transaction.c

linux 内核源代码

/**
 * void journal_lock_updates () - establish a transaction barrier.
 * @journal:  Journal to establish a barrier on.
 *
 * This locks out any further updates from being started, and blocks
 * until all existing updates have completed, returning only once the
 * journal is in a quiescent state with no updates running.
 *
 * The journal lock should not be held on entry.
 */
void journal_lock_updates(journal_t *journal)
{
	DEFINE_WAIT(wait);

	spin_lock(&journal->j_state_lock);
	++journal->j_barrier_count;

	/* Wait until there are no running updates */
	while (1) {
		transaction_t *transaction = journal->j_running_transaction;

		if (!transaction)
			break;

		spin_lock(&transaction->t_handle_lock);
		if (!transaction->t_updates) {
			spin_unlock(&transaction->t_handle_lock);
			break;
		}
		prepare_to_wait(&journal->j_wait_updates, &wait,
				TASK_UNINTERRUPTIBLE);
		spin_unlock(&transaction->t_handle_lock);
		spin_unlock(&journal->j_state_lock);
		schedule();
		finish_wait(&journal->j_wait_updates, &wait);
		spin_lock(&journal->j_state_lock);
	}
	spin_unlock(&journal->j_state_lock);

	/*
	 * We have now established a barrier against other normal updates, but
	 * we also need to barrier against other journal_lock_updates() calls
	 * to make sure that we serialise special journal-locked operations
	 * too.
	 */
	mutex_lock(&journal->j_barrier);
}

/**
 * void journal_unlock_updates (journal_t* journal) - release barrier
 * @journal:  Journal to release the barrier on.
 *
 * Release a transaction barrier obtained with journal_lock_updates().
 *
 * Should be called without the journal lock held.
 */
void journal_unlock_updates (journal_t *journal)
{
	J_ASSERT(journal->j_barrier_count != 0);

	mutex_unlock(&journal->j_barrier);
	spin_lock(&journal->j_state_lock);
	--journal->j_barrier_count;
	spin_unlock(&journal->j_state_lock);
	wake_up(&journal->j_wait_transaction_locked);
}

/*
 * Report any unexpected dirty buffers which turn up.  Normally those
 * indicate an error, but they can occur if the user is running (say)
 * tune2fs to modify the live filesystem, so we need the option of
 * continuing as gracefully as possible.  #
 *
 * The caller should already hold the journal lock and
 * j_list_lock spinlock: most callers will need those anyway
 * in order to probe the buffer's journaling state safely.
 */
static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
{
	int jlist;

	/* If this buffer is one which might reasonably be dirty
	 * --- ie. data, or not part of this journal --- then
	 * we're OK to leave it alone, but otherwise we need to
	 * move the dirty bit to the journal's own internal
	 * JBDDirty bit. */
	jlist = jh->b_jlist;

	if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
	    jlist == BJ_Shadow || jlist == BJ_Forget) {
		struct buffer_head *bh = jh2bh(jh);

		if (test_clear_buffer_dirty(bh))
			set_buffer_jbddirty(bh);
	}
}

/*
 * If the buffer is already part of the current transaction, then there
 * is nothing we need to do.  If it is already part of a prior
 * transaction which we are still committing to disk, then we need to
 * make sure that we do not overwrite the old copy: we do copy-out to
 * preserve the copy going to disk.  We also account the buffer against
 * the handle's metadata buffer credits (unless the buffer is already
 * part of the transaction, that is).
 *
 */
static int
do_get_write_access(handle_t *handle, struct journal_head *jh,
			int force_copy)
{
	struct buffer_head *bh;
	transaction_t *transaction;
	journal_t *journal;
	int error;
	char *frozen_buffer = NULL;
	int need_copy = 0;

	if (is_handle_aborted(handle))
		return -EROFS;

	transaction = handle->h_transaction;
	journal = transaction->t_journal;

	jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);

	JBUFFER_TRACE(jh, "entry");
repeat:
	bh = jh2bh(jh);

	/* @@@ Need to check for errors here at some point. */

	lock_buffer(bh);
	jbd_lock_bh_state(bh);

	/* We now hold the buffer lock so it is safe to query the buffer
	 * state.  Is the buffer dirty?
	 *
	 * If so, there are two possibilities.  The buffer may be
	 * non-journaled, and undergoing a quite legitimate writeback.
	 * Otherwise, it is journaled, and we don't expect dirty buffers
	 * in that state (the buffers should be marked JBD_Dirty
	 * instead.)  So either the IO is being done under our own
	 * control and this is a bug, or it's a third party IO such as
	 * dump(8) (which may leave the buffer scheduled for read ---
	 * ie. locked but not dirty) or tune2fs (which may actually have
	 * the buffer dirtied, ugh.)  */

	if (buffer_dirty(bh)) {
		/*
		 * First question: is this buffer already part of the current
		 * transaction or the existing committing transaction?
		 */
		if (jh->b_transaction) {
			J_ASSERT_JH(jh,
				jh->b_transaction == transaction ||
				jh->b_transaction ==
					journal->j_committing_transaction);
			if (jh->b_next_transaction)
				J_ASSERT_JH(jh, jh->b_next_transaction ==
							transaction);
		}
		/*
		 * In any case we need to clean the dirty flag and we must
		 * do it under the buffer lock to be sure we don't race
		 * with running write-out.
		 */
		JBUFFER_TRACE(jh, "Unexpected dirty buffer");
		jbd_unexpected_dirty_buffer(jh);
	}

	unlock_buffer(bh);

	error = -EROFS;
	if (is_handle_aborted(handle)) {
		jbd_unlock_bh_state(bh);
		goto out;
	}
	error = 0;

	/*
	 * The buffer is already part of this transaction if b_transaction or
	 * b_next_transaction points to it
	 */
	if (jh->b_transaction == transaction ||
	    jh->b_next_transaction == transaction)
		goto done;

	/*
	 * If there is already a copy-out version of this buffer, then we don't
	 * need to make another one
	 */
	if (jh->b_frozen_data) {
		JBUFFER_TRACE(jh, "has frozen data");
		J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
		jh->b_next_transaction = transaction;
		goto done;
	}

	/* Is there data here we need to preserve? */

	if (jh->b_transaction && jh->b_transaction != transaction) {
		JBUFFER_TRACE(jh, "owned by older transaction");
		J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
		J_ASSERT_JH(jh, jh->b_transaction ==
					journal->j_committing_transaction);

		/* There is one case we have to be very careful about.
		 * If the committing transaction is currently writing
		 * this buffer out to disk and has NOT made a copy-out,
		 * then we cannot modify the buffer contents at all
		 * right now.  The essence of copy-out is that it is the
		 * extra copy, not the primary copy, which gets
		 * journaled.  If the primary copy is already going to
		 * disk then we cannot do copy-out here. */

		if (jh->b_jlist == BJ_Shadow) {
			DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
			wait_queue_head_t *wqh;

			wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);

			JBUFFER_TRACE(jh, "on shadow: sleep");
			jbd_unlock_bh_state(bh);
			/* commit wakes up all shadow buffers after IO */
			for ( ; ; ) {
				prepare_to_wait(wqh, &wait.wait,
						TASK_UNINTERRUPTIBLE);
				if (jh->b_jlist != BJ_Shadow)
					break;
				schedule();
			}
			finish_wait(wqh, &wait.wait);
			goto repeat;
		}

		/* Only do the copy if the currently-owning transaction
		 * still needs it.  If it is on the Forget list, the
		 * committing transaction is past that stage.  The
		 * buffer had better remain locked during the kmalloc,
		 * but that should be true --- we hold the journal lock
		 * still and the buffer is already on the BUF_JOURNAL
		 * list so won't be flushed.
		 *
		 * Subtle point, though: if this is a get_undo_access,
		 * then we will be relying on the frozen_data to contain
		 * the new value of the committed_data record after the
		 * transaction, so we HAVE to force the frozen_data copy
		 * in that case. */

		if (jh->b_jlist != BJ_Forget || force_copy) {
			JBUFFER_TRACE(jh, "generate frozen data");
			if (!frozen_buffer) {
				JBUFFER_TRACE(jh, "allocate memory for buffer");
				jbd_unlock_bh_state(bh);
				frozen_buffer =
					jbd_alloc(jh2bh(jh)->b_size,
							 GFP_NOFS);
				if (!frozen_buffer) {
					printk(KERN_EMERG
					       "%s: OOM for frozen_buffer\n",
					       __FUNCTION__);
					JBUFFER_TRACE(jh, "oom!");
					error = -ENOMEM;
					jbd_lock_bh_state(bh);
					goto done;
				}
				goto repeat;
			}
			jh->b_frozen_data = frozen_buffer;
			frozen_buffer = NULL;
			need_copy = 1;
		}
		jh->b_next_transaction = transaction;
	}


	/*
	 * Finally, if the buffer is not journaled right now, we need to make
	 * sure it doesn't get written to disk before the caller actually
	 * commits the new data
	 */
	if (!jh->b_transaction) {
		JBUFFER_TRACE(jh, "no transaction");
		J_ASSERT_JH(jh, !jh->b_next_transaction);
		jh->b_transaction = transaction;
		JBUFFER_TRACE(jh, "file as BJ_Reserved");
		spin_lock(&journal->j_list_lock);
		__journal_file_buffer(jh, transaction, BJ_Reserved);
		spin_unlock(&journal->j_list_lock);
	}

done:
	if (need_copy) {
		struct page *page;
		int offset;
		char *source;

		J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
			    "Possible IO failure.\n");
		page = jh2bh(jh)->b_page;
		offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
		source = kmap_atomic(page, KM_USER0);
		memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
		kunmap_atomic(source, KM_USER0);
	}
	jbd_unlock_bh_state(bh);

	/*
	 * If we are about to journal a buffer, then any revoke pending on it is
	 * no longer valid
	 */
	journal_cancel_revoke(handle, jh);

out:
	if (unlikely(frozen_buffer))	/* It's usually NULL */
		jbd_free(frozen_buffer, bh->b_size);

	JBUFFER_TRACE(jh, "exit");
	return error;
}

/**
 * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
 * @handle: transaction to add buffer modifications to
 * @bh:     bh to be used for metadata writes
 * @credits: variable that will receive credits for the buffer
 *
 * Returns an error code or 0 on success.
 *
 * In full data journalling mode the buffer may be of type BJ_AsyncData,
 * because we're write()ing a buffer which is also part of a shared mapping.
 */

int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
{
	struct journal_head *jh = journal_add_journal_head(bh);
	int rc;

	/* We do not want to get caught playing with fields which the
	 * log thread also manipulates.  Make sure that the buffer
	 * completes any outstanding IO before proceeding. */
	rc = do_get_write_access(handle, jh, 0);
	journal_put_journal_head(jh);
	return rc;
}


/*
 * When the user wants to journal a newly created buffer_head
 * (ie. getblk() returned a new buffer and we are going to populate it
 * manually rather than reading off disk), then we need to keep the
 * buffer_head locked until it has been completely filled with new
 * data.  In this case, we should be able to make the assertion that
 * the bh is not already part of an existing transaction.
 *
 * The buffer should already be locked by the caller by this point.
 * There is no lock ranking violation: it was a newly created,
 * unlocked buffer beforehand. */

/**
 * int journal_get_create_access () - notify intent to use newly created bh
 * @handle: transaction to new buffer to
 * @bh: new buffer.
 *
 * Call this if you create a new bh.
 */
int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
{
	transaction_t *transaction = handle->h_transaction;
	journal_t *journal = transaction->t_journal;
	struct journal_head *jh = journal_add_journal_head(bh);
	int err;

	jbd_debug(5, "journal_head %p\n", jh);
	err = -EROFS;
	if (is_handle_aborted(handle))
		goto out;
	err = 0;

	JBUFFER_TRACE(jh, "entry");
	/*
	 * The buffer may already belong to this transaction due to pre-zeroing
	 * in the filesystem's new_block code.  It may also be on the previous,
	 * committing transaction's lists, but it HAS to be in Forget state in
	 * that case: the transaction must have deleted the buffer for it to be
	 * reused here.
	 */
	jbd_lock_bh_state(bh);
	spin_lock(&journal->j_list_lock);
	J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
		jh->b_transaction == NULL ||
		(jh->b_transaction == journal->j_committing_transaction &&
			  jh->b_jlist == BJ_Forget)));

	J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
	J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));

	if (jh->b_transaction == NULL) {
		jh->b_transaction = transaction;
		JBUFFER_TRACE(jh, "file as BJ_Reserved");
		__journal_file_buffer(jh, transaction, BJ_Reserved);
	} else if (jh->b_transaction == journal->j_committing_transaction) {
		JBUFFER_TRACE(jh, "set next transaction");
		jh->b_next_transaction = transaction;
	}
	spin_unlock(&journal->j_list_lock);
	jbd_unlock_bh_state(bh);

	/*
	 * akpm: I added this.  ext3_alloc_branch can pick up new indirect
	 * blocks which contain freed but then revoked metadata.  We need
	 * to cancel the revoke in case we end up freeing it yet again
	 * and the reallocating as data - this would cause a second revoke,
	 * which hits an assertion error.
	 */
	JBUFFER_TRACE(jh, "cancelling revoke");
	journal_cancel_revoke(handle, jh);
	journal_put_journal_head(jh);
out:
	return err;
}