[Mesa-dev] [PATCH 2/2] i965: Be resilient in the face of GPU hangs

[Chris Wilson]

Quoting Chris Wilson (2019-02-14 12:05:00)
> If we hang the GPU and end up banning our context, we will no longer be
> able to submit and abort with an error (exit(1) no less). As we submit
> minimal incremental batches that rely on the logical context state of
> previous batches, we can not rely on the kernel's recovery mechanism
> which tries to restore the context back to a "golden" renderstate (the
> default HW setup) and replay the batches in flight. Instead, we must
> create a new context and set it up, including all the lost register
> settings that we only apply once during setup, before allow the user to
> continue rendering. The batches already submitted are lost
> (unrecoverable) so there will be a momentarily glitch and lost rendering
> across frames, but the application should be able to recover and
> continue on fairly oblivious.
> 
> To make wedging even more likely, we use a new "no recovery" context
> parameter that tells the kernel to not even attempt to replay any
> batches in flight against the default context image, as experience shows
> the HW is not always robust enough to cope with the conflicting state.
> 
> Cc: Kenneth Graunke <kenneth at whitecape.org>
> ---
>  src/mesa/drivers/dri/i965/brw_bufmgr.c        | 25 +++++++++++++++++++
>  src/mesa/drivers/dri/i965/brw_bufmgr.h        |  2 ++
>  src/mesa/drivers/dri/i965/intel_batchbuffer.c | 19 ++++++++++++++
>  3 files changed, 46 insertions(+)
> 
> diff --git a/src/mesa/drivers/dri/i965/brw_bufmgr.c b/src/mesa/drivers/dri/i965/brw_bufmgr.c
> index b33a30930db..289b39cd584 100644
> --- a/src/mesa/drivers/dri/i965/brw_bufmgr.c
> +++ b/src/mesa/drivers/dri/i965/brw_bufmgr.c
> @@ -1589,6 +1589,16 @@ init_cache_buckets(struct brw_bufmgr *bufmgr)
>     }
>  }
>  
> +static void init_context(struct brw_bufmgr *bufmgr, uint32_t ctx_id)
> +{
> +   struct drm_i915_gem_context_param p = {
> +      .ctx_id = ctx_id,
> +      .param = 0x7, // I915_CONTEXT_PARAM_RECOVERABLE,
> +   };
> +
> +   drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CONTEXT_SETPARAM, &p);
> +}
> +
>  uint32_t
>  brw_create_hw_context(struct brw_bufmgr *bufmgr)
>  {
> @@ -1599,6 +1609,8 @@ brw_create_hw_context(struct brw_bufmgr *bufmgr)
>        return 0;
>     }
>  
> +   init_context(bufmgr, create.ctx_id);
> +
>     return create.ctx_id;
>  }
>  
> @@ -1621,6 +1633,19 @@ brw_hw_context_set_priority(struct brw_bufmgr *bufmgr,
>     return err;
>  }
>  
> +int
> +brw_hw_context_get_priority(struct brw_bufmgr *bufmgr, uint32_t ctx_id)
> +{
> +   struct drm_i915_gem_context_param p = {
> +      .ctx_id = ctx_id,
> +      .param = I915_CONTEXT_PARAM_PRIORITY,
> +   };
> +
> +   drmIoctl(bufmgr->fd, DRM_IOCTL_I915_GEM_CONTEXT_GETPARAM, &p);
> +
> +   return p.value; /* on error, return 0 i.e. default priority */
> +}
> +
>  void
>  brw_destroy_hw_context(struct brw_bufmgr *bufmgr, uint32_t ctx_id)
>  {
> diff --git a/src/mesa/drivers/dri/i965/brw_bufmgr.h b/src/mesa/drivers/dri/i965/brw_bufmgr.h
> index 32fc7a553c9..886b2e607ce 100644
> --- a/src/mesa/drivers/dri/i965/brw_bufmgr.h
> +++ b/src/mesa/drivers/dri/i965/brw_bufmgr.h
> @@ -356,6 +356,8 @@ uint32_t brw_create_hw_context(struct brw_bufmgr *bufmgr);
>  int brw_hw_context_set_priority(struct brw_bufmgr *bufmgr,
>                                  uint32_t ctx_id,
>                                  int priority);
> +int
> +brw_hw_context_get_priority(struct brw_bufmgr *bufmgr, uint32_t ctx_id);
>  
>  void brw_destroy_hw_context(struct brw_bufmgr *bufmgr, uint32_t ctx_id);
>  
> diff --git a/src/mesa/drivers/dri/i965/intel_batchbuffer.c b/src/mesa/drivers/dri/i965/intel_batchbuffer.c
> index 8097392d22b..afb6e2401e3 100644
> --- a/src/mesa/drivers/dri/i965/intel_batchbuffer.c
> +++ b/src/mesa/drivers/dri/i965/intel_batchbuffer.c
> @@ -748,6 +748,18 @@ execbuffer(int fd,
>     return ret;
>  }
>  
> +static void recreate_context(struct brw_context *brw)
> +{
> +   struct brw_bufmgr *bufmgr = brw->bufmgr;
> +   int prio;
> +
> +   prio = brw_hw_context_get_priority(bufmgr, brw->hw_ctx);
> +   brw_destroy_hw_context(bufmgr, brw->hw_ctx);
> +
> +   brw->hw_ctx = brw_create_hw_context(bufmgr);
> +   brw_hw_context_set_priority(bufmgr, brw->hw_ctx, prio);

Hmm, fwiw we can make this into a clone operation in the kernel. That
way, we can also do things like move across the ppgtt. We will have the
ability to that shortly...
-Chris