Re: [PATCH v2] misc: fastrpc: Fix channel resource access in device_open

[Greg KH <gregkh@linuxfoundation.org>]

On Wed, Jun 25, 2025 at 02:45:27AM +0300, Dmitry Baryshkov wrote:
> On Tue, Jun 24, 2025 at 04:38:25PM +0100, Greg KH wrote:
> > On Tue, Jun 24, 2025 at 04:36:35PM +0100, Greg KH wrote:
> > > On Tue, Jun 24, 2025 at 04:27:21PM +0300, Dmitry Baryshkov wrote:
> > > > On Thu, Jun 19, 2025 at 10:40:26AM +0530, Ekansh Gupta wrote:
> > > > > During rpmsg_probe, fastrpc device nodes are created first, then
> > > > > channel specific resources are initialized, followed by
> > > > > of_platform_populate, which triggers context bank probing. This
> > > > > sequence can cause issues as applications might open the device
> > > > > node before channel resources are initialized or the session is
> > > > > available, leading to problems. For example, spin_lock is initialized
> > > > > after the device node creation, but it is used in device_open,
> > > > > potentially before initialization. Move device registration after
> > > > > channel resource initialization in fastrpc_rpmsg_probe.
> > > > 
> > > > You've moved device init, however there is still a possibility for the
> > > > context devices to be created, but not bound to the driver (because all
> > > > the probings are async). I think instead we should drop the extra
> > > > platform driver layer and create and set up corresponding devices
> > > > manually. For example, see how it is handled in
> > > > host1x_memory_context_list_init(). That function uses iommu-maps, but we
> > > > can use OF nodes and iommus instead.
> > > 
> > > Is this a real platform device?  If so, why do you need a second
> > > platform driver, what makes this so unique?  If this isn't a platform
> > > device, then why not just use the faux bus instead?
> > > 
> > > It seems that "number of sessions" is a DT property, is that something
> > > that is really defined by the hardware?  Or is it just a virtual thing
> > > that people are abusing in the DT?
> 
> Purely software value.
> 
> > > 
> > > And if you really have all these sessions, why not make them real
> > > devices, wouldn't that make things simpler?
> > 
> > Oh wait, these are "fake" platform devices under the parent (i.e. real)
> > platform device.  That's not good, please don't do that, use the faux
> > bus code now instead to properly handle this.  Attempting to create a
> > device when open() is called is really really odd...
> 
> The driver doesn't created devices during open(). It creates them
> earlier, then another driver probes an populates the data. I suggest to
> follow Tegra approach, remove the sub-driver completely and instead of
> calling of_platform_populate() create necessary devices manually and set
> corresponding IOMMU configuration from the main driver's probe path.

That sounds much more reasonable.

thanks,

greg k-h