>Hi Ming,
>
>On 11/1/22 09:05, Ming Qian wrote:
>> driver need to cancel vpu before releasing the vpu instance, but it
>> doesn't want to release the m2m_ctx before the release callback of
>> instance is called, as driver may access the m2m_ctx in handling some
>> event which may be received in releasing instance.
>>
>> check and cancel the unstopped instance before release.
>>
>> Fixes: d91d7bc85062 ("media: amphion: release m2m ctx when releasing
>> vpu instance")
>> Signed-off-by: Ming Qian <ming.qian@xxxxxxx>
>> ---
>> drivers/media/platform/amphion/vpu_v4l2.c | 18 ++++++++++++++++++
>> 1 file changed, 18 insertions(+)
>>
>> diff --git a/drivers/media/platform/amphion/vpu_v4l2.c
>> b/drivers/media/platform/amphion/vpu_v4l2.c
>> index 99ad2f1c5a53..845fc53d8937 100644
>> --- a/drivers/media/platform/amphion/vpu_v4l2.c
>> +++ b/drivers/media/platform/amphion/vpu_v4l2.c
>> @@ -767,6 +767,23 @@ int vpu_v4l2_open(struct file *file, struct vpu_inst
>*inst)
>> return ret;
>> }
>>
>> +static void vpu_v4l2_check_and_cancel(struct file *file, struct
>> +vpu_inst *inst) {
>> + struct vb2_queue *vq;
>> +
>> + vpu_inst_lock(inst);
>> +
>> + vq = v4l2_m2m_get_src_vq(inst->fh.m2m_ctx);
>> + if (vb2_is_streaming(vq))
>> + v4l2_m2m_streamoff(file, inst->fh.m2m_ctx, vq->type);
>> +
>> + vq = v4l2_m2m_get_dst_vq(inst->fh.m2m_ctx);
>> + if (vb2_is_streaming(vq))
>> + v4l2_m2m_streamoff(file, inst->fh.m2m_ctx, vq->type);
>> +
>> + vpu_inst_unlock(inst);
>
>This feels weird. This is normally done in v4l2_m2m_ctx_release(), so if you
>need to do this here, doesn't that mean perhaps that the order of releasing
>things should be changed? I.e., first call v4l2_m2m_ctx_release() and only
>afterwards release the vpu instance.
>
>In other words, are you just fixing the symptom rather than the actual cause?
>
>Regards,
>
> Hans
>
Hi Hans,
I agree,
In the first I indeed call v4l2_m2m_ctx_release() and afterwards release the vpu instance,
But there may be some redundant event triggered by firmware. And they may led to kernel panic as I didn't put the whole event handler into the instance lock.
I just thought if I release the m2m ctx in releasing vpu instance, then thus case can be avoided. But it's really not well thought out.
I'll revert the patch " media: amphion: release m2m ctx when releasing vpu instance", then check the event handler carefully to make sure there is no issue if the m2m ctx has been fixed.
Ming
>> +}
>> +
>> int vpu_v4l2_close(struct file *file) {
>> struct vpu_dev *vpu = video_drvdata(file); @@ -774,6 +791,7 @@
>> int vpu_v4l2_close(struct file *file)
>>
>> vpu_trace(vpu->dev, "tgid = %d, pid = %d, inst = %p\n",
>> inst->tgid, inst->pid, inst);
>>
>> + vpu_v4l2_check_and_cancel(file, inst);
>> call_void_vop(inst, release);
>> vpu_inst_unregister(inst);
>> vpu_inst_put(inst);
