| Message ID | 545C6F67.9090600@huawei.com |
|---|---|
| State | Accepted, archived |
| Headers | show |
On Fri 07-11-14 15:06:15, alex chen wrote: > During commit process, journal_finish_inode_data_buffers will flush plug > list as follow: > jbd2_journal_commit_transaction > ->journal_finish_inode_data_buffers > ->filemap_fdatawait_range > ->wait_on_page_bit > ->__wait_on_bit > ->sleep_on_page > ->io_schedule > ->blk_flush_plug_list > > When ASYNC_COMMIT feature is set, the journal_finish_inode_data_buffers > separates the commit blocks from the rest of the journal blocks. So we > should finish inode data buffers immediately after submiting data > buffers, this allow most of the journal blocks to be written in a single > I/O operation and improve journal commit performance. So the combination of ASYNC_COMMIT and data=ordered mode is broken - it can happen that all journal writes make it to stable storage while data writes don't on power failure thus exposing stale data on next boot. So optimizing this combination is futile - we rather have to make sure we don't allow users to run such combination. I'll look into it tomorrow. Honza > > Signed-off-by: Alex Chen <alex.chen@huawei.com> > --- > fs/jbd2/commit.c | 20 ++++++++++---------- > 1 file changed, 10 insertions(+), 10 deletions(-) > > diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c > index b73e021..dc33d89 100644 > --- a/fs/jbd2/commit.c > +++ b/fs/jbd2/commit.c > @@ -555,6 +555,16 @@ void jbd2_journal_commit_transaction(journal_t *journal) > if (err) > jbd2_journal_abort(journal, err); > > + err = journal_finish_inode_data_buffers(journal, commit_transaction); > + if (err) { > + printk(KERN_WARNING > + "JBD2: Detected IO errors while flushing file data " > + "on %s\n", journal->j_devname); > + if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR) > + jbd2_journal_abort(journal, err); > + err = 0; > + } > + > blk_start_plug(&plug); > jbd2_journal_write_revoke_records(journal, commit_transaction, > &log_bufs, WRITE_SYNC); > @@ -752,16 +762,6 @@ start_journal_io: > } > } > > - err = journal_finish_inode_data_buffers(journal, commit_transaction); > - if (err) { > - printk(KERN_WARNING > - "JBD2: Detected IO errors while flushing file data " > - "on %s\n", journal->j_devname); > - if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR) > - jbd2_journal_abort(journal, err); > - err = 0; > - } > - > /* > * Get current oldest transaction in the log before we issue flush > * to the filesystem device. After the flush we can be sure that > -- > 1.8.4.3
Hi, Jan Kara On 2014/11/14 6:15, Jan Kara wrote: > On Fri 07-11-14 15:06:15, alex chen wrote: >> During commit process, journal_finish_inode_data_buffers will flush plug >> list as follow: >> jbd2_journal_commit_transaction >> ->journal_finish_inode_data_buffers >> ->filemap_fdatawait_range >> ->wait_on_page_bit >> ->__wait_on_bit >> ->sleep_on_page >> ->io_schedule >> ->blk_flush_plug_list >> >> When ASYNC_COMMIT feature is set, the journal_finish_inode_data_buffers >> separates the commit blocks from the rest of the journal blocks. So we >> should finish inode data buffers immediately after submiting data >> buffers, this allow most of the journal blocks to be written in a single >> I/O operation and improve journal commit performance. > So the combination of ASYNC_COMMIT and data=ordered mode is broken - it > can happen that all journal writes make it to stable storage while data > writes don't on power failure thus exposing stale data on next boot. So > optimizing this combination is futile - we rather have to make sure we > don't allow users to run such combination. I'll look into it tomorrow. > > Honza Thanks for your replay. In this patch, we will finish inode data buffers immediately after submiting data buffers, then write metadata buffer and journal blocks. So it can't happen that all journal blocks flush to disk while data don't on power failure. >> >> Signed-off-by: Alex Chen <alex.chen@huawei.com> >> --- >> fs/jbd2/commit.c | 20 ++++++++++---------- >> 1 file changed, 10 insertions(+), 10 deletions(-) >> >> diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c >> index b73e021..dc33d89 100644 >> --- a/fs/jbd2/commit.c >> +++ b/fs/jbd2/commit.c >> @@ -555,6 +555,16 @@ void jbd2_journal_commit_transaction(journal_t *journal) >> if (err) >> jbd2_journal_abort(journal, err); >> >> + err = journal_finish_inode_data_buffers(journal, commit_transaction); >> + if (err) { >> + printk(KERN_WARNING >> + "JBD2: Detected IO errors while flushing file data " >> + "on %s\n", journal->j_devname); >> + if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR) >> + jbd2_journal_abort(journal, err); >> + err = 0; >> + } >> + >> blk_start_plug(&plug); >> jbd2_journal_write_revoke_records(journal, commit_transaction, >> &log_bufs, WRITE_SYNC); >> @@ -752,16 +762,6 @@ start_journal_io: >> } >> } >> >> - err = journal_finish_inode_data_buffers(journal, commit_transaction); >> - if (err) { >> - printk(KERN_WARNING >> - "JBD2: Detected IO errors while flushing file data " >> - "on %s\n", journal->j_devname); >> - if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR) >> - jbd2_journal_abort(journal, err); >> - err = 0; >> - } >> - >> /* >> * Get current oldest transaction in the log before we issue flush >> * to the filesystem device. After the flush we can be sure that >> -- >> 1.8.4.3 -- 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 Fri 21-11-14 17:43:19, alex chen wrote: > Hi, Jan Kara > > On 2014/11/14 6:15, Jan Kara wrote: > > On Fri 07-11-14 15:06:15, alex chen wrote: > >> During commit process, journal_finish_inode_data_buffers will flush plug > >> list as follow: > >> jbd2_journal_commit_transaction > >> ->journal_finish_inode_data_buffers > >> ->filemap_fdatawait_range > >> ->wait_on_page_bit > >> ->__wait_on_bit > >> ->sleep_on_page > >> ->io_schedule > >> ->blk_flush_plug_list > >> > >> When ASYNC_COMMIT feature is set, the journal_finish_inode_data_buffers > >> separates the commit blocks from the rest of the journal blocks. So we > >> should finish inode data buffers immediately after submiting data > >> buffers, this allow most of the journal blocks to be written in a single > >> I/O operation and improve journal commit performance. > > So the combination of ASYNC_COMMIT and data=ordered mode is broken - it > > can happen that all journal writes make it to stable storage while data > > writes don't on power failure thus exposing stale data on next boot. So > > optimizing this combination is futile - we rather have to make sure we > > don't allow users to run such combination. I'll look into it tomorrow. > > > > Honza > > Thanks for your replay. In this patch, we will finish inode data buffers > immediately after submiting data buffers, then write metadata buffer and > journal blocks. So it can't happen that all journal blocks flush to disk > while data don't on power failure. It can. Until you flush disk's write caches disk it can happen that the IO is lost although it was reported as completed... So in this case, although we called journal_finish_inode_data_buffers() data needn't be stored on permanent storage until we send a cache flush. And that happens only after writing the commit block of the transaction (when async_commit is enabled). So the disk is free to first store the commit block of the transaction and only after that the data... Honza > >> > >> Signed-off-by: Alex Chen <alex.chen@huawei.com> > >> --- > >> fs/jbd2/commit.c | 20 ++++++++++---------- > >> 1 file changed, 10 insertions(+), 10 deletions(-) > >> > >> diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c > >> index b73e021..dc33d89 100644 > >> --- a/fs/jbd2/commit.c > >> +++ b/fs/jbd2/commit.c > >> @@ -555,6 +555,16 @@ void jbd2_journal_commit_transaction(journal_t *journal) > >> if (err) > >> jbd2_journal_abort(journal, err); > >> > >> + err = journal_finish_inode_data_buffers(journal, commit_transaction); > >> + if (err) { > >> + printk(KERN_WARNING > >> + "JBD2: Detected IO errors while flushing file data " > >> + "on %s\n", journal->j_devname); > >> + if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR) > >> + jbd2_journal_abort(journal, err); > >> + err = 0; > >> + } > >> + > >> blk_start_plug(&plug); > >> jbd2_journal_write_revoke_records(journal, commit_transaction, > >> &log_bufs, WRITE_SYNC); > >> @@ -752,16 +762,6 @@ start_journal_io: > >> } > >> } > >> > >> - err = journal_finish_inode_data_buffers(journal, commit_transaction); > >> - if (err) { > >> - printk(KERN_WARNING > >> - "JBD2: Detected IO errors while flushing file data " > >> - "on %s\n", journal->j_devname); > >> - if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR) > >> - jbd2_journal_abort(journal, err); > >> - err = 0; > >> - } > >> - > >> /* > >> * Get current oldest transaction in the log before we issue flush > >> * to the filesystem device. After the flush we can be sure that > >> -- > >> 1.8.4.3 >
diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c index b73e021..dc33d89 100644 --- a/fs/jbd2/commit.c +++ b/fs/jbd2/commit.c @@ -555,6 +555,16 @@ void jbd2_journal_commit_transaction(journal_t *journal) if (err) jbd2_journal_abort(journal, err); + err = journal_finish_inode_data_buffers(journal, commit_transaction); + if (err) { + printk(KERN_WARNING + "JBD2: Detected IO errors while flushing file data " + "on %s\n", journal->j_devname); + if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR) + jbd2_journal_abort(journal, err); + err = 0; + } + blk_start_plug(&plug); jbd2_journal_write_revoke_records(journal, commit_transaction, &log_bufs, WRITE_SYNC); @@ -752,16 +762,6 @@ start_journal_io: } } - err = journal_finish_inode_data_buffers(journal, commit_transaction); - if (err) { - printk(KERN_WARNING - "JBD2: Detected IO errors while flushing file data " - "on %s\n", journal->j_devname); - if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR) - jbd2_journal_abort(journal, err); - err = 0; - } - /* * Get current oldest transaction in the log before we issue flush * to the filesystem device. After the flush we can be sure that
During commit process, journal_finish_inode_data_buffers will flush plug list as follow: jbd2_journal_commit_transaction ->journal_finish_inode_data_buffers ->filemap_fdatawait_range ->wait_on_page_bit ->__wait_on_bit ->sleep_on_page ->io_schedule ->blk_flush_plug_list When ASYNC_COMMIT feature is set, the journal_finish_inode_data_buffers separates the commit blocks from the rest of the journal blocks. So we should finish inode data buffers immediately after submiting data buffers, this allow most of the journal blocks to be written in a single I/O operation and improve journal commit performance. Signed-off-by: Alex Chen <alex.chen@huawei.com> --- fs/jbd2/commit.c | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-)