On Thu, Dec 11, 2025 at 11:37:58AM +0530, Lazar, Lijo wrote: > > > On 12/11/2025 11:14 AM, Lazar, Lijo wrote: > > > > > > On 12/11/2025 10:52 AM, SHANMUGAM, SRINIVASAN wrote: > > > [AMD Official Use Only - AMD Internal Distribution Only] > > > > > > > -----Original Message----- > > > > From: Lazar, Lijo <[email protected]> > > > > Sent: Thursday, December 11, 2025 10:34 AM > > > > To: Deucher, Alexander <[email protected]>; amd- > > > > [email protected] > > > > Cc: SHANMUGAM, SRINIVASAN <[email protected]>; > > > > Prosyak, Vitaly <[email protected]>; Koenig, Christian > > > > <[email protected]>; Matthew Brost <[email protected]> > > > > Subject: Re: [PATCH V2] drm/amdgpu: fix a job->pasid access race in gpu > > > > recovery > > > > > > > > > > > > > > > > On 12/11/2025 1:53 AM, Alex Deucher wrote: > > > > > Avoid a possible UAF in GPU recovery due to a race between the sched > > > > > timeout callback and the tdr work queue. > > > > > > > > > > The gpu recovery function calls drm_sched_stop() and later > > > > > drm_sched_start(). drm_sched_start() restarts the tdr queue which > > > > > will eventually free the job. If the tdr queue frees the job before > > > > > time out callback completes, the job will be freed and we'll get a UAF > > > > > when accessing the pasid. Cache it early to avoid the UAF. > > > > > > > > > > Fixes: a72002cb181f ("drm/amdgpu: Make use of drm_wedge_task_info") > > > > > Cc: [email protected] > > > > > Cc: [email protected] > > > > > Cc: [email protected] > > > > > Suggested-by: Matthew Brost <[email protected]> > > > > > Signed-off-by: Alex Deucher <[email protected]> > > > > > --- > > > > > > > > > > v2: Check the pasid rather than job (Lijo) > > > > > Add fixes tag (Christian) > > > > > > > > > > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 10 ++++++++-- > > > > > 1 file changed, 8 insertions(+), 2 deletions(-) > > > > > > > > > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > > > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > > > > index 8a851d7548c00..c6b1dd95c401d 100644 > > > > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > > > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > > > > @@ -6634,6 +6634,8 @@ int amdgpu_device_gpu_recover(struct > > > > amdgpu_device *adev, > > > > > struct amdgpu_hive_info *hive = NULL; > > > > > int r = 0; > > > > > bool need_emergency_restart = false; > > > > > + /* save the pasid here as the job may be freed before > > > > > the end of the reset */ > > > > > + int pasid = job ? job->pasid : -EINVAL; > > > > > > > > > > /* > > > > > * If it reaches here because of hang/timeout and a RAS > > > > > error is @@ > > > > > -6734,8 +6736,12 @@ int amdgpu_device_gpu_recover(struct amdgpu_device > > > > *adev, > > > > > if (!r) { > > > > > struct amdgpu_task_info *ti = NULL; > > > > > > > > > > - if (job) > > > > > - ti = amdgpu_vm_get_task_info_pasid(adev, > > > > > job- >pasid); > > > > > + /* > > > > > + * The job may already be freed at this point > > > > > via the sched tdr > > > > workqueue so > > > > > + * use the cached pasid. > > > > > + */ > > > > > > > > amdgpu_device_gpu_recover() is run in tdr workqueue. > > > > > > > > Now if this is the case, someone has to explain the logic - > > > > > > > > Timeout is triggered here - > > > > https://github.com/torvalds/linux/blob/master/drivers/gpu/drm/ > > > > scheduler/sched_main > > > > .c#L559 > > > > > > > > This calls amdgpu_job_timedout() -> amdgpu_device_gpu_recover() > > > > > > > > After that, there is this access to the job - > > > > > > > > https://github.com/torvalds/linux/blob/master/drivers/gpu/drm/ > > > > scheduler/sched_main > > > > .c#L566
Yes, the DRM scheduler is broken. Luckily, if free_guilty is set, the job isn’t in the pending list, so it won’t disappear. It is actually safe to touch this in the above code example. Complete nonsense, yes. Is it safe, barely. I feel like there is concept in Linux which solves this whole thing. > > > > > > > > At least, in some condition, job is not expected to be freed. > > > > Then I'm not sure if this > > > > is the right fix. It isn’t. Fixing DRM scheduler is. But until then, I’m fairly certain your driver won’t explode with this patch. Matt > > > > > > What is that "someone", "some condition" you feel like? Its better > > > to bring proper justification, and take up this as separate > > > refactoring task > > > > > > > Basically, if scheduler code itself is not expecting job to be not freed > > after timedout callback, then why callback handler needs to assume the > > same? > > > > Taking out double 'not', following is what I meant - > > 'if scheduler code itself is expecting job to be not freed' > > Thanks, > Lijo > > > Now if callback handler does something else which in turn frees the job, > > the fix needs to be there instead of having this kind of fix. > > > > Thanks, > > Lijo > > > > > > > Best, > > > Srini > > >
