On 12/18/25 18:59, Boris Brezillon wrote:
On Thu, 18 Dec 2025 16:57:28 +0000
Steven Price <[email protected]> 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 <[email protected]>
Signed-off-by: Ketil Johnsen <[email protected]>
---
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;
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().
Yes, I also think we can safely drop the check. I even injected some
probe errors to double check, and I see that we still terminate the FW
VM successfully (as this is not executed in this case).
I will send a v2 shortly with this check (and braces) removed.
--
Thanks,
Ketil
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);
