[Intel-gfx] [PATCH v5 02/20] drm/msm: Fix drm/sched point of no return rules
Daniel Vetter
daniel.vetter at ffwll.ch
Fri Aug 6 18:41:47 UTC 2021
On Fri, Aug 6, 2021 at 7:15 PM Rob Clark <robdclark at gmail.com> wrote:
>
> On Fri, Aug 6, 2021 at 9:42 AM Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> >
> > On Fri, Aug 6, 2021 at 12:58 AM Rob Clark <robdclark at gmail.com> wrote:
> > >
> > > On Thu, Aug 5, 2021 at 3:47 AM Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> > > >
> > > > Originally drm_sched_job_init was the point of no return, after which
> > > > drivers must submit a job. I've split that up, which allows us to fix
> > > > this issue pretty easily.
> > > >
> > > > Only thing we have to take care of is to not skip to error paths after
> > > > that. Other drivers do this the same for out-fence and similar things.
> > > >
> > > > Fixes: 1d8a5ca436ee ("drm/msm: Conversion to drm scheduler")
> > > > Cc: Rob Clark <robdclark at chromium.org>
> > > > Cc: Rob Clark <robdclark at gmail.com>
> > > > Cc: Sean Paul <sean at poorly.run>
> > > > Cc: Sumit Semwal <sumit.semwal at linaro.org>
> > > > Cc: "Christian König" <christian.koenig at amd.com>
> > > > Cc: linux-arm-msm at vger.kernel.org
> > > > Cc: dri-devel at lists.freedesktop.org
> > > > Cc: freedreno at lists.freedesktop.org
> > > > Cc: linux-media at vger.kernel.org
> > > > Cc: linaro-mm-sig at lists.linaro.org
> > > > Signed-off-by: Daniel Vetter <daniel.vetter at intel.com>
> > > > ---
> > > > drivers/gpu/drm/msm/msm_gem_submit.c | 15 +++++++--------
> > > > 1 file changed, 7 insertions(+), 8 deletions(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/msm/msm_gem_submit.c b/drivers/gpu/drm/msm/msm_gem_submit.c
> > > > index 6d6c44f0e1f3..d0ed4ddc509e 100644
> > > > --- a/drivers/gpu/drm/msm/msm_gem_submit.c
> > > > +++ b/drivers/gpu/drm/msm/msm_gem_submit.c
> > > > @@ -52,9 +52,6 @@ static struct msm_gem_submit *submit_create(struct drm_device *dev,
> > > > return ERR_PTR(ret);
> > > > }
> > > >
> > > > - /* FIXME: this is way too early */
> > > > - drm_sched_job_arm(&job->base);
> > > > -
> > > > xa_init_flags(&submit->deps, XA_FLAGS_ALLOC);
> > > >
> > > > kref_init(&submit->ref);
> > > > @@ -883,6 +880,9 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
> > > >
> > > > submit->user_fence = dma_fence_get(&submit->base.s_fence->finished);
> > > >
> > > > + /* point of no return, we _have_ to submit no matter what */
> > > > + drm_sched_job_arm(&submit->base);
> > > > +
> > > > /*
> > > > * Allocate an id which can be used by WAIT_FENCE ioctl to map back
> > > > * to the underlying fence.
> > > > @@ -892,17 +892,16 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
> > > > if (submit->fence_id < 0) {
> > > > ret = submit->fence_id = 0;
> > > > submit->fence_id = 0;
> > > > - goto out;
> > > > }
> > > >
> > > > - if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
> > > > + if (ret == 0 && args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
> > > > struct sync_file *sync_file = sync_file_create(submit->user_fence);
> > > > if (!sync_file) {
> > > > ret = -ENOMEM;
> > > > - goto out;
> > > > + } else {
> > > > + fd_install(out_fence_fd, sync_file->file);
> > > > + args->fence_fd = out_fence_fd;
> > > > }
> > > > - fd_install(out_fence_fd, sync_file->file);
> > > > - args->fence_fd = out_fence_fd;
> > >
> > > I wonder if instead we should (approximately) undo "drm/msm/submit:
> > > Simplify out-fence-fd handling" so that the point that it could fail
> > > is moved up ahead of the drm_sched_job_arm()?
> >
> > Hm yeah. Up to you how you want to paint this shed, I think either is fine.
> >
> > > Also, does the dma_fence_get() work before drm_sched_job_arm()? From
> > > a quick look, it looks like it won't, but I'm still playing catchup
> > > and haven't had a chance to look at your entire series. If it doesn't
> > > work before drm_sched_job_arm(), then there is really no way to
> > > prevent a error path between the fence-init and job-submit.
> >
> > Yes. I thought I've checked that I put the _arm() in the right spot,
> > but I guess I screwed up and you need the fence before the point where
> > I've put the job_arm()? And yes the error path cannot be avoided for
> > out-fences, that's what I tried to explain in the commit message.
> >
> > > But, prior to your series, wouldn't a failure after
> > > drm_sched_job_init() but before the job is submitted just burn a
> > > fence-id, and otherwise carry on it's merry way?
> >
> > Maybe? I'm not sure whether the scheduler gets confused about the gap
> > and freak out abou that. I'm fairly new to that code and learning
> > (which is part why I'm working on it). Since you look up in
> > fences/syncobj after job_init() it should be pretty easy to whip up a
> > testcase and see what happens. Also as long as nothing fails you won't
> > see an issue, that's for sure.
>
> fair.. I'll try to come up with a test case.. pre-scheduler-conversion
> it wasn't a problem to fail after the fence seqno was allocated (well,
> I guess you might have problems if you had 2^31 failures in a row)
Yeah one thing drm/sched forces you to do is have a very clear notion
about the point of no return in your submit ioctl. Which I think is a
Very Good Thing, at least looking at i915 execbuf where the point of
no return is a multi-stage thing with such interesting intermediate
points like "we submit the ruquest but without actually running the
batchbuffer". The downside is that the submit ioctl isn't perfectly
transaction anymore, but I don't think that matters for tha tail
stuff, which is generally just some out-fence installing. That
generally never fails.
-Daniel
>
> BR,
> -R
>
> > -Daniel
> >
> > > BR,
> > > -R
> > >
> > > > }
> > > >
> > > > submit_attach_object_fences(submit);
> > > > --
> > > > 2.32.0
> > > >
> >
> >
> >
> > --
> > Daniel Vetter
> > Software Engineer, Intel Corporation
> > http://blog.ffwll.ch
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
More information about the Intel-gfx
mailing list