Re: regression introduced by "block: Add support for DAX reads/writes to block devices"

From: Jeff Moyer <jmoyer@redhat.com>
To: Linda Knippers <linda.knippers@hp.com>
Cc: Boaz Harrosh <boaz@plexistor.com>, "Wilcox\,
	Matthew R" <matthew.r.wilcox@intel.com>,
	"linux-kernel\@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"linux-fsdevel\@vger.kernel.org" <linux-fsdevel@vger.kernel.org>,
	"Vishal L. Verma" <vishal.l.verma@intel.com>
Subject: Re: regression introduced by "block: Add support for DAX reads/writes to block devices"
Date: Thu, 13 Aug 2015 14:19:05 -0400	[thread overview]
Message-ID: <x49mvxv3vc6.fsf@segfault.boston.devel.redhat.com> (raw)
In-Reply-To: <55CCD94D.7040807@hp.com> (Linda Knippers's message of "Thu, 13 Aug 2015 13:52:13 -0400")

Linda Knippers <linda.knippers@hp.com> writes:

>>> It causes the physical block size to be PAGE_SIZE but the
>>> logical block size is still 512.  However, the minimum_io_size
>>> is now 4096 (same as physical block size, I assume).  The
>>> optimal_io_size is still 0.  What does that mean?
>> 
>> physical block size - device's internal block size
>> logical block size - addressable unit
>
> Right, but it's still reported as 512 and that doesn't work.

Understood.  :)

>> optimal io size - device's preferred unit for streaming
>
> So 0 is ok.

Correct.

>> We can change the block device to export logical/physical block sizes of
>> PAGE_SIZE.  However, when persistent memory support comes to platforms
>> that support page sizes > 32k, xfs will again run into problems (Dave
>> Chinner mentioned that xfs can't deal with logical block sizes >32k.)
>> Arguably, you can use pmem and dax on such platforms using RAM today for
>> testing.  Do we care about breaking that?
>
> I would think so.  AARCH64 uses 64k pages today.

So does powerpc, but I guess nobody cares about that anymore.  ;-) If
the logical block size is smaller than the page size, we're going to
have to deal with sub-page I/O.  For now, we can do as Boaz suggested,
and just turn off dax for those configurations.  We could also just
revert the patch that introduced this problem.  I really don't know who
is going to care about O_DIRECT I/O performance to a persistent memory
block device.

Willy?  What was the real motivation there?

> I think Documentation/filesystems/dax.txt could use a little update
> too.  It has a section "Implementation Tips for Block Driver Writers"
> that makes it sound easy but now I wonder if it even works with the
> example ram drivers.  Should we be able to read any 512 byte
> "sector"?

If the logical block size is 512 bytes, then you have to be able to do
(direct) I/O to any 512 byte sector.  Simple as that.

Cheers,
Jeff