On 5/22/20 9:55 AM, Jason Gunthorpe wrote:
On Fri, May 22, 2020 at 09:29:57AM -0500, Pierre-Louis Bossart wrote:
+ ret = virtbus_register_device(vdev);
+ if (ret < 0)
+ return ret;
+
+ /* make sure the probe is complete before updating client list
*/
+ timeout = msecs_to_jiffies(SOF_CLIENT_PROBE_TIMEOUT_MS);
+ time = wait_for_completion_timeout(&cdev->probe_complete,
timeout);
This seems bonkers - the whole point of something like virtual bus is
to avoid madness like this.
Thanks for your review, Jason. The idea of the times wait here is to
make the registration of the virtbus devices synchronous so that the
SOF core device has knowledge of all the clients that have been able to
probe successfully. This part is domain-specific and it works very well
in the audio driver case.
This need to be hot plug safe. What if the module for this driver is
not available until later in boot? What if the user unplugs the
driver? What if the kernel runs probing single threaded?
It is really unlikely you can both have the requirement that things be
synchronous and also be doing all the other lifetime details properly..
Can you suggest an alternate solution then?
I don't even know what problem you are trying to solve.
The complete/wait_for_completion is a simple mechanism to tell that the
action requested by the parent is done. Absent that, we can end-up in a
situation where the probe may fail, or the requested module does not exist,
and the parent knows nothing about the failure - so the system is in a
zombie state and users are frustrated. It's not great either, is it?
Maybe not great, but at least it is consistent with all the lifetime
models and the operation of the driver core.
I agree your comments are valid ones, I just don't have a solution to be
fully compliant with these models and report failures of the driver
probe for a child device due to configuration issues (bad audio
topology, etc).
My understanding is that errors on probe are explicitly not handled in
the driver core, see e.g. comments such as:
/*
* Ignore errors returned by ->probe so that the next driver can try
* its luck.
*/
https://elixir.bootlin.com/linux/latest/source/drivers/base/dd.c#L636
If somehow we could request the error to be reported then probably we
wouldn't need this complete/wait_for_completion mechanism as a custom
notification.
This is not an hypothetical case, we've had this recurring problem when a
PCI device creates an audio card represented as a platform device. When the
card registration fails, typically due to configuration issues, the PCI
probe still completes. That's really confusing and the source of lots of
support questions. If we use these virtual bus extensions to stpo abusing
platform devices, it'd be really nice to make those unreported probe
failures go away.
I think you need to address this in some other way that is hot plug
safe.
Surely you can make this failure visible to users in some other way?
Not at the moment, no. there are no failures reported in dmesg, and the
user does not see any card created. This is a silent error.
This is probably domain-specific btw, the use of complete() is only part
of the SOF core where we extended the virtual bus to support SOF
clients. This is not a requirement in general for virtual bus users. We
are not forcing anyone to rely on this complete/wait_for_completion, and
if someone has a better idea to help us report probe failures we are all
ears.
