Quoting Tvrtko Ursulin (2019-10-11 14:55:00)
>
> On 10/10/2019 08:14, Chris Wilson wrote:
> > Normally, we rely on our hangcheck to prevent persistent batches from
> > hogging the GPU. However, if the user disables hangcheck, this mechanism
> > breaks down. Despite our insistence that this is unsafe, the users are
> > equally insistent that they want to use endless batches and will disable
> > the hangcheck mechanism. We are looking at perhaps replacing hangcheck
> > with a softer mechanism, that sends a pulse down the engine to check if
> > it is well. We can use the same preemptive pulse to flush an active
> > persistent context off the GPU upon context close, preventing resources
> > being lost and unkillable requests remaining on the GPU after process
> > termination. To avoid changing the ABI and accidentally breaking
> > existing userspace, we make the persistence of a context explicit and
> > enable it by default (matching current ABI). Userspace can opt out of
> > persistent mode (forcing requests to be cancelled when the context is
> > closed by process termination or explicitly) by a context parameter. To
> > facilitate existing use-cases of disabling hangcheck, if the modparam is
> > disabled (i915.enable_hangcheck=0), we disable persistence mode by
> > default. (Note, one of the outcomes for supporting endless mode will be
> > the removal of hangchecking, at which point opting into persistent mode
> > will be mandatory, or maybe the default perhaps controlled by cgroups.)
> >
> > v2: Check for hangchecking at context termination, so that we are not
> > left with undying contexts from a crafty user.
> >
> > Testcase: igt/gem_ctx_persistence
> > Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> > Cc: Joonas Lahtinen <joonas.lahtinen@xxxxxxxxxxxxxxx>
> > Cc: Michał Winiarski <michal.winiarski@xxxxxxxxx>
> > Cc: Jon Bloomfield <jon.bloomfield@xxxxxxxxx>
> > Reviewed-by: Jon Bloomfield <jon.bloomfield@xxxxxxxxx>
> > ---
> > drivers/gpu/drm/i915/gem/i915_gem_context.c | 132 ++++++++++++++++++
> > drivers/gpu/drm/i915/gem/i915_gem_context.h | 15 ++
> > .../gpu/drm/i915/gem/i915_gem_context_types.h | 1 +
> > .../gpu/drm/i915/gem/selftests/mock_context.c | 2 +
> > include/uapi/drm/i915_drm.h | 15 ++
> > 5 files changed, 165 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c
> > index 5d8221c7ba83..46e5b3b53288 100644
> > --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c
> > +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c
> > @@ -70,6 +70,7 @@
> > #include <drm/i915_drm.h>
> >
> > #include "gt/intel_lrc_reg.h"
> > +#include "gt/intel_engine_heartbeat.h"
> > #include "gt/intel_engine_user.h"
> >
> > #include "i915_gem_context.h"
> > @@ -269,6 +270,78 @@ void i915_gem_context_release(struct kref *ref)
> > schedule_work(&gc->free_work);
> > }
> >
> > +static inline struct i915_gem_engines *
> > +__context_engines_static(struct i915_gem_context *ctx)
> > +{
> > + return rcu_dereference_protected(ctx->engines, true);
> > +}
> > +
> > +static void kill_context(struct i915_gem_context *ctx)
> > +{
> > + intel_engine_mask_t tmp, active, reset;
> > + struct intel_gt *gt = &ctx->i915->gt;
> > + struct i915_gem_engines_iter it;
> > + struct intel_engine_cs *engine;
> > + struct intel_context *ce;
> > +
> > + /*
> > + * If we are already banned, it was due to a guilty request causing
> > + * a reset and the entire context being evicted from the GPU.
> > + */
> > + if (i915_gem_context_is_banned(ctx))
> > + return;
> > +
> > + i915_gem_context_set_banned(ctx);
> > +
> > + /*
> > + * Map the user's engine back to the actual engines; one virtual
> > + * engine will be mapped to multiple engines, and using ctx->engine[]
> > + * the same engine may be have multiple instances in the user's map.
> > + * However, we only care about pending requests, so only include
> > + * engines on which there are incomplete requests.
> > + */
> > + active = 0;
> > + for_each_gem_engine(ce, __context_engines_static(ctx), it) {
> > + struct dma_fence *fence;
> > +
> > + if (!ce->timeline)
> > + continue;
> > +
> > + fence = i915_active_fence_get(&ce->timeline->last_request);
> > + if (!fence)
> > + continue;
> > +
> > + engine = to_request(fence)->engine;
> > + if (HAS_EXECLISTS(gt->i915))
> > + engine = intel_context_inflight(ce);
>
> Okay preemption implies execlists, was confused for a moment.
>
> When can engine be NULL here?
The engine is not paused, and an interrupt can cause a schedule-out as
we gather up the state.
>
> > + if (engine)
> > + active |= engine->mask;
> > +
> > + dma_fence_put(fence);
> > + }
> > +
> > + /*
> > + * Send a "high priority pulse" down the engine to cause the
> > + * current request to be momentarily preempted. (If it fails to
> > + * be preempted, it will be reset). As we have marked our context
> > + * as banned, any incomplete request, including any running, will
> > + * be skipped following the preemption.
> > + */
> > + reset = 0;
> > + for_each_engine_masked(engine, gt->i915, active, tmp)
> > + if (intel_engine_pulse(engine))
> > + reset |= engine->mask;
>
> What if we were able to send a pulse, but the hog cannot be preempted
> and hangcheck is obviously disabled - who will do the reset?
Hmm, the idea is that forced-preemption causes the reset.
(See igt/gem_ctx_persistence/hostile)
However, if we give the sysadmin the means to disable force-preemption,
we just gave them another shovel to dig a hole with.
A last resort would be another timer here to ensure the context was
terminated.
-Chris
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
