On Wed, 22 Oct 2025 15:28:51 +0100 Steven Price <[email protected]> wrote:
> On 22/10/2025 15:00, Boris Brezillon wrote: > > On Wed, 22 Oct 2025 14:36:23 +0100 > > Steven Price <[email protected]> wrote: > > > >> On 22/10/2025 13:37, Boris Brezillon wrote: > >>> On Wed, 22 Oct 2025 12:30:13 +0200 > >>> Ketil Johnsen <[email protected]> wrote: > >>> > >>>> The function panthor_fw_unplug() will free the FW memory sections. > >>>> The problem is that there could still be pending FW events which are yet > >>>> not handled at this point. process_fw_events_work() can in this case try > >>>> to access said freed memory. > >>>> > >>>> This fix introduces a destroyed state for the panthor_scheduler object, > >>>> and we check for this before processing FW events. > >>>> > >>>> Signed-off-by: Ketil Johnsen <[email protected]> > >>>> Fixes: de85488138247 ("drm/panthor: Add the scheduler logical block") > >>>> --- > >>>> v2: > >>>> - Followed Boris's advice and handle the race purely within the > >>>> scheduler block (by adding a destroyed state) > >>>> --- > >>>> drivers/gpu/drm/panthor/panthor_sched.c | 15 ++++++++++++--- > >>>> 1 file changed, 12 insertions(+), 3 deletions(-) > >>>> > >>>> diff --git a/drivers/gpu/drm/panthor/panthor_sched.c > >>>> b/drivers/gpu/drm/panthor/panthor_sched.c > >>>> index 0cc9055f4ee52..4996f987b8183 100644 > >>>> --- a/drivers/gpu/drm/panthor/panthor_sched.c > >>>> +++ b/drivers/gpu/drm/panthor/panthor_sched.c > >>>> @@ -315,6 +315,13 @@ struct panthor_scheduler { > >>>> */ > >>>> struct list_head stopped_groups; > >>>> } reset; > >>>> + > >>>> + /** > >>>> + * @destroyed: Scheduler object is (being) destroyed > >>>> + * > >>>> + * Normal scheduler operations should no longer take place. > >>>> + */ > >>>> + bool destroyed; > >>> > >>> Do we really need a new field for that? Can't we just reset > >>> panthor_device::scheduler to NULL early enough in the unplug path? > >>> I guess it's not that simple if we have works going back to ptdev > >>> and then dereferencing ptdev->scheduler, but I think it's also > >>> fundamentally broken to have scheduler works active after the > >>> scheduler teardown has started, so we might want to add some more > >>> checks in the work callbacks too. > >>> > >>>> }; > >>>> > >>>> /** > >>>> @@ -1765,7 +1772,10 @@ static void process_fw_events_work(struct > >>>> work_struct *work) > >>>> u32 events = atomic_xchg(&sched->fw_events, 0); > >>>> struct panthor_device *ptdev = sched->ptdev; > >>>> > >>>> - mutex_lock(&sched->lock); > >>>> + guard(mutex)(&sched->lock); > >>>> + > >>>> + if (sched->destroyed) > >>>> + return; > >>>> > >>>> if (events & JOB_INT_GLOBAL_IF) { > >>>> sched_process_global_irq_locked(ptdev); > >>>> @@ -1778,8 +1788,6 @@ static void process_fw_events_work(struct > >>>> work_struct *work) > >>>> sched_process_csg_irq_locked(ptdev, csg_id); > >>>> events &= ~BIT(csg_id); > >>>> } > >>>> - > >>>> - mutex_unlock(&sched->lock); > >>>> } > >>>> > >>>> /** > >>>> @@ -3882,6 +3890,7 @@ void panthor_sched_unplug(struct panthor_device > >>>> *ptdev) > >>>> cancel_delayed_work_sync(&sched->tick_work); > >>>> > >>>> mutex_lock(&sched->lock); > >>>> + sched->destroyed = true; > >>>> if (sched->pm.has_ref) { > >>>> pm_runtime_put(ptdev->base.dev); > >>>> sched->pm.has_ref = false; > >>> > >>> Hm, I'd really like to see a cancel_work_sync(&sched->fw_events_work) > >>> rather than letting the work execute after we've started tearing down > >>> the scheduler object. > >>> > >>> If you follow my suggestion to reset the ptdev->scheduler field, I > >>> guess something like that would do: > >>> > >>> void panthor_sched_unplug(struct panthor_device *ptdev) > >>> { > >>> struct panthor_scheduler *sched = ptdev->scheduler; > >>> > >>> /* We want the schedu */ > >>> WRITE_ONCE(*ptdev->scheduler, NULL); > >>> > >>> cancel_work_sync(&sched->fw_events_work); > >>> cancel_delayed_work_sync(&sched->tick_work); > >>> > >>> mutex_lock(&sched->lock); > >>> if (sched->pm.has_ref) { > >>> pm_runtime_put(ptdev->base.dev); > >>> sched->pm.has_ref = false; > >>> } > >>> mutex_unlock(&sched->lock); > >>> } > >>> > >>> and > >>> > >>> void panthor_sched_report_fw_events(struct panthor_device *ptdev, u32 > >>> events) { > >>> struct panthor_scheduler *sched = READ_ONCE(*ptdev->scheduler); > >>> > >>> /* Scheduler is not initialized, or it's gone. */ > >>> if (!sched) > >>> return; > >>> > >>> atomic_or(events, &sched->fw_events); > >>> sched_queue_work(sched, fw_events); > >>> } > >> > >> Note there's also the path of panthor_mmu_irq_handler() calling > >> panthor_sched_report_mmu_fault() which will need to READ_ONCE() as well > >> to be safe. > > > > This could be hidden behind a panthor_device_get_sched() helper, I > > guess. Anyway, it's not so much that I'm against the addition of an > > extra bool, but AFAICT, the problem is not entirely solved, as there > > could be a pending work that gets executed after sched_unplug() > > returns, and I adding this bool check just papers over the real bug > > (which is that we never cancel the fw_event work). > > > >> > >> I agree having an extra bool is ugly, but it easier to reason about than > >> the lock-free WRITE_ONCE/READ_ONCE dance. It worries me that this will > >> be regressed in the future. I can't immediately see how to wrap this in > >> a helper to ensure this is kept correct. > > > > Sure, but you're not really catching cases where the work runs after > > the scheduler component has been unplugged in case someone forgot to > > cancel some works. I think I'd rather identify those cases with a > > kernel panic, than a random UAF when the work is being executed. > > Ultimately, we should probably audit all works used in the driver, to > > make sure they are properly cancelled at unplug() time by the relevant > > <component>_unplug() functions. > > Yes I agree, we should have a cancel_work_sync(&sched->fw_events_work) > call somewhere on the unplug path. That needs to be after the job irq > has been disabled which is currently done in panthor_fw_unplug(). Not necessarily. If we prevent any further FW events to queue the fw_events work, we can just cancel it in the sched_unplug() path, after we've transition to this "sched-is-gone" state.
