errors are bugs, so dump the stack */
1177        return ret;
1178}
1179
1180/*
1181 * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1182 * updates, if the update decided in the end that it didn't need access.
1183 *
1184 */
1185void
1186jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1187{
1188        BUFFER_TRACE(bh, "entry");
1189}
1190
1191/**
1192 * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1193 * @handle: transaction handle
1194 * @bh:     bh to 'forget'
1195 *
1196 * We can only do the bforget if there are no commits pending against the
1197 * buffer.  If the buffer is dirty in the current running transaction we
1198 * can safely unlink it.
1199 *
1200 * bh may not be a journalled buffer at all - it may be a non-JBD
1201 * buffer which came off the hashtable.  Check for this.
1202 *
1203 * Decrements bh->b_count by one.
1204 *
1205 * Allow this call even if the handle has aborted --- it may be part of
1206 * the caller's cleanup after an abort.
1207 */
1208int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1209{
1210        transaction_t *transaction = handle->h_transaction;
1211        journal_t *journal = transaction->t_journal;
1212        struct journal_head *jh;
1213        int drop_reserve = 0;
1214        int err = 0;
1215        int was_modified = 0;
1216
1217        BUFFER_TRACE(bh, "entry");
1218
1219        jbd_lock_bh_state(bh);
1220        spin_lock(&journal->j_list_lock);
1221
1222        if (!buffer_jbd(bh))
1223                goto not_jbd;
1224        jh = bh2jh(bh);
1225
1226        /* Critical error: attempting to delete a bitmap buffer, maybe?
1227         * Don't do any jbd operations, and return an error. */
1228        if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1229                         "inconsistent data on disk")) {
1230                err = -EIO;
1231                goto not_jbd;
1232        }
1233
1234        /* keep track of wether or not this transaction modified us */
1235        was_modified = jh->b_modified;
1236
1237        /*
1238         * The buffer's going from the transaction, we must drop
1239         * all references -bzzz
1240         */
1241        jh->b_modified = 0;
1242
1243        if (jh->b_transaction == handle->h_transaction) {
1244                J_ASSERT_JH(jh, !jh->b_frozen_data);
1245
1246                /* If we are forgetting a buffer which is already part
1247                 * of this transaction, then we can just drop it from
1248                 * the transaction immediately. */
1249                clear_buffer_dirty(bh);
1250                clear_buffer_jbddirty(bh);
1251
1252                JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1253
1254                /*
1255                 * we only want to drop a reference if this transaction
1256                 * modified the buffer
1257                 */
1258                if (was_modified)
1259                        drop_reserve = 1;
1260
1261                /*
1262                 * We are no longer going to journal this buffer.
1263                 * However, the commit of this transaction is still
1264                 * important to the buffer: the delete that we are now
1265                 * processing might obsolete an old log entry, so by
1266                 * committing, we can satisfy the buffer's checkpoint.
1267                 *
1268                 * So, if we have a checkpoint on the buffer, we should
1269                 * now refile the buffer on our BJ_Forget list so that
1270                 * we know to remove the checkpoint after we commit.
1271                 */
1272
1273                if (jh->b_cp_transaction) {
1274                        __jbd2_journal_temp_unlink_buffer(jh);
1275                        __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1276                } else {
1277                        __jbd2_journal_unfile_buffer(jh);
1278                        if (!buffer_jbd(bh)) {
1279                                spin_unlock(&journal->j_list_lock);
1280                                jbd_unlock_bh_state(bh);
1281                                __bforget(bh);
1282                                goto drop;
1283                        }
1284                }
1285        } else if (jh->b_transaction) {
1286                J_ASSERT_JH(jh, (jh->b_transaction ==
1287                                 journal->j_committing_transaction));
1288                /* However, if the buffer is still owned by a prior
1289                 * (committing) transaction, we can't drop it yet... */
1290                JBUFFER_TRACE(jh, "belongs to older transaction");
1291                /* ... but we CAN drop it from the new transaction if we
1292                 * have also modified it since the original commit. */
1293
1294                if (jh->b_next_transaction) {
1295                        J_ASSERT(jh->b_next_transaction == transaction);
1296                        jh->b_next_transaction = NULL;
1297
1298                        /*
1299                         * only drop a reference if this transaction modified
1300                         * the buffer
1301                         */
1302                        if (was_modified)
1303                                drop_reserve = 1;
1304                }
1305        }
1306
1307not_jbd:
1308        spin_unlock(&journal->j_list_lock);
1309        jbd_unlock_bh_state(bh);
1310        __brelse(bh);
1311drop:
1312        if (drop_reserve) {
1313                /* no need to reserve log space for this block -bzzz */
1314                handle->h_buffer_credits++;
1315        }
1316        return err;
1317}
1318
1319/**
1320 * int jbd2_journal_stop() - complete a transaction
1321 * @handle: tranaction to complete.
1322 *
1323 * All done for a particular handle.
1324 *
1325 * There is not much action needed here.  We just return any remaining
1326 * buffer credits to the transaction and remove the handle.  The only
1327 * complication is that we need to start a commit operation if the
1328 * filesystem is marked for synchronous update.
1329 *
1330 * jbd2_journal_stop itself will not usually return an error, but it may
1331 * do so in unusual circumstances.  In particular, expect it to
1332 * return -EIO if a jbd2_journal_abort has been executed since the
1333 * transaction began.
1334 */
1335int jbd2_journal_stop(handle_t *handle)
1336{
1337        transaction_t *transaction = handle->h_transaction;
1338        journal_t *journal = transaction->t_journal;
1339        int err, wait_for_commit = 0;
1340        tid_t tid;
1341        pid_t pid;
1342
1343        J_ASSERT(journal_current_handle() == handle);
1344
1345        if (is_handle_aborted(handle))
1346                err = -EIO;
1347        else {
1348                J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1349                err = 0;
1350        }
1351
1352        if (--handle->h_ref > 0) {
1353                jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1354                          handle->h_ref);
1355                return err;
1356        }
1357
1358        jbd_debug(4, "Handle %p going down\n", handle);
1359
1360        /*
1361         * Implement synchronous transaction batching.  If the handle
1362         * was synchronous, don't force a commit immediately.  Let's
1363         * yield and let another thread piggyback onto this
1364         * transaction.  Keep doing that while new threads continue to
1365         * arrive.  It doesn't cost much - we're about to run a commit
1366         * and sleep on IO anyway.  Speeds up many-threaded, many-dir
1367         * operations by 30x or more...
1368         *
1369         * We try and optimize the sleep time against what the
1370         * underlying disk can do, instead of having a static sleep
1371         * time.  This is useful for the case where our storage is so
1372         * fast that it is more optimal to go ahead and force a flush
1373         * and wait for the transaction to be committed than it is to
1374         * wait for an arbitrary amount of time for new writers to
1375         * join the transaction.  We achieve this by measuring how
1376         * long it takes to commit a transaction, and compare it with
1377         * how long this transaction has been running, and if run time
1378         * < commit time then we sleep for the delta and commit.  This
1379         * greatly helps super fast disks that would see slowdowns as
1380         * more threads started doing fsyncs.
1381         *
1382         * But don't do this if this process was the most recent one
1383         * to perform a synchronous write.  We do this to detect the
1384         * case where a single process is doing a stream of sync
1385         * writes.  No point in waiting for joiners in that case.
1386         */
1387        pid = current->pid;
1388        if (handle->h_sync && journal->j_last_sync_writer != pid) {
1389                u64 commit_time, trans_time;
1390
1391                journal->j_last_sync_writer = pid;
1392
1393                read_lock(&journal->j_state_lock);
1394                commit_time = journal->j_average_commit_time;
1395                read_unlock(&journal->j_state_lock);
1396
1397                trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1398                                                   transaction->t_start_time));
1399
1400                commit_time = max_t(u64, commit_time,
1401                                    1000*journal->j_min_batch_time);
1402                commit_time = min_t(u64, commit_time,
1403                                    1000*journal->j_max_batch_time);
1404
1405                if (trans_time < commit_time) {
1406                        ktime_t expires = ktime_add_ns(ktime_get(),
1407                                                       commit_time);
1408                        set_current_state(TASK_UNINTERRUPTIBLE);
1409                        schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1410                }
1411        }
1412
1413        if (handle->h_sync)
1414                transaction->t_synchronous_commit = 1;
1415        current->journal_info = NULL;
1416        atomic_sub(handle->h_buffer_credits,
1417                   &transaction->t_outstanding_credits);
1418
1419        /*
1420         * If the handle is marked SYNC, we need to set another commit
1421         * going!  We also want to force a commit if the current
1422         * transaction is occupying too much of the log, or if the
1423         * transaction is too old now.
1424         */
1425        if (handle->h_sync ||
1426            (atomic_read(&transaction->t_outstanding_credits) >
1427             journal->j_max_transaction_buffers) ||
1428            time_after_eq(jiffies, transaction->t_expires)) {
1429                /* Do this even for aborted journals: an abort still
1430                 * completes the commit thread, it just doesn't write
1431                 * anything to disk. */
1432
1433                jbd_debug(2, "transaction too old, requesting commit for "
1434                                        "handle %p\n", handle);
1435                /* This is non-blocking */
1436                jbd2_log_start_commit(journal, transaction->t_tid);
1437
1438                /*
1439                 * Special case: JBD2_SYNC synchronous updates require us
1440                 * to wait for the commit to complete.
1441                 */
1442                if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1443                        wait_for_commit = 1;
1444        }
1445
1446        /*
1447         * Once we drop t_updates, if it goes to zero the transaction
1448         * could start committing on us and eventually disappear.  So
1449         * once we do this, we must not dereference transaction
1450         * pointer again.
1451         */
1452        tid = transaction->t_tid;
1453        if (atomic_dec_and_test(&transaction->t_updates)) {
1454                wake_up(&journal->j_wait_updates);
1455                if (journal->j_barrier_count)
1456                        wake_up(&journal->j_wait_transaction_locked);
1457        }
1458
1459        if (wait_for_commit)
1460                err = jbd2_log_wait_commit(journal, tid);
1461
1462        lock_map_release(&handle->h_lockdep_map);
1463
1464        jbd2_free_handle(handle);
1465        return err;
1466}
1467
1468/**
1469 * int jbd2_journal_force_commit() - force any uncommitted transactions
1470 * @journal: journal to force
1471 *
1472 * For synchronous operations: force any uncommitted transactions
1473 * to disk.  May seem kludgy, but it reuses all the handle batching
1474 * code in a very simple manner.
1475 */
1476int jbd2_journal_force_commit(journal_t *journal)
1477{
1478        handle_t *handle;
1479        int ret;
1480
1481        handle = jbd2_journal_start(journal, 1);
1482        if (IS_ERR(handle)) {
1483                ret = PTR_ERR(handle);
1484        } else {
1485                handle->h_sync = 1;
1486                ret = jbd2_journal_stop(handle);
1487        }
1488        return ret;
1489}
1490
1491/*
1492 *
1493 * List management code snippets: various functions for manipulating the
1494 * transaction buffer lists.
1495 *
1496 */
1497
1498/*
1499 * Append a buffer to a transaction list, given the transaction's list head
1500 * pointer.
1501 *
1502 * j_list_lock is held.
1503 *
1504 * jbd_lock_bh_state(jh2bh(jh)) is held.
1505 */
1506
1507static inline void
1508__blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1509{
1510        if (!*list) {
1511                jh->b_tnext = jh->b_tprev = jh;
1512                *list = jh;
1513        } else {
1514                /* Insert at the tail of the list to preserve order */
1515                struct journal_head *first = *list, *last = first->b_tprev;
1516                jh->b_tprev = last;
1517                jh->b_tnext = first;
1518                last->b_tnext = first->b_tprev = jh;
1519        }
1520}
1521
1522/*
1523 * Remove a buffer from a transaction list, given the transaction's list
1524 * head pointer.
1525 *
1526 * Called with j_list_lock held, and the journal may not be locked.
1527 *
1528 * jbd_lock_bh_state(jh2bh(jh)) is held.
1529 */
1530
1531static inline void
1532__blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1533{
1534        if (*list == jh) {
1535                *list = jh->b_tnext;
1536                if (*list == jh)
1537                        *list = NULL;
1538        }
1539        jh->b_tprev->b_tnext = jh->b_tnext;
1540        jh->b_tnext->b_tprev = jh->b_tprev;
1541}
1542
1543/*
1544 * Remove a buffer from the appropriate transaction list.
1545 *
1546 * Note that this function can *change* the value of
1547 * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1548 * t_log_list or t_reserved_list.  If the caller is holding onto a copy of one
1549 * of these pointers, it could go bad.  Generally the caller needs to re-read
1550 * the pointer from the transaction_t.
1551 *
1552 * Called under j_list_lock.  The journal may not be locked.
1553 */
1554void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1555{
1556        struct journal_head **list = NULL;
1557        transaction_t *transaction;
1558        struct buffer_head *bh = jh2bh(jh);
1559
1560        J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1561        transaction = jh->b_transaction;
1562        if (transaction)
1563                assert_spin_locked(&transaction->t_journal->j_list_lock);
1564
1565        J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1566        if (jh->b_jlist != BJ_None)
1567                J_ASSERT_JH(jh, transaction != NULL);
1568
1569        switch (jh->b_jlist) {
1570        case BJ_None:
1571                return;
1572        case BJ_Metadata:
1573                transaction->t_nr_buffers--;
1574                J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1575                list = &transaction->t_buffers;
1576                break;
1577        case BJ_Forget:
1578                list = &transaction->t_forget;
1579                break;
1580        case BJ_IO:
1581                list = &transaction->t_iobuf_list;
1582                break;
1583        case BJ_Shadow:
1584                list = &transaction->t_shadow_list;
1585                break;
1586        case BJ_LogCtl:
1587                list = &transaction->t_log_list;
1588                break;
1589        case BJ_Reserved:
1590                list = &transaction->t_reserved_list;
1591                break;
1592        }
1593
1594        __blist_del_buffer(list, jh);
1595        jh->b_jlist = BJ_None;
1596        if (test_clear_buffer_jbddirty(bh))
1597                mark_buffer_dirty(bh);  /* Expose it to the VM */
1598}
1599
1600/*
1601 * Remove buffer from all transactions.
1602 *
1603 * Called with bh_state lock and j_list_lock
1604 *
1605 * jh and bh may be already freed when this function returns.
1606 */
1607static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1608{
1609        __jbd2_journal_temp_unlink_buffer(jh);
1610        jh->b_transaction = NULL;
1611        jbd2_journal_put_journal_head(jh);
1612}
1613
1614void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1615{
1616        struct buffer_head *bh = jh2bh(jh);
1617
1618        /* Get reference so that buffer cannot be freed before we unlock it */
1619        get_bh(bh);
1620        jbd_lock_bh_state(bh);
1621        spin_lock(&journal->j_list_lock);
1622        __jbd2_journal_unfile_buffer(jh);
1623        spin_unlock(&journal->j_list_lock);
1624        jbd_unlock_bh_state(bh);
1625        __brelse(bh);
1626}
1627
1628/*
1629 * Called from jbd2_journal_try_to_free_buffers().
1630 *
1631 * Called under jbd_lock_bh_state(bh)
1632 */
1633static void
1634__journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1635{
1636        struct journal_head *jh;
1637
1638        jh = bh2jh(bh);
1639
1640        if (buffer_locked(bh) || buffer_dirty(bh))
1641                goto out;
1642
1643        if (jh->b_next_transaction != NULL)
1644                goto out;
1645
1646        spin_lock(&journal->j_list_lock);
1647        if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1648                /* written-back checkpointed metadata buffer */
1649                if (jh->b_jlist == BJ_None) {
1650                        JBUFFER_TRACE(jh, "remove from checkpoint list");
1651                        __jbd2_journal_remove_checkpoint(jh);
1652                }
1653        }
1654        spin_unlock(&journal->j_list_lock);
1655out:
1656        return;
1657}
1658
1659/**
1660 * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1661 * @journal: journal for operation
1662 * @page: to try and free
1663 * @gfp_mask: we use the mask to detect how hard should we try to release
1664 * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1665 * release the buffers.
1666 *
1667 *
1668 * For all the buffers on this page,
1669 * if they are fully written out ordered data, move them onto BUF_CLEAN
1670 * so try_to_free_buffers() can reap them.
1671 *
1672 * This function returns non-zero if we wish try_to_free_buffers()
1673 * to be called. We do this if the page is releasable by try_to_free_buffers().
1674 * We also do it if the page has locked or dirty buffers and the caller wants
1675 * us to perform sync or async writeout.
1676 *
1677 * This complicates JBD locking somewhat.  We aren't protected by the
1678 * BKL here.  We wish to remove the buffer from its committing or
1679 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1680 *
1681 * This may *change* the value of transaction_t->t_datalist, so anyone
1682 * who looks at t_datalist needs to lock against this function.
1683 *
1684 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1685 * buffer.  So we need to lock against that.  jbd2_journal_dirty_data()
1686 * will come out of the lock with the buffer dirty, which makes it
1687 * ineligible for release here.
1688 *
1689 * Who else is affected by this?  hmm...  Really the only contender
1690 * is do_get_write_access() - it could be looking at the buffer while
1691 * journal_try_to_free_buffer() is changing its state.  But that
1692 * cannot happen because we never reallocate freed data as metadata
1693 * while the data is part of a transaction.  Yes?
1694 *
1695 * Return 0 on failure, 1 on success
1696 */
1697int jbd2_journal_try_to_free_buffers(journal_t *journal,
1698                                struct page *page, gfp_t gfp_mask)
1699{
1700        struct buffer_head *head;
1701        struct buffer_head *bh;
1702        int ret = 0;
1703
1704        J_ASSERT(PageLocked(page));
1705
1706        head = page_buffers(page);
1707        bh = head;
1708        do {
1709                struct journal_head *jh;
1710
1711                /*
1712                 * We take our own ref against the journal_head here to avoid
1713                 * having to add tons of locking around each instance of
1714                 * jbd2_journal_put_journal_head().
1715                 */
1716                jh = jbd2_journal_grab_journal_head(bh);
1717                if (!jh)
1718                        continue;
1719
1720                jbd_lock_bh_state(bh);
1721                __journal_try_to_free_buffer(journal, bh);
1722                jbd2_journal_put_journal_head(jh);
1723                jbd_unlock_bh_state(bh);
1724                if (buffer_jbd(bh))
1725                        goto busy;
1726        } while ((bh = bh->b_this_page) != head);
1727
1728        ret = try_to_free_buffers(page);
1729
1730busy:
1731        return ret;
1732}
1733
1734/*
1735 * This buffer is no longer needed.  If it is on an older transaction's
1736 * checkpoint list we need to record it on this transaction's forget list
1737 * to pin this buffer (and hence its checkpointing transaction) down until
1738 * this transaction commits.  If the buffer isn't on a checkpoint list, we
1739 * release it.
1740 * Returns non-zero if JBD no longer has an interest in the buffer.
1741 *
1742 * Called under j_list_lock.
1743 *
1744 * Called under jbd_lock_bh_state(bh).
1745 */
1746static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1747{
1748        int may_free = 1;
1749        struct buffer_head *bh = jh2bh(jh);
1750
1751        if (jh->b_cp_transaction) {
1752                JBUFFER_TRACE(jh, "on running+cp transaction");
1753                __jbd2_journal_temp_unlink_buffer(jh);
1754                /*
1755                 * We don't want to write the buffer anymore, clear the
1756                 * bit so that we don't confuse checks in
1757                 * __journal_file_buffer
1758                 */
1759                clear_buffer_dirty(bh);
1760                __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1761                may_free = 0;
1762        } else {
1763                JBUFFER_TRACE(jh, "on running transaction");
1764                __jbd2_journal_unfile_buffer(jh);
1765        }
1766        return may_free;
1767}
1768
1769/*
1770 * jbd2_journal_invalidatepage
1771 *
1772 * This code is tricky.  It has a number of cases to deal with.
1773 *
1774 * There are two invariants which this code relies on:
1775 *
1776 * i_size must be updated on disk before we start calling invalidatepage on the
1777 * data.
1778 *
1779 *  This is done in ext3 by defining an ext3_setattr method which
1780 *  updates i_size before truncate gets going.  By maintaining this
1781 *  invariant, we can be sure that it is safe to throw away any buffers
1782 *  attached to the current transaction: once the transaction commits,
1783 *  we know that the data will not be needed.
1784 *
1785 *  Note however that we can *not* throw away data belonging to the
1786 *  previous, committing transaction!
1787 *
1788 * Any disk blocks which *are* part of the previous, committing
1789 * transaction (and which therefore cannot be discarded immediately) are
1790 * not going to be reused in the new running transaction
1791 *
1792 *  The bitmap committed_data images guarantee this: any block which is
1793 *  allocated in one transaction and removed in the next will be marked
1794 *  as in-use in the committed_data bitmap, so cannot be reused until
1795 *  the next transaction to delete the block commits.  This means that
1796 *  leaving committing buffers dirty is quite safe: the disk blocks
1797 *  cannot be reallocated to a different file and so buffer aliasing is
1798 *  not possible.
1799 *
1800 *
1801 * The above applies mainly to ordered data mode.  In writeback mode we
1802 * don't make guarantees about the order in which data hits disk --- in
1803 * particular we don't guarantee that new dirty data is flushed before
1804 * transaction commit --- so it is always safe just to discard data
1805 * immediately in that mode.  --sct
1806 */
1807
1808/*
1809 * The journal_unmap_buffer helper function returns zero if the buffer
1810 * concerned remains pinned as an anonymous buffer belonging to an older
1811 * transaction.
1812 *
1813 * We're outside-transaction here.  Either or both of j_running_transaction
1814 * and j_committing_transaction may be NULL.
1815 */
1816static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1817{
1818        transaction_t *transaction;
1819        struct journal_head *jh;
1820        int may_free = 1;
1821        int ret;
1822
1823        BUFFER_TRACE(bh, "entry");
1824
1825        /*
1826         * It is safe to proceed here without the j_list_lock because the
1827         * buffers cannot be stolen by try_to_free_buffers as long as we are
1828         * holding the page lock. --sct
1829         */
1830
1831        if (!buffer_jbd(bh))
1832                goto zap_buffer_unlocked;
1833
1834        /* OK, we have data buffer in journaled mode */
1835        write_lock(&journal->j_state_lock);
1836        jbd_lock_bh_state(bh);
1837        spin_lock(&journal->j_list_lock);
1838
1839        jh = jbd2_journal_grab_journal_head(bh);
1840        if (!jh)
1841                goto zap_buffer_no_jh;
1842
1843        /*
1844         * We cannot remove the buffer from checkpoint lists until the
1845         * transaction adding inode to orphan list (let's call it T)
1846         * is committed.  Otherwise if the transaction changing the
1847         * buffer would be cleaned from the journal before T is
1848         * committed, a crash will cause that the correct contents of
1849         * the buffer will be lost.  On the other hand we have to
1850         * clear the buffer dirty bit at latest at the moment when the
1851         * transaction marking the buffer as freed in the filesystem
1852         * structures is committed because from that moment on the
1853         * buffer can be reallocated and used by a different page.
1854         * Since the block hasn't been freed yet but the inode has
1855         * already been added to orphan list, it is safe for us to add
1856         * the buffer to BJ_Forget list of the newest transaction.
1857         */
1858        transaction = jh->b_transaction;
1859        if (transaction == NULL) {
1860                /* First case: not on any transaction.  If it
1861                 * has no checkpoint link, then we can zap it:
1862                 * it's a writeback-mode buffer so we don't care
1863                 * if it hits disk safely. */
1864                if (!jh->b_cp_transaction) {
1865                        JBUFFER_TRACE(jh, "not on any transaction: zap");
1866                        goto zap_buffer;
1867                }
1868
1869                if (!buffer_dirty(bh)) {
1870                        /* bdflush has written it.  We can drop it now */
1871                        goto zap_buffer;
1872                }
1873
1874                /* OK, it must be in the journal but still not
1875                 * written fully to disk: it's metadata or
1876                 * journaled data... */
1877
1878                if (journal->j_running_transaction) {
1879                        /* ... and once the current transaction has
1880                         * committed, the buffer won't be needed any
1881                         * longer. */
1882                        JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1883                        ret = __dispose_buffer(jh,
1884                                        journal->j_running_transaction);
1885                        jbd2_journal_put_journal_head(jh);
1886                        spin_unlock(&journal->j_list_lock);
1887                        jbd_unlock_bh_state(bh);
1888                        write_unlock(&journal->j_state_lock);
1889                        return ret;
1890                } else {
1891                        /* There is no currently-running transaction. So the
1892                         * orphan record which we wrote for this file must have
1893                         * passed into commit.  We must attach this buffer to
1894                         * the committing transaction, if it exists. */
1895                        if (journal->j_committing_transaction) {
1896                                JBUFFER_TRACE(jh, "give to committing trans");
1897                                ret = __dispose_buffer(jh,
1898                                        journal->j_committing_transaction);
1899                                jbd2_journal_put_journal_head(jh);
1900                                spin_unlock(&journal->j_list_lock);
1901                                jbd_unlock_bh_state(bh);
1902                                write_unlock(&journal->j_state_lock);
1903                                return ret;
1904                        } else {
1905                                /* The orphan record's transaction has
1906                                 * committed.  We can cleanse this buffer */
1907                                clear_buffer_jbddirty(bh);
1908                                goto zap_buffer;
1909                        }
1910                }
1911        } else if (transaction == journal->j_committing_transaction) {
1912                JBUFFER_TRACE(jh, "on committing transaction");
1913                /*
1914                 * The buffer is committing, we simply cannot touch
1915                 * it. So we just set j_next_transaction to the
1916                 * running transaction (if there is one) and mark
1917                 * buffer as freed so that commit code knows it should
1918                 * clear dirty bits when it is done with the buffer.
1919                 */
1920                set_buffer_freed(bh);
1921                if (journal->j_running_transaction && buffer_jbddirty(bh))
1922                        jh->b_next_transaction = journal->j_running_transaction;
1923                jbd2_journal_put_journal_head(jh);
1924                spin_unlock(&journal->j_list_lock);
1925                jbd_unlock_bh_state(bh);
1926                write_unlock(&journal->j_state_lock);
1927                return 0;
1928        } else {
1929                /* Good, the buffer belongs to the running transaction.
1930                 * We are writing our own transaction's data, not any
1931                 * previous one's, so it is safe to throw it away
1932                 * (remember that we expect the filesystem to have set
1933                 * i_size already for this truncate so recovery will not
1934                 * expose the disk blocks we are discarding here.) */
1935                J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1936                JBUFFER_TRACE(jh, "on running transaction");
1937                may_free = __dispose_buffer(jh, transaction);
1938        }
1939
1940zap_buffer:
1941        jbd2_journal_put_journal_head(jh);
1942zap_buffer_no_jh:
1943        spin_unlock(&journal->j_list_lock);
1944        jbd_unlock_bh_state(bh);
1945        write_unlock(&journal->j_state_lock);
1946zap_buffer_unlocked:
1947        clear_buffer_dirty(bh);
1948        J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1949        clear_buffer_mapped(bh);
1950        clear_buffer_req(bh);
1951        clear_buffer_new(bh);
1952        clear_buffer_delay(bh);
1953        clear_buffer_unwritten(bh);
1954        bh->b_bdev = NULL;
1955        return may_free;
1956}
1957
1958/**
1959 * void jbd2_journal_invalidatepage()
1960 * @journal: journal to use for flush...
1961 * @page:    page to flush
1962 * @offset:  length of page to invalidate.
1963 *
1964 * Reap page buffers containing data after offset in page.
1965 *
1966 */
1967void jbd2_journal_invalidatepage(journal_t *journal,
1968                      struct page *page,
1969                      unsigned long offset)
1970{
1971        struct buffer_head *head, *bh, *next;
1972        unsigned int curr_off = 0;
1973        int may_free = 1;
1974
1975        if (!PageLocked(page))
1976                BUG();
1977        if (!page_has_buffers(page))
1978                return;
1979
1980        /* We will potentially be playing with lists other than just the
1981         * data lists (especially for journaled data mode), so be
1982         * cautious in our locking. */
1983
1984        head = bh = page_buffers(page);
1985        do {
1986                unsigned int next_off = curr_off + bh->b_size;
1987                next = bh->b_this_page;
1988
1989                if (offset <= curr_off) {
1990                        /* This block is wholly outside the truncation point */
1991                        lock_buffer(bh);
1992                        may_free &= journal_unmap_buffer(journal, bh);
1993                        unlock_buffer(bh);
1994                }
1995                curr_off = next_off;
1996                bh = next;
1997
1998        } while (bh != head);
1999
2000        if (!offset) {
2001                if (may_free && try_to_free_buffers(page))
2002                        J_ASSERT(!page_has_buffers(page));
2003        }
2004}
2005
2006/*
2007 * File a buffer on the given transaction list.
2008 */
2009void __jbd2_journal_file_buffer(struct journal_head *jh,
2010                        transaction_t *transaction, int jlist)
2011{
2012        struct journal_head **list = NULL;
2013        int was_dirty = 0;
2014        struct buffer_head *bh = jh2bh(jh);
2015
2016        J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2017        assert_spin_locked(&transaction->t_journal->j_list_lock);
2018
2019        J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2020        J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2021                                jh->b_transaction == NULL);
2022
2023        if (jh->b_transaction && jh->b_jlist == jlist)
2024                return;
2025
2026        if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2027            jlist == BJ_Shadow || jlist == BJ_Forget) {
2028                /*
2029                 * For metadata buffers, we track dirty bit in buffer_jbddirty
2030                 * instead of buffer_dirty. We should not see a dirty bit set
2031                 * here because we clear it in do_get_write_access but e.g.
2032                 * tune2fs can modify the sb and set the dirty bit at any time
2033                 * so we try to gracefully handle that.
2034                 */
2035                if (buffer_dirty(bh))
2036                        warn_dirty_buffer(bh);
2037                if (test_clear_buffer_dirty(bh) ||
2038                    test_clear_buffer_jbddirty(bh))
2039                        was_dirty = 1;
2040        }
2041
2042        if (jh->b_transaction)
2043                __jbd2_journal_temp_unlink_buffer(jh);
2044        else
2045                jbd2_journal_grab_journal_head(bh);
2046        jh->b_transaction = transaction;
2047
2048        switch (jlist) {
2049        case BJ_None:
2050                J_ASSERT_JH(jh, !jh->b_committed_data);
2051                J_ASSERT_JH(jh, !jh->b_frozen_data);
2052                return;
2053        case BJ_Metadata:
2054                transaction->t_nr_buffers++;
2055                list = &transaction->t_buffers;
2056                break;
2057        case BJ_Forget:
2058                list = &transaction->t_forget;
2059                break;
2060        case BJ_IO:
2061                list = &transaction->t_iobuf_list;
2062                break;
2063        case BJ_Shadow:
2064                list = &transaction->t_shadow_list;
2065                break;
2066        case BJ_LogCtl:
2067                list = &transaction->t_log_list;
2068                break;
2069        case BJ_Reserved:
2070                list = &transaction->t_reserved_list;
2071                break;
2072        }
2073
2074        __blist_add_buffer(list, jh);
2075        jh->b_jlist = jlist;
2076
2077        if (was_dirty)
2078                set_buffer_jbddirty(bh);
2079}
2080
2081void jbd2_journal_file_buffer(struct journal_head *jh,
2082                                transaction_t *transaction, int jlist)
2083{
2084        jbd_lock_bh_state(jh2bh(jh));
2085        spin_lock(&transaction->t_journal->j_list_lock);
2086        __jbd2_journal_file_buffer(jh, transaction, jlist);
2087        spin_unlock(&transaction->t_journal->j_list_lock);
2088        jbd_unlock_bh_state(jh2bh(jh));
2089}
2090
2091/*
2092 * Remove a buffer from its current buffer list in preparation for
2093 * dropping it from its current transaction entirely.  If the buffer has
2094 * already started to be used by a subsequent transaction, refile the
2095 * buffer on that transaction's metadata list.
2096 *
2097 * Called under j_list_lock
2098 * Called under jbd_lock_bh_state(jh2bh(jh))
2099 *
2100 * jh and bh may be already free when this function returns
2101 */
2102void __jbd2_journal_refile_buffer(struct journal_head *jh)
2103{
2104        int was_dirty, jlist;
2105        struct buffer_head *bh = jh2bh(jh);
2106
2107        J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2108        if (jh->b_transaction)
2109                assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2110
2111        /* If the buffer is now unused, just drop it. */
2112        if (jh->b_next_transaction == NULL) {
2113                __jbd2_journal_unfile_buffer(jh);
2114                return;
2115        }
2116
2117        /*
2118         * It has been modified by a later transaction: add it to the new
2119         * transaction's metadata list.
2120         */
2121
2122        was_dirty = test_clear_buffer_jbddirty(bh);
2123        __jbd2_journal_temp_unlink_buffer(jh);
2124        /*
2125         * We set b_transaction here because b_next_transaction will inherit
2126         * our jh reference and thus __jbd2_journal_file_buffer() must not
2127         * take a new one.
2128         */
2129        jh->b_transaction = jh->b_next_transaction;
2130        jh->b_next_transaction = NULL;
2131        if (buffer_freed(bh))
2132                jlist = BJ_Forget;
2133        else if (jh->b_modified)
2134                jlist = BJ_Metadata;
2135        else
2136                jlist = BJ_Reserved;
2137        __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2138        J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2139
2140        if (was_dirty)
2141                set_buffer_jbddirty(bh);
2142}
2143
2144/*
2145 * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2146 * bh reference so that we can safely unlock bh.
2147 *
2148 * The jh and bh may be freed by this call.
2149 */
2150void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2151{
2152        struct buffer_head *bh = jh2bh(jh);
2153
2154        /* Get reference so that buffer cannot be freed before we unlock it */
2155        get_bh(bh);
2156        jbd_lock_bh_state(bh);
2157        spin_lock(&journal->j_list_lock);
2158        __jbd2_journal_refile_buffer(jh);
2159        jbd_unlock_bh_state(bh);
2160        spin_unlock(&journal->j_list_lock);
2161        __brelse(bh);
2162}
2163
2164/*
2165 * File inode in the inode list of the handle's transaction
2166 */
2167int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2168{
2169        transaction_t *transaction = handle->h_transaction;
2170        journal_t *journal = transaction->t_journal;
2171
2172        if (is_handle_aborted(handle))
2173                return -EIO;
2174
2175        jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2176                        transaction->t_tid);
2177
2178        /*
2179         * First check whether inode isn't already on the transaction's
2180         * lists without taking the lock. Note that this check is safe
2181         * without the lock as we cannot race with somebody removing inode
2182         * from the transaction. The reason is that we remove inode from the
2183         * transaction only in journal_release_jbd_inode() and when we commit
2184         * the transaction. We are guarded from the first case by holding
2185         * a reference to the inode. We are safe against the second case
2186         * because if jinode->i_transaction == transaction, commit code
2187         * cannot touch the transaction because we hold reference to it,
2188         * and if jinode->i_next_transaction == transaction, commit code
2189         * will only file the inode where we want it.
2190         */
2191        if (jinode->i_transaction == transaction ||
2192            jinode->i_next_transaction == transaction)
2193                return 0;
2194
2195        spin_lock(&journal->j_list_lock);
2196
2197        if (jinode->i_transaction == transaction ||
2198            jinode->i_next_transaction == transaction)
2199                goto done;
2200
2201        /*
2202         * We only ever set this variable to 1 so the test is safe. Since
2203         * t_need_data_flush is likely to be set, we do the test to save some
2204         * cacheline bouncing
2205         */
2206        if (!transaction->t_need_data_flush)
2207                transaction->t_need_data_flush = 1;
2208        /* On some different transaction's list - should be
2209         * the committing one */
2210        if (jinode->i_transaction) {
2211                J_ASSERT(jinode->i_next_transaction == NULL);
2212                J_ASSERT(jinode->i_transaction ==
2213                                        journal->j_committing_transaction);
2214                jinode->i_next_transaction = transaction;
2215                goto done;
2216        }
2217        /* Not on any transaction list... */
2218        J_ASSERT(!jinode->i_next_transaction);
2219        jinode->i_transaction = transaction;
2220        list_add(&jinode->i_list, &transaction->t_inode_list);
2221done:
2222        spin_unlock(&journal->j_list_lock);
2223
2224        return 0;
2225}
2226
2227/*
2228 * File truncate and transaction commit interact with each other in a
2229 * non-trivial way.  If a transaction writing data block A is
2230 * committing, we cannot discard the data by truncate until we have
2231 * written them.  Otherwise if we crashed after the transaction with
2232 * write has committed but before the transaction with truncate has
2233 * committed, we could see stale data in block A.  This function is a
2234 * helper to solve this problem.  It starts writeout of the truncated
2235 * part in case it is in the committing transaction.
2236 *
2237 * Filesystem code must call this function when inode is journaled in
2238 * ordered mode before truncation happens and after the inode has been
2239 * placed on orphan list with the new inode size. The second condition
2240 * avoids the race that someone writes new data and we start
2241 * committing the transaction after this function has been called but
2242 * before a transaction for truncate is started (and furthermore it
2243 * allows us to optimize the case where the addition to orphan list
2244 * happens in the same transaction as write --- we don't have to write
2245 * any data in such case).
2246 */
2247int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2248                                        struct jbd2_inode *jinode,
2249                                        loff_t new_size)
2250{
2251        transaction_t *inode_trans, *commit_trans;
2252        int ret = 0;
2253
2254        /* This is a quick check to avoid locking if not necessary */
2255        if (!jinode->i_transaction)
2256                goto out;
2257        /* Locks are here just to force reading of recent values, it is
2258         * enough that the transaction was not committing before we started
2259         * a transaction adding the inode to orphan list */
2260        read_lock(&journal->j_state_lock);
2261        commit_trans = journal->j_committing_transaction;
2262        read_unlock(&journal->j_state_lock);
2263        spin_lock(&journal->j_list_lock);
2264        inode_trans = jinode->i_transaction;
2265        spin_unlock(&journal->j_list_lock);
2266        if (inode_trans == commit_trans) {
2267                ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2268                        new_size, LLONG_MAX);
2269                if (ret)
2270                        jbd2_journal_abort(journal, ret);
2271        }
2272out:
2273        return ret;
2274}
2275