On Fri, Nov 12, 2021 at 02:13:50PM +0000, Tvrtko Ursulin wrote:
>
> On 11/11/2021 16:49, Matthew Brost wrote:
> > On Mon, Nov 01, 2021 at 10:35:09AM +0000, Tvrtko Ursulin wrote:
> > >
> > > On 27/10/2021 21:10, Matthew Brost wrote:
> > > > On Wed, Oct 27, 2021 at 01:04:49PM -0700, John Harrison wrote:
> > > > > On 10/27/2021 12:17, Matthew Brost wrote:
> > > > > > On Tue, Oct 26, 2021 at 02:58:00PM -0700, John Harrison wrote:
> > > > > > > On 10/20/2021 14:47, Matthew Brost wrote:
> > > > > > > > A weak implementation of parallel submission (multi-bb execbuf
> > > > > > > > IOCTL) for
> > > > > > > > execlists. Doing as little as possible to support this
> > > > > > > > interface for
> > > > > > > > execlists - basically just passing submit fences between each
> > > > > > > > request
> > > > > > > > generated and virtual engines are not allowed. This is on par
> > > > > > > > with what
> > > > > > > > is there for the existing (hopefully soon deprecated) bonding
> > > > > > > > interface.
> > > > > > > >
> > > > > > > > We perma-pin these execlists contexts to align with GuC
> > > > > > > > implementation.
> > > > > > > >
> > > > > > > > v2:
> > > > > > > > (John Harrison)
> > > > > > > > - Drop siblings array as num_siblings must be 1
> > > > > > > >
> > > > > > > > Signed-off-by: Matthew Brost <matthew.br...@intel.com>
> > > > > > > > ---
> > > > > > > > drivers/gpu/drm/i915/gem/i915_gem_context.c | 10 +++--
> > > > > > > > drivers/gpu/drm/i915/gt/intel_context.c | 4 +-
> > > > > > > > .../drm/i915/gt/intel_execlists_submission.c | 44
> > > > > > > > ++++++++++++++++++-
> > > > > > > > drivers/gpu/drm/i915/gt/intel_lrc.c | 2 +
> > > > > > > > .../gpu/drm/i915/gt/uc/intel_guc_submission.c | 2 -
> > > > > > > > 5 files changed, 52 insertions(+), 10 deletions(-)
> > > > > > > >
> > > > > > > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c
> > > > > > > > b/drivers/gpu/drm/i915/gem/i915_gem_context.c
> > > > > > > > index fb33d0322960..35e87a7d0ea9 100644
> > > > > > > > --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c
> > > > > > > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c
> > > > > > > > @@ -570,10 +570,6 @@
> > > > > > > > set_proto_ctx_engines_parallel_submit(struct
> > > > > > > > i915_user_extension __user *base,
> > > > > > > > struct intel_engine_cs **siblings = NULL;
> > > > > > > > intel_engine_mask_t prev_mask;
> > > > > > > > - /* FIXME: This is NIY for execlists */
> > > > > > > > - if (!(intel_uc_uses_guc_submission(&i915->gt.uc)))
> > > > > > > > - return -ENODEV;
> > > > > > > > -
> > > > > > > > if (get_user(slot, &ext->engine_index))
> > > > > > > > return -EFAULT;
> > > > > > > > @@ -583,6 +579,12 @@
> > > > > > > > set_proto_ctx_engines_parallel_submit(struct
> > > > > > > > i915_user_extension __user *base,
> > > > > > > > if (get_user(num_siblings, &ext->num_siblings))
> > > > > > > > return -EFAULT;
> > > > > > > > + if (!intel_uc_uses_guc_submission(&i915->gt.uc) &&
> > > > > > > > num_siblings != 1) {
> > > > > > > > + drm_dbg(&i915->drm, "Only 1 sibling (%d)
> > > > > > > > supported in non-GuC mode\n",
> > > > > > > > + num_siblings);
> > > > > > > > + return -EINVAL;
> > > > > > > > + }
> > > > > > > > +
> > > > > > > > if (slot >= set->num_engines) {
> > > > > > > > drm_dbg(&i915->drm, "Invalid placement value,
> > > > > > > > %d >= %d\n",
> > > > > > > > slot, set->num_engines);
> > > > > > > > diff --git a/drivers/gpu/drm/i915/gt/intel_context.c
> > > > > > > > b/drivers/gpu/drm/i915/gt/intel_context.c
> > > > > > > > index 5634d14052bc..1bec92e1d8e6 100644
> > > > > > > > --- a/drivers/gpu/drm/i915/gt/intel_context.c
> > > > > > > > +++ b/drivers/gpu/drm/i915/gt/intel_context.c
> > > > > > > > @@ -79,7 +79,8 @@ static int
> > > > > > > > intel_context_active_acquire(struct intel_context *ce)
> > > > > > > > __i915_active_acquire(&ce->active);
> > > > > > > > - if (intel_context_is_barrier(ce) ||
> > > > > > > > intel_engine_uses_guc(ce->engine))
> > > > > > > > + if (intel_context_is_barrier(ce) ||
> > > > > > > > intel_engine_uses_guc(ce->engine) ||
> > > > > > > > + intel_context_is_parallel(ce))
> > > > > > > > return 0;
> > > > > > > > /* Preallocate tracking nodes */
> > > > > > > > @@ -563,7 +564,6 @@ void intel_context_bind_parent_child(struct
> > > > > > > > intel_context *parent,
> > > > > > > > * Callers responsibility to validate that this
> > > > > > > > function is used
> > > > > > > > * correctly but we use GEM_BUG_ON here ensure that
> > > > > > > > they do.
> > > > > > > > */
> > > > > > > > - GEM_BUG_ON(!intel_engine_uses_guc(parent->engine));
> > > > > > > > GEM_BUG_ON(intel_context_is_pinned(parent));
> > > > > > > > GEM_BUG_ON(intel_context_is_child(parent));
> > > > > > > > GEM_BUG_ON(intel_context_is_pinned(child));
> > > > > > > > diff --git
> > > > > > > > a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
> > > > > > > > b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
> > > > > > > > index bedb80057046..2865b422300d 100644
> > > > > > > > --- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
> > > > > > > > +++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
> > > > > > > > @@ -927,8 +927,7 @@ static void execlists_submit_ports(struct
> > > > > > > > intel_engine_cs *engine)
> > > > > > > > static bool ctx_single_port_submission(const struct
> > > > > > > > intel_context *ce)
> > > > > > > > {
> > > > > > > > - return (IS_ENABLED(CONFIG_DRM_I915_GVT) &&
> > > > > > > > - intel_context_force_single_submission(ce));
> > > > > > > > + return intel_context_force_single_submission(ce);
> > > > > > > I think this is actually going to break GVT.
> > > > > > >
> > > > > > > Not so much this change here but the whole use of single
> > > > > > > submission outside
> > > > > > > of GVT. It looks like the GVT driver overloads the single
> > > > > > > submission flag to
> > > > > > > tag requests that it owns. If we start using that flag elsewhere
> > > > > > > when GVT is
> > > > > > > active, I think that will cause much confusion within the GVT
> > > > > > > code.
> > > > > > >
> > > > > > > The correct fix would be to create a new flag just for GVT usage
> > > > > > > alongside
> > > > > > > the single submission one. GVT would then set both but only check
> > > > > > > for its
> > > > > > > own private flag. The parallel code would obviously only set the
> > > > > > > existing
> > > > > > > single submission flag.
> > > > > > >
> > > > > > Ok, see below.
> > > > > >
> > > > > > > > }
> > > > > > > > static bool can_merge_ctx(const struct intel_context *prev,
> > > > > > > > @@ -2598,6 +2597,46 @@ static void
> > > > > > > > execlists_context_cancel_request(struct intel_context *ce,
> > > > > > > > current->comm);
> > > > > > > > }
> > > > > > > > +static struct intel_context *
> > > > > > > > +execlists_create_parallel(struct intel_engine_cs **engines,
> > > > > > > > + unsigned int num_siblings,
> > > > > > > > + unsigned int width)
> > > > > > > > +{
> > > > > > > > + struct intel_context *parent = NULL, *ce, *err;
> > > > > > > > + int i;
> > > > > > > > +
> > > > > > > > + GEM_BUG_ON(num_siblings != 1);
> > > > > > > > +
> > > > > > > > + for (i = 0; i < width; ++i) {
> > > > > > > > + ce = intel_context_create(engines[i]);
> > > > > > > > + if (!ce) {
> > > > > > > > + err = ERR_PTR(-ENOMEM);
> > > > > > > > + goto unwind;
> > > > > > > > + }
> > > > > > > > +
> > > > > > > > + if (i == 0)
> > > > > > > > + parent = ce;
> > > > > > > > + else
> > > > > > > > + intel_context_bind_parent_child(parent,
> > > > > > > > ce);
> > > > > > > > + }
> > > > > > > > +
> > > > > > > > + parent->parallel.fence_context =
> > > > > > > > dma_fence_context_alloc(1);
> > > > > > > > +
> > > > > > > > + intel_context_set_nopreempt(parent);
> > > > > > > > + intel_context_set_single_submission(parent);
> > > > > > > Can you explain the need for setting single submission?
> > > > > > >
> > > > > > I think I can actually pull this out. This was needed when I tried
> > > > > > to
> > > > > > truely implement a guarante that all the parallel requests would be
> > > > > > running simultaneously. Couldn't ever to get that working because
> > > > > > of the
> > > > > > mess that is the execlists scheduler - a simple wait at the head of
> > > > > > queue until everyone joined just blew up for whatever reason. I
> > > > > > don't
> > > > > > believe this servers a purpose anymore, so I'll just drop it.
> > > > > >
> > > > > > Matt
> > > > > Is that not going to be a problem? I thought concurrent execution was
> > > > > a
> > > > > fundamental requirement?
> > > > >
> > > >
> > > > I don't think so. See the commit message. This implmementation is on par
> > > > with the bonding interface - there is no guarantee whatsoever that with
> > > > the bonding interface bonded requests actually run at the same time. It
> > > > says hopefully these submissions run together. That's what I do in this
> > > > patch too for execlists, hence the 'weak' clause in the commit message.
> > >
> > > With the new uapi definition implying a stricter guarantee - why not have
> > > this patch use special bb semaphore pre/post-ambles so scheduling
> > > behaviour
> > > is closer between the two backends?
> > >
> >
> > We could do that in a follow up if needed, as this bare minimum to get
> > this uAPI working. The real fix would be update the execlists scheduler
> > to be able to do a join of parallel requests and then schedule them
> > together. Should be fairly simple, tried to do this, but the execlists
> > scheduler is such a mess doing something simple is near impossible. IMO
> > there is little point wasting time on a legacy submission interface.
> > This implementation works as well as the old uAPI, let's get this in and
> > move on.
>
> Bashing aside, what downside do you see in just doing what I suggested right
> now? Code is there and all so it is a simple matter of adding a conditional
> somewhere to use it. And it would make the behaviour between the two
> backends closer. So it sounds like a no brainer to me. Or I am missing
> something?
For parallel submission, user batches should be inserting semaphore to
ensure that they are running together - the kernel inserting them is
redundant. The reason we do this for GuC submission is for safe
preemption, in execlists we just don't allow preemption while the
requests are running. As I said, the correct solution is update the
execlists scheduler to actually run these requests in parallel. Tried
that but proved difficult and landed on this patch. If someone wants to
fix the the execlists scheduler in a follow up they are welcome to but
in the meantime what I have in place is on par with the bonded
interface. I see no reason why this patch can't be merged.
Matt
>
> Regards,
>
> Tvrtko
