Message ID | 20091215053207.GA26541@thunk.org |
---|---|
State | Superseded, archived |
Headers | show |
Hello! On Dec 15, 2009, at 12:32 AM, tytso@mit.edu wrote: > Can you separate this patch into separate ones for each file system? Sure. > I think we can actually do better for ext4; for example, in the case > of data=journal or data=writeback, it's not necessary to flush the fs > data device. It's only the case of data=ordered that we need to send > a barrier. With that optimization, we do need to add a barrier in the > case of fsync() and when we are doing a journal checkpoint, but the > extra code complexity is worth not having to force a barrier for the > fs data device for every single commit, especially in the data=journal > mode with a heavy fsync workload. Indeed, this is a good idea too. I think we still can squeeze a bit more juice out of it if we can submit the data device flush right after submitting all the data, but only do the waiting for it before issuing journal device flush in normal journaling mode (we need to do the waiting before writing commit block in async journaling mode). > Do you have a test case that will allow you to easily try out this > patch, in all of ext4's various journalling modes? Thanks!! Not for vanilla kernel, but I'll see if I can construct something easily for it. > @@ -277,6 +278,16 @@ static int journal_finish_inode_data_buffers(journal_t *journal, > struct jbd2_inode *jinode, *next_i; > int err, ret = 0; > > + /* > + * If the journal is not located on the file system device, > + * then we must flush the file system device before we issue > + * the commit record > + */ > + if (commit_transaction->t_flushed_data_blocks && > + (journal->j_fs_dev != journal->j_dev) && > + (journal->j_flags & JBD2_BARRIER)) > + blkdev_issue_flush(journal->j_fs_dev, NULL); > + I am afraid this is not enough. This code is called after journal was flushed for async commit case, so it leaves a race window where journal transaction is already on disk and complete, but the data is still in cache somewhere. Also the callsite has this comment which is misleading, I think: /* * This is the right place to wait for data buffers both for ASYNC * and !ASYNC commit. If commit is ASYNC, we need to wait only after * the commit block went to disk (which happens above). If commit is * SYNC, we need to wait for data buffers before we start writing * commit block, which happens below in such setting. */ In fact it is only safe to wait here in ASYNC mode (and internal journal only) because the device was already flushed (and I hope all device drivers drain the queue too, if not, even this might be problematic) by the blkdev_issue_flush. Bye, Oleg -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, Dec 15, 2009 at 01:19:57AM -0500, Oleg Drokin wrote: > > + /* > > + * If the journal is not located on the file system device, > > + * then we must flush the file system device before we issue > > + * the commit record > > + */ > > + if (commit_transaction->t_flushed_data_blocks && > > + (journal->j_fs_dev != journal->j_dev) && > > + (journal->j_flags & JBD2_BARRIER)) > > + blkdev_issue_flush(journal->j_fs_dev, NULL); > > + > > I am afraid this is not enough. This code is called after journal > was flushed for async commit case, so it leaves a race window where > journal transaction is already on disk and complete, but the data is > still in cache somewhere. No, that's actually fine. In the ASYNC_COMMIT case, the commit won't be valid until the checksum is correct, and we won't have written any descriptor blocks yet at this point. So there is no race because during that window, the commit is written but we won't write any descriptor blocks until after the barrier returns. > Also the callsite has this comment which is misleading, I think: > /* > * This is the right place to wait for data buffers both for ASYNC > * and !ASYNC commit. If commit is ASYNC, we need to wait only after > * the commit block went to disk (which happens above). If commit is > * SYNC, we need to wait for data buffers before we start writing > * commit block, which happens below in such setting. > */ Yeah, that comment is confusing and not entirely accurate. I thought about cleaning it up, and then decided to do that in a separate patch. - Ted -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hello! On Dec 15, 2009, at 11:45 AM, tytso@mit.edu wrote: >>> + /* >>> + * If the journal is not located on the file system device, >>> + * then we must flush the file system device before we issue >>> + * the commit record >>> + */ >>> + if (commit_transaction->t_flushed_data_blocks && >>> + (journal->j_fs_dev != journal->j_dev) && >>> + (journal->j_flags & JBD2_BARRIER)) >>> + blkdev_issue_flush(journal->j_fs_dev, NULL); >>> + >> I am afraid this is not enough. This code is called after journal >> was flushed for async commit case, so it leaves a race window where >> journal transaction is already on disk and complete, but the data is >> still in cache somewhere. > No, that's actually fine. In the ASYNC_COMMIT case, the commit won't > be valid until the checksum is correct, and we won't have written any > descriptor blocks yet at this point. So there is no race because What do you mean by descriptor blocks? Actual logged blocks? > during that window, the commit is written but we won't write any > descriptor blocks until after the barrier returns. From my reading of the code, in jbd2_journal_commit_transaction: After first "JBD: commit phase 2" message we submit actual data buffers. Then after second "JBD: commit phase 2" message (in the "while (commit_transaction->t_buffers) {" loop) we actually iterate through all metadata changes and submit them to disk (after start_journal_io: label) Then journal_submit_commit_record writes commit record. At this point on we have entire transacton + commit block in the write queue (or partially already on disk) + all the data blocks in the write queue. Next blkdev_issue_flush on journal device is done, cementing the transaction on the permanent storage, but the actual data is still potentially not there abd it is not until the call to journal_finish_inode_data_buffers() where that data is flushed too. If you refer to the code after "JBD: commit phase 3" that waits on actual transaction blocks buffers to make sure they made it to the disk, my reading of blkdev_issue_flush is that it adds the flush at the end of the queue, i.e. after all of the journalled blocks that are sitting in there, so they all would be already copleted by the time blkdev_issue_flush returns. Bye, Oleg-- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/fs/ext4/fsync.c b/fs/ext4/fsync.c index 0b22497..98bd140 100644 --- a/fs/ext4/fsync.c +++ b/fs/ext4/fsync.c @@ -88,9 +88,21 @@ int ext4_sync_file(struct file *file, struct dentry *dentry, int datasync) return ext4_force_commit(inode->i_sb); commit_tid = datasync ? ei->i_datasync_tid : ei->i_sync_tid; - if (jbd2_log_start_commit(journal, commit_tid)) + if (jbd2_log_start_commit(journal, commit_tid)) { + /* + * When the journal is on a different device than the + * fs data disk, we need to issue the barrier in + * writeback mode. (In ordered mode, the jbd2 layer + * will take care of issuing the barrier. In + * data=journal, all of the data blocks are written to + * the journal device.) + */ + if (ext4_should_writeback_data(inode) && + (journal->j_fs_dev != journal->j_dev) && + (journal->j_flags & JBD2_BARRIER)) + blkdev_issue_flush(inode->i_sb->s_bdev, NULL); jbd2_log_wait_commit(journal, commit_tid); - else if (journal->j_flags & JBD2_BARRIER) + } else if (journal->j_flags & JBD2_BARRIER) blkdev_issue_flush(inode->i_sb->s_bdev, NULL); return ret; } diff --git a/fs/jbd2/checkpoint.c b/fs/jbd2/checkpoint.c index ca0f5eb..8868493 100644 --- a/fs/jbd2/checkpoint.c +++ b/fs/jbd2/checkpoint.c @@ -22,6 +22,7 @@ #include <linux/jbd2.h> #include <linux/errno.h> #include <linux/slab.h> +#include <linux/blkdev.h> #include <trace/events/jbd2.h> /* @@ -515,6 +516,20 @@ int jbd2_cleanup_journal_tail(journal_t *journal) journal->j_tail_sequence = first_tid; journal->j_tail = blocknr; spin_unlock(&journal->j_state_lock); + + /* + * If there is an external journal, we need to make sure that + * any data blocks that were recently written out --- perhaps + * by jbd2_log_do_checkpoint() --- are flushed out before we + * drop the transactions from the external journal. It's + * unlikely this will be necessary, especially with a + * appropriately sized journal, but we need this to guarantee + * correctness. Fortunately jbd2_cleanup_journal_tail() + * doesn't get called all that often. + */ + if ((journal->j_fs_dev != journal->j_dev) && + (journal->j_flags & JBD2_BARRIER)) + blkdev_issue_flush(journal->j_fs_dev, NULL); if (!(journal->j_flags & JBD2_ABORT)) jbd2_journal_update_superblock(journal, 1); return 0; diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c index 6a10238..5b78f9a 100644 --- a/fs/jbd2/commit.c +++ b/fs/jbd2/commit.c @@ -259,6 +259,7 @@ static int journal_submit_data_buffers(journal_t *journal, ret = err; spin_lock(&journal->j_list_lock); J_ASSERT(jinode->i_transaction == commit_transaction); + commit_transaction->t_flushed_data_blocks = 1; jinode->i_flags &= ~JI_COMMIT_RUNNING; wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING); } @@ -277,6 +278,16 @@ static int journal_finish_inode_data_buffers(journal_t *journal, struct jbd2_inode *jinode, *next_i; int err, ret = 0; + /* + * If the journal is not located on the file system device, + * then we must flush the file system device before we issue + * the commit record + */ + if (commit_transaction->t_flushed_data_blocks && + (journal->j_fs_dev != journal->j_dev) && + (journal->j_flags & JBD2_BARRIER)) + blkdev_issue_flush(journal->j_fs_dev, NULL); + /* For locking, see the comment in journal_submit_data_buffers() */ spin_lock(&journal->j_list_lock); list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) { diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h index f1011f7..638ce45 100644 --- a/include/linux/jbd2.h +++ b/include/linux/jbd2.h @@ -653,6 +653,7 @@ struct transaction_s * waiting for it to finish. */ unsigned int t_synchronous_commit:1; + unsigned int t_flushed_data_blocks:1; /* * For use by the filesystem to store fs-specific data