[v1] drm/panthor: always set fence errors on CS_FAULT

[PATCH] drm/panthor: always set fence errors on CS_FAULT

Posted by Chia-I Wu 3 months, 3 weeks ago

It is unclear why fence errors were set only for CS_INHERIT_FAULT.
Downstream driver also does not treat CS_INHERIT_FAULT specially.
Remove the check.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
---
 drivers/gpu/drm/panthor/panthor_sched.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/panthor/panthor_sched.c b/drivers/gpu/drm/panthor/panthor_sched.c
index a2248f692a030..1a3b1c49f7d7b 100644
--- a/drivers/gpu/drm/panthor/panthor_sched.c
+++ b/drivers/gpu/drm/panthor/panthor_sched.c
@@ -1399,7 +1399,7 @@ cs_slot_process_fault_event_locked(struct panthor_device *ptdev,
 	fault = cs_iface->output->fault;
 	info = cs_iface->output->fault_info;
 
-	if (queue && CS_EXCEPTION_TYPE(fault) == DRM_PANTHOR_EXCEPTION_CS_INHERIT_FAULT) {
+	if (queue) {
 		u64 cs_extract = queue->iface.output->extract;
 		struct panthor_job *job;
 
-- 
2.50.0.rc2.696.g1fc2a0284f-goog

Re: [PATCH] drm/panthor: always set fence errors on CS_FAULT

Posted by Boris Brezillon 3 months, 2 weeks ago

On Wed, 18 Jun 2025 07:55:49 -0700
Chia-I Wu <olvaffe@gmail.com> wrote:

> It is unclear why fence errors were set only for CS_INHERIT_FAULT.
> Downstream driver also does not treat CS_INHERIT_FAULT specially.
> Remove the check.
> 
> Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
> ---
>  drivers/gpu/drm/panthor/panthor_sched.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/panthor/panthor_sched.c b/drivers/gpu/drm/panthor/panthor_sched.c
> index a2248f692a030..1a3b1c49f7d7b 100644
> --- a/drivers/gpu/drm/panthor/panthor_sched.c
> +++ b/drivers/gpu/drm/panthor/panthor_sched.c
> @@ -1399,7 +1399,7 @@ cs_slot_process_fault_event_locked(struct panthor_device *ptdev,
>  	fault = cs_iface->output->fault;
>  	info = cs_iface->output->fault_info;
>  
> -	if (queue && CS_EXCEPTION_TYPE(fault) == DRM_PANTHOR_EXCEPTION_CS_INHERIT_FAULT) {
> +	if (queue) {
>  		u64 cs_extract = queue->iface.output->extract;
>  		struct panthor_job *job;
>  

Now that I look at the code, I think we should record the error when
the ERROR_BARRIER is executed instead of flagging all in-flight jobs as
faulty. One option would be to re-use the profiling buffer by adding an
error field to panthor_job_profiling_data, but we're going to lose 4
bytes per slot because of the 64-bit alignment we want for timestamps,
so maybe just create a separate buffers with N entries of:

struct panthor_job_status {
   u32 error;
};

Re: [PATCH] drm/panthor: always set fence errors on CS_FAULT

Posted by Chia-I Wu 3 months ago

On Sun, Jun 22, 2025 at 11:32 PM Boris Brezillon
<boris.brezillon@collabora.com> wrote:
>
> On Wed, 18 Jun 2025 07:55:49 -0700
> Chia-I Wu <olvaffe@gmail.com> wrote:
>
> > It is unclear why fence errors were set only for CS_INHERIT_FAULT.
> > Downstream driver also does not treat CS_INHERIT_FAULT specially.
> > Remove the check.
> >
> > Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
> > ---
> >  drivers/gpu/drm/panthor/panthor_sched.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/drivers/gpu/drm/panthor/panthor_sched.c b/drivers/gpu/drm/panthor/panthor_sched.c
> > index a2248f692a030..1a3b1c49f7d7b 100644
> > --- a/drivers/gpu/drm/panthor/panthor_sched.c
> > +++ b/drivers/gpu/drm/panthor/panthor_sched.c
> > @@ -1399,7 +1399,7 @@ cs_slot_process_fault_event_locked(struct panthor_device *ptdev,
> >       fault = cs_iface->output->fault;
> >       info = cs_iface->output->fault_info;
> >
> > -     if (queue && CS_EXCEPTION_TYPE(fault) == DRM_PANTHOR_EXCEPTION_CS_INHERIT_FAULT) {
> > +     if (queue) {
> >               u64 cs_extract = queue->iface.output->extract;
> >               struct panthor_job *job;
> >
>
> Now that I look at the code, I think we should record the error when
> the ERROR_BARRIER is executed instead of flagging all in-flight jobs as
> faulty. One option would be to re-use the profiling buffer by adding an
> error field to panthor_job_profiling_data, but we're going to lose 4
> bytes per slot because of the 64-bit alignment we want for timestamps,
> so maybe just create a separate buffers with N entries of:
>
> struct panthor_job_status {
>    u32 error;
> };
The current error path uses cs_extract to mark exactly the offending
job faulty.  Innocent in-flight jobs do not seem to be affected.

I looked into emitting LOAD/STORE after SYNC_ADD64 to copy the error
to panthor_job_status.  Other than the extra instrs and storage,
because group_sync_upd_work can be called before LOAD/STORE, it will
need to check both panthor_job_status and panthor_syncobj_64b.  That
will be a bit ugly as well.

Re: [PATCH] drm/panthor: always set fence errors on CS_FAULT

Posted by Boris Brezillon 1 month, 2 weeks ago

On Tue, 8 Jul 2025 14:40:06 -0700
Chia-I Wu <olvaffe@gmail.com> wrote:

> On Sun, Jun 22, 2025 at 11:32 PM Boris Brezillon
> <boris.brezillon@collabora.com> wrote:
> >
> > On Wed, 18 Jun 2025 07:55:49 -0700
> > Chia-I Wu <olvaffe@gmail.com> wrote:
> >  
> > > It is unclear why fence errors were set only for CS_INHERIT_FAULT.
> > > Downstream driver also does not treat CS_INHERIT_FAULT specially.
> > > Remove the check.
> > >
> > > Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
> > > ---
> > >  drivers/gpu/drm/panthor/panthor_sched.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/gpu/drm/panthor/panthor_sched.c b/drivers/gpu/drm/panthor/panthor_sched.c
> > > index a2248f692a030..1a3b1c49f7d7b 100644
> > > --- a/drivers/gpu/drm/panthor/panthor_sched.c
> > > +++ b/drivers/gpu/drm/panthor/panthor_sched.c
> > > @@ -1399,7 +1399,7 @@ cs_slot_process_fault_event_locked(struct panthor_device *ptdev,
> > >       fault = cs_iface->output->fault;
> > >       info = cs_iface->output->fault_info;
> > >
> > > -     if (queue && CS_EXCEPTION_TYPE(fault) == DRM_PANTHOR_EXCEPTION_CS_INHERIT_FAULT) {
> > > +     if (queue) {
> > >               u64 cs_extract = queue->iface.output->extract;
> > >               struct panthor_job *job;
> > >  
> >
> > Now that I look at the code, I think we should record the error when
> > the ERROR_BARRIER is executed instead of flagging all in-flight jobs as
> > faulty. One option would be to re-use the profiling buffer by adding an
> > error field to panthor_job_profiling_data, but we're going to lose 4
> > bytes per slot because of the 64-bit alignment we want for timestamps,
> > so maybe just create a separate buffers with N entries of:
> >
> > struct panthor_job_status {
> >    u32 error;
> > };  
> The current error path uses cs_extract to mark exactly the offending
> job faulty.  Innocent in-flight jobs do not seem to be affected.

My bad, I thought the faulty CS was automatically entering the recovery
substate (fetching all instructions and ignoring RUN_xxx ones), but it
turns out CS instruction fetching is stalled until the fault is
acknowledged, so we're good.

> 
> I looked into emitting LOAD/STORE after SYNC_ADD64 to copy the error
> to panthor_job_status.  Other than the extra instrs and storage,
> because group_sync_upd_work can be called before LOAD/STORE, it will
> need to check both panthor_job_status and panthor_syncobj_64b.  That
> will be a bit ugly as well.

Nah, I think you're right, I just had a wrong recollection of how
recovery mode works. The patch is

Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>