Re: [PATCH] dm-delay: support zoned devices

[Benjamin Marzinski <bmarzins@redhat.com>]

On Wed, Mar 26, 2025 at 08:55:48AM -0400, Damien Le Moal wrote:
> On 2025/03/21 13:52, Benjamin Marzinski wrote:
> > On Fri, Mar 21, 2025 at 08:18:16AM +0100, Christoph Hellwig wrote:
> >> Add support for zoned device by passing through report_zoned to the
> >> underlying read device.
> >>
> >> This is required to make enable xfstests xfs/311 on zoned devices.
> > 
> > On suspend, delay_presuspend() stops delaying and it doesn't guarantee
> > that new bios coming in will always be submitted after the delayed bios
> > it is flushing. That can mess things up for zoned devices. I didn't
> > check if that matters for the specific test. Setting
> > 
> > ti->emulate_zone_append = true;
> > 
> > would enforce write ordering, at the expense of adding a whole other
> > layer of delays to zoned dm-delay devices. Since this isn't really
> > useful outside of testing, I think that could be acceptable if necessary
> > (it would require us to support table reloads of zoned devices with
> > emulated zone append, since tests often want to change the delay).
> > However it would probably be better to see if we can just make dm-delay
> > preserve write ordering during a suspend.
> 
> delay_presuspend() calls flush_delayed_bios() with flush_all == true. So all
> BIOs will be flushed in the order they are queued in the delay list, which as
> far as I can tell is the order in which the user of dm-delay issued the BIOs. So
> for writes, the order is preserved as far as I can tell.

delay_presuspend() is called before we set the DMF_BLOCK_IO_FOR_SUSPEND
bit, which will stop incoming bio from getting mapped, and also before
lock_fs() is called. This means it's common for new bios to continue to
come into delay_map(), while delay_presuspend() is running.  The moment
delay_presuspend() sets dc->may_delay = false, those new bios will stop
getting queued by delay_bio(). They will get remapped immeditately to
the underlying device. flush_delayed_bios() doesn't even get called
until after dc->may_delay is set to false, and if there are a lot of
bios on the delayed_bios list, flush_delayed_bios() will schedule. So,
it's actually very common for new incoming bios to get passed to
underlying device before all the bios on the dc->delayed_bios list do.

Solving this without grabbing the dc->process_bios_lock mutex for every
bio sent to dm-delay probably involves keeping the incoming bios going
to dc->delayed_bios during suspend, at least until we can guarantee that
it's empty and no bios are being flushed.

-Ben

> 
> -- 
> Damien Le Moal
> Western Digital Research