Re: SDHCI Regression with 6ce3dd6eec11 ("blk-mq: issue directly if hw queue isn't busy in case of 'none'")

[Ming Lei <ming.lei@redhat.com>]

On Wed, Aug 22, 2018 at 09:06:40AM +0300, Jarkko Nikula wrote:
> On 08/21/2018 04:57 PM, Ming Lei wrote:
> > On Tue, Aug 21, 2018 at 04:45:41PM +0300, Jarkko Nikula wrote:
> > > On 08/21/2018 04:03 PM, Adrian Hunter wrote:
> > > > On 21/08/18 15:37, Jarkko Nikula wrote:
> > > > > Hi
> > > > > 
> > > > > I bisected some kind of SDHCI regression to commit 6ce3dd6eec11 ("blk-mq:
> > > > > issue directly if hw queue isn't busy in case of 'none'") causing dumps
> > > > > below and one or more systemd-udevd processes being in uninterruptible sleep
> > > > > state preventing safe reboot/shutdown.
> > > > > 
> > > > > This is from an Intel Baytrail based tablet with integrated eMMC but my
> > > > > up-to-date Debian/testing rootfs (with systemd) is on USB stick.
> > > > > 
> > > > > It doesn't revert cleanly on today's head 778a33959a8a but issue is gone if
> > > > > I go to a commit before 6ce3dd6eec11 and occurs at 6ce3dd6eec11.
> > > > 
> > > > This was discussed here:
> > > > 
> > > > 	https://marc.info/?l=linux-block&m=153334717506073&w=2
> > > > 
> > > > Coincidentally, I just sent the fix patch:
> > > > 
> > > > 	https://marc.info/?l=linux-mmc&m=153485326025301&w=2
> > > > 
> > > Cool, it fixed my regression. I tested both on top of 6ce3dd6eec11 and head
> > > 778a33959a8a. Maybe you would like to add into your patch another fixes tag
> > > and my tested by:
> > > 
> > > Fixes: 6ce3dd6eec11 ("blk-mq: issue directly if hw queue isn't busy in case
> > > of 'none'")
> > 
> > If you read the above links carefully, you'd see it is wrong to add the tag of
> > 'Fixes: 6ce3dd6eec11'.
> > 
> Hmm... why? That commit 6ce3dd6eec11 was clearly regressing on my setup
> while Adrian's fix for 81196976ed94 that has been present since v4.16 fixes
> my finding too.
> 
> I don't know well enough MMC and block layer but if commit 6ce3dd6eec11
> revealed an issue from MMC under my configuration I'd call it still a
> regression.

6ce3dd6eec11 just makes the original issue to happen easier, and the
root cause is that MMC's stack doesn't support concurrent dispatch for 
SDHCI.

Thanks, 
Ming