On Thu, Jun 24, 2021 at 09:59:08AM -0700, Emma Anholt wrote: > On Thu, Jun 24, 2021 at 7:00 AM Daniel Vetter <daniel.vetter@xxxxxxxx> wrote: > > > > Prep work for using the scheduler dependency handling. I'll add "We need to call drm_sched_job_init earlier so we can use the new drm_sched_job_await* functions for dependency handling here. That gives a bit more context on what's going on here. > > > > Signed-off-by: Daniel Vetter <daniel.vetter@xxxxxxxxx> > > Cc: Emma Anholt <emma@xxxxxxxxxx> > > Back when I wrote this, I think there were rules that there had to be > no failure paths between a job_init and a push. Has that changed? > > I really don't have the context to evaluate this, I'm not sure what > new "scheduler dependency handling" is given that there was already > something that I considered to be dependency handling! Full patch series link: https://lore.kernel.org/dri-devel/20210624140025.438303-1-daniel.vetter@xxxxxxxx/T/#t The job_init vs push_job is addressed here: https://lore.kernel.org/dri-devel/20210624140025.438303-1-daniel.vetter@xxxxxxxx/T/#mb6c4d75e1c57a5056d7b2ec8fbb9839fc5be41a7 I split job_init into job_init (which can fail, and be done earlier) and job_arm (which can't fail, and must be atomic with the push_job). The entire goal of this is to lift the dependency handling from "everyone copypastes v3d" to "drm_sched has it for you already", which is this patch here: https://lore.kernel.org/dri-devel/20210624140025.438303-1-daniel.vetter@xxxxxxxx/T/#m9f64aaf840cbf8815cd2ea9a68f99a51da9baa5f Cheers, Daniel -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
