[07/10] drm/i915/gem: Cancel non-persistent contexts on close
diff mbox series

Message ID 20191014220534.1662-7-chris@chris-wilson.co.uk
State New
Headers show
Series
  • [01/10] drm/i915/gem: Distinguish each object type
Related show

Commit Message

Chris Wilson Oct. 14, 2019, 10:05 p.m. UTC
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.
v3: Force context termination even if forced-preemption is disabled.

Testcase: igt/gem_ctx_persistence
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Michał Winiarski <michal.winiarski@intel.com>
Cc: Jon Bloomfield <jon.bloomfield@intel.com>
Reviewed-by: Jon Bloomfield <jon.bloomfield@intel.com>
---
 drivers/gpu/drm/i915/gem/i915_gem_context.c   | 189 ++++++++++++++++++
 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, 222 insertions(+)

Comments

Tvrtko Ursulin Oct. 15, 2019, 12:15 p.m. UTC | #1
On 14/10/2019 23:05, 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.
> v3: Force context termination even if forced-preemption is disabled.
> 
> Testcase: igt/gem_ctx_persistence
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
> Cc: Michał Winiarski <michal.winiarski@intel.com>
> Cc: Jon Bloomfield <jon.bloomfield@intel.com>
> Reviewed-by: Jon Bloomfield <jon.bloomfield@intel.com>
> ---
>   drivers/gpu/drm/i915/gem/i915_gem_context.c   | 189 ++++++++++++++++++
>   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, 222 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..49f37bba5693 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 +270,135 @@ void i915_gem_context_release(struct kref *ref)
>   		schedule_work(&gc->free_work);
>   }
>   
> +static inline struct i915_gem_engines *
> +__context_engines_static(const struct i915_gem_context *ctx)
> +{
> +	return rcu_dereference_protected(ctx->engines, true);
> +}
> +
> +static bool __reset_engine(struct intel_engine_cs *engine)
> +{
> +	struct intel_gt *gt = engine->gt;
> +	bool success = false;
> +
> +	if (!intel_has_reset_engine(gt))
> +		return false;
> +
> +	if (!test_and_set_bit(I915_RESET_ENGINE + engine->id,
> +			      &gt->reset.flags)) {
> +		success = intel_engine_reset(engine, NULL) == 0;
> +		clear_and_wake_up_bit(I915_RESET_ENGINE + engine->id,
> +				      &gt->reset.flags);
> +	}
> +
> +	return success;
> +}
> +
> +static void __reset_context(struct i915_gem_context *ctx,
> +			    struct intel_engine_cs *engine)
> +{
> +	intel_gt_handle_error(engine->gt, engine->mask, 0,
> +			      "context closure in %s", ctx->name);
> +}
> +
> +static bool __cancel_engine(struct intel_engine_cs *engine)
> +{
> +	/*
> +	 * 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.
> +	 *
> +	 * If there is no hangchecking (one of the reasons why we try to
> +	 * cancel the context) and no forced preemption, there may be no
> +	 * means by which we reset the GPU and evict the persistent hog.
> +	 * Ergo if we are unable to inject a preemptive pulse that can
> +	 * kill the banned context, we fallback to doing a local reset
> +	 * instead.
> +	 */
> +	if (CONFIG_DRM_I915_PREEMPT_TIMEOUT && !intel_engine_pulse(engine))
> +		return true;
> +
> +	/* If we are unable to send a pulse, try resetting this engine. */
> +	return __reset_engine(engine);

Open from last round is how likely is forced preemption to be compiled 
out, in which case the code will always fall back to reset immediately, 
even if workload would otherwise preempt just fine. (Given the dangers 
of reset hitting something unrelated as you explained.)

Could you always compile in forced preemption so it could be used on 
context close? I am thinking, allow it to be disabled via sysfs on its 
own, but on context close have it active for the context being closed so 
that if the pulse does not work it can kick in and reset. Sounds nicer 
than just resorting to engine reset, which as you described, can impact 
someone innocent.

Regards,

Tvrtko

> +}
> +
> +static struct intel_engine_cs *
> +active_engine(struct dma_fence *fence, struct intel_context *ce)
> +{
> +	struct i915_request *rq = to_request(fence);
> +	struct intel_engine_cs *engine, *locked;
> +
> +	/*
> +	 * Serialise with __i915_request_submit() so that it sees
> +	 * is-banned?, or we know the request is already inflight.
> +	 */
> +	locked = READ_ONCE(rq->engine);
> +	spin_lock_irq(&locked->active.lock);
> +	while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) {
> +		spin_unlock(&locked->active.lock);
> +		spin_lock(&engine->active.lock);
> +		locked = engine;
> +	}
> +
> +	engine = NULL;
> +	if (i915_request_is_active(rq) && !rq->fence.error)
> +		engine = rq->engine;
> +
> +	spin_unlock_irq(&locked->active.lock);
> +
> +	return engine;
> +}
> +
> +static void kill_context(struct i915_gem_context *ctx)
> +{
> +	struct i915_gem_engines_iter it;
> +	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.
> +	 */
> +	for_each_gem_engine(ce, __context_engines_static(ctx), it) {
> +		struct intel_engine_cs *engine;
> +		struct dma_fence *fence;
> +
> +		if (!ce->timeline)
> +			continue;
> +
> +		fence = i915_active_fence_get(&ce->timeline->last_request);
> +		if (!fence)
> +			continue;
> +
> +		/* Check with the backend if the request is still inflight */
> +		engine = active_engine(fence, ce);
> +
> +		/* First attempt to gracefully cancel the context */
> +		if (engine && !__cancel_engine(engine))
> +			/*
> +			 * If we are unable to send a preemptive pulse to bump
> +			 * the context from the GPU, we have to resort to a full
> +			 * reset. We hope the collateral damage is worth it.
> +			 */
> +			__reset_context(ctx, engine);
> +
> +		dma_fence_put(fence);
> +	}
> +}
> +
>   static void context_close(struct i915_gem_context *ctx)
>   {
>   	struct i915_address_space *vm;
> @@ -291,9 +421,47 @@ static void context_close(struct i915_gem_context *ctx)
>   	lut_close(ctx);
>   
>   	mutex_unlock(&ctx->mutex);
> +
> +	/*
> +	 * If the user has disabled hangchecking, we can not be sure that
> +	 * the batches will ever complete after the context is closed,
> +	 * keeping the context and all resources pinned forever. So in this
> +	 * case we opt to forcibly kill off all remaining requests on
> +	 * context close.
> +	 */
> +	if (!i915_gem_context_is_persistent(ctx) ||
> +	    !i915_modparams.enable_hangcheck)
> +		kill_context(ctx);
> +
>   	i915_gem_context_put(ctx);
>   }
>   
> +static int __context_set_persistence(struct i915_gem_context *ctx, bool state)
> +{
> +	if (i915_gem_context_is_persistent(ctx) == state)
> +		return 0;
> +
> +	if (state) {
> +		/*
> +		 * Only contexts that are short-lived [that will expire or be
> +		 * reset] are allowed to survive past termination. We require
> +		 * hangcheck to ensure that the persistent requests are healthy.
> +		 */
> +		if (!i915_modparams.enable_hangcheck)
> +			return -EINVAL;
> +
> +		i915_gem_context_set_persistence(ctx);
> +	} else {
> +		/* To cancel a context we use "preempt-to-idle" */
> +		if (!(ctx->i915->caps.scheduler & I915_SCHEDULER_CAP_PREEMPTION))
> +			return -ENODEV;
> +
> +		i915_gem_context_clear_persistence(ctx);
> +	}
> +
> +	return 0;
> +}
> +
>   static struct i915_gem_context *
>   __create_context(struct drm_i915_private *i915)
>   {
> @@ -328,6 +496,7 @@ __create_context(struct drm_i915_private *i915)
>   
>   	i915_gem_context_set_bannable(ctx);
>   	i915_gem_context_set_recoverable(ctx);
> +	__context_set_persistence(ctx, true /* cgroup hook? */);
>   
>   	for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++)
>   		ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES;
> @@ -484,6 +653,7 @@ i915_gem_context_create_kernel(struct drm_i915_private *i915, int prio)
>   		return ctx;
>   
>   	i915_gem_context_clear_bannable(ctx);
> +	i915_gem_context_set_persistence(ctx);
>   	ctx->sched.priority = I915_USER_PRIORITY(prio);
>   
>   	GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
> @@ -1594,6 +1764,16 @@ get_engines(struct i915_gem_context *ctx,
>   	return err;
>   }
>   
> +static int
> +set_persistence(struct i915_gem_context *ctx,
> +		const struct drm_i915_gem_context_param *args)
> +{
> +	if (args->size)
> +		return -EINVAL;
> +
> +	return __context_set_persistence(ctx, args->value);
> +}
> +
>   static int ctx_setparam(struct drm_i915_file_private *fpriv,
>   			struct i915_gem_context *ctx,
>   			struct drm_i915_gem_context_param *args)
> @@ -1671,6 +1851,10 @@ static int ctx_setparam(struct drm_i915_file_private *fpriv,
>   		ret = set_engines(ctx, args);
>   		break;
>   
> +	case I915_CONTEXT_PARAM_PERSISTENCE:
> +		ret = set_persistence(ctx, args);
> +		break;
> +
>   	case I915_CONTEXT_PARAM_BAN_PERIOD:
>   	default:
>   		ret = -EINVAL;
> @@ -2123,6 +2307,11 @@ int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
>   		ret = get_engines(ctx, args);
>   		break;
>   
> +	case I915_CONTEXT_PARAM_PERSISTENCE:
> +		args->size = 0;
> +		args->value = i915_gem_context_is_persistent(ctx);
> +		break;
> +
>   	case I915_CONTEXT_PARAM_BAN_PERIOD:
>   	default:
>   		ret = -EINVAL;
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.h b/drivers/gpu/drm/i915/gem/i915_gem_context.h
> index cfe80590f0ed..18e50a769a6e 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_context.h
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.h
> @@ -76,6 +76,21 @@ static inline void i915_gem_context_clear_recoverable(struct i915_gem_context *c
>   	clear_bit(UCONTEXT_RECOVERABLE, &ctx->user_flags);
>   }
>   
> +static inline bool i915_gem_context_is_persistent(const struct i915_gem_context *ctx)
> +{
> +	return test_bit(UCONTEXT_PERSISTENCE, &ctx->user_flags);
> +}
> +
> +static inline void i915_gem_context_set_persistence(struct i915_gem_context *ctx)
> +{
> +	set_bit(UCONTEXT_PERSISTENCE, &ctx->user_flags);
> +}
> +
> +static inline void i915_gem_context_clear_persistence(struct i915_gem_context *ctx)
> +{
> +	clear_bit(UCONTEXT_PERSISTENCE, &ctx->user_flags);
> +}
> +
>   static inline bool i915_gem_context_is_banned(const struct i915_gem_context *ctx)
>   {
>   	return test_bit(CONTEXT_BANNED, &ctx->flags);
> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context_types.h b/drivers/gpu/drm/i915/gem/i915_gem_context_types.h
> index fe97b8ba4fda..861d7d92fe9f 100644
> --- a/drivers/gpu/drm/i915/gem/i915_gem_context_types.h
> +++ b/drivers/gpu/drm/i915/gem/i915_gem_context_types.h
> @@ -137,6 +137,7 @@ struct i915_gem_context {
>   #define UCONTEXT_NO_ERROR_CAPTURE	1
>   #define UCONTEXT_BANNABLE		2
>   #define UCONTEXT_RECOVERABLE		3
> +#define UCONTEXT_PERSISTENCE		4
>   
>   	/**
>   	 * @flags: small set of booleans
> diff --git a/drivers/gpu/drm/i915/gem/selftests/mock_context.c b/drivers/gpu/drm/i915/gem/selftests/mock_context.c
> index 74ddd682c9cd..29b8984f0e47 100644
> --- a/drivers/gpu/drm/i915/gem/selftests/mock_context.c
> +++ b/drivers/gpu/drm/i915/gem/selftests/mock_context.c
> @@ -22,6 +22,8 @@ mock_context(struct drm_i915_private *i915,
>   	INIT_LIST_HEAD(&ctx->link);
>   	ctx->i915 = i915;
>   
> +	i915_gem_context_set_persistence(ctx);
> +
>   	mutex_init(&ctx->engines_mutex);
>   	e = default_engines(ctx);
>   	if (IS_ERR(e))
> diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
> index 63d40cba97e0..b5fd5e4bd16e 100644
> --- a/include/uapi/drm/i915_drm.h
> +++ b/include/uapi/drm/i915_drm.h
> @@ -1572,6 +1572,21 @@ struct drm_i915_gem_context_param {
>    *   i915_context_engines_bond (I915_CONTEXT_ENGINES_EXT_BOND)
>    */
>   #define I915_CONTEXT_PARAM_ENGINES	0xa
> +
> +/*
> + * I915_CONTEXT_PARAM_PERSISTENCE:
> + *
> + * Allow the context and active rendering to survive the process until
> + * completion. Persistence allows fire-and-forget clients to queue up a
> + * bunch of work, hand the output over to a display server and the quit.
> + * If the context is not marked as persistent, upon closing (either via
> + * an explicit DRM_I915_GEM_CONTEXT_DESTROY or implicitly from file closure
> + * or process termination), the context and any outstanding requests will be
> + * cancelled (and exported fences for cancelled requests marked as -EIO).
> + *
> + * By default, new contexts allow persistence.
> + */
> +#define I915_CONTEXT_PARAM_PERSISTENCE	0xb
>   /* Must be kept compact -- no holes and well documented */
>   
>   	__u64 value;
>
Chris Wilson Oct. 15, 2019, 12:19 p.m. UTC | #2
Quoting Tvrtko Ursulin (2019-10-15 13:15:49)
> 
> On 14/10/2019 23:05, 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.
> > v3: Force context termination even if forced-preemption is disabled.
> > 
> > Testcase: igt/gem_ctx_persistence
> > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> > Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
> > Cc: Michał Winiarski <michal.winiarski@intel.com>
> > Cc: Jon Bloomfield <jon.bloomfield@intel.com>
> > Reviewed-by: Jon Bloomfield <jon.bloomfield@intel.com>
> > ---
> >   drivers/gpu/drm/i915/gem/i915_gem_context.c   | 189 ++++++++++++++++++
> >   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, 222 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..49f37bba5693 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 +270,135 @@ void i915_gem_context_release(struct kref *ref)
> >               schedule_work(&gc->free_work);
> >   }
> >   
> > +static inline struct i915_gem_engines *
> > +__context_engines_static(const struct i915_gem_context *ctx)
> > +{
> > +     return rcu_dereference_protected(ctx->engines, true);
> > +}
> > +
> > +static bool __reset_engine(struct intel_engine_cs *engine)
> > +{
> > +     struct intel_gt *gt = engine->gt;
> > +     bool success = false;
> > +
> > +     if (!intel_has_reset_engine(gt))
> > +             return false;
> > +
> > +     if (!test_and_set_bit(I915_RESET_ENGINE + engine->id,
> > +                           &gt->reset.flags)) {
> > +             success = intel_engine_reset(engine, NULL) == 0;
> > +             clear_and_wake_up_bit(I915_RESET_ENGINE + engine->id,
> > +                                   &gt->reset.flags);
> > +     }
> > +
> > +     return success;
> > +}
> > +
> > +static void __reset_context(struct i915_gem_context *ctx,
> > +                         struct intel_engine_cs *engine)
> > +{
> > +     intel_gt_handle_error(engine->gt, engine->mask, 0,
> > +                           "context closure in %s", ctx->name);
> > +}
> > +
> > +static bool __cancel_engine(struct intel_engine_cs *engine)
> > +{
> > +     /*
> > +      * 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.
> > +      *
> > +      * If there is no hangchecking (one of the reasons why we try to
> > +      * cancel the context) and no forced preemption, there may be no
> > +      * means by which we reset the GPU and evict the persistent hog.
> > +      * Ergo if we are unable to inject a preemptive pulse that can
> > +      * kill the banned context, we fallback to doing a local reset
> > +      * instead.
> > +      */
> > +     if (CONFIG_DRM_I915_PREEMPT_TIMEOUT && !intel_engine_pulse(engine))
> > +             return true;
> > +
> > +     /* If we are unable to send a pulse, try resetting this engine. */
> > +     return __reset_engine(engine);
> 
> Open from last round is how likely is forced preemption to be compiled 
> out, in which case the code will always fall back to reset immediately, 
> even if workload would otherwise preempt just fine. (Given the dangers 
> of reset hitting something unrelated as you explained.)
> 
> Could you always compile in forced preemption so it could be used on 
> context close? I am thinking, allow it to be disabled via sysfs on its 
> own, but on context close have it active for the context being closed so 
> that if the pulse does not work it can kick in and reset. Sounds nicer 
> than just resorting to engine reset, which as you described, can impact 
> someone innocent.

I was thinking that those who care about interruptions would not
tolerate even having a preempt timer running. I would say its masochism
and would not expect it to be disabled, but I also say the same thing
about RT.

In short, I made it optional at compiletime just because I could.
-Chris
Tvrtko Ursulin Oct. 15, 2019, 12:49 p.m. UTC | #3
On 15/10/2019 13:19, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2019-10-15 13:15:49)
>>
>> On 14/10/2019 23:05, 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.
>>> v3: Force context termination even if forced-preemption is disabled.
>>>
>>> Testcase: igt/gem_ctx_persistence
>>> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
>>> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
>>> Cc: Michał Winiarski <michal.winiarski@intel.com>
>>> Cc: Jon Bloomfield <jon.bloomfield@intel.com>
>>> Reviewed-by: Jon Bloomfield <jon.bloomfield@intel.com>
>>> ---
>>>    drivers/gpu/drm/i915/gem/i915_gem_context.c   | 189 ++++++++++++++++++
>>>    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, 222 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..49f37bba5693 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 +270,135 @@ void i915_gem_context_release(struct kref *ref)
>>>                schedule_work(&gc->free_work);
>>>    }
>>>    
>>> +static inline struct i915_gem_engines *
>>> +__context_engines_static(const struct i915_gem_context *ctx)
>>> +{
>>> +     return rcu_dereference_protected(ctx->engines, true);
>>> +}
>>> +
>>> +static bool __reset_engine(struct intel_engine_cs *engine)
>>> +{
>>> +     struct intel_gt *gt = engine->gt;
>>> +     bool success = false;
>>> +
>>> +     if (!intel_has_reset_engine(gt))
>>> +             return false;
>>> +
>>> +     if (!test_and_set_bit(I915_RESET_ENGINE + engine->id,
>>> +                           &gt->reset.flags)) {
>>> +             success = intel_engine_reset(engine, NULL) == 0;
>>> +             clear_and_wake_up_bit(I915_RESET_ENGINE + engine->id,
>>> +                                   &gt->reset.flags);
>>> +     }
>>> +
>>> +     return success;
>>> +}
>>> +
>>> +static void __reset_context(struct i915_gem_context *ctx,
>>> +                         struct intel_engine_cs *engine)
>>> +{
>>> +     intel_gt_handle_error(engine->gt, engine->mask, 0,
>>> +                           "context closure in %s", ctx->name);
>>> +}
>>> +
>>> +static bool __cancel_engine(struct intel_engine_cs *engine)
>>> +{
>>> +     /*
>>> +      * 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.
>>> +      *
>>> +      * If there is no hangchecking (one of the reasons why we try to
>>> +      * cancel the context) and no forced preemption, there may be no
>>> +      * means by which we reset the GPU and evict the persistent hog.
>>> +      * Ergo if we are unable to inject a preemptive pulse that can
>>> +      * kill the banned context, we fallback to doing a local reset
>>> +      * instead.
>>> +      */
>>> +     if (CONFIG_DRM_I915_PREEMPT_TIMEOUT && !intel_engine_pulse(engine))
>>> +             return true;
>>> +
>>> +     /* If we are unable to send a pulse, try resetting this engine. */
>>> +     return __reset_engine(engine);
>>
>> Open from last round is how likely is forced preemption to be compiled
>> out, in which case the code will always fall back to reset immediately,
>> even if workload would otherwise preempt just fine. (Given the dangers
>> of reset hitting something unrelated as you explained.)
>>
>> Could you always compile in forced preemption so it could be used on
>> context close? I am thinking, allow it to be disabled via sysfs on its
>> own, but on context close have it active for the context being closed so
>> that if the pulse does not work it can kick in and reset. Sounds nicer
>> than just resorting to engine reset, which as you described, can impact
>> someone innocent.
> 
> I was thinking that those who care about interruptions would not
> tolerate even having a preempt timer running. I would say its masochism
> and would not expect it to be disabled, but I also say the same thing
> about RT.
> 
> In short, I made it optional at compiletime just because I could.

Okay it can be improved later if need be. Because nothing I am 
suggesting is not an ABI change.

Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

Regards,

Tvrtko

Patch
diff mbox series

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c
index 5d8221c7ba83..49f37bba5693 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 +270,135 @@  void i915_gem_context_release(struct kref *ref)
 		schedule_work(&gc->free_work);
 }
 
+static inline struct i915_gem_engines *
+__context_engines_static(const struct i915_gem_context *ctx)
+{
+	return rcu_dereference_protected(ctx->engines, true);
+}
+
+static bool __reset_engine(struct intel_engine_cs *engine)
+{
+	struct intel_gt *gt = engine->gt;
+	bool success = false;
+
+	if (!intel_has_reset_engine(gt))
+		return false;
+
+	if (!test_and_set_bit(I915_RESET_ENGINE + engine->id,
+			      &gt->reset.flags)) {
+		success = intel_engine_reset(engine, NULL) == 0;
+		clear_and_wake_up_bit(I915_RESET_ENGINE + engine->id,
+				      &gt->reset.flags);
+	}
+
+	return success;
+}
+
+static void __reset_context(struct i915_gem_context *ctx,
+			    struct intel_engine_cs *engine)
+{
+	intel_gt_handle_error(engine->gt, engine->mask, 0,
+			      "context closure in %s", ctx->name);
+}
+
+static bool __cancel_engine(struct intel_engine_cs *engine)
+{
+	/*
+	 * 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.
+	 *
+	 * If there is no hangchecking (one of the reasons why we try to
+	 * cancel the context) and no forced preemption, there may be no
+	 * means by which we reset the GPU and evict the persistent hog.
+	 * Ergo if we are unable to inject a preemptive pulse that can
+	 * kill the banned context, we fallback to doing a local reset
+	 * instead.
+	 */
+	if (CONFIG_DRM_I915_PREEMPT_TIMEOUT && !intel_engine_pulse(engine))
+		return true;
+
+	/* If we are unable to send a pulse, try resetting this engine. */
+	return __reset_engine(engine);
+}
+
+static struct intel_engine_cs *
+active_engine(struct dma_fence *fence, struct intel_context *ce)
+{
+	struct i915_request *rq = to_request(fence);
+	struct intel_engine_cs *engine, *locked;
+
+	/*
+	 * Serialise with __i915_request_submit() so that it sees
+	 * is-banned?, or we know the request is already inflight.
+	 */
+	locked = READ_ONCE(rq->engine);
+	spin_lock_irq(&locked->active.lock);
+	while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) {
+		spin_unlock(&locked->active.lock);
+		spin_lock(&engine->active.lock);
+		locked = engine;
+	}
+
+	engine = NULL;
+	if (i915_request_is_active(rq) && !rq->fence.error)
+		engine = rq->engine;
+
+	spin_unlock_irq(&locked->active.lock);
+
+	return engine;
+}
+
+static void kill_context(struct i915_gem_context *ctx)
+{
+	struct i915_gem_engines_iter it;
+	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.
+	 */
+	for_each_gem_engine(ce, __context_engines_static(ctx), it) {
+		struct intel_engine_cs *engine;
+		struct dma_fence *fence;
+
+		if (!ce->timeline)
+			continue;
+
+		fence = i915_active_fence_get(&ce->timeline->last_request);
+		if (!fence)
+			continue;
+
+		/* Check with the backend if the request is still inflight */
+		engine = active_engine(fence, ce);
+
+		/* First attempt to gracefully cancel the context */
+		if (engine && !__cancel_engine(engine))
+			/*
+			 * If we are unable to send a preemptive pulse to bump
+			 * the context from the GPU, we have to resort to a full
+			 * reset. We hope the collateral damage is worth it.
+			 */
+			__reset_context(ctx, engine);
+
+		dma_fence_put(fence);
+	}
+}
+
 static void context_close(struct i915_gem_context *ctx)
 {
 	struct i915_address_space *vm;
@@ -291,9 +421,47 @@  static void context_close(struct i915_gem_context *ctx)
 	lut_close(ctx);
 
 	mutex_unlock(&ctx->mutex);
+
+	/*
+	 * If the user has disabled hangchecking, we can not be sure that
+	 * the batches will ever complete after the context is closed,
+	 * keeping the context and all resources pinned forever. So in this
+	 * case we opt to forcibly kill off all remaining requests on
+	 * context close.
+	 */
+	if (!i915_gem_context_is_persistent(ctx) ||
+	    !i915_modparams.enable_hangcheck)
+		kill_context(ctx);
+
 	i915_gem_context_put(ctx);
 }
 
+static int __context_set_persistence(struct i915_gem_context *ctx, bool state)
+{
+	if (i915_gem_context_is_persistent(ctx) == state)
+		return 0;
+
+	if (state) {
+		/*
+		 * Only contexts that are short-lived [that will expire or be
+		 * reset] are allowed to survive past termination. We require
+		 * hangcheck to ensure that the persistent requests are healthy.
+		 */
+		if (!i915_modparams.enable_hangcheck)
+			return -EINVAL;
+
+		i915_gem_context_set_persistence(ctx);
+	} else {
+		/* To cancel a context we use "preempt-to-idle" */
+		if (!(ctx->i915->caps.scheduler & I915_SCHEDULER_CAP_PREEMPTION))
+			return -ENODEV;
+
+		i915_gem_context_clear_persistence(ctx);
+	}
+
+	return 0;
+}
+
 static struct i915_gem_context *
 __create_context(struct drm_i915_private *i915)
 {
@@ -328,6 +496,7 @@  __create_context(struct drm_i915_private *i915)
 
 	i915_gem_context_set_bannable(ctx);
 	i915_gem_context_set_recoverable(ctx);
+	__context_set_persistence(ctx, true /* cgroup hook? */);
 
 	for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++)
 		ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES;
@@ -484,6 +653,7 @@  i915_gem_context_create_kernel(struct drm_i915_private *i915, int prio)
 		return ctx;
 
 	i915_gem_context_clear_bannable(ctx);
+	i915_gem_context_set_persistence(ctx);
 	ctx->sched.priority = I915_USER_PRIORITY(prio);
 
 	GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
@@ -1594,6 +1764,16 @@  get_engines(struct i915_gem_context *ctx,
 	return err;
 }
 
+static int
+set_persistence(struct i915_gem_context *ctx,
+		const struct drm_i915_gem_context_param *args)
+{
+	if (args->size)
+		return -EINVAL;
+
+	return __context_set_persistence(ctx, args->value);
+}
+
 static int ctx_setparam(struct drm_i915_file_private *fpriv,
 			struct i915_gem_context *ctx,
 			struct drm_i915_gem_context_param *args)
@@ -1671,6 +1851,10 @@  static int ctx_setparam(struct drm_i915_file_private *fpriv,
 		ret = set_engines(ctx, args);
 		break;
 
+	case I915_CONTEXT_PARAM_PERSISTENCE:
+		ret = set_persistence(ctx, args);
+		break;
+
 	case I915_CONTEXT_PARAM_BAN_PERIOD:
 	default:
 		ret = -EINVAL;
@@ -2123,6 +2307,11 @@  int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
 		ret = get_engines(ctx, args);
 		break;
 
+	case I915_CONTEXT_PARAM_PERSISTENCE:
+		args->size = 0;
+		args->value = i915_gem_context_is_persistent(ctx);
+		break;
+
 	case I915_CONTEXT_PARAM_BAN_PERIOD:
 	default:
 		ret = -EINVAL;
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.h b/drivers/gpu/drm/i915/gem/i915_gem_context.h
index cfe80590f0ed..18e50a769a6e 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_context.h
+++ b/drivers/gpu/drm/i915/gem/i915_gem_context.h
@@ -76,6 +76,21 @@  static inline void i915_gem_context_clear_recoverable(struct i915_gem_context *c
 	clear_bit(UCONTEXT_RECOVERABLE, &ctx->user_flags);
 }
 
+static inline bool i915_gem_context_is_persistent(const struct i915_gem_context *ctx)
+{
+	return test_bit(UCONTEXT_PERSISTENCE, &ctx->user_flags);
+}
+
+static inline void i915_gem_context_set_persistence(struct i915_gem_context *ctx)
+{
+	set_bit(UCONTEXT_PERSISTENCE, &ctx->user_flags);
+}
+
+static inline void i915_gem_context_clear_persistence(struct i915_gem_context *ctx)
+{
+	clear_bit(UCONTEXT_PERSISTENCE, &ctx->user_flags);
+}
+
 static inline bool i915_gem_context_is_banned(const struct i915_gem_context *ctx)
 {
 	return test_bit(CONTEXT_BANNED, &ctx->flags);
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context_types.h b/drivers/gpu/drm/i915/gem/i915_gem_context_types.h
index fe97b8ba4fda..861d7d92fe9f 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_context_types.h
+++ b/drivers/gpu/drm/i915/gem/i915_gem_context_types.h
@@ -137,6 +137,7 @@  struct i915_gem_context {
 #define UCONTEXT_NO_ERROR_CAPTURE	1
 #define UCONTEXT_BANNABLE		2
 #define UCONTEXT_RECOVERABLE		3
+#define UCONTEXT_PERSISTENCE		4
 
 	/**
 	 * @flags: small set of booleans
diff --git a/drivers/gpu/drm/i915/gem/selftests/mock_context.c b/drivers/gpu/drm/i915/gem/selftests/mock_context.c
index 74ddd682c9cd..29b8984f0e47 100644
--- a/drivers/gpu/drm/i915/gem/selftests/mock_context.c
+++ b/drivers/gpu/drm/i915/gem/selftests/mock_context.c
@@ -22,6 +22,8 @@  mock_context(struct drm_i915_private *i915,
 	INIT_LIST_HEAD(&ctx->link);
 	ctx->i915 = i915;
 
+	i915_gem_context_set_persistence(ctx);
+
 	mutex_init(&ctx->engines_mutex);
 	e = default_engines(ctx);
 	if (IS_ERR(e))
diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
index 63d40cba97e0..b5fd5e4bd16e 100644
--- a/include/uapi/drm/i915_drm.h
+++ b/include/uapi/drm/i915_drm.h
@@ -1572,6 +1572,21 @@  struct drm_i915_gem_context_param {
  *   i915_context_engines_bond (I915_CONTEXT_ENGINES_EXT_BOND)
  */
 #define I915_CONTEXT_PARAM_ENGINES	0xa
+
+/*
+ * I915_CONTEXT_PARAM_PERSISTENCE:
+ *
+ * Allow the context and active rendering to survive the process until
+ * completion. Persistence allows fire-and-forget clients to queue up a
+ * bunch of work, hand the output over to a display server and the quit.
+ * If the context is not marked as persistent, upon closing (either via
+ * an explicit DRM_I915_GEM_CONTEXT_DESTROY or implicitly from file closure
+ * or process termination), the context and any outstanding requests will be
+ * cancelled (and exported fences for cancelled requests marked as -EIO).
+ *
+ * By default, new contexts allow persistence.
+ */
+#define I915_CONTEXT_PARAM_PERSISTENCE	0xb
 /* Must be kept compact -- no holes and well documented */
 
 	__u64 value;