On Wed, 23 Mar 2011 12:32:03 -0500
Steve French <smfrench@xxxxxxxxx> wrote:
> On Wed, Mar 23, 2011 at 11:14 AM, Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> > On Wed, 23 Mar 2011 12:27:31 +0300
> > Pavel Shilovsky <piastry@xxxxxxxxxxx> wrote:
> >
> >> 2011/3/23 Steve French <smfrench@xxxxxxxxx>:
> >> > Pavel,
> >> > Jeff and I talked about invalidate_mapping and its uses this
> >> > afternoon. ÂA few comments:
> >> > 1) Â Take a look at what happens on filemap_fdatawait call in
> >> > invalidate_mapping to see if it makes sense to return errors through
> >> > back to some of the callers of invalidate_mapping, in particular,
> >> > strict_fsync. ÂIf we can't write out the file data (ENOSPC, or host
> >> > down etc.), we want to make sure that the return code gets sent back
> >> > on any calls that can reasonably expect such an error.
> >>
> >> vfs_fsync_range has already done filemap_fdatawait call - so, in this
> >> case there is no need to do it again in cifs_invalidate_mapping. The
> >> only reason for calling this is to invalidate_inode_pages2 but any
> >> error there shouldn't affect fsync behavior, as I think.
> >>
> >> > 2) If invalidate_inode_pages2 fails (e.g. with EBUSY, because one of
> >> > the pages couldn't be freed from the mapping because it just got
> >> > redirtied right after we flushed it the line before) we set the
> >> > mapping to invalid but don't check
> >> > Â Âcifs_i->invalid_mapping
> >> > in many places. ÂShould we add checks for cifs_i->invalid_mapping in
> >> > more places?
> >>
> >> The one place where we should add such a check is read call, but it
> >> needs cifs_revalidate_file instead (that I am going to provide next)
> >> before generic_file_aio_read. In this case cifs_revalidate will check
> >> for invalid_mapping and needs to return a error if we could not
> >> invalidate all inode pages in cifs_invalidate_mapping (because it
> >> returns wrong data to the read call). But as you noticed lseek
> >> shouldn't think about this error and it uses cifs_revalidate_file too.
> >>
> >> So, we may add extra check for -EBUSY error code in callers of
> >> cifs_revalidate_{dentry,file} and cifs_invaliadate_mapping and
> >> separate them into two groups:
> >> 1) that aware about -EBUSY error code and return a error the it's caller.
> >> these are: cifs_d_revalidate, cifs_file_aio_read (future
> >> implementation), cifs_file_strict_mmap.
> >> 2) that doesn't aware about it and return ok in this case.
> >> these are: cifs_getattrs, cifs_lseek, cifs_fsync.
> >>
> >> Thoughts?
> >>
> >
> > The main reason I think we need to reconsider that error is that I
> > spent several months tracking down a rather nasty data corruption bug
> > relating to mmap on NFS in RHEL5 a few years ago:
> >
> > Â Âhttps://bugzilla.redhat.com/show_bug.cgi?id=435291
> >
> > Part of the problem there was that the NFS client ignored the return
> > code from invalidate_inode_pages2. The other part of the problem was a
> > lack of synchronization between mmap calls and the page fault handler.
> >
> > Needless to say, this was not a fun problem to track down. I think you
> > need to be very careful about ignoring errors from
> > invalidate_inode_pages2 and think carefully about what a failure there
> > means for all cases.
>
> I agree - need to be careful, but IIRC the NFS problem would be that
> it has a launder_page method which is returning an error through
> invalidate_inode_pages2 while in the cifs case the data is forced to
> be written out through filemap_fdatawrite or filemap_write_and_wait.
>
No. RHEL5 at that time didn't have a launder page method (it does now,
but that's another story).
--
Jeff Layton <jlayton@xxxxxxxxxx>
--
To unsubscribe from this list: send the line "unsubscribe linux-cifs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
