Re: [Intel-gfx] [PATCH] drm/i915: avoid concurrent writes to aux_inv

Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxxxxxxx> · Wed, 16 Mar 2022 10:03:22 +0000

On 04/03/2022 22:14, fei.yang@xxxxxxxxx wrote:
From: Fei Yang <fei.yang@xxxxxxxxx>

GPU hangs have been observed when multiple engines write to the
same aux_inv register at the same time. To avoid this each engine
should only invalidate its own auxiliary table. The function
gen12_emit_flush_xcs() currently invalidate the auxiliary table for
all engines because the rq->engine is not necessarily the engine
eventually carrying out the request, and potentially the engine
could even be a virtual one (with engine->instance being -1).
With this patch, auxiliary table invalidation is done only for the
engine executing the request. And the mmio address for the aux_inv
register is set after the engine instance becomes certain.

Signed-off-by: Chris Wilson <chris.p.wilson@xxxxxxxxx>
Signed-off-by: Fei Yang <fei.yang@xxxxxxxxx>
---
  drivers/gpu/drm/i915/gt/gen2_engine_cs.c      |  9 +++
  drivers/gpu/drm/i915/gt/gen6_engine_cs.c      |  9 +++
  drivers/gpu/drm/i915/gt/gen8_engine_cs.c      | 61 ++++++++-----------
  .../drm/i915/gt/intel_execlists_submission.c  | 35 +++++++++++
  drivers/gpu/drm/i915/i915_request.h           |  2 +
  5 files changed, 82 insertions(+), 34 deletions(-)

diff --git a/drivers/gpu/drm/i915/gt/gen2_engine_cs.c b/drivers/gpu/drm/i915/gt/gen2_engine_cs.c
index 1c82caf525c3..0ec4986e4805 100644
--- a/drivers/gpu/drm/i915/gt/gen2_engine_cs.c
+++ b/drivers/gpu/drm/i915/gt/gen2_engine_cs.c
@@ -37,6 +37,9 @@ int gen2_emit_flush(struct i915_request *rq, u32 mode)
  
  	intel_ring_advance(rq, cs);
  
+	/* hsdes: 1809175790. No fixup needed for gen2 */
+	rq->aux_inv_fixup = NULL;

Same thing that Stuart mentioned - would it not work for instance to 
initialize this in __i915_request_create?

+
  	return 0;
  }
  
@@ -123,6 +126,9 @@ int gen4_emit_flush_rcs(struct i915_request *rq, u32 mode)
  
  	intel_ring_advance(rq, cs);
  
+	/* hsdes: 1809175790. No fixup needed for gen4 rcs */
+	rq->aux_inv_fixup = NULL;
+
  	return 0;
  }
  
@@ -138,6 +144,9 @@ int gen4_emit_flush_vcs(struct i915_request *rq, u32 mode)
  	*cs++ = MI_NOOP;
  	intel_ring_advance(rq, cs);
  
+	/* hsdes: 1809175790. No fixup needed for gen4 vcs */
+	rq->aux_inv_fixup = NULL;
+
  	return 0;
  }
  
diff --git a/drivers/gpu/drm/i915/gt/gen6_engine_cs.c b/drivers/gpu/drm/i915/gt/gen6_engine_cs.c
index 5e65550b4dfb..efe51c4662fe 100644
--- a/drivers/gpu/drm/i915/gt/gen6_engine_cs.c
+++ b/drivers/gpu/drm/i915/gt/gen6_engine_cs.c
@@ -137,6 +137,9 @@ int gen6_emit_flush_rcs(struct i915_request *rq, u32 mode)
  	*cs++ = 0;
  	intel_ring_advance(rq, cs);
  
+	/* hsdes: 1809175790. No fixup needed for gen6 */
+	rq->aux_inv_fixup = NULL;
+
  	return 0;
  }
  
@@ -208,6 +211,9 @@ static int mi_flush_dw(struct i915_request *rq, u32 flags)
  
  	intel_ring_advance(rq, cs);
  
+	/* hsdes: 1809175790. No fixup needed for gen6 */
+	rq->aux_inv_fixup = NULL;
+
  	return 0;
  }
  
@@ -347,6 +353,9 @@ int gen7_emit_flush_rcs(struct i915_request *rq, u32 mode)
  	*cs++ = 0;
  	intel_ring_advance(rq, cs);
  
+	/* hsdes: 1809175790. No fixup needed for gen7 rcs */
+	rq->aux_inv_fixup = NULL;
+
  	return 0;
  }
  
diff --git a/drivers/gpu/drm/i915/gt/gen8_engine_cs.c b/drivers/gpu/drm/i915/gt/gen8_engine_cs.c
index b1b9c3fd7bf9..b6374cf53314 100644
--- a/drivers/gpu/drm/i915/gt/gen8_engine_cs.c
+++ b/drivers/gpu/drm/i915/gt/gen8_engine_cs.c
@@ -73,6 +73,9 @@ int gen8_emit_flush_rcs(struct i915_request *rq, u32 mode)
  
  	intel_ring_advance(rq, cs);
  
+	/* hsdes: 1809175790. No fixup needed for gen8 rcs */
+	rq->aux_inv_fixup = NULL;
+
  	return 0;
  }
  
@@ -106,6 +109,9 @@ int gen8_emit_flush_xcs(struct i915_request *rq, u32 mode)
  	*cs++ = 0; /* value */
  	intel_ring_advance(rq, cs);
  
+	/* hsdes: 1809175790. No fixup needed for gen8 xcs */
+	rq->aux_inv_fixup = NULL;
+
  	return 0;
  }
  
@@ -157,6 +163,9 @@ int gen11_emit_flush_rcs(struct i915_request *rq, u32 mode)
  		intel_ring_advance(rq, cs);
  	}
  
+	/* hsdes: 1809175790. No fixup needed for gen11 rcs */
+	rq->aux_inv_fixup = NULL;
+
  	return 0;
  }
  
@@ -165,30 +174,6 @@ static u32 preparser_disable(bool state)
  	return MI_ARB_CHECK | 1 << 8 | state;
  }
  
-static i915_reg_t aux_inv_reg(const struct intel_engine_cs *engine)
-{
-	static const i915_reg_t vd[] = {
-		GEN12_VD0_AUX_NV,
-		GEN12_VD1_AUX_NV,
-		GEN12_VD2_AUX_NV,
-		GEN12_VD3_AUX_NV,
-	};
-
-	static const i915_reg_t ve[] = {
-		GEN12_VE0_AUX_NV,
-		GEN12_VE1_AUX_NV,
-	};
-
-	if (engine->class == VIDEO_DECODE_CLASS)
-		return vd[engine->instance];
-
-	if (engine->class == VIDEO_ENHANCEMENT_CLASS)
-		return ve[engine->instance];
-
-	GEM_BUG_ON("unknown aux_inv reg\n");
-	return INVALID_MMIO_REG;
-}
-
  static u32 *gen12_emit_aux_table_inv(const i915_reg_t inv_reg, u32 *cs)
  {
  	*cs++ = MI_LOAD_REGISTER_IMM(1);
@@ -274,6 +259,9 @@ int gen12_emit_flush_rcs(struct i915_request *rq, u32 mode)
  		intel_ring_advance(rq, cs);
  	}
  
+	/* hsdes: 1809175790. aux_inv has been done above, no fixup needed */
+	rq->aux_inv_fixup = NULL;
+
  	return 0;
  }
  
@@ -288,7 +276,7 @@ int gen12_emit_flush_xcs(struct i915_request *rq, u32 mode)
  	if (mode & EMIT_INVALIDATE)
  		aux_inv = rq->engine->mask & ~BIT(BCS0);
  	if (aux_inv)
-		cmd += 2 * hweight32(aux_inv) + 2;
+		cmd += 4;
  
  	cs = intel_ring_begin(rq, cmd);
  	if (IS_ERR(cs))
@@ -319,16 +307,21 @@ int gen12_emit_flush_xcs(struct i915_request *rq, u32 mode)
  	*cs++ = 0; /* value */
  
  	if (aux_inv) { /* hsdes: 1809175790 */
-		struct intel_engine_cs *engine;
-		unsigned int tmp;
-
-		*cs++ = MI_LOAD_REGISTER_IMM(hweight32(aux_inv));
-		for_each_engine_masked(engine, rq->engine->gt, aux_inv, tmp) {
-			*cs++ = i915_mmio_reg_offset(aux_inv_reg(engine));
-			*cs++ = AUX_INV;
-		}
+		/*
+		 * We don't know which engine will eventually carry out
+		 * this request, so the mmio aux_inv register address is
+		 * unknown at this moment. We save the cs pointer supposed
+		 * to hold the aux_inv address in rq->aux_inv_fixup and set
+		 * it in execlists_dequeue() when the engine instance
+		 * carrying out this request becomes certain
+		 */
+		*cs++ = MI_LOAD_REGISTER_IMM(1);
+		rq->aux_inv_fixup = cs; /* save the pointer to aux_inv */
+		*cs++ = 0; /* mmio addr to be set at submission to HW */
+		*cs++ = AUX_INV;
  		*cs++ = MI_NOOP;
-	}
+	} else
+		rq->aux_inv_fixup = NULL;
  
  	if (mode & EMIT_INVALIDATE)
  		*cs++ = preparser_disable(false);
diff --git a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
index e1470bb60f34..7e8552414275 100644
--- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
+++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
@@ -1258,6 +1258,34 @@ static bool completed(const struct i915_request *rq)
  	return __i915_request_is_complete(rq);
  }
  
+static i915_reg_t aux_inv_reg(const struct intel_engine_cs *engine)
+{
+	static const i915_reg_t vd[] = {
+		GEN12_VD0_AUX_NV,
+		GEN12_VD1_AUX_NV,
+		GEN12_VD2_AUX_NV,
+		GEN12_VD3_AUX_NV,
+	};
+
+	static const i915_reg_t ve[] = {
+		GEN12_VE0_AUX_NV,
+		GEN12_VE1_AUX_NV,
+	};
+
+	if (engine->class == VIDEO_DECODE_CLASS) {
+		GEM_BUG_ON(engine->instance >= ARRAY_SIZE(vd));
+		return vd[engine->instance];
+	}
+
+	if (engine->class == VIDEO_ENHANCEMENT_CLASS) {
+		GEM_BUG_ON(engine->instance >= ARRAY_SIZE(ve));
+		return ve[engine->instance];
+	}
+
+	GEM_BUG_ON("unknown aux_inv reg\n");
+	return INVALID_MMIO_REG;
+}
+
  static void execlists_dequeue(struct intel_engine_cs *engine)
  {
  	struct intel_engine_execlists * const execlists = &engine->execlists;
@@ -1538,6 +1566,13 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
  			}
  
  			if (__i915_request_submit(rq)) {
+				/* hsdes: 1809175790 */
+				if (rq->aux_inv_fixup) {
+					*rq->aux_inv_fixup =
+						i915_mmio_reg_offset
+							(aux_inv_reg(engine));
+					rq->execution_mask = engine->mask;

Execution mask assignment needs a comment explaining what, why, and what 
are the consequences of fixing the mask here.

Also, from the top of my head for instance I can't remember why would 
aux_inv_fixup handling be just here and not in the first/virtual dequeue 
loop (the "while ... first_virtual_engine" one)? I thought the 2nd loop 
is non-virtual requests, but I am probably missing something.

+				}
  				if (!merge) {
  					*port++ = i915_request_get(last);
  					last = NULL;
diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h
index 28b1f9db5487..8e0764c43936 100644
--- a/drivers/gpu/drm/i915/i915_request.h
+++ b/drivers/gpu/drm/i915/i915_request.h
@@ -312,6 +312,8 @@ struct i915_request {
  	struct i915_capture_list *capture_list;
  #endif
  
+	u32 *aux_inv_fixup; /* hsdes: 1809175790 */
+
  	/** Time at which this request was emitted, in jiffies. */
  	unsigned long emitted_jiffies;
  

Question of affected platforms and how will this be handled in case of 
GuC? I don't see how it can without the firmware involvement. Only if 
i915 would nerf the virtual engine setup so each GuC virtual context 
gets only one engine from the set, but that would be beyond horrible. 
Best to involved GuC people here.

Regards,

Tvrtko






