RE: About a possibility of long latency to claim host

["Kiwoong Kim" <kwmad.kim@samsung.com>]

> On Tue, 9 Mar 2021 at 14:07, Ulf Hansson <ulf.hansson@linaro.org> wrote:
> >
> > On Tue, 9 Mar 2021 at 05:11, Kiwoong Kim <kwmad.kim@samsung.com> wrote:
> > >
> > > Dear All
> > >
> > > I want to discuss about this topic with you guys.
> > >
> > > There is an application to put many IO requests to a SD card right
> after completion of system resume.
> > > Current MMC stack invokes mmc_rescan asynchronously for
> PM_POST_SUSPEND.
> > > As reported to me, there could be a race between an IO thread and
> > > the kworker for mmc_rescan, especially when the application
> > > mentioned before is installed and the function of mmc_rescan is run
> later than expected For a series of IO requests, mmc_rescan, particularily
> at __mmc_claim_host called in mmc_sd_detect, didn't acquire a host for
> longer than expected.
> >
> > That's not a problem, in general.
> >
> > As long as the card is inserted and functional, the I/O operations
> > should be completed successfully. It doesn't matter if mmc_rescan() is
> > waiting to claim the host, as it's not important that it gets to run
> > as long as the card remains inserted/functional.
> >
> > >
> > > Below is the call stacks shown after the symptom happened and the
> system tried to enter into suspend again.
> > > In this case, mmc pm notifier is called with PM_SUSPEND_PREPARE, so it
> waits for completion or cancelling the work for mmc_rescan.
> > > For the latency, mobile users can see black screen for a long time
> sometimes, even with pushing a power button to wake up the system.
> >
> > I think I understand what you are saying, but please correct me if I am
> wrong.
> >
> > The I/O requests keep flowing into the blk queue even after
> > PM_SUSPEND_PREPARE has been fired, thus preventing the earlier
> > scheduled mmc_rescan() from claiming the host?
> >
> > This sounds quite reasonable that it could happen, at least
> > theoretically. Although, I am a bit surprised that nobody has reported
> > about this problem, until now. The design in the mmc core, has
> > remained unchanged in regards to this behaviour, for a very very long
> > time.
> >
> > Let me try to reproduce the problem, I will get back to you soon. In
> > the meantime, I would also appreciate it if you could share, more
> > exactly, how to trigger this problem at your side.
> 
> I have managed to reproduce the problem!  Wow, I wonder how this could
> have slipped through without anybody reporting about this, until now.
> So, thanks for bringing this to my attention!
> 
> My rough guess is that SD cards, used in these kinds of configurations
> (Android), are actually set to non removable. This would prevent
> mmc_rescan() from claiming the host.
> 
> In any case, this needs to be fixed properly. I will continue to work on a
> solution and get back to you with a patch. 

Okay and I'll also thing about its fix.

Thanks.
Kiwoong Kim

> 
> [...]
> 
> Kind regards
> Uffe