On Tue, Nov 23, 2021 at 7:29 PM Bjorn Andersson
<bjorn.andersson@xxxxxxxxxx> wrote:
>
> On Fri 29 Oct 14:48 PDT 2021, Tadeusz Struk wrote:
>
> > Venus video encode/decode hardware driver consists of three modules.
> > The parent module venus-core, and two sub modules venus-enc and venus-dec.
> > The venus-core module allocates a common structure that is used by the
> > enc/dec modules, loads the firmware, and performs some common hardware
> > initialization. Since the three modules are loaded one after the other,
> > and their probe functions can run in parallel it is possible that
> > the venc_probe and vdec_probe functions can finish before the core
> > venus_probe function, which then can fail when, for example it
> > fails to load the firmware. In this case the subsequent call to venc_open
> > causes an Oops as it tries to dereference already uninitialized structures
> > through dev->parent and the system crashes in __pm_runtime_resume() as in
> > the trace below:
> >
> > [ 26.064835][ T485] Internal error: Oops: 96000006 [#1] PREEMPT SMP
> > [ 26.270914][ T485] Hardware name: Thundercomm Dragonboard 845c (DT)
> > [ 26.285019][ T485] pc : __pm_runtime_resume+0x34/0x178
> > [ 26.286374][ T213] lt9611 10-003b: hdmi cable connected
> > [ 26.290285][ T485] lr : venc_open+0xc0/0x278 [venus_enc]
> > [ 26.290326][ T485] Call trace:
> > [ 26.290328][ T485] __pm_runtime_resume+0x34/0x178
> > [ 26.290330][ T485] venc_open+0xc0/0x278 [venus_enc]
> > [ 26.290335][ T485] v4l2_open+0x184/0x294
> > [ 26.290340][ T485] chrdev_open+0x468/0x5c8
> > [ 26.290344][ T485] do_dentry_open+0x260/0x54c
> > [ 26.290349][ T485] path_openat+0xbe8/0xd5c
> > [ 26.290352][ T485] do_filp_open+0xb8/0x168
> > [ 26.290354][ T485] do_sys_openat2+0xa4/0x1e8
> > [ 26.290357][ T485] __arm64_compat_sys_openat+0x70/0x9c
> > [ 26.290359][ T485] invoke_syscall+0x60/0x170
> > [ 26.290363][ T485] el0_svc_common+0xb8/0xf8
> > [ 26.290365][ T485] do_el0_svc_compat+0x20/0x30
> > [ 26.290367][ T485] el0_svc_compat+0x24/0x84
> > [ 26.290372][ T485] el0t_32_sync_handler+0x7c/0xbc
> > [ 26.290374][ T485] el0t_32_sync+0x1b8/0x1bc
> > [ 26.290381][ T485] ---[ end trace 04ca7c088b4c1a9c ]---
> > [ 26.290383][ T485] Kernel panic - not syncing: Oops: Fatal exception
> >
> > This can be fixed by synchronizing the three probe functions and
> > only allowing the venc_probe() and vdec_probe() to pass when venus_probe()
> > returns success.
> >
> > Changes in v2:
> > - Change locking from mutex_lock to mutex_trylock
> > in venc_probe and vdec_probe to avoid potential deadlock.
> >
>
> Rather than trying to synchronize away the side effects of
> of_platform_populate() I think we should stop using it.
>
> I had the very same problem in the qcom_wcnss remoteproc driver and
> in below change I got rid of that by manually initializing a struct
> device for the child node. In the event that the child probe defer I
> would just probe defer the parent as well.
>
> 1fcef985c8bd ("remoteproc: qcom: wcnss: Fix race with iris probe")
>
> The change might look a little bit messy, but the end result it much
> cleaner than relying on various locks etc.
>
>
> But in the qcom_wcnss case I have a child _device_ because I need
> something to do e.g. regulator_get() on. I fail to see why venc and vdec
> are devices in the first place.
I definitely agree with Bjorn that all this asynchronous component
probing feels overly complicated, and a rework is probably the better
solution.
Though my only question is: is someone planning to do this rework?
In the meantime, Tadeusz' patch does resolve a *very* frequent boot
crash seen when the venus driver is enabled.
So Stanimir, should we consider merging this as a stop gap until the
larger probe rework is done?
thanks
-john
