Quoting Tvrtko Ursulin (2019-05-07 13:46:45)
>
> On 03/05/2019 12:52, Chris Wilson wrote:
> > The handling of the no-preemption priority level imposes the restriction
> > that we need to maintain the implied ordering even though preemption is
> > disabled. Otherwise we may end up with an AB-BA deadlock across multiple
> > engine due to a real preemption event reordering the no-preemption
> > WAITs. To resolve this issue we currently promote all requests to WAIT
> > on unsubmission, however this interferes with the timeslicing
> > requirement that we do not apply any implicit promotion that will defeat
> > the round-robin timeslice list. (If we automatically promote the active
> > request it will go back to the head of the queue and not the tail!)
> >
> > So we need implicit promotion to prevent reordering around semaphores
> > where we are not allowed to preempt, and we must avoid implicit
> > promotion on unsubmission. So instead of at unsubmit, if we apply that
> > implicit promotion on adding the dependency, we avoid the semaphore
> > deadlock and we also reduce the gains made by the promotion for user
> > space waiting. Furthermore, by keeping the earlier dependencies at a
> > higher level, we reduce the search space for timeslicing without
> > altering runtime scheduling too badly (no dependencies at all will be
> > assigned a higher priority for rrul).
> >
> > v2: Limit the bump to external edges (as originally intended) i.e.
> > between contexts and out to the user.
> >
> > Testcase: igt/gem_concurrent_blit
> > Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > ---
> > drivers/gpu/drm/i915/gt/selftest_lrc.c | 12 ++++++++----
> > drivers/gpu/drm/i915/i915_request.c | 9 ---------
> > drivers/gpu/drm/i915/i915_scheduler.c | 11 +++++++++++
> > drivers/gpu/drm/i915/i915_scheduler_types.h | 3 ++-
> > 4 files changed, 21 insertions(+), 14 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/gt/selftest_lrc.c b/drivers/gpu/drm/i915/gt/selftest_lrc.c
> > index 4b042893dc0e..5b3d8e33f1cf 100644
> > --- a/drivers/gpu/drm/i915/gt/selftest_lrc.c
> > +++ b/drivers/gpu/drm/i915/gt/selftest_lrc.c
> > @@ -98,12 +98,14 @@ static int live_busywait_preempt(void *arg)
> > ctx_hi = kernel_context(i915);
> > if (!ctx_hi)
> > goto err_unlock;
> > - ctx_hi->sched.priority = INT_MAX;
> > + ctx_hi->sched.priority =
> > + I915_USER_PRIORITY(I915_CONTEXT_MAX_USER_PRIORITY);
> >
> > ctx_lo = kernel_context(i915);
> > if (!ctx_lo)
> > goto err_ctx_hi;
> > - ctx_lo->sched.priority = INT_MIN;
> > + ctx_lo->sched.priority =
> > + I915_USER_PRIORITY(I915_CONTEXT_MIN_USER_PRIORITY);
> >
> > obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
> > if (IS_ERR(obj)) {
> > @@ -958,12 +960,14 @@ static int live_preempt_hang(void *arg)
> > ctx_hi = kernel_context(i915);
> > if (!ctx_hi)
> > goto err_spin_lo;
> > - ctx_hi->sched.priority = I915_CONTEXT_MAX_USER_PRIORITY;
> > + ctx_hi->sched.priority =
> > + I915_USER_PRIORITY(I915_CONTEXT_MAX_USER_PRIORITY);
> >
> > ctx_lo = kernel_context(i915);
> > if (!ctx_lo)
> > goto err_ctx_hi;
> > - ctx_lo->sched.priority = I915_CONTEXT_MIN_USER_PRIORITY;
> > + ctx_lo->sched.priority =
> > + I915_USER_PRIORITY(I915_CONTEXT_MIN_USER_PRIORITY);
> >
> > for_each_engine(engine, i915, id) {
> > struct i915_request *rq;
> > diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
> > index 8cb3ed5531e3..065da1bcbb4c 100644
> > --- a/drivers/gpu/drm/i915/i915_request.c
> > +++ b/drivers/gpu/drm/i915/i915_request.c
> > @@ -468,15 +468,6 @@ void __i915_request_unsubmit(struct i915_request *request)
> > /* We may be recursing from the signal callback of another i915 fence */
> > spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
> >
> > - /*
> > - * As we do not allow WAIT to preempt inflight requests,
> > - * once we have executed a request, along with triggering
> > - * any execution callbacks, we must preserve its ordering
> > - * within the non-preemptible FIFO.
> > - */
> > - BUILD_BUG_ON(__NO_PREEMPTION & ~I915_PRIORITY_MASK); /* only internal */
> > - request->sched.attr.priority |= __NO_PREEMPTION;
> > -
> > if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
> > i915_request_cancel_breadcrumb(request);
> >
> > diff --git a/drivers/gpu/drm/i915/i915_scheduler.c b/drivers/gpu/drm/i915/i915_scheduler.c
> > index 380cb7343a10..ff0ca5718f97 100644
> > --- a/drivers/gpu/drm/i915/i915_scheduler.c
> > +++ b/drivers/gpu/drm/i915/i915_scheduler.c
> > @@ -391,6 +391,16 @@ bool __i915_sched_node_add_dependency(struct i915_sched_node *node,
> > !node_started(signal))
> > node->flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN;
> >
> > + /*
> > + * As we do not allow WAIT to preempt inflight requests,
> > + * once we have executed a request, along with triggering
> > + * any execution callbacks, we must preserve its ordering
> > + * within the non-preemptible FIFO.
> > + */
> > + BUILD_BUG_ON(__NO_PREEMPTION & ~I915_PRIORITY_MASK);
> > + if (flags & I915_DEPENDENCY_EXTERNAL)
> > + __bump_priority(signal, __NO_PREEMPTION);
> > +
>
> I don't really follow how can this be okay from here. It gives wait
> priority to every request which has a dependency now. Which sounds not
> far off from removing the priority bump for waiters altogether. Or
> reversing things and giving requests with no priority a boost.
It treats even inter-context wait equally, and so reduces the effect of
the user wait from the wait/set-domain ioctl and instead includes all
the waits they supply during execbuf.
It all stems from the way those priorities are ignored for preemption.
Currently we perform the same boost implicitly to all running requests,
which screws up timeslicing, so instead of doing it on the running
request we apply it to the queue and limit it to just the edges that are
susceptible to deadlocks (so in effect we are reducing the number of
requests that have the implicit boost).
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
