Re: [PATCH v3 1/3] pmem: only set QUEUE_FLAG_DAX for fsdax mode

From: Ross Zwisler
Date: Thu Jun 28 2018 - 15:04:35 EST


On Thu, Jun 28, 2018 at 05:42:34PM +0000, Kani, Toshi wrote:
> On Tue, 2018-06-26 at 16:04 -0600, Ross Zwisler wrote:
> > On Tue, Jun 26, 2018 at 02:51:52PM -0700, Dan Williams wrote:
> > > On Tue, Jun 26, 2018 at 2:31 PM, Kani, Toshi <toshi.kani@xxxxxxx> wrote:
> > > > On Tue, 2018-06-26 at 14:28 -0700, Dan Williams wrote:
> > > > > On Tue, Jun 26, 2018 at 2:23 PM, Kani, Toshi <toshi.kani@xxxxxxx> wrote:
> > > > > > On Tue, 2018-06-26 at 14:02 -0700, Dan Williams wrote:
> > > > > > > On Tue, Jun 26, 2018 at 1:54 PM, Kani, Toshi <toshi.kani@xxxxxxx> wrote:
> > > > >
> > > > > [..]
> > > > > > > > When this dm change was made, the pmem driver supported DAX for both raw
> > > > > > > > and memory modes (note: sector mode does not use the pmem driver). I
> > > > > > > > think the issue was introduced when we dropped DAX support from raw
> > > > > > > > mode.
> > > > > > >
> > > > > > > Still DAX with raw mode never really worked any way. It was also
> > > > > > > something that was broken from day one. So what happens to someone who
> > > > > > > happened to avoid all the problems with page-less DAX and enabled
> > > > > > > device-mapper on top? That failure mode detail needs to be added to
> > > > > > > this changelog if we want to propose this for -stable.
> > > > > >
> > > > > > My point is that the behavior should be consistent between pmem and
> > > > > > device-mapper. When -o dax succeeds on a pmem, then it should succeed
> > > > > > on a device-mapper on top of that pmem.
> > > > > >
> > > > > > Has the drop of dax support from raw mode made to -stable back to the
> > > > > > baseline accepted 545ed20e6df6? It will introduce inconsistency,
> > > > > > otherwise.
> > > > >
> > > > > That commit, 569d0365f571 "dax: require 'struct page' by default for
> > > > > filesystem dax", has not been tagged for -stable.
> > > >
> > > > Then, Fixes tag should be set to 569d0365f571 to keep the behavior
> > > > consistent.
> > >
> > > Sure, and the failure mode is...? I'm thinking the commit log should say:
> > >
> > > "Starting with commit 569d0365f571 "dax: require 'struct page' by
> > > default for filesystem dax", dax is no longer supported for page-less
> > > configurations. However, device-mapper sees the QUEUE_FLAG_DAX still
> > > being set and falsely assumes that DAX is enabled, this leads to
> > > <insert user visible failure mode details here>"
> >
> > Dan is correct that there is no user visible change for this. It is the right
> > thing to do for consistency and sanity, but it doesn't actually have user
> > visible behavior that needs to be backported to stable.
> >
> > Toshi is correct that this change is only for raw mode namespaces, not btt
> > namespaces.
> >
> > I'll adjust the changelog and remove the stable flag for v5, and I'll add a
> > Fixes: tag for patch 2.
>
> Hi Ross,
>
> Your patches look good. But I am still not clear about the Fixes &
> stable handling. Talking about user visible behavior, I do not think we
> had any issue until dax support was dropped from raw mode. Until then,
> the pmem driver supported dax for all modes, and the check for
> direct_access worked.

I agree that the fsdax + raw mode failure mode I mentioned in my cover letter
only started when we restricted filesystem DAX to having struct page, but I
think that the other failure mode, fsdax + some random block driver (I used
brd) was present in DM from the beginning.

In any case, I think both are fixed with the patches, and I think it's fine
that all 3 get thrown at stable. Thanks, Mike, for the help.