drm/i915: don't track relative-constants-mode
Subsystems (2)
PATCH 1/1 drm/i915: don't track relative-constants-mode
'relative_constants_mode' has always been tracked per-device, but this has actually been wrong ever since hardware contexts were introduced, as the INSTPM register is saved (and automatically restored) as part of the render ring context. The software per-device value could therefore get out of sync with the hardware per-context value.
Rather than track this correctly (per-context in LRC modes, per-device in legacy ringbuffer mode), a simpler solution is just to give up trying to track it at all, and always prefix each render batch with the single instruction needed to explicitly set it to the required mode for the current batch.
Test case (if anyone wants to write it) would be to create two contexts, submit a batch with a non-default mode in the first context, submit a batch with the default mode in the other context, submit another batch in the first context, but this time in default mode. Without this patch, the driver will fail to insert the instructions to reset INSTPM into the first context's ringbuffer.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92448
--- drivers/gpu/drm/i915/i915_drv.h | 2 -- drivers/gpu/drm/i915/i915_gem.c | 3 --- drivers/gpu/drm/i915/i915_gem_execbuffer.c | 32 +++++++++++++----------------- 3 files changed, 14 insertions(+), 23 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +0 -2
@@ -1728,8 +1728,6 @@ struct drm_i915_private {const struct intel_device_info info; -int relative_constants_mode; -void __iomem *regs;struct intel_uncore uncore;
drivers/gpu/drm/i915/i915_gem.c +0 -3
@@ -4615,9 +4615,6 @@ int i915_gem_init(struct drm_device *dev) i915_gem_idle_work_handler);init_waitqueue_head (&dev_priv->gpu_error.wait_queue);init_waitqueue_head (&dev_priv->gpu_error.reset_queue); - - dev_priv->relative_constants_mode =I915_EXEC_CONSTANTS_REL_GENERAL ; -init_waitqueue_head (&dev_priv->pending_flip_queue); dev_priv->mm.interruptible = true;
drivers/gpu/drm/i915/i915_gem_execbuffer.c +14 -18
@@ -1462,30 +1462,28 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,return -EINVAL ; } -if (instp_mode != dev_priv->relative_constants_mode) { -if (INTEL_INFO (dev_priv)->gen <4 ) { -DRM_DEBUG ("no rel constants on pre-gen4\n" ); -return -EINVAL ; - } - -if (INTEL_INFO (dev_priv)->gen >5 && - instp_mode ==I915_EXEC_CONSTANTS_REL_SURFACE ) { -DRM_DEBUG ("rel surface constants mode invalid on gen5+\n" ); -return -EINVAL ; - } +if (INTEL_INFO (dev_priv)->gen <4 ) { +DRM_DEBUG ("no rel constants on pre-gen4\n" ); +return -EINVAL ; + } -/* The HW changed the meaning on this bit on gen6 */ -if (INTEL_INFO (dev_priv)->gen >=6 ) - instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE ; +if (INTEL_INFO (dev_priv)->gen >5 && + instp_mode ==I915_EXEC_CONSTANTS_REL_SURFACE ) { +DRM_DEBUG ("rel surface constants mode invalid on gen5+\n" ); +return -EINVAL ; } + +/* The HW changed the meaning on this bit on gen6 */ +if (INTEL_INFO (dev_priv)->gen >=6 ) + instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE ; +break ;default :DRM_DEBUG ("execbuf with unknown constants: %d\n" , instp_mode);return -EINVAL ; } -if (params->engine->id ==RCS && - instp_mode != dev_priv->relative_constants_mode) { +if (params->engine->id ==RCS ) {struct intel_ring *ring = params->request->ring; ret =intel_ring_begin (params->request,4 );
@@ -1497,8 +1495,6 @@ static void eb_export_fence(struct drm_i915_gem_object *obj,intel_ring_emit_reg (ring,INSTPM );intel_ring_emit (ring, instp_mask <<16 | instp_mode);intel_ring_advance (ring); - - dev_priv->relative_constants_mode = instp_mode; }if (args->flags &I915_EXEC_GEN7_SOL_RESET ) {
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 11 lines · Dave Gordon
'
relative_constants_mode' has always been tracked per-device, but this has actually been wrong ever since hardware contexts were introduced, as the INSTPM register is saved (and automatically restored) as part of the render ring context. The software per-device value could therefore get out of sync with the hardware per-context value.Rather than track this correctly (per-context in LRC modes, per-device in legacy ringbuffer mode), a simpler solution is just to give up trying to track it at all, and always prefix each render batch with the single instruction needed to explicitly set it to the required mode for the current batch.
Agreed. However always emitting it wasteful for those who never touch it, i.e. everybody. Rather than track the last value, just track if it ever changed from default then reassign it before every batch?
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1 -1
@@ -1578,7 +1578,7 @@ execbuf_submit(struct i915_execbuffer_params *params,intel_ring_emit (ring, instp_mask <<16 | instp_mode);intel_ring_advance (ring); - dev_priv->relative_constants_mode = instp_mode; + dev_priv->relative_constants_mode =-1 ; }if (args->flags &I915_EXEC_GEN7_SOL_RESET ) {
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 13 lines · Chris Wilson
Quoted 11 lines · Dave Gordon
'
relative_constants_mode' has always been tracked per-device, but this has actually been wrong ever since hardware contexts were introduced, as the INSTPM register is saved (and automatically restored) as part of the render ring context. The software per-device value could therefore get out of sync with the hardware per-context value.Rather than track this correctly (per-context in LRC modes, per-device in legacy ringbuffer mode), a simpler solution is just to give up trying to track it at all, and always prefix each render batch with the single instruction needed to explicitly set it to the required mode for the current batch.
Quoted 3 lines
Agreed. However always emitting it wasteful for those who never touch it, i.e. everybody. Rather than track the last value, just track if it ever changed from default then reassign it before every batch?
And since nothing released has ever used it:
drivers/gpu/drm/i915/i915_drv.c +1 -0
@@ -283,7 +283,8 @@ static int i915_getparam(struct drm_device *dev, void *data, value =1 ;break ;case I915_PARAM_HAS_EXEC_CONSTANTS : +WARN_ONCE (1 ,"Planned obsolence: please report to bugs.freedesktop.org\n" ); value =INTEL_INFO (dev)->gen >=4 ;break ;case I915_PARAM_HAS_RELAXED_DELTA : value =1 ;
and deleting the code later would be an option.
-Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 13 lines · Chris Wilson
Quoted 11 lines · Dave Gordon
'
relative_constants_mode' has always been tracked per-device, but this has actually been wrong ever since hardware contexts were introduced, as the INSTPM register is saved (and automatically restored) as part of the render ring context. The software per-device value could therefore get out of sync with the hardware per-context value.Rather than track this correctly (per-context in LRC modes, per-device in legacy ringbuffer mode), a simpler solution is just to give up trying to track it at all, and always prefix each render batch with the single instruction needed to explicitly set it to the required mode for the current batch.
Quoted 3 lines
Agreed. However always emitting it wasteful for those who never touch it, i.e. everybody. Rather than track the last value, just track if it ever changed from default then reassign it before every batch?
But where to track whether it changed? If tracked globally, that will mean everyone pays if any one process ever submits a batch with a non-default value. If not global, then we still have the discrepancy between LRC modes (h/w value saved per-context) vs ringbuffer (maybe saved if device supports "hardware contexts", not saved otherwise).
Perhaps instead we should decide that between batches it shall always be left at the default value, so any batch that wants any other value must set it at the beginning and reset it when finished?
.Dave.
[PS: the reason I spotted this is because the tracking was broken with pre-emption enabled: you can no longer assume that the mode at the start of each batch is the same mode that the preceding batch used.]
Quoted 7 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +0
-0
drivers/gpu/drm/i915/i915_gem_execbuffer.c execbuf_submit(struct i915_execbuffer_params *params, · line 1584
Quoted 8 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1
-1@@ -1578,7 +1578,7 @@ execbuf_submit(struct i915_execbuffer_params *params,intel_ring_emit (ring, instp_mask <<16 | instp_mode);intel_ring_advance (ring); - dev_priv->relative_constants_mode = instp_mode; + dev_priv->relative_constants_mode =-1 ; }if (args->flags &I915_EXEC_GEN7_SOL_RESET ) {
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx