Re: [PATCH v2 3/7] dm: fix test for DAX device support

From: Mike Snitzer
Date: Fri Jun 01 2018 - 16:46:43 EST


On Fri, Jun 01 2018 at 4:19P -0400,
Mike Snitzer <snitzer@xxxxxxxxxx> wrote:

> On Tue, May 29 2018 at 3:51P -0400,
> Ross Zwisler <ross.zwisler@xxxxxxxxxxxxxxx> wrote:
>
> > Currently device_supports_dax() just checks to see if the QUEUE_FLAG_DAX
> > flag is set on the device's request queue to decide whether or not the
> > device supports filesystem DAX. This is insufficient because there are
> > devices like PMEM namespaces in raw mode which have QUEUE_FLAG_DAX set but
> > which don't actually support DAX.
>
> Isn't that a PMEM bug then?
>
> What is the point of setting QUEUE_FLAG_DAX if it cannot be trusted?
>
> > This means that you could create a dm-linear device, for example, where the
> > first part of the dm-linear device was a PMEM namespace in fsdax mode and
> > the second part was a PMEM namespace in raw mode. Both DM and the
> > filesystem you put on that dm-linear device would think the whole device
> > supports DAX, which would lead to bad behavior once your raw PMEM namespace
> > part using DAX needed struct page for something.
>
> The PMEM namespace in raw mode shouldn't be setting QUEUE_FLAG_DAX, if
> it didn't then the stacked-up linear DM wouldn't
>
> > Fix this by using bdev_dax_supported() like filesystems do at mount time.
> > This checks for raw mode and also performs other tests like checking to
> > make sure the dax_direct_access() path works.
>
> Sorry "This" does those things where?

I see you meant bdev_dax_supported() does these additional checks.

My previous question stands though. Why is QUEUE_FLAG_DAX getting set
if the device hasn't already passed these checks? Shouldn't setting
QUEUE_FLAG_DAX on request_queue depend on bdev_dax_supported() passing?

But looking at the drivers that do set QUEUE_FLAG_DAX: they
don't have the bdev readily available. Anyway, just strikes me as
bizarre that a driver can set QUEUE_FLAG_DAX without having to have
ensured bdev_dax_supported() passes (even if not programatically, but
that the developer has verified the hooks, et al exist).

But I'll give up on this line of questioning..

My dilemma now is: how do I take these changes without first rebasing
linux-dm.git ontop of Darrick's xfs tree?

I probably should've reviewed faster and been the one to take the entire
set (with appropriate acks obviously).