Re: [PATCH v4] drm/virtio: use uninterruptible resv lock for plane updates

["Christian König" <christian.koenig@amd.com>]

On 5/20/26 10:12, Dmitry Osipenko wrote:
> On 5/20/26 10:05, Christian König wrote:
>> On 5/20/26 08:50, Dmitry Osipenko wrote:
>>> On 5/19/26 11:27, Christian König wrote:
>>>> On 5/19/26 10:22, Deepanshu Kartikey wrote:
>>>>> virtio_gpu_cursor_plane_update() and virtio_gpu_resource_flush() lock
>>>>> the framebuffer BO's dma_resv via virtio_gpu_array_lock_resv() and
>>>>> ignore its return value. The function can fail with -EINTR from
>>>>> dma_resv_lock_interruptible() (signal during lock wait) or with
>>>>> -ENOMEM from dma_resv_reserve_fences() (fence slot allocation),
>>>>> leaving the resv lock not held. The queue path then walks the object
>>>>> array and calls dma_resv_add_fence(), which requires the lock held;
>>>>> with lockdep enabled this trips dma_resv_assert_held():
>>>>>
>>>>>   WARNING: drivers/dma-buf/dma-resv.c:296 at dma_resv_add_fence+0x71e/0x840
>>>>>   Call Trace:
>>>>>    virtio_gpu_array_add_fence
>>>>>    virtio_gpu_queue_ctrl_sgs
>>>>>    virtio_gpu_queue_fenced_ctrl_buffer
>>>>>    virtio_gpu_cursor_plane_update
>>>>>    drm_atomic_helper_commit_planes
>>>>>    drm_atomic_helper_commit_tail
>>>>>    commit_tail
>>>>>    drm_atomic_helper_commit
>>>>>    drm_atomic_commit
>>>>>    drm_atomic_helper_update_plane
>>>>>    __setplane_atomic
>>>>>    drm_mode_cursor_universal
>>>>>    drm_mode_cursor_common
>>>>>    drm_mode_cursor_ioctl
>>>>>    drm_ioctl
>>>>>    __x64_sys_ioctl
>>>>>
>>>>> Beyond the WARN, mutating the dma_resv fence list without the lock
>>>>> races with concurrent readers/writers and can corrupt the list.
>>>>
>>>> Well why are you trying to add a fence on an atomic mode set in the first place?
>>>>
>>>> That is usually an illegal operation here.
>>> That is pre-existing in the driver. It performs draw operation and in
>>> some cases waits for the completion during atomic. Whether all that
>>> syncing is correct is hard to say immediately as some of it may be
>>> historical edge cases.
>>
>> I'm not not so deeply in the atomic mode setting stuff but it strongly sounds like that this is seriously broken.
>>
>> The background is that the atomic mode set framework allows an output dma_fence which is signaled when the commit is finished.
>>
>> So when you allocate a fence slot and add a new fence to finish the atomic commit it is trivially possible that this cycles back and waits for the atomic commit to finish. In other words you have a deadlock.
>>
>> You probably need specially crafted userspace with the right timing to trigger that, but such issues are usually a rather big no-no and need to be fixed in the long term.
>>
>> Try to add dma_fence_begin_signaling() and dma_fence_end_signaling() annotation and enable lockdep, the tool should be able to point out if and what exactly goes wrong.
>>
>> The usual fix is to prepare everything before commit_tail is called (alloc memory, create, reserve slot, add dma_fence etc....) and then just send out the prepared commands later on.
> 
> We tried with moving resv alloc to prepare_fb() in a previous patch
> version, it resulted in a non-trivial deadlocks. The goal of this patch
> is to fix immediate problem with a minimal code change.

Yeah, totally fine with me to get that fixed first.

> What you're saying is correct, but it may require a rather big
> refactoring of the code. In general, everything works okay today, so not
> really an urgent problem.

It's just a potential issue and when the AI bots keep evolving like they already do they will sooner or later start to point that out as well.

Regards,
Christian.