Re: [PATCH v3 13/15] block, dax: make dax mappings opt-in by default

[Dan Williams <dan.j.williams@intel.com>]

On Tue, Nov 3, 2015 at 12:20 PM, Dave Chinner <david@fromorbit.com> wrote:
> On Mon, Nov 02, 2015 at 11:35:04PM -0800, Dan Williams wrote:
>> On Mon, Nov 2, 2015 at 4:32 PM, Dave Chinner <david@fromorbit.com> wrote:
[..]
>> Only in the mmap path:
>
> which means blkdev_direct_IO() is now always going to go down the
> dax_do_io() path for any driver with a ->direct_access method rather
> than the direct IO path, regardless of whether DAX is enabled on the
> device or not.
>
> That really seems wrong to me - you've replace explicit "is DAX
> enabled" checks with "is DAX possible" checks, and so DAX paths are
> used regardless of whether DAX is enabled or not. And it's not
> obvious why this is done, nor is it now obvious how DAX interacts
> with the block device.
>
> This really seems like a step backwards to me.

I think the reason it is not obvious is the original justification for
the bypass as stated in commit bbab37ddc20b "block: Add support for
DAX reads/writes to block devices" was:

    "instead of allocating a DIO and a BIO"

It turns out it's faster and as far as I can tell semantically
equivalent to the __blockdev_direct_IO() path.  The DAX mmap path in
comparison has plenty of sharp edges and semantic differences that
would be avoided by turning off DAX.

I'm not opposed to also turning off dax_do_io() when S_DAX is clear,
but I don't currently see the point.  At the very least I need to add
the above comments to the code, but do you still think opt-in DAX is a
backwards step?