Message-ID: <20251219083933.2dc9c009@fedora>
Date: Fri, 19 Dec 2025 08:39:33 +0100
From: Boris Brezillon <boris.brezillon@...labora.com>
To: Chia-I Wu <olvaffe@...il.com>
Cc: Steven Price <steven.price@....com>, Ketil Johnsen
 <ketil.johnsen@....com>, Liviu Dudau <liviu.dudau@....com>, Maarten
 Lankhorst <maarten.lankhorst@...ux.intel.com>, Maxime Ripard
 <mripard@...nel.org>, Thomas Zimmermann <tzimmermann@...e.de>, David Airlie
 <airlied@...il.com>, Simona Vetter <simona@...ll.ch>, Grant Likely
 <grant.likely@...aro.org>, Heiko Stuebner <heiko@...ech.de>,
 dri-devel@...ts.freedesktop.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] drm/panthor: Evict groups before VM termination

On Thu, 18 Dec 2025 14:54:35 -0800
Chia-I Wu <olvaffe@...il.com> wrote:

> On Thu, Dec 18, 2025 at 10:36 AM Boris Brezillon
> <boris.brezillon@...labora.com> wrote:
> >
> > On Thu, 18 Dec 2025 16:57:28 +0000
> > Steven Price <steven.price@....com> wrote:
> >  
> > > On 18/12/2025 16:26, Ketil Johnsen wrote:  
> > > > Ensure all related groups are evicted and suspended before VM
> > > > destruction takes place.
> > > >
> > > > This fixes an issue where panthor_vm_destroy() destroys and unmaps the
> > > > heap context while there are still on slot groups using this.
> > > > The FW will do a write out to the heap context when a CSG (group) is
> > > > suspended, so a premature unmap of the heap context will cause a
> > > > GPU page fault.
> > > > This page fault is quite harmless, and do not affect the continued
> > > > operation of the GPU.
> > > >
> > > > Fixes: 647810ec2476 ("drm/panthor: Add the MMU/VM logical block")
> > > > Co-developed-by: Boris Brezillon <boris.brezillon@...labora.com>
> > > > Signed-off-by: Ketil Johnsen <ketil.johnsen@....com>
> > > > ---
> > > >  drivers/gpu/drm/panthor/panthor_mmu.c   |  4 ++++
> > > >  drivers/gpu/drm/panthor/panthor_sched.c | 16 ++++++++++++++++
> > > >  drivers/gpu/drm/panthor/panthor_sched.h |  1 +
> > > >  3 files changed, 21 insertions(+)
> > > >
> > > > diff --git a/drivers/gpu/drm/panthor/panthor_mmu.c b/drivers/gpu/drm/panthor/panthor_mmu.c
> > > > index 74230f7199121..0e4b301a9c70e 100644
> > > > --- a/drivers/gpu/drm/panthor/panthor_mmu.c
> > > > +++ b/drivers/gpu/drm/panthor/panthor_mmu.c
> > > > @@ -1537,6 +1537,10 @@ static void panthor_vm_destroy(struct panthor_vm *vm)
> > > >
> > > >     vm->destroyed = true;
> > > >
> > > > +   /* Tell scheduler to stop all GPU work related to this VM */
> > > > +   if (refcount_read(&vm->as.active_cnt) > 0)
> > > > +           panthor_sched_prepare_for_vm_destruction(vm->ptdev);
> > > > +
> > > >     mutex_lock(&vm->heaps.lock);
> > > >     panthor_heap_pool_destroy(vm->heaps.pool);
> > > >     vm->heaps.pool = NULL;  
> Is it better to remove the panthor_heap_pool_destroy call here instead
> and let panthor_vm_free take care of it?

We can't because the heap_pool contains heap chunks (kernel BOs) that
are mapped in the very same VM, thus creating a circular ref. The whole
point of calling panthor_heap_pool_destroy() here is to kill this
circular ref. We could introduce the concept of weak and hard VM refs,
but I'm not sure it's worth it.

> 
> > > > diff --git a/drivers/gpu/drm/panthor/panthor_sched.c b/drivers/gpu/drm/panthor/panthor_sched.c
> > > > index f680edcd40aad..fbbaab9b25efb 100644
> > > > --- a/drivers/gpu/drm/panthor/panthor_sched.c
> > > > +++ b/drivers/gpu/drm/panthor/panthor_sched.c
> > > > @@ -2930,6 +2930,22 @@ void panthor_sched_report_mmu_fault(struct panthor_device *ptdev)
> > > >             sched_queue_delayed_work(ptdev->scheduler, tick, 0);
> > > >  }
> > > >
> > > > +void panthor_sched_prepare_for_vm_destruction(struct panthor_device *ptdev)
> > > > +{
> > > > +   /* FW can write out internal state, like the heap context, during CSG
> > > > +    * suspend. It is therefore important that the scheduler has fully
> > > > +    * evicted any pending and related groups before VM destruction can
> > > > +    * safely continue. Failure to do so can lead to GPU page faults.
> > > > +    * A controlled termination of a Panthor instance involves destroying
> > > > +    * the group(s) before the VM. This means any relevant group eviction
> > > > +    * has already been initiated by this point, and we just need to
> > > > +    * ensure that any pending tick_work() has been completed.
> > > > +    */
> > > > +   if (ptdev->scheduler) {
> > > > +           flush_work(&ptdev->scheduler->tick_work.work);
> > > > +   }  
> > >
> > > NIT: braces not needed.
> > >
> > > But I'm also struggling to understand in what situation ptdev->scheduler
> > > would be NULL?  
> >
> > I thought it could happen if the FW initialization fails in the middle,
> > and the FW VM is destroyed before the scheduler had a chance to
> > initialize, but it turns out the FW logic never calls
> > panthor_vm_destroy().
> >  
> > >
> > > Thanks,
> > > Steve
> > >  
> > > > +}
> > > > +
> > > >  void panthor_sched_resume(struct panthor_device *ptdev)
> > > >  {
> > > >     /* Force a tick to re-evaluate after a resume. */
> > > > diff --git a/drivers/gpu/drm/panthor/panthor_sched.h b/drivers/gpu/drm/panthor/panthor_sched.h
> > > > index f4a475aa34c0a..9a8692de8aded 100644
> > > > --- a/drivers/gpu/drm/panthor/panthor_sched.h
> > > > +++ b/drivers/gpu/drm/panthor/panthor_sched.h
> > > > @@ -50,6 +50,7 @@ void panthor_sched_suspend(struct panthor_device *ptdev);
> > > >  void panthor_sched_resume(struct panthor_device *ptdev);
> > > >
> > > >  void panthor_sched_report_mmu_fault(struct panthor_device *ptdev);
> > > > +void panthor_sched_prepare_for_vm_destruction(struct panthor_device *ptdev);
> > > >  void panthor_sched_report_fw_events(struct panthor_device *ptdev, u32 events);
> > > >
> > > >  void panthor_fdinfo_gather_group_samples(struct panthor_file *pfile);  
> > >  
> >  

Powered by blists - more mailing lists