Re: SCSI layer RPM deadlock debug suggestion

On 05/07/2021 00:45, Bart Van Assche wrote:

Hi Alan and Bart,

Thanks for the suggestions.

quoted

quoted

Removing commit e27829dc92e5 ("scsi: serialize ->rescan against ->remove")
solves this issue for me, but that is there for a reason.

Any suggestion on how to fix this deadlock?

This is indeed a tricky question.  It seems like we should allow a
runtime resume to succeed if the only reason it failed was that the
device has been removed.

More generally, perhaps we should always consider that a runtime
resume succeeds.  Any remaining problems will be dealt with by the
device's driver and subsystem once the device is marked as
runtime-active again.

Suppose you try changing blk_post_runtime_resume() so that it always
calls blk_set_runtime_active() regardless of the value of err.  Does
that fix the problem?

And more importantly, will it cause any other problems...?

That would cause trouble for the UFS driver and other drivers for which
runtime resume can fail due to e.g. the link between host and device
being in a bad state.

How about checking the SCSI device state inside scsi_rescan_device() and
skipping the rescan if the SCSI device state is SDEV_CANCEL or SDEV_DEL?

I find that the device state is SDEV_RUNNING for me at that point (so it 
cannot work).

Adding such a check inside __scsi_execute() would break sd_remove() and
sd_shutdown() since both use __scsi_execute() to submit a SYNCHRONIZE
CACHE command to the device.

Could we somehow signal from __scsi_remove_device() earlier that the 
request queue is dying or at least in some error state, so that 
blk_queue_enter() in the rescan can fail?

Currently we don't call blk_cleanup_queue() -> blk_set_queue_dying() 
until after the device_del(sdev_gendev) call in __scsi_remove_device().

Thanks,
John