Hi,
On Thu 28-01-10 09:23:43, 丁定华 wrote:
> As you wrote, if T2!=T1, then T2 is committing transaction while T1
> is running transaction, and if T1 complete commit, we don't care
> about the content of buffers. But there is a prerequisite -->"T1
> complete commit", if T1 start commit and another transaction T3
> becomes the new running transaction, T3 may need to reuse T2 log
> space and force checkpoint, and since we have clean the BH_dirty bit
> of buffers after T2 commits, so T2 may be freed before T1 complete
> commit, and unfortunately, T1 doesn't complete commit, so after
> replay, updates of T2 get lost, fs becomes inconsistent.
Hmm, you are right. That could probably happen. It only hits ext3/4 in
data=journaled mode but the bug is there. But it's hard to fix it in a
reasonable way - if we would just leave the dirty bit set, we will have
aliasing problems - buffer B is attached to some page which used to be from
file F, so unmap_underlying_metadata will not find it because it looks only
into page cache the block device, not to the pagecaches of individual
files. So if we reallocate the block of B for some other file G before the
buffer B is checkpointed, we have two dirty buffers representing the same
block and thus data corruption can happen.
Maybe we could handle them by setting b_next_transaction to the
transaction that deleted the buffer (in jbd2_journal_unmap_buffer) and
setting buffer freed like we do now. Commit code would handle freed
buffers like:
If b_next_transaction is set, file buffer to forget list of the next
transaction. If b_next_transaction isn't set, clear all dirty bits.
What do you think?
Honza
> 2010/1/27 Jan Kara <jack@xxxxxxx>:
> > Hi,
> >
> > On Wed 27-01-10 10:32:18, 丁定华 wrote:
> >> I'm a little confused about BH_Freed bit. The only place it is set
> >> is journal_unmap_buffer, which is called by jbd2_journal_invalidatepage when
> >> we want to truncate a file. Since jbd2_journal_invalidatepage is called
> >> outside of transaction, We can't make sure whether the "add to orphan"
> >> operation belongs to committing transaction or not, so we can't touch the
> >> buffer belongs to committing transaction, instead BH_Freed bit is set to
> >> indicate that this buffer can be discarded in running transaction. But i
> >> think we shouldn't clear BH_JBDdirty in jbd2_journal_commit_transaction, as
> >> following codes does:
> >> /* A buffer which has been freed while still being
> >> * journaled by a previous transaction may end up still
> >> * being dirty here, but we want to avoid writing back
> >> * that buffer in the future now that the last use has
> >> * been committed. That's not only a performance gain,
> >> * it also stops aliasing problems if the buffer is left
> >> * behind for writeback and gets reallocated for another
> >> * use in a different page. */
> >> if (buffer_freed(bh)) {
> >> clear_buffer_freed(bh);
> >> clear_buffer_jbddirty(bh);
> >> }
> >> Note that, *We can't make sure "current running transaction" can complete
> >> commit work.* If we clear BH_JBDdirty bit here, this buffer may be freed
> >> here, the log space of older transaction may be freed before the "current
> >> running transaction" complete commit work, and if this happends, filesystem
> >> will be inconsistent.
> > Let me sketch the situation here:
> > The file F gets truncated. The inode is added to orphan list in some
> > transaction T1, only then jbd2_journal_invalidatepage can be called.
> > As you wrote above, it can happen that jbd2_journal_invalidatepage on
> > buffer B runs when some transaction T2 containing B is being committed and
> > in that case we set BH_Freed. If T2 != T1 - i.e., T2 is being committed
> > and T1 is the running transaction, note that we clear the dirty bit only
> > when T2 is fully committed and we are processing forget list. So buffer has
> > been properly written to T2 and we just won't write it in the transaction
> > T1. And that is fine because as soon as transaction T1 finishes commit, we
> > don't care about what happens with buffers of F because the fact that F is
> > truncated is recorded and in case of crash we finish truncate during
> > journal replay. And if we crash before T1 finishes commit, we don't care
> > about contents of T1 either. If T2 == T1, the above reasoning applies as
> > well and the situation is even simpler.
> >
> > Honza
> > --
> > Jan Kara <jack@xxxxxxx>
> > SUSE Labs, CR
> >
>
>
>
> --
> 丁定华
--
Jan Kara <jack@xxxxxxx>
SUSE Labs, CR
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
