Re: [PATCH] xfs: fail dax mount if reflink is enabled on a partition

[Gotou, Yasunori/五島 康文]

On 2022/10/03 17:12, Darrick J. Wong wrote:
> On Fri, Sep 30, 2022 at 09:56:41AM +0900, Gotou, Yasunori/五島 康文 wrote:
> > Hello everyone,
> >
> > On 2022/09/20 11:38, Yang, Xiao/杨 晓 wrote:
> > > Hi Darrick, Brian and Christoph
> > >
> > > Ping. I hope to get your feedback.
> > >
> > > 1) I have confirmed that the following patch set did not change the test
> > > result of generic/470 with thin-volume. Besides, I didn't see any
> > > failure when running generic/470 based on normal PMEM device instaed of
> > > thin-volume.
> > > https://lore.kernel.org/linux-xfs/20211129102203.2243509-1-hch@xxxxxx/
> > >
> > > 2) I can reproduce the failure of generic/482 without thin-volume.
> > >
> > > 3) Is it necessary to make thin-volume support DAX. Is there any use
> > > case for the requirement?
> >
> >
> > Though I asked other place(*), I really want to know the usecase of
> > dm-thin-volume with DAX and reflink.
> >
> >
> > In my understanding, dm-thin-volume seems to provide similar feature like
> > reflink of xfs. Both feature provide COW update to reduce usage of
> > its region, and snapshot feature, right?
> >
> > I found that docker seems to select one of them (or other feature which
> > supports COW). Then user don't need to use thin-volume and reflink at same
> > time.
> >
> > Database which uses FS-DAX may want to use snapshot for its data of FS-DAX,
> > its user seems to be satisfied with reflink or thin-volume.
> >
> > So I could not find on what use-case user would like to use dm-thin-volume
> > and reflink at same time.
> >
> > The only possibility is that the user has mistakenly configured dm-thinpool
> > and reflink to be used at the same time, but if that is the case, it seems
> > to be better for the user to disable one or the other.
> >
> > I really wander why dm-thin-volume must be used with reflik and FS-DAX.
>
> There isn't a hard requirement between fsdax and dm-thinp.  The /test/
> needs dm-logwrites to check that write page faults on a MAP_SYNC
> mmapping are persisted directly to disk.  dm-logwrites requires a fast
> way to zero an entire device for correct operation of the replay step,
> and thinp is the only way to guarantee that.

Thank you for your answer. But I still feel something is strange.
Though dm-thinp may be good way to execute the test correctly,
I suppose it seems to be likely a kind of workaround to pass the test,
it may not be really required for actual users.

Could you tell me why passing test by workaround is so necessary?

Thanks,


> --D
>
> > If my understanding is something wrong, please correct me.
> >
> > (*)https://lore.kernel.org/all/TYWPR01MB1008258F474CA2295B4CD3D9B90549@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/