Re: [PATCH] scsi: fix hung_task when change host from recovery to running via sysfs

[Benjamin Block <bblock@linux.ibm.com>]

On Wed, Mar 22, 2023 at 09:24:32AM +0800, yebin (H) wrote:
> On 2023/3/21 22:22, Benjamin Block wrote:
> > On Tue, Mar 21, 2023 at 04:42:04PM +0800, Ye Bin wrote:
> >> From: Ye Bin <yebin10@huawei.com>
> >>
> >> When do follow test:
> >> Step1: echo  "recovery" > /sys/class/scsi_host/host0/state
> >
> > Hmm, that make me wonder, what potential use-case this is for? Just
> > testing?
>
> Thank you for your reply.
> Actually, I'm looking for a way to temporarily stop sending IO to the
> driver.
> Setting the state of the host to recovery can do this, but I changed
> the state to running and found that the process could not be woken up.
> I don't know what the purpose of designing this sysfs interface was.
> But this modification can solve the effect I want to achieve.

My first thought when seeing this was that maybe we should also limit
this interface to say `SHOST_RUNNING` and `SHOST_RECOVERY` (similar to
what is done in `store_state_field()`).
    That would limit the amount of corner cases drastically.

And in case of setting `SHOST_RUNNING` after the scsi host was set to
`SHOST_RECOVERY`, we could also make use of the already existing
function `scsi_restart_operations()` to handle the restart in the same
way as EH does.

But it's not up to me, to make that call.

> > For SDEVs we explicitly filter what states can be set from user-space.
> > Only `SDEV_RUNNING` and `SDEV_OFFLINE` can be set in
> > `store_state_field()`.
> >      There is probably quite a few other bad things you can do with this
> > interface by using any of the other states used for device destruction
> > or EH, and then trigger I/O or said destruction/EH otherwise.
> >      Not sure handling this one special case of `SHOST_RECOVERY` is quite
> > enough.
> >
> >> diff --git a/drivers/scsi/scsi_sysfs.c b/drivers/scsi/scsi_sysfs.c
> >> index ee28f73af4d4..ae6b1476b869 100644
> >> --- a/drivers/scsi/scsi_sysfs.c
> >> +++ b/drivers/scsi/scsi_sysfs.c
> >> @@ -216,6 +216,9 @@ store_shost_state(struct device *dev, struct device_attribute *attr,
> >>
> >>        if (scsi_host_set_state(shost, state))
> >>                return -EINVAL;
> >> +     else
> >> +             wake_up(&shost->host_wait);

In the very least, this should first check whether we really just made
the transition from `SHOST_RECOVERY` to `SHOST_RUNNING` before calling
this `wake_up()`.
    And for that - first get old state, then set the new state - we
probably would also need to grab the `host_lock` to make that race free.

Just calling `wake_up()` without knowing what state transition we just
made doesn't sound right to me.

> >> +
> >>        return count;
> >>   }
> >>
> >> --
> >> 2.31.1
> >>
> 

-- 
Best Regards, Benjamin Block        /        Linux on IBM Z Kernel Development
IBM Deutschland Research & Development GmbH    /   https://www.ibm.com/privacy
Vors. Aufs.-R.: Gregor Pillen         /         Geschäftsführung: David Faller
Sitz der Ges.: Böblingen     /    Registergericht: AmtsG Stuttgart, HRB 243294