Re: [PATCH 0/8] Make SCMI transport as standalone drivers

[Cristian Marussi <cristian.marussi@arm.com>]

On Tue, Jul 23, 2024 at 01:36:55PM +0000, Etienne CARRIERE wrote:
> Hi Cristian, Peng,
> 

Hi Etienne,

Thanks for giving this a go on your setup.

> On Thursday, July 11, 2024, Cristian Marussi worte:
> > On Thu, Jul 11, 2024 at 01:26:16PM +0000, Peng Fan wrote:
> > > > Subject: [PATCH 0/8] Make SCMI transport as standalone drivers
> > >
> > > You may need use V2 here :)
> > 
> > ...oh damn :P
> > 
> > > >
> > > > Hi all,
> > > >
> > > > Till now the SCMI transport layer was being built embedded into in the
> > > > core SCMI stack.
> > > >
> > > > Some of these transports, despite being currently part of the main
> > > > SCMI module, are indeed also registered with different subsystems like
> > > > optee or virtio, and actively probed also by those: this led to a few
> > > > awkward and convoluted tricks to properly handle such interactions at
> > > > boot time in the SCMI stack.
> > > >
> > > > Moreover some partner expressed the desire to be able to fully
> > > > modularize the transports components.
> > > >
> > > > This series aim to make all such transports as standalone drivers that
> > > > can be optionally loaded as modules.
> > > >
> > > > In order to do this, at first some new mechanism is introduced to
> > > > support this new capability while maintaining, in parallel, the old
> > > > legacy embedded transports; then each transport, one by one, is
> > > > transitioned to be a standalone driver and finally the old legacy
> > > > support for embedded transport is removed.
> > > >
> > > > Patch [1/8] is a mostly unrelated (but much needed) clean-up from
> > > > Peng, which I included in this series to avoid conflicts at merge.
> > > >
> > > > Patch [2/8] simply collects the existing datagram manipulation helpers
> > > > in a pair of function pointers structures, in preparation for later reworks.
> > > >
> > > > Patch [3/8] adds the bulk of the new logic to the core SCMI stack and
> > > > then each existing transport is transitioned to be a standalone driver in
> > > > patches 4,5,6,7 while shuffling around the compatibles. (no DT change
> > > > is needed of curse for backward compatibility) While doing this I kept
> > > > the module authorship in line with the main
> > > > author(S) as spitted out by git-blame.
> > > >
> > > > Finally patch [8/8] removes all the legacy dead code from the core
> > > > SCMI stack.
> > > >
> > > > No new symbol EXPORT has been added.
> > > >
> > > > The new transport drivers have been tested, as built-in and LKM, as
> > > > follows:
> > > >
> > > > - mailbox on JUNO
> > > > - virtio on emulation
> > > > - optee on QEMU/optee using Linaro setup
> > > >
> > > > Exercised using the regular SCMI drivers stack and the SCMI ACS suite,
> > > > testing commands, replies, delayed responses and notification.
> > > >
> > > > Multiple virtual SCMI instances support has been tested too.
> > > >
> > > > SMC has NOT been tested/exercised at run-time, only compile-tested.
> > > > (due to lack of hardware)
> > > >
> > > > Note that in this new setup, all the probe deferral and retries between
> > > > the SCMI core stack and the transports has been removed, since no
> > > > more needed.
> > > >
> > > > Moreover the new drivers have been tested also with a fully
> > > > modularized SCMI stack, i.e.:
> > > >
> > > >   scmi-core.ko + scmi-module.ko + scmi_transport_*.ko [ + vendor
> > > > modules ]
> > > >
> > > > ToBeDone:
> > > >  - completely remove any dependency at build time at the Kconfig level
> > > > between
> > > >    the SCMI core and the transport drivers: so that the transports will be
> > > >    dependent only on the related subsystems (optee/virtio/mailbox/smc)
> > > >    (easy to be done but maybe it is not worth...)
> > > >  - integrate per-platform transport configuration capabilities
> > > >    (max_rx_timeout_ms & friends..)
> > > >
> > > > Based on sudeep/for-next/scmi/updates.
> > > >
> > > > Any feedback, and especially testing (:D) is welcome.
> > > >
> > >
> > > For the v2 patchset:
> > > Tested-by: Peng Fan <peng.fan@nxp.com>  #i.MX95-19x19-EVK
> > >
> > 
> > Thanks a lot for the review and testing,
> > 
> > Cristian
> 
> 
> I've tested this v2 on stm32mp157c-scmi.dts. Using built-in modules
> works perfectly.  I've tweaked my platform setup to test the .ko and
> modprobe part. It works ok for the probe part but I faced kernel oops
> when unloading scmi-module after transport is loaded, used, then unoaded.
> The issue I saw is around calls to info->desc->ops->chan_free in
> scmi_cleanup_channels(). I wonder if there are some ops that were not
> unregistered when transport driver is unloaded.
> 

You are right, I could reproduce your oops in my QEMU/optee setup.

There was a bug in chan_free for optee that pre-dated this series....it
is exposed when unloading the scmi-module....I'll post a fix for this
as the initial patch of this series V3.

Moreover even once that was fixed, there was another bug in the
optee_remove of this new transport driver since I was calling the
platform_driver_unregister() too late (after the check for channel
empty)...as a result when you unload the scmi_transport_optee BEFORE the
scmi-module (which is another valid unload sequence option) the core
SCMI stack was NOT unbound like for the other transports.

Last but not least, I spotted another issue for all of these transport
drivers (and related WARN) when finally unloading the scmi-core module
(the last one to go) due to a missing device_release...this was easily
fixed just by using other platform drivers core helpers...so I
refactored more the DEFINE_SCMI_TRANSPORT_DRIVER macros internals...

Next week, on top of -rc1, I'll post a v3 with all the fixes I
mentioned.

Thanks,
Cristian