transaction.c

linux 内核源代码

/**
 * int journal_get_undo_access() -  Notify intent to modify metadata with
 *     non-rewindable consequences
 * @handle: transaction
 * @bh: buffer to undo
 * @credits: store the number of taken credits here (if not NULL)
 *
 * Sometimes there is a need to distinguish between metadata which has
 * been committed to disk and that which has not.  The ext3fs code uses
 * this for freeing and allocating space, we have to make sure that we
 * do not reuse freed space until the deallocation has been committed,
 * since if we overwrote that space we would make the delete
 * un-rewindable in case of a crash.
 *
 * To deal with that, journal_get_undo_access requests write access to a
 * buffer for parts of non-rewindable operations such as delete
 * operations on the bitmaps.  The journaling code must keep a copy of
 * the buffer's contents prior to the undo_access call until such time
 * as we know that the buffer has definitely been committed to disk.
 *
 * We never need to know which transaction the committed data is part
 * of, buffers touched here are guaranteed to be dirtied later and so
 * will be committed to a new transaction in due course, at which point
 * we can discard the old committed data pointer.
 *
 * Returns error number or 0 on success.
 */
int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
{
	int err;
	struct journal_head *jh = journal_add_journal_head(bh);
	char *committed_data = NULL;

	JBUFFER_TRACE(jh, "entry");

	/*
	 * Do this first --- it can drop the journal lock, so we want to
	 * make sure that obtaining the committed_data is done
	 * atomically wrt. completion of any outstanding commits.
	 */
	err = do_get_write_access(handle, jh, 1);
	if (err)
		goto out;

repeat:
	if (!jh->b_committed_data) {
		committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
		if (!committed_data) {
			printk(KERN_EMERG "%s: No memory for committed data\n",
				__FUNCTION__);
			err = -ENOMEM;
			goto out;
		}
	}

	jbd_lock_bh_state(bh);
	if (!jh->b_committed_data) {
		/* Copy out the current buffer contents into the
		 * preserved, committed copy. */
		JBUFFER_TRACE(jh, "generate b_committed data");
		if (!committed_data) {
			jbd_unlock_bh_state(bh);
			goto repeat;
		}

		jh->b_committed_data = committed_data;
		committed_data = NULL;
		memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
	}
	jbd_unlock_bh_state(bh);
out:
	journal_put_journal_head(jh);
	if (unlikely(committed_data))
		jbd_free(committed_data, bh->b_size);
	return err;
}

/**
 * int journal_dirty_data() -  mark a buffer as containing dirty data which
 *                             needs to be flushed before we can commit the
 *                             current transaction.
 * @handle: transaction
 * @bh: bufferhead to mark
 *
 * The buffer is placed on the transaction's data list and is marked as
 * belonging to the transaction.
 *
 * Returns error number or 0 on success.
 *
 * journal_dirty_data() can be called via page_launder->ext3_writepage
 * by kswapd.
 */
int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
{
	journal_t *journal = handle->h_transaction->t_journal;
	int need_brelse = 0;
	struct journal_head *jh;

	if (is_handle_aborted(handle))
		return 0;

	jh = journal_add_journal_head(bh);
	JBUFFER_TRACE(jh, "entry");

	/*
	 * The buffer could *already* be dirty.  Writeout can start
	 * at any time.
	 */
	jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);

	/*
	 * What if the buffer is already part of a running transaction?
	 *
	 * There are two cases:
	 * 1) It is part of the current running transaction.  Refile it,
	 *    just in case we have allocated it as metadata, deallocated
	 *    it, then reallocated it as data.
	 * 2) It is part of the previous, still-committing transaction.
	 *    If all we want to do is to guarantee that the buffer will be
	 *    written to disk before this new transaction commits, then
	 *    being sure that the *previous* transaction has this same
	 *    property is sufficient for us!  Just leave it on its old
	 *    transaction.
	 *
	 * In case (2), the buffer must not already exist as metadata
	 * --- that would violate write ordering (a transaction is free
	 * to write its data at any point, even before the previous
	 * committing transaction has committed).  The caller must
	 * never, ever allow this to happen: there's nothing we can do
	 * about it in this layer.
	 */
	jbd_lock_bh_state(bh);
	spin_lock(&journal->j_list_lock);

	/* Now that we have bh_state locked, are we really still mapped? */
	if (!buffer_mapped(bh)) {
		JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
		goto no_journal;
	}

	if (jh->b_transaction) {
		JBUFFER_TRACE(jh, "has transaction");
		if (jh->b_transaction != handle->h_transaction) {
			JBUFFER_TRACE(jh, "belongs to older transaction");
			J_ASSERT_JH(jh, jh->b_transaction ==
					journal->j_committing_transaction);

			/* @@@ IS THIS TRUE  ? */
			/*
			 * Not any more.  Scenario: someone does a write()
			 * in data=journal mode.  The buffer's transaction has
			 * moved into commit.  Then someone does another
			 * write() to the file.  We do the frozen data copyout
			 * and set b_next_transaction to point to j_running_t.
			 * And while we're in that state, someone does a
			 * writepage() in an attempt to pageout the same area
			 * of the file via a shared mapping.  At present that
			 * calls journal_dirty_data(), and we get right here.
			 * It may be too late to journal the data.  Simply
			 * falling through to the next test will suffice: the
			 * data will be dirty and wil be checkpointed.  The
			 * ordering comments in the next comment block still
			 * apply.
			 */
			//J_ASSERT_JH(jh, jh->b_next_transaction == NULL);

			/*
			 * If we're journalling data, and this buffer was
			 * subject to a write(), it could be metadata, forget
			 * or shadow against the committing transaction.  Now,
			 * someone has dirtied the same darn page via a mapping
			 * and it is being writepage()'d.
			 * We *could* just steal the page from commit, with some
			 * fancy locking there.  Instead, we just skip it -
			 * don't tie the page's buffers to the new transaction
			 * at all.
			 * Implication: if we crash before the writepage() data
			 * is written into the filesystem, recovery will replay
			 * the write() data.
			 */
			if (jh->b_jlist != BJ_None &&
					jh->b_jlist != BJ_SyncData &&
					jh->b_jlist != BJ_Locked) {
				JBUFFER_TRACE(jh, "Not stealing");
				goto no_journal;
			}

			/*
			 * This buffer may be undergoing writeout in commit.  We
			 * can't return from here and let the caller dirty it
			 * again because that can cause the write-out loop in
			 * commit to never terminate.
			 */
			if (buffer_dirty(bh)) {
				get_bh(bh);
				spin_unlock(&journal->j_list_lock);
				jbd_unlock_bh_state(bh);
				need_brelse = 1;
				sync_dirty_buffer(bh);
				jbd_lock_bh_state(bh);
				spin_lock(&journal->j_list_lock);
				/* Since we dropped the lock... */
				if (!buffer_mapped(bh)) {
					JBUFFER_TRACE(jh, "buffer got unmapped");
					goto no_journal;
				}
				/* The buffer may become locked again at any
				   time if it is redirtied */
			}

			/* journal_clean_data_list() may have got there first */
			if (jh->b_transaction != NULL) {
				JBUFFER_TRACE(jh, "unfile from commit");
				__journal_temp_unlink_buffer(jh);
				/* It still points to the committing
				 * transaction; move it to this one so
				 * that the refile assert checks are
				 * happy. */
				jh->b_transaction = handle->h_transaction;
			}
			/* The buffer will be refiled below */

		}
		/*
		 * Special case --- the buffer might actually have been
		 * allocated and then immediately deallocated in the previous,
		 * committing transaction, so might still be left on that
		 * transaction's metadata lists.
		 */
		if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
			JBUFFER_TRACE(jh, "not on correct data list: unfile");
			J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
			__journal_temp_unlink_buffer(jh);
			jh->b_transaction = handle->h_transaction;
			JBUFFER_TRACE(jh, "file as data");
			__journal_file_buffer(jh, handle->h_transaction,
						BJ_SyncData);
		}
	} else {
		JBUFFER_TRACE(jh, "not on a transaction");
		__journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
	}
no_journal:
	spin_unlock(&journal->j_list_lock);
	jbd_unlock_bh_state(bh);
	if (need_brelse) {
		BUFFER_TRACE(bh, "brelse");
		__brelse(bh);
	}
	JBUFFER_TRACE(jh, "exit");
	journal_put_journal_head(jh);
	return 0;
}

/**
 * int journal_dirty_metadata() -  mark a buffer as containing dirty metadata
 * @handle: transaction to add buffer to.
 * @bh: buffer to mark
 *
 * mark dirty metadata which needs to be journaled as part of the current
 * transaction.
 *
 * The buffer is placed on the transaction's metadata list and is marked
 * as belonging to the transaction.
 *
 * Returns error number or 0 on success.
 *
 * Special care needs to be taken if the buffer already belongs to the
 * current committing transaction (in which case we should have frozen
 * data present for that commit).  In that case, we don't relink the
 * buffer: that only gets done when the old transaction finally
 * completes its commit.
 */
int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
{
	transaction_t *transaction = handle->h_transaction;
	journal_t *journal = transaction->t_journal;
	struct journal_head *jh = bh2jh(bh);

	jbd_debug(5, "journal_head %p\n", jh);
	JBUFFER_TRACE(jh, "entry");
	if (is_handle_aborted(handle))
		goto out;

	jbd_lock_bh_state(bh);

	if (jh->b_modified == 0) {
		/*
		 * This buffer's got modified and becoming part
		 * of the transaction. This needs to be done
		 * once a transaction -bzzz
		 */
		jh->b_modified = 1;
		J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
		handle->h_buffer_credits--;
	}

	/*
	 * fastpath, to avoid expensive locking.  If this buffer is already
	 * on the running transaction's metadata list there is nothing to do.
	 * Nobody can take it off again because there is a handle open.
	 * I _think_ we're OK here with SMP barriers - a mistaken decision will
	 * result in this test being false, so we go in and take the locks.
	 */
	if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
		JBUFFER_TRACE(jh, "fastpath");
		J_ASSERT_JH(jh, jh->b_transaction ==
					journal->j_running_transaction);
		goto out_unlock_bh;
	}

	set_buffer_jbddirty(bh);

	/*
	 * Metadata already on the current transaction list doesn't
	 * need to be filed.  Metadata on another transaction's list must
	 * be committing, and will be refiled once the commit completes:
	 * leave it alone for now.
	 */
	if (jh->b_transaction != transaction) {
		JBUFFER_TRACE(jh, "already on other transaction");
		J_ASSERT_JH(jh, jh->b_transaction ==
					journal->j_committing_transaction);
		J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
		/* And this case is illegal: we can't reuse another
		 * transaction's data buffer, ever. */
		goto out_unlock_bh;
	}

	/* That test should have eliminated the following case: */
	J_ASSERT_JH(jh, jh->b_frozen_data == NULL);

	JBUFFER_TRACE(jh, "file as BJ_Metadata");
	spin_lock(&journal->j_list_lock);
	__journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
	spin_unlock(&journal->j_list_lock);
out_unlock_bh:
	jbd_unlock_bh_state(bh);
out:
	JBUFFER_TRACE(jh, "exit");
	return 0;
}

/*
 * journal_release_buffer: undo a get_write_access without any buffer
 * updates, if the update decided in the end that it didn't need access.
 *
 */
void
journal_release_buffer(handle_t *handle, struct buffer_head *bh)
{
	BUFFER_TRACE(bh, "entry");
}

/**
 * void journal_forget() - bforget() for potentially-journaled buffers.
 * @handle: transaction handle
 * @bh:     bh to 'forget'
 *
 * We can only do the bforget if there are no commits pending against the
 * buffer.  If the buffer is dirty in the current running transaction we
 * can safely unlink it.
 *
 * bh may not be a journalled buffer at all - it may be a non-JBD
 * buffer which came off the hashtable.  Check for this.
 *
 * Decrements bh->b_count by one.
 *
 * Allow this call even if the handle has aborted --- it may be part of
 * the caller's cleanup after an abort.
 */
int journal_forget (handle_t *handle, struct buffer_head *bh)
{
	transaction_t *transaction = handle->h_transaction;
	journal_t *journal = transaction->t_journal;
	struct journal_head *jh;
	int drop_reserve = 0;
	int err = 0;

	BUFFER_TRACE(bh, "entry");

	jbd_lock_bh_state(bh);
	spin_lock(&journal->j_list_lock);

	if (!buffer_jbd(bh))
		goto not_jbd;
	jh = bh2jh(bh);

	/* Critical error: attempting to delete a bitmap buffer, maybe?
	 * Don't do any jbd operations, and return an error. */
	if (!J_EXPECT_JH(jh, !jh->b_committed_data,
			 "inconsistent data on disk")) {
		err = -EIO;
		goto not_jbd;
	}

	/*
	 * The buffer's going from the transaction, we must drop
	 * all references -bzzz
	 */
	jh->b_modified = 0;

	if (jh->b_transaction == handle->h_transaction) {
		J_ASSERT_JH(jh, !jh->b_frozen_data);

		/* If we are forgetting a buffer which is already part
		 * of this transaction, then we can just drop it from
		 * the transaction immediately. */
		clear_buffer_dirty(bh);
		clear_buffer_jbddirty(bh);

		JBUFFER_TRACE(jh, "belongs to current transaction: unfile");

		drop_reserve = 1;

		/*
		 * We are no longer going to journal this buffer.
		 * However, the commit of this transaction is still
		 * important to the buffer: the delete that we are now
		 * processing might obsolete an old log entry, so by
		 * committing, we can satisfy the buffer's checkpoint.