On Tue, Mar 09, 2021 at 07:04:34PM -0800, Asutosh Das (asd) wrote:

Hello
I & Can (thanks CanG) debugged this further:

Looks like this issue can occur if the sd probe is asynchronous.

Essentially, the sd_probe() is done asynchronously and driver_probe_device()
invokes pm_runtime_get_suppliers() before invoking sd_probe().

But scsi_probe_and_add_lun() runs in a separate context.
So the scsi_autopm_put_device() invoked from scsi_scan_host() context
reduces the link->rpm_active to 1. And sd_probe() invokes
scsi_autopm_put_device() and starts a timer. And then driver_probe_device()
invoked from __device_attach_async_helper context reduces the
link->rpm_active to 1 thus enabling the supplier to suspend before the
consumer suspends.

I don't see a way around this. Please let me know if you
(@Alan/@Bart/@Adrian) have any thoughts on this.

How about changing the SCSI core so that it does a runtime_get before
starting an async probe, and the async probe routine does a
runtime_put when it is finished?  In other words, don't allow a device
to go into runtime suspend while it is waiting to be probed.

I don't think that would be too intrusive.

Alan Stern