Re: [PATCH v3 2/2] drm/virtio: Support sync objects

[Dmitry Osipenko]

On 3/24/23 00:18, Rob Clark wrote:
...
>> +static int
>> +virtio_gpu_parse_deps(struct virtio_gpu_submit *submit)
>> +{
>> +       struct drm_virtgpu_execbuffer *exbuf = submit->exbuf;
>> +       struct drm_virtgpu_execbuffer_syncobj syncobj_desc;
>> +       size_t syncobj_stride = exbuf->syncobj_stride;
>> +       struct drm_syncobj **syncobjs;
>> +       int ret = 0, i;
>> +
>> +       if (!submit->num_in_syncobjs)
>> +               return 0;
>> +
>> +       syncobjs = kcalloc(submit->num_in_syncobjs, sizeof(*syncobjs),
>> +                          GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
> I *think*, assuming I'm reading where this is called correctly (kinda
> wish git would show more lines of context by default) that these don't
> need to be NOWARN|NORETRY (same for post_deps).  I guess you inherited
> this from drm/msm, where I appear to have forgotten to update the
> syncobj path in commit f0de40a131d9 ("drm/msm: Reorder lock vs submit
> alloc").  I don't see anything obvious that would require NORETRY, but
> lockdep should be able to tell you otherwise if needed.

The NORETRY should prevent waking up OOM killer, it shouldn't help with
lockdep. Nothing prevents userspace from giving a big number of
num_in_syncobjs. But perhaps indeed not very practical to care about
this case, given that other similar memalloc paces of execbuffer_ioctl()
aren't using NORETRY. Alright, let's drop it in v4.

-- 
Best regards,
Dmitry